**NEUROPSYCHOPHARMACOLOGY OF PSYCHOSIS: RELATION OF BRAIN SIGNALS, COGNITION AND CHEMISTRY**

**Topic Editors André Schmidt and Stefan Borgwardt**

#### *FRONTIERS COPYRIGHT STATEMENT*

© Copyright 2007-2015 Frontiers Media SA. All rights reserved.

All content included on this site, such as text, graphics, logos, button icons, images, video/audio clips, downloads, data compilations and software, is the property of or is licensed to Frontiers Media SA ("Frontiers") or its licensees and/or subcontractors. The copyright in the text of individual articles is the property of their respective authors, subject to a license granted to Frontiers.

The compilation of articles constituting this e-book, wherever published, as well as the compilation of all other content on this site, is the exclusive property of Frontiers. For the conditions for downloading and copying of e-books from Frontiers' website, please see the Terms for Website Use. If purchasing Frontiers e-books from other websites or sources, the conditions of the website concerned apply.

Images and graphics not forming part of user-contributed materials may not be downloaded or copied without permission.

Individual articles may be downloaded and reproduced in accordance with the principles of the CC-BY licence subject to any copyright or other notices. They may not be re-sold as an e-book.

As author or other contributor you grant a CC-BY licence to others to reproduce your articles, including any graphics and third-party materials supplied by you, in accordance with the Conditions for Website Use and subject to any copyright notices which you include in connection with your articles and materials.

All copyright, and all rights therein, are protected by national and international copyright laws.

The above represents a summary only. For the full conditions see the Conditions for Authors and the Conditions for Website Use.

Cover image provided by Ibbl sarl, Lausanne CH

**ISSN** 1664-8714 **ISBN** 978-2-88919-335-6 **DOI** 10.3389/978-2-88919-335-6

## *ABOUT FRONTIERS*

Frontiers is more than just an open-access publisher of scholarly articles: it is a pioneering approach to the world of academia, radically improving the way scholarly research is managed. The grand vision of Frontiers is a world where all people have an equal opportunity to seek, share and generate knowledge. Frontiers provides immediate and permanent online open access to all its publications, but this alone is not enough to realize our grand goals.

## *FRONTIERS JOURNAL SERIES*

The Frontiers Journal Series is a multi-tier and interdisciplinary set of open-access, online journals, promising a paradigm shift from the current review, selection and dissemination processes in academic publishing.

All Frontiers journals are driven by researchers for researchers; therefore, they constitute a service to the scholarly community. At the same time, the Frontiers Journal Series operates on a revolutionary invention, the tiered publishing system, initially addressing specific communities of scholars, and gradually climbing up to broader public understanding, thus serving the interests of the lay society, too.

## *DEDICATION TO QUALITY*

Each Frontiers article is a landmark of the highest quality, thanks to genuinely collaborative interactions between authors and review editors, who include some of the world's best academicians. Research must be certified by peers before entering a stream of knowledge that may eventually reach the public - and shape society; therefore, Frontiers only applies the most rigorous and unbiased reviews.

Frontiers revolutionizes research publishing by freely delivering the most outstanding research, evaluated with no bias from both the academic and social point of view.

By applying the most advanced information technologies, Frontiers is catapulting scholarly publishing into a new generation.

## *WHAT ARE FRONTIERS RESEARCH TOPICS?*

Frontiers Research Topics are very popular trademarks of the Frontiers Journals Series: they are collections of at least ten articles, all centered on a particular subject. With their unique mix of varied contributions from Original Research to Review Articles, Frontiers Research Topics unify the most influential researchers, the latest key findings and historical advances in a hot research area!

Find out more on how to host your own Frontiers Research Topic or contribute to one as an author by contacting the Frontiers Editorial Office: researchtopics@frontiersin.org

# **NEUROPSYCHOPHARMACOLOGY OF PSYCHOSIS: RELATION OF BRAIN SIGNALS, COGNITION AND CHEMISTRY**

Topic Editors:

**André Schmidt,** University of Basel, Switzerland **Stefan Borgwardt,** University of Basel, Switzerland

Understanding complex mental illness from synapses to circuits to neural systems. Taken from Anticevic et al. (2013), Connectivity, pharmacology, and computation: toward a mechanistic understanding of neural system dysfunction in schizophrenia, Front. Psychiatry, doi: 10.3389/fpsyt.2013.00169.

# Table of Contents



Judith M. Ford and Deepak Cyril D'Souza *169 The Computational Anatomy of Psychosis*

Rick A. Adams, Klaas Enno Stephan, Harriet R. Brown, Christopher D. Frith and Karl J. Friston

*195 Reinforcement Learning and Dopamine in Schizophrenia: Dimensions of Symptoms or Specific Features of a Disease Group?*

Lorenz Deserno, Rebecca Boehme, Andreas Heinz and Florian Schlagenhauf

*211 Computational Neuropsychiatry – Schizophrenia as a Cognitive Brain Network Disorder*

Maria R. Dauvermann, Heather C. Whalley, André Schmidt, Graham L. Lee, Liana Romaniuk, Neil Roberts, Eve C. Johnstone, Stephen M. Lawrie and Thomas W. J. Moorhead

## *230 Dysconnectivity in the Frontoparietal Attention Network in Schizophrenia* Jonathan P. Roiser, Rebekah Wigton, James M. Kilner, Maria A. Mendez, Nicholas Hon, Karl J. Friston, Eileen M. Joyce

*243 Connectivity, Pharmacology, and Computation: Toward a Mechanistic Understanding of Neural System Dysfunction in Schizophrenia* Alan Anticevic, Michael W. Cole, Grega Repovs, Aleksandar Savic, Naomi R. Driesen, Genevieve Yang, Youngsun T. Cho, John D. Murray, David C. Glahn, Xiao-Jing Wang

and John H. Krystal, *264 Dysfunction and Dysconnection in Cortical–Striatal Networks During Sustained Attention: Genetic Risk for Schizophrenia or Bipolar Disorder and* **I***ts Impact on* 

*Brain Network Function* Vaibhav A. Diwadkar, Neil Bakshi, Gita Gupta, Patrick Pruitt, Richard White and Simon B. Eickhoff

# Neuropsychopharmacology of psychosis: relation of brain signals, cognition, and chemistry

## **André Schmidt \* and Stefan Borgwardt**

Department of Psychiatry (UPK), University of Basel, Basel, Switzerland \*Correspondence: andre.schmidt@unibas.ch

**Edited and reviewed by:**

Mihaly Hajos, Yale University School of Medicine, USA

**Keywords: psychosis, psychosis high-risk state, cognition, neuroimaging, pharmacology, computational psychiatry, brain connectivity**

Recent research has resumed the pivotal significance of cognitive impairmentsfor the development of psychosis spectrum disorders, proposing a shift in focus extending from psychotic symptoms as the key hallmarks (1, 2). Cognitive deficits are of particular interest as they precede the onset of psychosis by many years in the absence of any psychotic symptom and thus providing valuable predictions about the longitudinal course (3). Recognizing cognitive disturbances as the main promoter may allow early detection of the illness and might also lead to adequate and effective treatment.

Neuroscientific brain imaging techniques have essentially helped putting the attention back on cognition. In this research topic, we aimed at emphasizing the tremendous relevance of cognitive impairments for the early stages of psychosis and their neurobiological correlates as measured with a broad variety of brain imaging modalities such as electro- and magnetoencephalography, structural, functional, and resting state magnetic resonance imaging, near-infrared spectroscopy, and proton magnetic resonance spectroscopy. The topic begins with articles emphasizing the evidence of cognitive deficits in patients with schizophrenia, first-episode psychosis, and in personsfrom the general population with psychosis-like experiences and whether they are mirrored in brain signals as measured byfunctional near-infrared spectroscopy or structural magnetic resonance imaging (4–7). Further works review the underlying neuropharmacological mechanisms of cognitive impairments by focusing on different established domains and discuss potential drug targets for cognitive enhancement treatments (8, 9). This research topic also highlights the significance of the *N*-methyl-d-aspartate receptor for the development of psychosis and how glutamatergic metabolites are related to symptoms and cognitive function in clinical samples, suggesting promising new target pathways for the treatment of psychosis (10–12). Furthermore, electrophysiological modeling strategies in animals (13) and healthy subjects (14–16) are presented, which might help to establish neurobiological markers not only for the treatment of cognitive deficits but also for the prediction of psychosis and the development of preventive treatment schemes. The largest part of this issue unifies theoretical and experimental evidence reflecting the immense potential of computational neuroscience for shedding new light on the neurophysiological mechanisms underlying psychosis in general and on the formation of specific psychopathological signs and symptoms in particular. It starts with a normative consideration of psychotic symptoms as a result of

aberrant encoding of precision embedded with predictive coding framework (17). The topic ends up with several computational modeling approaches and reviews addressing the relation between neural network properties, pharmacological challenges, cognition, and genetic risk (18–22).

This issue is intended to provide a state-of-the-art cognitive perspective to consider developing psychosis and will serve as useful framework for further investigations inferring pathophysiological mechanisms of psychosis. Such sorts of analyses might help to predate the onset of psychosis in terms of abnormal brain signals and to improve and develop new therapeutic scenarios. We would like to thank all the authors and reviewers for their valuable contributions, as well as the Editorial Office for their help in the editing process.

## **REFERENCES**


of neural system dysfunction in schizophrenia. *Front Psychiatry* (2013) **4**:169. doi:10.3389/fpsyt.2013.00169


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 19 May 2014; accepted: 13 June 2014; published online: 01 July 2014.*

*Citation: Schmidt A and Borgwardt S (2014) Neuropsychopharmacology of psychosis: relation of brain signals, cognition, and chemistry. Front. Psychiatry 5:76. doi: 10.3389/fpsyt.2014.00076*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2014 Schmidt and Borgwardt. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

## Working memory capacity and psychotic-like experiences in a general population sample of adolescents and young adults

## **Tim B. Ziermans 1,2\***

<sup>1</sup> Department of Clinical Child and Adolescent Studies, Leiden University, Leiden, Netherlands

<sup>2</sup> Department of Neuroscience, Karolinska Institutet, Stockholm Brain Institute, Stockholm, Sweden

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Stella G. Giakoumaki, University of Crete, Greece Ricardo Carrion, North Shore-LIJ Health System, USA

#### **\*Correspondence:**

Tim B. Ziermans, Department of Clinical Child and Adolescent Studies, Leiden University, Postbus 9555, 2300 RB Leiden, Netherlands e-mail: t.b.ziermans@ fsw.leidenuniv.nl

Working memory (WM) impairment is a common feature in individuals with schizophrenia and high-risk for psychosis and a promising target for early intervention strategies. However, it is unclear to what extent WM impairment parallels specific behavioral symptoms along the psychosis continuum. To address this issue, the current study investigated the relation of WM capacity with psychotic-like experiences (PLEs) in a large Swedish population sample (N = 1012) of adolescents and young adults (M = 24.4 years, range 12–35). WM was assessed with two online computer tasks: a task where participants had to identify and remember the location of an odd shape and a task of remembering and following instructions. PLE scores were derived from a translated symptom questionnaire (Community Assessment of Psychic Experiences), which includes positive, negative, and depressive symptom scales. Positive and negative symptom scales were further subdivided into symptom clusters based on factor analyses. The results showed that low WM capacity was modestly associated with increased reports of bizarre experiences (BE) and depressive symptoms, after controlling for age, gender, and global symptom scores. Interestingly, when analyses were repeated for separate age groups, low WM was exclusively associated with a higher frequency of BE for young adults (20–27 years) and with depressive symptoms for older adults (28–35 years). These findings suggest that specific PLEs can be indicative of reduced WM capacity in early adulthood, which in turn may reflect an increased risk for psychosis and a greater need for targeted intervention. In contrast, during adolescence individual differences in cognitive development may influence the strength of the relationships and thereby mask potential vulnerabilities for psychopathology.

**Keywords: working memory, PLEs, schizotypy, psychosis proneness, CAPE, adolescence, internet**

## **INTRODUCTION**

Working memory (WM) refers to the cognitive function to retain and manipulate information over a brief period of time. Intact WM relies heavily on the dopaminergic regulation of neuronal networks, particularly in the prefrontal cortex [e.g., Ref. (1–3)]. Disrupted regulation has commonly been associated with behavioral expressions observed in a variety of neuropsychiatric disorders, including schizophrenia. Since the groundbreaking work of Goldman-Rakic and colleagues on prefrontal lobe function and the cellular mechanisms of WM (4, 5) the scientific community has largely embraced the notion that WM dysfunction constitutes a fundamental neurocognitive impairment pivotal to the pathogenesis of schizophrenia.

Indeed, meta-analytic evidence supports that impairments in WM, as well as in a wide range of other neurocognitive domains, are a common feature among individuals suffering from schizophrenia (6), schizoaffective disorders (7) and, to a lesser extent, young individuals at clinical high-risk for developing a psychotic disorder (8). The latter findings suggest that WM impairments may already be present during the earliest stages of the disorder when the first symptoms are starting to appear. However, most

high-risk studies only recruit help-seeking individuals and rely on categorical group comparisons defined by clinical cut-off scores. These strategies are unfit to address whether WM is associated with sub-clinical symptomatology in the population at large as well. Given the recent surge of interest in WM training programs as an intervention aid in early schizophrenia and high-risk populations (9, 10), it is critical to improve our understanding of the putative relationship between WM performance and psychotic-like symptomatology in order to optimize their use and efficacy.

The extended psychosis phenotype refers to the observation that psychotic symptoms or "psychotic-like experiences" (PLEs) exist on a continuum in the general population with clinical schizophrenia on one end and mild, non-clinical schizotypy on the other (11, 12). A recent estimate indicates that the prevalence of PLEs is 7.2% in the general population and for approximately 20% of these individuals the symptoms will persist over time (13). Although PLEs do not necessarily cause distress or affect daily functioning for a majority of individuals, epidemiological studies do indicate that a high intensity of PLEs is associated with an increased risk for psychosis (14–16). Interestingly, prevalence of symptoms may differ across genders and appears to be higher in adolescence than in adulthood (17, 18), a period when the first PLEs typically begin to appear and WM skills are still under maturation (19). Both gender and age may therefore represent moderating variables of the relation between WM and PLEs.

Very few large-scale studies have examined the relationship between WM and PLEs in non-clinical populations, in part due to the time- and resource-consuming aspects of on-site neurocognitive assessments. In addition, available studies have typically focused on one of three classical schizophrenia symptom groups: positive (e.g., hallucinations, delusions), negative [e.g., blunted affect, social withdrawal (SW)], or disorganized (e.g., odd speech and behavior) symptoms. While most findings indicate that reduced WM capacity is associated with increased symptoms, results are mixed for positive and negative symptoms (20–22). This could partly be due to use of different WM measures and clinical instruments, insufficient differentiation of symptoms clusters or sample bias (most studies recruited students). Consequently, it remains unclear whether WM is more commonly associated with specific schizotypal features.

In the current study the relationship between WM capacity and PLEs was further investigated with a fully automated online assessment procedure in a general population sample of Swedish adolescents and young adults. Previous population studies have demonstrated that PLEs do not represent a homogenous entity and can be divided into different subtypes for both positive and negative symptom dimensions (23, 24). These subtypes may convey a different level of risk for psychotic disorders and therefore also exhibit a different interplay with cognitive functions such as WM. It was expected that positive and negative PLEs would be best represented by previously established subdivisions of symptom clusters (18, 24–26) and that clusters associated most with an elevated risk of psychosis [bizarre experiences, persecutory ideas, perceptual abnormalities] (27) would show the strongest negative association with WM capacity.

## **MATERIALS AND METHODS PARTICIPANTS**

The study was carried out in a sample of 1087 Swedish citizens between 12 and 35 years of age. Participants were recruited via a company specialized in online data collection and population surveys (http://www.norstat.se). Individuals aged 15 years or older provided their informed consent via an online consent form. For individuals younger than 15 years one of their parents provided consent. The study was approved by the Central Ethical Review Board on Research Involving Humans at Karolinska Institutet. The role of the recruitment company was restricted to generating a study sample representative of the Swedish population and it was not involved in the design or execution of the study. Participants were drawn randomly from a voluntarily registered panel of over 100,000 Swedish citizens. A total of 9582 adult individuals received an invitation to participate and an additional 4652 parents of children between 12 and 18 years received an invitation for participation of their child (total response rate = 7.3%). In addition to meeting the age criterion, participants were required to be fluent in Swedish and to have access to a computer with internet connectivity and an operational sound system.

## **INSTRUMENTS AND TEST PROCEDURE Psychotic-like experiences**

Psychotic-like experiences were assessed with a Swedish version of the Community Assessment of Psychic Experiences (CAPE) (28). This self-report scale measures the lifetime prevalence of positive, negative, and depressive symptoms on both a frequency scale (0 = never to 3 = nearly always) and a distress scale (1 = not distressed to 4 = very distressed). The CAPE questions were translated from English into Swedish and back-translated to increase reliability. Three qualified Swedish researchers with a clinical background carried out the translation. An independent professional translator completed the back-translation, after which a consensus version was drawn up for implementation as an online survey. Frequency scores were transformed to range from 1 to 4 for further analysis and internal consistency of the total CAPE was high (α = 0.91). The translated questionnaire is freely available online at http://cape42.homestead.com/.

## **Working memory**

Two tasks of the Cogmed Progress Indicator (CPI) developed by Cogmed, Pearson Assessment, were used to assess WM capacity. The CPI tasks were originally designed to measure WM training improvements with non-trained tasks. The first WM task was the "Odd One Out"[based on a similar task in the Automated Working Memory Assessment (29)] and the second a digital variant of the "Following Instructions" task (30). **Figure 1** shows a single frame taken from both tasks.

In the Odd One Out, the participant is prompted to identify which shape out of three is the odd one and then remember its location. The procedure is then repeated with three new shapes after which three empty slots appear. The participant is required to respond by indicating *where* the odd shapes had appeared, in the correct order of appearance. Once the participant has successfully completed the practice trials with two items, the task begins with two items. Two correct trials on each level will lead to progression to the next level where the item load is increased by one. The test is completed when two trials on the same level are incorrect. The final score is calculated based on performance on the highest level achieved where at least one trial was passed, after which 0.3 would be subtracted for each incorrect answer on that level along with 0.15 subtraction for each incorrect answer on levels below the highest level achieved.

The Following Instructions task consists of common classroom items laid out on a table (e.g., eraser, crayon, box) and the task is to follow the verbal instructions given as accurately as possible. The instruction could for instance be "Click on the green eraser, then drag the black crayon to the yellow box," which would be a trial on span level three (because there are three items to remember what to do with). Practice trials with one and then two items are presented. The task has the same progression, stopping and scoring rules as described for the Odd One Out task above. Subjects exhibiting signs of floor effects (below a cut-off score of 2) were excluded from further data analysis. Finally, scores on both WM tasks were transformed into *z*-scores and an average *z*-score was calculated for each individual as a proxy for WM capacity.

### **Test procedure**

After consent was provided, participants received a link to the online assessment together with a unique username and password combination, valid for one assessment only. Participants were strongly advised to complete the assessment in a quiet environment. The order of the WM tasks was always the same [(1) Odd One Out, (2) Following Instructions]. After completion participants were automatically redirected to the web-based CAPE questionnaire. In order to progress each question required an answer, thereby preventing the occurrence of missing values. Participants that completed the assessment received a cinema voucher as a reward. The total duration of the recruitment period was 6 weeks.

#### **DATA ANALYSIS**

Statistical analyses were performed with IBM SPSS version 20.0. All data were checked for normality, homogeneity, duplicate cases, and outliers. To establish the presence of subtypes of PLEs in the present sample, it was first tested whether the positive symptoms scale contained a similar factor structure as reported in previous studies (18, 25, 26). For this purpose, a confirmatory factor analysis was conducted with IBM Amos 21 and three different models with a four-factor structure were compared. To determine the factor structure of negative symptoms an exploratory factor analysis was preferred, since only one known study (24) has previously

reported on this for the CAPE. Frequency scores were entered in a principal component analysis (PCA) with direct oblimin rotation. Three factors had an Eigenvalue ≥1 and the scree plot also indicated a cut-off of three factors. Subsequent parallel analysis with Monte Carlo simulation (31) confirmed that the three-factor structure best represented the data. Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy and Bartlett's test of sphericity were used to assess whether the factor model was appropriate. KMO was 0.91 and Bartlett's test significant (*p* < 0.001), which suggested that data properties were excellent for factor analysis (32).

Next, sum scores for all subscales (four positive, three negative) were generated. An additional subscale for depressive symptoms (eight items) was also added for explorative reasons. All subscales were first checked for gender and age differences with non-parametric Mann–Whitney *U* and Kruskal–Wallis tests. For age comparisons participants were divided into three groups with a total age span of 8 years each:"adolescents"(12–19 years),"young adults" (20–27 years), and "adults" (28–35 years). *Post hoc* testing was based on the assumption of unequal sample sizes and variances and checked with Tamhane's T2 tests. Next, it was assessed whether there were any linear associations between WM and CAPE subscale scores. Spearman's rho was used for bivariate correlations and the *p*-value was Bonferroni corrected for multiple comparisons (*p* < 0.0063). Multiple hierarchical regression models were used to further investigate the association of WM with symptom scores. For regression analyses *p*-values < 0.05 were considered to be statistically significant.

## **RESULTS**

#### **PARTICIPANTS**

A total of 1087 assessments were completed. Fifty-one duplicate cases were identified and excluded. Fourteen participants did not meet the age criterion and 10 additional adolescents were excluded after verification that a parent had either partially or completely filled out the CAPE on their behalf. This resulted in 1012 subjects for factor analysis of the CAPE, consisting of 547 female (54%) and 465 male (46%) individuals between 12 and 35 years (*M* = 24.4 years ± 6.6). The geographic distribution of participants was spread over all 21 counties (*län*) of Sweden and central tendency scores (mean, mode, median) for "highest education completed" indicated that a secondary education was the most common and average level of completed education in the sample.

## **CONFIRMATORY FACTOR ANALYSIS AND POSITIVE SYMPTOM SUBSCALES**

Model-fitting statistics are provided in **Table 1**. In sum, the results show that a four-factor solution for the positive symptoms dimension is plausible for our sample. Comparison of the three models indicates that the base model derived from Ref. (18), is the single model that would meet broadly used cut-off criteria for modelfitting indices (33). The model slightly improved when error terms of three item-pairs were correlated. The models based on the factor structure by Yung et al. (25) and Armando et al. (26), required an additional number of correlated error terms to obtain an adequate fit. Based on the comparison, the model of Ref. (18) was adopted in this study to generate sum scores for the subtypes: bizarre experiences (BE), persecutory ideas (PI), perceptual abnormalities


**Table 1 | Confirmatory factor analysis: fitting of three different models for the Community Assessment of Psychic Experiences – positive symptoms dimension**.

CFI, comparative fit index; RMSEA, root mean square error of approximation; AIC, akaike information criterion; BE, bizarre experiences; PA, perceptual abnormalities;

PI, persecutory ideas; MT, magical thinking; GR, Grandiosity; +/− = added/removed; r, correlated error terms; q, question; → = model pathway.

(PA), and grandiosity (GR). The factor loadings for both studies are displayed in **Table 2**.

## **EXPLORATORY FACTOR ANALYSIS AND NEGATIVE SYMPTOM SUBSCALES**

The three-factor solution explained 54% of the variance and all items loaded high (>0.4) on one factor variable and only one item (item 29) had a cross-loading ≥0.25 (**Table 3**). The first factor was related to avolition (AV) and consisted of seven items, three items loaded on a second factor related to affective flattening (AF) and the last factor consisted of four items related to social withdrawal (SW). Compared to the only available factor structure of CAPE negative symptoms (24) loadings were in high agreement (**Table 3**). One item (16; "Do you ever feel you have no interest to be with other people?") loaded highest on SW in the current study as opposed to AV in the previous study and was therefore included in the SW sum score only.

## **FREQUENCY OF PSYCHOTIC-LIKE EXPERIENCES**

Average prevalence of positive symptoms, determined by individuals reporting ≥2 on the frequency scales (=sometimes or more), was relatively high for PI and GR items (respectively 45 and 62%). Fewer individuals reported having experienced BE or PA (respectively 20 and 5%). All prevalence rates substantially decreased as the frequency rate increased (see **Table 2**). For negative symptoms the average prevalence rate was highest for SW (75%), then AV (70%), and lowest for AF (47%). In concordance with positive symptoms there was a steep decline in prevalence rates of negative symptoms with increased frequency (see **Table 3**). Depressive symptoms were also relatively common (*M* = 63%; range: 28–93%), with an average of 3% (range: 1–6%) of respondents indicating that the symptoms were almost always present.

## **PSYCHOTIC-LIKE EXPERIENCES BY GENDER AND AGE**

Average CAPE scores divided by gender and age groups are available in **Table 4**. There was no overall pattern of gender differences. Female individuals reported higher scores on PI (*U* = 115,642, *p* = 0.011) and depression (*U* = 102,606.5, *p* < 0.001) than males and there was a trend for higher PA in females as well (*p* = 0.053). In contrast, males reported higher frequencies of GR (*U* = 137,215, *p* < 0.026) and AF (*U* = 115,642, *p* = 0.011). Comparisons between age-groups showed that the reported frequency of all symptom scales differed significantly with age (all *p* < 0.05). *Post hoc* testing suggested that positive symptoms were more common in adolescents, while the highest frequency of negative and depressive symptoms was found in the young adult group (**Table 4**).

## **WORKING MEMORY CAPACITY**

Data of the WM assessment was missing for 16 subjects due to registration errors. Another 92 subjects were excluded based on suspected floor effects on at least one of both WM variables,leaving 904 remaining subjects for WM analysis (*M*age = 24.7 years ± 6.4; 420 males, 484 females). Average performance level was higher for the Odd One Out task (*M* = 5.81, SD = 1.21, range = 2.4–11.0) than for the Following Instructions task (*M* = 4.89, SD = 1.02, range = 2.0–8.4). The Pearson correlation coefficient between tasks was *r* = 0.31, *p* < 0.001. Correlations of both tasks with overall WM capacity (average *z*-score) was *r* = 0.81, *p* < 0.001. When checked, there was no gender difference in WM capacity (*t* <sup>902</sup> = 0.70, *p* = 0.48).

## **CORRELATIONS BETWEEN WORKING MEMORY AND PSYCHOTIC-LIKE EXPERIENCES**

Correlations of CAPE subscale frequency scores with WM capacity were calculated for 904 participants. After correction for multiple comparisons WM showed a negative association with positive symptom subtypes (BE: *r* = −0.13, *p* = 0.00005; PI: *r* = −0.10, *p* = 0.003) and depression (*r* = −0.10, *p* = 0.003). Partial correlations controlled for age and gender gave similar results (BE: *r* = −0.10, *p* = 0.004; PI: *r* = −0.09, *p* = 0.006; depression: *r* = −0.11, *p* = 0.001).

## **MULTIPLE REGRESSION WITH WORKING MEMORY AND PSYCHOTIC-LIKE EXPERIENCES**

Subscales correlating with WM capacity (BE, PI, and depression) were entered as dependent variables in hierarchical multiple

#### **Table 2 | Community assessment of psychic experiences: positive symptoms frequency scores and factor analysis**.


CFA, confirmatory factor analysis; EFA, exploratory factor analysis.

#### **Table 3 | Community assessment of psychic experiences: negative symptoms frequency scores and factor analysis**.


EFA, exploratory factor analysis.

regression models. Age (*z*-score) and gender were entered first, next WM capacity, and finally the models were checked for interactions with age and gender. For BE the procedure was followed without inclusion of gender terms in the model. An overview of the regression outcomes is provided in **Table 5**. All models were highly significant (*p* < 0.001) and WM was negatively associated with BE, PI, and depression, regardless of age*<sup>z</sup>* (BE, PI) or gender (PI, depression) effects (all *p* < 0.01). For depression there was an additional interaction effect of age*<sup>z</sup>* ×WM (*t* = −2.18, *p* = 0.03). All effect sizes (β) of significant predictors were considered small. To investigate the specificity of the association between WM and the symptom subscales, regression analyses were repeated with total CAPE score entered as an additional covariate. WM remained a significant predictor for BE and depression (both *p* < 0.05) and at trend-level for PI (*p* < 0.06). To further explore the presence of more specific age-effects, the sample was split into the three age-groups described above and regression analyses were repeated. WM significantly predicted BE in the young adult group [β = −0.12, *t*(419) = −2.43, *p* = 0.016] and depression in the adult group [β = −0.11, *t*(321) = −2.04, *p* = 0.046].

There were no significant associations between WM and symptom scores in the adolescent group.

## **DISCUSSION**

Understanding how cognitive functions relate to behavioral symptoms in the general population can provide further insight into underlying mechanisms of emerging psychopathology and lead to identification of early intervention targets. WM represents a cognitive function that is commonly compromised in schizophrenia spectrum disorders and in individuals atrisk for psychosis, though little is known about specific contributions of WM impairments to schizotypal symptomatology. This study aimed to investigate whether WM capacity was related to presence of PLEs in a large population sample of adolescents and young adults. WM was negatively associated with subtypes of positive (BE and persecutory ideas) and depressive symptoms, also after adjusting for age, gender, and global symptom scores (trend-level for persecutory ideas). However, effect sizes were small and when divided into different age groups, WM was exclusively associated with BE in young


#### **Table 4 | Community assessment of psychic experiences (CAPE) scores divided by gender and age group**.

BE, bizarre experiences; PA, perceptual abnormalities; PI, paranoid ideations; GR, grandiosity; social withdrawal; affective flattening; avolition; Post hoc results (p < 0.05).

<sup>a</sup>Adolescents and young adults > adults.

<sup>b</sup>Adolescents > young adults > adults.

<sup>c</sup>Young adults > adults.

<sup>d</sup>Young adults > adolescents (trend: p = 0.08).

<sup>e</sup>Young adults > adolescents.

#### **Table 5 | Regression models of symptom scores with working memory, age, and sex as predictors**.


SE, Standard Error; β = standardized beta values.

\*p < 0.05; \*\*p < 0.01.

adults (20–27 years) and with depression in the adult group (28–35 years).

Negative associations between WM and PLEs have previously been reported for clinical (34–36) and non-clinical samples (37, 38), but relatively few studies have assessed these relations for multiple schizotypal dimensions and simultaneously accounted for their shared variance in the analyses. Two studies that did apply this approach found that reduced WM is associated with more positive symptoms, which was confirmed by the current study. Schmidt-Hansen and Honey (20) reported a strong association for three out of four WM parameters derived from an *n*-back task (*N* = 289). Likewise, Matheson and Langdon (21) reported a negative association for performance on a Letter-Number sequencing task (*N* = 97). However, both studies also found evidence for a link between WM and negative symptoms. This discrepancy with the current findings may partially be caused by differences in study sample, as both previous studies were conducted in student populations. Furthermore, all three studies used different WM assessments and symptom questionnaires, which restricts direct comparisons. Here, performance on two computerized WM tasks was averaged to create a global estimate of each individual's WM capacity. For future studies the use of multiple parameters delineating different aspects of WM function could provide further detail about specific contributions to the underlying associations with PLEs.

Several additional findings in this study require further emphasis. First, the presented data confirm the presence of underlying subtypes of CAPE positive and negative symptom dimensions,corroborating earlier findings in clinical (23) and non-clinical (17, 18, 24–26) study cohorts. This strengthens the idea that the classical schizophrenia dimensions do not represent homogenous entities and can be subdivided into meaningful symptom clusters. Second, associations with low WM capacity complement previous observations of increased distress and decreased global functioning for high intensity of bizarre experiences and persecutory ideas (17, 18). It also supports the notion that high frequencies of these types of positive symptoms may reflect an increased vulnerability for psychosis (27). Third, there were marked age differences in symptom frequency, which also affected the strength of associations with WM capacity. Perhaps most surprising was the absence of significant WM associations for adolescent participants. This suggests that although PLEs are more frequent in adolescents (39) they are not necessarily a marker of decreased cognitive capacity. Speculatively, for some adolescents low WM capacity may simply reflect delayed cognitive maturation instead of underlying psychopathology. Fourth, the finding that depressive symptoms were uniquely associated with WM capacity was somewhat unexpected, although these scores tend to correlate highly with bizarre experiences and persecutory ideas on the CAPE questionnaire (18, 23) and may therefore share some overlap in associated cognitive dispositions. However, the association with WM was strongest in adult individuals, which could suggest an age-related shift in the idiosyncratic nature of WM impairment as a vulnerability marker for different types of psychopathology. Although this finding awaits replication, it underlines the divergent validity of the CAPE depression scale, which can be used independently to explore unique relations with cognitive factors of interest in future studies.

The computerized WM tasks used for this study were originally designed to measure WM training improvements or"transfer effects" on non-trained tasks, based on previous findings (40, 41). Despite ongoing debate regarding the actual extent of transfer effects, there is overall agreement that WM training can enhance performance on more complex cognitive tasks in addition to improved WM capacity (42–45). This apparent display of cognitive plasticity is accompanied by underlying functional brain changes, for example in the (striatal) dopamine system (46–48), dysregulation of which is particularly associated with presence of positive symptoms (49). Furthermore, cognitive remediation strategies that target WM and other cognitive functions have proven successful with regard to cognitive and functional outcome in chronic schizophrenia (50, 51), and bear promise for individuals at-risk for psychosis (52). However, cognitive training can be costly, labor intensive, and its efficacy may vary across individuals. Therefore it would be advantageous to select more narrowly defined participant groups with an increased chance of clinical improvement. Based on the current study results it is tempting to speculate that individuals with high expression of specific positive and depressive symptoms can potentially accomplish the greatest clinical gain from WM training. However, in this study WM only explained a very small proportion of variance in CAPE scores (1%). As such, it is deemed unlikely that stand-alone WM training could directly influence the presence of PLEs for most individuals. Notwithstanding, it is possible that WM training can channel its efficacy in a more indirect manner, e.g., by creating more optimal neurocognitive properties to benefit from other types of treatment. Additional research on actual training interventions is needed to further address this issue.

The current study has multiple methodological strengths and limitations that merit additional commenting. By implementing a fully automated online assessment procedure it was possible to reach a large number of participants within a relatively short time period. Moreover, it allowed for recruitment of a study sample that is considered more or less representative of the general Swedish population between 12 and 35 years. Although the sample may not be completely devoid of any selection bias, it contains, for example, more regional and occupational diversity than community or student populations, which are typically used for this type of large-scale studies. In addition, it has been demonstrated that online assessment of CAPE symptoms is robust against symptom simulation (53) and by requiring individuals to answer each individual question in order to proceed, the problem of dealing with missing data was omitted. Regardless, the assessment could have benefited from adding several items to help detect any malingering participants. Furthermore, the monetary incentive (cinema voucher) may have encouraged participants to register more than once. Although extra precaution was taken to prevent this from happening and data were carefully screened for double entries, it could not be verified completely that all registered participants represented unique individuals. Future studies using a similar approach are advised to apply more rigorous personal identification methods (e.g., via social security number) and to refrain from using a monetary incentive, which would circumstantially further reduce the study costs as well. Finally, even though the online cognitive assessments had considerable logistic advantages over on-site assessments, they did not allow for a fully controlled study environment, which may have influenced the outcome. In general, the adopted online assessment strategy has multiple caveats, though more traditional assessment procedures are by no means less prone to measurement error or sample bias. Furthermore, there are substantial advantages of online recruitment and assessment procedures regarding data completion, project duration, and costs.

To sum up, the current study has provided further evidence for the presence of discernible subtypes of positive and negative PLEs in a large population sample of Swedish adolescents and young adults. There was an inverse relation of WM capacity with presence of positive symptom subtypes (bizarre experiences and persecutory ideas) as well as depressive symptoms, which was moderated by age. However, these relations should be interpreted in light of small effect sizes and marginal impact of WM capacity on symptom variability. Together these findings suggest that specific schizotypal traits may be indicative of reduced WM capacity in different age groups and potentially harbor a greater need for targeted intervention.

## **AUTHOR CONTRIBUTIONS**

Tim B. Ziermans conceived the idea and methodology of this study, organized participant recruitment and data processing, conducted the statistical analyses, and wrote the final manuscript.

## **ACKNOWLEDGMENTS**

The author would like to thank the following people for their respective contributions: Torkel Klingberg, Megan Spencer-Smith, Henrik Ullman, Elin Lidman, Elin Helander, Henrik Malmqvist, Annie Möller, Maria Kareliusson, and Pedro Ribeiro (Klingberg lab); Jonas Beckeman, Sissela Bergman-Nutley, Lars Blåsjö, Michael Smietana, and Stina Söderqvist (Cogmed, Pearson Assessment); Per-Tomas Brettell, Johan Hermansson, and Jonas Horn (Norstat); Ida Enqvist, Rebecca Ahlfeldt. Dr. Ziermans was supported by a COFAS Marie Curie Fellowship.

## **REFERENCES**


**Conflict of Interest Statement:** The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 24 October 2013; paper pending published: 18 November 2013; accepted: 20 November 2013; published online: 03 December 2013.*

*Citation: Ziermans TB (2013) Working memory capacity and psychotic-like experiences in a general population sample of adolescents and young adults. Front. Psychiatry 4:161. doi: 10.3389/fpsyt.2013.00161*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Ziermans. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Neurocognitive functioning in schizophrenia and bipolar disorder: clarifying concepts of diagnostic dichotomy vs. continuum

#### **Carissa N. Kuswanto<sup>1</sup> , MinY. Sum<sup>1</sup> and Kang Sim1,2\***

<sup>1</sup> Research Division, Institute of Mental Health/Woodbridge Hospital, Singapore

<sup>2</sup> Department of General Psychiatry, Institute of Mental Health/Woodbridge Hospital, Singapore

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Drozdstoy Stoyanov Stoyanov, Medical University, Bulgaria Assen Veniaminov Jablensky, The University of Western Australia, Australia Kathryn Eve Lewandowski, McLean

Hospital, USA

#### **\*Correspondence:**

Kang Sim, Department of General Psychiatry, Institute of Mental Health/Woodbridge Hospital, 10, Buangkok View, 539747, Singapore e-mail: kang\_sim@imh.com.sg

The Kraepelinian dichotomy posits that patients with schizophrenia (SCZ) and bipolar disorder (BD) present as two separate psychotic entities such that they differ in terms of clinical severity including neurocognitive functioning. Our study aimed to specifically compare and contrast the level of neurocognitive functioning between SCZ and BD patients and identify predictors of their poor neurocognitive functioning.We hypothesized that patients with SCZ had a similar level of neurocognitive impairment compared with BD. About 49 healthy controls (HC), 72 SCZ, and 42 BD patients who were matched for age, gender, and premorbid IQ were administered the Brief Assessment of Cognition battery (BAC). Severity of psychopathology and socio-occupational functioning were assessed for both patients groups. Both BD and SCZ groups demonstrated similar patterns of neurocognitive deficits across several domains (verbal memory, working memory, semantic fluency, processing speed) compared with HC subjects. However, no significant difference was found in neurocognitive functioning between BD and SCZ patients, suggesting that both patient groups suffer the same degree of neurocognitive impairment. Patients with lower level of psychosocial functioning [F(1,112) = 2.661, p = 0.009] and older age [F(1,112) = −2.625, p = 0.010], not diagnosis or doses of psychotropic medications, predicted poorer overall neurocognitive functioning as measured by the lower BAC composite score. Our findings of comparable neurocognitive impairments between SCZ and BD affirm our hypothesis and support less the Kraepelinian concept of dichotomy but more of a continuum of psychotic spectrum conditions. This should urge clinicians to investigate further the underlying neural basis of these neurocognitive deficits, and be attentive to the associated socio-demographic and clinical profile in order to recognize and optimize early the management of the widespread neurocognitive deficits in patients with SCZ and BD.

**Keywords: schizophrenia, bipolar disorder, Kraepelinian dichotomy, neurocognitive functions, psychosocial functioning, psychotic spectrum**

## **INTRODUCTION**

The Kraepelinian dichotomy, a prominent paradigm in psychiatry, has influenced nosology for severe psychiatric disorders such as schizophrenia (SCZ) and bipolar disorder (BD) within diagnostic classification systems including the Diagnostic Statistical Manual of Mental Disorders and the International Classification of Diseases for many decades (1, 2). The paradigm posits that patients with dementia praecox (SCZ) and manic depression (BD) present as two separate psychotic disorders (3, 4). Evidence in support of the Kraepelinian notion includes volumetric neuroimaging differences in brain regions such as the amygdala, hippocampus, and lateral ventricles that appear to be disorder-specific (5–7). In addition, unlike BD, reduced gray matter volumes and neurocognitive neuromotor impairments occur in the prodromal stage of SCZ despite both having early onset of illness during adolescence and recent findings of common genetic predisposition in BD and SCZ (6, 8–13). Furthermore, patients with SCZ are thought to suffer more extensive brain morphological abnormalities and more severe neurocognitive deficits in comparison with patients with BD (14–18). Other than the apparent empirical evidence, the Kraepelinian dichotomy is appealing clinically due to its conceptual and diagnostic simplicity (19, 20).

However, several issues challenge the dichotomous paradigm of psychotic disorders with respect to SCZ and BD. These issues include the overlapping organic basis and blurred boundaries between psychoses from emerging biological data (21, 22), the non-specificity of psychopathological features such as perceptual disturbances and affective symptoms between SCZ and BD, and the shared genetic susceptibility loci uncovered within recent genome-wide association studies (GWAS) (23, 24). Accumulating evidence from clinical observations, antipsychotic treatment, and findings of neurocognitive functioning in SCZ and BD further argued against the Kraepelinian dichotomy. First, clinicians frequently encounter patients who do not fall neatly into either category, such as individuals with schizoaffective disorders and BD patients with prominent psychotic symptoms. Second, dopamine

dysregulation has been implicated in both disorders, making antipsychotic drugs useful in the management of SCZ as well as BD (6, 8). Third, with regards to neurocognitive functioning, extant studies have reported poorer neurocognitive functioning in patients with SCZ compared to those with BD, but there are also data suggesting that patients with SCZ are comparable with BD patients in terms of neurocognitive impairments (14, 25–31). These inconsistent findings on neurocognitive functioning may be due to differences in sample size, specific neurocognitive deficits measured, type of scales used, and presence of other confounders including age and premorbid intelligence (26, 27, 30, 32).

Thus, in this study, we aimed to specifically compare and contrast the level of neurocognitive functioning between patients with SCZ and BD as well as identify predictors of poor neurocognitive functioning in these patients. Based on clinical impression within a tertiary psychiatric hospital context and extant data, we hypothesized that patients with SCZ had a similar level of neurocognitive impairment compared with BD and the neurocognitive impairment is associated with particular socio-demographic and clinical factors after taking into account premorbid intelligence, level of education, and psychotropic medications prescribed.

## **MATERIALS AND METHODS**

## **PARTICIPANTS**

The cross-sectional study sample comprised of a total of 163 participants [72 patients diagnosed with SCZ, 42 patients with BD, and 49 healthy controls (HC)] who gave their written informed consent to participate in the study after a detailed explanation of the study procedures. Patients with SCZ and BD were recruited from the outpatient settings of the Institute of Mental Health, Singapore. All diagnoses were confirmed by a psychiatrist (Kang Sim) using information obtained from the existing medical record, clinical history, mental status examination, interviews with the patients and their significant others as well as the administration of the Structured Clinical Interview for DSM-IV disorders – Patient Version (SCID-I/P) (33). All patients were maintained on a stable dose of psychotropic medication for at least 2 weeks prior to the recruitment and did not have their medication withdrawn for the purpose of the study. HC were administered the Structured Clinical Interview for DSM-IV disorders – Non-Patient Version (SCID-I/NP) (34). None of the participants had a history of significant and/or unstable/untreated medical illnesses such as seizure disorder, head trauma, or cerebrovascular accidents. Moreover, no subjects had a current or past history of substance use or alcohol use disorder. This study was approved by the Institutional Review Boards of the Institute of Mental Health, Singapore.

## **NEUROCOGNITIVE ASSESSMENT**

## **The brief assessment of cognition battery**

The Brief Assessment of Cognition (BAC) battery (35) was administered to all subjects. The subjects were randomly assigned to a sequence of BAC version A and B. The BAC scales include:


We divided the token motor task into two measures, namely the number of tokens correctly placed into a container (tokens correct) and the number of tokens left after 60 s (tokens left). Although both measures tap motor speed function, tokens correct measure did not credit the subjects for tokens put into container incorrectly (i.e., not picked up and placed at the same time) which required eye-hand coordination and precision in order to perform the task correctly. Digit sequencing task, token motor task, symbol coding, and Tower of London measured working memory, motor speed, attention, and speed of information processing and executive functioning respectively (35, 36).

## **Wide range achievement test (reading scale)**

The wide range achievement test (WRAT) (37) reading subscale was used to measure premorbid intelligence in all the subjects.

## **CLINICAL ASSESSMENTS**

## **Global assessment of functioning**

The global assessment of functioning (GAF) (38) rated (from 0 to 90) the severity of symptoms, disability, and the total level of social and occupational functioning. The GAF was administered to both SCZ and BD groups.

## **Positive and negative syndrome scale**

The positive and negative syndrome scale (PANSS) (39) allowed characterization and quantification of psychotic psychopathology. The scale has 7 items for positive symptoms, 7 items for negative symptoms, and 16 items for general psychopathology. The PANSS was administered to SCZ and BD patients.

## **Young mania rating scales**

The Young mania rating scales (YMRS) (40) contains 11 items which were used to measure the severity of manic symptoms in BD patients. There are four items that are graded on a 0–8 scale (irritability, speech, thought content, and disruptive/aggressive behavior). The remaining seven items are graded on a 0–4 scale. The YMRS scale was administered to SCZ and BD patients.

## **STATISTICAL ANALYSES**

All statistical tests were performed using PASW for Windows, version 18.0 (SPSS, Inc., Chicago, IL, USA). The normality of distributions of continuous measures was checked with

the Kolmogorov–Smirnov one-sample test. Socio-demographic variables were compared using two sample student *t*-tests and *F*-test for continuous variables, and chi-square test for categorical variables. Two-way analysis of covariance (ANCOVA) was used to examine for any difference in neurocognitive functioning (BAC task performance) between the groups after adjusting for the covariates described below. We first compared the neurocognitive measures between all three groups, HC vs. SCZ and HC vs. BD, and adjusting for covariates of age, gender, years of education, and premorbid IQ. We then specifically compared the BAC item performance between SCZ and BD patients, by adjusting for premorbid IQ, age, gender, years of education, and clinical characteristics such as the duration of illness, duration of untreated psychosis, GAF scores, PANSS composite scores, YMRS scores, and antipsychotic dose in terms of mean daily chlorpromazine (CPZ) mg equivalents. Stepwise linear regression analyses were performed to determine predictors of neurocognitive functioning in patients including socio-demographic (such as age, gender) and clinical characteristics (such as diagnosis, duration of illness, GAF scores, PANSS composite scores, YMRS scores, and antipsychotic dose). The significance level for statistical tests was set at two tailed *p* < 0.05.

## **RESULTS**

## **SOCIO-DEMOGRAPHIC CHARACTERISTICS**

Across the three subject groups (HC, SCZ, BD), there were significant differences in subject's and mother's years of education but there was no significant difference in age, sex, handedness, and WRAT scores. Compared to patients with BD, patients with SCZ had longer duration of untreated psychosis and lower level of psychosocial functioning as reflected by the lower GAF scores. There was no significant difference in age, gender, handedness, years of education, WRAT scores, age of onset, and duration of illness between SCZ and BD. The socio-demographic and clinical features of the subjects are shown in **Table 1**.

## **NEUROCOGNITIVE FUNCTIONING**

**Table 2** showed the neurocognitive profile across the entire subject sample based on the BAC battery administration. HC subjects scored the highest in all BAC items compared to SCZ and BD patients. Compared to HC, patients with SCZ scored lower in BAC tasks of verbal memory, semantic and letter fluency, digit sequencing, tokens left, symbol coding, total BAC score indicating neurocognitive deficits in verbal memory and fluency, working memory, motor speed, and attention. Compared to HC, patients

## **Table 1 | Demographic and clinical characteristics of the participants**.


BD, bipolar disorder; CPZ, chlorpromazine; GAF, global assessment of functioning; HC, healthy controls; PANSS, positive and negative syndrome scale; SCZ, schizophrenia; WRAT, wide range achievement test; YMRS, young mania rating scale.\*p < 0.05.



BAC, brief assessment of cognition battery; BD, bipolar disorder; CPZ, chlorpromazine; GAF, global assessment of functioning; HC, healthy controls; SCZ, schizophrenia; WRAT, wide range achievement test. \*p < 0.05 \*\*p < 0.01.

<sup>a</sup>Adjusted for WRAT, age, gender, and years of education.

<sup>b</sup>Adjusted for WRAT, age, gender, years of education, duration of untreated psychosis, duration of illness, GAF (total, symptoms, and disability), antipsychotic dose, and PANSS and YMRS composite scores.

with BD scored lower in BAC tasks of verbal memory, semantic and letter fluency, digit sequencing, tokens correct, symbol coding, total BAC score indicating almost similar neurocognitive deficits as SCZ in verbal memory and fluency, working memory, motor speed, and attention. After taking into account multiple comparisons (at conservative threshold of *p* < 0.005), patients with SCZ scored lower on tasks of verbal memory, semantic fluency, symbol coding, and patients with BD scored lower on verbal memory, digit sequencing, and symbol coding when compared with HC.

However, there was no significant difference in all neurocognitive domains of BAC when SCZ and BD groups were compared and after controlling for covariates such as demographic characteristics (age, gender, years of education, WRAT scores) and clinical features (duration of untreated psychosis, antipsychotic dose, as well as GAF, PANSS, and YMRS composite scores). We also noted the possibility that a history of psychosis in BD patients may be associated with poorer performance in neurocognitive tasks compared to those BD subjects without history of psychosis (30). Thus, we performed a *post hoc* analysis whereby we compared BD patients with history of psychosis and those without and we found no significant difference in all neurocognitive performance in all BAC domains.

Multivariate linear regression analyses with the overall BAC score as the dependent variable and demographic (age, gender, education) and clinical factors (diagnosis, WRAT scores, GAF scores, PANSS composite scores, YMRS scores, duration of illness, duration of untreated psychosis, antipsychotic dose) as predictor variables revealed that patients with lower level of psychosocial functioning as indicated by lower GAF score [*F*(1,112) = 2.023, *p* = 0.046] and older age [*F*(1, 112) = −2.190, *p* = 0.031], but not diagnosis or doses of psychotropic medications, were associated with poorer overall neurocognitive functioning as measured by the lower BAC composite score.

## **DISCUSSION**

There were several main findings in this study. First, patients with BD and SCZ were found to have greater neurocognitive impairments in most of the BAC domains compared to the HC who were matched for age, gender and premorbid intelligence. Second, the extensive neurocognitive impairments found in SCZ were comparable with those observed in patients with BD. Third, the level of psychosocial functioning and age of patients with SCZ and BD were associated with neurocognitive impairment and not diagnosis or antipsychotic dose. Our findings highlighted that both SCZ and BD patients in our study sample suffered largely similar neurocognitive impairments, which support a continuum concept of psychosis rather than the Kraepelinean concept of dichotomous psychotic conditions.

Our findings of comparable and extensive neurocognitive impairments in both SCZ and BD are consistent with some earlier studies reporting findings of specific neurocognitive impairments across the patient groups. For example, BD and SCZ patients scored significantly worse than control subjects in verbal memory task (assessed by the California Verbal Learning Test), which did not differ when both patient groups were compared (25, 29). Likewise, motor speed was also impaired in patients with SCZ/BD during the Grooved Pegboard test when compared with HC (14, 31). Furthermore, other studies have shown comparable impairments in domains of verbal fluency and working memory within both SCZ and BD patients (14, 25, 28). In terms of executive functioning, a number of studies found comparable neurocognitive deficits in both patient groups irrespective of the tests used such as Wisconsin Card Sorting Test (WCST) (14, 29, 41), Trail Making Test B (TMT B) (28, 31), and Tower of Hanoi (29). Overall, our findings add to the literature to support the notion of comparable and pervasive neurocognitive impairments in patients with BD and SCZ (17, 29).

What may be the biological factors that underlie SCZ and BD which may be relevant to the observed neurocognitive impairments? First, recent evidence from large-scale GWAS has pointed to possible common susceptibility genes underlying SCZ and BD which may affect neurocognitive functioning. These common vulnerability cross disorder genes include ZNF804A, CACNA1C, NRGN, and PBRM1 (5, 19, 23, 24, 42–45). A recent study from the Psychiatric Genome Consortium reported that there are 14 genetic loci associated with both SCZ and BD (46), and shared genetic effects between SCZ and BD accounted for 52% of the genetic variance in SCZ and 69% for BD in the Swedish population (44). Second, neuroimaging aberrations of brain morphology and function have been noted. For example, voxel-based morphometry studies found that there was a substantial overlap in gray matter volumetric reductions such as the prefrontal cortex, thalamus, caudate, medial temporal lobe, insula, and the anterior cingulate regions in SCZ and BD (7, 15, 47). Third, while genetic risks for SCZ and BD are associated with different gray matter changes, McDonald et al. (48) suggested that they share white matter endophenotypes with the reduction of white matter volume in the left frontal and temporoparietal regions in both SCZ and BD.

Of note, older age and poorer psychosocial functioning were found to be associated with neurocognitive impairments in our patients. These findings are in line with that of other studies whereby older age was associated with decreased gray matter volume (49) and poorer neurocognitive performance in SCZ, and which can affect psychosocial functioning (50). The lifetime history of psychotic features in BD patients has also been found to worsen subsequent neurocognitive performance such as cognitive flexibility and working memory indicating age related effects on neurocognitive functioning in BD (10, 26, 27, 30, 32, 51). However, in contrast, Gildengers et al. (52) found no difference in the rate of neurocognitive decline between BD and healthy subjects with age.

We found that antipsychotic medication dose was not associated with neurocognitive impairment in SCZ and BD in this study although there is evidence of antipsychotic medications affecting changes in brain structures and neurocognitive performance. Previous studies have reported both improvements in neurocognitive impairments with antipsychotic treatment in SCZ and BD (26) but negative impact on cognition have also been reported in BD (53, 54). In SCZ patients, some studies reported that the administration of typical antipsychotic medication provides only modest-to-moderate gains in multiple cognitive domains (55, 56), while other studies found no correlation between antipsychotic medication dose and neurocognitive functioning in BD and SCZ (9, 18, 57).

Recent data also suggest that neurocognitive deficits may potentially serve as endophenotypical trait markers for BD and SCZ (58–60). Unlike a state marker whereby neurocognitive performance may vary with and be influenced by current clinical states such as affective change and psychotic phenomenology, a trait marker remains stable throughout the illness course and may be related to genetic factors (60, 61). For SCZ, a twin study by Pardo et al. (58) noted preservative errors on the WCST not only in the SCZ twin but also within non-SCZ monozygotic co-twin suggesting a spectrum of genetic liability. Furthermore, clinical symptoms do not seem to play a role in worsening cognitive performance in ultra-high risk of psychosis subjects (62). In BD, Sobzcak et al. (59) found deficits in memory function, psychomotor performance and attention in high risk relatives suggesting potential neurocognitive trait markers for BD. In a recent longitudinal study involving BD patients, the severity of depressive symptoms did not predict change of performance in any cognitive domains and all domains remained stable over the course of 6 years (60).

There are several limitations to the study. First, our sample size was rather small and the study solely recruited patients with remitted SCZ and BD seen at the outpatient setting of a tertiary psychiatric hospital and thus the results may not be generalizable to patients within the inpatient and in non-tertiary psychiatric settings. Second, this was a cross-sectional study and did not allow observations of changes of neurocognitive functioning over the course of time. Third, the administrators of the neurocognitive tests were not blinded to the diagnoses of the subjects. Lastly, we did not examine other biological correlates such as genetic factors, neurophysiological measures but we intend to do so as a follow up study of these findings. These multimodality studies can potentially elucidate biological underpinnings or correlates of these neurocognitive impairments in SCZ and BD.

In conclusion, we found that patients with BD and SCZ suffer widespread neurocognitive deficits involving almost all the examined cognitive domains. The severity of neurocognitive deficits has been found to be significantly associated with various factors such as the level of psychosocial functioning and age but not specific diagnoses or antipsychotic medications. The underlying biological basis of these neurocognitive deficits may be related to factors common to both conditions and awaits better definition. Our findings support the dimensional concept of psychotic spectrum conditions along a continuum rather than the dichotomous concept as proposed by Kraepelin. Although DSM-V (63) still retains the existing nosological boundaries between BD and SCZ, clinicians would have to remain vigilant in managing each case of SCZ or BD by taking into account their clinical, socio-demographic factors, and assessing for the presence of specific neurocognitive impairment so as to better optimize the care and cater for more individually tailored treatment plans for these patients with potentially crippling conditions.

## **ACKNOWLEDGMENTS**

This study was supported by NHG SIG (12003/12004/11003) research grants awarded to Dr. Kang Sim. We would like to thank all of the patients, their families, and our hospital staff for their support of this study.

## **REFERENCES**


in schizophrenia and bipolar I disorder. *Bipolar Disord* (2013) **15**:680–93. doi:10.1111/bdi.12096


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 August 2013; accepted: 21 November 2013; published online: 05 December 2013.*

*Citation: Kuswanto CN, Sum MY and Sim K (2013) Neurocognitive functioning in schizophrenia and bipolar disorder: clarifying concepts of diagnostic dichotomy vs. continuum. Front. Psychiatry 4:162. doi: 10.3389/fpsyt.2013.00162*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Kuswanto, Sum and Sim. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

## Neuroanatomical differences between first-episode psychosis patients with and without neurocognitive deficit: a 3-year longitudinal study

#### **Rosa Ayesa-Arriola1,2† , Roberto Roiz-Santiáñez 1,2† \*, Rocío Pérez-Iglesias 1,2,3, Adele Ferro<sup>4</sup> , Jesús Sainz <sup>5</sup> and Benedicto Crespo-Facorro1,2**

<sup>1</sup> Department of Psychiatry, School of Medicine, University of Cantabria, University Hospital Marqués de Valdecilla, IFIMAV, Santander, Spain

<sup>2</sup> CIBERSAM, Centro Investigación Biomédica en Red Salud Mental, Madrid, Spain

<sup>3</sup> Psychosis Studies Department, Institute of Psychiatry, London, UK

<sup>4</sup> Department of Experimental Clinical Medicine, Inter-University Center for Behavioural Neurosciences (ICBN), University of Udine, Udine, Italy

<sup>5</sup> CSIC, Spanish National Research Council, Institute of Biomedicine and Biotechnology of Cantabria, University of Cantabria, Santander, Spain

#### **Edited by:**

Stefan Borgwardt, University of Basel, Switzerland

#### **Reviewed by:**

Joaquim Radua, King's College London, UK Renata Smieskova, University Psychiatry Clinics, Switzerland

#### **\*Correspondence:**

Roberto Roiz-Santiáñez, Unidad Investigación Psiquiatría, Hospital Universitario Marqués de Valdecilla., Avda. Valdecilla s/n, 39008 Santander, Spain

e-mail: rroiz@humv.es

†Rosa Ayesa-Arriola and Roberto Roiz-Santiáñez have contributed equally to this work and should be considered co-first authors.

**Background:** The course of cognitive function in first-episode psychosis (FEP) patients suggests that some individuals are normal or near normal whereas some cases present a marked decline. The goal of the present longitudinal study was to identify neuroanatomical differences between deficit and non-deficit patients.

**Methods:** Fifty nine FEP patients with neuroimage and neurocognitive information were studied at baseline and 3 year after illness onset. A global cognitive function score was used to classify deficit and non-deficit patients at baseline. Analysis of covariances and repeated-measures analysis were performed to evaluate differences in brain volumes. Age, premorbid IQ, and intracranial volume were used as covariates.We examined only volumes of whole brain, whole brain gray and white matter, cortical CSF and lateral ventricles, lobular volumes of gray and white matter, and subcortical (caudate nucleus and thalamus) regions.

**Results:**At illness onset 50.8% of patients presented global cognitive deficit.There were no significant differences between neuropsychological subgroups in any of the brain regions studied at baseline [all F(1, 54) ≤ 3.42; all p ≥ 0.07] and follow-up [all F(1, 54) ≤ 3.43; all p ≥ 0.07] time points. There was a significant time by group interaction for the parietal tissue volume [F(1, 54) = 4.97, p = 0.030] and the total gray matter volume [F(1, 54) = 4.31, p = 0.042], with the deficit group showing a greater volume decrease.

**Conclusion:** Our results did not confirm the presence of significant morphometric differences in the brain regions evaluated between cognitively impaired and cognitively preserved schizophrenia patients at the early stages of the illness. However, there were significant time by group interactions for the parietal tissue volume and the total gray matter volume during the 3-year follow-up period, which might indicate that cognitive deficit in schizophrenia would be associated with progressive brain volume loss.

**Keywords: first-episode psychosis, neuroanatomical, neurocognition, schizophrenia, cognitive deficit**

## **INTRODUCTION**

Cognitive deficits are core features of schizophrenia that are already evident at early phases of the illness (1, 2), with more than 80% of patients showing deficits in one or more domains of cognitive function (3). There are, however, noticeable differences among patients, with a subgroup showing severe and debilitating cognitive dysfunctions, typical of Kraepelin's description of dementia praecox (4), and other subgroup considered to be "neuropsychologically normal" (5). These distinguishable subgroups probably lie at different levels of severity on a continuum of causes or of different factors that might be influencing outcome (6).

Cognitive functioning has been associated with measures of brain structures both in schizophrenia patients and healthy subjects (7). It is now well established that schizophrenia is also associated with structural brain abnormalities (8, 9). Although some longitudinal magnetic resonance imaging (MRI) studies have reported a progressive brain tissue loss during the early years after the first psychotic episode of schizophrenia compared to healthy subjects (10–12), others have failed to reveal such progressive brain volume loss (13–18). Nevertheless, only few studies have examined those brain abnormalities that characterize the disorder in cognitive subgroups and, to the best of our knowledge, there are no previous studies that have examined progressive brain changes associated with cognitive deficit in schizophrenia.

The aim of this study was to identify neuroanatomical differences that possibly underlie neurocognitive function deficit in first-episode psychosis (FEP) patients followed up for 3-years. Previous studies have associated white matter abnormalities with cognitive deficit (19, 20). Perez-Iglesias and colleagues (19), using diffusion tensor imaging, showed that deficits in executive and motor functioning in patients with FEP were associated with reductions in white matter integrity, and Wexler et al. (20) found that neuropsychologically impaired patients had significantly smaller white matter volumes in several regions. On the basis of these studies, we hypothesized that patients with cognitive deficit would present a white matter volume diminution.

## **MATERIALS AND METHODS**

### **STUDY DESIGN AND SETTING**

Data for the present investigation were obtained from a large epidemiological and 3-year longitudinal intervention program of FEP (PAFIP) conducted at the outpatient clinic and the inpatient unit at the University Hospital Marques de Valdecilla, Santander, Spain. It conformed to international standards for research ethics and was approved by the local institutional review board. Informed consent of the participants was obtained after the nature of the procedures had been fully explained. The referrals to the PAFIP came from the inpatient unit and emergency room at the University Hospital Marques de Valdecilla, community mental health services and other community health care workers in Cantabria. A more detailed description of our program has been previously reported (21, 22).

### **SUBJECTS**

All the patients included in PAFIP from February 2001 to December 2007 were invited to participate in this study. Patients referred to the program were selected if they met the following criteria: (1) 15–60 years; (2) living in the catchment area; (3) experiencing their first-episode of psychosis; (4) no prior treatment with antipsychotic medication or, if previously treated, a total life time of adequate antipsychotic treatment of less than 6 weeks; (5) DSM-IV criteria for brief psychotic disorder, schizophreniform disorder, schizophrenia, or schizoaffective disorder. Patients were excluded for any of the following reasons: (1) meeting DSM-IV criteria for drug dependence, (2) meeting DSM-IV criteria for mental retardation, (3) having a history of neurological disease or head injury. Our operational definition for a"first-episode of psychosis" included individuals with a non-affective psychosis (meeting the inclusion criteria defined above) who have not received previously antipsychotic treatment regardless the duration of psychosis. Individuals who entered the study received extensive clinical and psychopathological assessments and were examined by MRI scan. Only those patients who had a baseline neuropsychological assessment and completed both baseline and 3-year follow-up MRI scans were included in this study. Thus, 59 patients with a schizophrenia-spectrum disorder (schizophrenia *N* = 37, 62.7%; schizophreniform disorder, *N* = 15, 25.4%; schizoaffective disorder, *N* = 2, 3.4%; brief psychotic disorder, *N* = 2, 3.4%; not otherwise specified psychosis, *N* = 3, 5.1%) were included in the study.

The diagnoses were confirmed using the Structured Clinical Interview for DSM-IV (SCID-I) (23) carried out by an experienced psychiatrist 6 months after the baseline visit.

A group of healthy subjects (*N* = 43) who had no current or past history of psychiatric illness, including substance dependence, neurological or general medical disorders, as determined by using an abbreviated version of the Comprehensive Assessment of Symptoms and History (CASH) (24), was recruited from the local area. The controls underwent the same cognitive battery as the patients. After a detailed description of the study, each healthy subject gave their written informed consent to participate, in accordance with the local ethics committee (25).

There were no significant differences in a variety of variables (e.g., age at baseline, age of onset, academic level, alcohol, cannabis, or tobacco consumption, duration of untreated psychosis (DUP), duration of untreated illness (DUI), or symptomatology factors) between patients who were and those who were not included in the final analysis (all *p* > 0.236).

### **CLINICAL ASSESSMENT**

Clinical symptoms were assessed by using the Brief Psychiatric Rating Scale total (BPRS) (26), the Scale for the Assessment of Negative Symptoms (SANS) (27), and the Scale for the Assessment of Positive Symptoms (SAPS) (28). We also divided psychopathology into three dimensions of symptoms: positive (scores for hallucinations and delusions), disorganized (scores for formal thought disorder, bizarre behavior, and inappropriate affect), and negative (scores for alogia, affective flattening, apathy, and anhedonia) (29).

#### **PREMORBID AND SOCIODEMOGRAPHIC VARIABLES**

Premorbid and sociodemographic information was recorded from patients, relatives, and medical records. Age of onset of psychosis was defined as the age when the emergence of the first continuous (present most of the time) psychotic symptom occurred. DUI was defined as the time from the first unspecific symptoms related to psychosis (for such a symptom to be considered, there should be no return to previous stable level of functioning) to initiation of adequate antipsychotic drug treatment. DUP was defined as the time from the first continuous (present most of the time) psychotic symptom to initiation of adequate antipsychotic drug treatment. Information about alcohol, cannabis, and tobacco consume were converted into binary variables coding for either the presence or absence of use.

#### **MEDICATION ASSESSMENT**

The amount and type of medication being prescribed during the 3-year follow-up period was thoroughly recorded. Patients were randomized as part of an intervention program out of the scope of the present study. After written informed consent was obtained, patients were randomly assigned to Haloperidol (*N* = 8), Olanzapine (*N* = 12), Risperidone (*N* = 12), Quetiapine (*N* = 8), Ziprasidone (*N* = 10), and Aripiprazole (*N* = 9). At 3-year followup patients were on: Haloperidol (*N* = 3), Olanzapine (*N* = 8), Risperidone (*N* = 12), Quetiapine (*N* = 5) Ziprasidone (*N* = 4), Aripiprazole (*N* = 8), Amisulpride (*N* = 1), Clozapine (*N* = 2), and Risperidone depot (*N* = 4). Eight patients withdrew from the medication, 27 patients switched their medication during followup period, and 2 were taking more than one antipsychotic at the time of follow-up MRI. No reliable information on medication intake was available for four patients. Additional information about concomitant medications is available under request.

## **NEUROPSYCHOLOGICAL ASSESSMENT**

For the present study, baseline neuropsychological assessment was considered in both groups, patients, and normal control subjects. Baseline patients' assessment was carried out when clinical status permitted in order to maximize cooperation, and occurred at a mean of 10.5 weeks after intake followed by assessment after 3 years. They were never assessed during a period of clinical exacerbation. A detailed description has been reported elsewhere (1).

For the analysis in this study a subset of measures was selected to assess eight major cognitive domains: (1) for measuring verbal memory we used the Rey Auditory Verbal Learning Test [RAVLT (30)], delayed recall; (2) for measuring visual memory we used Rey Complex Figure [RCF (31)], delayed reproduction; (3) for measuring executive functions we used Trail Making Test [TMT (32)], time to complete TMT-B; (4) for measuring working memory we used the Backward Digits scale [WAIS III (33)], total subscore; (5) for measuring speed of processing we used WAIS III subtest Digit Symbol, standard total score; (6) for motor dexterity we used Grooved Pegboard Handedness (34), time to complete with dominant hand; (7) for measuring attention we used Continuous Performance Test [CPT (35)], total number of correct responses; (8) The WAIS III subtest of Vocabulary was used as measure of premorbid IQ (34), standard total score.

In order to calculate a measure of Global Cognitive Functioning (GCF), raw cognitive scores were reversed when appropriate before standardization so they all have the same direction (the higher, the better). According to previous methodology (36), the GCF was calculated as *T*-scores (mean = 50, SD = 10) with raw scores of the healthy comparison sample. *T*-scores were converted to deficit scores that reflect presence and severity of cognitive deficit. Deficit scores on all tests were then averaged to create the GCF, which according to Keefe and colleagues deficit criterion (37), was dichotomized into two patients' subgroups: "deficit" (GCF <1) and "non-deficit" (GCF ≥1) [see Ref. (6) for details].

### **MRI DATA ACQUISITION**

A multimodal MRI protocol [T1, T2, and proton density (PD) sequences] was acquired at the University Hospital Marques de Valdecilla, Santander, Spain, using a 1.5-T General Electric SIGNA System (GE Medical Systems, Milwaukee, WI, USA). This multimodal approach was designed to optimize discrimination between gray matter, white matter, and cerebrospinal fluid. The T1-weighted images, using a spoiled grass (SPGR) sequence, were acquired in the coronal plane with the following parameters: echo time (TE) = 5 ms, repetition time (TR) = 24 ms, number of excitations (NEX) = 2, rotation angle = 45°, field of view (FOV) = 26 cm × 19.5 cm, slice thickness = 1.5 mm, and a matrix of 256 × 192. The PD and transverse relaxation time (T2)-weighted images were obtained with the following parameters: 3.0 mm thick coronal slices, TR = 3000 ms, TE = 36 ms (for PD) and 96 ms (for T2), NEX = 1, FOV = 26 cm × 26 cm, matrix = 256 × 192. The in-plane resolution was 1.016 mm × 1.016 mm. MRIs of patients and controls were evenly acquired during follow-up time.

## **IMAGE PROCESSING**

Processing of the images after acquisition was done by using a family of software programs called BRAINS2 (38, 39). The T1-weighted images were spatially normalized and resampled to 1.0-mm<sup>3</sup> voxels so that the anterior-posterior axis of the brain was realigned parallel to the anterior commissure/posterior commissure line and the interhemispheric fissure aligned on the other two axes. The T2- and PD-weighted images were then aligned to the spatially normalized T1-weighted image. These images were then subjected to a linear transformation into standardized stereotaxic Talairach atlas space to generate automated measurements of frontal, temporal, parietal, and occipital lobes and also the cerebellum and subcortical regions (39). To further classify tissue volumes into gray matter, white matter, and CSF, we used a discriminant analysis method of tissue segmentation based on automated training class selection that utilized data from the T1-weighted, T2-weighted, and PD sequences (40). The discriminant analysis method permits to identify the range of voxel intensity values that characterize GM, WM, and CSF. An 8 bit number is assigned to each voxel indicating its partial volume tissue content (10–70 for CSF, 70–190 for GM, and 190–250 for WM). In this study we examined the volumes of whole brain (WB),whole brain gray matter (WBGM),whole brain white matter (WBWM), cortical CSF (CCSF), and lateral ventricles (LV), gray and white matter volumes of cortical (occipital, parietal, temporal, and frontal lobes) and subcortical (caudate nucleus and thalamus) regions volume. Caudate and thalamus were delineated using a reliable and validated semiautomated artificial neural network (41). The procedure for measuring the volume of caudate and thalamus are explained in detail in previous studies (42, 43).

#### **STATISTICAL ANALYSIS**

The Statistical Package for Social Science, version 19.0 (SPSS Inc., Chicago, IL, USA), was used for statistical analyses. Significance was determined at the 0.05 level.

To examine brain volumetric differences between neurocognitive subgroups (no deficit vs. deficit) at baseline and 3-year follow-up, 1-way ANCOVA was performed. In each general linear model, the dependent measures were MRI volumes and the independent measure was group (no deficit vs. deficit). To test the hypothesis that the two groups would result in different progressive brain volume changes, repeated-measures analysis of covariance (repeated-measures ANCOVA) was performed for each ROI variable. The between-subject factor was group (no deficit vs. deficit) and the within subject factor was time (baseline and 3 year). Effects of time by group (interaction effect) were examined. Age, ICV, and premorbid IQ were included as covariates. There were no differences between groups related to age and ICV. However, there was a wide age range in our sample and the use of these two variables has been suggested in brain volume studies (44). The sample size (*n* = 59) provided sufficient power (>80%) to detect large effect sizes (*d*Cohen > 0.8) but was underpowered (45%) to detect weak or modest effects (*d*Cohen < 0.5).

Pearson correlation coefficients with age, ICV, and IQ as covariates were used to investigate possible statistical relationships between brain volume and GCF.

A prior directional hypothesis had been made for the brain measure analyses, thereby lessening the need for Bonferroni corrections. The analyses examining the relationships between brain measures and GCF were performed without prespecified hypotheses, and therefore Bonferroni adjustments were applied.

## **RESULTS**

Demographic and clinical data are shown in **Table 1**. Neuropsychological baseline assessment showed that of the 59 patients included in the study,30 (50.8%) presented general neurocognitive deficit. There were no statistically significant differences in relevant demographic and clinical characteristics between patients with neurocognitive deficit (*N* = 30) and patients without it (*N* = 29) at baseline (**Table 1**). However, the general neurocognitive deficit group showed a significant higher BPRS total score and greater severity of positive (SAPS total and positive dimension) and disorganized (scores for formal thought disorder, bizarre behavior, and inappropriate affect) symptoms at follow-up.

Neuropsychological data is presented in **Table 2**. The general neurocognitive deficit group had worse premorbid IQ, and showed consistently greater deficits all over cognitive domains. Worse executive function, poor motor dexterity, and particularly attentional deficits marked the more severely deficit patients.

Brain volumes at baseline in FEP subjects are presented in **Table 3**. There were no significant differences between neuropsychological subgroups in any of the brain regions studied at baseline [all *F*(1, 54) ≤ 3.42; all *p* ≥ 0.070] and follow-up [all *F*(1, 54) ≤ 3.43; all *p* ≥ 0.07] time points (**Table 3**). Patients with cognitive deficit showed overall lower gray and white matter volumes but these differences did not reach statistical significance.

There was a significant time by group interaction for the parietal tissue volume [*F*(1, 54) = 4.97, *p* = 0.030], with the general

**Table 2 | Comparison of neurocognitive groups on neuropsychological variables (Student' s t -distribution with 58 degrees of freedom).**


**Table 1 | Sociodemographic and clinical characteristics of the two neurocognitive groups of patients.**


BPRS, brief psychiatric rating scale; DUP, duration of untreated psychosis; DUI, duration of untreated illness; ICV, intracranial volume; SANS, scale for the assessment of negative symptoms; SAPS, scale for the assessment of positive symptoms; SD, standard deviation. Bold values were statistically significant (p < 0.05).



neurocognitive deficit group showing a greater volume decrease (1.67%) than the non-deficit group (0.13%). Similarly, there was also a significant time by group interaction for the total gray matter volume [*F*(1, 54) = 4.31, *p* = 0.042], showing a greater reduction in the general neurocognitive deficit group (2.71%) than in the non-deficit group (1.45%) (see **Figure 1**). Interestingly, when the analyses were controlled by possible confounding variables (sex, DUP, tobacco, cannabis, and alcohol consumption) only the parietal lobe tissue showed a significant group by time interaction.

No significant correlations between brain volume at baseline and GCF were found. At follow-up period, there were significant negative correlations between GCF and parietal tissue lobe (*r* = −0.29, *p* = 0.031) and temporal lobe gray matter (*r* = −0.27; *p* = 0.049). However, these correlations were weak and did not remain significant after correcting for multiple testing (Bonferroni correction).

## **DISCUSSION**

In the present study, a GCF index was calculated to identify schizophrenia patients who had general neurocognitive deficit at baseline (6). Contrary to our expectations, there were no brain volume differences between the cognitively impaired and cognitively preserved groups at any of the time-point studied. However, there were significant time by group interactions for the parietal tissue volume and the total gray matter volume, with the general neurocognitive deficit group showing a greater reduction in both regions during the 3-year follow-up interval. This is, to the best of our knowledge, the first study to examine progressive brain changes in schizophrenic cognitive deficit.

Neuropsychological assessment carried out at baseline indicated that 50.8% of the patients included in the study presented general neurocognitive deficit. These results are in full agreement with previous studies in FEP patients (2).

Only two previous cross-sectional studies (20, 45) have examined brain volume differences between cognitive subgroups in schizophrenia and their results have been inconclusive. Supporting our results, Ortiz-Gil and colleagues (45) did not find differences in lateral ventricular volume or WB volume between cognitively intact and cognitively deficit schizophrenia patients. Using Voxelbased morphometry, they alsofailed to detect significant difference in volumes of gray and white matter between those groups. However, Wexler et al. (20), using VBM to compare neuropsychology near normal and neuropsychology impaired subgroups, found that these groups differed significantly from each other in white matter volumes of the sensorimotor and parietal-occipital regions, with the neuropsychology impaired group showing smaller volumes in these brain regions. Nonetheless, it is of note that this VBM study did not adopt any statistical procedures to control for multiple comparisons. It is important to take into account that mean duration of illness among patients was above 18 years in both studies, while our patients had a shorter duration of illness (non-deficit: 28.32 months, deficit: 30.45 months).

Only cross-sectional studies have addressed the relationship between cognitive deficit and brain structure in schizophrenia. However, it has been suggested that this relationship may not be adequately assessed in a cross-sectional study (45). We found significant time by group interactions for the parietal tissue volume and the total gray matter volume, with the general neurocognitive deficit group showing a greater reduction in both regions (parietal tissue volume and the total gray matter volume) during the 3-year follow-up interval. In a recent study (13) we found that brain tissue volumes decrease in patients at early years after the first episode was similar to that found in healthy controls. Although several longitudinal studies in schizophrenia have described a greater degree of brain tissue volumes decrease in the early stage of the illness (10–12), others have failed to confirm these findings (13–18). For a review see Olabi et al. (46). Taken this together, we might speculate that the progressive brain volume loss found in schizophrenia might be associated with this general cognitive deficit patients' subgroup.

## **LIMITATIONS**

A uniform follow-up interval using the same MRI scanner and protocol, and a thorough clinical investigation during the followup period add strength to the conclusions drawn from this study. However, several limitations should be taken into account when interpreting the results of the current study. First, the diagnostic heterogeneity of the sample might bias our findings. Second, and given the fact that schizophrenia is a life-long disease, a follow-up period of 3 years may be too short to demonstrate other subtle changes. Third, analyses taking into account if neurocognitive function maintained, declined, or improved during follow-up could not be conducted because of the small sample size in our study. Fourth, a major confounding factor could be the intake

of antipsychotic medication (47). Some studies have showed a relationship between antipsychotic medication use and longitudinal brain volume change in schizophrenia (10, 48), although others have failed to clearly demonstrate an influence of antipsychotic medication on brain volume change (49–51). Some patients withdrew from their medication, and most of them switched medication during the 3-year follow-up period, which makes the investigation of the effects of different types of antipsychotics an unfeasible study. Fifth, this study only measured a number of structures, and finally, brain volume changes in schizophrenia are subtle, so the sample size might be considered small to make any definitive assertions. While our data provided sufficient power to detect large effects, the detection of weak effects requires large study populations.

## **CONCLUSION**

In conclusion, our results, based on a representative sample of firstepisode schizophrenia-spectrum patients, do not confirm the presence of significant morphometric differences between cognitively impaired and cognitively preserved schizophrenia patients at the

## **REFERENCES**


R, Ayuso-Mateos JL, et al. Predictors of neurocognitive impairment at 3 years after a first episode non-affective psychosis. *Prog Neuropsychopharmacol Biol Psychiatry* (2013) **43**:23–8. doi:10.1016/j.pnpbp.2012.11.012


early stages of the illness. However, there were significant time by group interactions for the parietal tissue volume and the total gray matter volume during the 3-year follow-up period, which might indicate that cognitive deficit in schizophrenia would be associated with progressive brain volume loss. Further investigations are warranted to fully clarify the relationship between cognitive deficit and brain structure in schizophrenia.

## **ACKNOWLEDGMENTS**

The present study was carried out at the Hospital Marqués de Valdecilla, University of Cantabria, Santander, Spain, under the following grant support: Instituto de Salud Carlos III PI020499, PI050427, PI060507, PI1000183, SENY Fundació Research Grant CI 2005-0308007, and Fundación Marqués de Valdecilla API07/011. We wish to thank the PAFIP researchers. Adele Ferro was sustained by the funds of the 2007/2013 European Social Fund Operational Programme of the Autonomous Region Friuli Venezia Giulia. In addition, we should like to acknowledge the participants and their families for enrolling in this study.

(2003) **60**(6):585–94. doi:10.1001/ archpsyc.60.6.585


RR, Santos LC, et al. Lack of progression of brain abnormalities in first-episode psychosis: a longitudinal magnetic resonance imaging study. *Psychol Med* (2011) **41**(8):1677–89. doi:10.1017/ S0033291710002163


Gonzalez-Blanch C, et al. Reduced thalamic volume in first-episode non-affective psychosis: correlations with clinical variables, symptomatology and cognitive functioning. *Neuroimage* (2007) **35**(4):1613–23. doi:10.1016/j. neuroimage.2007.01.048


JM, Tordesillas-Gutierrez D, et al. Effect of antipsychotic drugs on brain morphometry. A randomized controlled one-year follow-up study of haloperidol, risperidone and olanzapine. *Prog Neuropsychopharmacol Biol Psychiatry* (2008) **32**(8):1936–43. doi:10.1016/j.pnpbp.2008.09.020

**Conflict of Interest Statement:** The study, designed and directed by Benedicto Crespo-Facorro, conformed to international standards for research ethics and was approved by the local institutional review board. Unrestricted educational and research grants from AstraZeneca, Pfizer, Bristol-Myers Squibb, and Johnson & Johnson provided support to PAFIP activities. No pharmaceutical industry has participated in the study concept and design, data collection, analysis and interpretation of the results, and drafting the manuscript. Prof. Crespo-Facorro has received unrestricted research funding from AstraZeneca, Pfizer, Bristol-Myers Squibb, and Johnson & Johnson that was deposited into research accounts at the University of Cantabria. Prof. Crespo-Facorro has received honoraria for his participation as a speaker at educational events from Pfizer, Bristol-Myers Squibb, and Johnson & Johnson and consultant fees from Pfizer. The rest of the authors report no additional financial.

#### *Received: 05 September 2013; accepted: 01 October 2013; published online: 17 October 2013.*

*Citation: Ayesa-Arriola R, Roiz-Santiáñez R, Pérez-Iglesias R, Ferro A, Sainz J and Crespo-Facorro B (2013) Neuroanatomical differences between first-episode psychosis patients with and without neurocognitive deficit: a 3-year longitudinal study. Front. Psychiatry 4:134. doi: 10.3389/fpsyt.2013.00134 This article was submitted to Schizophre-*

*nia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Ayesa-Arriola, Roiz-Santiáñez, Pérez-Iglesias, Ferro, Sainz and Crespo-Facorro. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, providedthe original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Near-infrared spectroscopy in schizophrenia: a possible biomarker for predicting clinical outcome and treatment response

#### **Shinsuke Koike1,2\*,Yukika Nishimura<sup>2</sup> , Ryu Takizawa2,3, NoriakiYahata<sup>4</sup> and Kiyoto Kasai <sup>2</sup>**

<sup>1</sup> Office for Mental Health Support, Division for Counseling and Support, The University of Tokyo, Tokyo, Japan

<sup>2</sup> Department of Neuropsychiatry, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan

<sup>3</sup> Social, Genetic and Developmental Psychiatry Centre, Institute of Psychiatry, King's College London, London, UK

<sup>4</sup> Department of Youth Mental Health, Graduate School of Medicine, The University of Tokyo, Tokyo, Japan

#### **Edited by:**

Stefan Borgwardt, University of Basel, Switzerland

#### **Reviewed by:**

Takahiro A. Kato, Kyushu University, Japan Masato Fukuda, Gunma University, Japan

#### **\*Correspondence:**

Shinsuke Koike, Office for Mental Health Support, Division for Counseling and Support, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan e-mail: skoike-tky@umin.ac.jp

Functional near-infrared spectroscopy (fNIRS) is a relatively new technique that can measure hemoglobin changes in brain tissues, and its use in psychiatry has been progressing rapidly. Although it has several disadvantages (e.g., relatively low spatial resolution and the possibility of shallow coverage in the depth of brain regions) compared with other functional neuroimaging techniques (e.g., functional magnetic resonance imaging and positron emission tomography), fNIRS may be a candidate instrument for clinical use in psychiatry, as it can measure brain activity in naturalistic position easily and non-invasively. fNIRS instruments are also small and work silently, and can be moved almost everywhere including schools and care units. Previous fNIRS studies have shown that patients with schizophrenia have impaired activity and characteristic waveform patterns in the prefrontal cortex during the letter version of the verbal fluency task, and part of these results have been approved as one of the Advanced MedicalTechnologies as an aid for the differential diagnosis of depressive symptoms by the Ministry of Health, Labor and Welfare of Japan in 2009, which was the first such approval in the field of psychiatry. Moreover, previous studies suggest that the activity in the frontopolar prefrontal cortex is associated with their functions in chronic schizophrenia and is its next candidate biomarker. Future studies aimed at exploring fNIRS differences in various clinical stages, longitudinal changes, drug effects, and variations during different task paradigms will be needed to develop more accurate biomarkers that can be used to aid differential diagnosis, the comprehension of the present condition, the prediction of outcome, and the decision regarding treatment options in schizophrenia. Future fNIRS researches will require standardized measurement procedures, probe settings, analytical methods and tools, manuscript description, and database systems in an fNIRS community.

**Keywords: near-infrared spectroscopy, verbal fluency task, biological markers, early intervention, clinical outcome**

## **INTRODUCTION**

Techniques that allow the easier and less invasive measurement of brain structure and activity, such as magnetic resonance imaging (MRI), functional MRI, and positron emission tomography (PET), have progressed rapidly over the past 20 years. There has been considerable expectation regarding the clinical application of neuroimaging techniques to psychological conditions and psychiatric illnesses (1, 2). Biological markers measured using neuroimaging instruments would clarify the pathophysiological features of psychiatric disorders, inform patients and family members regarding their actual conditions, and improve general impression of psychiatric disorders, by more easily giving the explanation of their conditions and discussing their assumed prognoses.

Functional near-infrared spectroscopy (fNIRS) is a functional brain imaging tool that can measure hemoglobin changes over the surface of the brain easily and non-invasively (**Figure 1**) (3–5). fNIRS technique was found in 1977 (6), and has been applied to measure brain hemodynamic activity through the scalp (7, 8). The release of commercial fNIRS machines that are small, movable, and work silently during the last decade (**Figure 1B**) has allowed the progress of fNIRS research in the field of psychiatry (9). In 2012, more than 100 studies were published on this subject; among these, about 20 articles pertained to the field of psychiatry (9). Part of the results of those fNIRS studies has been approved as one of the Advanced Medical Technologies as an aid for the differential diagnosis of depressive symptoms by the Ministry of Health, Labor and Welfare of Japan in 2009 (10–12), as the presence of different characteristic waveform patterns in the prefrontal cortex (PFC) during a verbal fluency task (VFT) has been reported among patients with major depressive illness (4, 12–17), bipolar disorder (12, 13), and schizophrenia (4, 5, 12). This was the first such approval in the field of psychiatry in Japan (see the "Application to supplementary diagnostic tool for psychiatric disorders" subsection). Here, we reviewed fNIRS research that focused on schizophrenia, which

is currently the most published topic in the application of fNIRS in the field of psychiatry, and address future investigations that are needed for the clinical application of this technique as an aid for the differential diagnosis, comprehension of present conditions, prediction of outcome, and decision regarding treatment options in schizophrenia.

## **PRINCIPLES OF BRAIN ACTIVITY MEASUREMENT USING NEAR-INFRARED LIGHT**

## **THEORETICAL BACKGROUND OF THE MEASUREMENT OF BRAIN ACTIVITY USING fNIRS INSTRUMENTS**

Near-infrared light, especially that with a wavelength of 650– 1000 nm, has characteristics that include a relatively high absorption through hemoglobin, as well as a relatively high penetration through bone and skin, compared with light with other wavelengths. The emission of near-infrared light from a source probe on the human scalp leads to its passing and scattering through brain tissues with relatively low absorption, followed by the absorption of part of this light by blood hemoglobin in small vessels (<1 mm) (18). A detector probe, which is normally placed 3 cm away from the source probe in adults, can detect scattered near-infrared light that is reflected by the surface of the cortex (**Figure 1A**). Therefore, the region located between the source and the detector probes is

generally set as a measurement area that is often called a "channel" (**Figure 1C**). fNIRS instruments can measure oxygenated hemoglobin (O2Hb) and deoxygenated hemoglobin (HHb), as well as total hemoglobin (tHb) by summing up O2Hb and HHb, by using near-infrared light with two or more different wavelengths that are slightly different from the absorption rates of O2Hb and HHb. In accordance with light manipulation, fNIRS instruments are roughly divided into three types: continuous wave (CW), frequency domain (FD), and time domain (TD) instruments. Although CW-type fNIRS machines are unable to measure absolute hemoglobin concentration in tissues, they are relatively small, have low installation and maintenance costs, and are able to perform measurements at a high sampling rate compared with FD- and TD-type fNIRS machines (18). Therefore, recent clinical studies have used CW-type fNIRS machines (19, 20). The theoretical details of all types of fNIRS instruments were reviewed elsewhere (18, 20, 21).

## **ADVANTAGES AND DISADVANTAGES OF fNIRS COMPARED WITH OTHER FUNCTIONAL IMAGING INSTRUMENTS**

The advantages and disadvantages of CW-type fNIRS compared with other functional imaging instruments (fMRI, PET, and EEG) are shown in **Table 1**. fNIRS has the following advantages: (1) non-invasiveness, which allows repetitive measurements, even in infants (21, 22); (2) easy setting; (3) small size and portability; (4) high temporal resolution compared with fMRI and PET (18, 20); (5) possibility of performing measurements in a non-restrained position, such that participants can sit on a chair, talk, and move their hands; and (6) possibility of relatively easily combining measurements with other neuroimaging techniques, such as EEG (23, 24), MRI (25–30), PET (31–33), and magnetoencephalography (MEG) (34, 35). Conversely, fNIRS has the following disadvantages: (1) low spatial resolution (10–30 mm); (2) possibility of performing measurements only at the surface of the cortex; (3) inability to measure absolute hemoglobin value (CW-type); and (4) the data obtained can be influenced by scalp, muscle, skull, and cerebrospinal fluid factors in addition to hemodynamic changes in the cortex (20).

Other than task-related hemoglobin changes derived from neural activities under neurovascular coupling, fNIRS signals vary depending on task-related blood pressure changes and skin blood flow, as well as spontaneous brain activity related to heart rate, respiration, and physiological oscillations (20, 21). Simultaneous measurements by fNIRS and pulse Doppler sonography or by fNIRS with different probe distances [e.g., shallow (5 mm) and deep (30 mm)] allow the distinction of fNIRS signals from the cortex from those of other brain tissues. A study based on fNIRS with different probe distances and laser Doppler velocimetry showed that hemoglobin changes from the cortex during the VFT may contain only 6% of fNIRS data in the PFC, and that most signal changes may arise from changes in skin blood flow (36). The results OPF simultaneous fNIRS, fMRI, MR angiography, and peripheral physiological measurements suggested that task-evoked sympathetic arterial vasoconstriction affected fNIRS signal changes substantially (37). However, a subsequent study in which simultaneous measurements using multiple fNIRS probe


**Table 1 | Comparison of CW-type fNIRS machines with other neuroimaging tools.**

Bold shows advantages compared to other neuroimaging tools.

<sup>a</sup>ETG-4000 (Hitachi Medical Corporation).

distances and laser Doppler flowmetry were performed suggested that about 50% of fNIRS signals contributed to the fNIRS signal component in the deep layer, mostly measured in the cortex, during several cognitive tasks, including the VFT (38). A study of simultaneous measurement using fMRI, fNIRS, and lase Doppler flowmeter showed that the prefrontal fNIRS signals were significantly correlated with the blood oxygenation level-dependent (BOLD) signals in the gray matter rather than those in the soft tissue or the laser Doppler signals (39). Previous studies have suggested that the ratio of fNIRS signal changes in the cortex varies from 6 to 60%, and the variation has been considered as being caused by differences in measuring instruments, estimation methods, and brain areas measured (20, 36–38, 40–44). Several studies have provided filtering methods that allow raising the ratio of task-related hemoglobin changes in the cortex to in the extracortical tissues; however, these systems require additional probes and complex analytical methodologies because of different nearinfrared absorption and scattering coefficients in each tissue and anatomic characteristics throughout the light path in each brain area (25, 38, 40, 41). Therefore, fNIRS is a reliable tool for research based on the group-level and/or channel-cluster-level investigations, although its reliability at the individual and single-channel levels is not sufficient (18, 45–48). Few studies in clinical psychiatry have considered these filtering methods because of the limitations of measurement time and setting. Future studies will be needed to improve the signal-to-noise ratio for application over wide measurement areas without losing the advantage of fNIRS instruments.

## **COMPARISON OF fNIRS HEMOGLOBIN CHANGES WITH BOLD SIGNALS IN fMRI**

The measurement of brain activity using the fMRI technique has progressed during the last two decades. The BOLD signal is thought to represent the differences in the magnetic properties of deoxygenated hemoglobin concentration under T2 weighted measurement conditions when cerebral blood flow (CBF) increases and HHb decreases in small vessels (49). The theoretical model of BOLD compared with fNIRS signals was reviewed in detail elsewhere (18). Briefly, localized O2Hb decreases and HHb increases occur (i.e., initial dip) when neurons are activated in a specific region. Several seconds later, the blood flow system is triggered to request glucose to the region (i.e., hemodynamic response), which is followed by a CBF increase and peripheral vascular bed dilation, leading to tHb increase and HHb decrease in small vessels, and O2Hb increase in blood capillaries and vascular bed. As this neurovascular coupling that occurs in the activated area is thought to vary according to brain area and vessel diameter, and to be sensitive to persistent neural activity, the relationship between hemoglobin changes and BOLD/fNIRS signals has a complex variation pattern.

Other than the difference in spatio-temporal measurement characteristics between BOLD and fNIRS signals, the BOLD signal is thought to detect mainly changes in the magnetic properties of small vessels, whereas the fNIRS signal is thought to detect changes in near-infrared light absorption in blood capillaries (50). Therefore, several discrepancies may occur between BOLD and fNIRS signals. Previous fNIRS studies have yielded inconsistent results compared with other imaging tools. Simultaneous fNIRS and fMRI studies have shown that the BOLD signal in a specific region was associated with the O2Hb of fNIRS signal changes in the corresponding region (25, 29, 39, 51) or with HHb changes (26–28, 30, 51). One of the explanations for these inconsistent correlation results is that most fNIRS studies have been conducted using block-designed tasks, whereas fMRI studies have mainly used an event-related design. In addition, the analysis of fMRI data by software such as Statistical Parametric Mapping (SPM) uses a hypothesis that fits a probable activation model, whereas most fNIRS studies have analyzed average signal intensity during the whole task period without any probable activation model. Differences in acquired signal handling may result in discrepancies regarding regions with significant brain activity. As O2Hb data generally exhibit larger changes compared with HHb during cognitive activity, most clinical fNIRS studies have mainly been analyzed using O2Hb data.

## **ESTIMATION OF BRAIN AREA**

As more studies have focused on the spatial characteristics of brain activity using multi-channel fNIRS instruments, the need to clarify the estimated location of each probe in the cortex and each channel on the scalp has arisen. As reviewed by Tsuzuki et al. recently (52), several methodologies can estimate brain areas at each probe and channel, such as structural MRI measurement using fNIRS probe marks for each participant (53, 54), a probabilistic registration method using a 3D digitizer (55), and a probabilistic virtual registration method without any additional instrument (56). As the virtual registration method enables the estimation of brain areas based on standard brain images at each probe and channel by defining only probe setting based on the 10–20 system electrode locations, and because this method has similar accuracy compared with other estimation methodologies (56), most fNIRS studies have applied this method to estimate the measurement brain areas of channels. fNIRS software can also use the virtual registration method as a toolbox (http://www.jichi.ac.jp/brainlab/tools.html) (57). There are several standard stereotaxic coordinate systems such as the Montreal Neurological Institute (MNI) and the Talairach Daemon. The LONI Probabilistic Brain Atlas (LPBA40) (58) system has been often used in fNIRS studies based on probabilistic registration methods (3, 52, 59, 60).

## **APPLICATION OF fNIRS IN SCHIZOPHRENIA RESEARCH**

Schizophrenia is a syndrome that is characterized by positive and negative symptoms and cognitive dysfunction with enduring social deficits. Moreover, it affects approximately 0.7% of the general population (2). The World Health Organization reported that the estimated burden of schizophrenia accounts for 2.3% of all diseases worldwide, and its disability-adjusted life year ranks ninth among all non-communicable diseases (61). However, effective treatments and objective indices for all symptoms and functions of schizophrenia have not been fully met.

Since the first fNIRS report of altered activation patterns in schizophrenia compared with healthy controls was published in 1994 (62), fNIRS research focusing on schizophrenia has been the most published topic in the field of psychiatry (9). We reviewed systematically research articles published up to April 1, 2013, by searching PubMed and Web of Science. As in previous systematic reviews (9, 18), we used "[(near infrared) OR (optical topography)] AND (schizo\* OR psycho\*) AND (brain OR cortex)" as a search term. Two hundred and sixty articles were extracted, among which 29 explored brain activity in patients with schizophrenia. Half of these articles (15 articles, including 4 studies of genetic variants) adopted a VFT as an activation cognitive battery during measurement.

## **VERBAL FLUENCY TASK**

The VFT is a popular cognitive task that is used in neuropsychological tests and neuroimaging measurements to explore various cognitive functions during verbal recall, retrieval, working memory, attention, and inhibition (avoiding inappropriate words) (63, 64). During the task, participants are instructed to say as many words from a given paradigm as possible in a given time (usually 60 s). This paradigm is roughly divided into semantic (category fluency task, CFT), such as fruits, or phonological (letter fluency task, LFT), such as words that begin with the letter "p." Neuropsychological studies have revealed that patients with schizophrenia have worse VFT performances compared with healthy controls (65). Although fMRI and fNIRS studies have shown that relatively global brain activity occurs during the VFT compared with a task that requires specific cognitive domains, such as the n-back working-memory task (28, 66) and the go/no-go task (67), most participants (including patients with chronic schizophrenia) can perform the task (10, 11, 19, 63, 64).

Eleven previous VFT studies are listed in **Table 2**. Watanabe and Kato firstly described that patients with schizophrenia had reduced O2Hb and HHb changes in the left PFC during the LFT compared with healthy controls (68). This study also demonstrated that patients who were medicated with atypical antipsychotics had better task performances and similar O2Hb changes compared with controls, suggesting that typical antipsychotics may impair task response and brain activity. However, that study did not explore whether impaired O2Hb changes were derived from impaired task response or functional impairment in the PFC, which may be ameliorated by atypical antipsychotics. Suto et al. firstly described the spatio-temporal activity patterns in the PFC and temporal cortex among patients with depression and schizophrenia and healthy controls using two 24-channel fNIRS machines (4). In that study, a modified task procedure was adopted in which three initial syllables changed in turn every 20 s during a 60 s task period, to help participants avoid silence and reduce differences in task performances among groups. Patients with schizophrenia had lower activity in the bilateral PFC and temporal cortex at the start of the task period compared with controls, whereas patients with depression had lower activity in the bilateral PFC and temporal cortex across the task period. These results were irrespective of task performance, and the task paradigm used in that study was used widely in further studies. In addition, this result was based on the Advanced Medical Technologies in Japan (10–12). Using a larger cohort, Takizawa et al. replicated the observation that patients with schizophrenia had slower and inappropriate activity after the task period compared with healthy controls (5). Furthermore, activities in the frontopolar prefrontal cortex (FPC) region were positively

#### **Table 2 | Previous fNIRS studies in schizophrenia using verbal fluency tasks.**



(Continued)

#### **Table 2 | Continued**


**Reference Correlational analysis between clinical variables and fNIRS signals**


Excluded previous gene association studies.

NA, not applicable; n.s., no significant correlation; FPC, the frontopolar prefrontal cortex; DLPFC, dorsolateral prefrontal cortex; VLPFC, ventrolateral prefrontal cortex; FDR, false discovery rate; GAF, the global assessment of functioning; PANSS, the positive and negative symptom scale; FEP, first-episode psychosis, UHR, ultra-high risk.

<sup>a</sup>OMM-3000/16, Shimadzu Corporation; ETG-4000 and ETG-100, Hitachi Medical Corporation; NIRO-300 and NIRO-200, Hamamatsu Photonics Corporation; HEO-200, Omron Healthcare Corporation.

<sup>b</sup>Exhibit case in the chronic schizophrenia group. Sixteen UHR and 2 FEP individuals were antipsychotics naïve and 8 UHR and 1 FEP individuals were drug naïve.

<sup>c</sup>As Azechi et al. (70) used the same sample from Ikezawa et al. (71) as the First group, we describe for the Second group.

<sup>d</sup>By Azechi et al. (70).

<sup>e</sup>Results from brain activity during the whole of the task period.

<sup>f</sup>Results in the chronic schizophrenia group.

<sup>g</sup>Results during the letter fluency tasks.

associated with global assessment of functioning scores in schizophrenia. Quaresuma et al. replicated the finding of reduced brain activity in schizophrenia during the LFT, whereas no significant change was observed during a visual spatial working memory task (69). Ikezawa et al. and Azechi et al. showed the efficacy of the LFT task in fNIRS (70, 71). Ikezawa et al. measured hemoglobin changes in the PFC using a two-channel fNIRS instrument during the LFT, CFT, Tower of Hanoi (TOH), the Sternberg task, and the Stroop task, and showed that brain activities during the LFT and TOH were significantly different between patients with schizophrenia and healthy controls (71). Azechi et al. explored this further using discrimination analysis and showed that 88.3% of participants correctly discriminated between patients and controls based on task performance on the TOH, LFT, and CFT, and on fNIRS signals during the VFT (70). This result confirmed that 75% of independent participants were able to discriminate correctly using the same procedure. Koike et al. explored the signal differences among different clinical stages of schizophrenia (ultrahigh risk, first-episode psychosis, and chronic schizophrenia) and showed that the activities in the FPC, ventrolateral PFC (VLPFC), and temporal cortex were lower in patients than they were in controls, whereas the activities in the dorsolateral PFC (DLPFC) decreased with advancing clinical stage (3). Those authors also replicated the finding that the activity in the FPC region was positively associated with global assessment of functioning scores in chronic schizophrenia, implying that it may be a candidate biomarker for the assessment of psychological condition in schizophrenia (3, 5). Takeshi et al. measured brain activity during the idea fluency task, which is thought to require more executive function, and showed that patients with schizophrenia had decreased activity in the ventral area of the PFC (72). Furthermore, these signal changes were positively associated with the global assessment of functioning scores, whereas O2Hb changes during the LFT were not. Shimodera et al. replicated the characteristic waveforms in schizophrenia, such as smaller initial activity at the start of the task period, reduced activity during the task period, and inappropriate activity after the task period, and used a numerical calculation (73).

The CFT has also been used in fNIRS studies, which often compare this task to the LFT. Kubota et al. used a two-channel fNIRS instrument to show for the first time that healthy controls had larger activity in the PFC during the LFT than in the CFT under similar task performances, whereas patients with schizophrenia had smaller activity during the LFT than during the CFT (75). Patients with schizophrenia had smaller activity than did healthy controls under similar task performances between the groups. This result is consistent with those of other studies (70, 71). Ehlis et al. used a multi-channel fNIRS instrument that covered the left frontotemporal region and found that healthy controls had larger and spatially wider activities during the LFT compared with the CFT (74), which was then replicated by measuring wider areas of the bilateral prefrontal/temporal cortices (76). Ehlis et al. also replicated the finding that patients with schizophrenia had significantly reduced activities compared with healthy controls during LFT, but not during the CFT.

## **APPLICATION OF fNIRS AS A SUPPLEMENTARY DIAGNOSIS TOOL FOR PSYCHIATRIC DISORDERS**

As described above, previous studies using a block-design LFT have indicated that patients with schizophrenia have not only reduced activity, but also inappropriate activity timing, especially at the start of the task period and post-task period, compared with healthy controls (4, 5). Subsequently, the Joint Project for Psychiatric Application of Near-Infrared Spectroscopy (JPSY-NIRS) Group has applied to these results and improved in the applicable way for clinical settings. The integral value (the size of the fNIRS signal area during the task period) and centroid value (the centroid time of the fNIRS signal area throughout the task) were determined by using averaged brain signals estimated in the frontopolar cortex. This group showed preliminarily that 69% of patients with MDD and 69% of patients with schizophrenia, and 69% MDD patients and 81% BP patients were correctly differentiated under an algorithm using these two values (10, 11). A part of these results was approved as one of the Advanced Medical Technologies as an aid for the differential diagnosis of depressive symptoms in 2009, which was the first such approval in the field of psychiatry in Japan (10, 11). Several criticisms have arisen regarding the limited replication in various clinical settings and the lack of consensus for application to mental health (77, 78). However, paper published recently on JPSY-NIRS replicated previous results (12). Using this algorithm, fNIRS can differentiate patients with depressive symptoms between major depressive disorder and psychotic disorders (bipolar disorder and schizophrenia), with a high classification rate (74.6 and 85.5%, respectively).

## **OTHER COGNITIVE TASKS AND MEASUREMENT SETTINGS**

Three studies have explored the differences in brain activity during a random generation task (RNG) between patients with schizophrenia and healthy controls. Sinba et al. used a two-channel fNIRS instrument to describe for the first time that patients with schizophrenia had reduced brain activity in the PFC and worse task performances during the RNG compared with healthy controls, which represented different features during ruler-catching and sequential finger-to-thumb tasks (79). As healthy controls with better RNG task performances had greater brain activity, it remained unclear whether low brain activity was derived from worse task execution and/or functional impairment in schizophrenia. Hoshi et al. used time-resolved spectroscopy and two-channel fNIRS instruments to show that patients with schizophrenia, particularly those with a longer duration of illness, had reduced hemoglobin concentration during the resting state, and that this may cause altered activity during the RNG task (80). Koike et al. used a multi-channel fNIRS instrument to show that patients with schizophrenia had significantly reduced activity in the bilateral DLPFC and VLPFC regions, and that the activity in the right DLPFC region was associated with an earlier age at onset (81).

Okada et al. firstly showed the presence of altered brain activity in schizophrenia using a two-channel fNIRS instrument (62). Patients with schizophrenia had an aberrant task-related response pattern during a mirror drawing task, which was thought to be derived from a disrupted interhemispheric integration. Fallgatter et al. also showed the presence of altered frontal lateralization in schizophrenia during a continuous performance task using a twochannel fNIRS instrument (82). Folley et al. explored brain activity in patients with schizophrenia, individuals with schizotypal personality, and healthy controls during a divergent-thinking task using a two-channel fNIRS instrument, and showed that individuals with schizotypal personality had enhanced divergent-thinking ability and greater brain activity in the right PFC compared with patients with schizophrenia and healthy controls (83). Lee et al. showed alternations in brain activity using fMRI and 24-channel fNIRS (not simultaneously) during the same event-related spatial working-memory task: patients with schizophrenia recruited the bilateral PFC, whereas healthy controls recruited only the right PFC (28). Zhu et al. used a 48-channel fNIRS instrument to show that patients with first-episode schizophrenia had reduced brain activity over the PFC during the Tower of London task (84). Nishimura et al. used a 52-channel fNIRS instrument during a go-no-go task to show that healthy controls had a significant decrease in activity in the DLPFC during the no-go condition, whereas patients with schizophrenia exhibited no changes (67). Furthermore, the high excitement score observed in patients with schizophrenia was associated with brain activity in the FPC and right DLPFC. Taniguchi et al. used a 24-channel fNIRS instrument to show that patients with schizophrenia had reduced brain activity in the PFC compared with healthy controls during a kana Stroop task, with similar task performances, whereas both patients and controls showed lack of activity during a kanji Stroop task, with significantly worse performance observed in the schizophrenia group (85).

Other than cognitive tasks, Fujita et al. explored hemoglobin changes through an electroconvulsive therapy using a two-channel fNIRS instrument, and showed that patients with schizophrenia had asymmetric hemoglobin changes in the PFC compared with patients with depression (86).

## **GENETIC VARIATION**

Although schizophrenia is a syndrome, has been considered as a consolidation of several pathophysiological features, and has high genetic heritability, no crucial genetic risk factor has been found (87, 88). The results of genome-wide association studies that used a large sample size have suggested that schizophrenia risk genes are unable to be determined by specific gene variants but are thought to consist of common variants; furthermore, these risk genes had a substantial influence on environmental effects that occurred before the onset of schizophrenia. To clarify the impact of specific genes related to schizophrenia on the brain, imaging/genetics studies were performed to explore the relationship between brain structure and activity and genetic variants.

Five studies have explored the relationship between genetic variants and brain activity using fNIRS instruments. Takizawa et al. firstly reported that the val108/158met polymorphism of the catechol-*O*-methyltransferase (*COMT*) gene affected brain activity in the PFC only in schizophrenia patients (and not in healthy controls) (89). Schizophrenia patients with the Met variant (Val/Met and Met/Met) had significantly greater activation in the bilateral FPC and DLPFC during the LFT than did Val/Val carriers, implying that the inverted U curve shift of dopamine availability in schizophrenia might have an effect on the brain activity in the PFC (89). Regarding the Gln/Pro polymorphism of the sigma-1 receptor gene, Takizawa et al. reported that patients with the Gln/Gln genotype had significantly greater brain activity in the FPC and left DLPFC during the LFT than did Pro allele carriers; however, no significant differences were observed in healthy controls (90). However, Ohi et al. later used a two-channel fNIRS machine and a larger sample set to show that this genotype effect occurred in both the schizophrenia and control groups (91). Ohi et al. also reported

that individuals with a longer cytosine/adenine/guanine (CAG) repeat in the spinocerebellar ataxia type 17 gene had reduced activity in the bilateral PFC during the TOH task in both the schizophrenia and control groups (92). Regarding the rs41279104 polymorphism of the nitric oxide synthase-I gene, Reif et al. showed that patients who were A allele carriers had significantly reduced activity in the right PFC during a VFT compared with those who had the GG genotype (93).

As recent methodological progress in gene analysis allows the exploration of whole genetic alterations between cases and controls using more than 10,000 samples, imaging/genetics studies should be performed using methods that enable the analysis of numerical data sets, such as bioinformatics and machine learning methods (94). Conversely, the investigation of the relationship between brain activity and relevant target genes (e.g., DLPFC function under dopamine regulation and *COMT* variants) may provide another imaging/genetics study strategy. The analysis of altered gene function in the brain among psychiatric illnesses may clarify the pathophysiology of specific psychiatric disorders and identify new treatment options.

## **EFFECT OF MEDICATION ON BRAIN ACTIVITY**

As previous clinical fNIRS studies have mostly explored patients in the chronic stage and receiving medication, the effect of medication on brain function was a limitation of these studies. Although an inconsistent effect of medication on fNIRS signal has been reported (68, 70, 71, 79), most previous fNIRS studies have reported an absence of association between brain activity and medication dose (3, 5, 12, 74, 81, 84–86) or different brain activity between individuals with ultra-high risk for psychosis with and without medication (3) (**Table 1**). However, all of those studies were cross-sectional, and a previous randomized and controlled trial showed that the administration of mirtazapine increased brain activity compared with trazodone and placebo in healthy volunteers (95). Controlled trials and/or longitudinal investigation to elucidate specific drug effects will be needed.

## **LIMITATIONS OF PREVIOUS STUDIES AND FURTHER DIRECTIONS**

The previous fNIRS studies of schizophrenia had several limitations; therefore, we propose further directions for future investigation (**Table 3**). First, as most previous studies have performed cross-sectional measurements in chronic and stable patients receiving medication, symptomatic, and functional changes were not fully explored. As fNIRS is able to perform measurements relatively easily in unstable patients, such as those with acute or recurrent conditions, longitudinal studies aimed at investigating changes in clinical symptoms and social function will be needed (3, 96). Although previous studies have revealed a negligible medication effect on fNIRS signals, investigations of drug-naïve patients or of those receiving controlling treatment (e.g., specific drugs, electroconvulsive therapy, and neurofeedback) will be needed to allow further clinical applications of fNIRS (60, 86, 97). The use of the easy portability of fNIRS machines may allow measurements in earlier clinical stages in cohort settings, such as adolescents with psychotic-like experiences, which may reveal the alterations in brain development in the PFC (98).


#### **Table 3 | Limitation of previous fNIRS studies and further implication.**

UHR, ultra-high risk; FEP, first-episode psychosis.

Second, half of the previous studies of schizophrenia adopted block-design VFTs for cognitive activation. Although one of the major disadvantages of fNIRS is the inability to measure brain activity in deep brain tissues, and the block-design VFT is appropriate for elucidating brain activity over the PFC, other cognitive tasks, and event-related design will be helpful to explore brain pathology in schizophrenia and to compare signal differences between fNIRS and other neuroimaging tools, such as fMRI and PET. Conversely, the exploration of brain activity in a more naturalistic position and during natural activities, such as driving and conversation, is suitable for future investigations using the advantages of fNIRS (99, 100).

Third, as fNIRS allows easy and repetitive measurements, a study including a large number of samples has been conducted (12). To analyze such a large data set, the same measurement procedure regarding task paradigm and probe setting and more sophisticated analytical methods, such as measurement tools (e.g., NIRS-SPM) (101), bioinformatics methods, and machine learning (94), will be needed. Standardized manuscript description and a database system will be needed for further comparisons and meta-analyses (102). The construction of an fNIRS community is expected to provide this type of background in NIRS research (18).

## **CONCLUSION**

Functional near-infrared spectroscopy has been progressing rapidly in the field of psychiatry, as it provides several advantages, such as small size, portability, silent functioning, and the achievement of easy and non-invasive measurements. A part of these results was approved in 2009 as one of the Advanced Medical Technologies as an aid for the differential diagnosis of depressive symptoms (10–12), which was the first such approval in the field of psychiatry in Japan. Future investigations aimed at exploring fNIRS differences in various clinical stages, longitudinal changes, drug effects, and variations during different task paradigms will be needed to develop more accurate biomarkers that can be used to aid differential diagnosis, the comprehension of the present condition, the prediction of outcome, and the decision regarding treatment options in schizophrenia.

Future fNIRS research environments will require standardized measurement procedures, probe settings, analytical methods and tools, manuscript description, and database systems in a n fNIRS community.

## **ACKNOWLEDGMENTS**

I would like to thank Mika Yamagishi and Hanako Sakurada for substantial support in article search. This study was supported by grants from the Ministry of Health, Labor, and Welfare (Health and Labor Science Research Grant for Comprehensive Research on Disability Health and Welfare H22-seishin-ippan-015 to Kiyoto Kasai) and from the JSPS/MEXT (No. 25870143 to Shinsuke Koike, and Grant-in-Aid for Scientific Research on Innovative Areas [Comprehensive Brain Science Network and Adolescent Mind and Self-Regulation (23118001 and 23118004) to Kiyoto Kasai], and National Bioscience Database Center (NBDC) of Japan Science and Technology Agency (JST) to Kiyoto Kasai. A portion of this study was also the result of a project entitled "Development of biomarker candidates for social behavior" carried out under the Strategic Research Program for Brain Sciences by the MEXT.

## **REFERENCES**


near-infrared spectroscopy study. *Neurosci Lett* (2010) **478**:136–40. doi:10. 1016/j.neulet.2010.05.003


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 September 2013; accepted: 25 October 2013; published online: 14 November 2013.*

*Citation: Koike S, Nishimura Y, Takizawa R, Yahata N and Kasai K (2013) Nearinfrared spectroscopy in schizophrenia: a possible biomarker for predicting clinical outcome and treatment response. Front. Psychiatry 4:145. doi: 10.3389/fpsyt.2013.00145 This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Koike, Nishimura, Takizawa, Yahata and Kasai. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

REVIEW ARTICLE published: 04 December 2013 doi: 10.3389/fpsyt.2013.00157

# Pharmacological interventions for the MATRICS cognitive domains in schizophrenia: what's the evidence?

#### **Wilhelmina A. M. Vingerhoets 1,2\*, Oswald J. N. Bloemen<sup>1</sup> , Geor Bakker 1,2 and Therese A. M. J. van Amelsvoort <sup>1</sup>**

<sup>1</sup> Department of Psychiatry and Psychology, Maastricht University, Maastricht, Netherlands

<sup>2</sup> Department of Nuclear Medicine, Academic Medical Centre, University of Amsterdam, Amsterdam, Netherlands

#### **Edited by:**

Stefan Borgwardt, University of Basel, Switzerland

#### **Reviewed by:**

Alkomiet Hasan, Ludwig-Maximilians-University Munich, Germany Rolf-Dieter Stieglitz, University of Basel, Switzerland

#### **\*Correspondence:**

Wilhelmina A. M. Vingerhoets, Department of Psychiatry and Psychology, Maastricht University, Vijverdalseweg 1, Maastricht 6226 NB, Netherlands e-mail: claudia.vingerhoets@ maastrichtuniversity.nl

Schizophrenia is a disabling, chronic psychiatric disorder with a prevalence rate of 0.5– 1% in the general population. Symptoms include positive (e.g., delusions, hallucinations), negative (e.g., blunted affect, social withdrawal), as well as cognitive symptoms (e.g., memory and attention problems). Although 75–85% of patients with schizophrenia report cognitive impairments, the underlying neuropharmacological mechanisms are not well understood and currently no effective treatment is available for these impairments. This has led to the Measurement and Treatment Research to Improve Cognition in Schizophrenia (MATRICS) initiative, which established seven cognitive domains that are fundamentally impaired in schizophrenia.These domains include verbal learning and memory, visual learning and memory, working memory, attention and vigilance, processing speed, reasoning and problem solving, and social cognition. Recently, a growing number of studies have been conducted trying to identify the underlying neuropharmacological mechanisms of cognitive impairments in schizophrenia patients. Specific cognitive impairments seem to arise from different underlying neuropharmacological mechanisms. However, most review articles describe cognition in general and an overview of the mechanisms involved in these seven separate cognitive domains is currently lacking. Therefore, we reviewed the underlying neuropharmacological mechanisms focusing on the domains as established by the MATRICS initiative which are considered most crucial in schizophrenia.

**Keywords: pharmacology, schizophrenia, cognition, MATRICS, neurotransmitters**

## **INTRODUCTION**

Schizophrenia is a disabling, chronic psychiatric disorder with a prevalence rate of 0.5–1% in the general population. Symptoms include positive and negative symptoms and disorganization. Approximately 75–85% of the schizophrenia patients report cognitive impairments as well (1). Although this aspect of schizophrenia was already described by Kraepelin nearly a century ago, cognitive impairments have long been under-identified as a symptom of schizophrenia and relatively little research has been conducted on this topic. Previous research showed that cognitive functioning is strongly associated with functional outcome in schizophrenia [e.g., skill acquisition in psychosocial rehabilitation treatment, demonstration of ability to solve simulated interpersonal problems, and community functioning (2, 3)]. Cognitive impairments often precede the onset of other symptoms and persist after other psychotic symptoms have been effectively treated (4). Furthermore, severity of cognitive impairments is predictive of poorer medication compliance (5), treatment adherence (6), and increased tendency for relapse in first-episode patients (7). At present, no effective treatment is available as existing (antipsychotic) medication mainly targets positive symptoms and does not improve cognition. Although it is sometimes assumed by clinicians (possibly due to marketing) that atypical antipsychotics are superior to typical antipsychotics, results from recent studies contradict this theory. Presently available guidelines for treatment

of schizophrenia, such as the NICE (8), PORT (9), and WFSBP (10) guidelines, indicate that there is no superiority of atypical antipsychotics on positive symptoms with the exception of clozapine, and instead state that use of antipsychotics should be evaluated based on their side effects. Since cognitive dysfunction is associated with functional outcome, development of an effective intervention strategy for these symptoms and corresponding guidelines is essential, as such guidelines are currently still lacking.

Lack of effective treatment strategies has over recent years led to an increase in studies investigating underlying neurobiological mechanisms of cognitive impairments and potential new pharmacological targets to enhance cognition in schizophrenia. Research has mainly focused on the role of neurotransmitters such as dopamine, serotonin, γ-aminobutyric acid (GABA), glutamate, and acetylcholine. Previous research indicated that specific cognitive impairments seem to arise from different underlying neurobiological mechanisms (11). For example, the prefrontal cortex (PFC) has been implicated in the executive functioning aspect of cognition (12) whereas the hippocampus has been linked to, e.g., episodic memory (13). This suggests that specific pharmacological agents could enhance domains of cognition differentially. Nonetheless, still little is known about the underlying neurobiology of cognition. Knowledge about these neurobiological mechanisms is highly necessary for development of new pharmacological intervention strategies.


#### **Table 1 | Pharmacological agents used and their main mechanism of action.**

(Continued)


To improve cognition research in schizophrenia the Measurement and Treatment to Improve Cognition in Schizophrenia (MATRICS) was developed. The MATRICS identified seven cognitive domains that are fundamentally impaired in schizophrenia: verbal learning and memory, visual learning and memory, working memory, attention and vigilance, processing speed, reasoning and problem solving, and social cognition (14). It was decided that cognition research in schizophrenia should mainly focus on these domains in order to identify the neurobiological mechanisms, ultimately to facilitate development of new pharmacological treatment strategies.

Although specific domains of cognition have been identified, most studies tend to describe cognition in general terms using a composite score. Currently, a review differentiating between separate MATRICS domains is lacking. Therefore, the aim of this review is to provide an outline of the underlying neuropharmacological mechanisms of each individual cognitive domain. We will focus on pharmacological intervention studies which used validated instruments to measure the effect on the MATRICS cognitive domains.

## **METHODS**

### **SEARCH STRATEGY**

A literature search was conducted in medical database PubMed. The following keywords were used: "pharmacology," "schizophrenia," and "Cognition." Subsequently, a separate search was conducted for each individual domain combining the keywords"pharmacology" and "schizophrenia" with the following keywords:"verbal learning,""verbal memory,""visual learning,""visual memory," "working memory," "attention," "vigilance," "processing speed," "reasoning,""problem solving," and "social cognition."

#### **INCLUSION AND EXCLUSION CRITERIA**

Papers that met the following inclusion criteria were included: (1) original research papers, both single challenge and clinical trials (full text available); (2) published in English; (3) use of a pharmacological intervention (pharmacological interventions used and their main mechanism of action are displayed in **Table 1**); (4) use of validated cognitive tests to measure one or more of the MATRICS domains; (5) subjects were patients with schizophrenia; and (6) were published between January 2000 and May 2013. Papers were excluded if: (1) only healthy subjects were included; (2) also schizoaffective disorders and other psychotic disorders were included; (3) cognitive domains were measured with nonvalidated tests; (4) cognitive domains other than the MATRICS domains were measured; and (5) only a composite score of cognition was reported.

## **RESULTS**

In total, the search strategy yielded 938 articles of which 293 articles were found using the keywords "pharmacology," "cognition," and "schizophrenia." The separate searches yielded 158 articles for verbal learning and memory, 85 for visual learning and memory, 100 for working memory, 234 for attention and vigilance, 22 for processing speed, 17 for reasoning and problem solving, and 29 for social cognition. After final screening 44 articles were included for verbal learning and memory, 26 for visual learning and memory, 43 for working memory, 22 for attention and vigilance, 31 for processing speed, 30 for reasoning and problem solving, and 7 for social cognition. Study characteristics and specifics are shown in**Table 2** (studies investigating antipsychotics) and **Table 3** (studies investigating non-antipsychotic intervention strategies).

## **VERBAL LEARNING AND MEMORY**

It has been proposed that impairments in verbal learning and memory are one of the most consistent cognitive deficits seen in schizophrenia (15) and is one of the most examined cognitive domains in these patients. A majority of the 44 included studies investigated the effects of typical and atypical antipsychotic medication on verbal learning and memory.

In the past, it was assumed that atypical antipsychotics are superior to typical antipsychotics in enhancing cognition in schizophrenia due to their affinity for both dopamine D<sup>2</sup> receptors and serotonin 5-HT2A receptors (16, 17). However large studies such as the Clinical Antipsychotic Trials of Intervention Effectiveness (CATIE) (18) and the European First Episode Schizophrenia Trial (EUFEST) (19) did not find differences between typical and atypical antipsychotics in cognitive enhancing effects. This theory was tested by Wagner et al. (16) who compared olanzapine, a dopamine D2, and serotonin 5-HT2A antagonist, to amisulpride, a dopamine D2/D<sup>3</sup> antagonist in 52 patients. No significant differences were found after 8 weeks between the two groups in terms verbal memory; both groups improved on verbal learning and memory tasks. Olanzapine was compared to risperidone and haloperidol by Purdon et al. (20). Although not significant after corrections for multiple comparisons, olanzapine was found to be superior to risperidone and haloperidol in enhancing verbal learning and memory.

The atypical antipsychotic ziprasidone was found to improve verbal learning and memory after 12 weeks of monotherapy (21). However, the study did not include a control group and practice effects were not adequately controlled for. Therefore the results have to be interpreted with caution. A study comparing the effects of risperidone and haloperidol on verbal memory found that performance on a verbal memory task remained essentially unchanged in both groups (22). Since the authors did not include a non-medicated schizophrenia group but included a healthy group as control this mainly provides evidence that there is no difference between risperidone and haloperidol. Nevertheless it suggests that both medications did not cause a clinically relevant enhancement of verbal memory. However,Harvey et al. (23) compared the effects of risperidone and quetiapine and found improvement in verbal memory in both treatment groups. The dropout rate of 57% in this trial was high which may have caused a selection bias. These results are not consistent with other findings. A small study with substantial methodological shortcomings by Purdon et al. (24) found no effect of haloperidol and quetiapine. Velligan et al. (25) also compared quetiapine to haloperidol and found that patients in the quetiapine group improved to a greater extent than the patients in the haloperidol group. However, these findings were not significant after correction for multiple comparisons. Moreover, Kivircik Akdede et al. (26) did not find improvement on a verbal learning and memory task after 8 weeks of quetiapine treatment.

Kim et al. (27) found that switching from oral atypical antipsychotics to depot risperidone led to significant improvement in verbal learning and memory. However, they also did not include a control group and did not correct for possible practice effects. Furthermore a high percentage of the patients dropped out in an early stage of the study which could have caused a bias in the sample. Despite these limitations, their results are supported by Suzuki and Gen (28) who also found an improvement in verbal learning and memory in patients treated with depot risperidone compared to patients who received haloperidol treatment (depot). However, the differences between these two treatments might be caused by the high rate of anticholinergic co-medication in the haloperidol group which is well known for its adverse effects on cognition (29). Additionally, improvement in verbal learning and memory was found by Surguladze et al. (30) with both depot risperidone and typical antipsychotics (depot). Performance did not significantly differ between the groups. However, they did not control for practice effects. Therefore improvement cannot reliably be attributed to medication effects. In addition, Kim et al. (31) found that compared to patients using risperidone, verbal memory improved more in patients using paliperidone extend release (ER), an active metabolite of risperidone with an extended delivery system that decreases fluctuations in serum drug concentrations.

Hence, the theory that atypical antipsychotic medications are superior to typical antipsychotics in enhancing verbal learning and memory is not supported by these findings. Nonetheless, some of the atypical antipsychotics, especially risperidone depot, have shown beneficial effects on this aspect of cognition.

Results regarding the effects of clozapine on verbal learning and memory are inconclusive. Three studies which investigated the effects of clozapine on verbal learning and memory were found. Purdon et al. (32) reported significant improvement in verbal learning and memory after treatment with clozapine. Their results were not replicated by Ertugrul et al. (33) and Sumiyoshi et al. (34) as both of these studies did not find improvement in verbal memory in clozapine-treated patients. This discrepancy in findings could be due to methodological differences such as different test batteries and trial duration.


(Continued)

**Table 2 |**

**Summary**

 **of included studies using** 

**antipsychotics**.


**www.frontiersin.org** December 2013 | Volume 4 | Article 157 |

**Table 2 |**

**Continued**

**48**

(Continued)


Moreover, the study by Purdon et al. (32) did not correct for possible practice effects and since they did not include a control group, improvement in verbal memory cannot be reliably attributed to the effects of clozapine. Thus, although results are inconclusive, at present no convincing evidence is available for the effectiveness of clozapine in enhancing verbal learning and memory in schizophrenia.

Geffen et al. (35) compared the effects of 10 and 20–30 mg of BL-1020, a new antipsychotic drug, to risperidone and placebo. BL-1020 is an antagonist for D<sup>2</sup> en 5-HT2a receptors and agonist for GABA<sup>A</sup> receptors (35). Post-mortem studies have found altered GABAergic transmission in schizophrenia, predominantly in the PFC (36, 37). Treatment with 20–30 mg BL-1020 was found to improve performance on a verbal learning memory task compared to placebo, whereas 10 mg of BL-1020 and risperidone did not have enhancing effects compared to placebo. A substantial number of patients, especially in the risperidone group, received anticholinergics as well, which could have influenced the results.

Although the findings described above do not support the theory that blocking the 5-HT2A receptors improves verbal learning and memory, a role for serotonin in this domain of cognition cannot be ruled out. Sumiyoshi et al. (38) found a significant improvement on verbal memory after adjunctive tandospirone (a 5-HT1a agonist) treatment compared to patients who did not receive adjunctive tandospirone. Riedel et al. (39) and Bervoets et al. (40) both found a significant improvement in verbal memory in patients treated with aripiprazole,an atypical antipsychotic drug with partial D<sup>2</sup> agonistic and antagonistic and 5-HT2a antagonistic properties in addition to partial agonistic activity at the 5-HT1a receptors. However, this is contradicted by Suzuki et al. (41) who compared the effects of aripiprazole, olanzapine, and perospirone on verbal memory. None of the three conditions showed significant changes in verbal memory scores and no differences were found between the three treatment conditions. Moreover, aripiprazole added to atypical antipsychotics did not improve verbal learning and memory in a trial by Yasui-Furukori et al. (42). This discrepancy in findings could be due to practice effects in the studies of Riedel et al. (39) and Bervoets et al. (40), since both studies did not include a control group and did not correct for possible practice effects. Furthermore, different tasks were used which makes it difficult to directly compare the results.

The role of the serotonin 5-HT3a receptor in verbal learning and memory has been examined as well. Akhondzadeh et al. (43) and Levkovitz et al. (44) investigated the effects of ondansetron, a serotonin 5-HT3a receptor antagonist, added to respectively risperidone and clozapine treatment, and found no enhancing effects on verbal learning and memory. However, the sample size of both studies was small. In contrast to these results, Zhang et al. (45) did find improvement in verbal memory after add-on tropisetron (a 5-HT3a receptor antagonist and nicotinic α<sup>7</sup> agonist) to risperidone treatment compared to placebo. This discrepancy in findings could be explained by the fact that the two studies used different pharmacological interventions. Both Akhondzadeh et al. (43) and Levkovitz et al. (44) administered ondansetron whereas Zhang et al. (45) used tropisetron. Since ondansetron has been found to have low affinity for the nicotinic α<sup>7</sup> receptors (46), the positive effects of tropisetron on verbal learning and memory

**Table 2 |**

**Continued**


**Table 3 |**

**Summary**

 **of included studies using** 

**non-antipsychotic**

**interventions**.



**Table 3 |**

**Continued**

could be due to its nicotinic α<sup>7</sup> agonistic properties rather than its antagonistic effects on the 5-HT3a receptors.

Effects of a selective serotonin reuptake inhibitor (SSRI) were examined by Friedman et al. (47) who added citalopram to treatment with atypical antipsychotics but found no improvement in verbal learning and memory.

Thus, although results are inconclusive there is preliminary evidence of positive effects of 5-HT1a agonists, but not 5-HT3a antagonists on verbal learning and memory in schizophrenia.

One study examined the effects of a dopamine D<sup>1</sup> *agonist* on verbal learning and memory in schizophrenia. George et al. (48) administered dihydrexidine as a pharmacological challenge and found no effects of dihydrexidine on verbal learning and memory.

The role of the neurotransmitter acetylcholine in cognition has been widely investigated and established (36). Post-mortem, a correlation was found between cognitive impairment and decreased levels of brain choline acetyltransferase in schizophrenia (49). Post-mortem studies have shown changes in both muscarinic and nicotinic acetylcholine receptors in patients with schizophrenia (50). Velligan et al. (51) conducted a trial to assess the effects adjunctive therapy with the selective nicotinic α4β<sup>2</sup> receptor agonist AZD3480 on verbal learning and memory in patients but found no improvement. An important limitation of this study is that only patients who smoked were included. Therefore, the lack of effect could be due to desensitization of the nicotinic receptors caused by chronic tobacco use. However, Freedman et al. (52) also did not find an improvement in verbal learning and memory after administration of the partial nicotinic α<sup>7</sup> agonist DMXB-A, added to antipsychotic treatment in patients, who abstained from nicotine at least 1 month prior to participation. Studies using acetylcholinesterase (enzyme that breaks down acetylcholine, thereby increasing acetylcholine levels) inhibitors (AChE-Is) as adjunctive therapy to antipsychotic medication also found no improvement on verbal learning and memory (53–58). These results suggest that acetylcholinesterase inhibitors do not effectively enhance verbal memory in patients with schizophrenia. However, all studies used a small sample size and one study included patients who were taking anticholinergic medication as well. This could have influenced the outcome of the study. Another important limitation is that none of these studies controlled for the effects of smoking. Hence, although there is theoretical evidence for a role of acetylcholine in cognition, various (possibly underpowered) intervention studies have not yielded positive results. Since Zhang et al. (45) found improvement in verbal learning and memory after adjunctive tropisetron treatment, it remains possible that enhancement of verbal learning and memory can be achieved with nicotinic α<sup>7</sup> agonists while simultaneously blocking the 5-HT3a receptors.

The role of the neurotransmitters GABA and norepinephrine in verbal learning in memory has also been investigated. Buchanan et al. (59) found no improvement in verbal learning and memory with adjunctive MK-0777, a partial GABA α2/α<sup>3</sup> agonist, therapy to antipsychotics. However, the authors argue that a lack of effect may be due to the fact that MK-0777 is a weak GABAα2/α<sup>3</sup> agonist. A pilot study by Friedman et al. (60) found no effects of the norepinephrine reuptake inhibitor atomoxetine, on verbal memory when added to treatment with atypical antipsychotics.

Hence, although positive effects were found with the antipsychotic BL-1020 with GABA<sup>A</sup> agonistic properties, the available studies were not able to detect positive effects of both a partial GABA α2/α<sup>3</sup> agonist and a norepinephrine reuptake inhibitor on verbal learning and memory. Nonetheless, these results suggest that GABA<sup>A</sup> receptors may be a potential target for future studies in verbal learning and memory.

Effects of the psychostimulant armodafinil were investigated by Kane et al. (61). Armodafinil is a longer-lasting isomer of modafinil, which is an alertness-promoting medication with mechanisms of action different from those of amphetamine (62) and has been found to improve cognition in healthy subjects and adults with ADHD (63). However, the exact mechanisms of action are complex and not entirely understood (64). Armodafinil, added to atypical antipsychotics, did not enhance verbal learning and memory.

The effects of several other pharmacological interventions on verbal learning and memory in schizophrenia have been investigated. Six months of daily creatine administration added to antipsychotic treatment did not have beneficial effects on verbal learning and memory (65). Effects of adjunctive mifepristone, a glucocorticoid receptor (GR) antagonist, on verbal learning and memory was examined by Gallagher et al. (66). They found no improvement in this domain of cognition. Goff et al. (67) examined the effects of sildenafil, a phosphodiesterase 5 (PDE5) inhibitor used to treat erectile dysfunction, on verbal learning and memory. PDE5 inhibitors increase cyclic guanine monophosphate (cGMP) which is thought to modulate long-term potentiation. They found single dose of 50 and 100 mg not to improve verbal learning and memory. On the other hand, Feifel et al. (15) found that patients who received intranasal oxytocin for 3 weeks performed better on a verbal learning and memory task. Additionally, Stone et al. (68) found that additional glucose administration in patients who were stable on clozapine, enhanced verbal learning and memory. All studies used a small sample and the dropout rate in the study by Feifel et al. (15) was 25%. Moreover, the majority of the patients in the study of Stone et al. (68) received additional medication besides clozapine. Therefore the effect cannot be attributed to a unique interaction of clozapine and glucose.

To summarize, the proposed superior effect on verbal learning and memory of atypical to typical antipsychotics, is not supported by the available data and any effect does not result from blocking the 5-HT2a receptors. Despite the fact that multiple studies describe a role of acetylcholine in cognition, nicotinic receptor agonist and acetylcholinesterase inhibitors do not seem to effectively improve verbal learning and memory in schizophrenia. However, some positive results have been found with risperidone depot, possibly because of medication compliance. In addition, serotonin 5-HT1a receptors and GABA<sup>A</sup> receptors may be molecular targets for enhancing this aspect of cognition. Additionally, there is preliminary evidence that adjunctive oxytocin and glucose treatment may be beneficial in enhancing verbal learning and memory. However, results have to be interpreted with care as the available studies are mostly pilot studies which, although important in early stages of research, inherently suffer from methodological shortcomings.

## **VISUAL LEARNING AND MEMORY**

Compared to verbal learning and memory, considerably less research has been conducted regarding visual learning and memory in schizophrenia. Results of studies that examined the effects of antipsychotic medication on visual learning and memory are, similar to verbal learning and memory, indecisive. Tyson et al. (17) compared the effects of atypical antipsychotic medication with high and low affinity for the 5-HT2a receptors on visual learning and memory. Patients treated with risperidone, olanzapine, and clozapine were assigned to the high-affinity group and patients using quetiapine and amisulpride were assigned to the low affinity group. Performance on visual memory tasks improved in patients in the low affinity group whereas performance in the high-affinity group declined. A limitation is that the authors did not correct for multiple testing. However, the effects were already present after 9 months of treatment and show a consistent pattern. Since performance declined in the high-affinity group, it is not likely that the improvement in the low affinity group is driven by practice effects. However, Rollnik et al. (69) found atypical antipsychotics (olanzapine, risperidone, or clozapine) to be superior to typical antipsychotics (haloperidol, chlorprothixene, perazine, or flupenthixol) in improving visual learning and after 3 weeks of treatment, although at the final assessment, this difference was no longer significant. Purdon et al. (20) compared the effects of olanzapine, risperidone, and haloperidol. After correction for multiple comparisons, no significant differences were found between the three treatment conditions; visual learning and memory did not improve in all three groups. As mentioned earlier, lack of effect could be caused by concomitant anticholinergic treatment and the small sample size.

Two studies compared the effects of haloperidol to respectively quetiapine (24) and risperidone (70). In both studies haloperidol and both quetiapine and risperidone were found not to enhance visual learning and memory. However, a substantial number of patients in both studies were treated with anticholinergics in addition to antipsychotics, which could have influenced the results. Surguladze et al. (30) found no differences in visual learning and memory between depot risperidone and a typical antipsychotic depot treatment, although performance in both groups improved. As discussed earlier, it cannot be ruled out that this improvement is caused by practice effects rather than the effects of medication.

Although results are inconclusive, clozapine does not seem to effectively enhance visual learning and memory. A study by Ertugrul et al. (33) did not find improvement with clozapine whereas Purdon et al. (32) found significant improvement on performance on one of three assessed visual memory subtests. However, this study has important limitations, which were described earlier.

Treatment with aripiprazole did not significantly improve visual learning and memory in the study by Riedel et al. (39). Tandospirone, which like aripiprazole has serotonin 5-HT1a agonistic properties, also did not significantly enhance visual learning and memory in the study of Sumiyoshi et al. (38). These findings suggest that stimulating the serotonin 5-HT1a receptors does not enhance visual learning and memory in schizophrenia. However, the serotonin 5-HT3a receptors may play a role in visual memory

in schizophrenia. Both Akhondzadeh et al. (43) and Levkovitz et al. (44) found visual learning and memory to improve with adjunctive ondansetron treatment. Although both studies did not correct for multiple testing and possible practice effects, visual memory was the only aspect of the cognitive domains measured that improved in both studies. Adjunctive tropisetron treatment on the other hand did not enhance visual memory (45). As mentioned earlier, tropisetron is both a 5-HT3a antagonist and partial nicotinic α<sup>7</sup> agonist whereas ondansetron has low affinity for the nicotinic α<sup>7</sup> receptors (46). Hence, the 5-HT3a receptors may play a role in visual learning and memory. However, enhancement of this aspect of cognition may depend on blockade of nicotinic α<sup>7</sup> receptors. The 5-HT1a receptor does not seem to be a promising target for enhancing visual learning and memory.

A preliminary study by Niitsu et al. (71) investigated the effect of the SSRI fluvoxamine, a sigma-1 receptor agonist, on cognition and found it not to enhance visual learning and memory. These preliminary results indicate sigma-1 receptor agonism does not enhance visual learning and memory in schizophrenia patients.

Studies investigating the role acetylcholine in visual learning and memory in schizophrenia have not yielded positive results. Selective nicotinic α4β<sup>2</sup> receptor agonist AZD3480 and partial nicotinic α<sup>7</sup> agonist DMXB-A both did not enhance visual learning and memory when added to antipsychotic treatment (51, 52). As mentioned earlier, the lack of effect in the study of Velligan et al. (51) could be due to desensitization of the nicotinic receptors caused by chronic tobacco use since only smokers were included. Studies conducted with acetylcholinesterase inhibitors do not report positive results either. Although Lee et al. (54) found improvement on the recognition subtest of Rey Complex Figure Test (RCTF) with galantamine, the other aspects of the task did not improve. Studies using donepezil also did not find improvement in visual learning and memory (50, 53, 55). Hence, both nicotinic receptor agonists and acetylcholinesterase inhibitors do not seem to effectively enhance visual learning and memory in schizophrenia.

Based on preliminary results of Buchanan et al. (59), GABA does not seem to enhance visual learning and memory in schizophrenia as they found no improvement with partial GABA<sup>A</sup> α2/α<sup>3</sup> agonist MK-0777. However, this is the only study investigating the role of GABA in visual learning and memory and MK-0777 is a weak GABA α2/α<sup>3</sup> agonist.

The possible enhancing effect of psychostimulant drugs on visual learning and memory has been investigated in schizophrenia as well. A single dose of d-amphetamine, an indirect dopamine D<sup>1</sup> agonist, did not improve visual learning and memory (72). A single administration of modafinil did not significantly enhance visual memory either (63). In addition, adjunctive armodafinil treatment did not improve visual learning and memory (61). Hence, based on these findings d-amphetamine and modafinil do not seem to be promising intervention strategies for enhancing visual learning and memory.

Results of other pharmacological interventions on visual learning and memory are mixed. Gallagher et al. (66) found no improvement after adjunctive mifepristone (a GR antagonist) treatment. The effect of adjunctive creatine treatment was investigated by Kaptsan et al. (65) but they found it not to enhance visual learning and memory. Levkovitz et al. (73) investigated the effects of minocycline, a second generation tetracycline with antiinflammatory and antimicrobial effects. Minocycline modulates the glutamate pathway by blocking nitric oxide induced neurotoxicity. Hyperactivity of glutamatergic neurotransmission [possibly caused by hypofunction of *N*-methyl-d-aspartate (NMDA) receptors] has been found in schizophrenia (36). Stimulated by glutamate, the NMDA receptors activate production of nitric oxide. The authors found improvement on a spatial recognition memory task after minocycline administration, added to antipsychotic treatment. Task performance showed a decrease after 10 weeks, but compared to baseline, performance improved at the final assessment. The drop-out rate in this study was high (due to adverse events and non-adherence) and although it did not differ between the two treatment groups, this could have biased the results.

To summarize, results regarding visual learning and memory are inconclusive. At present, there is some evidence for a role of the serotonin 5-HT3a receptors in this cognitive domain. 5-HT1a and sigma-1 receptor agonist on the other hand, did not yield promising results. Additionally, minocycline was found to have positive effects on visual learning memory when added to antipsychotic treatment. Both typical and atypical antipsychotics do not seem to enhance visual learning and memory. Studies investigating nicotinic receptor agonist, acetylcholinesterase inhibitors, GABA α2/α<sup>3</sup> agonist, psychostimulants, and GR antagonist did not yield positive results but are in need of replication as they were possibly underpowered. Overall, results need be interpreted with care as the described studies are often small pilot studies with inherent limitations.

## **WORKING MEMORY**

Working memory refers to a system with limited capacity for temporary storage and manipulation of information, necessary for cognitive tasks (74), and has been the subject of many studies investigating cognition in schizophrenia. Most of these studies focused on the effects of antipsychotic medication on working memory.

Although results are inconclusive at present, both typical and atypical antipsychotics show little beneficial effects on working memory. Rollnik et al. (69) compared the effectiveness of typical and atypical antipsychotics and found no differences between the two groups. Quetiapine and amisulpride, which both have low affinity for the 5-HT2a receptors, improved performance on a verbal working memory task compared to risperidone, olanzapine, and clozapine (high affinity for 5-HT2a) treatment (17). However, performance on a visual working memory tasks did not significantly improve in both groups. No differences were found in working memory performance between amisulpride and olanzapine by Wagner et al. (16). However, working memory performance did improve in both groups. Effects of olanzapine, risperidone, and haloperidol were compared by Purdon et al. (20). After correction for multiple comparisons, no significant differences were found between the three treatment conditions and working memory did not improve in any group. Mori et al. (75) examined the effects of switching from typical antipsychotics to olanzapine, quetiapine, risperidone, or perospirone on working memory and found improvement in working memory with risperidone, while

working memory performance of the patients treated with quetiapine decreased. Working memory performance did not change in the olanzapine and perospirone group. It must be noted that the mean age in the sample was high (59.9 years) and correlated negatively with working memory scores. Therefore age related cognitive decline may have confounded these results. On the contrary, Kivircik Akdede et al. (26) found improvement on working memory with quetiapine. However, this study did not include a control group and did not correct for possible practice effects. Risperidone treatment was found to be superior to haloperidol in the study by McGurk et al. (70), but after correcting for anticholinergic co-medication, this effect was no longer significant. Working memory performance did not improve in both groups. Paliperidone treatment was not superior to risperidone treatment in enhancing working memory in the study by Kim et al. (31) and working memory did not improve in both groups.

Results regarding clozapine are inconclusive. Papageorgiou et al. (76) compared the effects of clozapine and olanzapine on working memory and found that working memory improved in both groups. No differences were found between the two treatment conditions. Ertugrul et al. (33) conducted two working memory tasks (Digit Span Test and the Auditory Consonant Trigram Test) and found that performance on the Digit Span test improved with clozapine treatment, whereas performance on the Auditory Consonant Trigram Test did not. Both studies did not correct for possible practice effects or multiple testing. These results are contradicted by Purdon et al. (32) who did not find improvement in working memory with clozapine treatment. This discrepancy in findings could be due to methodological shortcomings and differences such as different test batteries and trial duration.

Riedel et al. found that aripiprazole did not significantly enhance working memory (39). Aripiprazole also did not improve working memory when added to atypical antipsychotics (42). Adjunctive sertindole, an atypical antipsychotic with high affinity for the D<sup>2</sup> and 5-HT2a receptors, did not improve working memory in patients treated with clozapine (77). Finally, working memory did improve after 20–30 mg of the novel antipsychotic BL-1020 (a dopamine antagonist with GABA agonistic properties) compared to placebo whereas 10 mg BL-1020 and risperidone did not (35).

Hence, although results are preliminary and inconclusive, at present no convincing evidence is available for the effectiveness of both typical and atypical antipsychotics on working memory in schizophrenia. However, the antipsychotic BL-1020 was found to improve working memory, possibly through its GABA<sup>A</sup> agonistic properties.

Single administration of the dopamine D<sup>1</sup> agonist dihydrexidine did not enhance working memory in a trial by George et al. (48). Given the small sample size and the fact that dihydrexidine was only administered a single time, beneficial effects of long term dihydrexidine administration on working memory cannot be ruled out.

Serotonin 5-HT3a receptors seem to be less important for working memory functioning. Both Akhondzadeh et al. (43) and Levkovitz et al. (44) found no improvement in working memory with adjunctive ondansetron (5-HT3a antagonist) treatment. Moreover, adjunctive tropisetron treatment also did not enhance working memory (45).

In addition, Niitsu et al. (71) found no improvement in working memory after adjunctive fluvoxamine treatment. Moreover, 12 weeks of adjunctive citalopram treatment did not enhance working memory (47). Hence, these results suggest that serotonin modulation pharmacological agents do not significantly enhance working memory in schizophrenia.

Both the selective nicotinic α4β<sup>2</sup> receptor agonist AZD3480 and the partial nicotinic α<sup>7</sup> agonist DMXB-A did not enhance working memory performance (51, 52). Acetylcholinesterase inhibitors donepezil and rivastigmine did not enhance working memory either (50, 53–55, 57, 58, 77). Jacobsen et al. (78) examined the effects of a single administration of nicotine [nicotinic acetylcholine receptor agonist (79)] on working memory in tobacco using patients and tobacco using healthy subjects. Although patients performed significantly worse on the N-back task compared to healthy controls, nicotine administration did improve performance on the two-back in patients while performance of the healthy controls declined. Thus, although the acetylcholinesterase inhibitors donepezil and rivastigmine and nicotinic receptor agonists do not show positive effects on working memory in schizophrenia, positive results were found with a single dose of nicotine administration. Therefore, a role of acetylcholine and nicotinic receptors in enhancing working memory cannot be ruled out.

Working memory did not improve in a study by Buchanan et al. (59) after 4 weeks of treatment with partial GABA α2/α<sup>3</sup> agonist MK-0777. Menzies et al. (80) studied the effects of lorazepam, a GABA<sup>A</sup> receptor agonist, and flumazenil, a GABA<sup>A</sup> receptor antagonist, on working memory in patients and healthy subjects. Working memory performance declined after a single dose of lorazepam whereas performance did not change significantly after flumazenil. Thus, although results are inconclusive, there is some evidence that in GABA<sup>A</sup> agonists may lower working memory. However positive effects on working memory were found with an antipsychotic with additional GABA<sup>A</sup> agonistic properties.

Friedman et al. (60) found no effects of atomoxetine (norepinephrine reuptake inhibitor) on working memory in patients. A lack of effect could be due to insufficient power.

Glutamatergic pathways have been examined by Duncan et al. (81). They added 50 mg of d-cycloserine, an antituberculous drug which, at low doses has agonistic properties at the NDMA receptors, to typical antipsychotics. After 4 weeks of treatment, working memory did not improve. However, these results need to be replicated before the role of NMDA receptors in working memory can be established.

Armodafinil, a longer-lasting isomer of modafinil, did not enhance performance on working memory task when added to atypical antipsychotics for 4 weeks (61). However, Turner et al. (63) found that a single dose of modafinil improved performance on a verbal working memory task compared to placebo. Performance on a spatial working memory task did not differ between modafinil condition and placebo condition. These results were not replicated by Spence et al. (82) who did not find differences in performance on a (verbal) working memory task after a single administration of modafinil compared to placebo. This discrepancy in findings could be due to differences in modafinil dose (200 and 100 mg, respectively). Moreover, both studies used different tasks which make direct comparison of results difficult. Nonetheless, a single dose of modafinil might enhance working memory in schizophrenia patients, although recurrent treatment with a longer-lasting variant showed no effect.

Add-on intranasal oxytocin treatment did not improve working memory (15). Furthermore, adjunctive mifepristone (GR antagonist) treatment did not enhance working memory (66). Goff et al. (67) did not find improvement in working memory after both 50 and 100 mg sildenafil administration. Furthermore, daily creatine administration added to antipsychotic treatment did not have beneficial effects on working memory either (65). Add-on minocycline treatment on the other hand, improved working memory in a study by Levkovitz et al. (73); the number of errors on a working memory task decreased significantly in the minocycline group whereas the number of errors in the placebo group increased. However, this study has some limitations which are described earlier. Thus, apart from minocycline, other lines of research did not identify possible new intervention strategies for working memory enhancement in schizophrenia.

To summarize, even though some positive results were found with minocycline, nicotine, and modafinil, the presently available studies did not identify promising molecular targets for enhancement of working memory in schizophrenia. GABA agonists have shown mixed results. Most of the studies examined the effects of antipsychotic treatment on working memory and only a limited number of studies used other pharmacological interventions, which makes it difficult to draw definite conclusions. More research needs to be conducted, especially on the potential role of GABA, norepinephrine, acetylcholine, glutamate, and psychostimulants.

## **ATTENTION AND VIGILANCE**

The construct of attention refers to a core cognitive function that relates to the ability to select, filter, focus, and process different stimuli in the environment. Attention is closely related to working memory and executive functioning (83) and therefore, it is difficult to distinguish between these constructs. Because of the broad definition, impairment in virtually any task can be considered as impairment in attention (83). Therefore, in this review, we only included articles which used the Continuous Performance Test (CPT) to specifically measure attention as this test is recommended by the MATRICS (84).

Studies using antipsychotics to enhance attention report little beneficial effects. The theory that atypical antipsychotics are superior to typical antipsychotics in enhancing cognition was tested for attention as well. A trial comparing olanzapine to amisulpride did not find differences in effects on attention; attention did not significantly improve in both groups (16). Olanzapine did not enhance attention in a study by Molina et al. (85), who examined the effects of switching from typical antipsychotics or risperidone to olanzapine. Attention did not improve with olanzapine treatment.

The effects of risperidone on attention were compared to the effects of respectively haloperidol (86) and quetiapine (23). In the study of Liu et al. (86) no improvement was found in both the haloperidol and the risperidone group. Additionally, although no differences were found between the risperidone and quetiapine in the study by Harvey et al. (23), within group analyses showed significant improvement in attention in the risperidone group, whereas the quetiapine group did not improve. Kim et al. (27) reported improvement with depot risperidone. However, paliperidone did not improve attention (31). To summarize, both typical and atypical antipsychotics have shown little beneficial effects on attention in schizophrenia patients.

The possible role of dopamine and serotonin in attention has also been examined with intervention strategies other than antipsychotic treatment. George et al. (48) examined the effects of a full dopamine D<sup>1</sup> agonist, dihydrexidine but found it not to enhance attention.

Studies using serotonin intervention strategies have not yielded positive results either. Golightly et al. (87) examined the role of serotonin in attention by using acute tryptophan, a precursor for serotonin, depletion. Depletion had no effect on performance on the CPT. Niitsu et al. (71) investigated the effect of adjunctive fluvoxamine treatment on attention and found no improvement in attention either. The effects of citalopram, added to treatment with atypical antipsychotics was examined by Friedman et al. (47). They found that adjunctive citalopram treatment did not enhance attention. Thus, although serotonin has been implicated in both verbal and visual memory, it does not seem to play a prominent role in attention.

Based on results of mostly preliminary studies, acetylcholine does not seem to play an important role in attention. Selective nicotinic α4β<sup>2</sup> receptor agonist AZD3480, did not enhance attention in a study by Velligan et al. (51). Nonetheless, Freedman et al. (52) found some beneficial effects of partial nicotinic α<sup>7</sup> agonist DMXB-A on attention. In their cross-over study, patients were treated with 75, 150 mg DMXB-A, and placebo. All treatment arms lasted 4 weeks, followed by a 1-week washout period. Over the course of the trial, attention improved in all treatment groups. The authors suggested that possible practice effects had obscured the potential effect of treatment, and therefore also examined the scores after the first 4 weeks of treatment. They found that attention improved with both doses of DMXB-A compared to the placebo. Two types of acetylcholinesterase inhibitors, rivastigmine, and donepezil, did not enhance attention in schizophrenia (56, 58, 88). On the other hand, a single nicotine administration improved attention in schizophrenia patients but not in healthy controls (89). Thus, despite negative results with acetylcholinesterase inhibitors, some promising results have been obtained with nicotine and a partial nicotinic α<sup>7</sup> agonist. These preliminary results suggest that the nicotinic receptors are a potential target for enhancement of attention in schizophrenia.

Partial GABA α2/α<sup>3</sup> agonist MK-0777 did not enhance attention when added to antipsychotic treatment (59). However, as mentioned earlier, MK-0777 is a weak GABA α2/α<sup>3</sup> agonist and this is the only study investigating the potential role of GABA in attention in schizophrenia. Therefore, these results need to be replicated using specific and more potent GABA modulatory agents.

Other lines of research have not yielded positive results. The psychostimulant drug armodafinil did not enhance attention (61). Furthermore, adjunctive mifepristone (GR antagonist) therapy did not improve attention (66). Duncan et al. (81) found no improvement in attention with adjunctive d-cycloserine treatment. Also, sildenafil (PDE5 inhibitor) did not enhance attention in schizophrenia in a trial by Goff et al. (67). Stone et al. (68) compared the effects of single administration of glucose to placebo in clozapine-treated patients and found attention to be significantly worse in the glucose condition. Thus, several other lines of research did not identify potential new intervention strategies enhancement of attention in schizophrenia.

To summarize, at present only a few pharmacological intervention strategies have been effective in enhancing attention in schizophrenia. Although results are not consistent, positive results have been found with both oral and depot risperidone treatment. Furthermore, a partial nicotinic α<sup>7</sup> agonist, and a single administration of nicotine did improve attention in schizophrenia, suggesting a role for acetylcholine and nicotinic receptors in attention. Overall, results need to be considered preliminary and more research needs to be conducted to replicate these findings as these studies have some important limitations.

## **PROCESSING SPEED**

Processing speed is considered a core cognitive function and refers to the speed at which the brain processes information. It can be measured as the number of correct responses during a task within a given amount of time (90).

As for other domains of cognition, a majority of the studies examined the effects of antipsychotic medication on processing speed. The theory that atypical antipsychotics are superior to typical antipsychotics in improving cognition in patients was tested for processing speed as well. A study comparing amisulpride with olanzapine found that processing speed did not improve in both groups (16). Moreover, performance on a processing speed task did not differ between patients treated with different typical antipsychotics and atypical antipsychotics and performance did not improve in both groups (69). Purdon et al. (20) compared the effects of olanzapine, risperidone, and haloperidol on processing speed. After correction for multiple comparisons, no significant differences were found between the three treatment conditions and processing speed did not improve in any of the groups. Thus, atypical antipsychotics do not appear to be superior to typical antipsychotics. Furthermore, no beneficial effects of the atypical antipsychotic ziprasidone were found on processing speed (21). Anticholinergics were allowed during the trial which might have influenced the results. Risperidone depot treatment did not enhance processing speed in the study by Kim et al. (27). Moreover, paliperidone ER was not found to be superior to risperidone (31). Unfortunately, within group comparisons were not reported. However, although no significant differences were found between two treatment groups, risperidone was found to improve processing speed whereas quetiapine did not in a study comparing these two antipsychotics (23). Contrary to those results, quetiapine was found to enhance processing speed in a trial by Kivircik Akdede et al. (26). As described earlier, this study did not include a control group and did not correct for possible practice effects. These results were not replicated by Purdon et al. (24). Although processing speed improved in both the quetiapine and haloperidol group, these results were not significant after correcting for multiple comparisons. Velligan et al. (25) compared

the effectiveness on processing speed between quetiapine and haloperidol treatment. They found no differences between the two treatment groups. Purdon et al. (32) found that processing speed improved with clozapine treatment. However, this improvement cannot reliably be attributed to the effects of clozapine as the study did not include a control group en they did not correct for possible practice effects. Beneficial effects of aripiprazole on processing speed were found by Riedel et al. (39). Yasui-Furukori et al. (42) found no improvement in processing speed with adjunctive aripiprazole treatment to olanzapine or risperidone. Furthermore, a new antipsychotic with GABA<sup>A</sup> receptor agonistic properties in addition to dopamine D<sup>2</sup> and serotonin 5-HT2a receptor blockade (BL-1020) did not improve processing speed (35).

To summarize, both typical and atypical antipsychotics do not seem to effectively enhance processing speed in schizophrenia patients. Although some studies reported positive effects of clozapine, quetiapine, and risperidone, these results should be interpreted with care as these studies have important limitations.

Studies examining the role of serotonin in processing speed did not yield positive results. Adjunctive ondansetron (5-HT3a receptor antagonist) therapy to antipsychotic medication did not enhance processing speed in a study by Levkovitz et al. (44). Adding citalopram to atypical antipsychotic treatment also did not enhance processing speed in the study by Friedman et al. (47). In addition, tryptophan depletion did not affect processing speed in the study by Golightly et al. (87). Thus, the available studies do not provide evidence for a prominent role of serotonin in processing speed in schizophrenia. However, all studies used small samples and the study of Golightly et al. (87) allowed concomitant anticholinergics which could have influenced the results.

Studies using acetylcholine related intervention strategies did not yield positive results either. Both the selective nicotinic α4β<sup>2</sup> receptor agonist AZD3480 and the partial nicotinic α<sup>7</sup> agonist DMXB-A did not enhance processing speed (51, 52). Moreover, three types of acetylcholinesterase inhibitors, donepezil, galantamine, and rivastigmine, did not improve processing speed (50, 54–58). Thus, although the role of acetylcholine in memory and attention has been well established, it does not seem to be a potential target in enhancing processing speed in schizophrenia. However, these studies have limitations such as small sample sizes.

Effects of GABA and norepinephrine on processing speed were also examined. Processing speed did not improve after adjunctive treatment with partial α2/α<sup>3</sup> agonist MK-0777 (59). As mentioned earlier, BL-1020 (antipsychotic with GABA<sup>A</sup> agonistic properties) enhanced antipsychotic also did not enhance processing speed (35). In the pilot study of Friedman et al. (60), processing speed did not improve after adjunctive treatment with the norepinephrine reuptake inhibitor atomoxetine. Thus, norepinephrine and GABA α2/α<sup>3</sup> receptors do not seem to play an important role in processing speed in schizophrenia as the available studies were not able to detect positive effects. However, lack of effect could be due to the small sample size of both studies.

Processing speed did increase after a single dose of damphetamine in the study of Pietrzak et al. (72). Adjunctive armodafinil treatment had no beneficial effects on processing speed in the study by Kane et al. (61). Hence, since d-amphetamine is an indirect dopamine D<sup>1</sup> agonist, dopamine D<sup>1</sup> receptors could be a possible target for enhancing processing speed in schizophrenia although the study by Pietrzak et al. (72) needs replication.

Lastly, there are some studies of agents that have no direct (known) effect on neurotransmission. Processing speed did not improve after single dose of 50 and 100 mg sildenafil (a PDE5 inhibitor commonly used for erectile dysfunction) in the crossover trial by Goff et al. (67). Furthermore, adjunctive dehydroepiandrosterone (DHEA) therapy, a corticosteroid that serves as a precursor for both androgenic and estrogenic steroids, did not enhance processing speed (91). Thus, other lines of research did not identify promising intervention strategies for enhancement of processing speed in schizophrenia.

To summarize, at present no convincing evidence exist for the effectiveness of both typical and atypical antipsychotics on processing speed in schizophrenia. Although, positive results were found with quetiapine, risperidone, clozapine, and aripiprazole, these studies have important limitations and therefore the effects cannot be reliably attributed to the effects of the medication. One positive result was found with single dose d-amphetamine, suggesting a potential role of dopamine D<sup>1</sup> receptors in processing speed. Studies investigating 5-HT3a receptor antagonist, SSRI, nicotinic receptor agonist, acetylcholinesterase inhibitors, a GABA α2/α<sup>3</sup> agonist, and a norepinephrine reuptake inhibitor did not find clinically relevant improvement in processing speed but are in need for replication as the described studies used small samples and were possibly underpowered.

## **REASONING AND PROBLEM SOLVING**

We found 30 studies that measured reasoning and problem solving after a pharmacological intervention. Reasoning and problem solving is considered an aspect of executive functioning. As for the other cognitive domains, a relatively large number of studies investigated the effects of antipsychotics on reasoning and problem solving.

Tyson et al. (17) compared antipsychotics with low affinity for 5-HT2a receptors to antipsychotics with high affinity for these receptors. Response latency on a reasoning and problem solving task decreased in the low affinity group, whereas latency increased in the high-affinity group. The authors did not correct for multiple comparisons. Since response latency increased in the high-affinity group, it is not likely that the improvement in the low affinity group is (only) driven by practice effects. Purdon et al. (20) compared the effects of olanzapine, risperidone, and haloperidol and found that olanzapine was superior in enhancing reasoning and problem solving. However, after correcting for multiple comparisons, none of the medications improved reasoning and problem solving. On the contrary, both Kim et al. (27) and Suzuki and Gen (28) found improvement in performance on a reasoning and problem solving task. However, Kim et al. (27) did not include a control group and did not correct for possible practice effects. Therefore, improvement cannot be reliably attributed to the effects of medication. Moreover, in the study of Suzuki and Gen (28) a high percentage of the patients in the haloperidol group was using concomitant anticholinergics whereas in the risperidone group anticholinergics were tapered during the first weeks of the trial. Therefore improvement in this group could be due to effects of diminishing anticholinergics.

Purdon et al. (24) investigated the effects of haloperidol and quetiapine and found that reasoning and problem solving did not improve in both groups. Olanzapine, perospirone, and aripiprazole did not improve reasoning and problem solving either (41). Moreover, adding aripiprazole to atypical antipsychotics also did not yield positive results (42). Furthermore, no improvement was found with clozapine by Ertugrul et al. (33) and Purdon et al. (32). Nielsen et al. (77) added sertindole to clozapine but found no improvement in reasoning and problem solving either. Finally, the new antipsychotic with GABA agonistic properties BL-1020 also did not improve reasoning and problem solving (35). Hence, although positive results were found with risperidone depot, both typical and atypical antipsychotics have shown little clinically relevant effects on reasoning and problem solving.

Serotonergic intervention strategies have not yielded positive results. Adjunctive ondansetron (5-HT<sup>3</sup> antagonist) treatment had no beneficial effects on reasoning and problem solving (43). Fluvoxamine and citalopram both did not enhance reasoning and problem solving either (47, 71). On the contrary, Golightly et al. (87) found that patients who underwent tryptophan depletion during the first session performed significantly worse on the Wisconsin Card Sorting Test (WCST) compared to placebo. However, this effect was not present during the second session. The WCST requires subjects to sort cards by a certain parameter. The subject is not told which parameter. The sorting principle changes after 10 correct responses in a row. Thus subjects have to acquire a certain strategy to sort the cards. The authors concluded that tryptophan depletion only affected strategy acquisition and that once this is learned; tryptophan depletion did not interfere with application of this strategy. Thus, based on the results described above, 5-HT3a antagonist and SSRI's did not effectively enhance reasoning and problem solving. However, a role for serotonin in this aspect of cognition cannot be ruled out since tryptophan depletion did interfere with strategy acquisition.

Partial nicotinic α<sup>7</sup> agonist DMXB-A had no enhancing effects on reasoning and problem solving in the study by Freedman et al. (52). In addition, studies investigating the effects of the acetylcholinesterase inhibitors donepezil and rivastigmine did not find improvement in performance on a reasoning and problem solving task either (50, 53, 56, 58). Hence, based on these results, nicotinic α<sup>7</sup> agonists and acetylcholinesterase inhibitors do not appear to significantly enhance reasoning and problem solving in schizophrenia.

In line with the results of Geffen et al. (35) described earlier, no positive results were obtained with other GABAergic interventions strategies. Adjunctive therapy with partial GABA α2/α<sup>3</sup> agonist MK-0777 did not enhance reasoning and problem solving abilities in the study by Buchanan et al. (59). In addition, adjunctive topiramate (an antiepileptic drug which potentiates GABAergic transmission probably through its AMPA/kainite receptor antagonistic properties) treatment did not enhance reasoning and problem solving either (92). Zoccali et al. (93) examined the effects of adjunctive lamotrigine, an anticonvulsant drug which reduces excessive glutamate release in the brain via inhibition of voltage-gated sodium and calcium channels (94). Reasoning and problem solving did not improve.

The potential role of norepinephrine in reasoning and problem solving in schizophrenia was examined in a pilot study by Friedman et al. (60) by adding atomoxetine to antipsychotic treatment. They found no improvement in this aspect of cognition.

In conclusion, the available evidence suggests that modulation of GABA and glutamate transmission and norepinephrine reuptake inhibition does not have enhancing effects on reasoning and problem solving abilities in schizophrenia.

Some positive results have been found with psychostimulant drugs. Compared to placebo, performance on a reasoning and problem solving task improved after a single dose of damphetamine [indirect dopamine D<sup>1</sup> agonist, (72)]. Contrary to these results, Turner et al. (63) found that, compared to placebo, response latency on the Tower of London task was significantly lower after modafinil administration. The number of attempts to obtain the correct solution did not differ between the two groups. Armodafinil, the longer-lasting isomer of modafinil, did not affect reasoning and problem solving abilities (61). Thus, although armodafinil and modafinil did not enhance reasoning and problem solving, d-amphetamine did improve this aspect of cognition. This suggests that dopamine D<sup>1</sup> agonists are a potential area of research for further study of reasoning and problem solving in schizophrenia.

Pharmacological intervention strategies that do not directly influence neurotransmission were also investigated. Levkovitz et al. (73) found that reasoning and problem solving abilities improved after add-on minocycline treatment whereas performance in the placebo group did not change. Both creatine and glucose had no enhancing effects on reasoning and problem solving; daily creatine administration in addition to antipsychotic treatment did not have beneficial effects on working memory in the study by Kaptsan et al. (65). Stone et al. (68) found no improvement in reasoning and problem solving after a single dose of glucose. Hence, despite of the study's limitations, minocycline's putative ability to improve reasoning and problem solving abilities should be studied further.

In conclusion, although positive results were found with depot risperidone treatment in studies with important limitations, antipsychotics do not seem to effectively improve reasoning and problem solving in schizophrenia. The available studies did not show enhancing effects of the neurotransmitters serotonin, acetylcholine, GABA, or norepinephrine on reasoning and problem solving. However, dopamine D<sup>1</sup> agonists may have potential in this cognitive domain since one positive result was found with a single dose d-amphetamine. Finally, minocycline had enhancing effects on reasoning and problem solving abilities. However, this study has important limitations. Therefore these results are in need for replication.

## **SOCIAL COGNITION**

In general terms, social cognition refers to the cognitive processes used to decode and encode the social world (95). Of all the MATRICS cognitive domains, social cognition has received the least attention in research. This is probably due to the fact that it is a relatively new area in schizophrenia research (96) and that the boundaries of this domain are not entirely clear (96). Final screening yielded no more than seven articles on social cognition and pharmacology using validated outcome measures.

A study by Mizrahi et al. (97) examined the effects of antipsychotic treatment on social cognition in schizophrenia. They investigated the effects of clozapine, risperidone, olanzapine, and loxapine on Theory of Mind (TOM). TOM is an aspect of social cognition which refers to the ability to understand intentions of others and to recognize that their actions are guided by beliefs about the world (97). They found that TOM improved after 2 weeks of medication use and continued to improve during the rest of the trial. Unfortunately, the authors did not differentiate between the four different antipsychotics. However, all four types of medication have high affinity for dopamine D<sup>2</sup> receptors as well as the 5-HT2a receptors (98, 99). These results are in line with the results of Sumiyoshi et al. (100) who found that perospirone treatment improved performance on a social cognition task. Nonetheless, both studies did not include a control group and used small samples. Hence, given the longitudinal design of this study, natural progression and practice effects cannot be excluded. Additionally, the high drop-out rate in the study of Sumiyoshi et al. (100) could have led to a bias in the sample. Behere et al. (101) found that performance on an emotion recognition task improved after risperidone treatment. However, it must be noted that they did not include a control group and that time of follow-up assessment was not equal for all patients. Opposed to these results, the study by Harvey et al. (23) discussed earlier did not find improvement in performance on an emotional recognition task in patients treated with either quetiapine or risperidone.

Hence, these results imply preliminary evidence that dopamine and serotonin are important for TOM related aspects of social cognition but not for emotion recognition.

The possible role of GABA in social cognition was investigated by Buchanan et al. (59). They found no improvement with the partial GABA α2/α<sup>3</sup> agonist MK-0777 when added to antipsychotic treatment. The psychostimulant armodafinil also did not improve social cognition (61). Hence, GABA α2/α<sup>3</sup> receptors are not prime candidates for enhancing social cognition in schizophrenia.

Another line of research focused on the role of oxytocin in social cognition. This neuropeptide is known for its role in positive social behavior (102). Pedersen et al. (103) found that intranasal oxytocin administration improved social cognition. This implicates a role for oxytocin in social cognition. Since previous research showed that plasma oxytocin levels are lower in schizophrenia compared to controls (104), it could be hypothesized that increasing oxytocin levels improves social cognition in schizophrenia. However, the study of Pedersen et al. (103) has some important limitations. First, the sample size was small. Moreover, they carried out multiple tests in this small sample; therefore the results have to be interpreted with caution. Studies investigating the role of oxytocin in social cognition are sparse. Therefore, to determine the exact role of oxytocin in social cognition more research needs to be conducted on this topic.

In conclusion, the available seven studies suggest preliminary evidence for a role of dopamine, serotonin, and oxytocin in social cognition in schizophrenia patients. GABA does not seem to be a promising target for enhancement of social cognition. However, these studies have important limitations. Therefore future research needs to confirm the role of these transmitters in social cognition.

## **DISCUSSION**

With this review we aimed to provide an outline of the underlying neuropharmacological mechanisms of the separate MATRICS domains. Although some potential targets were identified, overall, results of previous studies attempting to identify potential pharmacological targets for cognitive enhancement have been unsatisfactory. This review has shown that dysfunction in separate cognitive domains seems to arise from different underlying neuropharmacological mechanisms. This suggests that schizophrenia patients with different cognitive impairments could benefit from different (adjunctive) pharmacological agents. The identified potential molecular targets, which include dopamine D<sup>1</sup> receptors, serotonin 5-HT1a and 5-HT3a receptors, nicotinic α<sup>7</sup> receptors, GABA<sup>A</sup> receptors, and NMDA receptors, are described more extensively below.

## **DOPAMINE D<sup>1</sup>**

Although George et al. (48) were not able to detect positive results with a single dose of D<sup>1</sup> agonist dihydrexidine, Pietrzak et al. (72) found improvement in both processing speed and reasoning and problems solving abilities (both aspects of executive functioning) after a single dose of d-amphetamine, an indirect D<sup>1</sup> agonist. Decreased dopaminergic neurotransmission in the PFC has been hypothesized to be associated with cognitive dysfunction in schizophrenia (36). Since D<sup>1</sup> receptors are highly abundant in the PFC (105), this receptor subtype has been particularly associated with executive function and working memory. A PET study by Okubo et al. (106) found decreased D<sup>1</sup> receptors in the PFC in schizophrenia which was indeed associated with poorer executive functioning. Animal studies provide additional evidence for cognitive enhancing effects of D<sup>1</sup> receptor agonists as low doses of several D<sup>1</sup> agonists were found to enhance cognition in non-human primates (107, 108). Because a majority of dopamine receptors in the PFC belong to the D<sup>1</sup> subtype, and not to the D<sup>2</sup> subtype (109), which is related to psychotic symptom severity, D<sup>1</sup> receptors may be a feasible molecular target for enhancement of executive function related aspects of cognition without exacerbation of psychotic symptoms.

## **SEROTONIN 5-HT1a AND 5-HT3a RECEPTORS**

(Partial) 5-HT1a receptor agonism was found to improve verbal learning and memory (38–40). These enhancing effects may be due to the high density of 5-HT1a receptors in the hippocampus (36), which is an area in the brain well known for its role in memory. In addition, preclinical studies showed that both 5- HT1a agonists and antagonist enhanced cognition in rats (110, 111). 5-HT3a receptor antagonism was associated with improvement in visual learning and memory (43, 44). Contrary to other subtypes of serotonin receptors, the 5-HT3a receptor is the only ligand-gated ion channel subtype (43). To date, not many studies investigated the potential role of 5-HT3a receptors in cognition. However, since both Akhondzadeh et al. (43) and Levkovitz et al. (44) found improvement in visual learning and memory with the 5-HT3a receptor antagonist ondansetron, these receptors might be

a promising molecular target for enhancement of visual learning and memory in schizophrenia.

## **ACETYLCHOLINE NICOTINIC** α**<sup>7</sup> RECEPTORS**

The role of acetylcholine in cognition (particularly in learning, memory, and attention) has been widely established and central dysfunction of the cholinergic system has been found to be associated with cognitive symptoms in neurological diseases as Alzheimer and Parkinson's disease (112–114). However, with the exception of positive results found with nicotine and a partial nicotinic α<sup>7</sup> agonist on attention (52, 89), cholinergic interventions strategies did not affect any of the cognitive domains in the available studies. However, all the available studies that met the inclusion criteria used acetylcholinesterase inhibitors or nicotinic agonists. Although (mostly post-mortem) both nicotinic and muscarinic receptor abnormalities have been repeatedly found in schizophrenia (115–120), nicotinic receptor antagonists do not appear to impair cognition in the same manner as antimuscarinic drugs (121). This might explain the lack of improvement in the described studies. Indeed, a small pilot study by Shekar et al. (122) found that the muscarinic receptor agonist xanomeline improves verbal, visual, and working memory in patients with schizophrenia or schizoaffective disorder. Moreover, xanomeline improved cognition in Alzheimer disease (123). Thus, although positive effects on attention were found with nicotine and a partial nicotinic α<sup>7</sup> agonist, future studies should focus on muscarinic agents.

## **GABA<sup>A</sup> RECEPTORS**

Both GABA<sup>A</sup> receptor antagonists and agonists were found to affect working memory and verbal learning and memory (35, 80). Multiple studies have shown reduced GABAergic transmission, especially in the PFC, in schizophrenia (124). The PFC is strongly involved in working memory functioning (125) and animal studies have shown that appropriate GABA transmission in the dorsolateral prefrontal cortex (DPFC) is essential to adequate working memory functioning (36, 126). Lewis et al. (37) showed that altered GABA transmission in the DPFC is possibly limited to certain cell classes, such as the chandelier cells, which synchronize the activation of the pyramidal neurons via the GABA<sup>A</sup> receptor subtypes. Therefore, GABA<sup>A</sup> receptors may be a promising molecular target for enhancement of working memory.

## **GLUTAMATE NMDA RECEPTORS**

Glutamate, the primary excitatory neurotransmitter in the mammalian brain, has been linked to learning and memory because of its principal role in modulating long-term potentiation (127), and hyperactivity of glutamatergic neurotransmission has been implicated in schizophrenia (36). It has been hypothesized that cognitive dysfunction in schizophrenia is due to hypofunction of the NMDA receptor (81). Small increases in NMDA-dependent glutamate transmission might enhance cognition, whereas excessive stimulation might have neurodegenerative consequences (36). Activation of the NMDA receptors leads to synthesis of nitric oxide,which is able to further increase the excitotoxicity by increasing glutamate release from presynaptic neurons and inhibition of glial glutamate transporters (73). Indeed, minocycline (which blocks nitric oxide induced neurotoxicity) was found to improve

visual learning and memory, working memory, and reasoning and problem solving abilities (73). Although d-cycloserine (which non-competitively enhances NMDA neurotransmission) was not found to improve attention and working memory (81), pharmacological agents reducing the neurotoxic effects of extensive glutamate might still be a promising intervention strategy for cognitive enhancement.

## **CAN THE NEGATIVE RESULTS BE EXPLAINED BY STUDIES LIMITATIONS?**

A substantial number of the studies evaluated in this review addressing pharmacological cognitive enhancement in schizophrenia report negative results. It is debatable whether this is due to ineffectiveness of the agents used, or that potential results were obscured by methodological shortcomings as cognition research in schizophrenia deals with pertinent limitations. First, in many of the studies included in this review, cognition was not the primary outcome parameter. As a result, the design of these studies was not always optimal to measure cognitive enhancement. Second, the sample sizes are often too small to detect clinically relevant effects. Third, concurrent medications may interfere with the investigated pharmacological agents. Patients with schizophrenia are generally treated with antipsychotics. Because the entire mechanism of action of these medications is not completely understood, it cannot be ruled out that antipsychotics alter the effects of the added agents (62). Moreover, patients often use concomitant medications such as benzodiazepines, antidepressants, and anticholinergics, which could influence the effects of the adjunctive pharmacological agents as well. Especially anticholinergic medication is well known for its adverse effects on cognition (29). Therefore, studies including only medication-naïve patients at early stages of the disease are necessary. Fourth, studies often do not use fixed doses. It is possible that certain agents are only effective in a certain dose. If studies do use a fixed dose, this dose is often established for treating the illness or symptoms it was originally used for. However, it is not necessarily so that the same dose is required for cognitive enhancement in schizophrenia. Fifth, research seems to focus on the direct or semi-direct enhancing effect of modulating certain receptors while cognitive decline has most often been a process of a longer period of time possibly implying that enhancement studies should equally allow for more time to yield effects. Also, different paradigms could focus on neuroprotective targets and preventing cognitive decline early in the disorder in contrast to enhancing cognition after impairment. Sixth, not all studies report information about substance use of the participants. Patients with schizophrenia often abuse substances as tobacco, alcohol and cannabis, and epidemiological studies found high co-morbidity rates of substance abuse (40–60%) (128, 129). Particularly tobacco use can interfere with nicotinic receptor agents as it causes desensitization of this receptor subtype. Although most of the studies excluded patients with a (recent) co-morbid diagnosis of alcohol or drug abuse/dependence, tobacco using patients are almost never excluded. Seventh, studies often use different neuropsychological tasks to measure the same aspects of cognition which makes it difficult to compare the results. Therefore, studies should use a standardized cognitive battery, such as the MATRICS Consensus

Cognitive Battery [MCCB; (84)] or the CANTAB cognitive battery (130). The MCCB in particular, has been composed to reliably assess cognition in schizophrenia patients. At present, only three of the reported studies used this battery to measure cognition (50, 52, 59). Finally, not all studies used a control group and repeatedly conducted the same test battery without adequately correcting for potential practice effects. Therefore, improvement often cannot be reliably attributed to the effect of the pharmacological agent. Future research should take these limitations in to account in experimental setup and optimize their design for cognition measurement. Ideally, future studies should include a placebo or other control group and should ensure sufficient power by including enough participants. Furthermore, medication-naïve patients should be recruited, preferably non-smokers who do not use other substances as cannabis and alcohol. Regarding the design, future studies should choose cognition as primary outcome measure and optimize the design by using a fixed dose and make sure to choose a dose suitable for schizophrenia patients. Finally, future studies should make sure to choose sufficient trial duration and to use a standardized cognitive test battery which covers all the MATRICS domains. In this manner, studies will be better comparable and the effects of possible confounding variables will be limited. To summarize, the negative results of "pharmacological cognitive enhancement studies" could partially be explained by studies' limitations.

## **CONCLUSION**

Although some potential targets were identified, overall, results of previous studies attempting to identify potential pharmacological targets for cognitive enhancement have been disappointing. This review has shown that dysfunction in separate cognitive domains may arise from different underlying neuropharmacological mechanisms which suggests that schizophrenia patients with different cognitive impairments could benefit from different intervention strategies. Although development of effective cognitive enhancers is a complex process, it is an exciting challenge for this area of research as improvement of cognition contributes significantly to the quality of life of these patients.

## **AUTHOR CONTRIBUTIONS**

Wilhelmina A. M. Vingerhoets wrote the first draft of the manuscript. All authors contributed to and have approved the final manuscript.

### **REFERENCES**


for cognitive impairment in schizophrenia. *J Clin Psychopharmacol* (2008) **28**:59–63. doi:10.1097/jcp.0b013e318161318f


function in patients with schizophrenia.*Arch Gen Psychiatry* (2007) **64**:156–67. doi:10.1001/archpsyc.64.2.156


130. Levaux M-N, Potvin S, Sepehry AA, Sablier J, Mendrek A, Stip E. Computerized assessment of cognition in schizophrenia: promises and pitfalls of CANTAB. *Eur Psychiatry* (2007) **22**:104–15. doi:10.1016/j.eurpsy.2006.11.004

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 August 2013; accepted: 16 November 2013; published online: 04 December 2013.*

*Citation: Vingerhoets WAM, Bloemen OJN, Bakker G and van Amelsvoort TAMJ (2013) Pharmacological interventions for the MATRICS cognitive domains in schizophrenia: what's the evidence? Front. Psychiatry 4:157. doi: 10.3389/fpsyt.2013.00157 This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Vingerhoets, Bloemen, Bakker and van Amelsvoort. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

REVIEW ARTICLE published: 04 February 2014 doi: 10.3389/fpsyt.2014.00011

# A critical review of pro-cognitive drug targets in psychosis: convergence on myelination and inflammation

#### **Rune A. Kroken<sup>1</sup>\*, Else-Marie Løberg1,2,Tore Drønen<sup>1</sup> , Renate Grüner 3,4, Kenneth Hugdahl 1,2,3,5 , Kristiina Kompus <sup>2</sup> , Silje Skrede5,6,7 and Erik Johnsen1,8**

<sup>1</sup> Division of Psychiatry, Haukeland University Hospital, Bergen, Norway

<sup>3</sup> Department of Radiology, Haukeland University Hospital, Bergen, Norway


<sup>8</sup> Department of Clinical Medicine, University of Bergen, Bergen, Norway

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Joshua T. Kantrowitz, Columbia University, USA Wen-Jun Gao, Drexel University College of Medicine, USA Paolo Fusar-Poli, King's College London, UK

#### **\*Correspondence:**

Rune A. Kroken, Division of Psychiatry, Haukeland University Hospital, PB 1400, N-5021 Bergen, Norway

e-mail: rune.kroken@helse-bergen.no

Antipsychotic drugs have thus far focused on dopaminergic antagonism at the D2 receptors, as counteracting the hyperdopaminergia in nigrostriatal and mesolimbic projections has been considered mandatory for the antipsychotic action of the drugs. Current drugs effectively target the positive symptoms of psychosis such as hallucinations and delusions in the majority of patients, whereas effect sizes are smaller for negative symptoms and cognitive dysfunctions. With the understanding that neurocognitive dysfunction associated with schizophrenia have a greater impact on functional outcome than the positive symptoms, the focus in pharmacotherapy for schizophrenia has shifted to the potential effect of future drugs on cognitive enhancement. A major obstacle is, however, that the biological underpinnings of cognitive dysfunction remain largely unknown. With the availability of increasingly sophisticated techniques in molecular biology and brain imaging, this situation is about to change with major advances being made in identifying the neuronal substrates underlying schizophrenia, and putative pro-cognitive drug targets may be revealed. In relation to cognitive effects, this review focuses on evidence from basic neuroscience and clinical studies, taking two separate perspectives. One perspective is the identification of previously under-recognized treatment targets for existing antipsychotic drugs, including myelination and mediators of inflammation. A second perspective is the development of new drugs or novel treatment targets for well-known drugs, which act on recently discovered treatment targets for cognitive enhancement, and which may complement the existing drugs. This might pave the way for personalized treatment regimens for patients with schizophrenia aimed at improved functional outcome.The review also aims at identifying major current constraints for pro-cognitive drug development for patients with schizophrenia.

**Keywords: schizophrenia, cognition, glutamate, myelin, inflammation, immunology, connectivity, neuroimaging**

## **INTRODUCTION**

Schizophrenia is a severe mental disorder with typical onset in the late teens or early adulthood and a chronically relapsing remitting course in the majority of cases (1). Schizophrenia remains a leading cause of years lived with disability (2), and active psychosis has been ranked among the most disabling disorders by severity in the general population (3).

The clinical picture of schizophrenia is heterogeneous, but in the vast majority of patients, cognitive dysfunctions are present with adverse impact on daily functioning (1). Antipsychotic drugs have for six decades been a cornerstone in the treatment of the disorder but a fundamental limitation is their small effect sizes in the cognitive symptom domain (4) as the drugs are primarily effective against the positive symptoms of psychosis including hallucinations and delusions (4). Since the serendipitous discovery of chlorpromazine's antipsychotic properties, dopaminergic antagonism has served as a template for all subsequent antipsychotics through counteracting subcortical hyperdopaminergia (5). The drugs following chlorpromazine are usually classified as typical or first generation antipsychotics (FGAs), whereas clozapine much later became the prototype for the atypical or secondgeneration antipsychotic (SGA) drugs that were developed in the 1990s onward (4, 6). Only modest differences in antipsychotic efficacy have been found among FGAs and SGAs (7), but the SGAs seem to display a broader therapeutic repertoire including also pro-cognitive properties, although of a small magnitude (8). Pharmacologically, the FGAs have strong affinities for the dopamine type 2 (D2) receptor, whereas the SGAs are characterized by a

<sup>2</sup> Department of Biological and Medical Psychology, University of Bergen, Bergen, Norway

<sup>4</sup> Department of Physics and Technology, University of Bergen, Bergen, Norway

more pronounced 5-HT2A antagonism than FGAs, and a lower potency for the D2 receptor (9, 10). The narrow focus on the dopaminergic system and striatal hyperdopaminergia in particular has most likely contributed to the limited evolvement of more effective drug treatment options as it has become increasingly clear that dopaminergic disturbances account for only parts of the clinical picture (particularly for positive symptoms) (11– 13), and hyperdopaminergia is not present in all schizophrenia patients (14).

Recognizing the large impact that the schizophrenia associated cognitive dysfunctions have on functional outcome (15, 16), the efforts to improve the pharmacological treatment have moved toward cognitive enhancement, and pro-cognitive effects have been proposed as primary targets in future drug studies (17, 18). A major obstacle restricting hypothesis-driven pro-cognitive drug development has been the lacking knowledge of the structural and functional brain alterations underlying cognitive dysfunction in patients with schizophrenia (19, 20). Major scientific advances in unveiling the neurobiology of schizophrenia have emerged in recent years, however. The knowledge of transmitter dysfunctions especially is widened (21), accompanied by evidence of disturbances of myelination and inflammation/immunology in the pathophysiology of schizophrenia (22, 23). Thus, going beyond the neurotransmitter disturbances associated with schizophrenia, the present work reviews the latest evidence on myelin alterations, and the involvement of neuroinflammation/immunology in schizophrenia, with putative relevance for pro-cognitive drug effects in schizophrenia and related psychotic disorders. Moreover, these areas seem at least partly inter-related and the review aims at integrating the evidence from key publications in a joint model. Increased insight into the relevant functional and structural alterations might pave the way for more efficacious, targeted, and personalized pro-cognitive drug treatment opportunities aimed at improving real-world functioning. The review focuses on both previously unknown treatment targets for existing antipsychotic drugs, and potentially novel pro-cognitive drugs addressing extra-dopaminergic treatment targets, and gives an update also commenting on current constraints of pro-cognitive drug development for patients with schizophrenia.

## **COGNITIVE DYSFUNCTION IN PATIENTS WITH SCHIZOPHRENIA**

Cognitive function may be understood in terms of inter-related complex neural networks, influenced by different neurotransmitters, cytokines, and other substances including brain-derived neurotropic factor (BDNF) acting as neuromodulators (24).

Cognitive deficits are viewed as core characteristics of schizophrenia, with clinically significant cognitive impairments observed in a large majority of patients, and most patients experience lower cognitive functioning than would be expected from parental levels of education (16, 25). Cognitive deficits are seen before the development of psychosis (25–28), and cognitive decline and intellectual stagnation have been suggested to constitute a risk phenotype for developing schizophrenia (25). Furthermore, cognitive deficits are relatively stable after psychotic breakthrough and still present when the symptoms of psychosis have been treated (29, 30). A general cognitive dysfunction across cognitive domains is present, with additional selective deficits in working memory, executive function, attention, verbal fluency, episodic memory, and processing speed (15, 31). The effect size of the cognitive dysfunction compared to healthy subjects is close to 1.0 (32, 33), and in several neurocognitive dimensions the impairment is between 1 and 2 SDs, indicating a clinically significant loss of function (25, 34–37). Cognitive functioning has greater impact than positive psychotic symptoms in determining the patient's real-world functional outcome (15, 38–41), and relationships between more specific cognitive domains and functional domains have also been reported (38, 40, 42). Thus, cognitive impairment has been put forward as the main target for novel treatment interventions in schizophrenia (25).

A significant challenge has been to develop clinical meaningful measures of cognitive dysfunction for pro-cognitive drug advancement. A major contribution to establish a standardized and manageable test battery for measuring cognition in schizophrenia has been the development of the Measurement and Treatment Research to Improve Cognition in Schizophrenia (MATRICS) Consensus Cognitive Battery (MCCB) (43). The MCCB composite score has been suggested as the best outcome measure based on findings of low practice effects, high test–retest reliability, and good external validity (44), while a recent factor analysis identified a three factor model of processing speed, attention/working memory, and learning as the best representations of data in this particular study (45).

## **DOES ABNORMAL MYELINATION CAUSE COGNITIVE DYSFUNCTIONS IN PATIENTS WITH SCHIZOPHRENIA?**

Myelin, the main constituent of white matter, is essential for controlling and regulating conduction velocities along axons and thus the synchronicity of brain signals across different brain regions, and for controlling synapse plasticity (46). Decreased myelin integrity might accordingly alter normal signal transduction. In the following, we will focus on one major line of recent achievements in schizophrenia research, namely the rapidly expanding knowledge of abnormal central nervous myelination in patients with schizophrenia and the possibilities to prevent or repair myelin pathology. White matter brain alterations and functional connectivity problems that have been linked to cognitive dysfunctions in schizophrenia will be briefly described, followed by in-depth descriptions of new findings of myelin alterations linked to cognition. Again, a major problem is to identify manageable tests/measures of myelinization and myelin integrity that can be used in drug discovery programs in a reliable way. Finally, we will discuss candidate drug treatments counteracting deficits of myelinization in patients with schizophrenia

## **WHITE MATTER CHANGES AND COGNITION**

Several distinct alterations in gray and white matter are well established for schizophrenia patients as a group (47). Reviews of white matter tract pathology in patients with schizophrenia [mostly studies based on voxel-based morphometry (VBM) and diffusion tensor imaging (DTI)] have identified widespread and, in many studies, progressive changes (47). A meta-analysis of 15 DTI studies by Ellison-Wright and Bullmore (48) identified reductions (in all studies) of left frontal deep white matter and

left temporal white matter. Thus, the authors suggested that two different networks may be affected; one connecting the frontal lobe with thalamus and cingulate gyrus, the other connecting the temporal lobe with, e.g., insula, amygdala, hippocampus, and the frontal lobe. A review by Walterfang et al. (49) of white matter pathology in patients with schizophrenia identified changes in the left uncinate fasciculi as the most replicated finding, furthermore that deficits of information processing was associated with volume reductions in the uncinate and inferior longitudinal fasciculi, anterior internal capsule, and corpus callosum (49). White matter pathology related to cognition has also been found in first-episode patients with psychosis. For example, Rigucci and colleagues (50) reported a relationship between speed of processing and visual memory and white matter fractional anisotropy (FA) in fronto-temporal tracts. Focusing on intra-cortical white matter, several clusters of superficial white matter (SWM) composed of U-shaped fibers connecting neighboring gyri together with intra-cortical axons and fibers from deep matter pathways displayed reduced FA in a study comparing patients with schizophrenia to healthy controls. The correlations between SWM-FA and cognition displayed in healthy controls were missing in patients (51).

## **FUNCTIONAL CONNECTIVITY STUDIES**

Recently, also related to myelination, the search for brain alterations underlying the psychopathology of schizophrenia has refocused to examine neuronal dysconnectivity, i.e., abnormal communication between local and distributed neuronal circuits or disrupted integration of brain activation (52–54). Fornito et al. (53) reviewed data on brain connectivity, and concluded that schizophrenia is associated with: (1) a widespread and possibly context-independent connectivity deficits, (2) additional transient states of hyper- and/or hypo-connectivity related to specific tasks.

Disturbed connectivity within the frontal lobes has been found to be related to working memory and to executive functioning (52, 54). This includes abnormal connectivity in fronto-parietal, fronto-cerebellar, and fronto-hippocampal networks (52). Verbal processing deficits have been shown to be related to circuits in the language areas of the brain, particularly to fronto-temporal dysfunction (52, 54). Indeed, theories have been developed concentrating on the effect of reduced executive cognitive control over the language processing in the temporal lobes as central to the development of the symptoms of psychosis, for instance auditory verbal hallucination (55).

Of future interest may be the influence of antipsychotic drugs on the dynamic interaction between large-scale cortical networks related to resting and effort-mode states (56). Using fMRI, Hugdahl et al. (57) found overlapping activation in a large-scale network to occur with a range of different cognitive tasks implicating the prefrontal cortex [see also Ref. (58)] suggesting that effective cognitive processing would require down-regulation of the resting-state network in situations with increasing demands for the up-regulation of the effort-mode network, and that cognitive impairment in schizophrenia may result from failure of network up- and down-regulation dynamics. Thus, based on brain imaging results, therapeutic agents inducing stabilization and, ideally, normalization of large-scale cortical network dynamics could possibly have pro-cognitive effects. However, thus far no drug has been reported to enhance connectivity (52).

The major question related to the structural basis of functional dysconnectivity was addressed in the review by Fornito et al. (53). The authors concluded with major overlaps between functional and structural findings, but also differences, which could at least partly be caused by intrinsic problems of the methods of structural measurements (see Measuring Myelination).

## **MEASURING MYELINATION**

Several methodological issues need consideration regarding the use of imaging techniques for myelin assessments. The DTI signal may reveal changes in white matter integrity but has limitations in identifying the exact nature of these changes, as the FA reflects different processes like demyelination, axonal swelling, or atrophy, and is sensitive to confounding effects of fiber crossings (59). In a recent review by Du and Öngür (60), two novel MR-based approaches – diffusion tensor spectroscopy (DTS) and magnetization transfer ratio (MTR) – were reviewed as measures of axonal diameter (DTS) and myelin volume (MTR). Combining MTR with DTI, Palaniyappan et al. (61) recently showed that the degree of FA reduction in areas with decreased myelin volume predicted impaired processing speed, leading the authors to suggest that combined DTI/MTR could be used as a treatment measure in the development of pro-cognitive drugs.

A study applying graph theory to analyze data from DTI combined with MTR showed that patients with schizophrenia have a weaker globally integrated structural brain network when compared to healthy controls (62). The same study concluded that the frontal hubs had a less central role than in healthy controls, which is in accordance with a reduced structural capacity to integrate information across brain regions. Moreover, a new study combining DTI and resting-state fMRI comparing patients with schizophrenia to healthy controls identified a selective disruption of brain connectivity among central hub regions with a potential communication capacity reduction and brain dynamics alterations (63).

#### **MYELIN DEFICIENCIES IN SCHIZOPHRENIA**

Compelling evidence from several different lines of research has, in recent years, implicated dysfunctional myelin and oligodendrocytes, the cells responsible for wrapping myelin around neuronal axons in the brain, in the pathology of schizophrenia and related psychoses (22, 64–66). Sources of evidence include genetics and pharmacogenetic studies, post-mortem studies, and as briefly reviewed in the previous subsections of brain imaging studies, particularly DTI studies.

Normal myelination starts in the late part of fetal life and occurs most rapidly during the first 2 years after birth, but continues to adulthood in a region- and brain matter-specific manner (22, 65, 67). Initial myelination occurs subcortically, thus increasing the conduction velocities across different brain regions while leaving the intra-cortical portion of the axon unmyelinated during childhood (65). Interestingly, maturation of white matter in the prefrontal cortex and in fronto-temporal tracts occurs later, in late adolescence and early adult life, coinciding in time with the peak age of schizophrenia onset (1). Possibly, disturbances in the

maturation of myelin around the end of the myelination stage could be one of the events contributing to the peak debut incidence of schizophrenia in this age group (66). The prefrontal cortex is one of the regions most robustly implicated in schizophrenia pathology, and working memory performance, which engages this region, has been found to be positively associated with the level of white matter maturation (67). Later intra-cortical myelination occurs for the most part in adulthood and is associated with the refinement of cognitive functions and brain plasticity (65). Importantly, myelination is a dynamic process, and oligodendrocyte progenitor cells continue to differentiate throughout life (68). In schizophrenia, intra-cortical myelin deficits seems to be more pronounced than subcortical deficits, which have been found to be more strongly associated with the duration of illness. Indeed, evidence suggests that myelin integrity is closely related to both cognitive functioning and to the symptoms of a variety of psychiatric and neurological disorders (22, 46).

Alterations particularly in dopaminergic transmission (69, 70), and also in the glutamatergic system, (11, 71) have been robustly implicated in the pathology of schizophrenia for several decades. Glutamatergic *N*-methyl-d-aspartate (NMDA) receptor antagonists, such as phencyclidine (PCP) and ketamine, have the potential to induce not only positive symptoms of psychosis but also negative symptoms and cognitive dysfunction resembling those seen in schizophrenia (11, 71). This has given rise to the NMDA receptor hypofunction hypothesis of schizophrenia (21, 72). The hypothesis has been extensively investigated in animal models but several limitations related to how well evidence from preclinical studies translate into schizophrenia patients have been pointed to (72). Seemingly, paradoxically another model of schizophreniarelated cognitive dysfunction is that of glutamatergic hyperactivity (73). The model is based on other observations of increased prefrontal glutamate release following NMDA receptor antagonism. The apparent paradoxical glutamate findings might be explained by different NMDA receptor properties related to the different locations and/or sub-compositions of the different NMDA receptors as elaborated by Zhou et al. (74). Interestingly, there are several points of convergence between myelin deficits and neurotransmitter alterations. Firstly, glutamate is excitotoxic when in excess, and oligodendrocyte progenitor cells display great vulnerability for glutamatergic excitotoxicity (68). Increased glutamate levels have been found in some phases of schizophrenia (75–77). Secondly, there is emerging evidence that dysfunctional myelin and oligodendrocytes could increase striatal dopamine levels, see Ref. (22) for details. One example could be that if glutamatergic projections from the prefrontal cortex to brainstem areas suffer from myelin damage, this might lead to lower excitatory input at inhibitory brainstem GABAergic interneurons, resulting in less inhibition of dopaminergic mesolimbic projections and consequently striatal hyperdopaminergia as outlined by Schwartz et al. (78).

#### **MYELIN AS DRUG TARGET FOR PRESENT ANTIPSYCHOTICS**

Animal studies suggest that myelin-protecting and oligodendrocytestimulating properties may be a therapeutically relevant mechanism of action for at least some SGAs (61, 79, 80). In a DTI study of schizophrenia patients in exacerbation, Garver et al. (81) found increased diffusibility consistent with decreased myelin

integrity during acute psychosis. After initial examination, patients were treated with the SGAs risperidone or ziprasidone, or the FGA haloperidol, with partial restoration of myelin integrity after 4 weeks in the subgroup that responded to treatment. Bartzokis et al. found higher volumes of intra-cortical myelin in schizophrenia patients treated with oral risperidone compared to those treated with the FGA fluphenazine decanoate (82). In a recent study by the same group, the long acting injection formulation of risperidone was associated with increased intra-cortical myelin volume compared to the oral formulation of the same compound (64). Myelin primarily consists of lipids, particularly cholesterol, which are mainly synthesized *de novo* in the brain. Interestingly, several antipsychotic agents have been demonstrated to directly induce lipogenesis through the sterol regulatory element binding protein (SREBP) system, and these lipogenic effects have been suggested to contribute to myelin-stimulating effects of antipsychotic agents (83–85). In this regard, it is highly interesting that clozapine, with its superior clinical efficacy, is also among the antipsychotics associated with the most pronounced metabolic adverse effects; in fact, a correlation between clinical improvement and increase in serum lipid levels has repeatedly been demonstrated (86–88). Summing up, a small but consistent body of evidence indicates that some current SGAs have positive effects on myelin volume, with possible distinctions among drugs and drug formulations.

## **OTHER POTENTIAL "MYELIN-ENHANCING" TREATMENT OPTIONS**

In a clinical randomized controlled trial (RCT) by Amminger et al. (89), a markedly decreased progression rate to psychosis was found in at risk subjects receiving high-dose polyunsaturated fatty acids (PUFAs). PUFAs are involved in the myelination process, and peripheral PUFA levels have been found to be decreased in schizophrenia (90, 91). A recent DTI study in early-phase psychosis patients found an association between level of PUFAs in peripheral erythrocytes and white matter integrity (90). Possibly, PUFA distribution is altered in patients at risk for psychosis, with a link between PUFA levels and white matter integrity. Free radicals can damage membrane PUFAs, and disturbances in fatty acids and membrane phospholipid identified in patients with schizophrenia may be caused by increased oxidative stress according to a review by Yao and Keshavan (92). The same authors point to disruption of antioxidative systems related to schizophrenia, with reduced amounts of non-enzymatic plasma antioxidant components [e.g., albumin, bilirubin, uric acid, ascorbic acid (vitamin C), α-tocopherol (vitamin E)], see also the recent clinical study by Zhang et al. demonstrating a reduced plasma total antioxidant status in a sample of schizophrenia patients (93). Interestingly, PUFAs also have mild anti-inflammatory effects, see Section "Additional Drugs with Anti-Inflammatory Action as Add-on Treatments for Patients with Schizophrenia."

## **INFLAMMATION AND IMMUNOLOGY IN SCHIZOPHRENIA**

#### **IMPLICATING INFLAMMATORY SYSTEMS IN SCHIZOPHRENIA**

Several findings point to a link between inflammatory processes and the pathophysiology of schizophrenia: (1) activated peripheral inflammatory system and neuroinflammation in patients with schizophrenia (94, 95), (2) evidence from genetic studies of correlation between schizophrenia and genes encoding for different components of the immune system (96–98), (3) post-mortem studies demonstrating up-regulated immune genes in the prefrontal cortex of patients with schizophrenia (99), (4) findings that the raised risk of schizophrenia seen after maternal infections acts via immunological mechanism (23), and (5) psychotic symptoms and cognitive dysfunction caused by immunological neurological syndromes (100), e.g., the interesting line of pathophysiological evidence based on findings in autoimmune synaptic encephalitis (limbic encephalitis), where antibody formation against NMDA receptors is associated with a wide range of psychiatric symptoms, in some patients also with syndromes resembling schizophrenia (100). Binding of NMDA antibodies has been found predominantly in the hippocampus (101). Also, in systemic autoimmune disorders such as systemic lupus erythematosus (SLE), patients can suffer from various psychiatric syndromes, but the pathophysiology has yet to be characterized in detail (102). A recent remarkable finding is that the levels of soluble receptors for the pro-inflammatory cytokine tumor necrosis factor (TNF)-α correlates with function in patients with schizophrenia compared to healthy individuals (103), the same study also found increased levels of TNF-α in the treatment resistant group compared to treatment responders.

In the following sections, we highlight some of the main findings related to inflammation/immunology and schizophrenia, with the specific purpose of evaluating the status of immunemodulating treatment with cognitive-enhancing effects.

## **CYTOKINES**

In the context of inflammatory changes in schizophrenia, the majority of findings stem from the innate system, but components of the adaptive immunology system have also been implicated. The macrophages of the innate system induce the release of acutephase proteins [e.g., C-reactive protein (CRP) from hepatocytes] as well as producing and releasing cytokines (97), all markers of inflammation (104). Cytokines are also produced by peripherally activated endothelial cells and monocyte-derived dendritic cells (97).

#### **PERIPHERAL INFLAMMATORY FINDINGS IN SCHIZOPHRENIA**

Multiple studies have shown low-grade disturbances in the peripheral inflammatory system in patients with schizophrenia, see recent reviews (105–107). Pro-inflammatory changes include elevated serum measures of pro-inflammatory cytokines [interleukin (IL)-1β, IL-6, IL-8, TNF-α] and other pro-inflammatory factors [prostaglandin E<sup>2</sup> (PGE2), CRP] and elevated monocyte counts and activated immune cells have been associated with schizophrenia (108). Interestingly, some cytokines [IL-12, IFN-γ, TNF-α, soluble IL-2 receptor (sIL-2R)] may be trait markers for schizophrenia, while others are raised during phases of intensified symptoms (IL-1β, IL-6, and TGF-β) (106, 108). Furthermore, some anti-inflammatory substances are elevated in subgroups of patients with schizophrenia [soluble IL-1 receptor antagonist (sIL-1RA), soluble IL-2 receptor antagonist (sIL-2RA), soluble TNF receptors (sTNFRs)1, sTNFR2 (109), IL-10 (a cytokine with anti-inflammatory properties), and transforming growth factor (TGF)-β] (105). A review (110) identified decreased levels of IL-2

in patients with schizophrenia. IL-2 serves important functions in T-cell-mediated immunologic reactions. Although the findings of disturbances in the peripheral inflammatory system are not without inconsistencies, and the measurements of cytokines are sensible to a wide range of confounders (e.g., the influence of antipsychotic medication and metabolic status), there is firm evidence for a mixed picture of peripheral pro- and anti-inflammatory changes in a large proportion of patients with schizophrenia.

## **CENTRAL INFLAMMATORY FINDINGS IN PATIENTS WITH SCHIZOPHRENIA**

Brain neuroinflammation, in contrast to inflammation in other parts of the body, does not lead to leukocyte recruitment, but is characterized by activated microglia, the mononuclear phagocytes of the brain (111). Microglia develop from primitive myeloid progenitors, and have various roles in the developing human brain including synaptic pruning, remodeling of axons, neuronal differentiation, and programed cell death (112). PET studies using (R)[11C]PK11195 (95, 113) have found microglial activation in schizophrenia, in addition to post-mortem studies demonstrating increased density and activation of microglia, though some of these studies have yielded conflicting evidence (112). Activated microglia have increased levels of the translocator protein (TPSO) located at the mitochondrial membrane, and the binding of TPSO to the (R)[11C]PK11195 is utilized in PET studies as a measure of neuroinflammation, while other PET ligands for activated microglia are under development (113). Free-water imaging (114) is another mode of neuroinflammation imaging, and studies using this novel method have indicated widespread brain inflammation and early signs of neurodegeneration in first-episode patients with schizophrenia (115).

## **Microglia, astrocytes, and the kynurenine pathway**

Activated microglia have many important functions besides phagocytosis, including cytokine production (IL-1β, IL-12, and TNF-α) (94, 111), most importantly crosstalk with the serotonin and glutamate neurotransmitter systems, involving the kynurenine pathway (116). Activated microglia also induce functional changes in brain astrocytes, which release IL-6, IL-10, and TGF-β (94). In the presence of pro-inflammatory cytokines, e.g., IFNγ and TNF-α, tryptophan catabolism to kynurenins in microglia and astrocytes is induced, resulting in the production of quinolinic acid (QUIN) in microglia and KYNA in astrocytes. QUIN is an NMDA agonist and potentially excitotoxic, while KYNA is an NMDA antagonist and also blocks the α7 nicotinic acetylcholine receptor (α7nAChR) (94, 116). **Figure 1** summarizes major points of the activated inflammatory responses in schizophrenia. The demonstration of elevated levels of KYNA in the cerebrospinal fluid (CSF) of patients with schizophrenia by two independent groups (117, 118) in 2001 boosted research in brain inflammatory-related processes implicated in schizophrenia. Elevation of KYNA in the CSF and prefrontal cortex in patients with schizophrenia is now an established finding, and is probably a consequence of peripheral inflammation according to a recent review by Schwarcz et al. (119). Elevated levels of KYNA are associated with decreased cognitive performance in animal models, including altered pre-pulse inhibition of the auditory startle

response, deficits of working memory, and attentional set-shifting paradigms (120, 121), while reducing KYNA resulted in improved cognitive performance in a wide range of tasks (121). Adding to this, the kynurenic pathway also offers a possible explanation of how stress induces psychotic symptoms and cognitive dysfunction in schizophrenia by linking increased levels of glucocorticoids to the production of kynurenine in liver, subsequently converted to KYNA in the brain (121). The kynurenic pathway offers several potential drug targets for the reduction of brain KYNA levels, which could improve cognitive function in schizophrenia. Candidate pharmacological agents related to schizophrenia treatment block, e.g., the enzyme kynurenine aminotransferase (KAT), which transforms kynurenine into KYNA. Most KYNA-lowering drugs still lack behavioral data (122), while cerebrolysin, a peptidergic drug manufactured from purified porcine brain preparations, which lower brain tissue KYNA levels *in vitro* (123) also have been found to improve cognition (124). Interestingly, in a double blind RCT of patients with schizophrenia cerebrolysin (intravenous infusion) as add-on treatment to risperidone was found to improve scores at the Wechsler Adult Intelligence Scale and the Wechsler Memory Scale (124). Cerebrolysin is generally welltolerated (125) and has also demonstrated pro-cognitive effects in patients with dementia (126, 127) and traumatic brain injury (128). In another recent study, d-cycloserine (DSC), a partial agonist at the glycine site of the NMDA receptor, was also found to reduce the levels of KYNA in human post-mortem frontal cortex

homogenates by dose-dependently blocking KAT (129). DSC has also been found to improve memory consolidation and to reduce delusions in combination with cognitive behavioral therapy (130).

## **IL-1**β**, S100B, BDNF**

In addition to elevated KYNA/QUIN, other aspects of an activated brain immune system can also have clinically important effects in schizophrenia. The proinflammatory cytokine IL-1β has been linked to glutamatergic excitotoxicity (131), and close dynamic interactions between IL-1β and the NMDA receptor was the recently demonstrated (132). Pro-inflammatory cytokines (e.g., IL-1) are also candidate causal agents implicated in the cerebral myelin damages discussed above, see (133, 134). The protein S100B, produced by astrocytes, is a marker of astrocyte activation. Serving as an example of the pleiotropic effects associated with many inflammatory factors at low concentration, S100B has beneficial effects, limiting microglia activation, stress-related neuronal damage, and increase glutamate uptake by astrocytes, while at higher concentration it has detrimental effects, including induction of microglial TNF-α secretion and induced production of,e.g., COX2/PGE-2, IL-1β, etc. (135). A meta-analysis concluded that serum levels of S100B may be elevated in schizophrenia, and disruptions of the blood–brain barrier (BBB) and/or active secretion by astrocytes were suggested as possible causes (136).

The neurotrophin BDNF is among the substances secreted by immune cells during inflammation and immune responses (137), and an inverse relationship between levels of BDNF and certain cytokines has been found (138). BDNF has been implicated in widespread processes relevant to the pathophysiology of schizophrenia including neurodevelopment, synapse regulation, and neuroplasticity, as well as effects on cognition, see for example (138–140). BDNF has been found to be reduced in schizophrenia in both drug-naïve and chronic patients, although findings are not entirely consistent across studies. Furthermore, an association has been found between reduced neurocognitive function and reduced serum BDNF levels (139). BDNF interacts with dopaminergic, glutamatergic, and serotonergic systems in the brain, see Nurjono et al. (138) for a recent update. Regarding BDNF as a drug target, FGAs have been reported to lower peripheral BDNF, whereas SGAs either have no effect or increase BDNF, while antidepressants have been shown to increase BDNF (140).

## **STRUCTURAL AND FUNCTIONAL CORRELATES OF INFLAMMATORY DISTURBANCES IN PATIENTS WITH SCHIZOPHRENIA**

Structural brain alterations associated with immune-related genetic polymorphisms were recently reviewed by Fineberg and Ellman (107). The review identified associations between variability in the IL-1 gene complex and white and gray matter volume reductions in temporal and frontal brain regions, together with enlarged ventricles in patients with schizophrenia. Furthermore, this review discussed associations of IL-1 gene variations with decreased dorsolateral prefrontal function during different neurocognitive tasks (107). In another review, Meyer (105) found evidence for a positive correlation between the severity of cognitive deficits and the peripheral levels of IL-1β, IL-6, TNF-α, CRP, and S100B. While the correlation of the inflammatory disturbances with specific cognitive effects is not clarified in patients with schizophrenia, findings from animal studies and studies in non-schizophrenic individuals reveal associations of inflammatory disturbances with executive function, sustained attention, and working memory according to the review by Meyer et al. (105).

## **ANTI-INFLAMMATORY EFFECTS OF ANTIPSYCHOTIC DRUGS**

Monji et al. (141) recently reviewed studies (mostly *in vitro* and animal studies) concerning the capability of antipsychotic drugs to influence activated microglia. The review identified several studies showing that antipsychotic drugs significantly reduce the secretion of TNF-α, nitric oxide, IL-1β, and IL-2 from activated microglia. Furthermore, some drugs were found to have stronger inhibitory effects, e.g., risperidone, which inhibited the secretion of several cytokines from activated microglia more than haloperidol, and indications of clozapine-specific effects have also been reported. Moreover, in the context of the kynurenine pathway, one animal study demonstrated that chronic treatment with antipsychotic drugs reduces brain KYNA (142). Meyer et al. (94) reviewed the influence of antipsychotic treatment on peripheral cytokines in clinical populations and found that antipsychotics reduced the level of proinflammatory cytokines (IL-1β, IL-6, sIL-6R, TNF-α) while increasing peripheral production of anti-inflammatory substances such as (sIL-1RA, sIL-2R and IL-10). Here, SGAs seemed to have the most pronounced effects and a possible relation between the ability to normalize immune function and clinical effects were identified. Brain imaging studies investigating antipsychotic drug treatment in brain imaging paradigms capable of showing effects on neuroinflammation (e.g., PET/"free water" MRI) would truly improve this field of research, but we have not been able to identify any such studies.

## **THE EFFECTS OF ANTI-INFLAMMATORY/IMMUNE-MODULATING DRUGS IN THE TREATMENT OF SCHIZOPHRENIA Non-steroid anti-inflammatory drugs**

Based on the notion of activated immune responses in the brain and/or peripheral tissues, a few clinical studies have evaluated the effects of non-steroid anti-inflammatory drugs (NSAIDs) in schizophrenia. The pooled effects size in reducing total Positive and Negative Syndrome Scale (PANSS) score according to a recent meta-analysis of the effect of add-on treatment with NSAIDs [eight studies, two with aspirin (*n* = 270), six with celecoxib (*n* = 504)] to different antipsychotics showed non-significant benefit over placebo (143). Treatment with aspirin alone was superior to placebo, while treatment with celecoxib was not; these conclusions were also drawn in a meta-analysis of Sommer et al. (144). When the studies were categorized according to disease phase, the pooled analyses of all eight studies revealed positive effects of NSAIDs vs. placebo for first-episode patients. The authors concluded that the present data do not lend significant support to NSAIDs as add-on therapy to antipsychotic treatment for patients with schizophrenia. Another review agreed that indiscriminately suppressing inflammation seems not to be the optimal way of treating immunological disturbances associated with schizophrenia, and that treatment should aim at reversing glial loss, upregulating beneficial microglial activation and proliferation (MAP) together with other neuroprotective measures, while ideally downregulating harmful aspects of MAP (135). At present, studies more precisely aiming to elucidate the putative pro-cognitive effects of treatment with NSAIDs in patients with schizophrenia are lacking, although non-significant effects were demonstrated in a small study (145).

## **Erythropoietin**

Erythropoietin (EPO) is a cytokine with several identified functions in the brain (146). In a 12-week, placebo-controlled study (*n* = 39) with stable schizophrenia patients, treatment with recombinant human EPO (40000 IU/week) improved neurocognitive function compared to placebo in the domains of delayed memory, language-semantic fluency, attention and perseverative errors (147). An evaluation of gray matter development in the same study showed that gray matter loss was significantly less pronounced in the group treated with EPO (148). Animal studies have supported a pro-cognitive effect of EPO treatment (146). A systematic review identified multiple putative beneficial brain targets of EPO treatment, e.g., modulation of inflammation, neuroprotection, neurotransmission regulation, effects on BBB permeability, but the use of EPO can be limited by adverse effects such as thrombosis and cancer (149). EPO has also been found to improve cognitive function in studies of patients with Parkinson's disease (150) and multiple sclerosis (151).

## **Minocycline**

The tetracycline antibiotic drug minocycline has also been tested as an add-on treatment for patients with schizophrenia, based on evidence from animal studies that showed minocycline to have multiple actions including inhibition of microglial activation and the attenuation of adaptive immunity through the reduction of activity and expression of matrix metalloproteinases (MMPs). MMPs enable T-cells to migrate through the BBB by altering its permeability, a process associated with myelin degradation (152). In a placebo-controlled, 6-month trial in patients with schizophrenia (54 patients receiving 200 mg minocycline and 26 patients receiving placebo), cognitive function (and negative symptoms) improved significantly in the minocycline-treated group, particularly in the domains of working memory, cognitive flexibility, and planning (153). Chaudry et al. (154) reported the results of a 1-year RCT of 155 early schizophrenia patients performed in Brazil and Pakistan, where minocycline reduced negative symptoms (adjusted difference 95% confidence interval 1.55–5.51 at the PANSS negative subscale), but no significant effects of treatment were found in tests for cognitive function. The meta-analysis of Sommer et al. (144) identified four RCT studies with *n* = 182 patients given minocycline, and *n* = 166 on placebo, the effects on cognition were not reported, but no significant effects on overall symptom severity were identified.

## **Additional drugs with anti-inflammatory action as add-on treatments for patients with schizophrenia**

The meta-analysis by Sommer et al. (144) also investigated the effect of additional anti-inflammatory treatments. The outcome measure was the mean change in PANSS (155) or the Brief Psychiatric Rating Scale (156), and significant improvements of treatment with *n*-acetylcysteine (*n* = 140), aspirin (*n* = 270), and estrogens (*n* = 262) were identified, while studies with davunetide and PUFAs (eicosapentanoic acid and docosahexanoic acids) did not show significant effects.

## **SUMMING UP PRO-COGNITIVE EFFECTS OF IMMUNOMODULATING/ANTI-INFLAMMATORY DRUG TREATMENT FOR PATIENTS WITH SCHIZOPHRENIA: CRITICAL METHODOLOGICAL PROBLEMS REMAIN**

The involvement of inflammatory/immunological factors in the pathophysiology of a large portion of patients with schizophrenia is well established through different lines of research, and data point to several links between these factors and cognitive dysfunction. Presently, many drugs with anti-inflammatory or immunomodulatory action have been tested (as add-on to antipsychotic treatment) for patients with schizophrenia, and while some have shown positive effects for global schizophrenia symptoms, the effects on cognition were not possible to assess in the recent meta-analysis by Sommer et al. (144), as only 5 of 26 included studies reported data on cognitive tests and the heterogeneity of the cognitive tasks made it impossible to quantitatively review the effects. Furthermore, as highlighted by Sommer et al. the drugs tested that were shown to improve the symptoms of schizophrenia (*N*-acetylcysteine, estrogens, aspirin) have broad mechanisms of action, and it is not at all demonstrated that the improved symptoms are caused by anti-inflammatory actions. A further critical question to address in future drug research will be to establish reliable brain imaging methods to quantitatively assess neuroinflammation, as the levels of peripheral cytokines are only an indirect measure of brain neuroinflammation. PET studies with new ligands that specifically target activated microglia and the "free-water" MRI scanning seem promising, but extensive development remains before these methods are ready as outcome measures in drug discovery studies.

## **SUMMARY AND CONVERGENCE**

At present, pharmacological treatment of schizophrenia relies on antipsychotic drugs, which predominantly relieve positive psychotic symptoms and have little effect on the cognitive dysfunctions restricting the functional level of patients. However, recent developments in brain imaging, including novel specific PET ligands and MRI paradigms specifically investigating myelin integrity, neurotransmitter levels, and neuroinflammation, facilitate the unraveling of mechanisms underlying cognitive dysfunction in schizophrenia. Accumulating evidence supports a combined model including all these areas of pathology underlying the cognitive deficits in schizophrenia, schematically exemplified by **Figure 2** for the interplay between prefrontal cortex, striatum, and brain stem. This joint model represents several potential procognitive drug targets at the level of neurotransmitters, lipids, and inflammatory markers, however, no current drugs aimed at these targets are so far definitely proven to be clinically useful as pro-cognitive drugs for patients with schizophrenia. Studies concerning immunomodulatory drugs, including EPO, minocycline, celecoxib, aspirin, estrogens, *n*-acetylcysteine have shown promising results in single studies, but have not yielded straightforward support for the use in clinical practice to enhance cognition.

## **FUTURE PERSPECTIVES**

Drugs that would reduce brain levels of KYNA (e.g., KAT inhibitors) are expected to improve cognitive dysfunctions in schizophrenia based on animal studies, but remain to be developed for clinical use. The possibility of pro-cognitive effects of drugs that have been used to treat other medical conditions for years, e.g., EPO, minocycline, and aspirin is attractive, also because the safety profile of many of these drugs are well known in neurological patients populations, reducing the costs of introducing the drugs as cognitive enhancers for schizophrenia patients. Several critical questions currently constraining progress remain to be solved, however. Firstly, uniformly accepted and standardized cognitive assessments for repeated measures must be agreed on for pro-cognitive drug studies in schizophrenia patients. Secondly, the refinement of brain imaging methods unto sufficient reliability and specificity (e.g., measuring neuroinflammation and myelin integrity) for use as outcome measures in pro-cognitive drug is essential for studies of putative cognition enhancers in humans. Thirdly, and at least of prominent importance in the field of further developments of pro-cognitive glutamatergic substances are the problems of translating findings in animal studies to human drug treatment.

However, even though the process of developing novel pharmacological agents is highly sophisticated, identifying novel treatment targets through thorough clinical and preclinical studies remains pivotal in the continued search for improved treatment of schizophrenia. The convergence of a disordered inflammatory system, disturbances in myelination/oligodendrocyte function,

and transmitter dysfunctions in schizophrenia open new paths for pro-cognitive drug discovery with putative large functional improvements to be achieved for patients with schizophrenia.

## **ACKNOWLEDGMENTS**

Parts of the research were funded by grants to Erik Johnsen from the Research Council of Norway, project no. 213727, and several grants from Helse Vest. Parts of the research were funded by an ERC Advanced Grant and by the NORMENT SFF from the Research Council of Norway to Kenneth Hugdahl and Silje Skrede.

## **REFERENCES**


early-phase psychosis. *Schizophr Bull* (2013) **39**:830–8. doi:10.1093/schbul/ sbs089


interleukin-1beta and NMDA stimulation. *J Neuroinflammation* (2011) **8**:14. doi:10.1186/1742-2094-8-14


by recombinant human erythropoietin. *Mol Psychiatry* (2007) **12**:206–20. doi:10.1038/sj.mp.4001907


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 October 2013; accepted: 20 January 2014; published online: 04 February 2014.*

*Citation: Kroken RA, Løberg E-M, Drønen T, Grüner R, Hugdahl K, Kompus K, Skrede S and Johnsen E (2014) A critical review of pro-cognitive drug targets in psychosis: convergence on myelination and inflammation. Front. Psychiatry 5:11. doi: 10.3389/fpsyt.2014.00011*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2014 Kroken, Løberg , Drønen, Grüner, Hugdahl, Kompus, Skrede and Johnsen. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

REVIEW ARTICLE published: 10 June 2013 doi: 10.3389/fpsyt.2013.00052

## **Shaheen E. Lakhan1,2\*, Mario Caro<sup>2</sup> and Norell Hadzimichalis <sup>1</sup>**

<sup>1</sup> Biosciences Department, Global Neuroscience Initiative Foundation, Beverly Hills, CA, USA <sup>2</sup> Neurological Institute, Cleveland Clinic, Cleveland, OH, USA

#### **Edited by:**

André Schmidt, University of Basel, Switzerland; University Hospital of Basel, Switzerland

#### **Reviewed by:**

Paul C. Fletcher, University of Cambridge, UK Bernat Kocsis, Harvard Medical School, USA Henry H. Holcomb, University of Maryland School of Medicine, USA

#### **\*Correspondence:**

Shaheen E. Lakhan, Neurological Institute, Cleveland Clinic, 9500 Euclid Avenue, S100A, Cleveland, OH 44195, USA e-mail: slakhan@gnif.org

N-Methyl-d-aspartate (NMDA) receptors play a variety of physiologic roles and their proper signaling is essential for cellular homeostasis. Any disruption in this pathway, leading to either enhanced or decreased activity, may result in the manifestation of neuropsychiatric pathologies such as schizophrenia, mood disorders, substance induced psychosis, Huntington's disease, Alzheimer's disease, and neuropsychiatric systemic lupus erythematosus. Here, we explore the notion that the overlap in activity of at least one biochemical pathway, the NMDA receptor pathway, may be the link to understanding the overlap in psychotic symptoms between diseases. This review intends to present a broad overview of those neuropsychiatric disorders for which alternations in NMDA receptor activity is prominent thus suggesting that continued direction of pharmaceutical intervention to this pathway may present a viable option for managing symptoms.

**Keywords: NMDA, psychiatry, schizophrenia, mood disorders, substance induced psychosis, Huntington's disease, Alzheimer's disease, neuropsychiatric systemic lupus erythematosus**

#### **INTRODUCTION**

Diagnosis of psychiatric disorders is done clinically by focusing on observable symptoms and behaviors rather than on underlying psychodynamic processes or on the results of laboratory or imaging testing. Understanding the descriptive symptoms for mental disorders is vital in order to properly diagnose each psychiatric disease. The instruments most commonly used to diagnose and categorize mental illnesses, the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV-TR) and the World Health Organization's International Statistical Classification of diseases and Related Health Problems (ICD-10), focus on objective observations, without offering any discussion on the etiologies for these diseases. From a molecular, biochemical, and ultimately therapeutic perspective, it is equally as essential to characterize the role of various receptors, ligands, and neurotransmitters that when modified alter the manifestation of these symptoms.

Psychotic symptoms can be present in primary psychiatric disorders (schizophrenia, schizoaffective disorder, mood disorders, substance intoxication) and in psychiatric disorders that occur due to a medical condition such as Huntington's disease (HD), Alzheimer's disease (AD), and systemic lupus erythematosus (SLE). Several neurotransmitters have been linked to the development of psychotic symptoms, with dopamine and serotonin being the most widely studied due to the treatment effect of blocking certain subtypes of these receptors with antipsychotics. Unfortunately, long-term treatment with typical or atypical antipsychotics is limited due to side effects profile and high rates of discontinuation (Lieberman et al., 2005). *N*-Methyl-d-aspartate (NMDA) receptors have also been implicated in the development of psychotic symptoms and are a potential target for the development of novel treatments in the future.

#### **NMDA RECEPTORS IN NEUROPSYCHIATRIC DISORDERS**

*N*-Methyl-d-aspartate receptors are a class of glutamate receptor that when activated, mediate excitatory neurotransmission via passage of non-selective cations, including Ca2+, through the channel. They are abundantly and ubiquitously located throughout the brain and are understood to play a key role in synaptic plasticity and memory function (Stephenson et al., 2008; Li and Tsien, 2009). They are activated by binding the co-agonists glutamate and glycine, in addition to exposure to a positive change in membrane potential across the cell. Functional NMDA receptor heterotetramers are generally formed through a "dimer of dimers" mechanism and are conventionally made up of two glycine binding NR1 subunits and two glutamate binding NR2 subunits (**Figure 1**) (Dongen, 2009). While the NR1 subunit is considered essential to the formation of the complex, data indicates that the NR2 subunits may be interchangeable with either one or two NR3 subunits (Schuler et al., 2008).

The NMDA receptor is known to play an integral role in the regulation of signal transduction in multiple regions of the brain. Accordingly, any homeostatic dysfunction of NMDA receptor activity has the potential to result in a variety of pathologies. Previous authors have found a particularly high concentration of post-synaptic NMDA receptors in limbic structures (Kretschmer, 1999;Tsapakis and Travis, 2002), which is of uttermost importance in the pathogenesis of many psychiatric disorders.

**Abbreviations:** AA, arachidonic acid; Aβ, amyloid beta; AD, Alzheimer's disease; AMPA, α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid; BPAD, bipolar affective disorder; CONSIST, Cognitive and Negative Symptoms in Schizophrenia Trial; DSM, Diagnostic and Statistical Manual of Mental Disorders; DSM-IV-TR, Diagnostic and Statistical Manual of Mental Disorders, fourth-edition, text-revision; GABA, in γ-aminobutyric acid; HD, Huntington's disease; NMDA, *N*-Methyl-daspartate; SIP, substance induced psychosis; NPSLE, neuropsychiatric systemic lupus erythematosus; SLE, systemic lupus erythematosus.

Elucidation of the structural and mechanistic properties of NMDA receptors suggests a potential widespread use as a target for pharmacological intervention. Structurally, NMDA receptors provide a number of potentially viable sites for drug interaction. Initially, efforts focused on the development of broad-spectrum antagonists and ion channel blockers. However, more recent subtype selective targeting and strategic drug development aimed at these sites has resulted in the production of a variety of NMDA receptor agonists and antagonists that better address disease conditions ranging from mental disorders to pain management (Kemp and McKernan, 2002; Paoletti and Neyton, 2007).

*N*-Methyl-d-aspartate receptor antagonists describe a category of compounds that functionally inhibit or deactivate NMDA receptor activity. They can act broadly or specifically at various sites on the NMDA receptor including the agonist binding domains, allosteric sites, and the ion channel pore. This category of compounds has a potential use in any disease that results from glutamate-induced excitotoxicity ranging from cerebral ischemia and epilepsy to neurodegenerative disorders and neuropathic pain. More recent developments focus on subunit specific compounds, including NR2B-selective antagonists, and have resulted in minimization of side effects with improved therapeutic efficacy (Kemp and McKernan, 2002; Gogas, 2006; Paoletti and Neyton, 2007). NMDA receptor agonists, on the other hand, refer to the category of NMDA receptor targeted compounds that can enhance receptor activity. It is interesting to note, that some mental disorders can be treated with both NMDA receptor antagonists and agonists. These biphasic disorders in regards to NMDA receptor activity may require customized treatment protocols depending on the stage of the disease.

This review specifically focuses on neuropsychiatric disorders that manifest with psychotic symptoms and have the molecular commonality of NDMA receptor activity dysfunction. Here, we present a broad overview of the alterations in NMDA receptor activity in schizophrenia, bipolar disorder, HD, AD, substance induced psychosis, and neuropsychiatric systemic lupus erythematosus (NPSLE). We further emphasize the importance of continued pharmaceutical attention on this pathway as a target for the development of safer and more effective therapies.

## **SCHIZOPHRENIA**

Schizophrenia is a psychotic disorder characterized by abnormalities in thought processing and content, presence of delusions and/or hallucinations as well as the presence of negative symptoms. Schizophrenia is a worldwide public health issue that affects approximately 1% of the adult population. While some symptoms and molecular pathways overlap with other related disorders including bipolar disorder, schizophrenia is unique in its precise presentation of positive, negative, and cognitive symptoms. For many years the *dopamine hypothesis* prevailed to explain symptoms associated with schizophrenia (Meltzer and Stahl, 1976). Thus, conventional antipsychotic drugs used to treat these patients would act by interfering with and inhibiting dopamine neurotransmission. While individuals using those drugs experienced significantly reduced positive psychotic symptoms, the effects were not optimal and often resulted in adverse side effects (Levinson, 1991; Fleischhacker, 1995). More recent elucidation of other neurotransmitter pathways involved in the manifestation of schizophrenia, such as serotonin and glutamate has provided novel treatment targets aimed at more effectively inhibiting both positive and negative symptoms (Fitzgerald et al., 1995; Iqbal and van Praag, 1995; Lindenmayer, 1995). Atypical antipsychotics act by blocking of serotonin type 2 and D<sup>2</sup> receptors, initially, these medications were thought to be superior to typical antipsychotics due to less extrapyramidal side effects, however, recent evidence shows that there is also high discontinuation rate of atypical antipsychotics (Lieberman et al., 2005). Low treatment adherence has led investigators to study different and novel mechanisms for the development of novel medications for the management of psychosis.

Initial data in support of a more novel *glutamate hypofunction hypothesis* of schizophrenia arose from reports of low cerebrospinal fluid glutamate levels in patients with schizophrenia (Kim et al.,1980). Further studies corroborate this theory and indicate that administration of NMDA receptor antagonists including phencyclidine (PCP) and ketamine to patients with schizophrenia resulted in worsening of psychotic symptoms (Luby et al., 1959; Lahti et al., 1995; Gilmour et al., 2012). Additional studies reveal that administration of similar antagonists to healthy patients replicates symptoms of schizophrenia including positive, negative, and cognitive symptoms (Krystal et al., 1994; Gilmour et al., 2012).

Building on these data, more recent pharmacological approaches aimed at treating schizophrenia focus on the use of NMDA receptor agonists (Kemp and McKernan, 2002). However, direct activation of the receptor and reported excitotoxicity suggests the need to more specifically explore the glycine binding site as a potentially safer indirect target for treating glutamate hypofunction disorders (Lechner, 2006; Paoletti and Neyton, 2007). A number of studies are currently exploring this mechanism as a means of treating symptoms with minimal side effects. Both naturally occurring and synthetic glycine agonists including glycine, d-serine, and d-cycloserine are showing great promise for the treatment of positive and negative symptoms of schizophrenia (Coyle et al., 2003; Millan, 2005; Long et al., 2006). Following a similar mechanistic approach of indirectly targeting the glycine binding site, Glycine transport 1 (GLY-T1) inhibitors are being explored in order to modulate NMDA receptor function. The GLY-T<sup>1</sup> reuptake pump functions to remove excess glycine in the synaptic cleft and thus inhibitors are being actively explored to increase glycine at the synapse. Animal data from transgenic mice suggest that the GLY-T<sup>1</sup> inhibitor SSR103800 shows efficacy, decreased side effects, and suggests a use for GLY-T<sup>1</sup> inhibitor as an adjunct to conventional therapy for schizophrenia (Boulay et al., 2010).

One of the largest trials performed so far studying the effect of increased glutamate transmission is the Cognitive and Negative Symptoms in Schizophrenia Trial (CONSIST) (Buchanan et al., 2007). The trial's primary aim was to determine if coadministration of glycine (co-transmitter with glutamate at the NMDA receptor) or d-cycloserine (partial agonist at NMDA receptor) was associated with an improvement in cognitive impairment or in the negative symptoms of schizophrenia. During the trial, there was no improvement in the above mentioned symptoms with the experimental treatments. However, despite negative findings in this trial, there is clear evidence that NMDA receptor dysfunction is implicated in schizophrenia, and it is still an important research area for the development of future treatments.

Additional clinical evidence demonstrates that the GLY-T<sup>1</sup> inhibitor Org 25935 has been explored for its antipsychotic properties. Preliminary human data indicate that it can effectively counteract the effects of the NMDA receptor antagonist, ketamine (D'Souza et al., 2012). Promising Phase II clinical data corroborate these results and further suggest that the GLY-T<sup>1</sup> inhibitor RG1678 was a safe and effective compound for treating the negative symptoms of schizophrenia (Pinard et al., 2010).

The dopamine hypothesis and the glutamate hypofunction hypothesis of schizophrenia each separately explain specific aspects of the disease condition. However, some researchers argue that focusing on only one molecular pathway to characterize the complicated etiology of the disease is likely to narrow our understanding. In fact, some additional theories provide evidence that hypofunction of NMDA receptors results in dopaminergic abnormalities. Interestingly, this synergy between the two pathways best explains the positive, negative, and cognitive symptoms associated with the disease (Schwartz et al., 2012). Still, despite not agreeing on a molecular mechanism to explain the manifestation of schizophrenia, scientists do agree that NMDA receptor dysfunction plays an integral role and should continue to be studied as a therapeutic target.

## **MOOD DISORDERS**

Mood is described as the internal feeling tone that influences the way an individual perceives himself and the environment. The most widely studied mood disorders are major depressive disorder, and bipolar affective disorder (BPAD), the latter one is characterized by alternation between manic and depressive episodes. Psychotic symptoms can be present in severe episodes of depression, mania, or during mixed states (Stahl, 2008). Mood disorders were initially thought to be caused by alterations in the levels of norepinephrine and serotonin, but this theory has not been able to completely explain the cause of this complex illness, this is evidenced by the results from a large-randomized clinical trial where only two-thirds of the patients who were treated with several treatment courses of different antidepressants experienced remission of their depression (Gaynes et al., 2008). Other theories include the alteration of hormonal regulation (hypothalamic-pituitary and thyroid axis dysfunction), alterations in γ-aminobutyric acid (GABA) receptors, and alterations in the arachidonic acid pathway and NMDA receptors.

The *arachidonic acid (AA) hypothesis* of BPAD represents a wellstudied mechanism used to explain the pathophysiology of the disease. Support for this hypothesis gained popularity based on data showing that BPAD is associated with an upregulation in the AA cascade. These data were validated by reports that inhibitors of this cascade, including lithium, were successful mood stabilizers (Rapoport and Bosetti, 2002).

Similar to other mental disorders with psychotic manifestations, NMDA receptor activity seems to be intimately involved in the complicated web of pathways leading to the pathogenesis of mood disorders. The NMDA receptor antagonist, ketamine, has been shown to rapidly improve symptoms of depression (Zarate et al., 2006). Evidence from rat models has shown that the rapid improvement in depressive symptoms observed with ketamine is caused by an increase in the activation of the mammalian target of rapamycin (mTOR) pathway causing an increase in the number of neuronal synapses – opposite to changes seen during stress (Li et al., 2010). Other authors have described that ketamine also produces a disinhibition of GABAergic neurons, leading to increased presynaptic levels of glutamate; which then interacts with α-amino-3-hydroxy-5-methyl-4 isoxazolepropionic acid (AMPA) receptors, as the NMDA receptors are blocked by ketamine. This increased ratio of AMPA to NMDA activation and may be implicated in the antidepressant effects of ketamine (Sanacora et al., 2008; Andreasen et al., 2013).

Reports indicate there are elevated levels of glutamate in the left dorsolateral prefrontal cortex of adults with BPAD during the manic phase (Michael et al., 2003). More recent studies show that treatment with the NMDA receptor antagonist, MK-801, results in inhibition of the downstream AA pathway (Basselin et al., 2006). In addition, gene linkage analysis confirmed a role specifically for the NR2B subunit of the NMDA receptor in BPAD (Martucci et al., 2006). Together, these studies suggest that NMDA receptor activity plays a critical role in mood disorders likely through the in AA pathway.

## **HUNTINGTON'S DISEASE**

Huntington's disease is a progressive genetic neurodegenerative disorder that results in decreased muscle coordination and cognitive ability. It is typically diagnosed between 30 and 40 years of age and often results in death within 15–20 years. In many cases, patients also present with psychiatric symptoms including some

similar to those observed in patients with schizophrenia. On a molecular level, the underlying cause of HD is considered to be disruptions in the gene that encode the protein huntingtin, whose altered function ultimately leads to selective neuronal cell death (Gusella et al., 1983).

Some data suggest that the formation of nuclear protein aggregates plays a role in neuronal cell death associated with the disease (Carmichael et al., 2000). Others explore a variety of also likely contributing pathways including oxidative stress and impaired mitochondrial function (Davies and Ramsden, 2001). However, considering the focus of this review and the potential for manifestation of psychotic symptoms in HD, it is interesting to specifically note the role of the NMDA receptor pathway.

Early animal studies indicate that injections with kainic or quinolinic acids produce lesions similar to those observed in HD thus promoting the idea that NMDA receptor-mediated excitotoxicity may also contribute to the etiology of the disease (Coyle and Schwarcz, 1976; Beal et al., 1986). Related studies corroborate these early findings and report hyperactive NMDA receptors present in transgenic HD mice (Levine et al., 1999). Prolonged receptor activation results in excitotoxicity and cell death characteristic of many neurodegenerative disorders including HD. Further elucidation of this pathway explores a role for post-synaptic density protein 95 (PSD-95), a well-characterized scaffolding protein that binds to multiple cytoplasmic proteins including normal huntingtin and separately to the NR2 subunit of NMDA receptors. Once bound it causes clustering of receptors in the post-synaptic membrane and results in physiologic inhibition of NMDA receptor activity. In the case of HD, the presence of mutant huntingtin protein results in a disruption of PSD-95 binding to NMDA receptors, receptor hypersensitivity and resulting excitotoxicity, and ultimately increased neuronal cell death consistent with HD (Davies and Ramsden, 2001; Sun et al., 2001). Similar to schizophrenia and other complicated mental disorders, data point to the possibility of multiple and potentially parallel pathways giving rise to the variety of documented symptoms in HD. Nonetheless, the NMDA receptor pathway still remains a primary target for therapeutic intervention.

## **ALZHEIMER'S DISEASE**

Alzheimer's disease describes a type of dementia that results in serious and progressive cognitive impairment. Symptoms range from commonly observed memory loss to psychotic manifestations including delusions. Currently, clinicians diagnose patients based on DSM-outlined criteria; however, a definitive diagnosis can only be made post-mortem.

For many years, the *amyloid hypothesis* of AD was the dominating model thought to explain the pathophysiology of the disorder. This hypothesis gained early popularity based on data that identified amyloid β-peptide (Aβ) as an integral component in the plaques observed post-mortem in AD (Masters et al., 1985). Additional reports confirm the central role of the Aβ protein and suggest that its accumulation initiates downstream diseases symptoms and molecular manifestations including alteration in the tau protein and the formation of characteristic neurofibrillary tangles (Hardy et al., 1998; Hardy and Selkoe, 2002).

As mentioned above,NMDA receptors are well-studied for their crucial role in learning and memory, key areas that are affected in the manifestation of AD. Similar to other mental disorders, AD is likely the result of dysfunction in multiple neurotransmitter pathways. In fact, data suggest an overlap in pathways related to NMDA receptor activation and production of characteristic biochemical and symptomatic changes that occur in AD. However, the exact nature of that overlap is not yet determined. Consistent with the idea that accumulation of Aβ protein initiates the pathological cascade, some studies suggest that the NMDA receptor may function indirectly as a receptor for the Aβ protein (Cisse et al., 2011; Malinow, 2012). A recent report from Cisse et al., indicates that the Aβ protein binds directly to the tyrosine kinase receptor, EphB2, a known regulator of NMDA receptor function (Cisse et al., 2011; Nolt et al., 2011). This results in degradation of EphB2 and a downstream reduction in NMDA receptor-mediated long term potentiation (Cisse et al., 2011). Other studies suggest that enhanced NMDA receptor activity results in increased processing of the amyloid beta precursor protein thus producing an increase in the AD characteristic Aβ protein. This in turn results in a decrease in excitatory synaptic transmission and may contribute to cognitive effects observed in early stages of AD (Gordon-Krajcer et al., 2002; Butterfield and Pocernich, 2003). With regards to the glutamatergic pathway, AD is unique in that receptor dysfunction is dependent on stage of the disease. In contrast with theories describing the pathogenesis of early stage AD, late stage pathogenesis is thought to be attributed to loss of NMDA receptors and resulting hypoactivity (Butterfield and Pocernich, 2003).

Precise mechanism aside, researchers agree that in conjunction with Aβ protein accumulation, the NMDA receptor pathway is integral in the manifestation of AD pathophysiology. Thus data point to the notion of a strategic combination therapy regimen to most effectively treat the condition.

## **SUBSTANCE INDUCED PSYCHOSIS**

Substance induced psychosis (SIP) disorder represents another defined mental disorder that results in delusions or hallucinations that are linked to use or withdrawal from a variety of defined compounds. Interestingly, NMDA receptor antagonists PCP and ketamine, previously shown to produce an elevation of psychotic symptoms in patients with schizophrenia, are among the drugs reported to result in SIP (Luby et al., 1959; Lahti et al., 1995; Gilmour et al., 2012). Furthermore, there is evidence of potential for conversion from SIP to the more clearly defined psychotic disorder, schizophrenia (Rabe-Jablonska et al., 2012; Niemi-Pynttari et al., 2013). Although, the data in this space are mostly descriptive and the broad description of SIP gives rise to a poorly defined molecular mechanism, the likely role of NMDA receptors in the psychotic manifestation of this disease should not be left unnoted.

## **NEUROPSYCHIATRIC SYSTEMIC LUPUS ERYTHEMATOSUS**

Systemic lupus erythematosus is an autoimmune disorder that results in inflammation and threatens tissue damage to a number of different organ systems including the skin, heart, joints, lungs, and nervous system. Neuropsychiatric systemic lupus erythematosus is a poorly understood yet prevalent complication of SLE. It results when SLE specifically affects the central or



This table presents select drugs that, at minimum, have been approved for initial clinical testing (NLM, 2013). This approval provides evidence of the important of the NMDA receptor pathway in the potential treatment of various mental disorders.

peripheral nervous systems and patients manifest additional neurological symptoms such as psychosis, mood disorders, or cognitive dysfunction (Popescu and Kao, 2011). While the mechanism of NMDA receptor activity involvement is unique compared to the previously mentioned conditions, it is important to mention as it represents additional, albeit less clearly defined, evidence of an integral role for the NMDA receptor pathway in the pathogenesis of mental disorders.

Autoantibodies represent a type of antibody specifically targeted to endogenous proteins in an individual. They are understood to cause many of the symptoms associated with autoimmune disorders, including SLE. Previous data indicate that the cognitive manifestations of NPSLE may be related to medication and cardiovascular disease. However, additional data suggest that autoantibodies directed against various NMDA receptor subunits may also play a role in pathogenesis.

### **REFERENCES**

Andreasen, J. T., Gynther, M., Rygaard, A., Bøgelund, T., Nielsen, S. D., Clausen, R. P., et al. (2013). Does increasing the ratio of AMPAto-NMDA receptor mediated neurotransmission engender

antidepressant action? Studies in the mouse forced swim and tail suspension tests. *Neurosci. Lett.* doi:10.1016/j.neulet.2013.04. 045

Basselin, M., Chang, L., Bell, J. M., and Rapoport, S. I. (2006).

Studies show that NMDA receptor autoantibodies are prevalent in the manifestation of both mood symptoms and psychotic behaviors observed in patients with NPSLE (Mak et al., 2009). More specifically, autoantibodies to the NR2A and NR2B subunits of the NMDA receptor mediate neuronal death *in vitro* following breakdown of the blood brain barrier (Shefner et al., 1991; Katz et al., 1994; Gaynor et al., 1997; Kowal et al., 2004). Interestingly, the NMDA receptor antagonist memantine has been show to protect neurons from autoantibody-mediated cell injury, thus indicating that these autoantibodies may be acting as functional agonists (Lipton and Rosenberg, 1994; Kowal et al., 2004, 2006). Translating these results to human clinical trials, however, did not show similar efficacy (Petri et al., 2011). While these data may initially seem discouraging, there were some reported study limitations, including the use of physician confirmed, self-reported cognitive impairment. The authors indicate that memantine may still ultimately provide a therapeutic benefit in preventing the worsening of cognitive symptoms. Together, these data suggest that NMDA receptor antagonists represent a much underrated potential therapy for NPSLE.

## **CONCLUSION**

The role of the NMDA receptor pathway in a broad range of diseases and disorders is well established. However, how that "role" plays into the larger picture of disease manifestation is not entirely clear. Specifically, in the case of complicated mental disorders that manifest in a variety of symptoms including psychosis, the NMDA receptor pathway comes into play. Alterations in its activity lead to either hypofunction or hyperfunction and related excitotoxicity. To further complicate the matter, changes in NMDA receptor activity may be dependent on the stage of disease. For example, as indicated above, symptoms of AD can be initially attributed to glutamate-induced excitotoxicity. Late stage pathogenesis however, is thought to be a result, in part, of NMDA receptor hypoactivity.

**Table 1** presents select NMDA receptor-targeted drugs that have been used as a monotherapy in clinical trials to alleviate symptoms of schizophrenia, BPAD, HD, AD, and NPSLE. Still, many other drugs and various drug combinations are also being explored in preclinical studies. While research has not yet clearly elucidated how these changes in receptor activity tie into the overlapping symptoms between mental disorders, we can be certain that maintaining receptor homeostasis is integral for symptom management. Combination therapy directed at selective NMDA receptor antagonists and enhancers in addition to drugs that target other disease-specific pathways may be the key to controlling the symptoms of psychotic disorders. However, more thorough molecular comparisons between the various mental disorders with psychotic manifestations is still needed.

Chronic lithium chloride administration attenuates brain NMDA receptor-initiated signaling via arachidonic acid in unanesthetized rats. *Neuropsychopharmacology* 31, 1659–1674. doi:10.1038/sj.npp.1300920

Beal, M. F., Kowall, N. W., Ellison, D. W., Mazurek, M. F., Swartz, K. J., and Martin, J. B. (1986). Replication of the neurochemical characteristics of Huntington's disease by quinolinic acid. *Nature* 321, 168–171. doi:10.1038/321168a0


and subcortical glutamate receptor subunit expression by antipsychotic drugs. *J. Neurosci.* 15, 2453–2461.


5(Suppl.), 11–23. doi:10.1016/0924- 977X(95)00027-M


knockin mouse models of Huntington's disease. *J. Neurosci. Res.* 58, 515–532. doi:10.1002/(SICI)1097- 4547(19991115)58:4<515::AID-JNR5>3.0.CO;2-F


N., et al. (2006). N-methyl-Daspartate receptor NR2B subunit gene GRIN2B in schizophrenia and bipolar disorder: polymorphisms and mRNA levels. *Schizophr. Res.* 84, 214–221. doi:10.1016/j.schres.2006.02.001

	- methylethoxy)phenyl]methanone (RG1678), a promising novel medicine to treat schizophrenia. *J. Med. Chem.* 53, 4603–4614. doi:10.1021/jm100210p

Polyglutamine-expanded huntingtin promotes sensitization of N-methyl-D-aspartate receptors via post-synaptic density 95. *J. Biol. Chem.* 276, 24713–24718. doi:10.1074/jbc.M103501200


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 31 March 2013; accepted: 25 May 2013; published online: 10 June 2013.*

*Citation: Lakhan SE, Caro M and Hadzimichalis N (2013) NMDA receptor activity in neuropsychiatric disorders. Front. Psychiatry 4:52. doi: 10.3389/fpsyt.2013.00052*

*This article was submitted to Frontiers in Schizophrenia, a specialty of Frontiers in Psychiatry.*

*Copyright © 2013 Lakhan, Caro and Hadzimichalis. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in other forums, provided the original authors and source are credited and subject to any copyright notices concerning any third-party graphics etc.*

# Relationship between glutamate dysfunction and symptoms and cognitive function in psychosis

## **Kate Merritt \*, Philip McGuire and Alice Egerton**

Department of Psychosis Studies, Institute of Psychiatry, King's College London, London, UK

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

John H. Krystal, Yale School of Medicine, USA Anouk Marsman, Johns Hopkins University, USA

#### **\*Correspondence:**

Kate Merritt, Department of Psychosis Studies, Institute of Psychiatry, King's College London, De Crespigny Park, Denmark Hill, London SE5 8AF, UK e-mail: kate.merritt@kcl.ac.uk

The glutamate hypothesis of schizophrenia, proposed over two decades ago, originated following the observation that administration of drugs that block NMDA glutamate receptors, such as ketamine, could induce schizophrenia-like symptoms. Since then, this hypothesis has been extended to describe how glutamate abnormalities may disturb brain function and underpin psychotic symptoms and cognitive impairments.The glutamatergic system is now a major focus for the development of new compounds in schizophrenia. Relationships between regional brain glutamate function and symptom severity can be investigated using proton magnetic resonance spectroscopy (1H-MRS) to estimate levels of glutamatergic metabolites in vivo. Here we briefly review the 1H-MRS studies that have explored relationships between glutamatergic metabolites, symptoms, and cognitive function in clinical samples.While some of these studies suggest that more severe symptoms may be associated with elevated glutamatergic function in the anterior cingulate, studies in larger patient samples selected on the basis of symptom severity are required.

**Keywords: schizophrenia, psychosis, glutamate, NMDA, MRS, spectroscopy, imaging**

## **INTRODUCTION**

Accumulating evidence suggests that glutamatergic dysfunction may contribute to the pathogenesis of schizophrenia, and the symptoms and cognitive deficits associated with the disorder (1). The glutamatergic system presents an attractive therapeutic target, as dopaminergic antipsychotics have little effect on negative symptoms or cognitive impairment, yet these features are better predictors of social and functional outcome than positive symptoms (2, 3). Here we briefly review the existing evidence linking abnormal glutamatergic transmission to cognitive, negative, and positive symptoms of schizophrenia.

The observation that administration of antagonists at the *N*methyl-d-aspartate glutamate receptor complex (NMDAR) such as phencyclidine (PCP) or ketamine to healthy volunteers induces effects which resemble aspects of schizophrenia symptomatology forms a cornerstone of the glutamate hypothesis of schizophrenia (4–6). NMDAR antagonists also worsen positive, negative, and cognitive symptoms in patients with schizophrenia (7, 8). While dopamine-stimulating drugs such as amphetamine also produce positive "psychotic"-like effects, negative-type symptoms, and cognitive deficits are far more prominently elicited by ketamine than amphetamine administration (9). Neuroimaging studies indicate that these effects of ketamine are mediated by changes in activity in the frontal and cingulate cortices and the thalamus (10–13).

While pharmacological studies have provided evidence linking NMDAR dysfunction to these symptom domains, direct associations between glutamatergic function and symptom severity may be provided by neuroimaging studies. In a single photon emission tomography study using the NMDAR radiotracer <sup>123</sup>I-CNS-1261 in schizophrenia, the availability of NMDAR in the hippocampus was negatively associated with the severity of symptoms, especially negative symptoms (14). A ketamine challenge study in healthy volunteers using the same radiotracer linked regional NMDAR binding to the induction of negative (but not positive) symptoms, particularly in the thalamus (15). These approaches are currently limited by a lack of availability of suitable radiotracers (1).

An alternative is to use proton magnetic resonance spectroscopy (1H-MRS) to estimate the concentration of glutamatergic metabolites. MRI scanners with field strengths of 3 T or above can resolve glutamate, at least for the most part, from its metabolite glutamine (16). At lower field strengths glutamate and glutamine are reported in combination, as Glx. A limitation of 1H-MRS is that the glutamate concentration estimates are not specific to neuronal glutamate, and that changes in glutamate levels cannot be specifically attributed to altered neurotransmission over other metabolic processes (17). However the majority (∼80%) of glutamine synthesis reflects cycling of neurotransmitter glutamate (17), and clinical studies at 4 T in minimally treated patients with schizophrenia have reported higher Gln/Glu ratios (18) or higher glutamine levels (19) in the anterior cingulate cortex (ACC). This is consistent with increased glutamatergic neurotransmission, but could also result from a deficiency in the conversion of glutamine to glutamate. However the majority of 1H-MRS studies in psychosis have used field strengths <4 T and thereby were unable to accurately quantify glutamine concentrations. Thus, the results from these studies at lower field strengths cannot be specifically attributed to changes in glutamate neurotransmission. Nonetheless, increases in frontal glutamatergic neurotransmission in schizophrenia are broadly consistent with the NMDA receptor hypofunction hypothesis, as administration of NMDAR antagonists to rats increases glutamate release as detected by microdialysis (20), and 1H-MRS studies report elevated Gln/Glu ratios (21), or increased glutamate levels with no changes in glutamine (22). 1H-MRS studies with ketamine in man at 4 T reveal increased ACC glutamine (23), and at 3 T reveal increased ACC glutamate (24). One mechanism through which increases in glutamate release may occur is via NMDAR dysfunction on GABA-ergic interneurons, leading to the disinhibition of glutamatergic pyramidal cells, as described by the NMDAR hypofunction hypothesis (25).

Individual 1H-MRS glutamate studies in schizophrenia have produced some inconsistent findings. In general, studies at higher field strengths suggest that frontal glutamine function is elevated in the early stages of psychosis (18, 19), whereas the findings in chronic schizophrenia are more variable (26–30). This could reflect effects of antipsychotic medication. For example, one crosssectional and one longitudinal study reported higher Glx levels in the PFC of unmedicated, but not medicated schizophrenia, in comparison to controls (31, 32), although longitudinal studies have reported no effect of medication on anterior cingulate glutamate or glutamine levels (18, 33, 34). It could be that symptom severity contributes to variability in glutamatergic metabolites between patients early or late in the illness, as samples involving patients in the early phase of psychosis usually comprise non- or minimally medicated patients who are relatively symptomatic, whereas studies in chronic schizophrenia often involve patients who have been treated for long periods and have less severe or more stable symptoms (1). In relation to this, a recent meta-analysis of 1H-MRS glutamatergic studies in major depressive disorder found lower levels of Glx in the ACC which were only significant when remitted patients were excluded from the analysis (35).

In order to better understand the possible relationships between 1H-MRS glutamate, glutamine, and Glx levels and symptom severity, we identified publications that have reported associations with severity along symptom domains. These included studies in individuals at risk of psychosis, and patients with first-episode psychosis or established schizophrenia (**Table 1**).

## **RELATIONSHIPS BETWEEN GLUTAMATE FUNCTION AND POSITIVE SYMPTOMS**

Several studies have investigated associations between regional glutamatergic metabolite levels and the severity of positive psychotic symptoms (**Table 1**), and most found no association between regional Glx, glutamate or glutamine levels and positive symptom severity, in either high genetic or clinical risk populations (36, 37), first-episode psychosis (18, 19, 37–40) or chronic schizophrenia (26–28, 30, 41–46).

However, many of these studies involved small patient samples, relied on *post hoc* correlational analyses, and patients in whom the severity and/or the variance in severity of symptoms was low, due to either being sub-clinical threshold in the at risk studies or due to the presence of antipsychotic medication in established schizophrenia. A recent study pooling both medicated and unmedicated patients,where unmedicated patients possessed elevated Glx in medial prefrontal cortex (mPFC), detected an association between positive symptom severity and mPFC Glx, although this did not survive correction for multiple comparisons (31). Another study in the PFC found that treatment reduced the level of Glx/Cr in chronic patients, and associated the change in Glx with improvement in total BPRS score (32). Moreover a recent study found that 4 weeks of antipsychotic treatment in first-episode psychosis patients reduced Glx in the striatum, and this was associated with improvement in PANSS score (47).

A notable exception is the study of Ota et al. (48), which directly compared patients experiencing exacerbated psychotic symptoms to healthy controls and stable patients, and found that increases in Glx in inferior parietal white matter were specific to the group currently experiencing exacerbated psychotic symptoms (48). In line with this, our own studies which have compared glutamate levels in patients according to symptom severity have found higher ACC glutamate levels in first-episode psychosis patients who are still symptomatic following treatment compared to those in remission (38) and in patients with treatment-resistant schizophrenia compared to those who respond to medication (49). These differences may be independent of medication effects, as groups either did not differ according to medication (38, 49) or the symptomatic group were actually receiving higher medication doses (48), and as longitudinal studies have not reported changes in cortical glutamate in relation to changes in positive symptoms following antipsychotic treatment (18, 33, 34). This suggests that glutamate and Glx levels may be selectively elevated in patients whose positive symptoms are not well controlled by conventional antipsychotic medication.

## **RELATIONSHIPS BETWEEN GLUTAMATE FUNCTION AND NEGATIVE SYMPTOMS**

We recently reported that greater severity of PANSS negative symptoms was associated with higher levels of glutamate in the ACC in first-episode psychosis (38). Although there were several methodological differences, this contrasts with the study of Reid et al. (26), which reported that negative symptoms were associated with *lower* levels of ACC Glx in chronic schizophrenia. Other studies investigating correlations between ACC glutamatergic metabolites and negative symptoms have found no significant relationship (18, 19, 28, 42, 46, 50). We are not aware of any studies which have specifically compared regional levels of glutamatergic metabolites in patient groups selected according to differences in negative symptom severity.

As listed in **Table 1**, in brain regions other than the ACC many studies have failed to detect significant relationships between negative symptoms and regional glutamate, glutamine, or Glx levels in high genetic or clinical risk groups (36, 37), first-episode psychosis (18, 37, 39, 40), or chronic schizophrenia (27, 30, 31, 41, 44, 45, 48).

A general consideration of studies investigating relationships between glutamate markers and positive and negative symptoms is the scale used to score symptom severity. As detailed in **Table 1**, scores on a number of scales have been used and while these scales are highly correlated there are also some important differences in the clinical items included (51). As most studies examining



(Continued)


#### **Table 1 | Continued**

# Did not survive correction for the six comparisons (PANSS total, positive, and negative symptom subscales in two regions for each neurochemical).

Index: + denotes a positive relationship where higher levels of glutamate were associated with greater symptom severity or worse overall functioning; − denotes a negative relationship where higher levels of glutamate were associated with lesser symptom severity or better overall functioning. NS, not significant; ACC, anterior cingulate cortex; POC, parieto-occipital cortex; DLPFC, dorsolateral prefrontal cortex; MPFC, medial prefrontal cortex;Thal, thalamus; CSO, centrum semiovale; GHR, genetic high risk; RBANS, repeatable battery for the assessment of neuropsychological status; WCST, Wisconsin card sorting test; DSDT, the digit span distraction test;TMT, trail making test; IGT, Iowa gambling task; VF, verbal fluency; AVLT, auditory verbal learning test; BPRS, brief psychiatric rating scale; GAF, global assessment of functioning; CGI, clinical global impression scale; GAS, global assessment scale; SIPS, structured interview for prodromal symptoms; CSS, Chapman schizotypy scales; CDSS, Calgary Depression Scale for Schizophrenia; LSPR, life skills profile rating.

relationships between brain glutamate measures and symptoms have relied on *post hoc* correlational analysis, it is of note that additional items are included on the SAPS/SANS compared to the PANSS, which in turn has additional items compared to the BPRS. Therefore use of the SAPS/SANS may be preferable as they provide a more detailed assessment of symptoms.

## **RELATIONSHIPS BETWEEN GLUTAMATE MEASURES AND COGNITIVE DYSFUNCTION**

Relatively few studies have investigated relationships between glutamate measures and cognitive dysfunction in schizophrenia. The most commonly investigated task has been the Wisconsin card sort test (WCST): in schizophrenia, deficits on this task are associated with abnormal activation in the ACC and DLPFC (52). In a study of 19 patients with chronic schizophrenia at 3 T, poor performance on the WCST was associated with higher Gln/Glu ratios in the mPFC (including the ACC) (27). In contrast, a larger study of 43 patients with chronic schizophrenia at 1.5 T found that ACC, but not DLPFC, Glx levels were positively associated with WCST learning potential (42). In first-episode psychosis, an association between hippocampal, but not DLPFC, glutamate, and WCST errors was reported (53), and in a small sample of 16

genetic high risk individuals, no correlations between WCST performance and Glx levels in the caudate, ACC, or thalamus were detected (54).

Other tasks investigated include the Stroop, digit span distractibility test, auditory verbal learning test (AVLT), N-back task, Iowa gambling task and verbal fluency test (**Table 1**). Of these, there are reports of a positive association between mPFC Gln/Glu ratio and impairments on the digit span distraction test, which probes short-term memory and selective attention (27), and of DLPFC Glx and verbal learning and memory on the AVLT (55). Overall, the findings have been inconsistent and further studies with larger sample sizes are required. While the above studies used a voxel of interest method, using whole brain slice proton echo planar spectroscopy at 4 T in 30 patients Bustillo et al. (43) detected a positive correlation between general cognitive performance in schizophrenia and Glx. Furthermore, subsequent path analyses suggested that the relationship between glutamate and cognitive performance may be associated with negative symptoms and unemployment (43).

The relationship between glutamate and cognition in schizophrenia has been further investigated by combining 1H-MRS with functional magnetic resonance imaging (fMRI) of the blood oxygen level dependent (BOLD) response to measure changes in regional brain activation as participants perform cognitive tasks. In subjects at clinical high risk of developing psychosis, reductions in thalamic glutamate were correlated with an elevated BOLD response in the prefrontal cortex during a verbal fluency task, whereas the converse association was observed in controls (56). In a study by Valli et al. (57), medial temporal lobe glutamate levels were correlated with hippocampal BOLD response during an episodic memory task in controls, whereas this coupling was absent in clinical high risk subjects (57). Finally one study showed that hippocampal Glx was related to inferior frontal gyrus activation during episodic memory in controls, and suggested that the absence of this positive coupling in medicated schizophrenia patients may underlie episodic memory deficits, reflecting the results seen in Valli et al. (57, 58). Further investigation of the relationships between regional levels of glutamatergic metabolites and abnormalities in regional brain activation during cognitive tasks may provide a more sensitive means to characterize the relationship between glutamate and cognition in schizophrenia.

## **RELATIONSHIPS BETWEEN GLUTAMATE MEASURES AND SOCIAL AND OCCUPATIONAL FUNCTIONING**

Several studies have also reported correlations between brain glutamate levels and overall level of social and occupational functioning (**Table 1**). In genetically high risk subjects, lower levels of Glx/Cr in the mPFC were associated with lower levels of overall functioning (59). A longitudinal study at 4 T showed that loss of glutamate and glutamine in the thalamus, but not the ACC, over 7 years since first presentation correlated with impaired social functioning (34). In contrast, in first-episode psychosis higher levels of glutamate in the ACC, but not the thalamus, were associated with worse overall functioning (38), and in chronic schizophrenia higher DLPFC glutamate levels were also associated with worse overall functioning (60). Other studies have found no association

between glutamate measures and overall functioning in chronic schizophrenia (27) or genetic high risk subjects (36).

## **SUMMARY AND FUTURE DIRECTIONS**

1H-MRS studies relating regional glutamate measures to symptoms in schizophrenia have produced inconsistent findings. There may be several methodological reasons for this, including differences in the brain region investigated, and between samples such as medication, illness stage, and symptom severity. The use of sub-optimal technical approaches such as low field strengths together with small sample sizes may also underlie the conflicting findings. A recent study at 4 T indicates that elevations in the ratio of Gln/Glu are present in schizophrenia patients, consistent with elevated glutamatergic neurotransmission (18). However the majority of 1H-MRS studies in schizophrenia use field strengths <4 T which cannot reliably measure glutamine, and glutamate and Glx measures cannot be specifically attributed to glutamate neurotransmission (17). The increasing availability of higher field strength scanners may resolve some of the apparent inconsistencies in the literature.

The majority of studies that have investigated relationships between glutamate levels and symptom severity have applied correlational analysis, usually *post hoc* to the main study findings. A few studies have directly compared glutamatergic metabolite levels in patient groups on the basis of symptom severity, specifically those who were or were not currently experiencing symptom exacerbation (48), were or were not in symptomatic remission following initial treatment (38) or did or did not have treatmentresistant illness (49). All found higher levels of glutamate or Glx in the more symptomatic patient group.

The field would benefit from further studies pre-selecting groups of patients who differ in severity of negative symptoms or cognitive impairment, or longitudinal studies comparing within-subjects glutamate levels during periods of illness stability compared to relapse. Studies of the relationships between cognitive dysfunction and glutamate in schizophrenia may benefit from combination with fMRI to determine the efficiency of glutamate in supporting networks that subserve cognitive function. To date there are only a handful of published studies of this type. Finally, relationships between glutamate levels and symptom severity may be indirect; for example Stone et al. showed that hippocampal glutamate may interact with striatal dopamine to determine risk of psychosis (61), and a path analysis has suggested that negative symptoms may be secondary to poor cognition associated with low brain Glx (43).

One question is whether baseline levels of glutamate, glutamine, or Glx are predictive of subsequent outcome or response to treatment in schizophrenia. A recent study showed that elevated frontal Glx at baseline was associated with poor response after 4 weeks of antipsychotic treatment (62). This is consistent with the above findings of glutamatergic elevations in patients with schizophrenia whose symptoms have not responded well to treatment (38, 49).

The suggestion that symptoms that do not respond well to conventional antipsychotic treatment may have a glutamatergic basis (38, 49) warrants further investigation, as compounds that target the glutamatergic system may have particular efficacy in these patients. Meta-analyses conclude that of the agents which may improve NMDA receptor-mediated neurotransmission, the NMDAR co-agonist d-serine and the glycine transporter type 1 inhibitor sarcosine reduce total and negative symptoms as an adjuvant to antipsychotic medication (63, 64). Lamotrigine, which may inhibit glutamate release, blocks the psychomimetic effects of ketamine in healthy volunteers (65) and small trials of lamotrigine in chronic, often treatment-resistant or clozapine-treated patients found that it is beneficial in reducing symptoms (63, 66). Further work is required to determine the relationship between regional glutamate concentrations and the expression of symptoms at different stages of psychotic illness. This area may benefit from meta-analyses of the previously published findings and from new studies selecting patient samples *a priori* on the basis of symptom severity.

## **REFERENCES**


schizophrenia: a proton magnetic resonance spectroscopy study at 3 T. *Schizophr Res* (2007) **94**(1-3):328–31. doi:10.1016/j.schres.2007.05.008


after 4 weeks of antipsychotic treatment in first-episode psychosis: a longitudinal proton magnetic resonance spectroscopy study. *JAMA Psychiatry* (2013) **70**(10):1057–66. doi:10.1001/jamapsychiatry.2013.289


hyperglutamatergic effects of N-methyl-D-aspartate receptor antagonists. *Arch Gen Psychiatry* (2000) **57**(3):270–6. doi:10.1001/archpsyc.57.3.270


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 September 2013; paper pending published: 30 September 2013; accepted: 07 November 2013; published online: 26 November 2013.*

*Citation: Merritt K, McGuire P and Egerton A (2013) Relationship between glutamate dysfunction and symptoms and cognitive function in psychosis. Front. Psychiatry 4:151. doi: 10.3389/fpsyt.2013.00151*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Merritt, McGuire and Egerton. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited andthatthe original publication inthis journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Magnetic resonance imaging in studying schizophrenia, negative symptoms, and the glutamate system

## **Oliver Gruber <sup>1</sup>\*, Antonella Chadha Santuccione<sup>2</sup> and Helmut Aach<sup>2</sup>**

<sup>1</sup> Center for Translational Research in Systems Neuroscience and Psychiatry, Clinic for Psychiatry and Psychotherapy, University Medical Center Göttingen, Göttingen, Germany

<sup>2</sup> Medical Affairs – Psychiatry, Roche Pharma AG, Grenzach-Wyhlen, Germany

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Alice Egerton, King's College London, UK Naomi Robin Driesen, Yale University Medical School, USA

#### **\*Correspondence:**

Oliver Gruber, Clinic for Psychiatry and Psychotherapy, University Medical Center Göttingen, Von-Siebold-Str. 5, Göttingen 37075, Germany e-mail: ogruber@gwdg.de

Schizophrenia is characterized by positive, negative, and cognitive symptoms. While positive symptoms occur periodically during psychotic exacerbations, negative and cognitive symptoms often emerge before the first psychotic episode and persist with low functional outcome and poor prognosis. This review article outlines the importance of modern functional magnetic resonance imaging techniques for developing a stratified therapy of schizophrenic disorders. Functional neuroimaging evidence on the neural correlates of positive and particularly negative symptoms and cognitive deficits in schizophrenic disorders is briefly reviewed. Acute dysregulation of dopaminergic neurotransmission is crucially involved in the occurrence of psychotic symptoms. However, increasing evidence also implicates glutamatergic pathomechanisms, in particular N-methyl-D-aspartate (NMDA) receptor dysfunction in the pathogenesis of schizophrenia and in the appearance of negative symptoms and cognitive dysfunctions. In line with this notion, several gene variants affecting the NMDA receptor's pathway have been reported to increase susceptibility for schizophrenia, and have been investigated using the imaging genetics approach. In recent years, several attempts have been made to develop medications modulating the glutamatergic pathway with modest evidences for efficacy. The most successful approaches were those that aimed at influencing this pathway using compounds that enhance NMDA receptor function. More recently, the selective glycine reuptake inhibitor bitopertin has been shown to improve NMDA receptor hypofunction by increasing glycine concentrations in the synaptic cleft. Further research is required to test whether pharmacological agents with effects on the glutamatergic system can help to improve the treatment of negative symptoms in schizophrenic disorders.

#### **Keywords: schizophrenia, glutamate, negative symptoms, cognitive deficits, neuroimaging biomarkers, stratified therapy**

Neuroimaging techniques have been developed as important tools to investigate brain dysfunctions that underlie psychiatric disorders. In particular, modern functional magnetic resonance imaging (fMRI) holds the promise to provide neurofunctional biomarkers for improved diagnosis, prognosis, and optimized treatment of mental disorders [e.g., Ref. (1–4)]. In this brief, not exhaustive review, we will exemplify this translational neuroimaging research by focusing on schizophrenia and current challenges to advancing therapeutic approaches for this heterogeneous diagnostic category. First, the importance of negative symptoms and cognitive deficits for successful treatment of schizophrenic disorders will be highlighted,and an overview will be given on the neural correlates of these symptom domains as revealed by fMRI studies. Second, neuroimaging studies of glutamate levels and genetic risk factors pointing to an important role of glutamate–dopamine interactions in the pathophysiology of schizophrenic disorders will be briefly reviewed and will be related to recent findings of pharmacological and animal studies. In the final part of this review, we will present current approaches to develop therapeutic strategies that target the glutamatergic pathway in schizophrenic disorders.

## **"PSYCHOPATHOPHYSIOLOGY" OF SCHIZOPHRENIC DISORDERS AND THE IMPORTANCE OF NEGATIVE SYMPTOMS AND COGNITIVE DEFICITS**

In traditional psychiatry, mental disorders are diagnosed and classified on the basis of more or less specific psychopathological symptoms in the course of the disorder. Numerous recent findings from modern systems neuroscience and molecular neuroscience strongly suggest that diagnoses made on the basis of psychopathological criteria do not represent "natural disease entities" in the sense of diseases with uniform pathogenesis and pathology [see Ref. (1, 3–8)]. Therefore, schizophrenia represents a group of mental disorders that share more or less characteristic psychopathological symptoms, such as verbal hallucinations, delusions, and formal thought disorders.

During the last two decades, modern brain imaging techniques have allowed for the first time scientific investigations into the neural correlates of different symptom dimensions that characterize schizophrenic disorders. These studies of pathophysiological processes that underlie different psychopathological symptoms and syndromes (an approach that may be termed "psychopathophysiology") are briefly reviewed in the following.

Positive symptoms of schizophrenic disorders such as verbal hallucinations ("hearing voices") and delusional symptoms are certainly the most impressive characteristics of schizophrenic disorders. Nevertheless, much evidence has been provided that negative symptoms and cognitive disturbances are more strongly associated with the long-term functional outcome of patients suffering from schizophrenia than positive symptoms (9–14). Cognitive deficits are even present at first episode and remain relatively constant over the course of illness (15–17). In contrast to the dopamine model of schizophrenia, glutamatergic theories of schizophrenia account for negative symptoms and cognitive dysfunction as well, and may therefore lead to new treatment approaches specifically targeting the unmet medical need to improve negative symptomatology and cognitive deficits (18).

## **NEURAL CORRELATES OF POSITIVE SYMPTOMS IN SCHIZOPHRENIC DISORDERS**

The positive symptoms of schizophrenic disorders particularly include auditory–verbal hallucinations ("hearing voices") and delusional symptoms like paranoia. As regards auditory–verbal hallucinations, a number of functional brain imaging studies have been performed in the past. Overall, findings of these studies appear quite heterogeneous. However, a finding that has been replicated several times is overactivation of the superior temporal gyrus during experimental phases in which the patients exhibited the symptom of hearing voices [e.g., Ref. (19–21)]. In most of these studies, this activation of auditory association cortices was associated with additional activations in other brain regions. For instance, some of these studies also reported an increased brain activity in Broca's area, the anterior cingulate cortex, the hippocampus, and the amygdala [e.g., Ref. (20)]. Some of these findings have also been confirmed on the meta-analytical level. Jardri et al. (22) confirmed that phases of "hearing voices" are associated with increased activity in Broca's area, anterior insula, precentral gyrus, frontal opercular cortex, middle and superior temporal gyrus, inferior parietal lobule as well as in hippocampus and parahippocampal gyrus. It has to be noted, however, that patients suffering from this kind of intermittently occurring auditory–verbal hallucinations are probably not representative for most types of schizophrenic disorders in which the hallucinations persist for longer time periods. A second meta-analysis by Kühn and Gallinat (23) came to the conclusion that a current psychopathological state of experiencing auditory–verbal hallucinations may be associated with abnormal activation of brain regions that are also involved in speech production (e.g., Broca's area), whereas a subgroup of schizophrenic patients that exhibits the symptom of auditory–verbal hallucinations (in comparison to a subgroup without "life-time diagnosis" of auditory–verbal hallucinations) may be particularly characterized by abnormal brain activation in areas involved in speech processing and, more generally, the processing of auditory stimuli. In another, qualitative and quantitative review, Goghari et al. (24) showed an association between positive symptoms, in particular ideas of persecution, with the activity in medial prefrontal cortex, amygdala, hippocampus, and parahippocampal gyrus.

## **NEURAL CORRELATES OF NEGATIVE SYMPTOMS IN SCHIZOPHRENIC DISORDERS**

Negative symptoms of schizophrenic disorders are usually defined as symptoms representing a qualitative and/or quantitative reduction of mental capacities or qualities of experience. This class of symptoms is relatively heterogeneous and traditionally includes the five "A"s, which are affective flattening, apathy (reduced drive), anhedonia, asociality (social withdrawal), and alogia (impoverishment of thought) [e.g., Ref. (25)]. Early studies on the structural correlates of negative symptoms in schizophrenic disorders have shown gray matter reduction in temporal, medial frontal, insular, and hippocampal regions [e.g., Ref. (26)]. On the other hand, evidence on brain structural correlates of negative symptoms is very heterogeneous as there are also several studies that failed to find any correlation of brain volumes to negative symptoms in schizophrenia [e.g., Ref. (27–30)].

The development of functional neuroimaging techniques like PET and fMRI led to an increasing number of studies on the neurofunctional correlates of negative symptoms. Several studies have shown a reduced activation of the prefrontal cortex in schizophrenic patients with negative symptoms (25, 31–33). This principal finding of "hypofrontality" associated with negative symptoms has been confirmed by later studies though for different subregions of the prefrontal cortex. For example, in a study using memory retrieval tasks, Heckers et al. (34) found a significantly reduced recruitment of left frontal cortex (Brodmann area 44/9) in schizophrenic patients with deficit syndrome (i.e., patients with negative symptoms as primary and enduring features) as compared to both schizophrenic patients without deficit syndrome and healthy controls. Using auditory working memory tasks (*n*-back tasks), Menon et al. (35) found an inverse correlation of negative symptoms with activation in the frontal opercular cortex and in the right DLPFC. In contrast to that, another experiment using the *n*-back task (36) reported a correlation of activation deficits in the DLPFC with disorganization symptoms rather than with negative symptoms.

A number of functional neuroimaging studies have also reported associations of negative symptoms with activation in other brain regions including temporal cortices and the ventral striatum. For instance, Tamminga et al. (37) found both limbic system abnormalities and neocortical alterations associated with the deficit syndrome. A later series of studies have demonstrated a significant correlation of activation in temporal cortices with negative symptoms (38–40). Using the monetary incentive delay task, which activates the reward system, Juckel et al. (41) found a correlation of diminished ventral striatal activation with negative symptoms in schizophrenic patients. In another study by Simon et al. (42), a ventral striatal activation during reward anticipation was negatively correlated with symptoms of apathy, while activation during receipt of reward was negatively correlated with severity of depressive symptoms. Finally, in a recent systematic review and meta-analysis of 25 fMRI studies on schizophrenic symptomatology, Goghari and colleagues (24) have confirmed a relationship of negative symptoms with the functioning of the ventrolateral prefrontal cortex and the ventral striatum.

## **NEURAL CORRELATES OF COGNITIVE DYSFUNCTIONS IN SCHIZOPHRENIC DISORDERS**

Traditionally, psychiatric diagnosis of schizophrenic disorders is based on psychopathological, particularly positive and negative symptoms. Over the last few decades, the advent of modern experimental neuropsychological and functional neuroimaging techniques has switched the focus of interest toward brain dysfunctions in the domains of cognitive, emotional, and motivational processes that are highly prevalent in schizophrenic disorders. Especially cognitive deficits are of major interest as they have been proposed to represent core deficits of schizophrenic disorders [e.g., Ref. (43, 44)]. These core cognitive deficits of schizophrenic disorders include deficits in working memory, executive functions, episodic memory, and social cognitions.

Deficits in working memory in schizophrenic disorders have been found to be associated with dysfunctions of prefrontal cortices, especially of the dorsolateral prefrontal cortex, of the deep fronto-opercular cortex, and of the anterior cingulate cortex [e.g., Ref. (45–50)]. In the last few years, there have also been several reports of a disturbed connectivity between these prefrontal areas and the medial temporal lobe, particularly the hippocampus [e.g., Ref. (51, 52)].

Executive function is a construct that encompasses a variety of sub-functions [see for example Ref. (53, 54)], among them selective attention, background monitoring of the environment for potentially significant sensory events [e.g., Ref. (55)], and the adaptation of behavior to changing environmental conditions [e.g., Ref. (56–60)]. Patients with schizophrenic disorders exhibit multiple dysfunctions in these areas of executive control mechanisms that are associated with reduced activity in the posterior frontal medial cortex and the inferior frontal junction area (IFJA; a cortical subregion at the intersection of the precentral sulcus and the inferior frontal sulcus) (61, 62) as well as with abnormally increased activations in brain stem nuclei and the ventral striatum (63).

Episodic memory deficits in schizophrenic disorders have been found to be associated firstly with dysfunction of the extended hippocampal formation (including the hippocampus proper and the surrounding medial temporal structures) [e.g., Ref. (64)], and secondly with dysfunctions of prefrontal cognitive control mechanisms that are involved in encoding and retrieval processes (65).

Disturbed social cognitions in schizophrenic disorders include particularly impaired recognition of facial emotional expressions and reduced theory-of-mind capacities. Neuroimaging studies have revealed that deficits in these domains of social cognitions are associated, firstly, with reduced activation in the amygdala and the fusiform gyrus (66) and, secondly, with reduced activity in the fronto-median cortex, in the temporo-parietal junction cortex, and the amygdala–hippocampus complex [e.g., Ref. (67)]. Disturbed functional coupling between prefrontal areas and the amygdala has also been reported as a correlate of impaired emotional regulation mechanisms (68).

Taken together, neuroimaging studies on the neural correlates of different (positive, negative, cognitive) symptom domains in schizophrenic disorders suggest the involvement of different lateral and medial prefrontal, temporal (particularly including hippocampus and amygdala), and subcortical (in particular ventral striatum as part of the dopaminergic reward system) brain regions in the occurrence of these symptoms.

## **PATHOGENESIS, PATHOPHYSIOLOGY, AND TREATMENT OF SCHIZOPHRENIC DISORDERS: THE ROLE OF GLUTAMATE–DOPAMINE INTERACTIONS**

It is well-established that acute dysregulation of dopaminergic (and glutamatergic) neurotransmission is crucially involved in the occurrence of psychotic symptoms, whereas more chronic cellular neuropathology may be responsible for the development of cognitive deficits and negative symptoms (69). In more recent years, classical elements of the dopamine hypothesis such as the overactive mesolimbic dopamine system and a reduced mesocortical dopamine turnover are considered rather as a"final common pathway" (70), i.e., as the expression of upstream pathophysiological changes. In particular, it is postulated that *N*-methyl-d-aspartate (NMDA) receptor dysfunction may lead to dopaminergic dysregulation in schizophrenic disorders through a complex interaction between glutamatergic and dopaminergic, but also GABA-ergic mechanisms. Schwartz and co-workers (71), for example, explain both positive and negative symptoms of schizophrenia with dysfunctions of NMDA-glutamatergic synapses. Via effects on various complex circuits including GABA-ergic interneurons, these dysfunctions ultimately result both in hyperfunction of the mesolimbic dopamine system leading to positive symptoms, and also in hypofunction of the mesocortical dopamine system associated with negative symptoms and cognitive dysfunctions (**Figure 1**). Such pathomechanisms in functional interactions between prefrontal cortices and brain stem nuclei, in particular the ventral tegmental area (VTA), the striatum, and the thalamus, may be influenced by predisposing and/or protective genetic variants that exert effects on the glutamatergic synapse. Thus, the glutamatergic hypothesis is also compatible with our knowledge about the effects of susceptibility genes of schizophrenia (see below).

Modern pharmacological and animal research approaches are increasingly focusing on the interactions between the dopaminergic and glutamatergic system, especially within fronto-striatothalamo-frontal loops and in the interactions between frontal cortex, hippocampus, nucleus accumbens, and VTA. Within and between these brain regions, the dopamine and the glutamate system interact in a very complex way, and dysfunctions in these interactions are a central pathomechanistic explanation for the development of schizophrenic disorders [e.g., Ref. (73, 74)]. Particularly, the hippocampus plays a major role in regulating the dopaminergic reward system. Animal studies have shown that increased activity of the ventral hippocampus (subiculum) leads to increased dopamine turnover in the ventral striatum (nucleus accumbens) (75, 76). Other recent findings suggest that this glutamatergically mediated effect of the subiculum on the nucleus accumbens further leads to increased GABA-ergic projection onto the ventral pallidum with reduced tonic activity of the pallidum and consecutive disinhibition of dopamine neurons in the VTA (77). In this way, hyperactivity of the ventral hippocampus observed in schizophrenic disorders could lead to overstimulation and hyperactivity of dopaminergic VTA neurons which may account for various symptoms, in particular for delusional

phenomena. Further support for this theory is provided by studies using the MAM animal model of schizophrenia which have shown that the relevant pathophysiological changes such as VTA hyperactivity and increased response to amphetamines are no longer present after inactivation of the subiculum. This indicates that the subiculum is necessary to induce hyperdopaminergic states in this animal model (78).

The pathophysiological role of glutamatergic disbalances has also been investigated *in vivo* in patients with schizophrenia, for example using magnetic resonance spectroscopy (MRS). In very recent publications (79, 80), findings of such glutamatergic proton magnetic resonance spectroscopic imaging studies have been nicely reviewed, particularly with respect to their implications for drug discovery. Overall, neuroimaging studies support the current glutamate model of schizophrenia by suggesting a hypofunction of the NMDA receptor. In particular, proton magnetic resonance spectroscopic (1H-MRS) studies have provided evidence for altered levels of glutamate and glutamine in the medial prefrontal cortex and in the basal ganglia in early-stage, drug-naïve, or drug-free schizophrenia patients.

Some studies with unmedicated patients with schizophrenia have reported elevated glutamatergic levels in the medial prefrontal cortex as compared to healthy controls (81–84). More precisely, a recent meta-analysis by Marsman and colleagues (79) indicated that it is glutamine which is increased in the frontal cortex in schizophrenic patients, whereas glutamate is reduced. Such an elevated glutamine/glutamate ration may result from either a deficiency in glutaminase, which converts glutamine into glutamate, or from NMDA receptor hypofunction which has also been shown to increase glutamine levels and decrease glutamate levels (79). Further, glutamate levels in the medial prefrontal cortex have been found to be associated with negative symptoms and worse global functioning and to be decreased in remitted patients as compared to non-remitted patients (85). Consistent with that, most studies comparing medicated patients with healthy control subjects reported unchanged glutamate levels in the medial prefrontal cortex (81, 86–92). The meta-analysis by Marsman and colleagues (79) provided additional support for a progressive decrease of frontal glutamate and glutamine in patients with schizophrenia possibly indicating a progressive loss of synaptic activity. Finally, particularly in first episodes schizophrenic patients, increased glutamatergic levels have also been reported in the basal ganglia (93–95), and they appear to decrease to normal levels during antipsychotic treatment with risperidone (94).

Over the last 10–15 years, numerous potential susceptibility genes of schizophrenia have been identified, among them COMT, dysbindin-1, neuregulin-1, RGS4, GRM3, and DISC1. Many of these candidate genes have been shown to influence dopaminergic and/or glutamatergic neurotransmission, and effects on neuroplastic processes and particularly on synaptogenesis have also been reported. Imaging genetics is still a relatively novel approach that, however, has already made substantial contributions to our knowledge about genetic effects on brain structure and function. Early studies, for example, demonstrated the influence of variants in the COMT gene on working memory-related prefrontal activation (96) and on the functional interplay between dopamine synthesis in the midbrain and prefrontal function (97). Although the evidence for an association between the COMT gene and schizophrenia is not unequivocal, these findings nevertheless have high biological plausibility insofar as the influence of the COMT gene on the dopaminergic tone in the prefrontal cortex has been convincingly demonstrated (98). Further studies on the COMT genotype have shown more complex haplotype effects on prefrontal cerebral activations (99) and on gene–gene interactions between COMT and other genes such as RGS4, G72, DISC1, and GRM3 (49, 100, 101). Especially the latter finding is consistent with a role of glutamate–dopamine interactions in the pathophysiology and pathogenesis of schizophrenic disorders.

The number of genome-wide association studies between gene variants and diseases has markedly increased over the last few years due to the availability of modern chips. This has also inspired imaging genetics studies as genome-wide confirmed risk variants have also been investigated for their effects on brain structure and function. Two examples of this are the zinc finger protein 804A (ZNF804A), the function of which has not yet been more closely characterized, but which showed a genome-wide significant association with schizophrenia and also with bipolar disorder (102), as well as the CACNA1C gene, which was first discovered as a risk gene for bipolar disorder, but later also for schizophrenia (103). Studies on the ZNF804A polymorphism have shown an effect on the connectivity between the prefrontal cortex and the hippocampus (104–106). Effects of the CACNA1C gene have been reported with regard to activation of the hippocampus and the subgenual

Lewis and Gonzalez-Burgos (72)].

ACC (107) as well as activation of the amygdala during reward and fear recognition paradigms (108, 109).

Taken together, the studies summarized here support the important pathophysiological role of glutamate in schizophrenia and encourage further development of therapeutic strategies that target the glutamatergic pathway in schizophrenia.

## **THERAPEUTIC STRATEGIES TARGETING THE GLUTAMATERGIC PATHWAY IN SCHIZOPHRENIA**

A valid treatment for positive symptoms, based on the use of antipsychotic agents and their main capacity to modulate the dopaminergic system, is currently available for schizophrenia. However, antipsychotics are less effective in reducing negative symptoms or in ameliorating cognitive dysfunctions (110–112). Based on the novel findings that the glutamatergic system plays an important role in the pathogenesis of schizophrenia, several attempts have been made to identify drugs which, by modulating this system, could improve negative symptoms and cognitive dysfunction (113). Pharmacological targets are different types of glutamate receptors, which interact in a complex not yet fully understood way within glutamatergic synapses. These receptors include both ionotropic receptors [NMDA, α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA), and kainate receptors] and metabotropic glutamate receptors (mGluRs) [reviewed in Ref. (114)]. AMPA and kainate receptors are fast gating by glutamate binding and permeable for Na<sup>+</sup> and K<sup>+</sup> ions, whereas NMDA receptors exhibit a slow gating kinetic and need a predepolarization of the neuronal membrane for activation by glutamate binding. This predepolarization occurs when vicinal AMPA receptors are frequently activated and leads to the release of a blocking Mg2<sup>+</sup> ion from the NMDA receptor. Activated NMDA receptor channels are not ion specific and pass Ca2+, Na+, and K<sup>+</sup> ions, which leads not only to further depolarization but also intracellular processes making the synapse more sensible for signals from upstream and sending stronger signals downstream. This so-called long-term potentiation (LTP) provides the basis for synaptic and dendritic proliferation or pruning, learning, and memory (115).

Metabotropic glutamate receptors are G protein-coupled receptors influencing intracellular metabolic processes and are present on presynaptic and postsynaptic neurons as well as on glial cells near glutamatergic synapses. Currently, eight mGlu receptors are known of which the mGlu 2/3 receptors are investigated as targets for schizophrenia therapy, because they regulate presynaptic glutamate secretion.

After successful completion of preclinical trial programs of pharmacodynamic activity and safety, potential compounds are investigated in escalating clinical study programs. Phase 1 studies are open label, single to multiple dose trials with healthy volunteers, exploring pharmacokinetic and first safety characteristics. Phase 2 studies are explorative or proof-of-concept studies that are usually controlled and blinded with small to medium numbers of patients, and designed for dosis finding and first in patient proof of safety and efficacy. Phase 3 studies are large scale, double blinded controlled studies to confirm safety and efficacy. Some of the following described trial results are not yet published in peerreviewed journals, and we had to quote congress presentations or company statements.

#### **COMPOUNDS ENHANCING NMDA RECEPTOR FUNCTION**

Since the activation of the NMDA receptor also requires glycine as a co-agonist, the glycine binding site at the NMDA receptor is regarded as a promising pharmacological target to enhance its activity, thereby minimizing the risk of excitotoxicity that is associated with direct overactivation of the glutamate binding site (116). A first review of clinical trials published until 2003 with agonists of the glycine binding site of the NMDA receptor including glycine, d-cycloserine, and d-serine reported moderate effect sizes for glycine and d-serine on negative symptoms, but no effect of d-cycloserine (117). It included 358 small randomized trials with 6–51 participants and a maximum duration of 12 weeks. Up to now, the amount of data has grown confirming these early results. In almost all trials, the compounds were used as adjunctive treatment to regular antipsychotic therapy, and generally no effect was reported in patients taking clozapine.

Several clinical trials have been conducted using glycine, a substance that is endogenously produced and that can act as a co-agonist of NMDA receptors binding at the glycine modulatory site. The results of these studies suggest that glycine can significantly improve symptoms of schizophrenia (118), including negative symptoms, although there are also negative and equivocal studies (119).

d-Cycloserine is a selective co-agonist of NMDA receptors containing the NR2C subunit, and these receptors are involved in fear conditioning and memory consolidation (120). When used as a single dosis in rodents, d-cycloserine led to enhanced memory consolidation of novel information (120). In an exploratory clinical trial, once weekly dosing of d-cycloserine augmentation over 8 weeks did not improve cognitive symptoms but reduced negative symptoms and delusion severity in stable schizophrenic patients medicated with a range of different antipsychotics excluding clozapine (121).

d-Serine, another agonist of the glycine modulatory site within the NMDA receptor, has been shown in clinical trials to be capable of ameliorating several symptom domains in schizophrenia (122–124).

#### **GLYCINE TRANSPORTER-1 INHIBITORS**

Sarcosine is a potent natural glycine transporter-1 inhibitor (GlyT-1) (125). The inhibition of this transporter leads to increased levels of glycin in the synapsis and consequently enhanced NMDA receptor activation, which may represent a possible treatment mechanism for schizophrenic disorders in which a hypofunction of NMDA receptors is present (126). As recently shown, sarcosine may reduce negative symptoms in acutely ill schizophrenia patients receiving atypical antipsychotics, being more effective than the NMDA/glycine site agonist d-serine (125).

A meta-analysis including all abovementioned molecules found that, overall, the NMDA-enhancers were effective against most symptom domains of schizophrenia. Glycine, d-serine, and sarcosine significantly improved multiple symptom domains, whereas no symptom domain was improved by d-cycloserine. Furthermore, glycine, d-serine, and sarcosine were found to be superior to d-cycloserine in improving overall psychopathology [Ref. (118); see **Figure 2**]. However, these compounds have individual disadvantages to be developed to drugs licensed for long-term

use. Glycine is fast metabolized and passes the blood–brain barrier poorly. So it has to be applied with daily doses up to 60 g. Although d-serine is substantially metabolized, daily doses of 30–120 mg/kg were effective. But there is concern about nephrotoxicity at higher doses, although no significant adverse events have yet been observed at doses of ≤4 g/day (127). According to Sreekumar et al. (128), sarcosine might play a role in aggravating prostate cancer progression.

Bitopertin is a GlyT-1 inhibitor that increases levels of the glycine neurotransmitter by inhibiting its reuptake from the synaptic cleft. Preclinical evidence showed that this molecule is capable of ameliorating the symptoms of schizophrenia in animal models (129, 130). These preclinical findings encouraged a double blinded placebo-controlled phase IIb clinical trial which showed that adjunctive treatment with bitopertin in stable patients with predominant negative symptoms was capable of ameliorating negative symptoms and improving general clinical status (131). Currently, several phase III studies are underway with the hope that bitopertin may help in the treatment of currently unsatisfyingly responding stages of schizophrenia.

A major goal for future research combining psychopharmacology and modern functional neuroimaging techniques would be to understand how these molecules modulate the activity of pathophysiologically relevant neural structures as outlined above. Such studies could, for example, provide important information on whether these pharmacological agents can be successfully used to treat patient subgroups that are characterized by specific symptoms of schizophrenia.

## **AMPA RECEPTOR MODULATORS**

As described above, the fast trafficking of AMPA receptors in the synaptic cleft has an impact on NMDA receptor-mediated LTP and depression. These intracellular mechanisms influence synaptic strength and therefore constitute the basis of learning and memory (132). Therefore, modulation of AMPA activity could lead to amelioration of cognitive dysfunction in schizophrenia. To this aim and based on preclinical evidences, two molecules have been used in schizophrenia clinical trials: piracetam (133) and CX516 (121, 134).

Piracetam augmentation of haloperidol was capable of improving psychotic symptoms in schizophrenia, but had no effect on PANSS (133). Because only 30 patients (all receiving haloperidol) completed the placebo-controlled trial, more scientific evidence is needed to support such an effect. Trials with CX156 led to controversial results. In a small study, CX156 improved cognitive functions and negative symptoms in schizophrenic patients when compared to patients treated with clozapine (134). However, a larger study was unable to show any effect of CX516 on cognition or negative symptoms when compared to controls (135). Taken together, there is only little evidence about these molecules and their therapeutic effect.

## **mGlu RECEPTOR MODULATORS**

In contrast to the concept of improving symptoms of schizophrenia with ampakines, a line of evidence points to an overactivation of AMPA synapses in the prefrontal cortex downstream of NMDA receptor hypofunction (136). NMDA receptor blockade on GABAergic interneurons reduces inhibition of pyramidal cells and leads to excessive glutamate release in AMPA receptor synapses in the prefrontal cortex (137). Metabolic glutamate receptors 2 and 3 (mGlu 2/3) facilitate a feedback regulation of synaptic glutamate release (138). Consequently, mGlu 2/3 enhancing agents were developed to delimit pathologically enhanced glutamate release. In schizophrenia, clinical trials were conducted with the mGluR2/3 agonist, pomaglumetad methionil (LY2140023) and the mGlu2 positive allosteric modulator, ADX71149.

Pomaglumetad methionil was investigated as monotherapy in three clinical trials. At first, a phase 2 proof-of-concept trial showed significant improvement of positive and negative symptoms versus placebo (139). In a following phase 2 dose ranging trial, all of the four investigated dosing groups did not differ from placebo (140), which was also the case for a phase 2 trial comparing pomaglumetad methionil to olanzapine and a placebo group. In this trial, both active treatment groups did not separate regarding efficacy and safety parameters from the placebo group (141).

After demonstrating an augmentation of the efficacy of atypical antipsychotics in preclinical trials, a phase 1 study was conducted to prove the safety of the combination of pomaglumetad methionil with four different second generation antipsychotics in healthy subjects (142). A following placebo-controlled phase 2 study tested the substance as adjunctive to standard of care in patients with prominent negative symptoms of schizophrenia. This trial did not indicate a significantly greater reduction of negative symptoms or an improvement of secondary efficacy endpoints over placebo (143). Based on these results, a phase 3 study started in the meantime was stopped.

The mGluR2 selective positive allosteric modulator ADX71149 is co-developed by Addex Therapeutics (144) and Johnson & Johnson, who code it JNJ40411813. Results of a randomized placebo-controlled phase 2 study evaluating the safety, tolerability, and exploratory efficacy of the compound given in two different doses as adjunctive to an ongoing antipsychotic medication were reported at the 2013 annual meeting of the American Psychiatric Society. The study population comprised three groups: patients with residual negative symptoms, patients with residual positive symptoms, and patients with insufficient response to clozapine treatment. Tolerability results suggest that dose titration may be beneficial. An efficacy signal seen in the negative symptoms subgroup treated with the lower dose suggests this population warrants further evaluation in a formal proof-of-concept study (144, 145).

## **CONCLUSION AND FUTURE PERSPECTIVES**

In this brief review article, we have summarized recent findings from genetic, animal, and functional neuroimaging studies that together point to an important role of glutamate–dopamine interactions within cortico-striato-thalamo-cortical (CSTC) loops, which are modulated by hippocampal and amygdala inputs, in the pathophysiology of schizophrenic disorders. These findings provide the empirical basis for the development of novel treatment approaches for schizophrenic disorders that target glutamatergic mechanisms. The new findings from animal studies may also inspire analogous clinical neuroimaging investigations of neurofunctional interactions within CSTC loops as well as of the dynamic effects of mediotemporal structures such as hippocampus and amygdala on these CSTC loops in the human brain. Here, the development of valid animal model-supported experimental paradigms is of major importance as it may allow for the targeted *in vivo* investigation of these pathomechanisms involved in the pathophysiology of schizophrenic disorders [e.g., Ref. (4)]. Such targeted investigations may enable a future stratification of the heterogeneous group of schizophrenic disorders into pathophysiologically more homogenous "natural disease entities" (146). Moreover, the development of neurofunctional MRI biomarkers for sub-classification of patient groups and prediction of individual treatment responses may generally play an important role in a future individualized medicine in psychiatry [e.g., Ref. (4)].

## **ACKNOWLEDGMENTS**

This publication was supported by a grant of the Roche Pharma AG, Germany.

## **REFERENCES**


prefrontal cortex and basal ganglia during working memory performance. *Biol Psychiatry* (2000) **48**(2):99–109. doi:10.1016/S0006-3223(00)00227-4


**Conflict of Interest Statement:** Oliver Gruber was honorary speaker for the following companies: Astra Zeneca, Bristol Myers Squibb, Janssen Cilag, Lilly, Lundbeck, Otsuka, Servier. Oliver Gruber has been invited to scientific congresses by Astra Zeneca, Janssen Cilag, Pfizer, Servier. Oliver Gruber has received a research grant from Servier and a publication grant from Roche. Antonella Chadha Santuccione and Helmut Aach are full employes of Roche Pharma AG, Grenzach, Germany.

*Received: 01 October 2013; accepted: 14 March 2014; published online: 03 April 2014. Citation: Gruber O, Chadha Santuccione A and Aach H (2014) Magnetic resonance imaging in studying schizophrenia, negative symptoms, and the glutamate system. Front. Psychiatry 5:32. doi: 10.3389/fpsyt.2014.00032*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2014 Gruber, Chadha Santuccione and Aach. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Targeting neural synchrony deficits is sufficient to improve cognition in a schizophrenia-related neurodevelopmental model

## **Heekyung Lee<sup>1</sup>†‡, Dino Dvorak <sup>2</sup>‡ and André A. Fenton3,4\***

<sup>1</sup> Graduate Program in Neural and Behavioral Science, Downstate Medical Center, State University of NewYork, Brooklyn, NY, USA

<sup>2</sup> Graduate Program in Biomedical Engineering, Downstate Medical Center, State University of New York and New York University Polytechnic School of Engineering, Brooklyn, NY, USA

<sup>3</sup> The Robert F. Furchgott Center for Neural and Behavioral Science, Downstate Medical Center, State University of New York, Brooklyn, NY, USA

<sup>4</sup> Center for Neural Science, New York University, New York, NY, USA

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Vaibhav A. Diwadkar, Wayne State University School of Medicine, USA John Gigg, University of Manchester, UK

#### **\*Correspondence:**

André A. Fenton, Center for Neural Science, New York University, 4 Washington Place, New York, NY 10003, USA e-mail: afenton@nyu.edu

#### **†Present address:**

Heekyung Lee, Krieger Mind/Brain Institute, Johns Hopkins University, Baltimore, MD, USA

‡Heekyung Lee and Dino Dvorak have contributed equally to this work. Cognitive symptoms are core features of mental disorders but procognitive treatments are limited. We have proposed a "discoordination" hypothesis that cognitive impairment results from aberrant coordination of neural activity. We reported that neonatal ventral hippocampus lesion (NVHL) rats, an established neurodevelopmental model of schizophrenia, have abnormal neural synchrony and cognitive deficits in the active place avoidance task. During stillness, we observed that cortical local field potentials sometimes resembled epileptiform spike-wave discharges with higher prevalence in NVHL rats, indicating abnormal neural synchrony due perhaps to imbalanced excitation–inhibition coupling. Here, within the context of the hypothesis, we investigated whether attenuating abnormal neural synchrony will improve cognition in NVHL rats. We report that: (1) inter-hippocampal synchrony in the theta and beta bands is correlated with active place avoidance performance; (2) the anticonvulsant ethosuximide attenuated the abnormal spike-wave activity, improved cognitive control, and reduced hyperlocomotion; (3) ethosuximide not only normalized the task-associated theta and beta synchrony between the two hippocampi but also increased synchrony between the medial prefrontal cortex and hippocampus above control levels; (4) the antipsychotic olanzapine was less effective at improving cognitive control and normalizing place avoidance-related inter-hippocampal neural synchrony, although it reduced hyperactivity; and (5) olanzapine caused an abnormal pattern of frequency-independent increases in neural synchrony, in both NVHL and control rats. These data suggest that normalizing aberrant neural synchrony can be beneficial and that drugs targeting the pathophysiology of abnormally coordinated neural activities may be a promising theoretical framework and strategy for developing treatments that improve cognition in neurodevelopmental disorders such as schizophrenia.

**Keywords: neural synchrony, oscillations, schizophrenia, neurodevelopmental models, hippocampus, local field potentials, cognitive control, spike-wave discharge**

## **INTRODUCTION**

Cognitive impairment is a core feature of schizophrenia but treatments to improve cognition are limited, in part because the physiological mechanisms of cognitive dysfunction are unclear. Recent interest has been directed toward understanding the abnormal synchrony of neural oscillations as a pathophysiological mechanism that underlies the cognitive impairments in schizophrenia (1–4). Dysfunctional neural synchrony in a wide range of frequencies, including the delta (5), theta (6), beta (7), and gamma ranges (8, 9), have been characterized in patients with schizophrenia. In animal models, disruption of long-range hippocampal– prefrontal synchrony has been reported while performing a working memory task in a genetic mouse model of schizophrenia (10) and during an open-field task in the maternal immune activation (MIA) neurodevelopmental model of schizophrenia (11). Abnormal hippocampal–prefrontal synchrony was also observed in the neonatal ventral hippocampus lesion (NVHL) neurodevelopmental model but this synchrony abnormality was not itself associated with cognitive impairment (12). In fact, aberrant neural synchrony may be a common pathophysiology responsible for impaired cognition in a variety of disorders (2, 13). Indeed, treatments that normalize neural network dysfunction without addressing the underlying cellular pathology have been sufficient to improve cognitive function in animal models of neurodegenerative diseases (14, 15). Such studies demonstrate it may be a feasible strategy to improve cognitive functional outcomes by normalizing aberrant neural network synchrony in cases where the network dysfunction is part of the causal chain from the disease etiology to the manifest cognitive symptoms.

We previously reported that NVHL rats, an established neurodevelopmental animal model of schizophrenia (16), have a cognitive control impairment that was associated with abnormal synchrony between the two hippocampi (12). We also detected hippocampal–prefrontal synchrony abnormalities in these rats during home-cage resting behaviors, but these abnormalities were absent during place avoidance training, and so not related to performance on the cognitive test. Clinically, the presence of hippocampal dysfunction is apparent in schizophrenia. Postmortem and neuroimaging studies indicate that the hippocampus is one of the brain regions most consistently altered in schizophrenia (17–22). Furthermore, functional magnetic resonance imaging studies show dysfunctional hippocampal activities are associated with poor cognitive performance in schizophrenia patients (23–25). Perhaps similar to the findings in NVHL rats, reduced inter-hippocampal functional coupling has also been observed between the left- and right-hippocampus areas in healthy carriers of the CACNA1C risk variant in bipolar disorder (26), but to our knowledge these measurements have not been carried out in schizophrenia subjects. Because neuroimaging studies do not have sufficient temporal resolution to detect functional synchrony at the 100-ms or faster timescales of the electrophysiological synchrony that underlies information processing between neural networks, it is unclear whether or not the inter-hippocampal abnormalities in NVHL rats mimics abnormalities in schizophrenia. Regardless, it is our opinion that animal models used to study schizophrenia do not sensibly mimic disease features, instead they are better thought of as being models for testing hypotheses that are relevant to understanding schizophrenia. Thus, we hypothesized that the abnormal hippocampal synchrony plays a role in the pathophysiology of cognitive control impairment in a schizophrenia-related neurodevelopmental animal model. As outlined below, in the present study, we tested the idea that attenuating this pathophysiology can improve cognition.

We, like others, have proposed the discoordination hypothesis, which asserts that cognitive impairment arises because of aberrant coordination of what may otherwise be normal neural activities (12, 27). The discoordination hypothesis is rooted in the idea that disrupted physiological coordination and synchronization of unitary neural processes impair the ability to selectively and dynamically activate and suppress information for organizing knowledge and perception into useable representations (1, 2, 27, 28). We use the term (dis)coordination to refer to the collective temporal organization of unitary neural activity such as action and synaptic and local field potentials (LFPs) at individual sites. Thus, (dis)coordination can be measured at different temporal and spatial scales. In the present work, we examined inter-hippocampal synchrony of LFP oscillations. These oscillations are an example of a long-range spatial scale neural interaction that involves an approximate order of 10<sup>5</sup> neurons. These oscillations result from the multiple timescales of excitation–inhibition neural activity cycles that govern spiking within these networks (29). The discoordination hypothesis predicts that in the NVHL model, restoring the abnormal neural synchrony to normal will attenuate the cognitive control deficit. Accordingly, we took advantage of the hippocampal synchrony abnormality in NVHL rats, specifically the inter-hippocampal synchrony that was associated with cognitive performance,to evaluate the prediction. We examined whether normalizing the abnormal synchrony can attenuate the cognitive control deficit in the NVHL rats. We observed that hippocampal

and neocortical LFPs in the NVHL rats sometimes resembled spike-wave discharges (SWDs), a sign of network hyperexcitability. We hypothesized that these aberrant neural activities arise from an excitation–inhibition imbalance that may underlie abnormal neural synchrony in the NVHL rats. Since SWD in epilepsy patients is treated with ethosuximide (2-ethyl-2-methylsuccinimide), an anticonvulsant that acts by blocking T-type Ca2<sup>+</sup> channels, we studied whether ethosuximide would attenuate the abnormal SWD-like synchrony as well as impaired cognitive behavior and the associated neural synchrony that we have previously described in the NVHL model (12). We stress that our goal is to test the prediction of the discoordination hypothesis, not to suggest the use of ethosuximide as an antipsychotic. For comparison, we also examined the effects of the atypical antipsychotic olanzapine on neural synchrony and cognitive abnormalities in NVHL rats. Given that antipsychotics are not known to improve cognitive symptoms, the hypothesis predicts that olanzapine will not correct either the neural synchrony or cognitive abnormalities in NVHL rats.

We report that in NVHL rats, ethosuximide treatment attenuated the abnormal spike-wave activity, and most importantly, ethosuximide normalized the cognition-associated synchrony of theta and beta oscillations between the two hippocampi, and improved cognitive control in the place avoidance task. Olanzapine was less effective at reducing the abnormal spike-wave activity and normalizing the synchrony of oscillations between the two hippocampi and it did not improve cognitive control in NVHL rats, although it attenuated hyperlocomotion during the cognitive testing. These data are consistent with the discoordination hypothesis, and support the idea that drugs targeting the abnormal neural synchrony may be sufficient to rescue the cognitive deficits in schizophrenia-related neurodevelopmental animal models.

## **MATERIALS AND METHODS**

All procedures were in compliance with National Institutes of Health guidelines and were approved by Downstate Medical Center's Institutional Animal Care and Use Committee.

#### **NEONATAL VENTRAL HIPPOCAMPAL LESION**

The procedure followed the manual provided by Lipska and Weinberger. We used the exact protocol that is published in our prior work (12). Briefly, timed-pregnant (13 or 14 days in gestation) female Long-Evans rats were obtained from Charles River Laboratories (Wilmington, MA, USA). Pups were born at the Downstate animal facility. On postnatal day 7 (P7), male pups were anesthetized by hypothermia. Bilateral puncture holes (relative to bregma AP: −3.0 mm, ML: ±3.5 mm) were made in the skull with a 30-Ga injection needle. A bilateral infusion (0.3µl/side) of ibotenic acid solution (10µg/µl) was delivered to each ventral hippocampus (relative to the skull surface DV: −5.0 mm). Control animals were treated identically except saline was injected instead of ibotenic acid. A subset of the NVHL and control rats were given drug treatments as described below. The behavioral and neural synchrony measurements from these rats were compared to untreated NVHL and control rats, the data from which were reported previously (12), although all data collection was concurrent amongst the different treatment groups.

## **DRUGS**

Ibotenic acid, purchased from Sigma (St. Louis, MO, USA), was dissolved in artificial cerebrospinal fluid (ACSF) to a final concentration of 10µg/µl, pH 7.6. Ethosuximide, purchased from Sigma (St. Louis, MO, USA), was dissolved in sterile 0.9% NaCl and a concentration of 100 mg/kg was injected i.p. Olanzapine, purchased from Toronto Research Chemicals Inc. (Toronto, ON, Canada), was dissolved in dimethyl sulfoxide (DMSO) and 1 mg/kg was injected i.p. Both ethosuximide and olanzapine were administered just prior to the start of behavioral training (see below) and based on published work, they were expected to have been bioactive for several hours, beyond the duration of behavioral training.

## **TWO-FRAME AVOIDANCE TASK**

Five adult rats (P65) per group were trained on the two-frame task. The rat was placed on a circular 82 cm diameter, stainless steel rotating arena (Bio-Signal Group Corp., Brooklyn, NY, USA). The rat's position was tracked using an overhead camera and software (Tracker, Bio-Signal Group Corp., Brooklyn, NY, USA). All animals were habituated to the rotating arena the day before the start of active place avoidance training during two 10-min sessions without any shocks. On the following training trials, whenever the rat on the rotating arena entered a 60° shock zone that was defined in room coordinates, a mild constant current (60 Hz, <0.4 mA, 500 ms) foot shock was delivered. Foot shock was repeated every 1.5 s until the rat left the shock zone. Each place avoidance trial was 10 min and the interval between the trials was at least 10 min. A total of eight training trials were given on the training day. The drugs were injected i.p. just prior to the first avoidance training trial. Avoidance was measured by counting the number of times the rat entered the shock zone.

## **LOCAL FIELD POTENTIAL AND ELECTROENCEPHALOGRAM RECORDINGS**

All rats were housed individually, exposed to a 12-h light/dark circadian cycle, and had *ad libitum* access to water and food. When rats were P55-57, electrodes were surgically implanted bilaterally. Rats were given post-operative Ketofen (5 mg/kg) for 2 days after surgery and allowed to recover from the surgical procedure for 7 days. LFPs in the dorsal hippocampus (AP: −4 mm; ML: ±2.5 mm; DV: 3 mm) and in the medial prefrontal cortex (mPFC; AP:+3 mm; ML:±1 mm; DV: 4 mm) were recorded by implanting 75-µm Nichrome wires that were attached to a Millmax connector. The epidural cortical electroencephalogram (EEG) was recorded with a screw electrode in the frontal and parietal bones of each hemisphere. All electrodes were referred to an electrode implanted in the cerebellar white matter (AP: −10 mm; ML: +2 mm; DV: 3 mm), which seemed to be normal in NVHL rats. Recordings were made with a wireless digital telemetry system (Bio-Signal Group, Brooklyn, NY, USA). Recordings were made while the rats were performing the two-frame avoidance task and in the home-cage (for 2–3 h) during the light cycle.

The signals at the electrode connector were amplified (300 times), low-pass filtered (6 kHz) and digitized (24-bits, 12 kHz using delta-sigma analog-digital convertors). The digital signals were transmitted wirelessly to a recording system (dacqUSB, Axona Ltd., St. Albans, UK) for band-pass filtering (1–500 Hz), digital amplification and down-sampling (16-bits, 2000 Hz) using digital signal processors. The digital electrophysiological data were stored on computer hard drives for off-line analysis. The channels were marked if they had saturated signals and were excluded from further analysis if the proportion was substantial. At the end of training and recording, the rats were trans-cardially perfused and the brains removed for histological verification of the electrode locations.

## **POWER SPECTRA**

Power spectra measure the amplitude of frequency-specific oscillations that arise when spatially patterned synaptic potentials are coincident in time. These oscillations are locally generated phenomena. As such, an oscillation itself may not be a strong indication of neural synchrony as it relates to the concept of neural coordination (1, 27). Instead, power spectra are an estimate of the amount and magnitude of patterns of oscillation-generating synaptic potentials that are the fundamental components of neural coordinating processes. The patterns of synaptic potentials must themselves be temporally and spatially coordinated within and between neural networks to mediate the neural computations and information processing upon which cognition is based (29).

To compute power spectra, LFP signals were first normalized by dividing the signal by its RMS value. This step minimized electrode-specific signal properties caused by variability in electrode location or impedance. The signal was processed using 5 s long non-overlapping windows. For each window power spectral density (PSD) was computed using the "pwelch" Matlab function. The spectral peak at 60 Hz was removed and the PSD was interpolated between 58 and 62 Hz. The resulting PSD was normalized by dividing it by its sum so the total power was equal to 1. Based on the spectral profile, we then defined bands of interest for frequency-specific phase-locking analysis. We excluded the delta band (0–3 Hz) because of its attenuation by the hardware DC-removal filter (**Figure 1**). We defined the theta band as 5–15 Hz, beta as 20–30 Hz in order to exclude the strong second theta harmonic (**Figure 1**), slow gamma 30–55 Hz and fast gamma as 65–100 Hz (excluding the 60-Hz region). **Figure 1** shows that the left and right hippocampal spectra are similar, as expected, and that the spectra are attenuated in NVHL rats. Both ethosuximide and olanzapine only caused a marginal increase in the broadband spectral power. This confirmation of stable spectral power provides the starting point for the present study, which focuses on measuring how these component oscillations are coordinated between brain regions.

## **PHASE-LOCKING VALUE**

The coordination of neural signals at a pair of electrodes was estimated by computing the phase-locking value (PLV). Eight LFP channel pairs were selected for PLV analysis. They were the leftand right-hippocampi; left- and right-mPFC; left-hippocampus and left-mPFC; right-hippocampus and right-mPFC, as well as the inter-hemispheric and associational epidural electrode pairs. The PLV at a pair of electrodes was computed using custom software written in Matlab, based on a published algorithm (31). The phase of a signal at time *t* and sample *n* φ(*t*, *n*) was obtained by filtering the signal with a narrow-band finite input response (FIR)

filter using a zero phase-shift filtering algorithm followed by the Hilbert Transform. The filters were designed using the Matlab filter design toolbox. Filters in the range 0–100 Hz had fixed bandwidth of 5 Hz. For bands of interest extending over several 5 Hz bands the PLV across all such bands were averaged. Given a pair of LFP signals of *N* samples, the PLV was defined as follows:

$$\text{PLV} = \frac{1}{N} \left| \sum\_{n=0}^{N} \exp\left[i\theta\left(t, n\right)\right] \right|,$$

where θ(*t*, *n*) is the phase difference φ1(*t*, *n*) − φ2(*t*, *n*) between the two signals at time *t* sample *n*; *N* is the total number of samples; and *i* is the imaginary unit. Prior work showed that the intracranial electrode-pair specific PLVs at the four (theta, beta, slow gamma, fast gamma) frequency bands were different between NVHL and control rats (12). We made these comparisons between the NVHL and control treatment groups to determine if the drug treatment normalized neural synchrony between the particular electrode pairs. In addition, the PLVs for the full set of electrode pairs were studied to look for neural synchrony patterns that distinguished the NVHL and control and the drug-treatment groups. The set of electrode-pair and frequency-specific PLVs for each recording session was treated as a vector with dimension 32 (8 electrode pairs × 4 bands). Principal component analysis (PCA) was performed to reduce the dimensionality and identify the most useful measures of synchrony. Each session vector of PLVs (eight sessions per rat) was plot in the coordinate system of the first three principal components to visualize the group patterns. The first six principal components explained over 90% of the variance and so for quantification we only considered the first six principal components. We quantified the distance between the vectors in each session group by computing the 6-D Euclidean distance between all pairs of data points within a treatment group, and also between all pairs of data points against a reference treatment group. The untreated control and untrained NVHL groups as well as the ethosuximidetreated NVHL and the olanzapine-NVHL groups each served as reference groups for the comparisons. These pair-wise group distances were compared by two-sample Student's *t*-tests that were adjusted for multiple comparisons using the Bonferroni method.

#### **AUTOMATIC DETECTION OF SPIKE-WAVE ACTIVITY**

We used an algorithm to identify activity that resembled epileptiform SWDs, but are observed in normal Long-Evans rats without seizures (32). Each channel was first examined for signal saturation and only data segments with brief (<50 ms) saturations were considered for further analysis. Brief saturations in the LFP were allowed so spike-wave complexes with a large spike component were not removed from the analysis. We manually marked epochs of spike-wave events using the EEGLAB toolbox for Matlab. These epochs were used to calculate and optimize the detection parameters for the automatic classifier of spike-wave events. The classifier is a modified version of a published algorithm (33).

The algorithm worked in several steps (**Figure 2**):


**FIGURE 2 | Spike-wave activities in NVHL and control rats**. **(A)** A segment of the cortical EEG with spike-wave activity. An automatic classifier was used to detect these events. **(B)** The first step in the classifier was to transform the recorded signal using the continuous wavelet transform (CWT). **(C)** Then a sliding variance was computed for each wavelet scale. All the variances were summed across all scales resulting in a single variance profile over time (50 ms smoothing was applied to the resulting signal). SWD candidates were detected as threshold crossings [red line in **(C)**]. Each marked segment was tested for periodic spiking activity by detecting local maxima [red dots in **(C)**]. Only the detected segments that contained at least 50% of inter-spike intervals in the range of 2–10 Hz were included in further analysis.

inter-spike intervals for each pair of sequential maxima were computed. Only those marked segments that contained at least 50% of inter-spike intervals in the range of 2–10 Hz (the typical range of SWDs) were considered as spike-wave segments.

(7) The process was repeated several times with different parameters and the classified event segments were compared with manually marked events to tune the classifier.

## **ANALYSIS OF SPIKE-WAVE EVENT FEATURES**

Only manually marked segments were included in this analysis to avoid distortion by borderline spike-wave-like events that were detected by the automatic algorithm but did not necessarily share all the features such as clear spike and wave components. The subjectively selected events were identified on the basis of the classic spike-wave morphology, by an observer blind to the origin of the EEG. Thus the manual selection of spike-wave events resulted in a dataset of events with a bias to less specificity and more selectivity.

The manually marked events were first low-pass filtered with a cut-off frequency of 100 Hz. The filter removed the high frequency components that could cause errors in feature extraction. Local maxima were then detected in the signal and for each maximum, its preceding and following local minima were found. The spike components of the events were then defined as local maxima, surrounded by local minima with a maxima–minima amplitude of at least 400µV and with a spike width (defined below) of at least 50 ms. Spike sequences (for computing the inter-spike time) were defined as continuous spikes with a frequency between 5 and 13 Hz. The feature analysis was repeated several times to optimize the parameters for the component detection.

The following features were computed for both NVHL and control rats (**Figure 3**):


## **STATISTICAL ANALYSIS**

Group differences were compared by ANOVA or Student's *t*-test. The PLV between groups and across frequency bands were compared by two-way ANOVA. Repeated-measures were not used because the data were independently and differentially filtered for computing the synchrony estimates at each frequency band. Significance was set at *p* < 0.05, and Tukey's HSD *post hoc* tests were used as appropriate. The two-sample Kolmogorov–Smirnov test was used to compare the probability distributions of spike-wave event features in the NVHL and control groups. The distributions were generated by selecting a constant subsample of spike-wave events from each animal such that each subject contributed equally to the distribution. The resulting distribution was then divided by its sum to generate probability distributions that are normalized against different numbers of subjects in the NVHL and control groups.

## **RESULTS**

## **ABNORMAL SPIKE-WAVE ACTIVITIES IN THE EEG OF NVHL RATS**

Before focusing on the main aim of this study, to investigate cognition-related inter-hippocampal phase synchrony during active place avoidance cognitive behavior, we began by monitoring the field potential recordings from NVHL and control rats as they behaved in their home-cage. In the home-cage, both groups of rats were still or moving <0.5 cm/s about 80% of the time based on estimates from concurrent video tracking. The most conspicuous feature was the occasional presence of a large amplitude series of spike-wave events. We were surprised to find

**(A)** Three spike-wave cycles are shown to illustrate the features that were measured. The analyzed features include spike duration, inter-spike time, wave duration, spike-to-valley voltage, and inter-spike time. Comparison of the spike-wave complex features in NVHL and control rats: **(B)** inter-spike time,

**(C)** spike duration, **(D)** wave duration, **(E)** spike-to-valley voltage, and **(F)** wave energy. Distributions in each feature looked distinct, except in spike duration, and this was confirmed statistically by the two-sample Kolmogorov–Smirnov test in which the null hypothesis that the control and NVHL distributions were drawn from the same underlying distribution was rejected (p < 0.05).

these events and will call them spike-wave activity. They typically occurred in trains of ~6 Hz (**Figure 2A**). These events have been observed in the cortical EEG of normal Long-Evans rats during stillness and are a sign of heightened synchrony of cortical neural activity, perhaps the rat analog of the human alpha-like mu rhythm (32, 34). We automatically detected the individual events and confirmed by visual inspection that the algorithm properly detected the events (**Figures 2B,C**). Recordings from the epidural cortical screw electrodes were used for analysis because signals from the epidural cortical electrodes showed the strongest spike-wave waveforms across all animals. We analyzed a total of nine NVHL and seven control rats during home-cage behavior to estimate the prevalence of the spike-wave activity. Each rat was recorded for 3–5 h in the home-cage. The spike-wave activity was detected in six of nine NVHL rats and three of seven control rats. The prevalence was 3% across all the NVHL rats and 2% in all the control rats. In total, 1514 spike-wave events were detected in NVHL rats and 397 spike-wave events were detected in control rats, suggesting that these events were more common in NVHL rats.

We then measured various features of the spike-wave events in both NVHL and control rats (**Figures 3A–F**). The distributions of NVHL and control groups looked distinct in all the features except in spike duration. These were confirmed statistically by the two-sample Kolmogorov–Smirnov test because the null hypothesis that the control and NVHL distributions are drawn from the same underlying distribution was rejected (*p* < 0.05) for all features. These events tended to be larger amplitude and more frequent in NVHL rats in comparison to the control rats. Since we have previously characterized a cognition-related abnormality in neural synchrony in NVHL rats, in the framework of the discoordination hypothesis, these observations of abnormal synchrony suggested to us that reducing the spike-wave activities could improve cognition and related neural activity in NVHL rats.

## **ETHOSUXIMIDE REDUCED THE PREVALENCE OF ABNORMAL SPIKE-WAVE ACTIVITIES**

We then evaluated whether the spike-wave activities could be attenuated by ethosuximide, an anticonvulsant that is an antagonist of T-type calcium channels (35). It is used mainly to treat absence seizures. Given the framework that animal models are best utilized to test hypotheses about schizophrenia and that it is unwise to assume that models themselves mimic the disease (36), we reasoned that since ethosuximide is effective at suppressing epileptiform SWDs, then it may be also be effective on the spike-wave activities in the NVHL rats. Since it might suppress this specific form of abnormal neural synchrony by rebalancing excitation and inhibition, it occurred to us that ethosuximide could also act to reduce the cognition-related abnormal neural synchrony, which according to the discoordination hypothesis would attenuate the cognitive deficit observed in these rats.

For comparison, we also chose to study the effects of olanzapine, a compound with antagonist actions at dopaminergic, muscarinic, adrenergic, GABAergic, and histaminergic receptors. It is one of the most widely used second-generation antipsychotic drugs. These drugs were designed to suppress hyperlocomotion in animal models and are prescribed to reduce the positive symptoms of psychosis, but are not considered to be effective on cognitive symptoms.

We observed that ethosuximide reduced the spike-wave activity much more than olanzapine (**Figure 4A**). There was an eightfold reduction in the prevalence of spike-wave activity after ethosuximide (100 mg/kg i.p.) treatment, while only a twofold reduction in spike-wave activity in the NVHL rats (**Figure 4B**) after a dose of olanzapine (1 mg/kg i.p.) that is sufficient to attenuate sensorimotor deficits such as hyperlocomotion in the NVHL (see **Figure 7C**below) and acute PCP animal models (37). These observations motivated us to test whether ethosuximide can attenuate the abnormal cognition-related neural synchrony that we have previously characterized in NVHL rats (12).

## **ETHOSUXIMIDE ATTENUATED ABNORMAL INTER-HIPPOCAMPAL SYNCHRONY**

We then turned to the main effort of this research. We recorded LFPs while the rats performed the two-frame active place avoidance task. We previously identified that oscillations in the theta and beta ranges was less synchronized between the left and right dorsal hippocampi in NVHL rats than in control rats and that this inter-hippocampal synchrony was correlated with effective place avoidance behavior (12). Inter-hippocampal theta and beta synchrony was weaker in the NVHL rats compared to the controls (**Figure 5A**), which was already shown in prior work (12). We administered ethosuximide and olanzapine just prior to the first place avoidance training trials in new NVHL and control groups. The difference in inter-hippocampal synchrony between the NVHL and control rats was attenuated by ethosuximide (**Figure 5B**), suggesting that ethosuximide normalized interhippocampal synchrony in NVHL rats. Following olanzapine, inter-hippocampal synchrony was different between the NVHL and control groups (**Figure 5C**). However, this difference appeared because the drug increased synchrony in the NVHL rats and decreased synchrony in the control group, especially in the theta and beta bands. Indeed, after olanzapine, synchrony in the NVHL rats was greater than in the treated control rats. Olanzapine also increased inter-hippocampal synchrony in the faster frequency gamma bands. This increase was in excess of the synchrony seen in untreated control rats, which in contrast to ethosuximide, is why olanzapine did not restore normal synchrony.

We also examined inter-hemispheric synchrony between electrodes in the mPFC (**Figure 6A**). The synchrony of LFP oscillations between the left and right-mPFC was greater for slower frequencies than faster frequencies but not different between the NVHL and control groups. After ethosuximide, the frequency dependence on the magnitude of synchrony between the left and right-mPFC was no longer observed. This was in large part because ethosuximide increased synchrony at the faster gamma frequencies, almost to the level of the slower beta frequencies. This effect was more pronounced in the NVHL animals. Olanzapine had a similar effect on synchrony between the left and right-mPFC sites such that there was no longer a detectable difference in the magnitude of synchrony in the different frequency bands and the NVHL and control groups were also indistinguishable.

Finally, we examined synchrony between the mPFC and dorsal hippocampus of one hemisphere (**Figure 6B**). Synchrony was generally less than it was between the two hippocampi or between the two mPFC. There was greater synchrony in slower frequency bands than the faster gamma bands. There was, however, no difference between NVHL and control animals. After ethosuximide, synchrony increased, especially in the faster bands, for both the NVHL and control animals. This removed the effect that synchrony was greater at slow frequencies. The NVHL and control groups were also indistinguishable after ethosuximide. Olanzapine increased synchrony in both NVHL and control rats but the increase was greater for the NVHL rats. Synchrony increased in all bands so there was no effect of the frequency band. Clearly, both ethosuximide and olanzapine altered the synchrony of oscillations in the LFPs of control and NVHL rats.

## **ETHOSUXIMIDE ATTENUATED THE COGNITIVE CONTROL IMPAIRMENT OF NVHL RATS**

We examined whether these changes in synchrony were associated with changes in behavior using the active place avoidance task variant that allows evaluation of the ability to use relevant information and ignore irrelevant information, what is known as cognitive control. We first examined the effects of the drugs on locomotor hyperactivity (**Figure 7A**), which is a feature of NVHL rats that we have previously observed during the place avoidance task (12). Indeed, hyperactivity has been widely used to model the positive symptoms of schizophrenia and the effect of olanzapine in reducing hyperactivity has been shown in a variety of schizophrenia-related animal models (38, 39). Ethosuximide significantly reduced hyperactivity in NVHL rats compared to control rats (**Figure 7B**). As expected, olanzapine also reduced hyperactivity in NVHL rats compared to controls rats, so that the two groups were indistinguishable (**Figure 7C**).

**FIGURE 6 | Inter-mPFC or inter-hippocampus-mPFC phase synchrony was not different in the NVHL and control groups during active place avoidance**. **(A)** Inter-mPFC synchrony was frequency-dependent but not different between the NVHL and control rats in the untreated groups (two-way ANOVA group: F1,32 = 0.92, p = 0.34; frequency: F3,32 = 8.97, p = 0.0001; interaction: F3,32 = 0.33, p = 0.81). The inter-mPFC synchrony was greater in the NVHL rats in the ethosuximide-treated group and the frequency dependence was lost due to increased synchrony at the gamma frequencies (middle: two-way ANOVA group: F1,32 = 5.42, p = 0.03; frequency: F3,32 = 1.41, p = 0.26; interaction: F3,32 = 0.16, p = 0.92). In the olanzapine-treated rats, the group and frequency dependencies were no longer observed because the drug increased synchrony at all frequencies we assessed (right: two-way ANOVA group: F1,28 = 1.43, p = 0.24; frequency: F3,28 = 0.40, p = 0.76;

We then examined the effects on cognitive behavior. NVHL rats were slower to learn the place avoidance (**Figure 8A**), as we have previously reported (12). Ethosuximide attenuated the deficit because under ethosuximide NVHL rats performed at a level that was statistically indistinguishable from control rats, although there was a trend for them to be slower to learn in the initial trials (**Figure 8B**). The olanzapine treatment did not improve the cognitive performance of NVHL rats, which remained impaired compared to control rats (**Figure 8C**). In fact, olanzapine seemed to slightly impair avoidance of the control rats during the initial trials. This was confirmed by comparing the three groups of control rats from **Figure 7** by two-way ANOVA (treatment: *F*2,96 = 9.70, *p* = 0.0001; trials: *F*7,96 = 33.86, *p* < 0.0001; interaction: *F*14,96 = 1.24, *p* = 0.26; *post hoc* tests: only olanzapine-treated control rats differed from ethosuximide-treated and untreated control rats).

Next we investigated the relationship between place avoidance performance and inter-hippocampal synchrony in the theta and beta bands. In the untreated NVHL and control rats, greater phaselocking at theta and beta frequencies was significantly correlated (*r*'s > 0.8; *p*'s < 0.01) with better place avoidance measured as lower total entrances (**Figure 9A**) but not with hyperactivity, measured as running speed (**Figure 9B**). The relationship between higher synchrony and lower entrances was apparent but not interaction: F3,28 = 0.11, p = 0.95) groups. **(B)** Inter-area synchrony (between mPFC and hippocampus) was frequency-dependent but not different between the NVHL and control rats in the untreated group (left: two-way ANOVA group: F1,32 = 0.43, p = 0.51; frequency: F3,32 = 5.29, p = 0.005; interaction: F3,32 = 2.05, p = 0.13). There were no group or frequency dependencies in the ethosuximide-treated rats because the drug increased synchrony at all frequencies to the level of the maximal (theta) inter-hippocampal synchrony (middle: two-way ANOVA group: F1,32 = 0.11, p = 0.74; frequency: F3,32 = 0.30, p = 0.83; interaction: F3,32 = 0.07, p = 0.97). The olanzapine-treated NVHL rats had greater synchrony at all frequencies and there was no frequency dependence because the drug also increased synchrony at all frequencies (right: two-way ANOVA group: F1,28 = 1.22, p = 0.28; frequency: F3,28 = 0.26, p = 0.85; interaction: F3,28 = 0.19, p = 0.90) groups.

significant in the ethosuximide-treated animals (theta: *r* = 0.37; *p* > 0.1; beta: *r* = 0.14; *p* > 0.1). Because performance was good in most of these rats, there was limited variability in behavior for the synchrony measures to explain. The trend of a relationship between synchrony and cognition was reversed in the olanzapinetreated rats. Greater synchrony was associated with worse performance, but this trend was not significant (theta: *r* = 0.50; *p* > 0.1; beta: *r* = 0.35; *p* > 0.1).

Finally, in the effort to describe neural coordination during the place avoidance task, we examined the full set of 32 synchrony measures (8 electrode pairs × 4 frequency bands) as a pattern to investigate if there were group specific patterns of synchrony (**Figure 10**). The 32-dimensional space defined by the synchrony vectors (**Figure 10A**) was reduced to a 3-dimensional space to visualize the data using the first three principal components (**Figure 10B**). The data formed several clusters instead of filling the parameter space, which suggests neural activity may preferentially occupy preferred locales of this parameter state space. The untreated NVHL and control animals appeared to form two clusters. The ethosuximide and olanzapine-treated groups occupied distinct regions of the space away from the untreated groups. To estimate how distinct the groups might be, we computed the average distance amongst all pairs of vectors within a treatment group. The first six principal components accounted

F7,64 = 0.48, p = 0.85); inset: average speed: t <sup>8</sup> = 0.30, p = 0.77).

for over 90% of the variance in the data set (**Figure 10C**), so we computed the pair-wise distances as the 6-D Euclidean distance. The average distance from the untreated NVHL group was significantly greater for each group than it was for the vectors in the untreated NVHL group (**Figure 10D**). Similarly, the

distances from the untreated control vectors were significantly greater for all other groups except the control group treated with ethosuximide, suggesting that ethosuximide did not significantly

change control synchrony patterns. Because place avoidance behavior in the NVHL group was improved by ethosuximide, we estimated the distance between this group and the others. The distance to the untreated control group was no different than the distance within the NVHL-ethosuximide group, suggesting that the two groups were relatively similar. The complementary estimate of distance, this time to the olanzapine–NVHL group had a different answer. All groups had greater distances to the olanzapine–NVHL group than the distances amongst the group itself, consistent with the synchrony pattern caused by olanzapine being unique, rather than normalizing.

## **DISCUSSION**

## **SUMMARY**

The main finding of this study is that a drug that normalizes aberrant cognition-related neural synchrony in NVHL rats can also attenuate the cognitive control impairment in this schizophreniarelated neurodevelopmental model of mental dysfunction. This work demonstrates that the pattern of abnormal cognition-related neural synchrony is acutely normalized by ethosuximide, and furthermore shows that the synchrony alterations were associated with improved cognitive performance after ethosuximide but not olanzapine, despite both drugs being effective at reducing hyperactive locomotion in the NVHL model. Although olanzapine increased theta and beta synchrony between the two hippocampi, which is correlated with task performance (**Figure 8**), the drug also increased gamma synchrony beyond baseline levels, which may itself be associated with poor place avoidance performance in mutant mice (40, 41). Note that under ethosuximide, and also under olanzapine, the overall neural synchrony pattern in NVHL rats was distinct from the untreated control pattern, but the pattern after ethosuximide was closer to the control pattern than the pattern after olanzapine (**Figure 10D**) and perhaps this is one of the reasons that place avoidance performance was improved by ethosuximide but not olanzapine.

We have used the NVHL model to examine abnormal synchrony and cognitive impairment. Although NVHL rats are not a model of schizophrenia *per se*, these animals are an experimental

**Frontiers in Psychiatry** | Schizophrenia February 2014 | Volume 5 | Article 15 |

corresponds mostly to frequency band-non-specific phase synchrony between the contralateral cortical electrode pairs as well as ipsilateral cortical electrode pairs, which are measures of global cortical phase synchrony. PC #3 corresponds mostly to frequency band-non-specific phase synchrony between the left and right-hippocampi as well as between the left and right prefrontal cortex, which is an estimate of region-specific inter-hemispheric synchrony. PC #6 corresponds to frequency-specific synchrony in the beta,

distinctiveness. The distances between all pairs of vectors were computed within a treatment group and between all treatment groups and a specific treatment group to estimate similarity/distinctiveness of the treatment groups. The average distances are given to the untreated NVHL group (top

ethosuximide-treated NVHL group (third row), and to the olanzapine-treated

row), to the untreated control group (second row), to the

NVHL group (bottom row).

model with neural synchrony and related cognitive abnormalities, which allows a prediction of the discoordination hypothesis to be tested. This unifying hypothesis was proposed to account for the schizophrenia syndrome independent of etiology (1, 2, 12, 27, 28). This view acknowledges that schizophrenia may turn out to be heterogeneous and that multiple factors contribute, including genetic alterations,infectious,toxic, and stressful events (42,43). Whatever the etiology, the discoordination hypothesis asserts that disruption of the physiological synchronization of neural activities produces cognitive abnormalities in schizophrenia, despite otherwise seemingly normal expression of unitary physiological phenomena. The hypothesis is based specifically on the physiological coordination of neural activities, defined as the set of neural processes that control the timing of spike discharge across ensembles of neurons (1, 29, 44, 45). Our group has demonstrated this neural coordination during cognitive control in active place avoidance tasks (46, 47). The current findings support the discoordination hypothesis by demonstrating that correcting the neural synchrony abnormality in NVHL rats can be associated with cognitive improvement, which was the case with ethosuximide but not with olanzapine. Olanzapine did not normalize either neural synchrony or cognitive behavior (**Figures 4**, **5**, and **8**). In further support of the hypothesis, olanzapine caused abnormal neural synchrony in control rats, which seemed to worsen the place avoidance.

### **SPIKE-WAVE ACTIVITY**

A secondary finding in this study is that spike-wave like activity was observed in the epidural cortical EEG of both NVHL and control rats during stillness in the home-cage and only rarely during active place avoidance behavior, consistent with prior observations (32, 34). This observation may have implications for how deficits in NVHL animals are interpreted when those deficits are characterized in tasks during which there is behavioral stillness, like pre-pulse inhibition, inhibitory avoidance, and food-motivated radial-arm mazes. Complex interactions of inhibitory and excitatory systems between the thalamus and cortical structures generate a variety of normal and abnormal rhythmic states including spindle waves and SWDs during absence seizures. Spindle waves are characterized by 1–4 s periods of 6–14 Hz oscillations during normal sleep (48). The characteristic frequency of the SWD is 3–4 Hz in humans (49) and is normally higher (4–12 Hz) in rodents (50, 51). Our analysis of cortical spike-wave activity shows that the features of the spike-wave activity in NVHL and control rats are different (**Figure 2**). The duration of the activity was greater in NVHL rats compared to control rats, further suggesting that this activity in the NVHL rats may be of a different neural origin than it is in control rats. In NVHL rats it may reflect a form of pathological network synchrony and even a sign of epileptiform propensity (52).

Indeed, during several hours of continuous monitoring, we have observed rare, but nonetheless *bona fide* epileptiform discharges and behavioral seizures in NVHL rats (52). The NHVL rats may have increased susceptibility to seizures due to dysregulation of the GABA system (53, 54). There is widespread reduction of GAD67 expression (54, 55), and decreased expression of PV interneurons in the prefrontal cortex and the hippocampus (54) in these rats. There is also increased firing in pyramidal neurons in the NVHL rats. The mPFC pyramidal neurons of NVHL rats respond to stimulation of the ventral tegmental area, the origin of dopamine projections to the PFC, with an increase in firing rates (56) instead of the normal decrease (57), indicating hyperexcitability. It has been shown that NVHL rats developed epileptiform SWDs at a twofold lower dose of pentylenetetrazol (PTZ, a GABA<sup>A</sup> antagonist) than sham-operated rats, further evidence that the NVHL rats have increased susceptibility to absence-like seizures (54). Indeed, PD7 injections into ventral hippocampus with kainate, an excitotoxin with a different mechanism of action than ibotenate (58), is used as a model of temporal lobe epilepsy (59). Perturbations of the GABA system, together with the aberrant response reflecting the inability of dopamine to activate interneurons, may result in enhanced cortical excitability and abnormal spike-wave activity in NVHL rats.

## **ETHOSUXIMIDE TREATMENT**

We now discuss the effects of ethosuximide in the contexts of its pharmacological actions, typical use as well as the relevance of the findings with ethosuximide to understanding the cognitive impairments in the NVHL rat model and mental illness such as schizophrenia. To be clear, neither the findings nor the discussion that follows suggest that ethosuximide might be an effective treatment for schizophrenia. Rather, the procognitive and normalizing neural coordination effects of ethosuximide in the NVHL model provides new evidence that attenuating neural synchrony deficits can itself be procognitive.

Ethosuximide is used extensively to treat absence (petit mal) seizures (60–62). The mechanism of action is to block lowthreshold Ca2<sup>+</sup> currents in the T-type Ca2<sup>+</sup> channels and enhance GABAergic tone (63, 64). We observed that ethosuximide reduced spike-wave activity in NVHL rats, indicating it is effective on abnormal synchrony in NVHL rats during home-cage behaviors (**Figure 3**). The drug also increased inter-hippocampal synchrony in NVHL rats so that neural coordination between the two hippocampi approached normal levels in the theta- and betafrequency bands. Nonetheless, taken together, after ethosuximide, the overall pattern of neural synchrony in the NVHL rats was modified but the pattern of synchronies was not restored to the pattern of the untreated controls. Indeed, ethosuximide also changed the synchrony pattern in the controls animals too, but these changes were not associated with cognitive disability.

The effectiveness of ethosuximide may be contemplated in relation to the fact that the synchronization of oscillations reflects the temporally precise interaction of neural activities (65). Such coordinated interactions result from an appropriate balance between GABA-mediated inhibition and glutamate-mediated excitation, which may be unbalanced in adult NVHL rats (56). Alterations in one or both of these systems can result in abnormal network synchrony. Schizophrenia has been associated with dysregulation of cortical GABAergic neurotransmission (66) and abnormalities in NMDA-receptor mediated neurotransmission (67). There are significantly decreased levels of GABAergic neuronal markers like parvalbumin mRNA in the prefrontal cortex (68, 69) and in the dorsolateral prefrontal cortex of schizophrenia patients, decreased levels have been reported of glutamic acid decarboxylase (GAD67) mRNA, the GABA-synthesizing enzyme (70,71). Indeed, NMDA-receptor hypofunction in schizophrenia (72, 73) is now thought to primarily affect GABA inhibition by reducing GABAergic tone. Thus dysregulation of excitatory and inhibitory neurotransmission contribute to abnormalities in the coordination of neural activities, which according to the hypothesis, is the basis for the impaired cognitive functions that are central to schizophrenia and a logical target for developing cognition promoting therapies. It was recently proposed that aberrant low frequency delta/theta oscillations emerge as a consequence of NMDA-receptor hypofunction (3, 74). The reduced excitation would deinactivate T-type Ca2<sup>+</sup> channels and exaggerate bursting, especially in thalamic neurons, which may be the origin of the low frequency thalamocortical oscillations that could disrupt cortical function in schizophrenia (75). In this context, given ethosuximide's mechanism of action, it is perhaps not surprising that the drug both normalized neural synchrony and improved cognition in NVHL rats. Ethosuximide reduces excitability primarily by blocking activity-dependent lowthreshold Ca2<sup>+</sup> currents as well as by acting as a partial agonist at the picrotoxin GABA-blocking receptor (64).

In the absence of the forgoing discussion, it may seem surprising to some readers that ethosuximide can improve interhippocampal synchrony and place avoidance behavior. Some readers may have assumed that because ethosuximide is antiepileptic it must decrease neural synchrony because epileptiform activity is thought of as excessive synchrony. According to these opinions, ethosuximide would decrease synchrony, not increase it, and thus worsen place avoidance behavior in control as well as NVHL rats. As we have pointed out in the Section "Introduction," epileptiform activity is not simply excessive synchrony, and in the NVHL rat, we have measured the evidence of widespread decreases of synchrony between brain areas, resembling functional disconnection (52). Embedded in this loosening of neural coordination, perhaps by excitation–inhibition uncoupling, we also observed that some brain areas can exhibit relatively increased synchrony. Thus, there are alternative conceptualizations of the ethosuximide effects but they are not supported by the data on the subject in general or on NVHL rats in particular.

Ethosuximide treatment also reduced hyperactivity in NVHL rats, which in animal models, is widely studied as a behavioral analog of the positive symptoms of schizophrenia, although the reduction was not as effective as after olanzapine. Blocking T-type Ca2<sup>+</sup> channels can produce antipsychotic effects in rats by attenuating the psychomotor effects of both the NMDA antagonist MK-801 and amphetamine, the dopaminergic psychostimulant (76). T-type Ca2<sup>+</sup> channels are widely expressed in the brain, including areas such as the thalamus, the prefrontal cortex, the hippocampus, and the nucleus accumbens (77), which are regions that have been reported to function abnormally in schizophrenia (74, 78, 79). Interestingly, a number of clinically validated antipsychotics, including haloperidol, pimozide, flunarizine, and clozapine, are potent T-type channel antagonists (76, 80–82). A recent genome-wide study of single-nucleotide polymorphisms (SNP) found that variations in calcium channel genes are associated with schizophrenia as well as a range of major psychiatric disorders (83). These data implicate calcium channels in the pathophysiology of schizophrenia, and while calcium channel abnormality may help to account for why ethosuximide was effective on NVHL rats, it is also possible that ethosuximide was effective because of its fundamental effects on excitation–inhibition balance through the action on otherwise normal calcium channels. Nonetheless, the present work does not imply and we do not propose to use ethosuximide as a treatment for cognitive deficits in schizophrenia. Rather the present findings demonstrate that procognitive effects may be possible if treatments can normalize the coordination of neural activity between cognitive processing centers like the two hippocampi.

## **OLANZAPINE TREATMENT**

Second-generation (atypical) antipsychotic medications, including olanzapine, are effective in treating psychosis, hallucinations, and delusions (84) but ineffective in treating the cognitive impairments that strongly debilitate schizophrenia patients (85). Olanzapine, a thienobenzodiazepine derivative, is classified as a multi-acting receptor-targeted antipsychotic, showing high affinity for dopaminergic, serotonergic, cholinergic, histaminergic, and muscarinic receptors (86, 87). As expected, we observed that olanzapine reduced hyperactivity in NVHL rats to control levels during the place avoidance task, indicating we used an effective antipsychotic dose, at least by traditional measures. The dose was however, ineffective at improving cognitive performance of NVHL rats in the two-frame task. This is of course consistent with the drug's lack of effect on cognitive symptoms in patients.

Olanzapine treatment in NVHL and control rats showed a reversed effect on neural synchrony. Olanzapine increased interhippocampal synchrony in NVHL rats to the level of untreated controls in the theta and beta bands and in excess of the control levels in the gamma bands. In contrast, the drug decreased inter-hippocampal synchrony in control rats. The increased synchrony in olanzapine-treated NVHL rats was not associated with improved cognitive performance in the NVHL rats, which would be at odds with the discoordination hypothesis, except that after olanzapine, the gamma synchrony was excessive, indicating that olanzapine changed neural coordination, but as demonstrated with the multidimensional synchrony analysis (**Figure 10**), the synchrony changes did not normalize neural synchrony in NVHL rats. Conversely, but also in support of the discoordination hypothesis, the decreased synchrony in olanzapine-treated control rats was associated with a higher number of errors in the two-frame task when compared to untreated control rats, which is consistent with the observation that the drug changed the neural synchrony pattern in control animals to a pattern that is substantially distinct (**Figures 10B,D**) from the untreated control pattern. It is well documented that antipsychotic drugs cause EEG abnormalities associated with general slowing of background activity, an increase in paroxysmal theta or delta activity and the development of epileptiform discharges (88). In patients on olanzapine, there is an increased diffuse and intermittent slowing of the EEG (89, 90) and an increased high risk of EEG abnormalities such as theta and delta slowing, sharp waves or phase reversal, and/or spikewave activities (88, 91). Healthy subjects on olanzapine showed an increased power in the theta band and a decrease in the beta band (92). In summary, olanzapine normalized inter-hippocampal synchrony in the theta and beta bands but it also caused abnormal synchrony changes in both the NVHL and control rats that were not

associated with improved cognitive performance in the two-frame task.

## **CONCLUSION**

The findings reported here are consistent with the main assertion and prediction of the discoordination hypothesis that abnormal neural synchrony during cognitive effort is associated with impaired cognition and that restoring normal synchrony will promote cognition. As shown here, targeting the pathophysiology of abnormal neural coordination, regardless of the etiology, may be both a rational and effective program for developing a muchneeded generation of procognitive treatments for mental illness in general and schizophrenia in particular.

## **AUTHOR CONTRIBUTIONS**

Heekyung Lee collected and analyzed data, Dino Dvorak analyzed data, Heekyung Lee and André A. Fenton wrote the manuscript, and André A. Fenton planned and organized the research.

## **ACKNOWLEDGMENTS**

This work was supported by NIMH grant R01MH084038.

## **REFERENCES**


hippocampal-lesioned rats. *Int J Neuropsychopharmacol* (2009) **12**:1097–110. doi:10.1017/S1461145709009985


and reduces stimulant-induced glutamate release in the nucleus accumbens of rats. *Neuropharmacology* (2010) **62**(3):1413–21. doi:10.1016/j.neuropharm. 2010.11.015


neurobiology of schizophrenia. *CNS Drugs* (2006) **20**:389–409. doi:10.2165/ 00023210-200620050-00004


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 September 2013; accepted: 29 January 2014; published online: 14 February 2014.*

*Citation: Lee H, Dvorak D and Fenton AA (2014) Targeting neural synchrony deficits is sufficient to improve cognition in a schizophrenia-related neurodevelopmental model. Front. Psychiatry 5:15. doi: 10.3389/fpsyt.2014.00015*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2014 Lee, Dvorak and Fenton. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

REVIEW ARTICLE published: 18 December 2013 doi: 10.3389/fpsyt.2013.00171

## Mismatch negativity: translating the potential

## **Juanita Todd1,2,3,4\*, Lauren Harms 1,2,3,4, Ulrich Schall 2,3,4,5 and Patricia T. Michie1,2,3,4**

<sup>1</sup> School of Psychology, University of Newcastle, Callaghan, NSW, Australia

<sup>2</sup> Priority Research Centre for Brain and Mental Health, University of Newcastle, Callaghan, NSW, Australia

<sup>3</sup> Schizophrenia Research Institute, Darlinghurst, NSW, Australia

<sup>4</sup> Hunter Medical Research Institute, Newcastle, NSW, Australia

<sup>5</sup> School of Medicine and Public Health, University of Newcastle, Callaghan, NSW, Australia

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Philip R. Corlett, Yale School of Medicine, USA Karsten Heekeren, University Hospital for Psychiatry Zurich, Switzerland Andreea Oliviana Diaconescu, University of Zurich, Switzerland; ETH Zurich, Switzerland

#### **\*Correspondence:**

Juanita Todd, School of Psychology, University of Newcastle, University Drive, Callaghan, NSW 2308, Australia e-mail: juanita.todd@ newcastle.edu.au

The mismatch negativity (MMN) component of the auditory event-related potential has become a valuable tool in cognitive neuroscience. Its reduced size in persons with schizophrenia is of unknown origin but theories proposed include links to problems in experiencedependent plasticity reliant on N-methyl-d-aspartate glutamate receptors. In this review we address the utility of this tool in revealing the nature and time course of problems in perceptual inference in this illness together with its potential for use in translational research testing animal models of schizophrenia-related phenotypes. Specifically, we review the reasons for interest in MMN in schizophrenia, issues pertaining to the measurement of MMN, its use as a vulnerability index for the development of schizophrenia, the pharmacological sensitivity of MMN and the progress in developing animal models of MMN. Within this process we highlight the challenges posed by knowledge gaps pertaining to the tool and the pharmacology of the underlying system.

**Keywords: mismatch negativity, auditory event-related potential, schizophrenia, NMDA, synaptic plasticity, NMDAR**

With more than 150 papers discussing or reporting smaller mismatch negativity (MMN) amplitude in schizophrenia, changes in this component of the auditory event-related potential (ERP) are now recognized as one of the most replicable electrophysiological abnormalities in this group. This review begins with an explanation of why this finding holds so much potential as a tool in the study of biological changes associated with the schizophrenia illness. However, the sections that follow expose the many challenges to the endeavor to translate this research – both in terms of understanding the meaning, relevance, and cause of smaller MMN amplitude and in terms of building animal models that can provide insight into etiology.

## **THE POTENTIAL – WHY IS THERE SO MUCH INTEREST IN MMN IN SCHIZOPHRENIA?**

Auditory MMN is evident in scalp-recorded evoked potentials when an unexpected event or sound transition occurs in a regular repeating pattern (1). MMN is not a response to novelty *per se* but rather to how unlikely a particular sound transition is given a preceding sequence (2). It is therefore a very context-dependent evoked potential that only occurs when a pre-existing *predictionmodel* exists specifying the most likely sound transitions in the present environment.

The nature of sound sequences used to elicit and study MMN range from very simple to highly complex. The vast majority of studies in schizophrenia employ the former in which a regular repeating identical sound occurs with high probability (the *standard*) and a physically deviant sound (the *deviant*) interrupts these repetitious trains on rare occasions (estimates suggest max *p* ≤ 0.30) (3). Sequences of this kind, known as oddball sequences, promote the formation of a prediction-model specifying that

acoustic input is best explained by standard-to-standard transitions and the rare occurrence of a standard-to-deviant transition is generally used to index MMN (although note that deviant-tostandard transitions also elicit a smaller MMN-like response). Regularities are implicitly learned as MMN does not require attention to sounds: sound sequences are usually presented via headphones to participants who are asked to ignore sounds and direct attention to an alternate task (3).

The classical derivation of MMN involves a deviant-minusstandard difference waveform with MMN quantified as the most negative peak evident between ~100 and 250 ms following the point of deviance (4). The quintessential finding in studies of this kind is that the averaged response to standard stimuli is similar in schizophrenia and matched control groups but the response to the deviant is significantly smaller resulting in smaller MMN in the difference waveform (5).

The major cortical sources contributing to scalp-recorded MMN are located bilaterally in primary and secondary auditory cortices with the precise locations dependent upon the sound characteristics (6). Intra-cortical recordings in primates suggest that the additional negativity in scalp-recorded response to deviations has its origin in lamina II/III of the auditory cortex (7). The neural mechanisms contributing to the difference in response to predicted versus deviating sound are the topic of considerable debate. There is general acceptance that processes such as neuronal adaptation make an important contribution; that is, the regular stimulation of the same afferent pathways will result in adaptation which reduces the response to sounds matching these properties, while in contrast deviating tones that are physically different will stimulate new afferent pathways (8–10). However, both sophisticated sequence designs and computational modeling

support the existence of additional processes subserving prediction. Although some of these will be reviewed in more detail below (see The Measurement of MMN – Is there an Optimal Paradigm with Which to Study MMN in Schizophrenia?), extensive discussions can be found elsewhere (2, 3, 11–13). In summary, there is accumulating evidence that the altered responsiveness to sounds (reduced to expected and sensitized to deviant) cannot be explained by neuronal adaptation alone. Computational models suggest that predictions are expressed in the input from higher to lower levels within a hierarchical network (11, 12, 14). Changed response to sound therefore becomes a function of both neuronal adaptation and top-down inputs to auditory cortex (e.g., from secondary to primary cortex and from prefrontal to secondary cortex) that modify responsiveness to sound reflecting the predicted continuation of a learned pattern.

A network-level appreciation of MMN generation is valuable because it emphasizes that understanding the smaller MMN amplitude in schizophrenia could require consideration of both cause and consequence, acknowledging the functional role of the signal. According to free-energy conceptualizations of brain function, prediction-modeling (the use of regularity to extrapolate patterns) provides a regulatory service allowing the brain to conserve energy (15). Reduced response to the predicted causes of sensory stimulation, such as diminished response to standard input, reserves resources for events that contradict expectations and that may signal important changes in the environment prompting new learning. MMN signals a *prediction-error* indicating that the model presently influencing cortical response has failed to account for the current sensory event. The error communicates a need to adjust the current model to facilitate more accurate predictions, and simultaneously alerts the system to the possible importance of the eliciting event (15, 16). The error is communicated upward (via feed-forward connections) within a hierarchical network which enables future predictions to be weighted by error frequency (17). When a model is highly reinforced (low error frequency) predictions specified in top-down connections are weighted strongly and errors elicit large MMN. Conversely, when error frequency is higher the weighting on top-down predictions is reduced and MMN elicited by errors is smaller in amplitude. This concept has elsewhere been described as *predictive confidence* in a model which is of course inversely related to the probability of deviations (2). MMN amplitude therefore reflects quantification of confidence that the eliciting event violates a contextual regularity. Large MMN can trigger an orienting response with consequences for performance of concurrent tasks (18–23). These observations demonstrate the way in which a deviation from predictions can draw upon resources engaged in other activities.

In schizophrenia, MMN as a process appears to be largely intact in that it obeys similar principles to those evident in healthy comparison groups – MMN is larger for deviations that are rarer and/or more physically different from the current prediction-model (24). However, MMN amplitude reaches asymptote at a lower amplitude resulting in the most pronounced group differences generally being observed where MMN is very large in controls (25–27). Understanding the functional relevance of MMN amplitude is pivotal to understanding the potential impact of smaller MMN in schizophrenia. It has been proposed that MMN amplitude in part communicates the change in cortical responsiveness required to accommodate the new event into prediction-models (16). To the extent that this is true, the lower plateau in MMN amplitude in schizophrenia means that the adjustment to predictions prompted by larger or rare deviations may be equivalent to that for smaller or more frequent deviations. So if the size of the error-signal itself influences the adjustment in predictions, then equivalence of error signals across a range of deviant probabilities or physical differences (25–27) would perpetuate insensitivity to these different contexts.

This circular challenge becomes particularly relevant when considering how smaller MMN is related to biological changes associated with the illness,such as gray matter changes and involvement of the *N*-methyl-d-aspartate (NMDA) glutamate receptor (NMDAGluR) system. A key biological change observed in schizophrenia is loss of cortical gray matter volume (28–35). Within the auditory system this volume loss is due to a combination of reduced pyramidal somal cell size, reduced dendritic spine density, and (a correlated) reduction in axon terminal density in lamina III (36–39). Molecular-level studies in schizophrenia therefore place a core pathological process in the very same cortical layer implicated in the generation of MMN. The projected functional consequence of reduced volume in this cell layer is a net "diminished excitatory synaptic connectivity" consistent with an impact on the spread of activation [(36), p. 384]. The generation of MMN includes spreading activation across lamina III within A1 (Heschyl's gyrus or primary auditory cortex) but also importantly from A1 to STG [superior temporal gyrus or auditory association cortex (11, 17)]. So in terms of MMN generation, this pathology in auditory cortical regions is expected to lead to smaller MMN generated in response to error and/or a lower limit to MMN size [consistent with experimental observations reviewed (36, 38)]. This predicted consequence is also consistent with observations that in schizophrenia (but not in matched controls) limited MMN size correlates with reduced gray matter volume in auditory regions (35, 40) and in the only published longitudinal study, progressive loss of gray matter in auditory cortices correlated with progressive reduction in MMN (40).

The etiology of this auditory cortical pathology in schizophrenia is unclear but two hypotheses put forward include a developmental origin and/or failure in sustained support of structural integrity (36). The first supposes that there is an over-elimination of excitatory synapses in lamina III during development. Since this process is very protracted [continuing into the third decade in auditory cortices (41)], it maps well onto data showing progressive reductions in STG gray matter volume in schizophrenia over this period (42). However, experience-dependent plasticity modifies dendritic spine structure throughout life (through long term potentiation and depression) and this activity has a stabilizing effect on these structural elements. The same pathology could therefore also arise (uniquely or in conjunction with developmental over-pruning) through failure in the factors that should support structural integrity.

The second and not unrelated biological factor with relevance for smaller MMN in schizophrenia is the impairment in NMDAGluR system function proposed in glutamatergic models of schizophrenia which have become increasingly accepted as etiopathological models of schizophrenia. These models arose from observations that the drug phencyclidine (PCP) has (i) psychotomimetic effects and (ii) non-competitively blocks NMDAGluRs [for detailed reviews see Ref. (43–45)]. NMDAGluRs have a key role to play in experience-dependent synaptic plasticity, and in particular long term potentiation and depression. Importantly (as reviewed later in Section "How do Pharmacological Manipulations Alter the MMN Process?"), MMN is reduced in healthy individuals administered an NMDAGluR antagonist, such as ketamine. So the question that emerges is this: is smaller MMN in schizophrenia a consequence of gray matter reduction and/or NMDAGluR hypofunction in auditory cortex, or could the process of prediction-modeling in audition provide clues as to why structural integrity in this region is inadequately supported? In summary, the reasons to suppose that MMN can be used as tool to provide insights into schizophrenia-related brain changes are many and derive from both empirical and theoretical origins. In the following sections we review MMN research from several perspectives to identify some important challenges to realizing its potential in the study of biological processes linked to schizophrenia.

## **THE MEASUREMENT OF MMN – IS THERE AN OPTIMAL PARADIGM WITH WHICH TO STUDY MMN IN SCHIZOPHRENIA?**

Smaller MMN in schizophrenia has been observed over a number of different experimental paradigms so the underlying reasons for MMN-reduction are likely to reflect the common demands inherent in processing the various sound sequence types. Different paradigms each have advantages and disadvantages (reviewed below) and it is the authors' opinion that there is no ideal paradigm. In fact, to nominate an ideal could be detrimental to the field – the advantage in doing so (increased comparability between studies) is outweighed by the disadvantage which is neglect of the opportunities to use this tool to address specific questions about the integrity of the underlying system.

#### **CONSIDERATIONS IN MEASURING A DIFFERENCE WAVEFORM**

There are at least three contributing factors to MMN obtained in the classical deviant-minus-standard waveform: (i) differences in physical attributes of the deviant and standard sounds (this is particularly likely when the deviant contains physical features that vary in acoustic energy such duration or intensity), (ii) differences in refractoriness (or more properly called adaptation) of the neural generators of responses to the more frequently presented standard versus rare deviant stimulus and (iii) a genuine deviancedetection process or the *true MMN* resulting from the deviant violating predictions derived from contextual regularities. This means that a group difference in classically derived MMN could arise from any one or a combination of these factors. A number of experimental control protocols have been developed to refine a measure of *true MMN* – the primary motivation for these studies being to determine whether a true MMN exists or whether what is measured as MMN could be accounted for entirely by adaptation effects on other components of the ERP. In a series of papers, Schröger and his colleagues devised a random control stimulus sequence, the *many-standards* sequence. For example in the case

of a frequency deviant, a many-standards control would involve random presentations of tones of different frequencies (including tones with the frequencies of the deviant and standard tone) each having the same probability as the deviant in the oddball sequence [Frequency/pitch: (46–48); Duration: (49); Location: (50)]. A comparison of responses to the deviant from an oddball sequence and to the same sound in a many-standards sequence controls for physical differences and adaptation contributions to the classically derived MMN. In this way the resultant difference waveform can be attributed to a novel event violating a *stored neural representation* of regularity in recent stimulation. This is because the random presentation order of the many-standards sequence *prevents the development of a representation of regularity in recent stimulation* but preserves the rate of occurrence and therefore some level of equivalence in terms of adaptation effects.

To our knowledge there are no studies that have implemented conservative control procedures to extract true MMN in schizophrenia. As noted above, the vast majority of the published literature in schizophrenia has employed MMN derived using the classical oddball derivation. The common finding of similar responses to standards in schizophrenia and matched control groups is reassuring in that it suggests that adaptation effects on the standard response at least are not detectably impaired in patients and implies that the major group differences are in how the brain responds to deviations. However whether there are group differences in the extent of cross adaptation to other frequencies (or durations/intensities, etc., in the case of simple deviance paradigms) is unknown and this could impact on the response to a physically different deviant tone. It is also reassuring to note that in healthy individuals, even when stringent controls such as the many-standards control are implemented, the classical derivation gives a reasonable approximation of the true MMN (49). But once again – whether this is equally true in patients is unknown. Below we review some of the literature employing non-classical MMN derivations indicating that there may indeed be some group differences in response to repetitive sound as well as the response to deviations.

#### **EXAMPLES OF PARADIGMS USED IN SCHIZOPHRENIA**

A few patient studies have employed methods that can differentiate components of the MMN process and these have yielded mixed findings. Some have controlled for the physical differences between standard and deviants sounds, either by presenting two oddball sequences with the roles of deviant and standard sounds reversed [a flip-flop design: (51, 52)] or by presenting a sequence consisting of repeated presentations of deviant sounds only (53), referred to as a deviant-as-standard sequence. In both instances, it is possible to derive difference waveforms from comparing responses to the *same sound* as a repetitive standard and a deviant event. While these studies report no group difference in response to the repetitive oddball standard presentations, Todd et al. (54) recently found evidence of clear group differences in the morphology of ERPs to the repetitious presentation of deviants and these differences had a significant impact on computation of the MMN. The group difference appeared in a negative component of the ERP to repetitive sounds occurring ~200–300 ms post-stimulus. While its origin and functional role are unknown, it suggests that such differences in cortical response to a repetitive sound more generally warrant further investigation. However, neither the flip-flop control nor the deviant-as-standard sequence removes the contribution of adaptation effects to the difference waveform. Nor does it ensure that the computed MMN represents true deviance detection of a violation of contextual regularity. That is, these procedures do not allow the extraction of the *true MMN*.

The roving paradigm is another method that can partially address group differences in various contributions to MMN. In a typical roving paradigm a string of sounds of the same pitch are eventually interrupted by a sound with a higher or lower pitch which then continues to repeat. Predictive processing is highly dynamic so within two to three repetitions, the new sound becomes a standard and deviations from its properties will now elicit MMN (55–57). In the roving paradigm, it is possible to study the way the response to a new standard changes after incremental repetitions (e.g., 6 vs. 12 vs. 24, etc.). An increasing positivity as a function of repetition length, termed "repetition positivity," is apparent in the ERP to standards approximately 50–200 ms poststimulus. The increase in MMN amplitude with the length of repetition is referred to as a memory trace effect and is a function of both this apparent increased positivity (which is in fact a decrease in negativity) to the standard and increased negativity to a subsequent deviant [although see Ref. (58)]. In schizophrenia, the increment in both components is smaller than that in controls (59), but more notably in this study, the positivity in response to standard repetitions failed to increment at all. So in this paradigm smaller MMN in schizophrenia appears to indicate less change in responsiveness to sound generally.

In a recent magnetoencephalographic (MEG) study in schizophrenia (60), the MMN<sup>m</sup> data in a similar roving paradigm was explored using a technique called dynamic causal modeling (DCM). DCM differs from conventional source modeling and brain connectivity methods by utilizing a biologically informed causal model placing constraints on model inversion such that the parameters of reconstruction describe specific processes like change in synaptic coupling strength between source locations and postsynaptic gain. Rather than estimating dipole activity at a particular point in time; it models dipole activity over a period of time to identify parameters that change (11). When applied to the roving paradigm, DCM has provided evidence supporting the conceptualization of MMN as an active contextual perceptual inference process. The"best fit"to experimental data is achieved by a model that incorporates both local intrinsic adaptation effects as well as plastic changes in extrinsic inputs to auditory cortex (i.e., from STG to A1 and also from areas of the prefrontal cortex to STG). When applied to schizophrenia data, DCM has provided evidence for problems in two of three components: the largest effect size for group differences was for reduced change in intrinsic connections within primary auditory cortex A1. Such changes are considered evidence of impaired feature specific adaptation. There was also a reversed polarity in changes to connectivity between prefrontal and auditory areas which was interpreted as a failure in the normal influence of these top-down inputs in modifying auditory cortical response. The authors also comment on reduced modulation of the forward connection from the A1

to the STG (but this was not significant according to **Table 2**, p. 26). Although this is the first study of its kind in schizophrenia, the results conform to the view that impaired signaling of error (smaller MMN) could be indicative of impairment in encoding the contextual memory against which deviance is registered (potentially including impaired adaptation), but also consequently impaired ongoing modification of cortical responsiveness by feedback projections. It is therefore possible that the roving standard paradigm is more sensitive to any problems in forming a contextual memory based on tone repetitions. However, when interpreting results it is important to consider the assumptions of the roving paradigm carefully before drawing this conclusion. The MMN elicited to a change in frequency in a roving paradigm signals that the current prediction-model failed to account for the present stimulus properties and the model may require updating. With repetition of the new frequency, an updated prediction-model is built. The degree to which the model requires updating will be a function of the difference between the representation of the new and former standard frequencies. Given that the changes in frequency can be quite subtle and that frequency discrimination is impaired in schizophrenia (53, 61, 62), it is possible that less evidence of updating after a new standard (repetition positivity) could in part reflect less distinct representations of the new and former frequency. So the roving paradigm places considerable weight on stimulus-specific adaptation to a new frequency, which therefore may augment the importance of intrinsic adaptation in A1. Whether DCM of a classic oddball paradigm would replicate major group differences in intrinsic A1 connections remains to be determined.

At present the literature certainly suggests that altered (smaller) response to sequence deviations is a major contributor to smaller MMN in schizophrenia. It is clear that the reduction in classical MMN is not only robust across many different cohorts, laboratories, ethnic groups, etc. (5), but also exhibits substantial stability over time (63). However there is reason to suppose that there may also be differences in responding to repetitive sound more generally that could be contributing to problems in prediction-modeling and/or MMN computation. Novel paradigms and novel data processing approaches are likely to provide valuable data with which to address these contributions. Despite these differences it should be noted that many elements of the predictive process underlying MMN remain intact in schizophrenia. In addition to those covered above we recently demonstrated that persons with schizophrenia, despite producing smaller MMN amplitude to deviants, are equally able to *reduce* MMN size to a deviant if the occurrence of that deviant could be inferred from the identity of the prior tone. The equivalent use of this predictive information in both schizophrenia and control group reinforces the position that abnormalities in the MMN process in schizophrenia primarily involve limited gain in the differential response to a very rare versus common event. So while the choice of paradigm used to study MMN in schizophrenia will depend not only on the questions driving your research but also the time you have available, the one principal recommendation that we do put forward is to ensure that you adopt a very rare deviant event (<15%) to maximize your power to expose group differences.

## **DOES MMN-REDUCTION INDICATE VULNERABILITY TO SCHIZOPHRENIA?**

In this section we review studies addressing whether small MMN may be a vulnerability index for schizophrenia. It should be noted that research into schizophrenia has always been limited by its diagnostic heterogeneity and phenomenological overlap with related developmental, affective, or personality disorders which all can share some of the clinical features of schizophrenia, such as psychosis, cognitive impairment, or therapeutic response to certain classes of pharmacological agents (64). Not surprisingly, attempts to find a diagnostic marker for schizophrenia – that not only represents an *endophenotype* of the disorder but can also serve as a tool helping to unveil its pathology – continues to be limited by the lack of a pathogenomic definition of the condition. Diagnostic limitations notwithstanding, a number of groups have indeed investigated MMN's status as an endophenotype for schizophrenia as defined by current diagnostics tools and this research is pertinent to evaluating the translational potential of MMN in schizophrenia. The definition upon which the evaluation is based is that proposed by Gottesman and Gould (65) where an endophenotype is defined as an intermediate phenotype along the pathway between genotype and the observable established aspects of the illness. Specific criteria to be met include: (i) it is associated with illness in the population; (ii) it is heritable; (iii) it is primarily stateindependent (manifests in an individual whether or not illness is active); (iv) within families, small MMN and the illness cosegregate; and (v) smaller MMN is evident in non-affected family members at a higher rate than in the general population (65).

## **CRITERION (I) AND (II)**

Mismatch negativity amplitude reduction in schizophrenia is a very robust finding with an effect size of 0.99 observed by Umbricht and Krljes (5) in their meta-analysis of 32 studies published prior to 2004. Reduced MMN is less prevalent in related conditions, such as bipolar affective disorder independent of the presence of psychotic symptoms (52, 66–68) and in major depression (66) but see (69–71). That is, there is considerable evidence that criterion (i) is met in terms of its association with the illness, although its specificity to schizophrenia is less clear than initially thought. MMN appears to be heritable based on twin studies with heritability estimates ranging from 0.48 to 0.68 when using a duration increment deviant (67, 72) but not for a frequency deviant (73). However, even duration MMN shows only weak phenotypic association (0.39) with schizophrenia (67). Nonetheless, there is preliminary evidence supporting criterion (ii) for endophenotypic status of MMN.

## **CRITERION (III)**

Across the large number of studies on MMN in schizophrenia there are no consistent relationships between MMN size and the severity of symptoms of psychosis. Although impaired predictionerror signaling is implicated in the genesis of delusions (74) and although there is preliminary but consistent evidence from one group of a relationship with auditory hallucinations (75–77), the literature fails to demonstrate consistent relationships across studies. The meta-analysis by Umbricht and Krljes (5) emphasized that the majority of studies did not find correlations (either with positive or negative symptoms) and observed no change in MMN when symptoms improved. Further (63), found duration MMN in very large sample of patients (*N* = 163) exhibited substantial stability across a 1 year retest interval, and to be independent of fluctuations in clinical symptoms, positive or negative. So at face-value, these results from cross-sectional designs seem to support state-independence of presence of smaller MMN [criterion (iii)]. However, the evidence about state-independence of degree to which MMN is reduced in patients in an acute phase vs. a postacute phase is mixed (78, 79) although differences due to medication changes cannot be eliminated (see later Section "How do Pharmacological Manipulations Alter the MMN Process?" for further discussion of state vs. trait effects on MMN as a vulnerability marker).

The failure to demonstrate state-dependence is paralleled by more consistent relationships between MMN amplitude and relatively stable features of the illness such as level of functioning (80) and cognitive impairments (81) despite substantial changes in negative and positive symptom severity. Various measures of current functioning have been shown to be associated with MMN amplitude in patients: global assessment of function [GAF: (82–84)], social and occupational functioning assessment scale [SOFAS; (35) but not in first episode patients (85)], the independent living scales [ILS: (86)], and work functioning and independent living ratings from the role functioning scale (87). It has been suggested that the relationship between reduced MMN and impaired functioning might be mediated by anatomical changes such as gray matter loss in relevant brain regions (35). Impaired cognition (whether independent of gray matter declines or consequential) could also mediate the relationship between MMN and functional status since there is a wealth of evidence now that the strongest predictor of functional outcomes in patients is cognition [also often called neurocognition: (88)]. However, the number of published MMN studies in schizophrenia that examine not only cognition in the same sample but correlations between MMN and cognitive performance as well, is quite limited. This is despite reasonable expectations that the reliance of MMN generation on the the NMDAGluR system (see How do Pharmacological Manipulations Alter the MMN Process? and Translation to Animal Models below for further discussion) should lead to relationships with those aspects of cognition that are also reliant on the unique characteristics of the NMDAGluR, such as context-dependent effects, integration of information over time and new learning (44). However, there are some data that suggest such relationships do exist at least in patients (but not necessarily in healthy controls).

NMDAGluRs have a number of unique features. Firstly, activation of NMDAGluRs currents is conditional in that channels only gate following presynaptic release of glutamate and concurrent postsynaptic membrane depolarization which relieves Mg2<sup>+</sup> blockade. This conditional characteristic of the NMDAGluR is likely to be particularly important where responses are determined by context, as is the case for MMN, but also in situations where flexibility of response is required dependent on context. One specific example of contextual processing is the AX version of the continuous performance (AX-CPT) task where a response to the letter X on screen should only executed when the X is preceded by the letter A. It is well established that patients are impaired on the AX-CPT task, producing often fewer correct responses to AX sequences (impaired priming of response by a target-consistent cue) and higher rates of false alarms to BX sequences [impaired inhibition of a response prompted by a target-inconsistent cue (89)]. However, despite reports of concurrent smaller MMN and AX-CPT impairments, there appear to be no reports of a correlation between the two (90) and one explicit report of no association between the two (54).

Secondly, although NMDGluRs exhibit complex kinetics with evidence of multiple gating modes characterized by different mean open times (91–93), it is generally accepted that they mediate long duration excitatory postsynaptic currents in the brain and participate in synaptic integration and certain forms of synaptic plasticity (92). Prefrontal cortex NMDAGluRs in particular have slower kinetics than sensory regions (94) and therefore are potentially involved in maintaining activity in prefrontal neurons (95), for example during the delay periods of working memory tasks. There are reports of co-occurrence of working memory deficits and reduced MMN in patients with schizophrenia (63, 90) but to our knowledge, only one report of a correlation between working memory (measured using the digit sequencing task from the Japanese version of the brief assessment of cognition) and (duration) MMN in patients (96). Both classic oddball MMN amplitude (97) and longer term effects on growth in MMN amplitude (98) have been shown to correlate with digit span (the ability to maintain and or manipulate acoustic presentation of digits) in a healthy control group consistent with the requirement to store auditory information over time for both indices.

Thirdly, NMDAGluR activation leads to a cascade of events that initiate long term potentiation and depression, the primary processes responsible for new memory formation and learning in hippocampus and cortex. Retention or storage of information is less reliant on NMDAGluR. One of the most robust cognitive deficits exhibited by patients are deficits in memory (verbal declarative memory in particular) with the majority of evidence suggesting that the largest deficit, as measured by effect size, occurs for encoding or new learning of material in comparison to retention [although there is still a small deficit in retention even when initial learning differences are taken into account (99)]. Four studies report relationships between MMN and memory performance but in patients only. Baldeweg et al. (59) using a roving oddball paradigm found that the MMN trace effect (the increase in MMN that occurs with increasing repetition of prior standards) correlated with performance on an everyday memory test (Rivermead behavioral memory test). Kawakubo et al. (100) in a study that reports data on patients only found that MMN elicited by a phoneme duration deviant (but not a tone duration deviant) was correlated with immediate free recall (initial learning or encoding measure) from a list learning task, the Rey auditor verbal learning task (RAVLT). In contrast, Kaur et al. (101) found that tone duration MMN correlated with verbal memory assessed using the RAVLT in first episode psychosis patients with a schizophrenia spectrum diagnosis of either schizophrenia, schizoaffective, and schizophreniform illness. No correlations were reported for controls. Kiang et al. (84) in a standard oddball paradigm found duration MMN was correlated in patients only with short-delay as well as long delay free recall but not (significantly) with immediate

free recall on another list learning task, the California verbal learning task, but neither of the free recall measures was adjusted for initial or prior learning differences. So while there is evidence of MMN correlations with memory, these correlations may not be restricted to initial learning or encoding phase.

In addition, there is evidence of relationships between MMN amplitude and other cognitive domains that are commonly shown to be impaired in patients but are less clearly dependent on NMDAGluR properties, such as executive functions [anti-saccade performance: (102); proverb interpretation: (84); perseverative errors on the Wisconsin Card Sorting Task: (103); mental control subtest of WMS-III: (85)].

In summary – MMN does not appear to be consistently related to the severity of either positive or negative symptoms experienced by the patient at the time of recording (either examined across patients or assessed within patients at different time points), consistent with (symptom) state-independence of MMN. However, MMN amplitude does appear to vary across individuals as a function of more stable features of their illness. The relationship of MMN to functional measures is relatively robust and there is growing evidence of correlations not only between MMN amplitude and those aspects of cognition that are likely to be reliant on unique aspects of NMDAGluR but other domains of cognition that are reliably impaired in patients such as executive functions. However, these are issues that deserve more attention and in particular, more systematic investigations are required before any strong assertions can be made about contributions from a common mechanism of NMDAGluR-dysfunction to MMN-reduction and cognitive deficits in schizophrenia.

#### **CRITERION (IV) AND (V)**

Mismatch negativity has also been investigated as a potential predictor of developing psychosis or schizophrenia in populations considered at-risk mental state (ARMS). For instance, the comprehensive assessment of at-risk mental state (CAARMS) criteria (104) defines ARMS as (1) a significant drop of global functioning over a period of 12 months and having a close biological relative with a psychotic disorder and/or (2) experiencing attenuated or very brief episodes of psychotic symptoms in combination with functional decline. MMN has been investigated in both groups; however, to date there appear to have been no MMN longitudinal studies of transition to psychosis specifically within a clinically unaffected genetic group.

Whether MMN is reduced in unaffected first-degree biological relatives of patients with schizophrenia is controversial. The initial studies by Jessen et al. (105) 1 and Michie et al. (106) reported reduced oddball frequency and duration MMN respectively in first-degree relatives when compared to healthy controls, but these results with one exception were not replicated in later publications, neither for duration deviants – (72, 107–109) and (102), [based on a larger sample from the same source as (106)], nor frequency deviants – (73, 108, 109). The one exception is (110) which interestingly used an identical MMN duration deviant paradigm as (106). While there are design and other methodological

<sup>1</sup>A somewhat puzzling result in the Jessen et al. study, is that patient MMN amplitude was not significantly reduced in contrast to relatives.

differences between those studies (three in total) that show significant MMN-reductions in first-degree relatives and those that do not (six in total), the bulk of the evidence suggests that at-risk but clinically unaffected family members do not exhibit reduced MMN (minimal support for criterion (iv) and (v) of endophenotype criteria).

The evidence that MMN is reduced in the ARMS group, is more consistent, although data on whether reduced MMN predicts transition to psychosis is still preliminary. It is important to note that while the term *prodromal* is sometimes used to describe those clinically defined at-risk groups (111), strictly speaking whether they are prodromal or not at the time of assessment can only be determined subsequently by whether they develop a schizophrenia spectrum disorder within the follow-up period (usually 12–24 months). All of the MMN investigations in ARMS have used duration deviants, either a duration increment (112–117), or a duration decrement (111, 118) or both (112) but some also report data on frequency deviants in the same sample (111, 116–118) or a double deviant [deviant on both frequency and duration (117)]. Of these eight papers, five found that duration increment MMN was significantly reduced in the ARMS group (112–114, 116, 117) whereas duration decrement MMN was either not significant (111, 118) or showed a smaller effect size (112). The findings for frequency MMN are mixed: (117) found that frequency MMN was reduced (as was the double deviant) but (116) did not. To date therefore, the evidence seems to suggest that deviants that differ from standards by being longer in duration are more sensitive to the at-risk mental state that other deviant types. It seems unlikely that these findings are due to the effects of medication since in each sample the numbers of ARMS individuals who were medicated with anti-psychotics at the time of testing was small as were the dosages.

Most investigations of clinical high risk groups also report transition data and examine whether MMN predicts those who will subsequently develop a schizophrenia disorder – although in two cases, the number of transitions was too small for statistical analysis (112, 113). Bodatsch et al. (118) were the first to report transition data. Interestingly they found that duration (decrement) MMN predicted those who converted to schizophrenia within a 24 month period of the assessment date whereas frequency MMN did not. Higuchi et al. (115) and Shaikh et al. (114) observed similar results for duration (increment) MMN. Neither included a frequency MMN deviant. However (117), found that the best predictor of later transition to schizophrenia was MMN to their double deviant. Neither duration alone nor frequency alone was significant. Perhaps importantly, the double deviant elicited the largest MMN (larger than that to frequency or duration alone) which perhaps reflects the importance of challenging the upper limits on MMN size in at-risk groups as well as patients with an established illness. In summary, evidence for smaller MMN within families in general is not strong but the evidence for smaller MMN in clinical high risk groups is quite compelling. Therefore endophenotype criterion (iv) and (v) are only partially met at best. It remains to be seen whether the reduction in the atrisk groupings is really about risk status *per se* or a reflection of schizophrenia-related pathology that has begun to impact on brain function.

## **HOW DO PHARMACOLOGICAL MANIPULATIONS ALTER THE MMN PROCESS?**

Pharmacology as a field of research offers a unique avenue to study MMN both in terms of how different chemicals can perturb the perceptual inference process and how they may relate to schizophrenia pathology. When considering the pharmacological sensitivity of MMN, it is clear that a change in MMN could reflect an effect on any one of a number of constituent processes described in Section "The Measurement of MMN – Is there an Optimal Paradigm with Which to Study MMN in Schizophrenia?"Surprisingly few studies provide details on how a substance affected response to standard repetitious sounds with the majority reporting on the difference waveforms only (16/27 studies, see **Tables 1**–**3**). For the purpose of this review we have restricted Tables to acute drug effects on healthy adult populations. One of the key foci in pharmacological research on MMN is how it is affected by alterations to NMDAGluR activity (**Table 1**).

Several groups have argued that impaired plasticity linked to NMDAGluRs is a core feature of the schizophrenia illness (45, 128, 129). The first study to demonstrate this link was actually in the macaque where Javitt and colleagues demonstrated a dose-dependent reduction in MMN following local infusion of the NMDAgluR antagonist, PCP (130). Using a flip-flop control design (see The Measurement of MMN – Is there an Optimal Paradigm with Which to Study MMN in Schizophrenia? above for description and Translation to Animal Models below for additional detail) the authors report no significant effect of phencyclidine on response to the repetitive sound (at least in these local field potentials, see Translation to Animal Models for further discussion) but a pronounced dose-dependent effect on the response to the deviant. The authors conclude that NMDAGluR activity is critical to forming the associative links between stimuli (i.e., accumulating information about transition statistics) that define the context against which a rare deviant sound is recognized as an aberrant event. Studies in humans are largely consistent with this initial study. Of eight published studies, seven report significant dose-dependent reduction of MMN amplitude after ketamine [although one only in combination with the CB1 inverse agonist rimonabant (125)]. The one exception (a low-dose study) used a selective attention paradigm in which participants were asked to attend and respond to stimuli in one ear while simultaneously hearing stimuli in the other ear (unattended) from which MMN was derived (119). Of the studies showing an effect of ketamine on MMN, four explicitly picture and/or discuss response to the repetitive standard tones with only one (120) reporting a slight but significant increase in the obligatory N1 component to repetitive sounds in the presence of the drug. The low-dose ketamine study by Oranje and colleagues also show enhanced N1. The most recent study reports on a roving standard paradigm demonstrating that, although MMN was reduced under ketamine overall, the effect of ketamine was more pronounced with increased repetition of the standard [i.e., when predictive confidence was highest (126)]. The paper, however, did not report on whether this pattern was due to effects on the positivity to repetitive standards or increased negativity to deviants or both. The grand-averaged (i.e., across repetitions) ERP to standards is presented in a subsequent paper reporting a DCM analysis of the same data (14). A visual

#### **Table 1 | N-methyl-d-aspartate receptor studies.**


#### **Table 2 | Nicotine receptor studies.**


comparison of Figure 1 (placebo) and Figure 2 (ketamine) from the paper suggests that the effect of ketamine on MMN was perhaps a combination of reduced positivity in the standard waveform and reduced negativity in response to the deviant. However, DCM analyses applied to the data indicated the ketamine had a selective effect on parameters representing synaptic plasticity with no effect on indices reflecting adaptation. Additionally, analysis indicated that the major effect of ketamine was to reduce the normal increase in synaptic plasticity in the forward connection between left A1 and left STG in response to the deviant tones. Therefore while ketamine administration and schizophrenia are both associated with reduced MMN, the pattern of change (at least for this roving paradigm) is quite distinct. Ketamine-induced a prominent change in feed-forward connections between A1 and STG only (14) while schizophrenia was associated principally with reduced intrinsic connections within A1 and significant alteration in prefrontal-STG connections (60).

In Schmidt et al. (14, 126), and an earlier observation by Umbricht and colleagues (131), MMN measures were associated with the intensity of psychotic-like reactions to ketamine. Schmidt et al. (126) reported that participants with the most restricted growth in MMN with increased standard repetition experienced the most pronounced disturbances noted on a "control and cognition" subscale of the altered state of consciousness questionnaire in the presence of ketamine. Similarly Umbricht and colleagues observed that the participants who produced the smallest MMNs

at baseline experienced the highest symptom ratings (on a variety of measures) under ketamine. In the subsequent DCM analysis of Schmidt et al. (14), those who experienced the most pronounced "control and cognition" subscale disturbances in the presence of ketamine also showed the most pronounced decrease in plasticity in the left A1-STG connections under ketamine. Although the measures differ between studies, these observations invite the intriguing conclusion that small MMN amplitude (and/or limited growth in MMN) may indicate a limitation in synaptic plasticity that is linked to vulnerability to psychotic-like phenomena. Interestingly, this vulnerability does not appear to be generic as studies have shown no relationship between MMN amplitude and psychotic-like response to psilocybin and no significant effect of psilocybin on MMN amplitude (122, 126). Furthermore, there is evidence that this effect on NMDAGluR-mediated plasticity shows some specificity to ketamine-induced antagonism. Memantine (also an NMDAGluR antagonist) actually augmented MMN amplitude (123). The downstream effect of ketamine (and MK-801) has recently been demonstrated to be quite different to that of memantine. One interesting observation is that these compounds have opposing effects on postsynaptic density proteins – namely ketamine and MK-801 reliably increased *Homer1a* relative to *Homer1b* expression while memantine has the reverse effect. Authors suggest that the former impacts the expression of genes related to response to neuronal injury and preservation of homeostatic scaling of synaptic response. The latter, in contrast,

#### **Table 3 | Monoamine receptor studies.**


strengthens synaptic transmission. These very different effects on plasticity may go some way to explaining the opposing effects of these NMDAGluR-antagonists on MMN and invite speculation as to whether individual differences in these same ratios may in fact confer differential susceptibility to ketamine-induced psychoticlike experiences (and disruption to the MMN process, however see also discussion of animal research using memantine in Section "Translation to Animal Models"). In summary, antagonism of NMDAGluRs in the presence of ketamine produces quite consistent reduction in MMN amplitudes and continues to hold promise in furthering our understanding of the plasticity underlying MMN as well as vulnerability to psychotic phenomena. Of course it should be remembered that acute disruption under ketamine is unlikely to mirror the full consequences of adjustment to a more chronic compromise in function (if present) in schizophrenia.

Just as there are multiple elements to the MMN process, there are multiple ways to pharmacologically influence NMDAGluRmediated synaptic plasticity [for a relevant review see Ref. (132)]. NMDAGluR antagonism also occurs under acute exposure to alcohol and consistent with this, acute administration of ethanol has been shown to reduce MMN amplitude (133, 134). Nicotine in contrast enhances synaptic plasticity with mechanisms linked to effects on presynaptic NMDAGluRs (135). Nicotine exerts its effects on the central nervous system via acetylcholine receptors (135). Galantamine has been used to test theories about how augmentation of cholinergic neurotransmission can modulate gain in MMN (136). The results of both an empirical study and a simulation experiment indicate that enhanced cholinergic neurotransmission alters precision in prediction-modeling – i.e., changes confidence in the current inference model. More specifically, under these conditions the system places a greater emphasis on bottom-up input and "boosts" the response to deviants while also attenuating the usual reduction in confidence in a model following the occurrence of a deviance. The authors suggest that acetylcholine plays a key role in modulating gain in superficial pyramidal neurons in early sensory brain areas. Consistent with this action, five of the seven studies listed in **Table 2** support nicotine enhancement of MMN. Significant enhancement of MMN has been demonstrated under both acute (137–141) and more prolonged exposure (139). Of the seven studies listed, five present or report on response to standard tones but only one indicates significant drug effects. Baldeweg et al. employed the roving standard paradigm and revealed that the increase in MMN amplitude with nicotine was due to a selective augmentation of the positivity

to repeated standards with no significant effect on response to deviants.

Cannabis use is considered by some to be a risk factor in the development of psychosis. Furthermore, chronic cannabis use has been associated with gray matter volume changes and cognitive deficits reminiscent of those in schizophrenia (144–147). The action of endogenous cannabanoids is also linked to NMDAGluRs in protecting against excessive stimulation at glutamatergic synapses. The cannabanoids are released from postsynaptic neurons and exert their action on CB1 receptors located on presynaptic neurons which transiently decreases neurotransmitter release [(148) for review]. There are currently two published studies on the acute effects of cannabis [in fact the latter is a reanalysis of the former with genetic data included (149, 150)]. The first of these explored the effect of administering ∆9-tetrahydrocannabinol (THC) alone (the psychoactive component of cannabis) versus in combination with the other cannabinoids present in cannabis extract. The results indicated no significant impact of ∆9-THC alone, but significant augmentation of MMN at central sites in the presence of cannabis extract compared to placebo. However the study also demonstrated a significant correlation such that higher concentration of the ∆9- THC-metabolite 11-OH-THC was associated with smaller MMN amplitude (*r* = 0.62, *p* = 0.002). In their later combination with genetic data the same group revealed that susceptibility to MMNreduction in the presence of ∆9-THC was a function of genotype for neuregulin 1 (MMN reduced in the presence of ∆9-THC for those with NRG1 rs7834206 polymorphism). Neuregulin 1 is a gene implicated in schizophrenia that influences synaptic plasticity via multiple pathways, including those involving NMDAGluRs (128). The possibility that genes confer vulnerability to ∆9- THC effects on cognitive processing also finds support in animal research where heterozygous Neuregulin 1 knockout mice have been observed to show differential sensitivity to the acute effects of ∆9-THC on behavioral phenotypes of schizophrenia (151).

Despite dopamine being a modulator of NMDAGluRs and a central focus of treatment and models of schizophrenia (37, 132), there are no studies supporting a significant effect of altered dopamine levels on MMN amplitude in healthy adults (see **Table 3**). Furthermore, with one recent exception (reviewed below), studies within schizophrenia do not support a significant effect of medication type or dose on MMN [see Ref. (5) for review], and no significant differences in MMN amplitude between medicated and unmedicated patients (52, 152). However, Zhou and colleagues (79) have recently reported a significant progressive increment in MMN in persons with schizophrenia treated with aripiprazole (larger at 4 and 8 weeks of treatment than at baseline). MMN was measured using a traditional oddball paradigm with two deviant types (frequency and duration) and effects are presented and reported for the difference waveforms only. Aripiprazole differs from other second generation anti-psychotics in that its action at dopamine D2 receptors shows *functional selectivity* (153–155). It has been proposed that this selectivity may be related to observations that it has differential effects on the two main dopaminergic pathways, namely a predominant effect on the mesolimbic pathway (156). Although MMN remained significantly smaller than that in matched controls, the authors argue

that the effect of the drug treatment on MMN does raise questions about whether MMN amplitude is really a trait or state marker. Finally studies on benzodiazepines, often prescribed to persons with schizophrenia, have consistently failed to demonstrate any effect on MMN amplitude (see **Table 3**).

In summary, the literature to date on the pharmacology of MMN generally reflects its obvious relationship with experience-dependent synaptic plasticity. The observed effects of ketamine, nicotine and cannabis provide support for the NMDAGluR-susceptibility of the system underlying MMN and therefore utility to schizophrenia research on MMN as they offer useful insights into the neurobiological processes that can influence or modulate MMN amplitude. However the memantine studies clearly caution that the relationship between perturbation of NMDAGluRs and MMN amplitude is not a simple one. There is insufficient information within most published studies to determine exactly how the various agents are altering the underlying processes with most reporting on difference waveforms only. Clarity regarding which elements of changed responsiveness are affected by drugs is particularly important to a thorough understanding of the process and these issues are discussed further in Section "Translation to Animal Models" below. Where possible it may be advantageous to add genotyping to pharmacology studies as it is well known that susceptibility to drug effects can be dependent on genetic profiles but of course the cost and sample size requirements are often prohibitive.

## **TRANSLATION TO ANIMAL MODELS**

Animal models have the potential to inform investigation of the physiological basis of MMN and potentially how it is disrupted in persons with schizophrenia. However, the primary issue of debate in animal models of MMN is to determine which components of the MMN process can be observed (and under which conditions they can be observed) in animals, there being some skepticism expressed in the past over whether the rodent brain exhibits "true MMN" (165). Debate over how exactly to measure MMN (see The Measurement of MMN – Is there an Optimal Paradigm with Which to Study MMN in Schizophrenia?) is therefore particularly relevant to animal work. Results from several different control designs are reviewed in this section and for ease of communication they are presented diagrammatically in **Figure 1**.

The presence of adaptation to repeated stimuli (or stimulusspecific adaptation, SSA) in A1 is well-described in the cat (8, 166), rat (167–170), and macaque (171). A1 neuron populations will adapt to repetitive stimuli and will exhibit relatively large responses to a rare deviant compared to a common standard. However, as reviewed in Section "The Measurement of MMN – Is there an Optimal Paradigm with Which to Study MMN in Schizophrenia?" evidence of "true MMN" must include something more than just adaptation effects. Terminology used is not consistent across animal and human studies so for simplicity we adopt the following nomenclature: oddball mismatch response (MMR) is used when referring to the difference between the response to a rare deviation and the regular (usually physically different) repeating sound; and controlled MMR is used to define evidence of contextual deviance detection when a study has included some kind of control for adaptation and physical characteristics of the stimuli (see **Figure 1**).

## **WHAT IS THE EVIDENCE FOR A "CONTROLLED MMR" IN LOCAL FIELD POTENTIALS?**

**Table 4** summarizes the animal studies that utilize oddball paradigms to test for evidence that a contextually deviant sound has been detected. The location of the recording electrode plays a large role in the findings of these studies, as epidural-placed electrodes receive input from a larger and more distributed network of cortical neurons than local field potential electrodes (LFP; measuring changes in synaptic potentials) or electrodes that register multi- or single-unit spike activity [MUA, SUA; see Ref. (165)]. The majority of studies using LFP recordings have searched for evidence of contextual deviance detection by recording from A1, but some also have recorded from the hippocampus (172–174). The majority of investigations using LFP or MUA


#### **Table 4 | Summary of papers investigating mismatch responses in animal models.**


#### **Table 4 | Continued**

A1, primary auditory cortex; CA1, CA1 region of hippocampus; DG, dentate gyrus; LFP, local field potentials; MGB, medial geniculate body of thalamus; MUA, multi-unit activity; SUA, single-unit activity.

(intra-cortical recordings), that also use many-standards or deviant-alone control, have not found any evidence for true MMR in the A1 (8, 167, 169, 171), whereas the majority of studies using *epidural* recordings and appropriate controls *have* found evidence for contextual deviance detection. This indicates that while adaptation to stimulus frequency/pitch does occur in the A1, the deviance-detection component of the MMN has not been identified in this region, and if it is present in the A1, it may be generated by more distributed networks than can only be observed using LFP recordings.

One of the first animal model studies (130) found larger responses to deviant sounds in an oddball paradigm originate in the supragranular layers (II–III) and are dependent upon normal functioning of the NMDAGluR (covered in Section"How do Pharmacological Manipulations Alter the MMN Process?"). However, Javitt et al. did not utilize a many-standard control and in a more recent study that did (171), larger responses *were not* observed to the deviant compared to the same tone in the many-standards sequence (no evidence of true MMR). The many-standards control in this study may have overestimated adaptation, however, as the deviant was presented at a probability of 10% in the oddball sequence, and the tones used for the many-standards control were presented at a maximum rate of 5% (171).

There are, to our knowledge, only two studies thus far utilizing LFP recording that have found some evidence for what might be considered controlled MMR in the rat brain (170, 175). Imada et al. (175), using a frequency deviant paradigm, found that the difference between the ERPs to deviant and standard stimuli was significantly elevated from 0 in an oddball MMR. However as a control they included separate sequences preserving the timing of standard sounds (without deviants) and deviant sounds (without standards) from the original oddball sequence. The ERP to deviant-alone and the standard-alone did not differ (175). According to the authors, this indicates that there was a statistically significant change in the response to the deviant stimuli in the oddball measure, over and above adaptation, which may hint at the presence of true MMR, albeit one that is not large enough to produce a statistically significant difference in responses to the oddball deviant compared to the deviant-alone. Taaseh et al. (170) utilized both LFP and MUA recording and found that the response to the oddball deviant tones did not exceed the response to the same tone in a many-standards control sequence (their *diversebroad* sequence). However, the response to the deviant sound *was* larger than the many-standards control in some recordings (although not at a population level), indicating that perhaps a subset of neurons do exhibit sensitivity to the contextual deviance of the sound (170). In addition, their thorough characterization of adaptation effects was used to generate a model which predicted that if only adaptation were occurring, the response to the deviant would be smaller than the response to the same tone in the many-standards control. However, these responses did not differ significantly, suggesting (albeit rather indirectly) that contextual deviance detection is occurring in the auditory cortex. There is therefore some evidence that recognition of contextual deviance occurs in the rat A1. Such hints of response to contextual aberrance may represent the first step in a cascade that results in full MMN responses (170). This study, however, did not identify the significantly elevated response to the deviant compared to the control that is similar to the MMN typically seen in humans– a pattern only observed in animal models when recorded from larger brain volumes (165).

## **WHAT IS THE EVIDENCE FOR A "CONTROLLED MMR" IN EPIDURAL RECORDINGS?**

Studies using epidural recordings in animals (mainly rats) use electrodes implanted in the skull or sitting on the dura over the auditory cortex (or close to the auditory cortex) to record ERPs elicited during oddball sequences. **Table 2** summarizes these studies. Of these, four use the many-standards control. Ahmed et al. (182) used speech sounds as stimuli, and found larger responses to the deviant sound/ba/in an oddball sequence compared to the same sound presented at the same rate in a many-standards control in anesthetized rats (182). Nakamura et al. (186) and Jung et al. (178) both investigated responses to frequency/pitch deviants in awake rats, and found larger responses to deviant stimuli compared to those in the many-standards control sequence (178, 186). Astikainen et al. (183) examined responses to frequency deviants in *anesthetized* rats and found evidence of deviance detection when the deviant was a high frequency (4.2 kHz) stimulus, not a low frequency stimulus (3.8 kHz), as was also found in Nakamura et al. (3.6 Vs. 2.5 kHz) (183). Nakamura et al. also investigated the effect of duration deviants and identified larger responses to long (but not short) duration deviants compared to the same tones in the many-standards control sequence (186). Therefore all four studies using the many-standards control have indeed observed a response reminiscent of controlled MMR.

Although all these studies all report deviance detection in the rat brain, the morphology of the ERPs differs between each of the studies, with the most dramatic difference in ERPs being observed between the awake recordings (178, 186) and the anesthetized recordings (182, 183). Awake ERPs comprise 3–4 negative and positive components, and the anesthetized recordings feature only one large positive component. The polarity of the deviance-detection effects is also different, with negative deflections in response to the deviant in the awake animals, and positive deflections in anesthetized animals. Such trends can be similarly observed in other studies (**Table 4**),with the majority differences between the deviant and the standard being positive in anesthetized rats and negative in awake rats. These findings illustrate that deviance-detection need not be a response that exactly mimics the human MMN (insofar as being of the same polarity). Indeed human mismatch responses (MMR) change from negative to positive in polarity when the recording location is moved from fronto-centrally sites to the mastoid when recorded using a nose reference (189).

The remainder of studies examining epidural ERPs in oddball paradigms use either the deviant-alone to control for adaptation, or use no adaptation control at all. Although the deviantalone sequence may overestimate the contribution of adaptation (165), three of the five studies utilizing this control sequence find evidence for controlled MMR recorded from epidural electrodes (179, 180, 187). Of those epidural studies that did not implement a control for adaptation, several did not use simple frequency deviants (which are highly predisposed to be affected by adaptation mechanisms in the A1), but rather used more complex deviants that presumably elicited responses that would not be so readily perturbed by neural adaptation. For instance, Roger et al. (181) measured responses to duration deviants [which are not affected by adaptation in the A1 to the same degree as frequency deviants; (169)], and found larger responses to duration deviants (181). In addition, Ruusuvirta et al. (174) examined responses to duration deviants in combination with stimulus onset asynchrony deviants, and found evidence for larger deviant responses (compared to standards) only in the condition in which the deviant occurred earlier than expected and deviated in duration, indicating a possible threshold level for deviance that needs to be reached before a deviance-detection-like response iselicited (174). However, without the use of a many-standards (or even a deviant-alone) control, it remains difficult to conclude that such responses are completely independent of adaptation effects. Although tone duration is coded in a different way to the frequency, and previous studies have demonstrated that adaptation mechanisms for duration are not as robust as they are for frequency (169), this was examined in the auditory cortex only, and it cannot be ruled out that another region (or network of regions) adapts to stimulus duration in the same way as the auditory cortex does for frequency. These studies show that when using methods that detect shifts in potential at a network-level (i.e., a large spatial scale), signs of human-like MMR (independent of adaptation) are evident in animals.

## **HOW DOES ANIMAL RESEARCH AUGMENT HUMAN RESEARCH ON MMN?**

Animal models are an ideal tool with which to investigate the underlying neurobiology of MMR because of the ability to perform more invasive and selective neurobiological manipulations (e.g., drug microinjection to specific regions). In addition, the comprehensive genetic toolkit available for mouse models will enable researchers to determine the role of specific genes and neural populations in the generation of MMR using transgenic animals and optogenetics. With a consensus emerging regarding the ideal recording method (epidural) and the current preferred control to use (many-standards) to examine MMR (adaptation and deviance detection) in animal models, the time is ripe for further studies examining the neurobiology of these elements of MMR, using pharmacological, developmental, and genetic manipulations. Unfortunately, few studies implement a control for adaptation effects (or if they do, do not adequately report drug effects on these), and it is therefore difficult to determine whether or not the agents given affect adaptation, deviance-detection, or both. The preferable way to investigate the pharmacology of MMN in animal models would be test different manipulations in a model that exhibits both adaptation and deviance detection [e.g., Ref. (178)], and to compare how different agents affect the responses to standards, deviants, and control stimuli; with the difference between the control and standard representing adaptation and the difference between the deviant and the control representing MMR and the effects of drug interventions on these two separate components can be examined. To our knowledge, no studies have tested such a model thus far.

Studies in rats and mice (like those on the macaque described above) have been used to study the role of the NMDAGluR system in MMN. These studies, in contrast to Javitt et al. (130), suggest that perturbations in NMDAR signaling can also alter responses to standard stimuli. In the absence of a control for adaptation effects, Ehrlichman et al. found that ketamine (a non-competitive NMDAGluR antagonist) concurrently *increased* the response to the standard (which was small to start with) and *reduced* the response to the deviant, albeit not to a significant degree, in awake mice (172). Tikhonravov et al. used a low (0.1 mg/kg) and high (0.3 mg/kg) dose of intraperitoneal MK-801 to examine the effects of NMDAGluR perturbation on responses to oddball stimuli, in addition to a deviant-alone control in pentobarbital-anesthetized rats (179). Responses to the deviant were more positive than the responses to the standard for the saline condition and were relatively unchanged by the low-dose of MK-801. However, the highdose of MK-801 lowered the response to the deviant and increased the response to the standard, effectively reversing the polarity of the oddball effect. The degree to which MK-801 altered the *magnitude* of the difference between the oddball deviant and the deviant-alone was not reported. However, the period over which the deviant was significantly different from the deviant-alone was reported: there was a positive deflection in response to the deviant (relative to deviant-alone) in saline-treated animals (indicating deviance detection), which was absent after low-dose MK-801 and reversed in polarity after high-dose MK-801 (179). The finding that low-dose MK-801 can reduce the difference between the oddball deviant and deviant-alone, without a change in the differences between deviant and standard could indicate that this dose can preferentially disrupt deviance detection, while sparing adaptation. However, the high-dose of MK-801 not only disrupted both deviance detection and adaptation but reversed the polarity of these changes. Such reversals, caused by an increase in the response to the standard and a decrease in the response to the deviant, are similar to those found after ketamine by Ehlrichman et al.

In a second study, Tikhonravov et al. investigated the effects of a low (3 mg/kg) and a high (10 mg/kg) dose of memantine on responses to oddball stimuli and deviant-alone stimuli in anesthetized rats (180). Like MK-801, memantine is an uncompetitive NMDAGluR antagonist, but unlike MK-801, is a very low-affinity antagonist and has potential as a cognitive enhancer [(190) see also discussion of memantine in Section "How do Pharmacological Manipulations Alter the MMN Process?"]. Similar to previous studies in anesthetized rats, the oddball MMR was positive in saline-treated rats. The low-dose of memantine resulted in a significant increase in the response to the deviant and the oddball MMR was significantly prolonged compared to the saline group. The high-dose of memantine, on the other hand, reduced the time over which the oddball MMR was significantly different from 0 and reversed the late phase of the difference waveform, possibly due to an increase in the response to the standard (180). With regard to deviance-detection-specific changes, the deviant was more positive than the deviant-alone in saline-treated rats, an effect that was prolonged in the low-dose memantine group. In rats treated with the high-dose of memantine, however, no

deviance-detection response was observed. These findings indicate that low-dose memantine potentiates the deviance-detection response, with no significant effect on the adaptation response, but that high-dose memantine acts similarly to high-dose MK-801, affecting both adaptation and deviance detection and reversing the polarity of the oddball MMR.

Overall,these pharmacological studies suggest that NMDAGluRantagonists act in a dose-dependent fashion, with low-dose/lowaffinity antagonists facilitating deviance detection (by increasing the response to the deviant), then as NMDAGluR perturbations are increased with low-dose/high affinity antagonists, deviance detection is inhibited while adaptation is spared. However, with high-doses of high affinity blockers (ketamine and MK-801), the response to the standard is increased, thus indicating impaired adaptation. The selective impairment of deviance detection and not adaptation after MK-801 is also highlighted in (169), where it was reported that subcutaneous MK-801 (maximum dose, 0.1 mg/kg) reduced responses to both the deviant and the standard together, while preserving the difference between them, indicating that this dose *did not* affect neural adaptation in A1, similar to findings in the study in which the same drug was given intraperitoneally (179). These studies indicate that the mechanism by which NMDAGluR-antagonists may reduce the MMN may be rather complex, with the dose and affinity of the antagonist interacting with adaptation and deviance-detection resulting in varied effects on these elements of the MMN. This animal work highlights a complexity in the role of NMDAGluRs in the generation of MMRs that was not discovered in human studies using NMDAGluR-antagonists, most likely due to the smaller dose-range, smaller sample size, and lack of consistent reporting of the response to standards seen in the human work. This animal work therefore highlights the need to examine a larger dose-range, as well as the need for consistent reporting of both standard and deviant responses in human pharmacological MMN studies.

Several of the previously mentioned animal pharmacological studies have similar weaknesses highlighted for the human studies. While (179, 180) used a control for adaptation (the deviantalone control), they did not report directly on the magnitude of responses to the standard, the deviant and the deviant-alone control, thus making the interpretation of the mechanisms affected by pharmacological agents very problematic. While these animal model investigations are still in their infancy, the promise of such models is far-reaching. Future investigations will be able to focus on a range of doses of NMDAGluR-antagonists in paradigms in which adaptation and deviance detection are welldescribed, to further explore the role of NMDAGluR signaling in both of these components. In addition, NMDAGluR-antagonists can be infused directly into regions of interests [as performed by Javitt et al. (130)], to determine where NMDAGluR signaling is important for MMR. Further, high-doses of muscimol [a Gamma-Amino Butyric Acid (A) agonist] can be infused to completely inactivate regions to determine their contribution to deviance detection. For example (174) found evidence of MMR in rat hippocampus – the degree to which these contribute to epidural-recorded MMR can be determined by infusing muscimol into selected hippocampal regions. In addition, animal models of schizophrenia-like reductions in MMN could possibly be developed. (173) used transgenic mice heterozygous for *neuregulin 1* (*nrg1*; an hypothesized schizophrenia-susceptibility gene) and found a reversal in polarity of the oddball MMR in *nrg1* mutants compared to wild type mice. However, this study did not adopt a control for adaptation effects, so it is unknown in *nrg1* plays a role in adaptation or adaptation-independent deviance-detection. Experiments in our lab are currently underway to determine the effect of maternal immune activation on MMR (both adaptation and deviance detection) in rats. Maternal immune activation is a risk factor for schizophrenia and when modeled in rats and mice, is associated with numerous schizophrenia-like behavioral and neurodevelopmental outcomes, particularly those related to NMDAGluR-dysfunction (191–193). This model therefore may also exhibit schizophrenia-like changes in MMR and could be used as a potential experimental platform to examine the pathology underlying schizophrenia-like reductions in the mismatch response, and potential treatments for such alterations.

## **CONCLUSION**

The above review supports several conclusions regarding the potential of MMN as a tool to study the biological processes taking place in those with (and potentially those at-risk for) schizophrenia.

### **MEASUREMENT AND REPORTING OF MMN**

There are pros and cons to any experimental design and it is the authors' opinion that there is no optimal design for use in schizophrenia studies. To recommend an optimal design, while improving consistency in the literature, would come at the expense of the unique contributions that can be made by novel designs [see Ref. (24) for review] that can enrich our understanding of the perceptual inference within the MMN process and how gain in this signal is controlled. Furthermore, study design is often limited by mundane and yet important considerations of test duration that could render particular paradigms less feasible. However, two recommendations arise from this review. The first is that authors and journals facilitate a closer adherence to publication standards. A failing of the current literature is that a large number of studies (including some by our group) do not adhere to recommended publication standards for ERP research which stipulates the display of the original ERPs from which difference waveforms have been derived (194). By comparing research findings across species and drug studies (as above) the importance of at least reporting standard and deviant ERPs becomes clear. Secondly, while the traditional oddball paradigm provides a robust measure of the reduction in MMN amplitude in schizophrenia, protocols designed to identify constituents of the MMN process (or oddball combined with rigorous controls) introduce the capacity to begin disentangling which components of the process are in fact compromised. This advantage becomes crucial when attempting the translation of this research into animal models (see Translation to Animal Models) and arguably also to pharmacological studies (see How do Pharmacological Manipulations Alter the MMN Process?). Where feasible, future studies should consider the advantage of designs (or analysis techniques) that facilitate the differentiation between adaption effects and something more akin to true contextual deviance detection.

## **VULNERABILITY TO SCHIZOPHRENIA**

Our review of the literature to date indicates limited support for small MMN conferring some vulnerability to schizophrenia but we consider this an open question best addressed by large longitudinal studies. It is the authors' opinion that there is much to gained by continued efforts to understand factors that influence MMN size (using novel designs, pharmacological, and animal research) in parallel to such efforts.

## **PHARMACOLOGY AND ANIMAL MODELS OF MMN**

It seems clear from the pharmacological studies (in animals and humans) that adaptation and sensitivity to contextual deviance may show differential effects in the presence of compromised NMDAGluR function. Pharmacological, like animal studies, require indices of both to be informative about the effect of agents on the MMN process. Similar to schizophrenia studies, animal, and pharmacological studies inconsistently report on standard and deviant effects. While our review of pharmacological studies is more supportive of some agents (e.g., glutamatergic and cholinergic) than others (e.g., monoamines), our knowledge of influences on perceptual inference and learning underlying MMN continues to grow and challenge existing models [e.g., Ref. (98, 195)] and it remains possible that future paradigms may be sensitive to agents that current paradigms are not. In other areas of learning, animal models and pharmacology have provided great insight into schizophrenia [associative learning (196)] and there is good reason to suppose that this will also be true of MMN research as well as offering the potential to examine commonalities in underlying pathology. While there are many challenges in translating the potential of MMN in elucidating the pathophysiology of the schizophrenic illness, we believe the current state of research encourages scientists to pursue the many numerous potentially fruitful avenues available to achieve this goal.

## **AUTHOR CONTRIBUTIONS**

Juanita Todd, – lead writer and overall editor. Primary contributions Sections "The Potential – Why is there so Much Interest in MMN in Schizophrenia?" "The Measurement of MMN – Is there an Optimal Paradigm with Which to Study MMN in Schizophrenia?" "How do Pharmacological Manipulations Alter the MMN Process?" and "Conclusion." Lauren Harms – Primary contribution Section "Translation to Animal Models." Reference manager. Patricia T. Michie – Primary contributions Sections "The Measurement of MMN – Is there an Optimal Paradigm with Which to Study MMN in Schizophrenia?"and"Does MMN-Reduction Indicate Vulnerability to Schizophrenia?" Ulrich Schall – co-writer of all sections and preparation of **Tables 1**–**3**.

## **ACKNOWLEDGMENTS**

Our thanks are extended to Ms. Lisa Whitson and Mr. Alexander Provost who sourced many of the papers reviewed for the pharmacology section. Lauren Harms is supported by a National Health and Medical Research Council Project grant (ID1026070). Research referenced in this paper includes that supported by additional National Health and Medical Research Council Project grants (IDs 569259 and 1002995). Ulrich Schall was supported by the Schizophrenia Research Institute utilizing infrastructure funding from the New South Wales Ministry of Health and New South Wales Ministry of Trade and Investment (Australia).

## **REFERENCES**


mismatch negativity and related responses. *Psychophysiology* (2011) **48**:4–22. doi:10.1111/j.1469-8986.2010.01114.x


patients with first-episode schizophrenia. *Am J Psychiatry* (2003) **160**:156–64. doi:10.1176/appi.ajp.160.1.156


schizophrenia. *Neurosci Lett* (2001) **309**:185–8. doi:10.1016/S0304-3940(01) 02072-9


of auditory distraction in cigarette smokers. *Nicotine Tob Res* (2006) **8**:263–73. doi:10.1080/14622200600576669


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 September 2013; accepted: 04 December 2013; published online: 18 December 2013.*

*Citation: Todd J, Harms L, Schall U and Michie PT (2013) Mismatch negativity: translating the potential. Front. Psychiatry 4:171. doi: 10.3389/fpsyt.2013.00171*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Todd, Harms, Schall and Michie. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Contributions of experimental psychiatry to research on the psychosis prodrome

## **Mitja Bodatsch\*, Joachim Klosterkötter and Jörg Daumann**

Department of Psychiatry and Psychotherapy, University of Cologne, Cologne, Germany

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Nicolas A. Crossley, King's College London, UK Jun Soo Kwon, Seoul National University College of Medicine, South Korea Risto Näätänen, University of Helsinki, Finland

#### **\*Correspondence:**

Mitja Bodatsch, Department of Psychiatry and Psychotherapy, University of Cologne, Kerpener Street 62, Cologne 50924, Germany e-mail: mitja.bodatsch@uk-koeln.de

In the recent decades, a paradigmatic change in psychosis research and treatment shifted attention toward the early and particularly the prodromal stages of illness. Despite substantial progress with regard to the neuronal underpinnings of psychosis development, the crucial biological mechanisms leading to manifest illness are yet insufficiently understood. Until today, one significant approach to elucidate the neurobiology of psychosis has been the modeling of psychotic symptoms by psychedelic substances in healthy individuals. These models bear the opportunity to evoke particular neuronal aberrations and the respective psychotic symptoms in a controlled experimental setting. In the present paper, we hypothesize that experimental psychiatry bears unique opportunities in elucidating the biological mechanisms of the prodromal stages of psychosis. Psychosis risk symptoms are attenuated, transient, and often only retrospectively reported. The respective neuronal aberrations are thought being dynamic. The correlation of unstable psychopathology with observed neurofunctional disturbances is thus yet largely unclear. In modeling psychosis, the experimental setting allows not only for evoking particular symptoms, but for the concomitant assessment of psychopathology, neurophysiology, and neuropsychology. Herein, the glutamatergic model will be highlighted exemplarily, with special emphasis on its potential contribution to the elucidation of psychosis development. This model of psychosis appears as candidate for modeling the prodrome by inducing psychotic-like symptoms in healthy individuals. Furthermore, it alters pre-attentive processing like the Mismatch Negativity, an electrophysiological component which has recently been identified as a potential predictive marker of psychosis development. In summary, experimental psychiatry bears the potential to further elucidate the biological mechanisms of the psychosis prodrome. A better understanding of the respective pathophysiology might assist in the identification of predictive markers, and the development of preventive treatments.

**Keywords: experimental psychiatry, psychosis modeling, PCP/NMDA, mismatch negativity, clinical high-risk, prodrome, schizophrenia**

## **INTRODUCTION**

Since more than 10 decades, researchers aim at understanding the neurobiological mechanisms of psychosis. Concomitant with the debut of modern nosology in the late nineteenth century, almost all pioneer thinkers of psychiatry provided theoretical models regarding the underlying biological mechanisms of psychosis even though the experimental techniques at the time did not allow for any empirical evidence. The paradigmatic claim that "all mental disorders are brain disorders"dates back to Griesinger's works (1). Later on, Kraepelin, the ancestor of modern nosology, proposed that "dementia praecox" originates in a misdirected neurodevelopment and insisted on the idea that psychosis is identical to a brain disease (2). Bleuler, who gave birth to the name "schizophrenia," provided a sophisticated theoretical model of disrupted neural association networks contributing to the schizophrenic *Grundsymptome* (3). In this tradition, one of the last hypotheses can be found in the seminal works of the German psychiatrist Gerd Huber, who proposed an elaborated model to trace back subtle psychopathological changes to neurofunctional

disturbances of the limbic system (4, 5). Although not all hypotheses of that kind led to fruitful insights, the increasingly elaborated methods of biological psychiatry partially provided empirical evidence for some models in demonstrating, e.g., reductions in brain volume (6), functional aberrations of association cortices (7), and limbic neuropathology (8), thereby justifying the modeling approach as one primary route to guide empirical and experimental work.

However, the empirical work characterized above necessarily represents a kind of "backwards engineering." Always starting from phenomenology, biological psychiatry is principally conditioned to a verification/falsification dichotomy regarding the primary hypothesis. Thus, this backwards approach is limited by the demarcations given by psychopathology and phenomenological nosology.

In this regard, experimental psychiatry, understood as the modern continuation of ancient theoretical modeling, provides a completive approach to the elucidation of the biological mechanisms of psychosis. The experimental evocation of a psychotic syndrome by psychedelics allows for forward predictions on neurofunctional, cognitive, and psychopathological changes (9). Following this route, experimental psychiatry allows for tracing forwards the consequences of targeted manipulations of neurochemical pathways and for the subsequent comparison with empirical and phenomenological findings.

The birth of experimental psychiatry dates back to the seminal studies of Luby and colleagues (10) who demonstrated that a "schizophrenomimetic" drug evokes psychotic symptoms resembling schizophrenia in healthy individuals (10). Following this approach, Domino et al. were able to evoke comparable symptoms in healthy persons by applying the "dissociative" phencyclidine (PCP)-derivate ketamine (11). In the sequel, empirical findings on the neurochemical features of *N*-methyl-d-aspartate receptor (NMDAR) antagonists as PCP and ketamine first supported the dominating hypothesis of dopamine hyperfunction in schizophrenia (11). However, further evidence demonstrating that *N*-methyl-d-aspartate (NMDA) antagonists critically interact with various regulatory mechanisms of corticolimbic functions that are relevant to schizophrenia led to the establishment of the glutamate hypothesis (11). This hypothesis implies that NMDAmediated dysfunctions play a critical role not only in dopaminergic regulation, which has been reconceptualized as the final common pathway to psychosis, but particularly predict impairments in cortical, sensory, and associative brain regions that contribute to cognitive and negative symptom dimensions (9). Today, it is widely accepted that the glutamate/NMDA pathway represents a discrete pathophysiological aspect of schizophrenia (9, 12, 13).

However, a yet widely neglected aspect of experimental psychiatry represents its potential contribution to the understanding of psychosis development and pre-psychotic, i.e., prodromal stages [but see Ref. (14, 15)]. This aspect, however, may be of particular interest since the neurobiological mechanisms at the very first, subclinical beginning of psychosis development are yet insufficiently understood (16). Thereby, experimental psychiatry may significantly contribute to the identification of targets for preventive treatments.

In the present review, we aim to investigate if and how experimental psychiatry may further elucidate the biological mechanisms of the psychosis prodrome. Thereto, we will firstly give a short overview on the crucial psychopathological, neurocognitive, and neurofunctional findings in the prodrome with special emphasis on neural information processing [functional magnetic resonance imaging (fMRI), electrophysiology]. Secondly, we will exemplarily focus on the well established PCP/NMDA model of psychosis and its potential to mimic the psychosis prodrome.

## **METHODS**

We carried out a computer search of the MEDLINE database. No limits were set regarding the publication date. We used the following Medical Subject Heading (MeSH) categories: (1) (PCP OR NMDA) AND (psychosis OR schizophrenia), (2) [prediction OR ultra-high-risk (UHR) OR clinical high-risk OR at-risk mental state (ARMS) OR basic symptoms (BS)] AND [psychosis OR schizophrenia], and (3) [neurocognition OR cognition OR fMRI OR P50 OR N100 OR sensory gating OR mismatch negativity (MMN) OR P300] AND [UHR OR prodrome] AND [psychosis OR schizophrenia]. Studies on PCP/NMDA were restricted to human subjects. With regard to the psychosis prodrome, studies were included if current at-risk criteria (COPER/COGDIS, UHR) were employed in the respective studies.

## **THE PSYCHOSIS PRODROME**

## **PSYCHOPATHOLOGY**

The prodrome of psychosis is at first a phenomenological concept. It originates in the observation that mental disorders, and particularly schizophrenia, mostly do not appear at a sudden. In general, manifest psychosis represents the severe end of a long-term development in which subtle psychopathological changes appear years before a diagnosis can be validly made (17). Since full-blown psychosis represents a severe disorder with critical long-term consequences, the establishment of prediction and prevention based on prodromal signs of psychosis has become a main goal of clinical research. However, although early clinical signs of psychosis development can validly be traced backwards after an individual has already developed full-blown psychosis, the forward prediction of transition to psychosis is difficult and the respective approaches bear substantial uncertainty regarding their predictions (18, 19). In sum, two predictive approaches are currently implemented.

The BS approach points to subtle, subjectively experienced changes of mental functions that are thought to mark the earliest stages of psychosis development (20, 21). Empirical research has led to two well-defined criteria, pointing either to a collection of highly predictive cognitive and perceptive disturbances (COPER) or to predominantly cognitive disturbances (COGDIS), respectively (22). Subjects qualifying for the COPER criterion developed psychosis in 34.9% within 11 months on average (range 1–37, median 9 months) (23).

The currently most widely used clinical criteria of psychosis prediction point to so called UHR symptoms that are thought to mark the latest stages of psychosis development (24, 25). According to this approach, either attenuated psychotic symptoms or brief, spontaneously remitting psychotic symptoms or a genetic liability in combination with an actual loss of functioning indicate a markedly increased risk for an imminent onset of full-blown psychosis (26). Transition rates in samples identified by the aforementioned criteria amount to 30% on average within the available observation periods (27).

Taken together, the prediction of psychosis development based on clinical criteria inherits significant uncertainty, as mirrored by non-conversion rates of more than 50%, at least within feasible observation periods (27). This observation has led to a paradigm shift in that the aforementioned criteria are thought to identify a "risk-state"probably leading to psychosis rather than a"prodrome" mandatorily leading to manifest psychosis (17).

Although the prospective identification of individuals making the transition to full-blown psychosis thus faces major challenges, it is undoubted that the prodromal development commonly starts from subtle changes in perception and cognition and ends up with attenuated and transient psychotic symptoms, respectively, at the verge of manifest psychosis (28). Furthermore, even though not highly predictive of the further course, negative

symptoms appear at very early stages of the prodromal development, thereby even preceding (pre-)psychotic symptoms (29). Besides psychopathology, however, recent research has suggested that the prodrome can also be validly characterized on other domains, i.e., neurocognition and neurofunctioning (28).

## **NEUROCOGNITIVE FINDINGS**

A huge number of studies demonstrated neurocognitive deficits in individuals at-risk and prodromal subjects, respectively. In particular, studies focusing on working memory, executive functions and verbal fluency/learning were able to provide discriminative statistics for the prospective identification of future converters (30–37). Thereby, investigations employing language based tasks demonstrated that verbal fluency deficits precede psychosis onset up to 30 months (30), and that disturbances of working memory can be found up to 64 months prior to psychosis (31). Executive dysfunctions in the prodrome comprise attention and processing speed which appear as well years before psychosis onset (32, 33, 35, 37).

## **NEUROFUNCTIONAL FINDINGS**

Regarding fMRI investigations, yet two studies compared fMRI correlates of neurocognitive functions in converters (i.e., prodromal subjects) to non-converters. Sabb et al. demonstrated a higher activation of temporal lobes, the frontal operculum, the left precentral gyrus, the caudate, and striatal regions of future converters during the semantic logic condition of a language processing task (38). Allen et al. demonstrated an increased activation in future converters, too, with regard to the left superior frontal gyrus, the middle frontal gyrus, parts of the brainstem, and the left hippocampus in a verbal fluency task (39). Taken together with suggestions of a gradual decline in frontal and striatal activation from the clinical risk state to chronic psychosis (40, 41), particularly regions contributing to language processing seem to be involved in prodromal stages (28). Progressive structural changes during transition to psychosis have been found in the superior temporal gyrus (42).

Regarding electrophysiology, the prodrome seems to be characterized by neuronal disturbances in sensory processing domains. Ziermans and colleagues investigated the Pre-Pulse Inhibition (PPI), a startle response, and suggested a differential deficit in converters vs. non-converters (43, 44). Sensory gating measures (P50/N100) seem to be less relevant to the prodromal development since two out of three studies did not find significant differences between converters and non-converters (45–47). The P3 amplitude, which correlates to memory and attentive processes, has been demonstrated to be exclusively disturbed in future converters by one study (48). Of the published studies evaluating the MMN, a correlate of pre-attentive stimulus discrimination presumably sensitive to the stage of illness (49–54), the majority consistently demonstrated MMN deficits in future converters but not in non-converters (55–60). Bodatsch et al. and later on Perez et al. provided evidence that MMN amplitude deficits predict psychosis onset and allow for an estimation of the remaining time until transition (55, 61). Taken together, correlates of sensory processing and pending higher order functions indicate significant disturbances of neural information processing that may characterize the prodrome (28). Thereby, the MMN might be of particular

interest regarding future research (62, 63) and early intervention strategies (64).

## **THE PCP/NMDA MODEL OF PSYCHOSIS**

### **PCP/NMDA AND PHARMACOLOGY**

Phencyclidine is a non-competitive antagonist of the NMDA glutamate receptor (NMDAR) (65). Comparable substances are MK801 and ketamine, respectively (66). The binding of PCP at the receptor is state dependent, thereby limited to the open channel state, and shows stereo-selectivity (67, 68). Other channels that can be blocked by PCP are voltage-dependent sodium and potassium channels as well as, with different binding features, the nicotinic acetylcholine receptor (69–71). Interactions with membrane proteins have been identified with regard to opioid receptors, dopamine, and noradrenaline transporters, respectively (72–74). However, the main action site seems to be the NMDAR, since all other effects are less potent and only of minor importance in the clinically relevant doses of PCP (66).

Since NMDA antagonists have been demonstrated to produce schizophrenia-like symptoms,clinical investigations aimed at evaluating the potential therapeutic benefit of glutamatergic agonists (75). Studies investigating naturally occurring agonists employed glycine, d-serine, and d-alanine, respectively (75). The results demonstrate that the combination of one of these agonists with an antipsychotic leads to significant improvements in positive, negative, and cognitive symptom ratings (76–82). In particular, two studies provided preliminary evidence that glycine might lead to partial symptom remission in subjects clinically at-risk of developing psychosis (83). Furthermore, it has been demonstrated that the glutathione precursor *N*-acetyl-cysteine improves MMN deficits in schizophrenia patients (84).

## **PCP/NMDA AND PSYCHOPATHOLOGY**

Since PCP has first been described as "schizophrenomimetic" in the first publications on that topic (10), subsequent research has been able to quantify the respective positive, negative, and cognitive symptoms by psychopathological rating scales (9, 75). Krystal et al. (85) demonstrated that ketamine produces behavior similar to schizophrenia as assessed with the Brief Psychiatric Rating Scale (BPRS) (85). Moreover, individuals suffering from schizophrenia display increases in positive and negative symptom ratings after administration of NMDA antagonists (85, 86). Taken together, the psychopathological observations suggest that NMDA antagonists affect a brain system that is vulnerable to psychotic experiences (9, 75).

However, positive symptoms provoked by ketamine administration have been demonstrated being less severe as those observed in clinical psychosis (85–87). Moreover, some psychopathological characteristics of clinical psychosis seem to be underrepresented in the PCP/NMDA model since, e.g., hallucinations are relatively rare in experimental psychosis (85, 86). However, perception distortion is a typical symptom after ketamine administration (85, 86). In particular, ketamine affects the intensity and integrity of sensory stimuli (85, 88), and salience (85, 88–90), respectively. In turn, negative and disorganized symptoms as alogia and formal thought disorder, respectively, represent specific psychopathological effects of ketamine (91–93).

## **PCP/NMDA AND COGNITION**

*N*-methyl-d-aspartate antagonists have been demonstrated to produce a wide range of cognitive deficits. Cognitive functions that can be addressed by ketamine comprise predominantly working memory and executive processing (85, 86, 88, 92–96). Moreover, particular performance deficits after administration of subanesthetic doses of ketamine have been shown for learning/cognitive flexibility and verbal fluency (87), which corresponds to the clinical observation of poverty of speech and circumstantiality after ketamine administration (9). The cognitive deficits produced by NMDA antagonists seem to be rather specifically comparable to schizophrenia, since, e.g., a dissociation between disturbed learning ability but intact ability to retain material once learned can be observed in schizophrenia as well as after administration of NMDA antagonists (9, 75, 97).

Taken together, studies demonstrated particularly the induction of working memory impairments and verbal fluency dysfunction in healthy volunteers. These deficits have been pinpointed to ketamine-induced dysfunctions of frontal and temporohippocampal parts of the brain (87).

## **PCP/NMDA AND NEURAL INFORMATION PROCESSING**

Deficits in information processing have been reliably demonstrated across methods in schizophrenia. NMDA antagonists have been demonstrated to induce changes in surrogate markers of neural information processing in terms of behavioral and performance changes (9, 15, 75, 87). Furthermore, disturbances of information processing have directly been observed by neurofunctional measures after NMDA antagonist administration. Schizophrenialike deficits in MMN generation can be induced by local application of NMDA antagonists as well as by systemic administration in healthy individuals (98–102). In contrast, MMN is not modulated by serotonergic or dopaminergic agonists (103, 104). In turn, other brain potentials, e.g., P300, are significantly affected by administration of other psychotomimetic drugs as psilocybin (105). In normal volunteers, however, reduced MMN amplitudes predict susceptibility to ketamine-induced psychosis (98). In schizophrenia, deficits in pre-attentive tone matching might lead to disturbances of higher order functions as the detection of prosody and auditory emotion recognition (9). In turn, in contrast to the results obtained in animal experiments, NMDA antagonists enhance PPI and startle magnitude (106, 107).

With regard to brain imaging studies, it has been demonstrated that ketamine affects metabolic activity in frontal areas, the cingulum, and the thalamus, respectively (90, 108). Furthermore, ketamine induces an increased dopamine release in the striatum of healthy volunteers (109).

## **DISCUSSION**

## **PCP/NMDA AND THE PRODROME**

The PCP/NMDA model of psychosis has been demonstrated to be particularly suited to mimic certain aspects of psychosis (9, 11, 75, 87). These aspects straddle basic neurophysiological aberrations as well as neurocognitive and psychopathological features. The psychopathological symptom patterns observed after NMDA antagonist administration have been shown to be largely comparable to clinical psychosis (85, 86). Cognitive deficits produced by

NMDA antagonists can similarly be found in schizophrenia (85, 86, 92, 93, 96). Deficits in neurophysiological measures that have been conceptualized as an endophenotype of psychotic disorders have been specifically evoked by the NMDA antagonist ketamine (98, 99).

Although experimental psychiatry has thus proven its ability to advance the neurobiological understanding of schizophrenia, it still faces many criticism. The neurofunctional and psychopathological changes, respectively, evoked in an experimental setting lack many aspects of clinical psychosis (11). Furthermore, experimental psychiatry falls short of modeling the complex brain network disturbances that underlie schizophrenia (11). The evoked changes are moreover transient, thus not able to mimic long term, reciprocal neurodynamics, and represent in sum only partial aspects of the pathophysiological picture. However, mimicking the vast complexity of psychosis in general or schizophrenia in particular shall certainly not be the goal of experimental psychiatry (11). At the foremost, such a modeling approach would distract any well-defined forward predictions. Instead, experimental psychiatry should aim at providing insights in the effects of particular synaptic functions in psychosis (11).

However, it has yet been almost neglected that particularly these alleged imitations of such models provide some unique opportunities in themselves. Psychosis models promoted our understanding of the crucial pathophysiological features of clinical psychosis. Since the prodrome represents "early psychosis," it can be assumed that the neurobiological properties of the prodrome are at large the very same as in full-blown psychosis, merely at an earlier stage. However, the phenotype of manifest psychosis results not only from the latest differentiation of certain neurobiological alterations, but represents a complex interplay of accumulating developmental factors and progressive pathophysiological changes. This pathophysiological progression defines the later stages of psychosis and makes them distinct from their clinical and biological precursors (110). For example, the pattern of MMN deficits in later stages of schizophrenia has been demonstrated to be different from the early and particularly prodromal stages (111). From a clinical perspective, neurobiological factors at early stages of illness might inform primarily about disease trajectories and prognosis, whereas factors at later stages might inform foremost about persistent pathophysiological mechanisms (110). Experimental modeling of the pre-psychotic stages might thus assist to understand the crucial neurobiological pathways that turn the spontaneously remitting clinical high-risk state into a prodromal development, although these states are clinically indistinguishable (28, 55). However, although models that explicitly address the prodrome of psychosis are yet to come, the existing well established models of psychosis may already provide opportunities for research in this respect.

In synopsis of the literature, the PCP/NMDA model of psychosis displays some properties that brings it close to the prodromal stages of psychosis (see **Table 1** for overview). As in the clinical at-risk state, the produced symptoms might resemble in large parts the full-blown psychotic symptomatology (85, 86), but are, however, transient like at least in some high-risk conditions (brief limited intermittent psychotic symptoms). With regard to psychopathology, it has been demonstrated that the


**Table 1 | Comparison of the PCP/NMDA model and the psychosis prodrome**.

positive symptoms induced by NMDA antagonists are less severe than in clinical psychosis (85, 86). This is reminiscent of the attenuated psychotic symptoms that can be found in the prodrome (26). Moreover, changes in the integrity and intensity of sensory perception as well as aberrant salience, as produced by NMDA antagonists (85, 88, 89), might be analogous to the respective BS (21). Besides that, subanesthetic doses of ketamine have been demonstrated to induce particularly negative and disorganized symptoms (91–93) that have been found to precede positive symptoms in the prodromal development (29). Moreover, deficits in verbal fluency and working memory, respectively, which can be evoked by subanesthetic doses of ketamine as well (85, 86, 92–94), have been demonstrated in individuals at-risk and in prodromal stages by neuropsychological and neurofunctional (fMRI) investigations (30–33, 35–39). In clinical studies, the aforementioned cognitive deficits have been demonstrated being predictive of future transition to psychosis in at-risk samples (30–33, 35–37). Finally, the MMN, which has been demonstrated to be predictive of psychosis development (55–58), is rather specifically affected by PCP/NMDA antagonists (9, 75) and altered MMN amplitudes in healthy individuals predicted the individual's susceptibility to PCP induced psychotic experiences (98). The latter aspect suggests that vulnerability as well as resilience to psychotic experiences might be further understood by a detailed elucidation of NMDA antagonist actions in sensory domains. A close relationship between EEG measures, structural brain changes, and glutamate neurotransmission in the psychosis prodrome has already been demonstrated (112, 113). Counterintuitively, however, a PPI deficit as observed in schizophrenia can not be evoked by NMDA antagonists (107), which might illustrate the limitations of modeling. Following the implications of the glutamate hypothesis, at least two studies have yet been able to demonstrate beneficial effects of the naturally occurring agonist glycine in the at-risk state (83).

## **CONTRIBUTIONS OF EXPERIMENTAL PSYCHIATRY TO THE UNDERSTANDING OF PSYCHOSIS DEVELOPMENT**

At least with regard to the PCP/NMDA model of psychosis, some features advocate in favor of the models potential to advance the understanding of psychosis development. As clinical psychosis, however, not all aspects of the prodrome might be sufficiently represented in such a model. In particular, it is yet only speculative

if the symptoms provoked by NMDA antagonists are comparable to prodromal psychopathology (21, 24). Moreover, since psychosis development might proceed via different psychopathological syndromes (21, 24, 26) and more than one pathophysiological pathway (110), it is an open question if and which of these could be best mimicked by psychedelic substances. Furthermore, although much of the neurocognitive and neurofunctional disturbances observed in the prodrome might be evoked by NMDA antagonists, the representation might be rather incomplete.

However, the potential contribution of experimental psychiatry to the understanding of the psychosis prodrome should not be underestimated. As in schizophrenia, prodrome modeling might allow for strong forward predictions and assist in the identification of crucial pathophysiological mechanism as illustrated by the glutamate hypothesis (9, 75), which has been significantly promoted by the results of PCP/NMDA research. Regarding the at-risk state of psychosis, the application of glutamatergic agonists might be understood as a logical consequence of the implications derived from the PCP/NMDA model. Experimental psychiatry might thus not only advance basic research, but assist in the identification of targeted pharmacological interventions in putative prodromal stages of illness.

## **CONCLUSION**

A synopsis of the literature shows that the prodrome of psychosis has been almost neglected by experimental psychiatry and the focus has yet been on manifest psychotic disorders. Since prevention of mental disorders became increasingly relevant in the recent decades, it might be fruitful to further evaluate the potential contribution of experimental psychiatry to this goal. As exemplarily illustrated by the PCP/NMDA model of psychosis, however, many aspects advocate that prodromal stages might be validly mimicked by psychedelic substances. In particular,psychopathological as well as neurocognitive and neurofunctional findings in the prodrome seem to be well represented by the PCP/NMDA model. In this regard, future research should aim at comparing the psychopathological properties of putative prodrome models to the respective clinical observations. Furthermore, neurocognitive and neurofunctional effects, respectively, of psychedelics should be evaluated with regard to those deficits that have been demonstrated being predictive of psychosis development.

In summary, experimental psychiatry bears the potential to further elucidate the biological mechanisms of the psychosis prodrome. A better understanding of the respective pathophysiology might assist in the identification of predictive markers, and the development of preventive treatments.

## **REFERENCES**


113. Stone JM, Day F, Tsagaraki H, Valli I, McLean MA, Lythgoe DJ, et al. Glutamate dysfunction in people with prodromal symptoms of psychosis: relationship to gray matter volume. *Biol Psychiatry* (2009) **66**(6):533–9. doi:10.1016/j. biopsych.2009.05.006

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 31 August 2013; paper pending published: 25 September 2013; accepted: 04 December 2013; published online: 17 December 2013.*

*Citation: Bodatsch M, Klosterkötter J and Daumann J (2013) Contributions of experimental psychiatry to research on the psychosis prodrome. Front. Psychiatry 4:170. doi: 10.3389/fpsyt.2013.00170*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Bodatsch, Klosterkötter and Daumann. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Effects of nicotine on the neurophysiological and behavioral effects of ketamine in humans

#### **Daniel H. Mathalon1,2\*, Kyung-Heup Ahn3,4,5, Edward B. Perry Jr.3,4,5, Hyun-Sang Cho3,6, Brian J. Roach<sup>2</sup> , Rebecca K. Blais <sup>3</sup> , Savita Bhakta<sup>3</sup> , Mohini Ranganathan<sup>3</sup> , Judith M. Ford1,2 and Deepak Cyril D'Souza3,4,5**

<sup>1</sup> Department of Psychiatry, University of California San Francisco, San Francisco, CA, USA

<sup>2</sup> Mental Health Service (116D), San Francisco VA Medical Center, San Francisco, CA, USA

<sup>3</sup> Department of Psychiatry, Yale University School of Medicine, New Haven, CT, USA

<sup>4</sup> Schizophrenia Biological Research Center (116A), VA Connecticut Healthcare System, West Haven, CT, USA

<sup>5</sup> Abraham Ribicoff Research Facilities, Connecticut Mental Health Center, New Haven, CT, USA

<sup>6</sup> Department of Psychiatry, Yonsei University College of Medicine, Seoul, South Korea

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Kevin M. Spencer, Harvard Medical School, USA Verner Knott, University of Ottawa Institute of Mental Health Research, Canada

#### **\*Correspondence:**

Daniel H. Mathalon, Mental Health Service (116D), San Francisco VA Medical Center, 4150 Clement Street, San Francisco, CA 94121, USA e-mail: daniel.mathalon@ucsf.edu

**Background:** N-methyl-D-aspartate (NMDA) receptor hypofunction has been implicated in the pathophysiology of schizophrenia and its associated neurocognitive impairments. The high rate of cigarette smoking in schizophrenia raises questions about how nicotine modulates putative NMDA receptor hypofunction in the illness. Accordingly, we examined the modulatory effects of brain nicotinic acetylcholine receptor (nAChR) stimulation on NMDA receptor hypofunction by examining the interactive effects of nicotine, a nAChR agonist, and ketamine, a non-competitive NMDA receptor antagonist, on behavioral and neurophysiological measures in healthy human volunteers.

**Methods:** From an initial sample of 17 subjects (age range 18–55 years), 8 subjects successfully completed 4 test sessions, each separated by at least 3 days, during which they received ketamine or placebo and two injections of nicotine or placebo in a doubleblind, counterbalanced manner. Schizophrenia-like effects Positive and Negative Syndrome Scale, perceptual alterations Clinician Administered Dissociative Symptoms Scale, subjective effects Visual Analog Scale and auditory event-related brain potentials (mismatch negativity, MMN; P300) were assessed during each test session.

**Results:** Consistent with existing studies, ketamine induced transient schizophrenia-like behavioral effects. P300 was reduced and delayed by ketamine regardless of whether it was elicited by a target (P3b) or novel (P3a) stimulus, while nicotine only reduced the amplitude of P3a. Nicotine did not rescue P300 from the effects of ketamine; the interactions of ketamine and nicotine were not significant. While nicotine significantly reduced MMN amplitude, ketamine did not.

**Conclusion:** Nicotine failed to modulate ketamine-induced neurophysiological and behavioral effects in this preliminary study. Interestingly, ketamine reduced P3b amplitude and nicotine reduced P3a amplitude, suggesting independent roles of NMDA receptor and nAChR in the generation of P3b and P3a, respectively.

**Keywords: schizophrenia, nicotine, ketamine, N-methyl-D-aspartate receptor, nicotinic acetylcholine receptor, event-related potential, mismatch negativity, P300**

### **INTRODUCTION**

Several lines of evidence support a glutamatergic hypothesis of schizophrenia involving *N*-methyl-d-aspartate (NMDA) receptor hypofunction (1–7). Studies with the NMDA receptor antagonist, ketamine, in healthy human subjects have been a cornerstone of the glutamatergic hypothesis of schizophrenia, producing clinical symptoms and cognitive impairments similar to those observed in schizophrenia (4, 5, 7–13). Nicotine has been shown to have cognitive enhancing effects in some (14–16), although not all (17, 18), studies. It has been suggested that the high rate of cigarette smoking in schizophrenia patients may reflect their efforts to use nicotine to "self-medicate" (19–21). In the present, exploratory

study, we examined the interactive effects of ketamine and nicotine in healthy volunteers in order to determine whether impairments in brain function mediated by NMDA receptor hypofunction can be ameliorated by nicotine.

## **PARALLEL EFFECTS OF KETAMINE AND SCHIZOPHRENIA ON CLINICAL, COGNITIVE, AND NEUROPHYSIOLOGICAL FUNCTION**

When administered at subanesthetic doses, ketamine produces an array of transient effects in healthy humans that resemble the positive and negative symptoms and cognitive deficits of schizophrenia (4, 7–13, 22–26). Ketamine has also been shown to produce some of the event-related brain potential (ERP) abnormalities observed in schizophrenia including reductions of two ERP components associated with controlled and automatic processing of deviant stimuli, P300 and mismatch negativity (MMN), as described below.

P300 is a positive voltage ERP component occurring about 300 ms following infrequent deviant stimuli interspersed among frequent "standard" stimuli, typically elicited in auditory or visual "oddball" tasks (27). P300 has been posited to reflect allocation of attentional resources (28–30), stimulus categorization (31), and contextual updating of working memory (32). Multiple brain regions have been implicated as neural generators of the P300, including the temporo-parietal junction and prefrontal cortex (33, 34). P300 comprises two subcomponents, the P3b and P3a that are differentially present depending on task conditions (27, 35–38). The P3b is primarily elicited by infrequent target stimuli, reflects top-down allocation of attention, and has a parietal scalp maximum (27, 36, 37). The P3a is primarily elicited by an infrequent non-target distractor or novel stimuli in an oddball sequence, reflects bottom-up orienting of attention, and has a fronto-central scalp maximum (27, 38–41).

P300 amplitude reduction, particularly auditory P3b, is one of the most widely replicated brain abnormalities observed in patients with schizophrenia (42, 43), although it is also reduced in a number of other psychiatric and neurological disorders (44, 45). Furthermore, several studies have shown that ketamine reduces P300 amplitude in healthy volunteers (46–51), consistent with the possibility that NMDA receptor hypofunction contributes to P300 amplitude reduction in schizophrenia. Interestingly, in the Knott et al. (47) study, the reduction of P300 amplitude by ketamine was only evident in the subgroup of non-smokers, consistent with a possible protective effect of nicotine.

Mismatch negativity is a negative ERP component elicited automatically by infrequent deviant auditory stimuli randomly interspersed among frequent "standard" stimuli (52, 53). MMN has been widely interpreted to reflect auditory sensory echoic memory because the detection of deviance requires an online representation of what has recently been "standard" in the auditory stream (52). More recently, interpretations of the MMN have emphasized its reflection of both short-term (seconds) and long-term (minutes to hours) synaptic plasticity in the service of auditory sensory/perceptual learning since the amplitude of the MMN to a deviant stimulus increases as a function of the number of repetitions of the preceding standard stimulus (54, 55). From this perspective, memory traces of the recent auditory past code predictions of future auditory events, with the MMN signaling a prediction error, when the auditory expectancy is violated by a deviant stimulus (54–56). Auditory deviance along a number of dimensions elicits MMN, including pitch, duration, intensity, and location, among others (52, 57). MMN generators have been localized to the auditory cortex and to the frontal lobes (52). MMN is considered to be largely pre-attentive (52), and it is typically elicited while subjects perform a distractor task.

Mismatch negativity has been shown to be reduced in schizophrenia, particularly the duration-deviant MMN (58–60). Moreover, MMN abnormalities, relative to P300 abnormalities, appear to be more specific to schizophrenia (61, 62). Noncompetitive NMDA receptor antagonism by phencyclidine (PCP)

and ketamine has been shown to reduce MMN in non-human primates (63, 64) and healthy volunteers (26, 46, 65–68), respectively. However, some studies failed to find an effect of ketamine on MMN (50) or showed the effect to depend on the type of auditory deviance or the underlying cortical source examined (65, 67). Recently, Knott and colleagues (66) reported that reduction of MMN with ketamine was only seen in people with a predisposition to experience auditory hallucinations and delusions. Moreover, this effect was blocked when these subjects were chewing nicotine gum, consistent with a possible protective effect of nicotine (66).

## **EFFECTS OF NICOTINE ON NEUROCOGNITIVE AND NEUROPHYSIOLOGICAL FUNCTION**

Nicotine has been shown to enhance cognitive functions including attention, episodic memory, and working memory in humans in some (14–16, 69, 70), but not all (17, 18), studies. The effects of nicotine administration or smoking on neurophysiological measures related to processing of deviant stimuli have been examined in a number of studies (71, 72). In experienced tobacco users, cigarette smoking or nicotine administration has been shown to increase P300 amplitude (73–75), especially in non-smokers (47), and/or to reduce P300 latency (74, 76, 77). However, some studies have failed to show such effects (78–81) or have found effects on visual, but not auditory, P300 (77). In addition, some studies (75, 82, 83), but not others (84), have shown reduced P300 amplitudes in chronic smokers, suggesting a distinction between the effects of acute and chronic exposure to nicotine. Indeed, Knott and colleagues reported that the enhancement of P300 produced by nicotine administration was only evident in non-smokers (47).

Regarding MMN, some studies have reported that nicotine or nicotinic agonists increase MMN amplitude (85–88) or shorten MMN latency (86, 89, 90), whereas others have failed to show such amplitude increases (89–91) or latency reductions (66, 88, 91) with nicotine administration.

## **NICOTINE AND SCHIZOPHRENIA**

There is a high prevalence of cigarette smoking among patients with schizophrenia (92–95). This may reflect an effort by schizophrenia patients to "self-medicate" clinical symptoms and a number of neurocognitive impairments including deficits in attention and memory (19) and deficient sensory gating (96). Thus, there is significant interest in developing nicotinic acetylcholine receptor (nAChR) agonists to target the neurocognitive symptoms of schizophrenia. Indeed, an alpha7 nicotinic agonist, 3-[(2,4-dimethoxy) benzylidene] Anabaseine (DMXAB) has been shown to produce significant improvements in cognitive function and P50 sensory gating in patients with schizophrenia (97).

Several studies have examined the effects of nicotine administration on MMN in patients with schizophrenia. Regarding MMN amplitude, a study of schizophrenic smokers showed that nicotine increases the amplitude and the latency of duration-deviant MMN, but shows no effect on frequency-deviant MMN (98). An earlier study from the same group examined the effects of nicotine on a mixed sample of smoking and non-smoking schizophrenia patients with high levels of auditory hallucinations and found no effects on MMN amplitudes in response to duration, frequency, and intensity deviants, although the latency of the

intensity-deviant MMN was shortened by nicotine (99). Another study of non-smoking schizophrenia patients failed to show any effect of nicotine on a frequency-deviant MMN (90).

## **NICOTINIC RECEPTOR MODULATION OF GLUTAMATERGIC NEUROTRANSMISSION**

One possible mechanism by which nAChR agonists might enhance neurocognitive and neurophysiological function is facilitating glutamatergic neurotransmission via presynaptic nAChR (100) or via GABA interneurons (101, 102). Nicotine or nAChR agonists have been shown to facilitate glutamatergic transmission in rat prefrontal cortex (103, 104) and hippocampus (105). Specifically relevant to this study, nicotine attenuates or reverses memory and attentional deficits induced by the NMDA receptor antagonist MK-801 (dizocilpine) in rats (106, 107). Moreover, dizocilpine blocked nicotinic enhancement of memory consolidation in mice (108). Knott et al. (47) examined the effects of nicotine and a sub-perceptual dose of ketamine on P300 in men and women, smokers and non-smokers. In non-smokers, ketamine reduced P300, an effect that did not interact with nicotine. However, in the third assessment block, following the drug infusion, nicotine increased P300 amplitude on its own but further reduced P300 when combined with ketamine. In a subsequent study, Knott and colleagues (66) found that ketamine reduces the amplitude of MMN in healthy individuals with a high propensity toward hallucinatory experiences and/or delusional thinking, an effect that was blocked by nicotine. Similar effects were not evident in individuals with a low propensity toward these psychotic symptoms.

### **THE RATIONALE AND HYPOTHESES FOR CURRENT STUDY**

It is difficult to isolate and study the NMDA receptor hypofunction and its consequences in schizophrenia. The ketamine paradigm in healthy subjects offers a pharmacological model for investigating nicotine's effect on putative NMDA receptor hypofunction in schizophrenia. Specifically, this study examined the effects of nAChR activation on NMDA receptor hypofunction by investigating the interactions of ketamine, a non-competitive NMDA receptor antagonist, and nicotine, a nAChR agonist, when administered to healthy volunteers in a placebo-controlled study over four test days. The neurophysiological outcome measures, chosen based on their established sensitivity to schizophrenia, consisted of: (1) two variants of the auditory P300, the P3b elicited by target stimuli and the P3a elicited by novel distractor stimuli, and (2) the MMN elicited automatically by a duration-deviant auditory stimulus. The primary and secondary hypotheses were that nicotine would attenuate the neurophysiological abnormalities and schizophrenia-like clinical symptoms induced by ketamine, respectively.

## **MATERIALS AND METHODS**

#### **RESEARCH PARTICIPANTS**

The study was approved by the Institutional Review Boards of VA Connecticut Healthcare System (West Haven, CT, USA) and Yale University School of Medicine (New Haven, CT, USA). Subjects were recruited via public advertisements and were paid for their study participation. Written informed consent was obtained from all subjects. Smokers who were not interested in quitting and lifetime non-smokers who had tried nicotine in the past were both invited to participate. Subjects were medically healthy by physical examination, history, electrocardiography, and laboratory testing. They had no history of a DSM-IV Axis-I disorder (other than nicotine dependence), major current or recent (<6 weeks) life stressors, and first-degree relative with a history of psychosis. Screening procedures included the Structured Clinical Interview for DSM-IV, Non-Patient Edition (109), selected sub-tests from the Wechsler Adult Intelligence Scale (Information, Vocabulary, Block Design, Picture Completion) to provide an estimate of general level of cognitive ability (110), and the Fagerström Test for Nicotine Dependence (111) to measure the severity of nicotine dependence in smokers. Subjects were instructed to refrain from consuming psychoactive substances from 1 week prior to testing through completion of the study. An outside informant identified by the subject was interviewed to corroborate information provided by potential subjects. Urine toxicology testing was performed at screening and on the morning of each test day to rule out recent illicit substance use. Subjects were instructed to fast overnight and abstain from smoking after 11:00 p.m. prior to arrival for each test day. They were excluded if breath carbon monoxide levels were higher than 10 ppm.

Seventeen subjects participated in at least one test day. Eight of the 17 subjects scheduled for four test days did not complete testing: 3 of 5 (60%) non-smokers and 5 of 12 (42%) smokers. The reasons for discontinuation were mostly related to adverse effects of ketamine or nicotine (*n* = 6). Adverse events and study discontinuations were reported to the VA Connecticut Human Studies Subcommittee. As with all of our prior ketamine studies, clinical follow-ups indicated that all adverse events associated with acute ketamine administration resolved spontaneously without any late appearing or persistent adverse effects. There were no significant differences in age, sex, education, smoking status, or Fagerström Nicotine Dependence scores between study completers and non-completers. Only nine subjects completed all four test days. Demographic data for these nine completers are presented in **Table 1**. Of the nine completers (five men, four women), one woman was excluded from the ERP analyses because she was too somnolent and impaired to perform the oddball task or the MMN distractor task during the Ketamine Alone test day. In addition, one man was dropped from the ERP MMN analysis because of technical problems running the MMN paradigm during his Nicotine Alone test day.

#### **Table 1 | Demographic data for study completers (N** = **9)**.


## **TEST DAYS**

Across the four test days, subjects received ketamine and nicotine in a double-blind, randomized, 2 × 2 crossover design. On each test day, subjects received ketamine (a bolus of 0.26 mg/kg over 1 min, followed by maintenance infusion at 0.65 mg/kg/h × 2 h) or placebo (normal saline). Fifteen minutes after the ketamine bolus, subjects received an intravenous infusion of nicotine (1.0µg/kg/min × 10 min) or placebo followed by another nicotine or placebo injection 70 min later. The reason for two spaced nicotine injections was to attempt to minimize the possibility of desensitization that is known to occur with nicotine exposure. The dose of nicotine administered with each infusion (1.0µg/kg/min × 10 min) = 0.7 mg in a 70-kg individual. A regular cigarette contains about 1.2–1.4 mg nicotine and an average of 0.88 mg of nicotine is delivered to a smoker from each cigarette. The timing of procedures is detailed in **Table 2**. Behavioral ratings were obtained at baseline and repeated periodically after the administration of ketamine and nicotine, but ERP data were collected only once per test day. Plasma ketamine and nicotine levels were measured after each infusion to rule out any pharmacokinetic interactions.

## **BEHAVIORAL MEASURES**

The schizophrenia-like clinical symptoms induced by ketamine were assessed using the Positive and Negative Syndrome Scale (PANSS) (112). Perceptual alterations were assessed using the Clinician Administered Dissociative Symptoms Scale (CADSS) (113). The following subjective states were rated by participants using 100-mm Visual Analog Scales (VASs) (10): Talkative, Happy, Drowsy, Tense, Dad, Calm, Depressed, Anxious, Energetic, Fearful, Mellow, High,Angry, Mania, Irritable, Tired, Hungry, and Craving.

## **ANALYSIS OF NICOTINE AND KETAMINE LEVELS**

Plasma ketamine and norketamine were assayed using the identical method as described in detail elsewhere (114). Plasma nicotine concentrations were assayed using reversed-phase highperformance liquid chromatography (HPLC) based on a modification of a previously described method (115, 116). The nicotine assay involved HPLC/MS operated in the APCI/SIM mode using deuterated nicotine as an internal standard. After addition of the internal standard the plasma is deproteinized with sulfosalicylic acid and the supernatant made alkaline and extracted with heptane methylene chloride 85:15. This solvent is then dried down via vacuum centrifuge. The residue is redissolved in ethanol and an aliquot is injected into the HPLC. The HPLC column (Nova Pak C18 30 cm × 3.9 mm, 4µm) is run in the isocratic mode using methanol acetonitrile ammonium formate (pH 4.0) 32.5:42.5:35.0 as mobile phase. The standard curve encompassing a range of 1– 200 ng/ml was linear with negligible intercept. Plasma controls containing 4, 40, and 80 ng/ml nicotine in six consecutive runs demonstrated an inter-assay relative standard deviation RSD of 8.6, 7.4, and 8.3%, respectively.

## **ERP MEASURES**

Electroencephalography (EEG) data were recorded using a 23 channel Physiometrix electrode cap. The cap included one ground electrode placed on the forehead (FPz), and the mean of freely

#### **Table 2 | Study procedures**.


PANSS, Positive and Negative Syndrome Scale; CADSS, Clinician Administered Dissociative Symptoms Scale; VAS, Visual Analog Scale; VS, vital signs; MMN, mismatch negativity; AX-CPT, AX-Continuous Performance Test.

placed bilateral earlobe electrodes served as the reference channel. Vertical and horizontal electro-oculograms (VEOGs and HEOGs) were recorded and used to correct EEG data for eye blink and eye-movement artifacts. Electrode impedances were maintained at <5 kΩ. The data were recorded using Neuroscan Synamps amplifiers, which were calibrated prior to each session. Data were acquired using a 0.05–100-Hz band pass filter, and the sampling rate was 1000 Hz.

P300 was elicited during an auditory oddball target detection task. Three types of stimuli were delivered through Etymotic ER-3A insert earphones: (1) standard tones: 500 Hz pure tones (rise/fall 5 ms; 50 ms duration), (2) target tones: 1000 Hz pure tones (rise/fall 5 ms; 50 ms duration), and (3) novel distractor sounds, selected from a corpus of novel sounds (average duration of 250 ms) developed by Friedman et al. (117). All auditory stimuli were presented at an identical sound pressure level (~80 dB SPL C scale). The task was presented in three blocks. Each block comprised of 150 pseudo-randomized stimuli (80% standards, 10% targets, 10% novel distractors) presented with a stimulus onset asynchrony (SOA) of 1250 ms. Subjects were instructed to press a response button with the index finger of their dominant hand each time a target tone occurred, giving equal emphasis to speed and accuracy.

For the MMN paradigm, subjects were presented with a pseudorandom sequence of standard tones (90% probability; 633 Hz; 5 ms rise/fall time; 50 ms duration) and duration-deviant tones (10% probability; 633 Hz, 5 ms rise/fall time, 100 ms duration) presented at 80 dB SPL, with a 510-ms SOA. A long durationdeviant MMN paradigm was chosen because of some evidence that it may more sensitive to the effect of schizophrenia than other types of MMN (58, 60, 61). The MMN paradigm was presented in two blocks, with each block comprising 783 standard tones and 87 deviant tones. These tones were presented binaurally through earphone inserts, while subjects performed a visual AX-Continuous Performance Task (AX-CPT) (118). Because several of the behavioral performance files from this task were irretrievably corrupted resulting in many subjects with missing behavioral performance data, the AX-CPT performance data were not analyzed in the current study. Thus, the AX-CPT essentially served as the distractor task during MMN recording.

## **ERP DATA PROCESSING**

As the MMN and P300 measures generally achieve their maximum amplitudes along the midline and do not typically show hemispheric lateralization, EEG data from the midline fronto-central sites (Fz, Cz) and fronto-central-parietal sites (Fz, Cz, Pz) were analyzed for the MMN and P300 components, respectively. The processing pipeline for the P300 elicited during the three-stimulus auditory oddball task involved the following steps: continuous data were separated into 1000 ms epochs time-locked to stimulus onset, with a 100-ms pre-stimulus baseline. VEOG and HEOG data were used to correct EEG for eye-movements and blinks with a regression-based algorithm (119). After baseline correction, epochs containing artifacts (voltages exceeding ±100µV) were rejected. P300 was identified as the most positive peak in a 235- to 400-ms time window following stimulus onset; however, because target P3b and novelty P3a have different topographies, different rules were used for identifying their peaks. The target P3b peak was first identified at Pz, then a 50-ms window (±25 ms) surrounding this peak's latency was used to identify target P3b peaks at other sites. Novelty P3a showed more scalp variability in peak latency than target P3b, particularly at frontal sites, leading us to adopt a more flexible peak identification approach. Novelty P3a peaks were first identified at all central and parietal sites. From the range of peak latencies obtained at central sites (C3, Cz, C4), minimum and maximum latencies were identified. By subtracting 50 ms from the minimum and adding 50 ms to the maximum, the search window for identification of P3a peaks at frontal sites was defined. Somewhat early latency cut-off (400 ms) for auditory P300s was chosen to avoid picking the second late positive component, which peaked around 550 ms (see **Figure 1**). Peak amplitudes and latencies for target P3b and novelty P3a were extracted from electrodes Fz, Cz, and Pz for statistical analyses.

The same eye-movement correction and artifact rejection criteria used in the P300 data processing pipeline were applied to the MMN standard and deviant trials, but these data were segmented into 550 ms epochs and baseline corrected using the 50-ms

**FIGURE 1 | Event-related brain potential grand average waveforms (left) and corresponding topographic maps (right) are shown for placebo (black), ketamine alone (red), nicotine alone (blue), and ketamine** + **nicotine (magenta) days**. ERPs, overlaid for each test day, are shown to oddball targets at Pz (top row), to oddball novels at Cz (middle row), and to difference waveforms (deviants-standards) at Cz. The oddball target elicited a P3b, the oddball novel elicited a P3a, and the

deviant elicited a MMN, with each peak denoted by an arrow on the ERP waveforms. Amplitude in microvolts is on the y-axis, and latency in milliseconds is on the x-axis. Stimulus onset is at 0 ms. Negativity is plotted down. Scalp topography maps are shown for each test day for each stimulus, at the peak latency for P3b (top), P3a (middle), and MMN (bottom). Hot colors indicate positive voltage and cool colors, negative voltage.

preceding tone onset. Standard and deviant trials remaining after artifact rejection were averaged separately, and the resulting ERP for the standard was subtracted from the deviant to create a difference wave. MMN was identified as the most negative peak between 90 and 270 ms post-tone onset in the difference wave at electrodes Fz and Cz. Peak amplitudes and latencies were extracted for statistical analyses.

## **STATISTICAL ANALYSES Behavioral data**

Initially, behavioral data were examined descriptively using means, standard deviations, and graphs. Each measure was tested for normality using Kolmogorov-Smirnov test statistics and normal probability plots. All PANSS, CADSS, and VAS measures were highly skewed. Thus, these non-normal behavioral data were first ranked and then fitted using a mixed-effects model with an unstructured variance-covariance matrix and *p*-values adjusted for Analysis of Variance (ANOVA)-type statistics (ATS). In these models, Ketamine (active vs. placebo), Nicotine (active vs. placebo), and Time (−110, −19, +35, +110, and +180 min) were included as withinsubjects explanatory factors, while Subject was the clustering factor. Time reflected the time point, in minutes, relative to Time 0 when the first intravenous infusion of active-nicotine or placebonicotine was initiated (see **Table 2**). All two- and three-way interactions were modeled. Significant interactions were followed by appropriate *post hoc* tests and graphical displays to interpret the effects. All results were considered statistically significant at *p* < 0.05. Bonferroni correction was applied within but not across

domains. Thus, for example, a cut-off alpha level of 0.05/2 = 0.025 was used to declare effects significant for the two CADSS ratings (Subject- and Clinician-Rated).

## **ERP data**

For the ERP data, which were collected once per test day, repeatedmeasures ANOVAs were conducted with Ketamine (active vs. placebo) and Nicotine (active vs. placebo) as within-subjects factors. The ANOVA models assessing P300 amplitudes and latencies included two additional within-subjects factors: Deviant Type (target vs. novel) and Lead (Fz vs. Cz vs. Pz). The ANOVA model for MMN amplitude and latency included one additional withinsubjects factor: Lead (Fz vs. Cz). Analyses proceeded in a hierarchical fashion, with higher order interactions being parsed by examining lower order simple main effects and interactions. Ultimately, condition comparisons were tested with single degree of freedom contrasts.

## **RESULTS**

### **BEHAVIORAL DATA**

Ketamine produced significant increases in PANSS Total, CADSS Subject-Rated and Clinician-Rated Perceptual Alterations, and VAS subjective "High" ratings (see **Figure 2**). All Ketamine main effects and Ketamine × Time interactions were significant at *p* < 0.0001 (**Table 3**). *Post hoc* analyses showed significant effects of Ketamine at time points −19, +35, and +110 (all *p* < 0.05) for each of these measures. There were no significant main effects of Nicotine. Nor were there any significant Ketamine × Nicotine

**FIGURE 2 | Mean and standard errors are plotted for Positive and Negative Syndrome Scale (PANSS) total scores (upper left), Visual Analog Scale of subjective states (upper right), subject-rate (lower-left) and clinician-rated (lower-right) perceptual alterations using the**

**Clinician Administered Dissociative Symptoms Scale (CADSS)**. For each plot, values for each of the four test days are overlaid, for Placebo (black), Ketamine Alone (red), Nicotine Alone (blue), and Ketamine + Nicotine (magenta) days.


in italics. Significant p-values are shown in bold. Num df, numerator degrees of freedom.

or Ketamine × Nicotine × Time interactions for any of these outcome measures.

### **P300 AMPLITUDE**

Event-related brain potential overlays showing P300 waves and topographic maps are shown in **Figure 1**, P300 peak amplitude and latency means are presented in **Table 4** and histograms showing the effects of the drug conditions on P3b and P3a are shown in **Figure 3**. Results of the Ketamine × Nicotine × Deviant Type × Lead repeated-measures ANOVA for P300 amplitude and latency are presented in **Table 5**.

In terms of main effects, only the effect of Lead was significant, with contrasts indicating equivalent P300 amplitudes at Pz and Cz that were both larger than P300 amplitude at Fz. In terms of two-way interactions, there were significant Ketamine × Lead (*p* = 0.02) and Deviant Type × Lead (*p* = 0.001) effects, with a trend (*p* = 0.057) toward a Deviant Type × Nicotine effect. The Ketamine × Lead effect was parsed by examining the main effects of Ketamine separately for each of the three midline leads, with results showing ketamine to significantly reduce P300 amplitude at electrode Cz (*p* = 0.046), but not at Fz (*p* = 0.384) or Pz (*p* = 0.11). This ketamine-induced reduction of midline vertex

**Table 4 | Means and standard errors for auditory oddball P300 amplitude and latency across the four test sessions**.


P300 amplitude did not significantly depend on Deviant Type (*p* = 0.389). The Deviant Type × Lead effect was parsed by examining lead effects separately for targets and novels, both of which were significant. These Lead effects reflected the typical midline scalp topographies of target P3b amplitude (i.e., Fz < Cz < Pz) and novelty P3a amplitude (i.e., Fz < Cz = Pz). The Deviant Type × Nicotine trend was parsed by examining the main effect of Nicotine for each Deviant Type separately. Nicotine significantly reduced the amplitude of novelty P3a (*p* = 0.02) but not target P3b (*p* = 0.737). No other main effects or interactions were significant.

## **P300 LATENCY**

There was a significant main effect of Ketamine (*p* = 0.043) indicating that ketamine delayed P300 latency by 25.44 ms relative to placebo (**Tables 4** and **5**). There was also a significant main effect of Lead (*p* = 0.018) indicating that P300's peak latency was significantly shorter at Cz than at Pz and Fz. No other main effects or interactions were significant (see **Table 5**).

## **MISMATCH NEGATIVITY AMPLITUDE**

Event-related brain potential overlays showing MMN difference waves and topographic maps are shown in **Figure 1**, MMN peak amplitude and latency means are presented in **Table 6**, and histograms showing the effects of the drug conditions on MMN are shown in **Figure 4**. The ANOVA results for MMN amplitude are presented in **Table 7**. None of the main effects were significant, but there were significant Nicotine × Ketamine and Nicotine × Ketamine × Lead interactions. The Nicotine × Ketamine × Lead three-way interaction was parsed by examining the Nicotine × Ketamine effect separately for Fz and Cz. The Nicotine × Ketamine effect was significant at Cz



ANOVA results are based on mutlivariate assumptions for repeated-measures, and all F-tests are based on Wilks' Lambda. Follow-up ANOVA results are shown in italics. Significant p-values are shown in bold.

## **Table 6 | Means and standard errors for mismatch negativity amplitude and latency across the four test sessions**.


(*p* = 0.015) but not at Fz (*p* = 0.347). The Nicotine × Ketamine effect at Cz and the overall Nicotine × Ketamine two-way interaction (averaged over leads) were parsed by examining the main effect of each drug condition separately for the active and placebo days of the other drug condition. Nicotine Alone produced a trend level reduction of MMN amplitude relative to Placebo (*p* = 0.058 for Cz; *p* = 0.084 for average of Fz and Cz), but this Nicotine effect was not evident when Nicotine + Ketamine was compared to Ketamine Alone. In contrast, Ketamine did not significantly affect MMN amplitude when administered alone or along with Nicotine. No other interaction effects were significant.

## **MISMATCH NEGATIVITY LATENCY**

Analysis of variance results for MMN latency are presented in **Table 7**. There was a significant main effect of ketamine (*p* = 0.025) indicating that ketamine shortened MMN latency by 19.64 ms relative to placebo. No other main effects or interactions were significant.

## **PLASMA DRUG LEVELS**

Mean plasma levels for ketamine, norketamine, dehydroketamine, and nicotine are presented in **Table 8**. There were no significant differences in levels of plasma ketamine, norketamine, or dehydroketamine levels between the ketamine alone condition and the ketamine–nicotine condition. In addition, there were no significant differences in plasma nicotine levels between the nicotine alone condition and the ketamine–nicotine condition.

## **DISCUSSION**

The principal findings of the current study are the differential effects of ketamine and nicotine on MMN and P300, and their interactive effects on MMN.

#### **EFFECTS OF KETAMINE**

Consistent with previous studies (4, 5, 7–13, 120), ketamine induced transient schizophrenia-like behavioral effects in healthy subjects. In terms of the electrophysiological measures, ketamine decreased the amplitude and delayed the latency of P300, regardless of whether P300 was elicited by a target or novel stimulus. The decrease in amplitude is consistent with the other ERP studies showing ketamine to reduce P300 amplitude at parietal leads (46– 51), although we did not observe the previously reported increase in novelty P3a at frontal leads with ketamine (51). Our results are also consistent with a prior study showing ketamine to decrease the amplitude of the late positive potential in a working memory task (9). These results provide evidence that glutamatergic neurotransmission at NMDA receptors contribute to P300 generation, both in response to infrequent target stimuli (P3b) and infrequent novel stimuli (P3a). Moreover, inasmuch as P300 amplitude reduction and latency delay are well established in schizophrenia (42, 43, 121, 122), our findings are consistent with the NMDA receptor hypofunction model of schizophrenia (1–3, 5, 6, 123, 124) and its possible role in mediating P300 deficits.

The current study did not find ketamine to significantly reduce MMN amplitude. This conflicts with a number of previous studies (26, 46, 65–68), but is consistent with some studies that either failed to show a ketamine effect on MMN (50) or showed the ketamine-induced MMN reduction to be limited to a subset of task conditions or cortical source locations (26, 65, 67), or to a subgroup of subjects (66). The discrepant results across these studies may be due to differences in the dosage and dosing schedule of ketamine. For example, the study with the most robust effects (26) used a high dose of ketamine (0.9 mg/kg), while the study with a non-significant result (50) used a relatively low dose (0.3 mg/kg). Heekeren and colleagues (65) also demonstrated dose-dependent changes in MMN amplitude using two different doses of ketamine (0.1–0.15 and 0.15–0.20 mg/kg). The absence of a significant ketamine effect on MMN in our study may also have been related to the relatively small size of our subject sample, resulting in limited power to detect an effect. For the ketamine vs. placebo effect on MMN amplitude during the placebo-nicotine day, the effect size (Cohen's *d*) was estimated to be −0.43. This effect size, which appears to be smaller than the effects reported in prior studies showing ketamine to reduce MMN, would reach significance (*p* < 0.05) with a sample of about 25 subjects. This underscores the limited power in our current study, and points to an effect of ketamine on MMN that may emerge with moderate sample sizes.

However, it is noteworthy that our sample size was sufficiently large to detect robust psychotomimetic effects of ketamine, as well as significant ketamine-induced reductions and delays in the P300 ERP component. Thus, consistent with the report of Oranje and colleagues (50), the P300 ERP component appears to be more sensitive to the effects of NMDA receptor blockade than the MMN component. In terms of MMN latency, ketamine reduced MMN latency by about 19 ms. This effect has not been previously reported to our knowledge, and therefore should be regarded as preliminary pending replication in future studies.

#### **EFFECTS OF NICOTINE**

To our knowledge, this is the first study to show nicotine to reduce the novelty P300 (P3a) in humans. This unexpected finding appears to conflict with the plethora of evidence showing nicotine to enhance cognitive functions, including attention (14–16). It is possible to construe the novelty P3a reduction by nicotine as a reflection of enhanced focus on the target detection task and reduced susceptibility to distraction by non-target distractors. However, such an interpretation is not consistent with other evidence showing P3a reduction in patients with schizophrenia (43, 121, 125–128) and patients with frontal lobe lesions (129–131), two conditions known to be associated with attentional impairments and increased distractibility. Thus, nicotine's reduction of the P3a response to novel distractors is unlikely to be a reflection of its cognitive enhancing effects.

We did not observe significant effects of nicotine on target P300 (P3b), inconsistent with some prior reports showing nicotine to increase P300 amplitude (47, 73, 74) and decrease P300 latency (75, 77) in smokers. However, our results are consistent with other studies reporting no effects of nicotine on P300 (76, 78–80). Our study was relatively unique in its use of the intravenous route for nicotine administration, which may partially account for inconsistencies between our results and some prior studies. More generally, inconsistencies among studies may also be related to differences in the type of nicotinic agonist and dosage used, differences in sensory modality of the oddball task used to elicit the P300, and different representations of smokers and non-smokers in the study samples.

The differential effects of nicotine on P3a and P3b in our study is consistent with other evidence that the neuroanatomical (33)

### **Table 7 | Analysis of variance (ANOVA) of MMN peak amplitudes and latencies**.


ANOVA results are based on mutlivariate assumptions for repeated-measures, and all F-tests are based on Wilks' Lambda. Follow-up ANOVA results are shown in italics. Significant p-values are shown in bold.



Ketamine levels were assayed only on the days that subjects received ketamine (ketamine alone and ketamine + nicotine) and nicotine was assayed only on the days that subjects received nicotine (nicotine alone and nicotine + ketamine).

and neurochemical (71) underpinnings of P3a and P3b are at least partially dissociable. Polich and Criado (71) proposed dopaminergic/frontal processes for P3a generation and locus coeruleusmediated noradrenergic/parietal processes for P3b generation. Evidence for this includes demonstrations that chronic abuse of different street drugs are associated with differential effects on P3a and P3b amplitudes (71). Nonetheless, roles for nicotinic cholinergic neurotransmission, as well as glutamatergic neurotransmission, have not figured prominently in prior neurochemical models of P300 generation.

Nicotine alone produced a trend level reduction of MMN amplitude, but this effect was not evident when comparing the Nicotine + Ketamine condition to Ketamine alone. These results conflict with some prior studies showing nicotine or nicotinic agonists to enhance MMN amplitude in response to duration (85), frequency (86, 87), or inter-stimulus interval (88) deviants in healthy volunteers, and similarly failed to corroborate studies showing nicotine to enhance duration-deviant MMN amplitude in schizophrenia patients (98). One possible reason for the discrepancy between our findings and those reported previously is that our study used an intravenous route of nicotine administration whereas prior studies used either gum (85, 88, 98) or a transdermal patch (87). While differences in the pharmacokinetics and pharmacodynamics between intravenous vs. gum and transdermal routes of administration have not been systematically studied', it is likely that time to onset of action and peak levels achieved would differ between these modes of nicotine delivery, and such differences could account for variability in the effects of nicotine on MMN. At the same time, it should be noted that a number of studies using gum (47, 99) or transdermal patches (89, 90) failed to demonstrate an enhancement of MMN amplitude by nicotine. Moreover, the fact that our study focused on duration-deviant MMN, in part because of evidence that it is more sensitive to schizophrenia than other types of MMN (58, 60, 61) cannot be the reason we failed to observe enhancement by nicotine, since at least two prior studies have shown the amplitude of the durationdeviant MMN to be increased by nicotine [Ref. (85, 98); but, see Ref. (99)].

#### **COMBINED EFFECT OF KETAMINE AND NICOTINE**

Discordant with the study hypothesis, nicotine did not improve either the behavioral or neurophysiological abnormalities induced by ketamine. Of the many drugs tested in the ketamine model, few have been shown to reduce the schizophrenia-like behavioral and cognitive effects of ketamine in healthy human subjects. Lamotrigine, but not haloperidol or lorazepam, has been shown to reduce some of the behavioral and cognitive symptoms induced by ketamine in healthy volunteers (132–134). With previous findings from animal and human studies documenting cognitive enhancing effects of nicotine in humans (135–138) and animals (139, 140), including animal data showing nicotine to ameliorate NMDA-antagonist induced cognitive deficits (106, 107) or NMDA-antagonists to block cognitive enhancing effects of nicotine (108), it was surprising that nicotine did not show any tendency to reverse ketamine's psychotomimetic or cognitive ERP effects. Inconsistencies among studies may be due to differences in nicotine dose, rate, and route of nicotine administration, and the smoking status of the subjects tested. Importantly, our results are consistent with other studies showing that nicotine did not block ketamine's deleterious effects on P300 (47) or on neurocognitive test performance (141) in humans, suggesting that any pro-cognitive effects of nicotine may not be able to overcome the impairments produced by NMDA receptor blockade. However, our results were not consistent with a prior study that showed nicotine to prevent ketamine's reduction on MMN amplitude, an effect that was only observed in the subgroup of healthy volunteers with a high propensity to have hallucinatory experiences (66). However, this prior study used a substantially lower dose of ketamine than used in the current study, and nicotine was administered with chewing gum rather than the intravenous route employed here.

## **LIMITATIONS**

The main limitations of the current study include the small sample size, the high dropout rate, the heterogeneous smoking status of our sample, and the use of only one dose of nicotine. Future studies aimed at elucidating the effects of nicotine on patients with schizophrenia by using pharmacological models of psychosis in healthy volunteers must consider the fact that the large majority of schizophrenia patients are significantly dependent on nicotine. Accordingly, for studies about nAChR function to be relevant to schizophrenia, heavy smokers need to be included in the subject sample. However, the inclusion of nicotine-dependent heavy smokers in such studies raises the question of when to schedule the nicotine challenge relative to the timing of their last cigarette. Studying smokers who have been asked to abstain from smoking for several hours or more prior to study onset would mean studying them in a nicotine-withdrawal state. On the other hand, studying smokers who have smoked recently and are in a nicotine-satiated state may obscure the effects of intravenous nicotine. Further complicating this issue, studying non-smokers would result in high dropout routes because nicotine is generally unpleasant to non-smokers. Moreover, data from non-smokers may not generalize to schizophrenia patients, most of whom are heavy smokers.

In conclusion, the results of this study suggest that activation of nACH receptors does not influence ketamine's psychotomimetic effects or physiological effects on MMN and P300 in healthy human volunteers. However, ketamine and nicotine appear to have independent effects on P3a, P3b, and MMN suggesting differential effects of nACH and NMDA receptor systems on these ERP components. Moreover, this is the first study to report a significant reduction in P3a amplitude by nicotine.

## **ACKNOWLEDGMENTS**

The authors also thank Angelina Genovese, R.N.C., M.B.A.; Elizabeth O'Donnell, R.N.; Brenda Breault, R.N., B.S.N.; Sonah Yoo, R.Ph.; Robert Sturwold, R.Ph.; and Willie Ford of the Neurobiological Studies Unit at the VA Connecticut Healthcare System,West Haven Campus for their central contributions to the success of this project. The authors also acknowledge the contributions of John Krystal, MD for his role in providing the impetus in conducting the study. The Department of Veterans Affairs Schizophrenia Biological Research Center and a NARSAD Young Investigator Award to Hyun-Sang Cho supported this study.

## **REFERENCES**


in humans. Psychotomimetic, perceptual, cognitive, and neuroendocrine responses. *Arch Gen Psychiatry* (1994) **51**:199–214. doi:10.1001/archpsyc.1994. 03950030035004


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 25 October 2013; paper pending published: 19 November 2013; accepted: 07 January 2014; published online: 24 January 2014.*

*Citation: Mathalon DH, Ahn K-H, Perry EB Jr, Cho H-S, Roach BJ, Blais RK, Bhakta S, Ranganathan M, Ford JM and D'Souza DC (2014) Effects of nicotine on the neurophysiological and behavioral effects of ketamine in humans. Front. Psychiatry 5:3. doi: 10.3389/fpsyt.2014.00003*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2014 Mathalon, Ahn, Perry Jr, Cho, Roach, Blais, Bhakta, Ranganathan, Ford and D'Souza. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# The computational anatomy of psychosis

#### **Rick A. Adams <sup>1</sup>\*, Klaas Enno Stephan1,2,3, Harriet R. Brown<sup>1</sup> , Christopher D. Frith<sup>1</sup> and Karl J. Friston<sup>1</sup>**

<sup>1</sup> Wellcome Trust Centre for Neuroimaging, Institute of Neurology, University College London, London, UK.

<sup>2</sup> Translational Neuromodeling Unit, Institute for Biomedical Engineering, University of Zurich, ETH Zurich, Zurich, Switzerland

<sup>3</sup> Laboratory for Social and Neural Systems Research, University of Zurich, Zurich, Switzerland

#### **Edited by:**

Stefan Borgwardt, University of Basel, Switzerland

#### **Reviewed by:**

Andrea Mechelli, King's College London, UK Christian G. Huber, Universitäre Psychiatrische Kliniken Basel, Switzerland

#### **\*Correspondence:**

Rick A. Adams, Wellcome Trust Centre for Neuroimaging, 12 Queen Square, London WC1N 3BG, UK e-mail: rick.adams@ucl.ac.uk

This paper considers psychotic symptoms in terms of false inferences or beliefs. It is based on the notion that the brain is an inference machine that actively constructs hypotheses to explain or predict its sensations. This perspective provides a normative (Bayes-optimal) account of action and perception that emphasizes probabilistic representations; in particular, the confidence or precision of beliefs about the world. We will consider hallucinosis, abnormal eye movements, sensory attenuation deficits, catatonia, and delusions as various expressions of the same core pathology: namely, an aberrant encoding of precision. From a cognitive perspective, this represents a pernicious failure of metacognition (beliefs about beliefs) that can confound perceptual inference. In the embodied setting of active (Bayesian) inference, it can lead to behaviors that are paradoxically more accurate than Bayes-optimal behavior. Crucially, this normative account is accompanied by a neuronally plausible process theory based upon hierarchical predictive coding. In predictive coding, precision is thought to be encoded by the post-synaptic gain of neurons reporting prediction error.This suggests that both pervasive trait abnormalities and florid failures of inference in the psychotic state can be linked to factors controlling post-synaptic gain – such as NMDA receptor function and (dopaminergic) neuromodulation. We illustrate these points using biologically plausible simulations of perceptual synthesis, smooth pursuit eye movements and attribution of agency – that all use the same predictive coding scheme and pathology: namely, a reduction in the precision of prior beliefs, relative to sensory evidence.

**Keywords: free energy, active inference, precision, sensory attenuation, illusions, psychosis, schizophrenia**

## **INTRODUCTION**

This paper attempts to explain the positive and negative symptoms of schizophrenia in terms of false inference about states of the world producing sensations – and to link this explanation to neuromodulatory dysconnections at the synaptic level. In brief, we take a normative approach to action and perception – namely, active inference and the Bayesian brain hypothesis. We then consider neuronally plausible implementations of active inference to see how particular failures of neuromodulation would be expressed in terms of perceptual inference and behavior. The main conclusion is that a wide range of psychotic symptoms can be explained by a failure to represent the precision of beliefs about the world – and that this failure corresponds to abnormal neuromodulation of the post-synaptic gain of superficial pyramidal cells in cortical hierarchies. This may sound like a very specific assertion; however, there are many converging lines of evidence that point to this conclusion – lines that we try to draw together in this paper.

The basic idea is that faulty inference leads to false concepts (delusions) or percepts (hallucinations) and that this failure is due to a misallocation of precision to hierarchical representations in the brain. In what follows, we will refer to beliefs, inference, priors, and precision in a Bayesian sense. In this setting, a *belief* is a probability distribution over some unknown state or attribute. Beliefs, in this sense, may or may not be consciously accessible. A belief can be held with great *precision*, such that the probability distribution is

concentrated over the most likely value – the mean or *expectation*. This means the precision (inverse variance) corresponds to the confidence or certainty associated with a belief. In Bayesian inference, beliefs prior to observing data are called *prior beliefs*, which are updated to *posterior beliefs* after seeing the data. This updating rests upon combining a prior belief with sensory evidence or the *likelihood* of the data. In hierarchical Bayesian inference, the *sufficient statistics* of a belief (like the expectation and precision) are themselves treated as unknown quantities. This means that one can have beliefs about beliefs; for example, one can have an expectation about a precision (c.f., expected uncertainty). Heuristically, this leads to the distinction between fixed and random effects in classical statistics; or between risk (known uncertainty) and ambiguity (unknown uncertainty) in economics. Beliefs about beliefs are inevitable in hierarchical inference and are sometimes referred to as *empirical priors*, because they provide constraints on beliefs at lower levels of the hierarchy. Behaviorally, precision and beliefs about precision (including subjective confidence in beliefs) are to some extent dissociable (Fleming et al., 2012). Beliefs about precision are particularly important in hierarchical Bayesian inference, because they can have a profound effect on posterior expectations – and inappropriate beliefs about precision can easily lead to false inference.

The nature of this failure can be understood intuitively by considering classical statistical inference: imagine that we are using a *t*-test to compare the mean of some data, against the null hypothesis that the mean is zero. The sample mean provides evidence against the null hypothesis in the form of a *prediction error*: namely, the sample mean minus the expectation under the null hypothesis. The sample mean provides evidence against the null but how much evidence? This can only be quantified in relation to the precision of the prediction error. The *t*-statistic is simply the prediction error weighted by its precision (i.e., divided by its standard error). If this precision-weighted prediction error is sufficiently large, one rejects the null hypothesis. Clearly, if we overestimate the precision of the data, the *t*-statistic will be too large and we expose ourselves to false positives. Analogous rules apply to Bayesian inference, in that the optimal combination of a prior belief with some evidence is a posterior belief whose mean is a mixture of the prior and data means, weighted according to their precision. If the precision of the data is overestimated, or if the precision of the prior is underestimated, the posterior expectation will shift from the prior mean to the data mean (**Figure 1**).

So how could this lead to false beliefs and delusions? The following scenario (Frith and Friston, 2012) illustrates this: imagine the temperature warning light in your car is too sensitive (precise), reporting the slightest fluctuations (prediction errors) above some temperature. You naturally infer that there is something wrong with your car and take it to the garage. However, they find no fault – and yet the warning light continues to flash. Your first instinct may be to suspect the garage has failed to identify the fault – and even to start to question the Good Garage Guide that recommended it. From your point of view, these are all plausible hypotheses that accommodate the evidence available to you. However, from the perspective of somebody who has never seen your warning light, your suspicions would have an irrational and slightly paranoid flavor. This anecdote illustrates how delusional systems may be elaborated as a consequence of imbuing sensory evidence with too much precision. Note that the primary pathology here is quintessentially metacognitive in nature: in the sense that it rests on a belief (the warning light reports precise information) about a belief (the engine is overheating). Crucially, there is no necessary impairment in forming predictions or prediction errors – the problem lies in the way they are used to inform inference or hypotheses.

In what follows, we will consider the brain as performing inference using predictive coding, in which the evidence for hypotheses is reported by precision-weighted prediction errors. In these schemes, certain neurons compare bottom-up inputs with topdown predictions to form a prediction error that is weighted in proportion to its expected precision. Crucially, this weighting corresponds to the gain or sensitivity of prediction error units. This means that abnormalities in the modulation of postsynaptic gain could, in principle, lead to false inferences of the sort described above. We will illustrate this in a concrete fashion using biologically plausible simulations of false inference, all of which use exactly the same predictive coding scheme and intervention; namely, a decrease in the precision (post-synaptic gain of prediction error units) at higher levels of cortical hierarchies, relative to the precision at sensory levels. Some of these simulations have been reported previously in different contexts (Friston and Kiebel, 2009a; Adams et al., 2012; Brown et al., in press). Here, we

frame these simulations in terms of false inference and emphasize their common mechanisms. There are several other examples that we could have used; for example, the relationship between state-dependent precision and attention or the role of dopamine in encoding the precision of affordance and its effects on action selection. However, the examples chosen are sufficient to illustrate the diverse phenomenology that can be explained by one simple abnormality – a reduction in the precision of empirical prior beliefs, relative to sensory precision.

This paper focuses on false inference. However, the normative principles we appeal to cover both *inference* and *learning*. Neurobiologically, this corresponds to the distinction between updating neuronal representations in terms of synaptic activity and learning causal structure through updating synaptic efficacy (i.e., synaptic plasticity). The important thing here is that abnormal beliefs about precision also lead to *false learning*, which produces – and is produced by – false inference. This circular causality follows inevitably from the nature of inference, which induces posterior dependencies among estimates of hidden quantities in the world (encoded by synaptic activity and efficacy respectively). The point here is that a simple failure of neuromodulation (and implicit encoding of precision) can have far-reaching and knock-on effects that can be manifest at many different levels of perceptual inference, learning, and consequent behavior.

This paper comprises six sections. We start with a brief review of the symptoms and signs of schizophrenia, with a special focus on how *trait* and *state* abnormalities can be cast in terms of false inference. The second section reviews the psychopharmacology of psychosis with an emphasis on the synaptic (neuromodulatory) mechanisms that we suppose underlie false inference. The third establishes the normative theory (active inference) and its biological instantiation in the brain (generalized Bayesian filtering or predictive coding). The resulting scheme is used in the final three sections to illustrate failures of perceptual inference in the context of omission paradigms, abnormalities of active inference in the context of smooth pursuit eye movements and misattribution of agency in the context of deficits in sensory attenuation.

## **PSYCHOSIS AND FALSE INFERENCE**

In this section, we briefly review the state and trait abnormalities of schizophrenia to emphasize a common theme; namely, a failure of inference about the world that arises from an imbalance in the precision or confidence attributed to beliefs. We distinguish between state and trait abnormalities because the evidence suggests that trait abnormalities may be associated with a relative decrease in prior precision, while some state abnormalities can be explained by a (possibly compensatory) increase in prior precision (or reduction in sensory precision). In this setting, state abnormalities include the florid (Schneiderian or first rank) symptoms of acute psychosis, while trait abnormalities are more pervasive and subtle. The diagnostic criteria for schizophrenia are based largely on state abnormalities, because they are easily and reliably detected. These include:


before the first psychotic episode, and tend to covary with disorganization symptoms (reviewed in Silverstein and Keane, 2011). A decreased influence of context can sometimes lead to perceptions that are more veridical than those of normal subjects. Important examples here include a resistance to the hollow mask illusion – which is also state-dependent (Keane et al., in press) – and size-weight illusion (Williams et al., 2010).

These symptoms can occur episodically and – with the possible exception of catatonia-respond well to anti-dopaminergic drugs in the majority of patients. We use the term "trait" abnormalities to refer to more constant features of the disorder, which are less responsive to dopamine blockade (although these responses have not been explored as thoroughly as those of state symptoms). Some are found in first-degree relatives and high-risk groups, and may qualify as endophenotypes of schizophrenia. Despite their prevalence, they are less diagnostic because they are found in other diagnostic categories (and to some extent in the normal population). They include (among others):


Many trait abnormalities have been considered as the result of a failure to adequately predict sensory input, rendering all percepts surprising (e.g., the P50) and reducing differential responses to oddball stimuli (e.g., the MMN and P300). Predictive coding in particular has been used in recent formulations of these deficits in schizophrenia (Fletcher and Frith, 2009). Specifically, it is suggested that the main problem in schizophrenia lies not with the prediction of sensory input *per se*, but in the delicate balance of precision ascribed to prior beliefs and sensory evidence (Friston, 2005; Corlett et al., 2011). Later, we will use simulations to demonstrate how a relative increase in – or failure to attenuate – sensory precision can explain abnormal responses to surprising events.

In terms of cognitive paradigms, the "beads task" has been used to characterize formal beliefs and probabilistic reasoning in schizophrenic subjects. In this paradigm, subjects are told that red and green beads are drawn at random from an urn that contains (for example) 85% of one color and 15% of the other. The subject

must decide which color predominates. In reality, all subjects are shown the same sequence of beads. In the *draws to decision* version of the task, the subject has to answer as soon as they are certain. In the *probability estimates* version, the subject can continue to draw and change their answer. Interestingly, delusional patients "jump to conclusions" in the first version, while they are more willing to revise their decision in light of contradictory evidence in the second (Garety and Freeman, 1999). Bayesian modeling suggests that jumping to conclusions may reflect greater "cognitive noise" in delusional patients (Moutoussis et al., 2011), which may speak to reduced precision of higher level (cognitive) representations and consequently a greater influence of new sensory evidence (Speechley et al., 2010).

Can state abnormalities also be explained by imbalances in the precisions of prior beliefs and sensations? The short answer is yes. For example, delusional mood describes a state in which patients feel the world is strange and has changed in some way – where their attention is drawn to apparently irrelevant stimuli and odd coincidences. A loss of precise prior beliefs is consistent with a sense of unpredictability and greater attention to sensory events. Indeed, this line of thinking has been used to explain the loss of Gestalt or central coherence in autism (Pellicano and Burr, 2012). In terms of formal models, the top-down control of sensory precision has been shown to explain several psychophysical and physiological aspects of attention (Feldman and Friston, 2010); thereby providing a formal link between precision and attention. The key insight from these models is that posterior beliefs about states of the world can direct attention to sensory features by top-down modulation of sensory precision. A failure of top-down attenuation of sensory precision (sensory attenuation) therefore fits comfortably with abnormalities of sensory attention in this context.

State abnormalities include the cardinal psychotic symptoms, such as hallucinations and delusions. Hallucinations could be understood as the result of an *increase* in the relative precision of prior beliefs, such that the posterior beliefs are impervious to contradictory – but imprecise – sensory evidence. This has been discussed as an explanation for visual hallucinosis in organic psychosyndromes (Friston, 2005). However, the hallucinations associated with psychosis may be better understood as a failure to attenuate the sensory consequences (corollary discharge) of self-made acts; for example, a failure to attenuate the auditory consequences of sub-vocal or inner speech (Frith et al., 1998; Allen et al., 2007). Delusions are probably more complex and their emergence may be better understood as secondary phenomena: several authors have proposed that they could arise as rational (Bayes-optimal) posterior beliefs that explain away precise sensory prediction errors: e.g.,Fletcher and Frith (2009). These explanations relate to earlier "empiricist" accounts such as Maher (1974), Gray et al. (1991), and Kapur (2003), who emphasizes aberrant salience (c.f., sensory precision). Implicit in these secondary accounts is a compensatory increase in the precision of explanations for sensory cues that are imbued with too much precision or salience. This is consistent with their peculiar resistance to rational argument. In the final section, we will consider an example of a compensatory increase in the precision of high-level beliefs that is necessary to compensate for a failure of sensory attenuation.

## **SUMMARY**

In summary, the symptoms and signs of schizophrenia are not inconsistent with a reduction of high-level precision or a failure of sensory attenuation (the top-down attenuation of sensory precision), with compensatory (secondary) changes in the precision of (empirical) prior beliefs. In particular, some psychotic states may reflect a compensatory response to trait abnormalities that bias inference toward sensory evidence that is imbued with too much precision or salience. A further mechanistic dissociation between state and trait abnormalities is suggested by the fact that the former generally respond to antipsychotic (antidopaminergic) treatment, while trait abnormalities do not. Before considering the computational anatomy of hierarchical inference in the brain, we will briefly review the psychopharmacology and neuropathology of schizophrenia.

## **THE PSYCHOPHARMACOLOGY OF PRECISION**

This section considers the neuromodulatory processes implicated in schizophrenia, with a special focus on the laminar specificity of cortical neuromodulation. Our premise here is that psychotic abnormalities are manifestations of false inference, caused by the aberrant encoding of precision. This precision is thought to be encoded by post-synaptic gain of neuronal populations reporting prediction errors – the principal or pyramidal cells of superficial cortical layers (Mumford, 1992; Feldman and Friston, 2010). Synaptic gain modulation is a change in the response amplitude of a neuron that is independent of its selectivity or receptive field characteristics (Salinas and Thier, 2000). In other words, post-synaptic gain is a factor that quantifies the effect of a presynaptic input on post-synaptic output (e.g., depolarization at the soma). Changes in synaptic gain are generally thought to be mediated by non-linear (e.g., multiplicative) synaptic mechanisms; for example, NMDA receptor activation.

Of all the receptors that determine synaptic gain, the most ubiquitous is the glutamatergic NMDA receptor (NMDA-R). NMDA-Rs have several important functions that are expressed over different timescales. First, they can drive (i.e., induce an excitatory post-synaptic potential) post-synaptic cells like other ionotropic glutamatergic (AMPA and Kainate) receptors. However, the driving effect of NMDA-Rs is only possible if the cell is already depolarized; otherwise, the NMDA-R is blocked by a magnesium ion. This non-linear property makes them synaptic coincidence detectors or "AND gates." Second, NMDA-Rs have time constants that are much longer than that of AMPA-Rs and Kainate-Rs. This enables integration of synaptic inputs over tens to hundreds of milliseconds – increasing the gain of synaptic inputs to distal dendrites. Finally, NMDA-Rs are famous for their role in plasticity: at longer timescales, the influx of calcium ions through NMDA-R channels causes a cascade of intracellular events that result in longterm synaptic depression or potentiation (LTD or LTP). However, NMDA-Rs also have a major impact on the short-term plasticity of glutamatergic synapses. This is because they regulate the functional state and number of AMPA-Rs – by phosphorylation or by changing the trafficking of AMPA-R subunits to and from the cell membrane (Passafaro et al., 2001; Montgomery and Madison, 2004; Bagal et al., 2005). Together, these properties make a significant contribution to the dynamics of neural networks, especially

to oscillatory behavior and sustained firing patterns (Durstewitz, 2009).

Other key determinants of synaptic gain are the classical neuromodulator receptors; e.g., dopamine (DA-Rs), acetylcholine (in particular muscarinic AChRs), and serotonin (5-HTRs). With the exception of nicotinic AChRs (which are ionotropic) these are all metabotropic receptors – they do not activate ion channels but are coupled to signal transduction mechanisms (via G proteins) that affect intracellular second messengers, such as cyclic adenosine monophosphate (cAMP) or cyclic guanosine monophosphate (cGMP). Fluctuations in cAMP/cGMP concentration affect the activity of protein kinases, which – through phosphorylation – alters neuronal excitability via changes in the production, surface expression or activity of voltage or ligand-gated ion channels, including the NMDA-R itself. This mechanism is also used by another glutamatergic receptor – with pronounced modulatory effects on synaptic gain – the metabotropic glutamate receptor (mGluR; De Pasquale and Sherman, 2012). It is important to note that DA-R subtypes have opposite effects on synaptic gain: D1R activation stimulates cAMP production and increases the excitability of depolarized neurons, whereas D2R activation inhibits cAMP production and reduces gain (reviewed in Frank, 2005).

Synaptic gain is not just determined by receptor activity but also by network dynamics, like the synchronization of fast oscillations, especially in the 40–100 Hz or gamma frequencies (c.f., synchronous gain: Chawla et al., 1999). The fast acting inhibitory γ-amino butyric acid receptor (GABAA-R) is instrumental in this synchronization process. In the cortex, a GABAergic (parvalbuminpositive basket cell or PVBC) interneuron contacts many pyramidal cells, which it transiently hyperpolarizes. When this hyperpolarization wears off, all the cortical pyramidal cells can then fire together, leading to synchronous firing across the network and oscillations as the cycle recurs (Gonzalez-Burgos and Lewis, 2012).

Abnormalities in at least three of these synaptic gain mechanisms have been proposed to be a primary pathology in schizophrenia – those of NMDA, GABA, and dopamine receptors. NMDA-Rs play a central role in theories of schizophrenia (Olney and Farber, 1995; Abi-Saab et al., 1998; Goff and Coyle, 2001; Stephan et al., 2006; Corlett et al., 2011). Studies of genetic risk in schizophrenia have highlighted the role of genes related to glutamatergic transmission, with GABA and dopamine related genes implicated to a lesser extent (Harrison and Weinberger, 2005; Stephan et al., 2006; Hall et al., 2009; Greenwood et al., 2012). Neuropathological evidence indicates abnormalities of the glutamate and GABA systems: both pre- and post-synaptic markers, morphometric, and biochemical measures of glutamatergic transmission are reduced, as is the expression of the GABA synthesizing enzyme glutamic acid decarboxylase (GAD), parvalbumin-immunoreactive GABAergic interneurons and their synaptic markers (reviewed in Harrison et al., 2011). These neuropathological changes are particularly apparent in hippocampus and frontal cortex, both at high levels in the cortical hierarchy (Felleman and Van Essen, 1991).

Conversely, the evidence for dopaminergic abnormalities in schizophrenia is neither neuropathological nor structural, but functional. The most widely replicated abnormality is that of elevated striatal dopamine availability – in acute psychoses of both schizophrenia (Laruelle et al., 1996; Breier et al., 1997) and epilepsy (Reith et al., 1994). A recent review concluded that dopamine dysregulation is more closely linked to the state of psychosis than the trait of schizophrenia (Howes and Kapur, 2009), although there are some important caveats: presynaptic dopamine is also raised to a lesser degree in those who are prone to psychosis but not floridly psychotic, and patients with symptoms resistant to dopamine blockade do not have elevated striatal dopamine synthesis (Demjaha et al., 2012).

Is aberrant glutamatergic and GABAergic transmission linked to the trait abnormalities of the previous section? The psychotomimetic effects of ketamine suggest a strong association. Ketamine blocks NMDA-Rs and also potentiates AMPA-R signaling, leading to decreased burst firing of pyramidal neurons, with subsequent impairment of activation of GABAergic interneurons (Shi and Zhang, 2003). Ketamine administration can reproduce a whole spectrum of trait phenomena: such as SPEM abnormalities (Radant et al., 1998; Weiler et al., 2000); impaired P50 suppression (Oranje et al., 2002); diminished P300 (Gunduz-Bruce et al., 2012); reduced MMN (Umbricht et al., 2000; Schmidt et al., 2012); cognitive impairments (Kantrowitz and Javitt, 2010); and negative symptoms (Krystal et al., 1994). In fact, the only trait phenomenon that ketamine does not reproduce is a reduced susceptibility to the hollow mask illusion (Passie et al., 2003). This is in contrast to dopaminergic agonists, which do not reproduce perceptual, SPEM (Reilly et al., 2008), P50 (Oranje et al., 2004) or MMN (Leung et al., 2007) abnormalities – and have only small effects on the P300 (Luthringer et al., 1999). Indeed, prefrontal D1R hypoactivity has been associated with cognitive deficits and negative symptoms in animal models (Goldman-Rakic et al., 2004).

Ketamine's reproduction of state symptoms is less consistent: its effects include loss of perceptual organization (Uhlhaas et al., 2007) and induction of a delusional mood (Corlett et al., 2011), but it does not cause a loss of attenuation of self-induced sensations (PC Fletcher, personal communication) or lead to auditory verbal hallucinations. It is interesting to note that while the negative symptoms induced by ketamine are correlated with its NMDA-R binding, the positive symptoms are not (Stone et al., 2008). Conversely, D2R levels in cortical and striatal areas correlate with positive but not negative symptom scores (Kessler et al., 2009). Nevertheless, some trait-like phenomena can be reproduced by both ketamine and dopaminergic agonists, such as reduced latent inhibition (Young et al., 2005; Razoux et al., 2007), blocking (O'Tuathaigh et al., 2003; Freeman et al., 2013), and the body ownership illusion (Albrecht et al., 2011; Morgan et al., 2011). This is not surprising, as there are complex interactions between glutamatergic, GABAergic, and dopaminergic neurotransmission, within and between the brain stem, striatum, and prefrontal cortex (see **Figure 2**). For example, hypofunction of NMDA-Rs in cortical projections to the ventral tegmental area (which are themselves regulated by D<sup>2</sup> autoreceptors, nACh-Rs, and 5- HT-Rs) are in a position to reduce the activity of mesofrontal D1R-projecting dopaminergic neurons (that potentiate prefrontal NMDA-Rs) and increase activity (via decreased GABAergic inhibition) of mesostriatal D2R-projecting neurons (Stephan et al., 2009). NMDA-Rs and D1Rs within the same cell potentiate each

other in numerous ways (Cepeda and Levine, 2006). In the prefrontal cortex,NMDA-R impairments may lead to hypofunction of GABAergic PVBC's, disinhibition of pyramidal cells, and reduced prefrontal gamma activity (Gonzalez-Burgos and Lewis, 2012). Alternatively, NMDA-R hypofunction could impact directly on the excitability of prefrontal pyramidal cells.

The neuropathology of schizophrenia is usually associated with higher cortical systems; e.g., prefrontal cortex and the medial temporal lobe. For example, perceptual deficits in schizophrenics (and normal subjects) have been shown to correlate with frontal and temporal volume loss (Dazzan et al., 2006). The hierarchical level of a cortical area is defined in terms of extrinsic (ascending and descending) connections that have a laminar specificity: ascending (extrinsic) projections target the granular layer 4, which

sends forward (intrinsic) connections to (supragranular) layers 2 and 3. These then either send further forward (extrinsic) projections up to the next hierarchical level, or pass signals down via (infragranular) layers 5 and 6 to the level below. See Bastos et al. (2012) for a review of this canonical circuitry from the point of view of predictive coding. In prefrontal cortex – as in the rest of the cerebrum – NMDA-Rs are distributed throughout the cortical layers but are most concentrated in superficial layers 2 and 3 (Jansen et al., 1989), as are D1Rs (Lidow et al., 1991). By contrast, D2Rs are much less prevalent than D1Rs in the cortex (by an order of magnitude) and their peak concentration is in layer 5 (Lidow et al., 1991). Nevertheless, Opris et al. (2012) have recently shown in primates that cocaine (which increases dopaminergic transmission) reduces the activity of superficial pyramidal cells (perhaps via D2Rs) and thereby their synchronization with layer 5 pyramidal cells in the same minicolumn – impairing performance in a working memory task.

Many of the neuropathological changes in schizophrenia are found in supragranular layers 2 and 3, with additional abnormalities in layer 5: see Harrison et al. (2011) for a fuller treatment of this complex and sometimes inconsistent literature. In brief, the somal volume of layer 3 DLPFC pyramidal cells has been found to be reduced (Rajkowska et al., 1998; Pierri et al., 2001), and these neurons have smaller basal dendrites (Glantz and Lewis, 2000; Kalus et al., 2000) and lower dendritic spine density (Kolluri et al., 2005). These changes may be caused by the neurotrophic effects of reduced NMDA-R inputs (Rajan and Cline,1998;Monfils and Teskey, 2004) and a loss of synaptic connectivity (Perrone-Bizzozero et al., 1996; Glantz and Lewis, 1997) – perhaps with the thalamus (Lewis et al., 2001) or association cortex (Sweet et al., 2007). Others have found losses of interneurons in layer 2 in both prefrontal and cingulate cortex (Benes et al., 1991) and smaller dendritic fields of prefrontal layer 5 pyramidal cells (Black et al., 2004). In the medial temporal lobe, most abnormalities are again found in the superficial layers; such as atypical clustering of neurons in layer 2 of entorhinal cortex (Jakob and Beckmann, 1986; Arnold et al., 1991; Falkai et al., 2000).

## **SUMMARY**

In summary, the main neuropathological abnormalities appear to be expressed in high hierarchical levels (prefrontal cortex and the medial temporal lobe), particularly in supragranular layers that contain superficial pyramidal cells. The main neuromodulatory (trait) abnormalities include the hypofunction of cortical NMDA-Rs and GABAergic neurons (and possibly D1Rs) – in contrast to the elevation of striatal D2R activity in (the state of) psychosis. In short, the neuropharmacological and neuropathological evidence points to abnormal neuromodulation of superficial pyramidal cells. This is important because – in predictive coding schemes – the post-synaptic gain of these cells encodes the precision of prediction error. The next section explains why this is the case, starting from basic principles.

## **NEUROBIOLOGICAL IMPLEMENTATION OF ACTIVE INFERENCE**

This section introduces the theory behind inference in the brain. This normative account provides key constraints on the functional (computational) anatomy of action and perception. This allows one to understand (and simulate) inference in a principled way – that is also grounded in neuroanatomy and neurophysiology. We will use the formalism below to simulate some of the schizophrenic abnormalities reviewed above. These simulations rest on descriptions of the neuronal processes (differential equations) that underwrite inference in the brain. These equations are based on three assumptions:


• Neuronal firing rates encode the expected state of the world, under this model.

The first assumption is the free energy principle, which leads to active inference in the embodied setting of action (Friston et al., 2010a). This provides a normative (Bayes-optimal) account of action and perception, in which both minimize a free energy bound on the (negative log) evidence for the brain's model of the world. Free energy is a quantity from statistics that measures the quality of a model in terms of the probability that it could have generated observed outcomes. This means that minimizing free energy maximizes the Bayesian evidence for the generative model (Ballard et al., 1983; Hinton and van Camp, 1993; Dayan et al., 1995). The second assumption is motivated by noting that the world is both dynamic and non-linear and that hierarchical causal structure emerges inevitably from a separation of temporal scales (Ginzburg, 1955; Haken, 1983). The final assumption is the Laplace assumption that, in terms of neural codes, leads to the *Laplace code* that is arguably the simplest and most flexible of all neural codes (Friston, 2009).

Given these assumptions, one can simulate a whole variety of neuronal processes by specifying the particular equations that constitute the brain's generative model. Action and perception are then specified completely by the above assumptions and can be implemented in a biologically plausible fashion. In brief, these simulations use differential equations that minimize the free energy of sensory input using a generalized gradient descent (Friston et al., 2010b).

$$\begin{aligned} \dot{\tilde{\mu}}\left(t\right) &= \mathcal{D}\tilde{\mu}\left(t\right) - \partial\_{\tilde{\mu}}F\left(\tilde{s}, \tilde{\mu}\right) \\ \dot{a}\left(t\right) &= -\partial\_{a}F\left(\tilde{s}, \tilde{\mu}\right) \end{aligned} \tag{1}$$

These coupled differential equations describe perception and action respectively. They say that neuronal activity encoding posterior expectations about (generalized) hidden states of the world µ˜ = µ, µ 0 , µ 00 , . . . and action *a* reduce free energy – where free energy *F* (*s*˜,µ˜) is a function of (generalized) sensory inputs *s*˜ = *s*, *s* 0 , *s* 00 , . . . and neuronal activity. The first differential equation is known as generalized predictive coding or Bayesian filtering: see also Rao and Ballard (1999). The first term is a prediction based upon a differential matrix operator D that returns the generalized motion of expected hidden states. The second (correction) term is usually expressed as a mixture of prediction errors that ensures the changes in posterior expectations are Bayesoptimal predictions about hidden states of the world. The second differential equation says that action also minimizes free energy. The differential equations above are coupled because sensory input depends upon action, which depends upon perception through the posterior expectations. This circular dependency leads to a sampling of sensory input that is both predicted and predictable, thereby minimizing free energy and, implicitly, prediction errors.

To perform neuronal simulations under this scheme, it is only necessary to integrate or solve Eq. 1 to simulate the neuronal dynamics that encode posterior expectations and associated action. Posterior expectations depend upon the brain's generative model of the world, which we assume has the following hierarchical form:

$$\begin{aligned} s &= \mathcal{g}^{(1)}\left(\mathbf{x}^{(1)}, \boldsymbol{\nu}^{(1)}\right) + \boldsymbol{\alpha}\_{\boldsymbol{\nu}}^{(1)} \\ \dot{\mathbf{x}}^{(1)} &= f^{(1)}\left(\mathbf{x}^{(1)}, \boldsymbol{\nu}^{(1)}\right) + \boldsymbol{\alpha}\_{\boldsymbol{\nu}}^{(1)} \\ &\vdots \\ \boldsymbol{\nu}^{(i-1)} &= \mathcal{g}^{(i)}\left(\mathbf{x}^{(i)}, \boldsymbol{\nu}^{(i)}\right) + \boldsymbol{\alpha}\_{\boldsymbol{\nu}}^{(i)} \\ \dot{\mathbf{x}}^{(i)} &= f^{(i)}\left(\mathbf{x}^{(i)}, \boldsymbol{\nu}^{(i)}\right) + \boldsymbol{\alpha}\_{\boldsymbol{\nu}}^{(i)} \\ &\vdots \\ \boldsymbol{\alpha}\_{\boldsymbol{\nu}}^{(i)} &\sim N\left(0, \Pi\_{\mathbf{x}}^{(i)-1}\right) \\ \boldsymbol{\alpha}\_{\boldsymbol{\nu}}^{(i)} &\sim N\left(0, \Pi\_{\mathbf{x}}^{(i)-1}\right) \\ \Pi\_{\mathbf{x}}^{(i)} &= \exp\left(\pi\_{\mathbf{x}}^{(i)}\left(\mathbf{x}^{(i)}, \boldsymbol{\nu}^{(i)}\right)\right) \\ \Pi\_{\boldsymbol{\nu}}^{(i)} &= \exp\left(\pi\_{\mathbf{x}}^{(i)}\left(\mathbf{x}^{(i)}, \boldsymbol{\nu}^{(i)}\right)\right) \end{aligned} \tag{2}$$

This equation describes a probability density over the sensory and hidden states that generate sensory input. Here, the hidden states have been divided into hidden states and causes (*x* (*i*) , *v* (*i*) ), with (*i*) denoting their level within the hierarchical model. Hidden states and causes are abstract variables that the brain uses to explain or predict sensations – like the motion of an object in the field of view. In these models, hidden causes link hierarchical levels, whereas hidden states link dynamics over time. Here (*g* (*i*) , *f* (*i*) ) are non-linear functions of hidden states and causes that generate hidden causes for the level below and – at the lowest level – sensory inputs. Random fluctuations in the motion of hidden states and causes ω (*i*) *<sup>x</sup>* , ω (*i*) *v* enter each level of the hierarchy. Gaussian assumptions about these random fluctuations make the model probabilistic. They play the role of sensory noise at the first level and induce uncertainty at higher levels. The amplitudes of these random fluctuations are quantified by their precisions Π (*i*) *<sup>x</sup>* , Π (*i*) *v* that may depend upon the hidden states or causes through their log precisions π (*i*) *<sup>x</sup>* , π (*i*) *v* 

#### **PERCEPTION AND PREDICTIVE CODING**

Given the form of the generative model Eq. 2 we can now write down the differential Eq. 1 describing neuronal dynamics in terms of (precision weighted) prediction errors on the hidden causes and states. These errors represent the difference between posterior expectations and predicted values, under the generative model (using *A* × *B*: =*A <sup>T</sup>B* and omitting higher-order terms):

$$\begin{split} \dot{\tilde{\mu}}\_{\mathbf{x}}^{(i)} &= \mathcal{D}\tilde{\mu}\_{\mathbf{x}}^{(i)} + \left( \frac{\partial \tilde{\mathcal{g}}^{(i)}}{\partial \tilde{\mu}\_{\mathbf{x}}^{(i)}} - \frac{1}{2} \tilde{\varepsilon}\_{\mathbf{v}}^{(i)} \frac{\partial \tilde{\pi}\_{\mathbf{v}}^{(i)}}{\partial \tilde{\mu}\_{\mathbf{x}}^{(i)}} \right) \cdot \dot{\xi}\_{\mathbf{v}}^{(i)} \\ &+ \left( \frac{\partial \tilde{f}^{(i)}}{\partial \tilde{\mu}\_{\mathbf{x}}^{(i)}} - \frac{1}{2} \tilde{\varepsilon}\_{\mathbf{x}}^{(i)} \frac{\partial \tilde{\pi}\_{\mathbf{x}}^{(i)}}{\partial \tilde{\mu}\_{\mathbf{x}}^{(i)}} \right) \cdot \dot{\xi}\_{\mathbf{x}}^{(i)} \\ &+ \frac{\partial \operatorname{tr}\left( \tilde{\pi}\_{\mathbf{v}}^{(i)} + \tilde{\pi}\_{\mathbf{x}}^{(i)} \right)}{\partial \tilde{\mu}\_{\mathbf{x}}^{(i)}} - \mathcal{D}^{T} \xi\_{\mathbf{x}}^{(i)} \end{split}$$

$$\dot{\tilde{\mu}}\_{\nu}^{(i)} = \mathcal{D}\tilde{\mu}\_{\nu}^{(i)} + \left(\frac{\partial \tilde{\mathcal{g}}\_{\nu}^{(i)}}{\partial \tilde{\mu}\_{\nu}^{(i)}} - \frac{1}{2} \tilde{\varepsilon}\_{\nu}^{(i)} \frac{\partial \tilde{\pi}\_{\nu}^{(i)}}{\partial \tilde{\mu}\_{\nu}^{(i)}}\right) \cdot \dot{\xi}\_{\nu}^{(i)} \tag{3}$$

$$+ \left(\frac{\partial \tilde{\ell}^{(i)}}{\partial \tilde{\mu}\_{x}^{(i)}} - \frac{1}{2} \tilde{\varepsilon}\_{x}^{(i)} \frac{\partial \tilde{\pi}\_{x}^{(i)}}{\partial \tilde{\mu}\_{\nu}^{(i)}}\right) \cdot \dot{\xi}\_{x}^{(i)}$$

$$+ \frac{\partial \text{tr}\left(\tilde{\pi}\_{\nu}^{(i)} + \tilde{\pi}\_{x}^{(i)}\right)}{\partial \tilde{\mu}\_{\nu}^{(i)}} - \dot{\xi}\_{\nu}^{(i+1)}$$

$$\begin{aligned} \boldsymbol{\xi}\_{\boldsymbol{x}}^{(i)} &= \boldsymbol{\Pi}\_{\boldsymbol{x}}^{(i)} \widetilde{\boldsymbol{e}}\_{\boldsymbol{x}}^{(i)} = \boldsymbol{\Pi}\_{\boldsymbol{x}}^{(i)} \left( \mathcal{D} \widetilde{\boldsymbol{\mu}}\_{\boldsymbol{x}}^{(i)} - \widetilde{\boldsymbol{f}}^{(i)} \left( \widetilde{\boldsymbol{\mu}}\_{\boldsymbol{x}}^{(i)}, \widetilde{\boldsymbol{\mu}}\_{\boldsymbol{v}}^{(i)} \right) \right) \\ \boldsymbol{\xi}\_{\boldsymbol{v}}^{(i)} &= \boldsymbol{\Pi}\_{\boldsymbol{v}}^{(i)} \widetilde{\boldsymbol{e}}\_{\boldsymbol{v}}^{(i)} = \boldsymbol{\Pi}\_{\boldsymbol{v}}^{(i)} \left( \widetilde{\boldsymbol{\mu}}\_{\boldsymbol{v}}^{(i-1)} - \widetilde{\boldsymbol{g}}^{(i)} \left( \widetilde{\boldsymbol{\mu}}\_{\boldsymbol{x}}^{(i)}, \widetilde{\boldsymbol{\mu}}\_{\boldsymbol{v}}^{(i)} \right) \right) \end{aligned}$$

Equation 3 can be derived by computing the free energy for the hierarchical model in Eq. 2 and inserting its gradients into Eq. 1. This produces a relatively simple update scheme, in which posterior expectations are driven by a mixture of prediction errors, where prediction errors are defined by the equations of the generative model.

It is difficult to overstate the generality of Eq. 3: its solutions grandfather nearly every known statistical estimation scheme, under parametric assumptions about additive or multiplicative noise (Friston, 2008). These range from ordinary least squares to advanced variational deconvolution schemes. The scheme is called *generalized Bayesian filtering* or *predictive coding* (Friston et al., 2010b). In neural network terms, Eq. 3 says that error units ξ (*i*) *v* compute the difference between expectations at one level µ˜ (*i*−1) *v* and predictions from the level above *g*˜ (*i*) µ˜ (*i*) *<sup>x</sup>* , µ˜ (*i*) *v* . Conversely, posterior expectations (encoded by the activity of state units) are driven by prediction errors from the same level and the level below. These constitute bottom-up and lateral messages that drive posterior expectations toward a better prediction to reduce the prediction error in the level below. This is the essence of recurrent message passing between hierarchical levels to optimize free energy or suppress prediction error: see Friston and Kiebel (2009b) and Feldman and Friston (2010)for a more detailed discussion. Crucially, in neurobiological implementations of this scheme, the sources of bottom-up prediction errors have to be superficial pyramidal cells, because it is these – and only these – cells that send forward (ascending) connections to higher cortical areas. Conversely, predictions are conveyed from deep pyramidal cells, by backward (descending) connections, to target the superficial pyramidal cells encoding prediction error (Mumford, 1992; Bastos et al., 2012): see **Figure 3**.

Note that the precisions depend on the expected hidden causes and states. We have proposed that this dependency mediates attention and action selection in hierarchical processing (Feldman and Friston, 2010; Friston et al., 2012). Equation 3 tells us that the (state-dependent) precisions Π (*i*) *<sup>x</sup>* , Π (*i*) *v* modulate the responses of prediction error units to their presynaptic inputs. This modulation depends on the posterior expectations about the states and suggests something intuitive – attention is mediated by activitydependent modulation of the synaptic gain of principal cells that

convey sensory information (prediction error) from one cortical level to the next. This translates into a top-down control of synaptic gain in principal (superficial pyramidal) cells elaborating prediction errors and fits comfortably with the modulatory effects of top-down connections in cortical hierarchies that have been associated with attention and action selection.

## **ACTION**

In active inference, posterior expectations elicit behavior by sending top-down predictions down the hierarchy that are unpacked into proprioceptive predictions at the level of the cranial nerve nuclei and spinal cord. These engage classical reflex arcs to suppress proprioceptive prediction errors and produce the predicted motor trajectory

$$\dot{a} = -\frac{\partial}{\partial a} F = -\frac{\partial \tilde{\mathfrak{s}}}{\partial a} \times \xi\_{\nu}^{(1)} \tag{4}$$

The reduction of action to classical reflexes follows because the only way that action can minimize free energy is to change sensory (proprioceptive) prediction errors by changing sensory signals; cf., the equilibrium point formulation of motor control (Feldman and Levin, 1995). In short, active inference can be regarded as equipping a generalized predictive coding scheme with classical reflex arcs: see Friston et al. (2009) and Adams et al. (2013) for details. The actual movements produced clearly depend upon top-down predictions that can have a deep and complex structure, as we will see later.

## **SUMMARY**

In summary, starting with the assumption that the brain is trying to maximize the evidence for its model of the world, one can derive plausible equations describing neuronal dynamics in terms of message passing among different levels of a (cortical) hierarchical model. These messages comprise precision-weighted prediction errors that are passed forward from one level to the next and top-down predictions that are reciprocated to minimize prediction error. In this scheme, precision is encoded by the gain of superficial pyramidal cells reporting prediction error, which is implicated in the synaptic pathology of schizophrenia. This is a straightforward consequence of the mathematical form of predictive coding and the fact that superficial pyramidal cells are the source of ascending connections in the brain. At the proprioceptive level, prediction errors can be reduced either by changing predictions (perception) or by changing sensations (action). In the last three sections, we use Eqs 3 and 4 to simulate active inference under a number of generative models, while manipulating the precision at different hierarchical levels. These models are described completely by the Eq. 2, which are provided in figures that summarize the generative model used in each example.

## **PERCEPTUAL INFERENCE AND HALLUCINATIONS**

This section focuses on perceptual inference to show how reducing the precision at high levels of a generative model can confound perception and distort perceptual synthesis. We will examine a non-trivial problem; namely, recognizing structure and syntax in communication, using a well studied model – birdsong. This is an interesting problem because it calls upon both the dynamics modeled by hidden states and a hierarchical structure that entails a separation of temporal scales (Kiebel et al., 2009). We first describe our generative model of birdsong and then examine the sorts of inference that arise when prior precision is reduced. We then model a compensatory reduction in sensory precision. In brief, we will see a loss of responses to violations – of the sort that characterize psychotic traits (e.g., reduced MMN) – and the emergence of hallucinosis with compensatory changes in sensory precision.

## **ATTRACTORS IN THE BRAIN**

The basic idea behind the generative model in this section is that the environment unfolds as an ordered sequence of dynamics, whose equations of motion have an attractor manifold that contains sensory trajectories. Crucially, the shape of this manifold is itself changed by other dynamical systems that have their own attracting sets. If the brain has a generative model of these hierarchically coupled dynamics, then we would expect to see cascades of neuronal attractors (c.f., central pattern generators) that are trying to predict sensory input. In this hierarchical setting, one would expect higher attractors to predict the changing shape of lower attractors, thereby modeling a separation of temporal scales

of the sort seen in language (e.g., from formants to phonemes, from phonemes to words, from words to phrases, from phrases to sentences, and so on).

The example used here deals with the generation and recognition of birdsongs (Laje and Mindlin, 2002). We imagine that birdsongs are produced by two time-varying control parameters that control the frequency and amplitude of vibrations of a songbird's syrinx (see **Figure 4**). There has been an extensive effort using attractor models at the biomechanical level to understand the generation of birdsong; e.g., Laje et al. (2002). Here, we use attractors at higher levels to provide time-varying control over the resulting sonograms. To produce synthetic stimuli, we drove the syrinx with two states of a Lorenz attractor, one controlling the frequency (between 2 and 5 kHz) and the other controlling the amplitude or volume. The parameters of the Lorenz attractor were chosen to generate a short sequence of chirps every second or so. To endow the generative model with a hierarchical structure, we placed a second Lorenz attractor – whose dynamics were an order of magnitude slower – over the first. The states of the slower attractor entered as control parameters (the Rayleigh and Prandtl number) to control the shape of the lower attractor.

We generated a single song, corresponding roughly to a cycle of the higher attractor and then filtered the ensuing sonogram (summarized as peak amplitude and volume) using the message-passing scheme described in the previous section Eq. 3. The results are shown in **Figure 4** (lower panels), in terms of the predicted sonogram and prediction error at the sensory level. These results show that – after several hundred milliseconds – the veridical hidden states and causes can be recovered and provide accurate predictions of auditory sensations. Note that the percept or predictions are not an exact copy of the stimulus – the mismatch is reflected in the prediction errors on the lower right. These prediction errors provide contextual guidance for posterior expectations about hidden causes and states. Note that prediction errors coincide with the onset of each chirp, where the prediction errors for the third chirp are more protracted – suggesting that this chirp was less easy to predict than the others.

## **OMISSION-RELATED RESPONSES**

To examine responses to surprising stimuli or violations – and how they depend upon precision – we repeated the simulation but omitted the last three chirps. The corresponding percepts are shown with their prediction errors in **Figure 5** (top row). These results illustrate two important phenomena. First, there is a vigorous expression of prediction error with the first missing chirp. This reflects the dynamical nature of perception: at this point, there is no sensory input to predict and the prediction error is generated entirely by top-down predictions. Second, it can be seen that there is a transient (illusory) percept, when the missing chirp should have occurred. Its frequency is too low, but its timing is preserved in relation to the expected chirp. This is an interesting stimulation from the point of view of ERP studies of omission-related responses that provide clear evidence for the predictive capacity of the brain (e.g., Nordby et al., 1994; Yabe et al., 1997).

This simulation models neuronal responses to unpredicted or surprising stimuli of the sort used in oddball paradigms to elicit the MMN or P300. These electrophysiological markers are particularly

expectations (right).

pertinent here, because the same cells reporting prediction error (superficial pyramidal cells) are thought to be the primary source of electrophysiological measurements. In these simulations, the sensory log precision was two, the log precision of (first level) hidden states was eight and the log precision of second level prediction errors was high (16). These precisions correspond to the true uncertainty or amplitude of random fluctuations used to generate the song. So what would happen if we reduced the precision of prediction errors at the second level that provides top-down predictions about the syntax and timing of the chirps?

that have to be inferred, given the stimulus. This stimulus is represented as a

## **PRECISION AND ODDBALL RESPONSES**

The middle row of **Figure 5** shows the results of repeating the simulation when the log precision at the second level was reduced to two. This has two remarkable effects: first, there is a failure to detect the third chirp (that previously elicited the greatest prediction error – white arrow) and, second, there is a marked attenuation of the omission–related response. The explanation for these phenomena is straightforward: because we have reduced the precision at higher levels, there is less confidence in top-down predictions and therefore every stimulus is relatively surprising. In fact, the third stimulus is so unpredictable it is not perceived, eliciting a large prediction error (black arrow in the middle right panel). Similarly, a high amplitude prediction error is seen shortly afterward in response to the surprising omission. However, it is attenuated in comparison to responses under precise top-down predictions. This allows sensory evidence to resolve prediction errors more quickly, thereby reducing their amplitude. This may speak to the attenuation of oddball responses as a psychotic trait. In particular, the attenuation of the MMN can be seen in terms of the difference between the prediction errors to the omitted chirp, relative to the third (standard) chirp (red arrows). These simulations highlight

based upon posterior expectations, while the right-hand panels show the associated (precision weighted) prediction error at the sensory level. The top panels show a normal omission-related response using log precisions of 16 at the second (higher) level. This response is due to precise top-down predictions that are violated when the first missing chirp is not heard. This response is attenuated, when the log precision of the second level is reduced to two (middle row). This renders top-down predictions more sensitive to

have been predicted on the basis of top-down (empirical) prior beliefs – is missed, leading to sensory prediction errors that nearly match the amplitude of the prediction errors elicited by the omission. The lower row shows predictions and prediction errors when there is a compensatory decrease in sensory log precision from two to minus two. Here, there is a failure of sensory prediction errors to entrain high-level expectations and subsequent false inference that persists in the absence of any stimuli.

an important but intuitive point: attenuated mismatch or violation responses in chronic schizophrenia may not reflect a failure to detect surprising events but reflect a failure to detect unsurprising (predictable) events. In other words, they may reflect the fact that *every event is surprising*. In summary, a reduced precision of (confidence in) top-down predictions means that everything is mildly surprising and may provide an explanation for the failure to confidently infer regularities in the sensorium (and for larger P50 responses to repeated stimuli). As noted above, abnormal P50, P300, and MMN responses have also been demonstrated in firstdegree relatives, and do not normalize with anti-dopaminergic treatment (Winterer and McCarley, 2011) – consistent with their status as trait phenomena. So what would happen if we tried to compensate for reduced prior precision by reducing sensory precision?

## **PRECISION AND HALLUCINATIONS**

The lower row of **Figure 5** shows the results of the simulation with a compensatory reduction in sensory log precision from two to minus two. Here, the omission–related response is abolished; however there is a complete failure of perceptual inference, during the song and after its termination. Although the tempo of the percept is roughly the same as the stimulus, there is loss of frequency tracking and syntax. This false percept emerges because sensory information is not afforded the precision needed to constrain or entrain top-down predictions. The structured and autonomous nature of these predictions is an inevitable consequence of a generative model with deep structure – that is required to explain the dynamic and non-linear way in which our sensations are caused. The ensuing false inference can be associated with hallucinosis in the sense that there is a perceptual inference in the absence of sensory evidence. Clearly, the computational anatomy of hallucinations in the psychotic state is probably much more complicated – and specific to the domain of self-made acts (such as speech and movement). We will turn to the misattribution of agency in the final section. Here, it is sufficient to note that a compensatory reduction of sensory precision could produce hallucinosis of the sort seen in organic psychosyndromes. Note that the prediction error persists throughout the stimulus train and has, paradoxically, lower amplitude than in the previous simulations. This is because the prediction error is precision weighted – and we have reduced its precision.

## **SUMMARY**

In summary, we have used a fairly sophisticated generative model with dynamical and hierarchical structure to recognize sequences of simulated chirps in birdsong. This is a difficult Bayesian filtering problem that the brain seems to solve with ease. The key thing to take from these simulations is that some of the trait abnormalities associated with psychosis (schizophrenia) can be explained by a loss of precise top-down predictions – rendering everything relatively surprising (c.f., delusional mood), and reducing the difference between responses to standard and oddball stimuli. The loss of precise top-down (empirical) priors can also be invoked to explain a resistance to illusions (Silverstein and Keane, 2011) that depend upon prior beliefs. We will revisit this in the context of the force-matching illusion in the last section. One can compensate for relatively precise sensory prediction errors by reducing sensory precision – but at the expense of dissociating from the sensorium and false (hallucinatory) inference. This compensated state could be a metaphor for some psychotic states. Having said this, the fact that the hallucinations of schizophrenia respond to antipsychotics suggests that they are associated with a hyper-dopaminergic state and may involve a failure of sensory attenuation of corollary discharge (see last section). In the next section, we ask what would happen if perceptual deficits of this sort occurred during active inference and affected motor behavior.

## **ABNORMALITIES OF SMOOTH PURSUIT UNDER VISUAL OCCLUSION**

This section uses a generative model for smooth oculomotor pursuit to illustrate the soft neurological signs that result from changing the precision of prediction errors in active inference. This example is particularly pertinent to schizophrenia where, arguably, some of the most reproducible signs are found in terms of eye movements. To simulate anticipatory smooth pursuit eye movements, we require a hierarchical model that generates hidden motion. One such model is summarized in **Figure 6** (see figure legend for details). In brief, this model produces smooth pursuit eye movements because it embodies prior beliefs that gaze *x* (1) *<sup>o</sup>* and the target *x* (1) *t* are attracted by the same invisible point *v* (1) in the visual field. Target motion then provides evidence that the attracting (invisible) point is moving, which induces posterior beliefs that the eye will be attracted to that moving point. These posterior beliefs create proprioceptive predictions that descend to the oculomotor system, where they are fulfilled by oculomotor reflexes (see **Figure 6**). Crucially, we also equipped the subject with (veridical) prior beliefs that the invisible point moves with sinusoidal motion (equations at the second level in **Figure 6**) – so that, during periods of visual occlusion, the subject can anticipate where the target will reappear. This part of the model constitutes the highest hierarchical level and allowed us to simulate smooth pursuit of a target with sinusoidal motion that passes temporarily behind a visual occluder.

## **SIMULATING PSYCHOPATHOLOGY**

We modeled a putative deficit in schizophrenia by reducing the precision on the prediction errors of hidden states at the second level. Lowering this precision (the precision of ω (2) *<sup>x</sup>* in **Figure 6**) reduces the contribution of prediction errors to the posterior expectations modeling (hidden) periodic motion of the target. This results in a slowing of the (prior beliefs about the) target trajectory, as confidence in the prediction errors about its motion falls. This would normally place more emphasis on bottom-up prediction errors to guide inference; however, during occlusion these prediction errors are not available and we should see a behavioral effect of reducing precision.

To test for these behavioral effects, we reduced the log precision on the second level from −1 to −1.25. Neurobiologically, this corresponds to a reduction in the post-synaptic gain of superficial pyramidal cells encoding prediction error in cortical areas responsible for representing regularities in target motion. **Figure 7** shows the resulting active inference (upper panels) and trajectories of the target (solid black line) and eye (broken red lines) in the middle and bottom panels respectively. Comparison with the equivalent results under normal precision (broken black lines) reveals some characteristic properties of schizophrenic pursuit. First, with reduced precision, pursuit is disproportionately affected by target occlusion: at the end of occlusion, the lag behind the target is increases. This is despite the fact that when the target is visible and pursuit is stabilized, the tracking is normal (1200– 1400 and 2000–2200 ms). This reproduces empirical findings in schizophrenia at modest speeds (see Thaker et al., 1999). Second, pursuit under reduced precision is less accurate on the third cycle than the first, consistent with a deficit in inferring the target trajectory. Indeed, it lags so much just prior to 2700 ms that it has to make a catch-up saccade when the target re-emerges (saccades exceed 30 °/s). Overall, these results are consistent with findings in schizophrenia that suggest an impaired ability to maintain veridical pursuit eye movements in the absence of visual information. Furthermore, they suggest that the computational mechanism that underlies this failure rests on a failure to assign precision or certainty to (empirical) prior beliefs about hidden trajectories.

The relative loss of certainty about top-down predictions may also explain the ability of schizophrenics to respond to unpredicted changes in direction of the target. To demonstrate this, we removed the occluder, decreased the target period to around 500 ms, and introduced an unexpected reversal in the motion of the target – at the beginning of the second cycle of motion (at around 780 ms). The results of these simulations are shown in **Figure 8**. The traces in black correspond to normal pursuit and the traces in red show the performance under reduced precision. Although the effect is small (as it is in real subjects – Hong et al., 2005), the schizophrenic simulation (red lines) shows *more accurate* pursuit performance, both in terms of the displacement between the target and center of gaze, and in terms of a slight reduction in the peak velocity during the compensatory eye movement – a movement that is nearly fast enough to be a saccade. These differences are highlighted by red circles.

inputs is much simpler and is summarized by the equations specifying the generative process (lower left). The real-world provides sensory input in two modalities: proprioceptive input from cranial nerve nuclei reports the (horizontal) angular displacement of the eye **s**<sup>o</sup> and corresponds to the center of gaze in extrinsic coordinates **x**o. Exteroceptive (retinal) input reports the angular position of a target in a retinal (intrinsic) frame of reference **s**<sup>t</sup> . This input models the response of 17 visual channels, each equipped with a Gaussian receptive field deployed at intervals of one angular unit – about 2° of visual angle. This input can be occluded by a function of target location O(**x**<sup>t</sup> ), which returns values between zero and one, such that whenever the target location **x**<sup>t</sup> is behind the occluder retinal input is zero. The response of each visual channel depends upon the distance of the target from the center of gaze. This is just the difference between the oculomotor angle and target location. The hidden states of this model comprise the oculomotor states – oculomotor angle and velocity **x**<sup>o</sup> , **x** 0 o and the target location. Oculomotor velocity is driven by action and

## **SUMMARY**

In summary, a reduction in the precision of high-level prediction errors can account for both impaired smooth pursuit eye movements during occlusion and the paradoxical improvement of responses to unpredictable changes in target direction. This dissociation makes perfect sense from the point of view of the computational anatomy we have modeled here – reducing synaptic gain (precision) at high levels of a hierarchical predictive coding scheme reduces confidence in predictions that impairs performance when these predictions are needed (during occlusion) and

sinusoid), with a time constant of one time bin or 16 ms. The random fluctuations on sensory input and the motion of hidden states had a log precision of 16. The generative model (lower right) has a similar form to the generative process but with two important exceptions: there is no action and the motion of the hidden oculomotor states is driven by the same hidden cause that moves the target. In other words, the agent believes that its gaze is attracted to the same fictive point in visual space that is attracting the target. Second, the generative model is equipped with a deeper (hierarchical) structure that can represent periodic trajectories in the hidden cause of target motion: hidden causes are informed by the dynamics of

(2)

any amplitude and a frequency – that is determined by a second level hidden

beliefs about the frequency of periodic motion. The log precisions on the random fluctuations in the generative model were three at the first (sensory) level and minus one at the higher level, unless stated otherwise.

(2) with a prior expectation of η. This prior expectation corresponds to

hidden states at a second level x˙

cause v

. These model sinusoidal fluctuations of

**FIGURE 7 | Smooth pursuit of a partially occluded target with and without high-level precision**. These simulations show the results of applying Bayesian filtering Eq. 3 using the generative process and model of the previous figure. Notice, that in these simulations of active inference, there is no need to specify any stimuli explicitly – active sampling of the visual field means that the subject creates their own sensory inputs. The upper panels shows the responses of each of the (17) photoreceptors in image format as a function of peristimulus time. They illustrate the small fluctuations in signal that are due to imperfect pursuit and consequent retinal slip at the onset of target motion. Later, during periods of occlusion, the sensory input disappears. The lower panels show the angular displacement (top) and velocity (bottom) of the target (solid lines) and eye (broken lines) as a function

that improves performance when they are not (during unpredicted motion). In the final simulations, we retain a focus on active inference but instead of attenuating high-level precision we examine the effects of failing to attenuate low-level sensory precision.

## **SENSORY ATTENUATION, ATTRIBUTION OF AGENCY, AND DELUSIONS**

This section uses a generative model of (somatosensory) sensations that could be generated internally or externally. This model is used to illustrate the perceptual consequences of sensory attenuation, in terms of estimating the magnitude of externally of peristimulus time. They illustrate the remarkably accurate tracking behavior that is produced by prior beliefs that the center of gaze and target are drawn to the same fictive point – beliefs that action fulfils. The gray area corresponds to the period of visual occlusion. The upper right panel shows sensory input when the precision of prediction errors on the motion of hidden states at the second level was reduced from a log precision of −1 to −1.25. The associated behavior is shown with red broken lines in the lower panels. The dashed horizontal line in the lower panel corresponds to an angular velocity (30°); at which the eye movement would be considered saccadic. This simulation illustrates the loss of Bayes-optimal tracking when the motion of the target corresponds to high-level posterior beliefs but the precision of these beliefs is attenuated.

and internally generated events. In brief, we reproduce the force-matching illusion (Shergill et al., 2003, 2005) by yoking externally applied forces to the perceived level of self-generated forces. Finally, we demonstrate the disappearance of the illusion and the emergence of false inferences about (antagonistic) external forces, when there is a failure to attenuate sensory precision and a compensatory increase in the precision of empirical prior beliefs.

## **ACTIVE INFERENCE AND SENSORY ATTENUATION**

Sensory attenuation refers to a decrease in the intensity of a perceived stimulus when it is self generated (Blakemore et al., 1998).

broken traces in red show the performance under reduced precision. Although the effect is small, reducing the precision about prior beliefs produces more accurate pursuit performance, both in terms of the displacement between the target and center of gaze and in terms of a slight reduction in the peak velocity during the compensatory eye movement (red circles). This illustrates the paradoxical improvement of performance that rest upon precise sensory information that cannot be predicted a priori (and is characteristic of syndromes like schizophrenia and autism).

We have suggested that sensory attenuation is necessary to allow reflex arcs to operate (Brown et al., in press). The argument is simple: proprioceptive prediction errors can only be resolved by moving – via motor reflexes – or by changing predictions. This means the effects of ascending prediction errors on posterior expectations must be attenuated to allow movement: if proprioceptive sensations are conveyed by ascending primary (Ia and Ib) sensory afferents with too much precision, then they would subvert descending predictions that create prediction errors and therefore prevent movement. It is therefore necessary to temporarily suspend the precision of sensory reafference to permit movement. If we associate the perceived intensity or detectability of the sensory consequences of action with a lower bound on their posterior confidence interval, attenuation of sensory precision provides a simple explanation for the attenuation of the perceived intensity of selfgenerated sensations. In what follows, we present simulations of sensory attenuation by simulating the force-match illusion and then demonstrate how overly precise prior beliefs can compensate for a failure of sensory attenuation but expose the actor to somatic delusions.

## **THE GENERATIVE PROCESS AND MODEL**

**Figure 9** summarizes the generative process and model (using the form of Eq. 2). This model is as simple as we could make it, while retaining the key ingredients that are required to demonstrate inference about or attribution of agency. The equations on the left describe the real world, while the equations on the right constitute the subject's generative model. In the real world, there is one hidden state **x***<sup>i</sup>* modeling self-generated force that is registered by both proprioceptive *s<sup>p</sup>* and somatosensory *s<sup>s</sup>* inputs. Externally generated forces **v***<sup>e</sup>* are added to internally generated forces to provide somatosensory input. The key thing about this model is that somatosensory sensations are caused ambiguously, by either internally or externally generated forces:*s<sup>s</sup>* = **x***<sup>i</sup>* + **v***<sup>e</sup>* . The only way that the underlying cause of the sensations can be inferred is by reference to proprioceptive input – that is only generated internally. This is a very simple model, where the somatosensory input is used metaphorically to represent the sensory consequences of events that could be caused by self or others, while proprioceptive input represents signals that can only be caused by self-made acts. Active inference now compels the subject to infer the causes of its sensations.

The generative model used for this inference is shown on the right. In this model, internally and externally generated forces (*xi* , *xe*) are modeled symmetrically, where changes in both are attributed to internal and external hidden causes (*v<sup>i</sup>* , *ve*). The hidden causes trigger the dynamics associated with the hidden states, much like the push that sets a swing in motion. This means that proprioceptive and somatosensory inputs are explained in terms of hidden causes, where proprioceptive sensations are caused by internally generated forces and somatosensory consequences report a mixture of internal and external forces. Crucially, the precision afforded sensory prediction errors depends upon the internally generated force (and its hidden cause). This dependency is controlled by a parameter γ that mediates the attenuation of sensory precision: as internally generated forces rise, sensory precision falls, thereby attenuating the amplitude of (precision weighted) sensory prediction errors. These context or state-dependent changes in precision enable the agent to attend to sensory input, or not – depending upon the relative precision of prediction errors at the sensory and higher levels. This context sensitive sensory precision is shown in **Figure 10** as π (cyan circles).

## **FUNCTIONAL ANATOMY**

**Figure 10** illustrates how this generative model could be transcribed into a plausible neuronal architecture. In this example, we have assigned sensory expectations and prediction errors to the thalamus, while corresponding expectations and prediction

**used in the simulations of sensory attenuation**. The generative process (on the left) models real-world states and causes, while the model on the right is the generative model used by the subject. In the real world, the hidden state x<sup>i</sup> corresponds to self-generated pressures that are sensed by both somatosensory s<sup>s</sup> and proprioceptive s<sup>p</sup> input channels. External forces are modeled with the hidden cause v<sup>e</sup> and are sensed only by the somatosensory channel. Action causes the self-generated force x<sup>i</sup> to increase and is modified by a sigmoid squashing function σ. The hidden state decays slowly over four time bins.

errors about hidden states (forces) are associated with the sensorimotor cortex. The expectations and prediction errors about the hidden causes of forces have been placed – somewhat agnostically – in the prefrontal cortex. Notice how proprioceptive predictions descend to the spinal cord to elicit output from alpha motor neurons (playing the role of proprioceptive prediction error units) that cause movements through a classical reflex arc. Red connections denote ascending prediction errors, black connections descending predictions (posterior expectations), and the cyan connection denotes descending neuromodulatory effects that mediate sensory attenuation. The ensuing hierarchy conforms to the functional form of the predictive coding scheme in Eq. 3. In this architecture, predictions based on expected states of the world can either be fulfilled by reflex arcs or they can be corrected by ascending sensory prediction errors. Which of these alternatives occurs depends on the relative precisions along each pathway – that are set by the descending modulatory connection to sensory prediction errors. We now use this model to demonstrate some key points.

## **SENSORY ATTENUATION AND THE FORCE-MATCHING ILLUSION**

To produce internally generated movements, we simply supplied the subject with prior beliefs that the internal hidden cause increased transiently to a value of one, with high sensory attenuation γ = 6. We then followed this self-generated movement with v<sup>i</sup> and external causes ve. The hidden cause excites dynamics in hidden states x<sup>i</sup> and xe, which decay slowly. Internal force is perceived by both proprioceptive and somatosensory receptors, as before, while external force is perceived only by somatosensory receptors. Crucially, the precision of the sensory input ω<sup>s</sup> is influenced by the level of internal force, again modulated by a squashing function, and controlled by a parameter γ that governs the level of attenuation of precision. The generalized predictive coding scheme associated with this generative model is shown schematically in the next figure.

an exogenously generated force that matched the self-generated force. The left-hand panels in **Figure 11** show the results of this simulation. The lower left panel shows the internal hidden cause (blue line) with relatively tight 90% confidence intervals (gray areas). Prior beliefs about this hidden cause excite posterior beliefs about internally generated forces, while at the same time attenuating the precision of sensory prediction errors. This is reflected by the rise in the posterior expectation of the internal force (blue line in the upper right panel) and the transient increase in the confidence interval about this expectation. The resulting proprioceptive predictions are fulfilled by action (bottom right panel) to produce the predicted sensations (upper left panel). Note that proprioceptive prediction (blue line) corresponds to somatosensory prediction (green line) and that both are close to the real values (broken black line). This simulation shows normal self-generated movement under permissive sensory attenuation.

The right-hand panels of **Figure 11** show exactly the same results as in the left-hand panels; however here, we have yoked the exogenous force *x<sup>e</sup>* to the self-generated force *x<sup>i</sup>* perceived at 90% confidence (dotted line in the top right graph) – as opposed to the true force exerted by the subject. In other words, the external force corresponds to the force that would be reported by the subject to match the perceived force at 90% confidence. The 90% confidence interval was chosen as a proxy for the percept to reconcile

the perceived intensity literature with results from signal detection paradigms (Cardoso-Leite et al., 2010). Experimental work in the auditory domain has demonstrated that perceived intensity can be attenuated by increasing sensory noise (decreasing precision) (Lochner and Burger, 1961; Richards, 1968). When coupled to the 90% confidence interval, the internally generated force is now much greater than the matched external force (shown on the upper left graph). This is the key finding in the force-matching illusion and is entirely consistent with sensory attenuation. In this setting, the loss of confidence in posterior estimates of hidden states that

are self-generated translates into an illusory increase in the force applied, relative to the equivalent force in the absence of sensory attenuation.

We repeated these simulations under different levels of selfgenerated forces by modulating the prior beliefs about the internal hidden cause (from a half to twice the normal amplitude). The results are shown as the blue circles in the left panel of **Figure 12**, which plots the self-generated force against the yoked or matched external force with a corresponding 90% confidence interval. These results are remarkably similar to those obtained empirically (right panel – reproduced from Shergill et al., 2005) and reveal sensory attenuation through an illusory increase in the self-generated force, relative to matched forces over a wide range of forces. The red line in the left panel comes from the final simulations, in which we asked what would happen if subjects compensated for a failure in sensory attenuation by increasing the precision of their prior beliefs?

## **FALSE INFERENCE AND FAILURES OF SENSORY ATTENUATION**

We now demonstrate two pathologies of sensory attenuation: first, a loss of sensory attenuation resulting in a catatonic state and second, how compensation for such a loss could allow movement but result in a somatic delusion. The consequences of reducing sensory attenuation (from six to two) are illustrated in the left panels of **Figure 13**. Here, the loss of sensory attenuation maintains the precision of the hidden states above the precision of prior beliefs about hidden causes (lower left panel). This means that bottom-up sensory prediction errors predominate over top-down predictions and expectations about internally generated forces are profoundly suppressed. Because there are no predictions about proprioceptive changes, there is a consequent akinesia. This state is reminiscent of the catatonic symptoms of schizophrenia such as immobility, mutism, catalepsy and waxy flexibility, in which the patient may maintain a fixed posture for a long time, even though (in the case of waxy flexibility) their limbs can be moved easily by someone else.

We shall now examine how a loss of sensory attenuation might be compensated for by increasing the precision of prediction errors at higher levels in the hierarchy (by increasing the log precision of prediction errors on hidden states and causes by four log units). This compensatory increase is necessary for movement and ensures the precision of top-down predictions is greater than bottom-up sensory prediction errors. These manipulations permit movement but abolish the force-matching illusion, as indicated by the line of red circles in the left panel of **Figure 12**. One might ask – why don't subjects adopt this strategy and use precise prior beliefs about hidden causes all the time?

The answer is evident in the right panels of **Figure 13**, which show the results of a simulation with low sensory attenuation and compensatory increases in precision at higher levels. Here, there is an almost perfect and precise inference about internally and externally generated sensations. However, there is a failure of inference about their hidden causes. This can be seen on the lower left, where the subject has falsely inferred an antagonistic external hidden cause that mirrors the internal hidden causes. Note that this false inference does not occur during

time in 100 ms time bins; the y axes force in Newtons. Left panels: in the first part of this simulation an internal force is generated from a prior belief about the cause v<sup>i</sup> , followed by the presentation of an external force. Posterior beliefs about the hidden states (upper right panel) are similar, but the confidence interval around the force for the internally generated state is much broader. This is because sensory level precision must be attenuated in order to allow proprioceptive predictions to be fulfilled by reflex arcs instead of being corrected by sensory input: i.e., the confidence intervals around v<sup>i</sup> must be narrower than those around x<sup>i</sup> to allow movement to proceed. If perceived

of the estimate of hidden state (highlighted by the dotted line), it will be lower when the force is self generated than when the force is exogenous (the difference is highlighted by the arrow). Right panels: the simulation was repeated but the external force was matched to the lower bound of the 90% confidence interval of the internal force. This means that internally generated force is now greater than the externally applied force (double-headed arrow, upper left panel). This reproduces the normal psychophysics of the force-matching illusion that can be regarded as entirely Bayes-optimal, under appropriate levels of precision.

normal sensory attenuation (see **Figure 11**), where the true external hidden cause always lies within the 90% confidence intervals. The reason for this false inference or delusion is simple: action is driven by proprioceptive prediction errors that always report less force than that predicted. However, when these prediction errors are very precise they need to be explained – and can only be explained by falsely inferring an opposing exogenous force. This only occurs when both the predictions and their consequences are deemed to be very precise. This false inference could be interpreted as a delusion in the same sense that the sensory attenuation is an illusion. Having said this, it should be noted that – from the point of view of the subject – its inferences are Bayes-optimal. It is only our attribution of the inference as false that gives it an illusory or delusionary aspect.

This simulation has some face validity in relation to empirical studies of the force-matching illusion. The illusion is attenuated in normal subjects that score highly on ratings of delusional beliefs (Teufel et al., 2010). Furthermore, subjects with schizophrenia – who are prone to positive symptoms like delusions – do not show the force-matching illusion (Shergill et al., 2005). In other words, there may be a trade-off between illusions at a perceptual level and delusions at a conceptual level that is mediated by a (failure of) sensory attenuation.

## **SUMMARY**

The ideas reviewed in this section suggest that attribution of agency – in an ambiguous situation – can be resolved by attenuating the precision of sensory evidence during movement: in other words, attending away from the consequences of self-made

acts. When implemented in the context of active inference, this provides a Bayes-optimal explanation for sensory attenuation and attending illusions. The simulations show how exacerbations of a trait loss of sensory attenuation could subvert movement and even cause catatonia. This can be ameliorated by compensatory increases in high-level precision, which in turn necessarily induce false (delusional) inferences about agency. This is important, given the negative correlation between sensory attenuation and predisposition to delusional beliefs in normal subjects and the reduced force-matching illusion in schizophrenia. On a physiological level, increased dopaminergic transmission in the striatum could reflect a putative increase in high-level precision, compensating for hypofunction of cortical NMDA-Rs. In summary, we have shown how active inference can explain the fundamental role of sensory attenuation, and how its failure could lead to not only catatonic states but also compensatory changes that induce delusions. This is one illustration of how psychotic state abnormalities might be secondary compensations for trait abnormalities.

## **CONCLUSION**

Bayesian computations enable inference and learning under uncertainty. Furthermore, they prescribe the optimal integration of prior expectations (amassed over a lifetime or indeed evolution) with the sensory evidence of a moment; this integration is optimal because it embodies the relative uncertainty (precision) of each source of information. For this reason, the accurate representation of precision in a hierarchical Bayesian scheme is crucial for inference. The aberrant encoding of precision can therefore lead to false inference by overweighting prior expectations or sensory evidence. This paper has described how various trait abnormalities in schizophrenia could result from a decrease in prior precision (or a failure to attenuate sensory precision); and how some psychotic states could result from compensatory increases in prior precision (or decreases in sensory precision). We have outlined several physiological mechanisms for encoding precision (such as neuromodulation and neuronal oscillations) that are abnormal in schizophrenia. Genetic and neuropathological evidence suggest that NMDA-R (and GABA to some extent) may play a role in trait abnormalities, whereas the physiological evidence points toward dopaminergic pathology in the psychotic state. Clearly, a strict dichotomy is unlikely, since these neurotransmitter systems have complex interactions.

Using a biologically plausible predictive coding scheme, we have shown how a reduction of high-level (prior) precision can account for two trait phenomena: abnormal ERP responses to predictable and unpredictable stimuli and SPEM abnormalities. We have also shown how a failure to attenuate sensory precision might explain a resistance to (force-matching) illusions and (in severe cases) catatonia. Using these model systems, we were able to explain the delusional and hallucinatory inference characteristic of the psychotic state by compensatory increases (resp. decreases) in prior (resp. sensory) precision.

One might ask how specific these "trait" and "state" simulations are to schizophrenia, as opposed to psychotic symptoms *per se*. An important point to take from the formal arguments in this paper is that the common factor underlying psychotic phenomena is computational, not physiological: i.e., the key to understanding these

symptoms is as disorders of precision encoding, and not – for example – necessarily of a particular neuromodulator. Another important message is that these simulations undermine a clear division between "normal" and "psychotic" brains, as even bizarre phenomena such as somatic delusions can occur in a normal inferential architecture in which precision encoding is awry. To what extent the physiological (or pharmacological) causes of transient psychotic symptoms in healthy people overlap with similar symptoms in schizophrenia is an interesting question, which physiologically informed models may help us to address (Moran et al., 2011).

Simulations of the sort used above clearly require empirical validation: this should be possible as the models make quantitative predictions about the dynamics of cortical populations that can be tested with dynamic causal modeling (Friston et al., 2003). Indeed, dynamic causal modeling studies of schizophrenic subjects have already demonstrated changes in effective connectivity consistent with decreased high level – and increased low-level – precision in

the hollow mask paradigm (Dima et al., 2009, 2010). We conclude with some of the many interesting and outstanding questions in the computational modeling of schizophrenia:


the contribution of developmental and environmental stressors (Giovanoli et al., 2013) or a combination of the above.


## **REFERENCES**


cortex in schizophrenia. *Arch. Gen. Psychiatry* 48, 625–632. doi:10.1001/archpsyc.1991.0181031 0043008


are formalized with appropriate models (as in Moran et al., 2011).

• Finally, schizophrenia is unlikely to be the only pathology of precision – it is notable that most current treatments for psychiatric disorders target neuromodulatory systems. Aberrant precision estimation may also prove to be a simple but powerful explanation for other psychiatric disorders; e.g., the loss of central coherence in autism (Pellicano and Burr, 2012). In future, many psychiatric disorders may be distinguished by their sites of – and causes of – variation in their encoding of precision; e.g., the precision of distributions over future action outcomes modeling helplessness in depression (Huys and Dayan, 2009).

## **ACKNOWLEDGMENTS**

This work was funded by the Wellcome Trust. Klaas Enno Stephan would like to acknowledge support by the René and Susanne Braginsky Foundation. The authors would also like to thank Steve Silverstein for his very useful comments on the manuscript, and also our reviewers whose suggestions have improved the paper.


21, 1740–1745. doi:10.1177/ 0956797610389187


affordance and active inference. *PLoS Comput. Biol.* 8:e1002327. doi:10.1371/journal.pcbi.100 2327


Targeting the dopamine D1 receptor in schizophrenia: insights for cognitive dysfunction. *Psychopharmacology (Berl.)* 174, 3–16. doi:10.1007/s00213-004-1793-y


sustained visual tracking in schizophrenic patients. *Exp. Brain Res.* 165, 125–131. doi:10.1007/s00221- 005-2276-z


schizophrenic subjects. *Proc. Natl. Acad. Sci. U.S.A.* 93, 9235–9240. doi:10.1073/pnas.93.17.9235


decreased dendritic length and spine density in layers III and V of sensorimotor neocortex. *Synapse* 53, 114–121. doi:10.1002/syn.20039


for executive control microcircuits. *J. Cogn. Neurosci.* 24, 2334–2347. doi:10.1162/jocn\_a\_00307


in the prefrontal cortex of subjects with schizophrenia. *Arch. Gen. Psychiatry* 58, 466–473. doi:10.1001/archpsyc.58.5.466


errors predicts s-ketamine-induced cognitive impairments. *Neuropsychopharmacology* 37, 865–875. doi:10.1038/npp.2011.261


and P. J. Harrison (Oxford: Wiley-Blackwell), 311–333.


doi:10.1097/00001756-199705260- 00035

Young, A. M. J., Moran, P. M., and Joseph, M. H. (2005). The role of dopamine in conditioning and latent inhibition: what, when, where and how? *Neurosci. Biobehav. Rev.* 29, 963–976. doi:10.1016/j.neubiorev. 2005.02.004

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 15 April 2013; paper pending published: 25 April 2013; accepted: 16 May 2013; published online: 30 May 2013.*

*Citation: Adams RA, Stephan KE, Brown HR, Frith CD and Friston KJ (2013) The computational anatomy of psychosis. Front. Psychiatry 4:47. doi: 10.3389/fpsyt.2013.00047*

*This article was submitted to Frontiers in Schizophrenia, a specialty of Frontiers in Psychiatry.*

*Copyright © 2013 Adams, Stephan, Brown, Frith and Friston. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in other forums, provided the original authors and source are credited and subject to any copyright notices concerning any third-party graphics etc.*

# Reinforcement learning and dopamine in schizophrenia: dimensions of symptoms or specific features of a disease group?

#### **Lorenz Deserno1,2\*, Rebecca Boehme<sup>2</sup> , Andreas Heinz <sup>2</sup> and Florian Schlagenhauf 1,2**

<sup>1</sup> Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany

<sup>2</sup> Department of Psychiatry and Psychotherapy, Charité – Universitätsmedizin Berlin, Berlin, Germany

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

James A. Waltz, University of Maryland School of Medicine, USA Guillermo Horga, Columbia University Medical Center, USA

#### **\*Correspondence:**

Lorenz Deserno, Department of Psychiatry and Psychotherapy, Charité – Universitätsmedizin Berlin, Campus Mitte, Charitéplatz 1, 10117 Berlin, Germany e-mail: lorenz.deserno@charite.de

Abnormalities in reinforcement learning are a key finding in schizophrenia and have been proposed to be linked to elevated levels of dopamine neurotransmission. Behavioral deficits in reinforcement learning and their neural correlates may contribute to the formation of clinical characteristics of schizophrenia. The ability to form predictions about future outcomes is fundamental for environmental interactions and depends on neuronal teaching signals, like reward prediction errors. While aberrant prediction errors, that encode nonsalient events as surprising, have been proposed to contribute to the formation of positive symptoms, a failure to build neural representations of decision values may result in negative symptoms. Here, we review behavioral and neuroimaging research in schizophrenia and focus on studies that implemented reinforcement learning models. In addition, we discuss studies that combined reinforcement learning with measures of dopamine. Thereby, we suggest how reinforcement learning abnormalities in schizophrenia may contribute to the formation of psychotic symptoms and may interact with cognitive deficits. These ideas point toward an interplay of more rigid versus flexible control over reinforcement learning. Pronounced deficits in the flexible or model-based domain may allow for a detailed characterization of well-established cognitive deficits in schizophrenia patients based on computational models of learning. Finally, we propose a framework based on the potentially crucial contribution of dopamine to dysfunctional reinforcement learning on the level of neural networks. Future research may strongly benefit from computational modeling but also requires further methodological improvement for clinical group studies.These research tools may help to improve our understanding of disease-specific mechanisms and may help to identify clinically relevant subgroups of the heterogeneous entity schizophrenia.

**Keywords: schizophrenia, dopamine, computational modeling, reinforcement learning, aberrant salience, prediction error, fMRI, PET imaging**

## **INTRODUCTION AND OUTLINE**

The "dopamine-hypothesis" of schizophrenia was initially built upon the observation that dopamine receptor antagonists, such as haloperidol, attenuate psychotic symptoms (1). Evidence showing that elevated dopamine levels are indeed involved in the pathophysiology of psychotic symptoms and schizophrenia is primarily derived from neurochemical studies using positronemission-tomography (PET) with radioactive ligands targeting the brain's dopamine system. Such studies clearly indicate elevated levels of presynaptic dopamine function (2, 3) with particularly strong evidence from meta-analyses for elevated dopamine synthesis capacity (4, 5). A hallmark of dopamine research was the observation that phasic releases of dopaminergic neurons code a temporal-difference prediction error, which was later shown to be causally involved in learning (6–8). This ability to form predictions about future outcomes is fundamental for interactions with the environment and depends on neuronal representations of such teaching signals. Behavioral impairments in reinforcement learning are a key finding in schizophrenia patients and have been proposed to be closely linked to reports of elevated presynaptic dopamine neurotransmission. Influential theoretical work suggests that dysfunctional reinforcement learning may contribute to the formation of the prominent clinical characteristics of schizophrenia patients, namely positive and negative symptoms (9–11). Furthermore, prediction errors are involved in learningrelated changes in synaptic plasticity (12), and aberrant plasticity has been suggested as a potential common biological mechanism characterizing the schizophrenia spectrum (13, 14).

Embedded in this context, the central attempt of this article is to review studies on reinforcement learning in schizophrenia and to disentangle dimensions of symptom formation and potential disease-specific mechanisms in the existing literature. The primary focus of this article is to provide an up-to-date overview of the existing literature with the aim to review existing evidence for two influential theories. Therefore, we only include a brief introduction (see Reinforcement Learning in Schizophrenia: Theoretical Considerations) to these hypotheses and refer to the original publications for more detailed theoretical descriptions. The empirical studies reviewed here comprise behavioral and functional neuroimaging studies [restricted to functional magnetic resonance imaging (fMRI) and PET] in patients suffering from schizophrenia. In the first part, we start with studies on reward anticipation and processing based on pre-learned contingencies. Subsequently, we focus on studies that directly examine learning over time with a focus on studies that implemented reinforcement learning models. Finally, we summarize studies that combined experimental perturbations of the brain's dopamine system, such as pharmacological challenges and molecular imaging (PET), with measures of reinforcement learning.

## **FEEDBACK ANTICIPATION AND PROCESSING**

A series of studies used the monetary incentive delay task (MID), a paradigm invented by Knutson and colleagues [(15), see also Ref. (16)]. The initial study demonstrated that participants speed up motor responses to obtain rewards and that anticipation as well as delivery of rewards evoke ventral striatal activation. The first application of this task in schizophrenia patients was carried out by Juckel and colleagues: they found reduced ventral striatal activation in unmedicated patients (17). This finding was subsequently replicated in a larger cohort of drug-naïve, first-episode patients (18, 19). In the study by Juckel et al. (17), it was demonstrated that blunting of anticipatory ventral striatal activation elicited by monetary reward reflected the individual degree of negative symptoms (17). This association was also present in patients treated with typical or first generation antipsychotics (FGAs, or "typical" antipsychotics), who showed reduced ventral striatal activation during reward anticipation, while patients treated with atypical or second generation antipsychotics (SGAs, or "atypical" antipsychotics) showed intact activation during anticipation of monetary reward in the same region (20, 21). This effect of SGAs was recently replicated in a larger cohort of patients (22). In line with these results, further studies applying the MID task in chronic schizophrenia patients medicated predominantly with SGAs did not find reduced ventral striatal anticipation of monetary reward in the patient group, as a whole (23–25). Two studies replicated the association of reward anticipation with negative symptoms (25) and apathy (24), while two other studies reported a correlation of ventral striatal activation during reward anticipation with positive symptoms (18, 19).

Although the static MID task is thought to mirror aspects of animal experiments studying reinforcement learning in the dopaminergic system [e.g., Ref. (6)], the gross time scale of fMRI compared to neurophysiological studies needs to be taken into account (26). Nevertheless, it has been demonstrated that ventral striatal activation during reward anticipation is indeed modulated by dopamine: a positive correlation between the anticipatory activation in core dopamine areas and reward-induced dopamine release was observed via competition of endogenous dopamine with a PET D2/3-receptor radioligand (27). In a study by Knutson et al. (28), diminished ventral striatal reward anticipation was reported when comparing healthy participants receiving amphetamine (resulting in a massive release of dopamine) to placebo (28). The latter study coincides with the results reported above in schizophrenia patients during reward anticipation and the wellestablished finding of elevated presynaptic dopamine function in schizophrenia using PET with FDOPA and similar tracers [for meta-analyses see: Ref. (4, 5)]. Based on this,it appears conceivable that event-related responses to reward-indicating cues disappear in the noise of elevated dopaminergic activity observed in schizophrenia patients and that this may ultimately contribute to a failure of salience attribution to environmentally relevant stimuli (9, 10, 29). Interestingly, Esslinger et al. (18) implemented the MID task in combination with another task possibly reflecting salience and showed in an exploratory correlation analysis that more pronounced ventral striatal hypoactivation during reward anticipation was associated with more salience attribution to neutral stimuli (18). In line with this, a recent study using emotional picture stimuli demonstrated that schizophrenia patients rate neutral pictures as more salient (30). These results provide some rather indirect support for the idea of aberrant salience in schizophrenia, which we will briefly introduce in the following section.

In contrast to reward anticipation, fewer studies used the MID task to examine the delivery of monetary outcome. One study (31) found that violations of outcome expectancies triggered abnormal neural responses in unmedicated patients: While medialprefrontal activation was exaggerated when an expected-reward was omitted, ventral striatum (VS) displayed reduced activation for successful versus unsuccessful loss avoidance. The degree of delusion severity was found to be associated with activation in medial-prefrontal cortex (PFC) for successful versus unsuccessful loss avoidance. Moreover, functional connectivity between VS and medial PFC was reduced in patients. In a similar vein, Waltz et al. (25) found reduced activation in the medial PFC and lateral PFC when comparing win versus loss trials in schizophrenia patients medicated with SGAs. Activation to reward delivery in lateral PFC was negatively correlated with the degree of positive and negative symptoms (25). Another study (23) tested high and low rewards together with high and low punishments against neutral events and found significant activation in lateral PFC of healthy controls, most likely reflecting salience. This activation pattern was diminished in patients treated with SGAs. Interestingly, a recent study showed exaggerated activation in dorsolateral PFC elicited by neutral outcomes in unmedicated patients (19).

Two studies examined classical conditioning that actually took place outside the MRI scanner (32, 33). These designs might be thought of as extensions to studies using the MID task: Contingencies were pre-learned before scanning, but allow one to distinguish between expected-rewards, unexpected-rewards (presumably mirroring positive prediction errors), and unexpected omissions of rewards (presumably mirroring negative prediction errors; (32)). Juice was used as a primary reinforcer in 18 medicated patients (32). Attenuated neural responses in dopaminergic core areas (midbrain and striatum) to expected and unexpected-reward deliveries were observed,while activation in reward omission trials was largely intact. Morris et al. (33) completed this approach in a full 2 × 2 design, thereby enabling an orthogonalization of the factors "rewards" and "surprise" as well as the interaction of both factors, which is assumed to mirror prediction-error-related brain activation. In 21 schizophrenia patients medicated with SGAs, this revealed a disrupted differentiation between expected and unexpected events in a way that ventral striatal activation is not coding prediction errors: while response to expected events in right VS was exaggerated, response to unexpected outcomes in left VS was found to be blunted (33).

In summary, fMRI studies in reward processing using the MID task have so far provided important insights into the neural processes underlying outcome anticipation and delivery in schizophrenia. In particular, the finding of reduced ventral striatal activation during outcome anticipation was consistently replicated across three studies involving a total of 68 unmedicated patients. An association of anticipatory ventral striatal activation with negative symptoms was reported in five studies involving 10 unmedicated patients and 52 medicated patients. Antipsychotic medication remains a crucial issue since these drugs specifically block those striatal D2-receptors that are (among others) activated by potentially prediction-error-associated dopamine release [e.g., Ref. (34, 35)] and moreover affect presynaptic dopamine synthesis (36, 37). Therefore, assessing unmedicated patients is key to understanding dopamine dysfunction in schizophrenia and to avoiding confounds by medication effects, which also appear to differ depending on FGAs versus SGAs (20, 21). Furthermore, one important limitation of the studies discussed thus far is the fact that all reward contingencies are pre-learned (i.e., before participants enter the MRI scanner and perform the task). Anticipatory brain activation during the MID task is likely to capture some aspects of reinforcement learning in particular with respect to cueor action-related value signals. Some kind of value quantification is usually the main outcome variable of reinforcement learning models. It is important to note that these functions evolve over time, which is also a fundamental principle of brain signals. This points out an important limitation of the MID studies which may therefore provide a rather coarse proxy of value-related brain activation and consequently emphasizes the necessity to study learning over the course of time. Thus, studying the temporal dynamics underlying the actual learning process may provide more insights into symptom- and disease-specific processes associated with schizophrenia. In contrast to studies which used the MID task or similar designs, all studies discussed in Section "Behavioral Studies of Reinforcement Learning in Schizophrenia" refer to experimental paradigms that investigate learning on a trial-by-trial basis. Detailed computational modeling of such temporal dynamics may be particularly helpful to elucidate dysfunctional processes in patients and to improve characterization of a heterogeneous disease entity that is so far still based on symptoms (38–41).

## **REINFORCEMENT LEARNING IN SCHIZOPHRENIA: THEORETICAL CONSIDERATIONS**

Reinforcement learning represents a promising, theory-driven tool (42) which aims to quantify learning on a trial-by-trial basis and has so far been implemented in a limited number of clinical group studies [e.g.,Ref. (43),**Table 1**].Although there are several different variants of models, most of them separate two main contributors to the learning process and both of them change on trial-by-trial (**Box 1**): first, the delivered outcome which refers to the time point when prediction errors arise. This teaching signal is thought to be crucially involved in driving any learning process. Second, the values of environmental cues or actions which are learned via this teaching signal. Concepts of motivational or incentive salience are closely linked with values of actions or environmental cues (44) that can be acquired during prediction-error-driven trial-anderror learning. Differences in the perceived properties of feedback stimuli *per se* (e.g., shifts in hedonic experience or salience) may also influence the elicitation of prediction errors and thus potentially corrupt learning processes. Based on these two main time points, we will proceed with a brief summary of two influential hypotheses with respect to the potential contribution of reinforcement learning to symptom dimensions and disease-specific features in schizophrenia.

We begin with the"aberrant salience"hypothesis: schizophrenia patients may attribute salience to otherwise neutral environmental stimuli, and those stimuli may ultimately appear meaningful and evoke delusional mood in patients (9, 10). This process has been described as closely linked to a dysregulation of the dopamine system where both chaotic dopamine firing (45) and elevated baseline dopamine levels (46, 47) have been proposed to be involved. Whether this process actually reflects reinforcement learning in the same way as it was theoretically and mechanistically defined for healthy people (42) remains an open and exciting question. If this is the case, then neutral events should elicit prediction errors which may consequently train values for the associated cues or actions, and these values may finally exceed incentive values associated with rewarding or otherwise reinforcing events. In other words, patients are assumed to attribute importance to stimuli ignored by healthy volunteers and thereby learn something else. The degree of this alteration should be related to positive symptom levels, in particular delusions. It is important to note, that a prerequisite for the latter idea is that misattributed salience to certain neutral events remains stable over a period of time. Alternatively, it may also be possible that the process of misattributing salience is fluctuating permanently, resulting in a random pattern (a state where "everything is salient") that would formally result in no learning at all and might therefore be harder to quantify. It is also conceivable that aberrant aspects of reinforcement learning have not yet been formulated correctly. Here, the role of unsigned prediction errors, as a valence-unspecific salience signal, might be of interest and could possibly be integrated in models of reinforcement learning (48–50).

The second hypothesis focuses on a deficit in the representation of learned values (11). This hypothesis posits that prediction errors are not adequately used to learn values even though hedonic experience itself remains mainly intact. This concept relates closely to the idea that reward feedback is not adequately transformed into motivational drive for goal-directed behavior (51) and has been proposed as a potential mechanism for the origin of negative symptoms (11). In general, a failure to learn any value may also be based on a reduction of hedonic experience, in which case no prediction errors are elicited and therefore no values can be learned; based on studies reviewed in the next section, this appears to be unlikely in schizophrenia patients. On the other hand, a deficit in using monetary and primary rewards for motivated behavior would appear similar to what was proposed in the incentive-sensitization theory of addiction disorders, which assumes a shift from non-drug rewards to drug-related rewards (44). In schizophrenia, such a shift may predominantly concern neutral stimuli and therefore result in aberrant learning as pointed out in the aberrant salience hypothesis.



RW, Rescorla–Wagner-model; TD, temporal-difference model; SARSA, state action response state action; RPE, reward prediction error; VS, ventral striatum; RT, reaction time.

As indicated, the two hypotheses are only partially independent. It is possible that both mentioned mechanisms exist in parallel and converge in producing a behavioral deficit but diverge in their differential contribution to symptom formation. In the following, we will review studies that aimed to test these hypotheses. Thereby, we try to build a coherent picture of how reinforcement learning may contribute to the formation of psychotic symptoms and if this appears to be dimensional or categorical. Finally, we endeavor to interpret previous studies with regard to their disease specificity by summarizing and discussing those studies that examined learning over time.We start with behavioral studies followed by a section on imaging studies. We also mention if studies implemented models of reinforcement learning and how parameters underlying these models were inferred.

## **BEHAVIORAL STUDIES OF REINFORCEMENT LEARNING IN SCHIZOPHRENIA**

Behavioral deficits in associative learning, particularly in instrumental tasks where feedback is used to guide behavior, are frequently replicated in schizophrenia patients. So far, only seven studies have implemented models of reinforcement learning (see **Table 1**), and although reinforcement learning modeling quantifies the observed behavior, only two of these studies were purely behavioral; the other five studies also collected fMRI data and regressed model-derived learning time-series (e.g., prediction errors) against imaging data. Studies on classical conditioning are reported in the subsequent section, because all the clinical studies conducted so far have assessed classical conditioning effects via physiological measures. In the following we will summarize studies that used instrumental tasks. We will also describe modeling studies in detail, because this approach represents a powerful tool to provide a more fine-grained understanding of learning mechanisms and psychopathology (40, 41, 52, 53).

Based on the direct involvement of dopamine in both reinforcement learning and the neurobiology of schizophrenia, more systematic experimental examinations of alterations in reinforcement learning have been reported in the last decade.With regard to aberrant salience and the described ideas about aberrant learning,

#### **Box 1 Reinforcement learning models.**

A prediction error is defined as the difference between a delivered reward R and an expected value, here denoted as Q. t and a denote indices that refer to time and the value associated with a chosen action, respectively.

$$
\delta\_{\text{Cu},t} = \mathcal{R}\_t - \mathcal{Q}\_{\text{d},t} \tag{1}
$$

In model-free learning, this error signal can be used to update values:

$$
\Omega\_{\mathfrak{a},t+1} = \Omega\_{\mathfrak{a},t} + a\delta\_{\Omega\mathfrak{a},t} \tag{2}
$$

Here, α represent a learning rate, which weighs the influence of δQa,t on Qa,t <sup>+</sup> <sup>1</sup> with natural boundaries between 0 and 1. For examples of clinical studies using this algorithm, please compare Murray et al. (43) or Schlagenhauf et al. (77). Equation 2 refers to environments, in which each time point or trial t consists of one stage, e.g., one action, which results in feedback delivery. This can be extended to sequential decision tasks, where each trial consists of multiple numbers of stages and for example only the final stage is associated with feedback delivery. For an extension of the Eqs 1 and 2 for sequential decisions, please compare the work by Daw et al. (80) or Glascher et al. (79).

Still referring to model-free learning, we can define δ and the update equation differently, as for example in actor-critic models. The same error signal, generated by the critic, updates values of the critic and the actor:

$$
\delta\_{C, \mathfrak{s}, t} = \mathcal{R}\_t - \mathcal{L}\_{\mathfrak{s}, t} \tag{3}
$$

$$\mathcal{C}\_{\mathbf{s},t\neq\mathbf{t}} = \mathcal{C}\_{\mathbf{s},t} + a\_{\mathbf{s}}\boldsymbol{\aleph}\_{\mathbf{C}\mathbf{s},t} \tag{4}$$

Notably, the critic Eqs 5 and 6 neglects the specific action that was chosen in trial t. The actor learns specific action values via the same error signal δCs,t:

$$\mathcal{A}\_{\mathfrak{s},\mathfrak{a},t\upharpoonrightt} = \mathcal{A}\_{\mathfrak{s},\mathfrak{a},t} + \mathfrak{a}\_{\mathfrak{s}}\delta\_{\mathfrak{C},t} \tag{5}$$

This approach was applied in one clinical study (64).

So far, all presented models are examples for model-free learning. Subsequently, we present one example, which touches the ground of model-based learning. Depending on task structure, it is possible to implement certain aspects of the environment. For instance, in an environment with two choice options prediction errors may also be used to update values of unchosen actions ua; this can be done by an additional extension of Eq. 2:

$$
\Omega\_{\omega\mathfrak{u},t\vim} = \Omega\_{\omega\mathfrak{u},t} - a\mathfrak{d}\_{\Omega\mathfrak{u},t} \tag{6}
$$

Equation 8 represents a full double-update learner (77), while it is also possible to weigh the influence of the double-update by adding another free parameter:

$$
\Omega\_{\omega a, t+1} = \Omega\_{\omega a, t} - \kappa a \delta\_{\Omega a, t} \tag{7}
$$

Here, we use chosen prediction errors to update unchosen values. Based on the task design, it may be possible to use unchosen prediction errors (143). An elegant approach is to mix values learned by two different algorithms. This can be achieved by introducing a weighing parameter, for example as in Eq. 7. Please note that the contribution of additional free parameters (e.g., different learning rates for rewards and punishments in Eq. 2 or different learning rates for the critic and the actor in Eqs 4 and 5) needs to be quantified and that this is ultimately a question answered by model selection procedures [e.g., Ref. (115)].

For all the described models, learned values need to be transformed into choice probabilities to generate behavior. One commonly used approach is the softmax equation, which can be written as:

$$\mathfrak{p}(\mathfrak{a},t) = \frac{\mathsf{e}\exp(\mathfrak{k}\times\varOmega\_{\mathfrak{a},t})}{\sum\_{\mathfrak{a}'} \mathsf{e}\exp(\mathfrak{k}\times\varOmega\_{\mathfrak{a}',t})}\tag{\mathfrak{B}}$$

Here, all models refer to instrumental tasks. Most of the equations are applicable in similar forms to classical conditioning. For detailed reading, we refer to the scholarly book by Sutton and Barto (42).

so far only one experiment has been developed which specifically tests changes in adaptive (speeding up of reaction times for relevant cues) and aberrant salience (speeding up for irrelevant cues). This work by Roiser et al. (54) showed reduced adaptive salience in schizophrenia patients mostly medicated with SGAs but no general group difference in reaction time measures of aberrant salience. Within patients only, the individual degree of delusions was positively correlated with explicit measures of aberrant salience (54). Furthermore, using the same task, it was demonstrated that unmedicated people with an at-risk mental state for psychosis exhibit greater measures of aberrant salience, and this bias was correlated with their severity of delusion-like

symptoms (55). Imaging results from this multimodal study (55) are reported in the next section of this article. These findings point toward the expected direction but rather support a dimensional perspective on positive symptoms, in particular delusions, in a way that the presence of aberrant learning may fluctuate with changes in clinical symptoms. Nevertheless, the findings require further validation in unmedicated patients, since antipsychotic medication directly affects dopamine neurotransmission and primarily attenuates positive symptoms. Other evidence for aberrant learning primarily comes from classical conditioning during fMRI and is reported in the next section on fMRI studies.

Studies from Gold and colleagues have contributed an important body of work to the field. These studies provide evidence for the second hypothesis that postulates a deficit in value representation (11). With regard to hedonic experience, they demonstrated that stable-medicated, chronic patients do not differ in ratings on affective picture material nor do they differ in terms of speeded motor responses to repeat or to endure viewing of these pictures. It was observed that patients respond slightly faster to repeat viewing of neutral pictures (56). These results are in line with behavioral ratings in other studies using similar affective pictures (30, 57, 58).

Together, these findings indicate that schizophrenia patients are surprisingly unimpaired in short hedonic experiences. It is important to ask how these experiences are used to learn values that may guide behavior. Studies showed that delay discounting is altered in schizophrenia in such a way that immediate rewards are preferred over larger rewards in the future and with the degree of this difference being associated with working memory deficits (59–62). A study by Heerey et al. (63) found that in two separate tasks stable-medicated, chronic patients show intact reward sensitivity but impaired weighing of potential outcomes in a decision making task: only potential losses were weighed less by patients (63). Again, the ability to use potential outcomes to guide behavior was correlated with working memory function in patients.

Hypothetically, this deficit may be based on a shift from a goaldirected to a more inflexible learning system. Even in relatively simple tasks learning speed may increase based on additional use of a goal-directed system that accurately maps separate stimulus values to their potential outcome consequences, which may then be used for appropriate action selection. Models of reinforcement learning do not map perfectly on this distinction. Instead, several agents that update values based on prediction errors can be summarized as model-free controllers of learning and decision processes, because they neglect the contribution of additional environmental features (task structure) to the learning process (compare **Box 1**). Nevertheless, the kind of teaching signal used to update values can even be varied within the group of modelfree agents. Formally, one class includes model-free Q-learning algorithms, where each possible action becomes associated with a single value and these specific values are used to compute a prediction error. In contrast, a more rigid model-free system may learn values based on teaching signals that convey information about rewarded or punished states (e.g., a pair of stimuli) as, for example, formulated in actor-critic learning (42). This appears to be accompanied by slower learning compared to the more precise mapping of one Q-value to each stimulus associated with a certain value. Gold et al. (64) approached this question by applying

a task that requires learning from rewards in one condition and the avoidance of punishment in another condition in a sample of 47 stable-medicated, chronic patients. Patients were split into two subgroups with high and low levels of negative symptoms, respectively. Only patients with high levels of negative symptoms were shown to be selectively impaired in the reward-approach condition but demonstrated intact loss avoidance learning. This dissociation was also confirmed in a post-acquisition transfer test (64). A deficit in reward-based learning, but not in the avoidance of punishment, which was associated with negative symptoms, was also found in two other independent studies, both in patients treated with antipsychotic medication (65, 66). In the study by Gold et al. (64), an actor-critic model, a Q-learner, and a hybrid of these two models were fitted to the observed data and parameters were inferred using maximum-likelihood estimation. Based on model selection, data of the high-negative-symptom group was better explained by the actor-critic model, while healthy participants and the low-negative-symptom group of patients were better explained by the Q-learner. Such a deficit in value-based learning may also be closely connected to a deficit in cost computation of effortful behavior (67). The impact of this shift to a more rigid and rather imprecise learning system may depend on task demands and may in some rare cases be advantageous – if tasks require participants to behave rigid and at low levels of exploration (68). Again, it is important to note that most of the summarized studies were conducted in stable-medicated, rather chronic patients. The important question as to what extent these findings generalize remains to be examined.

The deficit of using outcomes to guide behavior may exacerbate when patients are confronted with situations where they are required to adapt their behavior flexibly. This can be examined in tasks like the Wisconsin Card Sorting Task or reversal learning. Indeed, a deficit in such tasks has been reported repeatedly in chronic, medicated states of schizophrenia (69–73). Studies in medication-free, first-episode patients indicate that such impairments are already present at the beginning of the disease and are stable over time (for at least 6 years), independent of general IQ effects (74, 75). Two recent studies demonstrate that the deficit in rapid behavioral adaptation is most likely due to an increased tendency to switch in schizophrenia patients (76, 77). A study by Schlagenhauf et al. (77) implemented detailed computational modeling of learning – ranging from standard Rescorla–Wagner-Models to Double-Update-Models (**Box 1**) and finally belief-based Hidden–Markov-Models (78) – to the data of 24 unmedicated patients. While the used Rescorla–Wagner-Models clearly provide a model-free account of reinforcement learning, the Double-Update- and the Hidden–Markov-Models can both be regarded as a model-based account of reinforcement learning because both incorporate important aspects of the experimental environment of the given task but in different ways: the Double-Update-Model simply integrates the dichotomy of the two choice options in the reversal learning task by updating each action value with the same prediction error but in different directions; the Hidden–Markov-Model approaches this differently by updating the probability of being in one of the two states and thereby actually building an internal model of the task's states (in the following, this is referred to as the participant's belief about the visited trial being informative about the state or not). Maximum-a-posteriori estimates of model parameters were inferred using random-effects Bayesian techniques complemented by model selection at the population and at the individual level. Random-effects parameters refer to individual parameter estimates per participant in contrast to fixed-effects parameters, which assume one set of parameters for a population. Note that random-effects fitting of models and model selection are crucially important to compare how models map to learning processes across groups and to compare parameters between groups. Also, individual model comparison is important because the meaning of underlying parameters remains unclear if the probability that a participant's data is given by the inferred parameters (the likelihood) is around chance (please also compare Section "Methodological Remarks"). Based on these methods, it was demonstrated that the beliefbased model explained the observed data best. This is in line with another study on reversal learning in healthy participants (78). Modeling results revealed increased switching in patients due to false beliefs with respect to feedback-conveyed information about the state of the task, which are based on reversals of reward contingencies (77). The study by Schlagenhauf et al. (77) was conducted in 24 unmedicated patients, of whom a substantial number was not able to apply the belief-based strategy. In these patients (*n* = 11), the reversal learning deficit was more pronounced. This was best explained by the actual presence of their positive symptoms, which is a remarkable contrast to several studies examining stable-medicated, chronic patients with attenuated positive symptoms. This subgroup of patients was additionally characterized by the model in terms of reduced reward sensitivity and showed a relatively better (although still poor) fit by the simple, model-free Rescorla–Wagner algorithm. Parameters of the models were used to generate regressors for the analysis of fMRI data and the results are discussed in the subsequent section.

There is convincing support that deficits in flexible behavioral adaptation and reversal learning, in particular, are important features of schizophrenia patients with an increased tendency to switch as a potential specific mechanism (76, 77). This is in line with an important assumption concerning the hypothesis of a deficit in value representation: an impaired functioning of the socalled rapid learning system that is assumed to rely on prefrontal and orbitofrontal brain structures deeply involved in cognitive functions such as working memory, which allows for flexible adaptation of decisions (47). This system is thought to interact with a more rigid learning system supposedly implemented in the basal ganglia pathways. As already mentioned above, these complementary learning systems may also be associated with the distinction of model-free and model-based controllers of learning, where the latter is implicated in using an internal model of the environment to optimize choice behavior (79, 80). It appears plausible that potential deficits in the model-based domain may be closely linked to well-established findings of impaired cognitive control with most evidence from measures of working memory and cognitive processing speed. Model-based learning relies on precise mapping of the environment and uses this map for forward planning of decisions. This process requires individuals to keep online values of multiple stimuli to allow for flexible decision making.

There is indeed evidence that working memory capacity limits the ability to learn multiple stimulus values to guide decisions and the degree of model-based behavior (81, 82), while, at the same time, possibly directing patients toward more inflexible aspects of learning, which themselves may be affected or spared in schizophrenia. There is additional evidence that patients learn reward contingencies, but that they may need more time depending on task demands (68, 83, 84). Interestingly, in a post-acquisition testphase,Waltz et al. (83) observed that medicated patients learned to avoid previously punished stimuli, while preference for the previously rewarded cues was weakened compared to controls. In a next step, Waltz et al. (85) studied stable-medicated, chronic patients with an established go-nogo learning task (86). During the training phase, patients showed an overall go-bias but no gradual adaptation to the more frequently rewarded stimuli, while the gradual adaptation to negative outcomes appeared to be intact (85). In line with deficits in reversal learning, rapid trial-to-trial adjustments were impaired in patients. This analysis was compared with predictions from a neurocomputational model of dopamine-induced basal ganglia-cortex interactions proposed by Frank et al. (87): high levels of presynaptic dopamine accompanied by alterations in D1-receptor density may specifically impair go-pathways which are proposed to facilitate reward-approach rather than punishment avoidance (47). This idea is also supported by recent optogenetic animal research (88). In accordance,it was also demonstrated that patients are less able to speed up responses to approach reward and show reduced exploration. Both effects were most pronounced in a subgroup of high-level negative symptoms (89).

In this section, we summarized results from studies on behavioral impairments during performance of instrumental tasks and only three studies, to date, have implemented reinforcement learning modeling to the observed behavioral data (64, 77, 89). Two of those studies demonstrated the ability to identify subgroups of the heterogeneous clinical entity referred to as schizophrenia. Further studies with similar experiments are needed across different disease states (e.g., first-episode) and medication states (in particular unmedicated patients as well as different medications to rule out the possibility that alterations in learning mechanisms are secondary to medication effects). This may be a potentially helpful route toward an identification of patient subgroups based on generative computational models of behavior and neural mechanisms. Recent methodological progress shows improved classification accuracy and allows for clustering within patients based on parameters of generative models of brain connectivity (90, 91), and this may also apply to generative models of behavior.

## **FUNCTIONAL IMAGING STUDIES OF REINFORCEMENT LEARNING IN SCHIZOPHRENIA**

This section will summarize studies that collected fMRI data during reinforcement learning to examine neural substrates of the behavioral alterations discussed in the previous section of this article. First, we summarize studies that examined classical conditioning. This process of associative learning has not been discussed in the previous section because classical conditioning paradigms do not usually require an instrumental response. Nevertheless, physiological responses reflect associative changes in stimulus contingencies, namely the unconditioned and the conditioned stimuli

(US and CS). Second, we report studies that investigated instrumental conditioning during fMRI. In both parts, we explicitly describe the application of reinforcement learning models, how parameters underlying these models were inferred, and how these measures were further applied to the imaging data.

## **CLASSICAL CONDITIONING**

Jensen et al. (92) studied aversive classical conditioning in 13 medicated patients. Their analysisfocused on the onset of CS associated with a neutral or an aversive event. In patients, they found elevated left ventral striatal activation to CS preceding neutral events compared to CS preceding aversive events (92). This aberrant attribution of salience was confirmed in skin conductance measures and post-learning self-reports. In a slightly different aversive conditioning paradigm neural responses to CS and US were studied in 20 medicated patients, and similar findings were demonstrated (93): attenuated activation to CS but intact responses to US were reported in the amygdala. Within patients, CS-related activation in the midbrain was correlated with delusion severity in a way that stronger CS-related responses in neutral trials predicted a higher degree of delusional symptoms (93). The authors additionally implemented a temporal-difference model to quantify neural correlates of prediction errors. Notably, the model's free parameter, the learning rate, was fixed for the entire sample and not fitted individually to behavioral or physiological responses [which were shown to vary, according reaction times and skin conductance e.g., Ref. (94, 95)]. Romaniuk and colleagues found no aversive prediction error correlate in the midbrain of schizophrenia patients as was observed in healthy controls. When modeling prediction errors for neutral events, they found a neural correlate of these prediction errors in patients' midbrain but not in controls (93).

With regard to appetitive classical conditioning with monetary reward, one study investigated neural activation to rewardassociated CS in 25 medicated patients. They reported that relatively lower ventral-striatal and ventro-medial-prefrontal activation depended on the degree of anhedonia (96), which is in line with previous findings using the MID task (17). Another study examined appetitive classical conditioning in thirsty participants (15 medicated patients) using water as reward. The analysis focused on reward delivery and found blunted ventral striatal activation in patients to be correlated with negative symptoms (97). Further, functional connectivity of the dopaminergic midbrain with the insula was reduced in patients.Another appetitive classical conditioning paradigm with monetary reward was used in a study by Diaconescu et al. (98) in 18 medicated patients. While patients and controls were similarly able to recall reward contingencies in explicit ratings, implicit measures (skin conductance) did not differ between reward CS and neutral CS in patients. The analysis of fMRI data also focused on CS and revealed that increased activation in striatal and prefrontal areas of healthy controls to reward CS was accompanied by stronger effective connectivity between VS and orbitofrontal cortex as assessed using structural equation modeling (98). Crucially, this pattern was reversed in patients for the neutral CS. This is an important finding, as it has long been described that neural correlates of learning spread over nodes of a network and thereby drive changes in plasticity. A disturbance of such a mechanism was also proposed to be at the heart of the pathophysiology of schizophrenia (99–101). We will return to this issue in the final section.

#### **INSTRUMENTAL LEARNING**

We now proceed with further studies that investigated neural correlates during instrumental learning. In line with evidence for aberrant learning from classical conditioning, a recent multimodal imaging study using the instrumental "salience attribution task" [(55); for behavioral results see previous section] found that ventral striatal activation to irrelevant stimulus features were positively correlated to delusion-like symptom severity in 18 unmedicated people with an at-risk mental statefor psychosis (55). Furthermore, hippocampal responses to irrelevant features were differently correlated with dopamine synthesis capacity in VS revealing a positive relationship in controls and a negative relationship in people with an at-risk mental state.

One exemplary study that assessed the association between impaired reinforcement learning and brain activation in dopaminergic target brain areas of first-episode schizophrenia patients (*n* = 13, 8 medicated) used an instrumental learning task with two choice options: one signaled a potential monetaryfeedback and the other a potential neutral feedback (43). In contrast to several other studies (see previous section), the groups did not differ in terms of acquisition of reward contingencies, which may be due to the rather small sample size of this pioneer study. In line with another study (59), patients responded faster on neutral trials in the study by Murray et al. (43). A Q-learner was fitted to the observed data based on maximum-likelihood estimates of parameters. Both groups did not differ in terms of model parameters. To generate regressors for fMRI data analysis, one set of parameters was fitted for the entire sample (fixed-effects). Model-derived prediction errors were used as a parametric modulator of feedback events. Prediction error correlates in bilateral midbrain, right VS, hippocampus, insula, and cingulate cortex were significantly stronger in controls than in patients. In patients,midbrain correlates of prediction errors appeared slightly augmented in neutral trials (43). A more complicated "allergy prediction" task design enabled Corlett et al. (102) to investigate different stages of learning in 14 patients, most of whom were medicated. For event-related fMRI analysis, an event was defined to start at the beginning of each stimulus presentation and to end after outcome delivery lasting a total time of 4 s. Compared to controls, patients did not activate the left caudate during the training stage, which was followed by revaluation of stimuli pairs that were either ambiguous or well learned pairs of cues during training. The comparison of these pairs revealed a failure to activate substantia nigra and right PFC. In the last phase, expectations about the outcome based on the trained stimulus pairs were violated. Here, predictable events elicited an augmented response in right PFC in patients versus controls, while an attenuated response was found for unexpected events (102). This lack of differentiation between expected- and unexpectedness events correlated with the level of unusual thought content. Notably, the analysis strategy chosen in this design makes it hard to interpret the findings in terms of prediction error or expected value signals because the whole trial period was modeled in the singlesubject of the fMRI data. Similar results were reported in another

study that investigated 20 medicated patients while performing a guessing–gambling paradigm at different levels of uncertainty but analyzed expectation-related and reward-related activation separately (103). Expectation-related brain activation at time of motor responses revealed increased activation with lower predictability in a fronto-parietal network, and this effect was diminished in dorsolateral PFC and anterior cingulate cortex of schizophrenia patients. Reward-associated activation was analyzed in relation to levels of predictability (assumed to mirror prediction error related brain activation), and patients showed reduced activation in putamen, dorsal cingulate, and superior frontal cortex (104). One study assessed probabilistic category learning ("weather prediction task") in medicated schizophrenia patients (*n* = 40) during fMRI. Albeit impaired performance in all patients, a small number of patients were able to apply a similar strategy to the task as controls did (105). When comparing fMRI data of these matched groups (*n* = 8 patients) during the presentation of stimulus combinations, patients displayed reduced activation in striatum and dorsolateral PFC. Patients exhibited stronger activation in a more rostral region of dlPFC and parietal cortex. Results from this task are hard to compare with instrumental reinforcement learning tasks due to the experimental design that primarily tests classification learning at different levels of difficulty.

In another study on instrumental learning, Gradin et al. (106) examined 15 medicated patients. Temporal-difference modeling was applied to the task that delivered water as reward. Randomeffects parameters were initially estimated with maximumlikelihood, and the obtained parameters were subsequently used as empirical priors to regularize the possible range parameters to avoid extreme values of parameter estimates [also compare: Ref. (53, 106)]. Although patients differed in the amount of delivered water, no difference on model parameters was observed. To generate regressors for fMRI analysis, a single set of parameters was fitted for the entire sample (fixed-effects). Modelderived prediction errors were analyzed as parametric modulators of reward delivery, and model-derived values were included as modulators of expectation-related activation at the trial onset. Compared to controls, no correlation with prediction errors was observed in striatum, thalamus, amygdala-hippocampal complex, and insula of medicated schizophrenia patients. A trendwise reduction in midbrain correlated with positive symptoms in patients. Patients also displayed reduced coding of valuerelated activation in the amygdala-hippocampal complex and this, again, was correlated with positive symptoms. Importantly, this study also included another psychiatric patient group, medicated depressed patients, and this group also exhibited blunted neural correlates of expected-reward values and prediction errors in slightly different regions. The strength of this reduction was correlated with anhedonia severity in dopaminergic core areas. In combination with detailed computational modeling, Schlagenhauf et al. (77) studied reversal learning (compare previous section) in 24 unmedicated patients. Analysis of fMRI focused on the time of reward delivery and included different model-derived modulations of this onset. The authors found reduced ventral striatal coding of model-derived reward prediction errors in patients. This finding remained trend-wise significant when restricting the group comparison to patients who had insight into the underlying

task structure as defined by their beliefs about the states of the task based on a Hidden–Markov-Model (*n* = 12). A second fMRI analysis based on the latter model was applied to define subjective informative punishment trials,i.e.,when participants believed that a change in reward contingencies had appeared. Both patients with good and poor task insight showed reduced ventral striatal activation during these trials (77). Reduced ventral striatal activation was also reported in another recent fMRI study on reversal learning in 28 medicated, chronic schizophrenia patients (76). In the study by Schlagenhauf et al. (77), patients with good task insight displayed relatively stronger activation of ventro-lateral and dorso-medial PFC than patients with poor insight. Well performing patients were not distinguishablefrom controls with respect to activation in these prefrontal regions. This result may reflect compensatory PFC processes in schizophrenia patients similar to that which has been described for the neural correlates of working memory deficits (107, 108).

In summary, several studies revealed reduced activation of brain areas typically encoding errors of reward prediction, most prominently the VS. This was reported consistently across classical and instrumental conditioning tasks, despite the fact that most of these studies differ enormously with regard to experimental designs and analysis strategies. Prediction errors arise when a reward is delivered and are typically thought to train expected values of stimuli or associated actions (42). Therefore, functional neuroimaging studies that studied learning during scanning have so far helped to elucidate the underlying dynamics of previous findings derived from studies using the MID or similar tasks. That is, neuronal teaching signals are not coded in ventral striatal activation of medicated and unmedicated patients to a similar extent as in controls. Only five imaging studies have applied reinforcement learning models to describe this process on a trial-by-trial level and these vary considerably in terms of the implemented models,inference of model parameters and the application of model-derived measures to the imaging data. We will further comment on these issues in the subsequent section. These studies comprised 78 medicated patients and 24 unmedicated patients. Studies in unmedicated patients are still rare. Nevertheless, the finding of reduced prediction error coding in dopaminergic core areas may indeed build a common ground for impaired learning of stimulus or decision values. In addition, such impaired coding might be closely related to the elevated levels of presynaptic dopamine synthesis capacity in schizophrenia reported in meta-analyses of PET studies (4, 5, 109). An important question remains how this stable marker of the dopamine system, probably reflecting tonic or rather stable aspects of dopaminergic neurotransmission (3), relates to eventrelated changes during learning. Studies approaching this question are discussed in Section "Functional Imaging Studies of Reinforcement Learning with Additional Neurochemical Measures or Pharmacological Challenges of the Dopamine System" of this article. Furthermore, it has been proposed that a hyperdopaminergic state in schizophrenia may result in imprecise and inefficient cortical information processing as a potential mechanism for cognitive impairments observed in patients as well as their first-degree relatives and in people at-risk mental states (9, 110, 111). This idea is compatible with the proposal of a deficit in prefrontal value representation shown to be related to negative symptoms. However, exact cognitive and affective correlates of such deficits remain to be explored. We will return to this in the final section.

The emerging picture is less clear with regard to evidence provided in favor of the aberrant salience hypothesis, in particular regarding the extent to which reduced neural correlates of prediction errors are linked to processes of aberrant salience attribution. Notably, the idea of aberrant salience may also account for reduced value-related anticipatory dopaminergic signals, in patients who exhibit high levels of positive and negative symptoms (for example). In this case, a lack of activation to cues associated with monetary as well as, probably, social reward may reflect reduced motivational or incentive salience in terms of apathy or other dimensions of negative symptoms, which may be a result of aberrant salience attribution. However, this requires more systematic studies along symptom dimensions. Evidence for neural correlates of aberrant learning was demonstrated in fMRI studies on classical conditioning that showed elevated striatal activation to cues indicating the delivery of a neutral event (92, 93, 98) and in one specific instrumental task design, the "salience attribution task" (55, 112). Studies using this specifically designed task point toward a relationship with positive symptoms, particularly delusions. Consequently, symptom and medication states of included patients may be crucially important. Indeed, a study on reversal learning in unmedicated patients with more pronounced positive symptoms showed that a subgroup of patients was not able to infer the task structure and this was best explained by individual levels of positive symptoms (77). Therefore, it is important to consider the amount of variance in symptom ratings and different medication states to better understand variability related to aberrant aspects of neural learning signals. Furthermore, when reviewing clinical data of several studies summarized in this article, it is compelling that even in medicated patients there is considerable variability in the extent of positive symptoms across studies varying from high levels to nearly no positive symptoms. Future studies are needed to address the question whether blunted learning signals indeed reflect aberrant salience attribution – and if this is a schizophrenia specific feature or a dimension of positive psychotic symptoms – which may then consequently also emerge in other psychiatric diseases and to some extent even in the at-risk healthy population or healthy people with some degree of psychotic experience.

## **METHODOLOGICAL REMARKS**

The combination of model-derived learning signals with functional brain measures is very promising. This mechanistically informed quantification of signals reflecting learning processes provides a more fine-grained insight into neural trial-by-trial correlates of learning mechanisms and disease-specific alterations as compared to standard event-related fMRI analyses which rather rely on event definitions such as correct responses or experimenter-defined changes in reward contingencies. In fact, the latter may not always reflect the way study participants solve these tasks. On the other hand, a small number of healthy volunteers, in most studies, exhibit behavior that cannot be described better than chance by any reinforcement learning model. This may indicate the need to extend from standard reinforcement learning models to other types of models, for example Bayesian learners (94, 113, 114). Such non-fitters should be reported more clearly, in particular in clinical between-group studies, because

this may crucially impair the between-group analysis of model parameters and comparisons of neural correlates based on modelderived measures between groups: in fact, underlying parameters of non-fitters are meaningless in terms of the mechanism that is described by the model [compare Ref. (77)]. Although studies which actually apply reinforcement learning modeling are the minority of those reported in this review article (seven studies, for an overview see **Table 1**), there is considerable variability on how these few studies inferred the models' parameters (some did and others did not fit parameters) and how (or if any) model selection was applied.

Further, the generation of trial-by-trial model-derived timeseries for fMRI data analysis is sometimes performed based on random-effects parameters (individual parameters for each subject) or based on one set of parameters (fixed-effects). One group recommends the latter approach for studies in healthy volunteers by arguing for more robust correlations of BOLD signal with model-derived regressors (53). On the other hand, this appears questionable for group studies in which group differences in parameters may be causally linked to the disease status. We have the impression that model comparison techniques are of key importance (115). Even in the simple case that no alternative models are fitted, it may be informative to include a report of model fit based on the likelihood that the observed data is given by the parameters. To our mind, a situation where the individual model fit (expressed via the likelihood of the data given by the parameters) does not differ between groups exemplifies a desirable case: even if parameters differ between groups in this case, modelderived regressors are readily applicable to fMRI data because they do not differ in terms of the likelihood that the modeled strategy captures important aspects of the observed raw responses. Based on the sparsely available papers on these issues, the application of fixed-effects parameters to fMRI data rather appears as a workaround based on the observation that noisy parameters based on maximum-likelihood estimates potentially add further noise when fitting a hemodynamic model with model-derived time-series as parametric modulators to the imaging data [compare Ref. (53)]. In the case of clinical between-group studies, the use of fixed-effects parameters results in a situation where the observed behavior is relatively well explained by those parameters. Consequently, differences in terms of model parameters will then be expressed via the correlation between the regressor and the signal. This can be minimized by using parameters that closely match the observed individual's behavior to generate regressors. Unfortunately, no systematic studies of these questions are available involving either healthy volunteers only, or comparisons between psychiatric patients and healthy controls. Consequently, it appears to be desirable to develop methodological guidelines for these techniques, as it was published for other modeling approaches, for example for dynamic causal modeling for fMRI (116).

## **FUNCTIONAL IMAGING STUDIES OF REINFORCEMENT LEARNING WITH ADDITIONAL NEUROCHEMICAL MEASURES OR PHARMACOLOGICAL CHALLENGES OF THE DOPAMINE SYSTEM**

In this last section, we describe research that pharmacologically manipulated the dopamine system during reinforcement learning or acquired an additional molecular measurement (PET) of dopamine. There are a substantial number of groups researching these questions in healthy volunteers, and this section does not aim to present a complete picture of all such studies. We will instead refer to studies that are particularly important for a better understanding of the above reviewed studies in patients.

A highly influential study was conducted by Pessiglione et al. (117). An instrumental learning task with a reward-approach (win or no win) and a punishment–avoidance (loss or no loss) condition was used (117). A similar design was recently applied in a behavioral study of medicated schizophrenia patients (64). Pessiglione et al. (117) demonstrated that dopamine medication, l-DOPA, and haloperidol, have opposing effects on behavior and neural correlates of model-derived prediction errors, and that these effects are selective for the reward-approach condition: l-DOPA administration enhanced reward-approaching behavior and associated ventral striatal reward prediction errors whereas haloperidol reduced such effects (117). The same direction of medication effects was reported in a study using aversive Pavlovian conditioning under amphetamine, haloperidol, or placebo (118). In addition, ventral striatal reward anticipation as in the MID task appears to be conveyed by reward-induced dopamine release (27) and can be blunted by massive dopamine release, based on dose-dependent effects of psychostimulants (26, 28). In line with this, a recent study applied the same reward-approach task as in Murray et al. (43) and found that methamphetamine blunts both reward prediction errors in VS and expected value representation in ventro-medial PFC (119). The strength of the disruption of value representation in ventro-medial PFC was correlated with amphetamineinduced psychotic symptoms. A third condition, amisulpridepretreatment, did not affect amphetamine-induced blunting of learning signals. It is important to note that the reducing effects of haloperidol on striatal reward prediction error encoding can explain reduced prediction error related activation in medicated schizophrenia patients, whereas the blunting effects of amphetamine may potentially mirror a subcortical hyperdopaminergic state, as was demonstrated in unmedicated schizophrenia patients (2, 3). Therefore studies in unmedicated patients are crucially important to remove this confound. FGAs and SGAs strongly differ in their dopamine receptor affinity, and, based on two MID studies, it was shown that they also affect reward anticipation differently. These results point toward the idea that SGAs may help to remediate reward-related anticipatory brain activation (20–22) which nevertheless requires random assignment in a clinical-trialtype design. Such studies have not yet been conducted with learning tasks. In unmedicated patients, a reduction of ventral-striatal prediction error coding was recently demonstrated during reversal learning (77). Elevated presynaptic dopamine levels may account for this reduced activation, similar to that observed for Parkinson patients on l-DOPA medication, affecting the VS (early in the illness less degenerated) in an overdosing manner (120). Here, a long-lasting increase of presynaptic dopamine function may "drown" value-related and error-related phasic learning signals. Multimodal imaging studies combining fMRI with PET radiotracers that assess presynaptic dopamine function can link individual differences in neurochemical parameters with functional activation. For example, PET with FDOPA may be an important target

for application in multimodal imaging studies, since this measure has been demonstrated in meta-analyses to best characterize the subcortical hyperdopaminergic state of patients [for metaanalyses see: Ref. (4, 5)]. Supporting this idea, another study demonstrated that ventral-striatal prediction errors are indeed negatively correlated with dopamine synthesis capacity in healthy controls (121). This negative correlation suggests that long-lasting increases in presynaptic dopamine function, as observed in schizophrenia patients, may reduce phasic learning signals, hypothetically via presynaptic D2-autoreceptors which regulate presynaptic activity of DOPA-decarboxylase activity to ensure homeostasis within the dopaminergic system (46, 121). Animal studies (122, 123) and other functional human imaging studies (36, 124, 125) also support the idea of this interplay of differential aspects of dopamine neurotransmission. In line with this, it has also been shown that behavioral effects of a dopamine-enhancing drug during reversal learning indeed depend on baseline levels of dopamine synthesis capacity (126): Participants with lower dopamine synthesis capacity benefit behaviorallyfrom a dopamine agonist,while the same drug dose seemed to be disadvantageous for participants with rather high levels of dopamine synthesis capacity. Therefore, dopamine effects in learning and cognition appear to be a fine-tuned and optimized non-linear system where rather low and rather high levels result in inefficient neural processing (127, 128). This view is also supported by one of the few clinical multimodal imaging studies using FDOPA PET in combination with a working memory task during fMRI in controls and in people with an at-risk mental state for psychosis (129). At the same working memory load, they found a positive linear relationship of dopamine synthesis capacity and working memory related activation in dorsolateral PFC of healthy controls, while this relationship was negative in people with an at-risk mental state, indicating that potentially "too" high levels of dopamine synthesis may promote lower dorsolateral PFC activation during the same working memory load at which both groups coped with comparable performance. This observation can be reconciled with the observation of prefrontal efficiency during working memory when examining cognitive performance and dorsolateral PFC activation (130, 131): different dorsolateral PFC activation may primarily reflect different performance. Patients are assumed to reach maximum limits of dorsolateral PFC activation earlier reflecting a general impairment in this cognitive domain [see also: Ref. (108)]. A step further, there is also evidence that a reduction of working-memory-dependent effective connectivity from dorsolateral PFC to parietal cortex may be the potential mechanism underlying this inefficiency (108). Connectivity may indeed be an important target and it has also been demonstrated that functional connectivity during aversive conditioning is shifted differently by a dopamine agonist versus a dopamine antagonist (132).

## **CONCLUSION, REMARKS, OUTLOOK**

In this review article we summarized studies that provide evidence for behavioral and neural correlates of impaired reinforcement learning in schizophrenia. Two main hypotheses guided this review: (1) Aberrant prediction errors drive learning of otherwise irrelevant stimuli and actions in schizophrenia, and that there is a potentially close link between this mechanism and the emergence of positive psychotic symptoms, in particular delusions. (2) A deficit of expected value representation may characterize patients suffering from schizophrenia, and this may fundamentally contribute to the formation of negative symptoms.

There is evidence for aberrant learning with most specific findings derived from the salience attribution task (54, 55). Although there is still limited evidence and future studies are needed for clarification, it seems conceivable that aberrant learning is involved in the formation of delusions and can therefore be observed in patients with prominent positive symptoms. Whether this sensitization to otherwise neutral stimuli is indeed dopamine mediated and actually blunts learning signals elicited by regularly salient cues remains to be further explored (9).

Our summary of fMRI studies during reinforcement learning clearly demonstrates that a reduction of these learning signals, namely blunted coding of ventral-striatal prediction errors, was consistently observed across studies. This deficit may actually be involved in aberrant learning as well as in a failure of value representation depending on fluctuating symptom states. During acute psychotic clinical states this may provide a conduit for aberrant learning, while the persistence of a noisy learning signals may provide the ground for a failure of building value expectations, ultimately contributing to the development and progress of detrimental negative symptoms. A large body of evidence from behavioral studies supports the hypothesis of a deficit in value representation and that this impairment is pronounced in patients with high levels of negative symptoms (11, 64). Nevertheless, antipsychotic medication was shown to contribute to the severity of negative symptoms based on the degree of striatal D2-receptor blockade (133) and may therefore also exacerbate impairments in value representation.

The psychosis spectrum has been characterized by imprecise and inefficient cortical information processing as a potential mechanism for cognitive impairments observed in patients and their first-degree relatives as well as in at-risk mental states. As a potential mechanism behind this, a disrupted cortico-cortical synaptic plasticity was suggested by a comprehensive biological hypothesis of schizophrenia, the "dysconnectivity" hypothesis (100, 101). This hypothesis focuses on aberrant experience-driven control of synaptic plasticity via *N*-methyl-d-aspartate receptors (NMDAR). Abnormal modulation of NMDAR-induced plasticity by neurotransmitter systems such as dopamine, acetylcholine, or serotonin are at the heart of this idea. In the present article,we have described a close link between reinforcement learning, symptom dimensions of schizophrenia and dopamine,which acts as a neuromodulator of NMDAR-function: Animal research demonstrated that D1-receptor agonists and D2-receptor antagonists facilitate NMDAR-dependent long-term plasticity while D2-receptor agonist suppress it (134, 135). Earlier in the manuscript, we have discussed the role of these receptors during reward and punishment as well as go-nogo learning (87, 88). Further, these receptors are targets of current antipsychotic treatment strategies. Stephan and colleagues conclude ". . . it is not plasticity *per se* that is abnormal but its modulation during reinforcement and perceptual learning." (100) These modulatory influences of NMDAR-function are thought to contribute to cortical representations of environmental states (136) and the consistently described reduction of ventral-striatal prediction errors could be crucially involved in a deficient shaping of such cortical representations (137, 138). Here, it is important to underline that aberrant neuromodulation can indeed be formulated via computational models of learning-induced plasticity.

So far, there is converging evidence that dysconnectivity may indeed account for the repeatedly described prefrontal inefficiency observed in schizophrenia during the performance of cognitive tasks: Using models of effective connectivity for fMRI, reduced working-memory-dependent prefrontal-parietal connectivity was reported, initially in medicated patients (108) and subsequently replicated in medication-naïve first-episode patients and in people with an at-risk mental state (139). Based on parameters of these models, a clustering analysis was able to identify three mechanistically informed subgroups of patients (91). These subgroups were found to be biologically distinct in terms of connectivity profiles and mapped on different levels of negative symptom severity (91). This observation also appears to be in accordance with the proposal of a deficit in value representation, which was demonstrated to be pronounced at high levels of negative symptoms (64). Therefore, it appears desirable to study the effect of neural learning signals at various levels in neural networks. It is important to note that studying the interaction of model-free learning signals and model-based neural representations of cognitive processes on the level of neural networks in a computational framework clearly has the potential to move beyond the evidence provided by standard cognitive tasks, such as working memory, even if some of the identified deficits overlap. The contribution of such an approach can be to gain more mechanistic information when studying these processes by applying detailed computational modeling to behavioral and neurobiological data. The focus of this idea is that different types of computational processes described in terms of different models may help us to improve our understanding of how patients actually solve certain tasks beyond the observation of being impaired or not. This may offer a unique source for mechanistically informed subtyping based on how patient subgroups deal differently with challenging tasks and in particular how these abilities are implemented in neural networks. Such subgroups require clinical validation in terms of longitudinal predictions (e.g., treatment responses). Promising future research in this field may strongly benefit from an integration of different modeling techniques for reinforcement as well as perceptual learning and brain connectivity [e.g., dynamic causal modeling; (140)]. It has been demonstrated in healthy volunteers that such an experimental approach is feasible (94, 141, 142) and therefore presents a highly promising venue for schizophrenia research. Finally, this may result in a dissection of the heterogeneous clinical entity of schizophrenia into biologically informed subgroups, thereby providing a framework for a better understanding of cognitive deficits, where a deficit of learning expectations about sensory inputs and future actions may constitute a potential key mechanism of the disorder.

#### **ACKNOWLEDGMENTS**

This study was supported by funding from the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG) to Florian Schlagenhauf SCHL1969/1-1.

## **REFERENCES**


in schizophrenia. *Arch Gen Psychiatry* (2012) **69**:1195–204. doi:10.1001/ archgenpsychiatry.2012.847


striatal-cortical dysfunction. *Biol Psychiatry* (2007) **62**:756–64. doi:10.1016/j. biopsych.2006.09.042


with a probabilistic learning task. *Learn Mem* (2008) **15**:728–32. doi:10.1101/ lm.1106408


positron emission tomography: implication of dopaminergic tone. *J Neurosci* (2011) **31**:7886–90. doi:10.1523/JNEUROSCI.6024-10.2011


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 10 September 2013; paper pending published: 01 October 2013; accepted: 07 December 2013; published online: 23 December 2013.*

*Citation: Deserno L, Boehme R, Heinz A and Schlagenhauf F (2013) Reinforcement learning and dopamine in schizophrenia: dimensions of symptoms or specific features of a disease group? Front. Psychiatry 4:172. doi: 10.3389/fpsyt.2013.00172*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Deserno, Boehme, Heinz and Schlagenhauf. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Computational neuropsychiatry – schizophrenia as a cognitive brain network disorder

**Maria R. Dauvermann<sup>1</sup>\*, Heather C.Whalley <sup>1</sup> , André Schmidt 2,3, Graham L. Lee<sup>4</sup> , Liana Romaniuk <sup>1</sup> , Neil Roberts <sup>5</sup> , Eve C. Johnstone<sup>1</sup> , Stephen M. Lawrie<sup>1</sup> and ThomasW. J. Moorhead<sup>1</sup>**

<sup>1</sup> Division of Psychiatry, Royal Edinburgh Hospital, University of Edinburgh, Edinburgh, UK

<sup>2</sup> Department of Psychiatry, University of Basel, Basel, Switzerland

<sup>3</sup> Medical Image Analysis Center, University Hospital Basel, Basel, Switzerland

<sup>4</sup> McGovern Institute for Brain Research, Massachusetts Institute of Technology, Cambridge, MA, USA

<sup>5</sup> Clinical Research Imaging Centre, QMRI, University of Edinburgh, Edinburgh, UK

#### **Edited by:**

Stefan Borgwardt, University of Basel, Switzerland

#### **Reviewed by:**

Philip R. Corlett, Yale School of Medicine, USA Lorenz Deserno, Max-Planck-Institute for Human Cognitive and Brain Sciences, Germany

#### **\*Correspondence:**

Maria R. Dauvermann, Division of Psychiatry, Royal Edinburgh Hospital, University of Edinburgh, Kennedy Tower, Morningside Park, Edinburgh EH10 5HF, UK e-mail: m.r.dauvermann@sms.ed.ac. uk,maria.dauvermann@childrens. harvard.edu

Computational modeling of functional brain networks in fMRI data has advanced the understanding of higher cognitive function. It is hypothesized that functional networks mediating higher cognitive processes are disrupted in people with schizophrenia. In this article, we review studies that applied measures of functional and effective connectivity to fMRI data during cognitive tasks, in particular working memory fMRI studies.We provide a conceptual summary of the main findings in fMRI data and their relationship with neurotransmitter systems, which are known to be altered in individuals with schizophrenia. We consider possible developments in computational neuropsychiatry, which are likely to further our understanding of how key functional networks are altered in schizophrenia.

**Keywords: computational neuropsychiatry, schizophrenia, fMRI, dynamic causal modeling, cognition, neurotransmitter, dopamine, glutamate**

## **INTRODUCTION**

Schizophrenia is a severe psychiatric disorder, which is initially manifested through positive symptoms including delusions, hallucinations, and disorganized thoughts. As the illness progresses negative symptoms such as avolition, alogia, and apathy may occur. Prior to diagnosis of illness, cognitive deficits can occur and illness progression can also be associated with cognitive deficits (1, 2). It is widely established that such cognitive deficits are considered a core symptom of schizophrenia (3) and are associated with reductions in working memory performance. Working memory deficits are one of the main neurocognitive impairments found in subjects

**Abbreviations:** AC/ACC, anterior cingulate/anterior cingulate cortex; ARMS, atrisk mental state; BMS, Bayesian model selection; BOLD response, blood-oxygenlevel dependent response; D<sup>1</sup> receptor, D<sup>1</sup> subtype of the dopamine receptor; D<sup>2</sup> receptor, D<sup>2</sup> subtype of the dopamine receptor; D2/3 receptors, D2/3 subtype of the dopamine receptor; DCM, Dynamic Causal Modeling; DMN, default-mode network; DLPFC, dorsolateral prefrontal cortex; EC, effective connectivity; EST, subjects with EST; FC,functional connectivity; FEP, subjects with first episode psychosis; FES, subjects with first episode schizophrenia; GABA, γ-aminobutyric acid; GBC, global-based connectivity; HR, subjects at high risk of schizophrenia; HR+, subjects at high familial risk of schizophrenia with transient psychotic symptoms; HR−, subjects at high familial risk of schizophrenia without transient psychotic symptoms; HRill, subjects at high familial risk of schizophrenia who subsequent to scanning developed schizophrenia; HSCT, Hayling sentence completion task IFG inferior frontal gyrus; MFG, middle frontal gyrus; MRS, magnetic resonance spectroscopy; NMDA, *N*-Methyl-d-aspartate acid; PC, parietal cortex; PET, positron emission tomography; rCBF, regional cerebral blood flow; SPECT, single-photon emission computed tomography; SPL, superior parietal lobe; STG, superior temporal gyrus.

with first episode schizophrenia (FES) (4, 5) and in people with established schizophrenia (EST) (6). Similar deficits also occur in individuals at high risk of schizophrenia [HR; Ref. (2)]. Furthermore, recent evidence has been presented, which indicates a relationship between severity of working memory deficits and the severity of negative symptoms (7). The severity of working memory deficits that is evident at the first episode of schizophrenia can predict the quality of life at the established stage of the illness (8, 9).

Two major neurotransmitter circuits have been implicated in clinical and cognitive symptoms in subjects with schizophrenia: these are the dopamine and glutamate neurotransmitter circuits. Evidence has been presentedfor separate alterations/disruptions of dopamine and glutamate as well as an interactive role between both neurotransmitters<sup>1</sup> . The two main neurobiological hypotheses in schizophrenia are based on the theories of altered dopaminergic transmission ("dopamine hypothesis of schizophrenia") and altered glutamatergic transmission ("glutamate hypothesis of schizophrenia"). It is thought that both dopamine and glutamate modulate the dorsolateral prefrontal cortex (DLPFC) and in schizophrenia alter the performance in cognitive processes such as in working memory (10–13). Such work supports the notion of schizophrenia as a brain disorder. FMRI and positron emission tomography (PET) findings of altered functional activation and functional connectivity (FC) during working memory have been

<sup>1</sup> It is noted that other neurotransmitter circuits are interacting with dopaminergic and/or glutamatergic circuits such as serotonin and GABA (24, 29, 149).

reported in people with schizophrenia when they are compared to healthy controls (14, 15). Furthermore, PET studies have presented evidence for indirect markers of altered dopamine transmission, which was correlated with working memory performance (2, 16). Alterations of indirect measures of glutamate concentrations have been reported by proton magnetic resonance spectroscopy (MRS) studies (17).

One subfield within the emerging field of computational neuropsychiatry is based on modeling fMRI networks and the evidence of (i) altered dopaminergic and/or glutamatergic transmission in (ii) cognitive function (i.e., working memory) in people with schizophrenia. Therefore, the objectives are the investigation of impaired cognitive function mediated by large-scale networks in combination with underlying neurobiological circuits such as dopamine and glutamate. Researchers in computational neuropsychiatry examine and model altered cognitive brain function in terms of functionally integrated regions [i.e., effective connectivity (EC)] (18), which may be mediated by genetic factors and neurotransmitter circuits (19–21). Mechanistic responses can be inferredfrom the computational modeling of cognitive brainfunction where the localized brain function is monitored through the BOLD response (22). This modeling approach allows computational neuropsychiatry to further our understanding of the neurobiological processes, which underlie altered cognitive brain function in individuals with schizophrenia. Thus, advancing our knowledge of schizophrenia as a cognitive brain network disorder.

In this review, we summarize fMRI findings in verbal/numeric working memory<sup>2</sup> in context of (i) the understanding of schizophrenia as a cognitive brain disorder (from clinical and cognitive neurosciences) and (ii) the understanding of schizophrenia as a cognitive brain network disorder (from computational neuropsychiatry). We discuss these sets of findings in context of the dopamine and the glutamate hypotheses of schizophrenia. We consider two key research questions for the discussion of each set of findings:


The review is structured as followed: first, the dopamine and glutamate hypotheses of schizophrenia are summarized (Section Schizophrenia as a Brain Disorder). Second, exemplary findings of verbal/numeric working memory deficits from fMRI studies in subjects with schizophrenia are summarized. These findings are discussed in context of the dopamine hypothesis and the glutamate hypothesis of schizophrenia (Section Schizophrenia as a Cognitive Brain Disorder). Third, we present a brief introduction to computational neuropsychiatry.We provide examples from computational neuropsychiatry and the application to the investigation of cognitive brain large-scale networks in people with schizophrenia<sup>3</sup> . Finally,we consider current methodological limitations of the methods (Section From Computational Neuropsychiatry Towards Schizophrenia as a Cognitive Brain Network Disorder). We outline potential future influences of computational advances in schizophrenia that may shape our understanding of schizophrenia with the aim of developing more effective treatments for the disorder (Section Understanding of Schizophrenia).

## **SCHIZOPHRENIA AS A BRAIN DISORDER**

Neurobiological research into alterations of dopaminergic and/or glutamatergic neurotransmission has paved the way for the understanding of schizophrenia as a disorder of the brain. The dopamine hypothesis posits that dopamine function is altered in schizophrenia and that this dysfunction may be the pathophysiological pathway leading to clinical and cognitive symptoms (23, 24). The glutamate hypothesis proposes that the altered dopaminergic dysfunction may be secondary to aberrant glutamatergic dysregulation, which may contribute to clinical and cognitive symptoms in schizophrenia (25–27).

## **DOPAMINE HYPOTHESIS OF SCHIZOPHRENIA**

The origin of the dopamine hypothesis of schizophrenia is based on the discovery of antipsychotic drugs by Delay et al. (28) in 1952. Carlsson and Lindqvit reported the first findings of an effect of antipsychotic drugs on the metabolism of dopamine (29). The dopamine hypothesis posits that alterations of dopaminergic receptors may underlie the clinical symptoms of schizophrenia (30). Over last three decades, the dopamine hypothesis of schizophrenia has undergone reformulations in light of newly available preclinical and clinical findings. Here, we consider the three main hypotheses: (i) the "dopamine receptor hypothesis," (ii) the "modified dopamine hypothesis of schizophrenia," and (iii) the "dopamine hypothesis: version III."

The dopamine receptor hypothesis goes back to studies reporting antipsychotics affecting the affinity of dopamine receptors (31–33). Further evidence for the hypothesis was presented with increased synaptic monoamine levels during the induction of psychotic symptoms (34). The focus of this hypothesis rests on the excess of dopamine receptors. Thus, the clinical treatment is aimed at blocking the dopamine D<sup>2</sup> subtype of the dopamine receptors (35).

The modified dopamine hypothesis of schizophrenia has been formulated to integrate new findings (36). Preclinical and clinical studies (i.e., post-mortem, metabolite, and dopamine receptor neuroimaging studies) have advanced the understanding of relationships between affinity and occupancy of D<sup>2</sup> and D<sup>1</sup> subtypes of the dopamine receptors and regional specificity (37). Furthermore, it was assumed that findings of altered regional dopaminergic receptor function from preclinical and indirect clinical studies could be linked to clinical symptomatology in schizophrenia (36). The hypothesis suggests that "hypofrontality," as measured with reduced regional cerebral blood flow (rCBF) in the PFC may indicate low dopamine levels in the PFC (36). Findings from preclinical

<sup>2</sup> In this review, we focus on the "2-back" task [verbal "2-back", (104); numeric "2-back", (97)] to review/discuss brain function and PET findings of comparable experimental paradigms, psychological/ cognitive domains/components and activated brain regions.

<sup>3</sup>Exemplary studies on verbal fluency findings are presented.

lesion studies proposed that prefrontal "hypodopaminergia" lead to striatal "hyperdopaminergia" (38, 39). In addition, it is hypothesized that prefrontal "hypodopaminergia" could cause negative symptoms, whereas striatal "hyperdopaminergia" could lead to positive symptoms (36).

The dopamine hypothesis: version III synthesizes published findings on dopamine and its potential role in schizophrenia from the main fields into one unifying hypothesis. The hypothesis aims to provide a framework for findings from developments in clinical research into genetic (risk) factors, environmental risk factors, neurochemical and neuroimaging studies, and preclinical studies, which may be related to increased presynaptic striatal dopaminergic function in schizophrenia (23). The hypothesis is comprised of four components: (i) The interaction of "hits" such as frontotemporal dysfunction, genes, stress, and drugs may lead to striatal dopamine dysregulation (i.e., increased presynaptic dopamine synthesis capacity) and therefore to psychosis. (ii) It is hypothesized that the primary dopaminergic dysfunction is located at the presynaptic dopaminergic level instead of the D<sup>2</sup> receptor level. (iii) The hypothesis assumes that the dopamine dysregulation combined with cultural and societal factors could lead to future clinical diagnosis of "psychosis" rather than schizophrenia. (iv) It is proposed that the dopamine dysfunction could change the perception and judgment of stimuli (possibly through aberrant salience), which could result in cognitive deficits (40, 41).

Recent meta-analyses, which examined markers of striatal dopamine alterations in schizophrenia, reported evidence of different types of elevated dopamine dysfunction. Supporting evidence for the dopamine hypothesis has been shown by increased striatal presynaptic dopaminergic function in medication-free or medication-naïve patients with schizophrenia contrasted to healthy controls (42) and increased striatal dopamine synthesis capacity (43). Contradictory findings have however been reported by Fusar-Poli and Meyer-Lindenberg (44), who found no difference in striatal dopamine active transporter density between patients with schizophrenia and healthy controls.

In summary, while both the dopamine receptor hypothesis and the modified dopamine hypothesis of schizophrenia have their origins in the neurobiological investigation of the mode of action of antipsychotics, the dopamine hypothesis: version III aims at integrating advances in research of schizophrenia into one unifying dopamine hypothesis. The scope of understanding of dopaminergic dysregulation has become more defined, ranging from the whole brain perspective, through the perspective of regional specificity between (DL)PFC and striatum, to the current perspective of elevated presynaptic striatal dopaminergicfunction. The development of the dopamine hypothesis over the three versions has helped shape the understanding of schizophrenia as a brain disorder.

## **GLUTAMATE HYPOTHESIS OF SCHIZOPHRENIA**

The origin of the glutamate hypothesis of schizophrenia was based on the discovery of psychotomimetic effects of ketamine and phencyclidine, which elicited psychotic symptoms in healthy people. Symptoms such as delusions and hallucinations experienced by healthy individuals were compared to positive symptoms seen in FES (45, 46). The glutamate hypothesis postulates a mechanistic

process of altered interacting glutamatergic and/or dopaminergic neurotransmitter circuitries implicated in the pathophysiology of clinical and cognitive symptoms in schizophrenia (47–50). In this review, we consider three models of the glutamate hypothesis with relevance to the investigation of altered working memory function in people with schizophrenia: (i) the "*N*-Methyl-d-aspartate acid (NMDA) receptor hypofunction model" of schizophrenia, (ii) the "acute ketamine model," and (iii) the "dysconnection hypothesis" of schizophrenia.

The NMDA receptor hypofunction model of schizophrenia posits that the subtype of the glutamate receptor is implicated in multiple pathological brain mechanisms of schizophrenia ranging across cellular, chemical, and neuronal levels (51–54). It has been proposed that NMDA receptor hypofunction could underlie the pathophysiology of negative and cognitive symptoms in schizophrenia (29, 51, 55, 56). Clinical trials with agents modulating NMDA receptor in addition to treatment with first-generation antipsychotics (FGA; such as chlorpromazine, haloperidol, perphenazine) and second-generation antipsychotics (SGA; such as clozapine and olanzapine) presented supporting evidence for amelioration of negative and cognitive symptoms (51, 57, 58). Evidence for the involvement of NMDA receptor hypofunction through interactions among different neurotransmitters such as γaminobutyric acid (GABAergic) interneurons (51) and dopamine (59, 60) has also been reported.

Evidence for the glutamate hypothesis in humans is based on clinical studies with ketamine in healthy subjects. Results suggest that glutamatergic alterations could explain the pathophysiological mechanisms resulting in positive symptoms predominantly experienced by FES and those with first episode psychosis (FEP) (45,61).While findings from ketamine injection studies have aided the understanding of glutamatergic signaling in the development of delusions and hallucinations, evidence for altered glutamatergic transmission in negative and cognitive symptoms is scarce. FMRI findings from ketamine studies in healthy subjects propose that altered glutamatergic signaling could be implicated in working memory (12, 45, 62). These findings are in keeping with evidence from glutamatergic animal models, which report aberrant working memory function after the inhibition of glutamatergic receptors (63–66).

The dysconnection hypothesis of schizophrenia posits that altered NMDA receptor-mediated synaptic plasticity may be the underlying pathophysiological mechanism in individuals with schizophrenia (20, 21, 67). The authors propose that altered synaptic plasticity may explain both clinical symptoms and cognitive deficits in people with schizophrenia neurobiologically by altered NMDA receptor neuromodulation. Therefore, the dysconnection hypothesis synthesizes neurobiological findings (i.e., dopamine as one of the main neuromodulators leading to aberrant NMDA receptor function) with clinical and cognitive neuroscientific findings (i.e., cognitive impairment) in individuals with schizophrenia. One of the main objectives of the dysconnection hypothesis is to offer a new approach and therefore new interpretation of neurophysiological and neuroimaging data. This may be used to assist in the understanding of altered cognitive function in people with schizophrenia. For functional neuroimaging data, the biophysical modeling approach of dynamic causal modeling [DCM; Ref. (18)] has been proposed to infer biophysical processes (namely, NMDA receptor-dependent synaptic plasticity) underlying the blood-oxygen-level-dependent (BOLD) responses. In addition, the authors provide arguments that the development of positive symptoms such as delusions can be explained by a "failure of selfmonitoring mechanism" or "corollary discharge" (20). Abnormal EC findings from EEG and fMRI studies across a range of cognitive tasks in subjects with schizophrenia in contrast to healthy controls have been reported (68–70). These lead to a new insight into altered connectivity above those provided by FC studies, which are formulated under different theoretical frameworks, specifically DCM findings enable the inference of biophysical processes underlying neural responses (18, 19, 71).

In summary, the three hypotheses, the NMDA receptor hypofunction model, the acute ketamine model, and the dysconnection hypothesis, have motivated researchers to investigate biophysical circuit processes implicated in glutamatergic and dopaminergic interaction in negative symptoms and cognitive function in schizophrenia. These circuit mechanisms are thought to underlie altered working memory function in schizophrenia. Research on the NMDA receptor hypofunction model has its roots in the pharmacological examination of antipsychotics, the development of new agents, and its effects on clinical and cognitive symptoms in preclinical and clinical research in schizophrenia. The focus of researchers examining the acute ketamine model and the dysconnection hypothesis lies on elucidating proposed neurobiological processes of blockade of NMDA receptor underlying altered cognitive brain function in schizophrenia. The study designs of both versions differ in the investigation of (i) the pharmacological effect of ketamine on altered cognitive brain function and clinical symptomatology in healthy controls (the acute ketamine model) and (ii) altered synaptic plasticity during altered cognitive brain function in subjects with schizophrenia. Despite the different approaches, researchers of both versions of the glutamate hypothesis share the common aim of increasing our insight into schizophrenia by the translation of neurobiological knowledge from basic research to clinical research in schizophrenia. Furthermore, researchers share the common methodological approach of large-scale network analysis of fMRI data. Taken together, development over the three versions of the glutamate hypothesis of schizophrenia have presented promising evidence for shaping the understanding of schizophrenia as a cognitive brain network disorder.

## **SCHIZOPHRENIA AS A COGNITIVE BRAIN DISORDER**

Clinical and cognitive neuroscience studies have applied *in vivo* neuroimaging techniques of fMRI, PET, and single-photon emission computed tomography (SPECT) to assess neurobiological processes that underlie working memory function in people with schizophrenia. Techniques such as PET and SPECT use injections of positron-emitting radionuclide as tracer (for PET) or gamma-emitting radionuclide as tracer (for SPECT) in the living brain. Although these nuclear medical imaging techniques are non-invasive they require the administration of tracers. FMRI provides non-invasive *in vivo* imaging,which measures brain function by means of the BOLD response (72).

In the last two decades, the fields of clinical and cognitive neurosciences merged to provide a multidisciplinary approach to research into schizophrenia. This approach has created the notion of schizophrenia as a cognitive brain disorder (15, 73, 74).

## **EXAMPLES OF fMRI AND PET STUDIES INVESTIGATING ALTERED WORKING MEMORY FUNCTION IN SUBJECTS WITH SCHIZOPHRENIA**

Working memory tasks were initially investigated with fMRI in healthy subjects (75–78). These initial findings led to the use of fMRI as a tool for examining neurobiological markers that could be related to working memory performance. The examination of working memory function was extended to individuals with schizophrenia.

Reported findings of brain function during working memory (among several domains and components of working memory tasks) in healthy controls have led to the understanding that dopamine modulates working memory in healthy controls (79–81). This evidence of dopaminergic involvement in working memory was extended by the findings of altered dopaminergic modulation in schizophrenia (74, 82). Subsequently, converging findings were reported that regions such as DLPFC, anterior cingulate cortex (ACC), and parietal cortex (PC) are activated in working memory in both healthy controls and in subjects with schizophrenia (83–86). However, in those with schizophrenia, these regions exhibit increased or reduced functional activations and FC between prefrontal and parietal regions as well as between prefrontal and temporal regions in contrast to healthy controls. Alterations in FC occur at all stages of the illness (87, 88): (i) in HR subjects (89); (ii) in FES and FEP (90), and (iii) in subjects with EST (91).

Systematic reviews and meta-analyses of working memory fMRI studies in people with schizophrenia do not report consistent findings (92–95). Some studies report increased activation of the DLPFC, commonly referred to as "hyperfrontality," however, others report decreased activation or "hypofrontality." This picture of differing functional activation in terms of the direction, extent, and/or pattern of BOLD responses was attributed to the variation of domains and components of working memory tasks (92–95). Also it was considered that methodological factors in the applied analyses would contribute to these variations in functional activation (93, 95, 96). In addition, differences in medication could contribute to variation in the reported functional activation between studies.

Here, we review exemplary fMRI studies using the numeric or verbal "N-back" task in subjects with EST and healthy controls, which reported functional activation and FC findings (**Table 1**). The reviewed studies present group differences between subjects with schizophrenia and healthy controls. In functional activation studies, evidence was reported for increased activation in DLPFC, PFC, ventral PFC, medial frontal gyrus, and AC during high working memory load in subjects with EST (89, 97–101). However, reduced activation in prefrontal regions, such as ventral PFC, DLPFC, AC, and parietal regions was found during high working memory load in subjects with EST (97, 98, 102). One study in FES found a reduction of activation in inferior frontal gyrus (IFG), superior frontal gyrus, and AC during high working memory load (103). We note three factors, which contributed to difficulties in comparing the findings across the reviewed studies: (i) missing information of phase of schizophrenia (100), (ii)



second-generation

antipsychotics;

 SFG,

frontal gyrus; IPL, inferior parietal lobe; LP,

superior frontal gyrus; SPL, superior parietal lobule; vPFC, ventral prefrontal cortex; vlPFC, ventrolateral

aPatients with different schizophrenia bPhase of illness, illness onset, and illness duration not reported. Phase of illness based on symptoms

cChlorpromazine

dSeed-based

 connectivity

 only reported here.

 equivalents

 in milligrams per day.

 subtypes, such as paranoid subtype,

low-performers;

 OFG, orbitofrontal

 gyrus; PC, posterior cingulate; PHG,

schizoaffective

 subtype,

 PFC; WM, working memory.

undifferentiated

 scores.

 subtype.

parahippocampal

 gyrus; ROI, region of interest; SGA,

heterogeneous groups of subjects with EST (97, 101, 103), and (iii) limited information on antipsychotic treatment (89, 97–101, 103). Fundamentally, none of the functional activation findings was interpreted in context of the dopamine or glutamate hypothesis. The lack of a clear understanding in terms of neural activation and pathophysiological mechanism suggests there is a need for studies examining wider prefrontal circuitry underlying working memory deficits in schizophrenia (93, 95).

Functional connectivity studies applied voxel-based seed approaches to the BOLD response (89, 100, 103), with the exception of one study, which applied an ROI-to-ROI approach (101). Despite equivalent methodological approaches, the FC findings are not entirely comparable due to the use of different seed locations. Findings of reduced connectivity involving subregions of the PFC were found in FES and EST. Reduced FC findings in subjects with schizophrenia and EST were reported in the majority of studies: (i) Reduced prefrontal–parietal<sup>4</sup> FC in subjects with schizophrenia (100); (ii) Reduced prefrontal–hippocampal, prefrontal–striatal, and within-PFC FC in EST (89); and (iii) Reduced parieto-prefrontal FC and between putamen and ventrolateral PFC in EST (101). Further evidence for reduced FC between medial frontal gyrus and putamen was found in FES (103). In contrast to most studies that report reduced FC in the early and late phases of the illness, increased FC between the ventral PFC and posterior PC was shown in subjects with schizophrenia (100). The findings of both reduced and increased FC between subregions of the PFC and the posterior PC may be related to variations in behavioral response to task load for subjects with schizophrenia (100). Similar difficulties in comparing the FC findings among the studies are present as in the comparison of the functional activation studies due to unclear and missing information regarding the illness phase, diagnosis, and medication treatment. Similarly, no reference is made to the dopamine or glutamate hypothesis in interpreting the FC findings.

In summary, findings presented by FC studies during the "Nback" task have paved the way for the understanding of largescale functional networks in working memory. Furthermore, the insight of brain alterations in subjects with schizophrenia has advanced with FC from individually activated regions to connectivity between brain regions. The perspective of circuit-based neurobiology and cognitive brain function opens the doors for translational research from preclinical and clinical research in schizophrenia. However, FC is limited as the connection assessments are based upon regional correlations and this approach does not allow inferences of directions or causality between connected regions (18).

Positron emission tomography and SPECT imaging in schizophrenia research are used to investigate indirect markers of dopamine measures such as D2/3 receptors, presynaptic dopaminergic function, dopamine synthesis capacity, dopamine release, and dopamine transporters. Three [H215O] PET studies consistently reported reduced rCBF in DLPFC and PC in verbal/numeric "2-back"in subjects with EST in contrast to healthy controls (104– 106). Reduced prefrontal–hippocampal FC findings in subjects

with schizophrenia in contrast to healthy controls (105, 106) confirmed the hypothesis of reduced functional connections in working memory. Correlational PET studies provided support for dopaminergic alterations and measures of the "2-back" task in subjects with schizophrenia (2, 16).

In summary, fMRI and PET studies in the field of clinical and cognitive neurosciences have been used to investigate brain function during working memory in people with schizophrenia (**Figure 1**). Both fMRI and PET findings have advanced the understanding of altered working memory performance and brain function in subjects with schizophrenia. This has led to better insight into the interaction between altered working memory function and experimental/clinical factors (such as cognitive domains of working memory function, performance level, phases of illness, clinical symptomatology, and effects of antipsychotic medication) in individuals with schizophrenia.

## **EXAMPLES OF fMRI STUDIES INVESTIGATING ALTERED SPATIAL WORKING MEMORY FUNCTION – GLUTAMATE HYPOTHESIS OF SCHIZOPHRENIA**

The role of the DLPFC in working memory deficits has been associated with glutamatergic alterations and more specifically in dopamine–glutamate interactions (10, 50, 51). Furthermore, it has been reported that ketamine, a NMDA receptor antagonist, can induce psychosis-like symptoms in healthy subjects (45). Here, we briefly summarize the main functional activation and FC findings of fMRI studies on the spatial "N-back" task in the context of the glutamate hypothesis of schizophrenia (**Table 2**).

Anticevic et al. (12) presented ketamine-induced reduced functional activation in task-activated regions (such as the DLPFC and the precuneus) and task-deactivated regions of the default-mode network (DMN). In addition, the combination of a spiking localcircuit model of performance during the spatial "N-back" task and the functional activation findings revealed that the modulation of

<sup>4</sup>Reduced FC between the dorsal PFC and posterior PC.


**Table 2 |Schizophrenia as a cognitive brain disorder II – summary of main findings in spatial working**

 **memory.** ketamine alters the association between the task-activated and the task-deactivated networks. Finally, it was shown that ketamine modulates FC between the fronto-parietal and DMN networks. In a recent study, Driesen et al. (62) provided further support for ketamine-induced reduced prefrontal FC during the spatial "Nback" task. Two FC approaches with the same seed regions were employed, seed-based FC and global-based connectivity (GBC), which revealed both decreased FC within the DLPFC. The seedbased analysis resulted in reduced FC between DLPFC and middle frontal gyrus [MFG, IFG, and insula (among other regions) under ketamine in contrast to saline]. The GBC analysis showed decreased FC of the DLPFC under ketamine.

In summary, these studies on altered spatial working memory function inform on the glutamate hypothesis, through the acute ketamine model (**Figure 2**). In this, they have advanced the understanding of NMDA receptor-modulated brain function in healthy subjects.

## **FROM COMPUTATIONAL NEUROPSYCHIATRY TOWARD SCHIZOPHRENIA AS A COGNITIVE BRAIN NETWORK DISORDER**

Clinical and cognitive neurosciences have advanced the understanding of altered working memory function in subjects with schizophrenia. FMRI studies in working memory among other neuroimaging and electrophysiological techniques, have reported on functional activation and FC findings in subjects with schizophrenia. Both findings of functional activation and FC revealed methodological, cognitive, and clinical factors related to our understanding of altered working memory function in patients with schizophrenia. In particular, FC findings mark the beginning of the notion of "disconnection" and "dysconnection" (20, 21, 67, 107) in working with people with schizophrenia. FC is defined as the statistical association or dependency among two or more anatomically distinct time-series (107). FC findings cannot be interpreted in terms of causal effects between connected regions and thus, does not allow for a mechanistic inference of the BOLD responses.

The modeling of functional large-scale networks<sup>5</sup> during working memory function in schizophrenia could provide mechanistic explanations for altered brain function in individuals with schizophrenia. The advantage of modeling functional large-scale networks in terms of EC over FC is that inferences can be drawn on mechanistic processes, which are not directly observable in the BOLD response.

## **COMPUTATIONAL NEUROSCIENCE AND COMPUTATIONAL NEUROPSYCHIATRY**

Marr proposes a theoreticalframeworkfor computational research on the brain on three levels (1976). At the first level, researchers should aim to gain knowledge of the high-level computations of the brain such as working memory ("computational level"). At the next level, the testing of the brain's methods and algorithms for the high-level working memory function is led by hypotheses derived from the acquired knowledge and testing how appropriate an algorithm such as Bayesian inference is for modeling the working

memory brain function ("algorithmic level"). Finally, when an algorithm is found, which is valid and more likely than alternative algorithms to predict known brain function/behavior, then the investigation of the biological implementation can be pursued ("physical level").

Computational neuropsychiatry is an emerging field within computational neuroscience. Computational neuropsychiatry aims to provide an explanatory bridge between altered cognitive function and neurobiological mechanisms associated with the development of mental illness (108, 109, Huys, unreferred preprint). Computational neuropsychiatry in humans has been defined by outlining a set of components, which include biophysical modeling and computational modeling (109). Different types of computational models at different neural levels are used dependent on the study hypothesis (108, Huys, unreferred preprint).

## **COMPUTATIONAL NEUROPSYCHIATRY AND MODELING OF FUNCTIONAL LARGE-SCALE NETWORKS IN SUBJECTS WITH SCHIZOPHRENIA**

Connectionist and neural network models in working memory/cognitive control in subjects with schizophrenia have added to our understanding of both the brain function and the neurobiological mechanism underlying working memory (74, 76). The strength of these models is based on the translational link between human brain function (i.e., functional activation) and preclinical neurobiological evidence (namely, dopaminergic modulation) during working memory.

Following on from the work of Cohen and Braver, evidence for the understanding of schizophrenia as a cognitive network disorder has been presented by both preclinical studies (8, 10, 110–113) and human FC studies in working memory (89, 100, 101, 103). Recent studies examining biophysical mechanisms underlying

<sup>5</sup>As one subfield within computational psychiatry.

altered functional large-scale networks aim to bridge the gap between the human functional network used in working memory and the preclinical neurobiological processes. Examples of such computational neuropsychiatric studies, including EC during working memory in subjects with schizophrenia, are reviewed. In this, we focus on DCM studies investigating the numeric/verbal "N-back" task in subjects with schizophrenia and healthy controls. This is considered in the context of the dopamine and glutamate hypotheses of schizophrenia. Both neurobiological hypotheses have contributed to the formulation of research objectives in computational neuropsychiatry (114) and the development of computational modeling techniques of fMRI data in subjects with schizophrenia (20).

## **Dynamic causal modeling for fMRI – examples of modeling functional large-scale networks**

Dynamic causal modeling for fMRI has been introduced as a method to provide insight into the notion of "functional integration" during cognitive performance. "Functional integration" has been advanced from the historic notion of "functional specialization" (115), which is defined by context-dependent interactions among different brain regions (18, 116–118).

Dynamic causal modeling has been described as a biophysical modeling of neuronal dynamic processes (18, 19) 6 , which can be used as a method for the computation of synaptic plasticity from fMRI task-based studies (20, 21). Together biophysical modeling and Bayesian inference analysis form the framework for DCM (71, 117, 118). Thus, DCM is a modeling approach, which combines defined network models (i.e., hypotheses) with Bayesian inversion methods (19, 117). Specifically, DCM assesses inter-regional EC through assessment of experimentally induced changes (18) and therefore allows for mechanistic inferences from neuronal function.

Bilinear DCM infers dynamics at the neuronal level by translating modeled neuronal responses into predicted BOLD measurements (18). Non-linear DCM for fMRI (71, 119) is an advanced approach for increasing the biological plausibility of DCMs by the means of modeling "gain modulation" (i.e., non-linear modulation of neuronal connections) (19, 117, 118). In non-linear DCM, the modulation of connection strengths by experimental inputs is supplemented by direct modulation of neural activity in one or more network regions (18, 119). The computations for gating in neural networks use the multiplicative computation of non-linear modulation (120, 121). Accordingly, non-linear DCM can be used for inferring that the strength of a connection is modulated by activity of other neuronal populations (119, 122).

*Findings of altered effective connectivity during working memory in subjects with schizophrenia.* The first DCM studies in healthy controls described large-scale networks in working memory and a similar task [continuous performance test; (123–125)]. A recent study in healthy controls built the linkage between EC results and underlying dopaminergic modulation of large-scale networks comprising of the DLPFC and PC during verbal memory performance (126).

To date four DCM studies have examined the verbal/numeric "N-back" task in subjects with schizophrenia using bilinear DCM (127–130) (**Table 3**). These provide novel insights into reduced task-dependent EC and increased task-independent EC measures through modeling large-scale networks in schizophrenia.

In the first study, increased fronto-temporal intrinsic connectivity was found to be associated with increased functional activation of the superior temporal gyrus (STG) during the numeric "N-back" task in the subjects at the prodromal and at the acutely psychotic stage of schizophrenia in contrast to the healthy controls. This suggests a potential marker for vulnerability to the disorder (127). Furthermore, progressively decreased intrinsic connectivity between the STG and the MFG in subjects at-risk mental state (ARMS) and FES subjects in contrast to the healthy controls was reported. This finding suggested that functional activation may resemble increased task-independent EC between the PFC and the STG. However, the results of the study are not comparable to other DCM studies because (i) only one model was examined and (ii) the biological plausibility of the EC measures is not clearly accessible. No reference to the dopamine or glutamate hypotheses was made.

The second study investigated the working memory-dependent modulatory effect for the prefrontal–parietal connectivity in subjects with EST and healthy subjects during the numeric "N-back" task (128). The large-scale networks included the right DLPFC, the PC, and the visual cortex with bidirectional connection between all regions. The main finding was decreased task-dependent EC from the DLPFC to the PC in the subjects with EST. Thus, this finding could resemble evidence for the glutamate hypothesis of schizophrenia, specifically the NMDA receptor hypofunction model and the dysconnection hypothesis.

The third study examined possible vulnerability markers for psychosis from the verbal "N-back" task in ARMS subjects, FES subjects, and healthy subjects (129). This study examined reduced task-dependent EC measures as well as relationships between connectivity parameters and antipsychotic medication received by subjects. In this study, EC in interhemispheric large-scale networks between the bilateral superior parietal lobes (SPL) and the bilateral MFG was assessed. This study reported novel findings of progressively decreased working memory and induced modulation of connectivity between the MFG and the SPL (from healthy subjects to ARMS). Additionally, further decreased EC of modulatory effects were observed in non-medicated subjects with FEP contrasted to healthy controls. Evidence for amelioration of reduced EC between the MFG and the SPL in subjects with FES, who received SGA medication, could reflect alterations of dopaminergic regulation of NMDA receptor-dependent synaptic plasticity of fronto-parietal connections. However, this interpretation is limited by the lack of a control group of FES who are treated with different types of antipsychotic medication. These findings across different subpopulations of schizophrenia together with the effect of antipsychotic medication may reflect support for the NMDA receptor hypofunction model and the dysconnection hypothesis.

<sup>6</sup>We consider DCM as the generative model approach as introduced in the seminal article by Friston et al. (18).


**|SchizophreniaasacognitivebrainnetworkdisorderIIsummaryofmainfindingsinverbal/numericworkingmemoryneuroimagingandbiophysical**

episode schizophrenia; FEP, subjects with first episode psychosis; FGA, first-generation antipsychotics; HC, healthy controls; HR, subjects at high risk of schizophrenia; IC, intrinsic connectivity; IFG, inferior frontal gyrus; INS, insula; MFG, middle frontal gyrus; PC, parietal cortex; PCC, posterior parietal cortex; PFC, prefrontal cortex; SGA, second-generation antipsychotics; SMA, supramarginal area; SPL, superior parietal lobe; STG, superior temporal gyrus; VC, visual cortex.

aSubjects at high clinical risk of schizophrenia.

bWith high suicidal risk.

cWith low suicidal risk.

dChlorpromazine equivalents in

 mg/day.

eBMS at the group-level.

fBMS at the model family level.

In the fourth study, Zhang et al. (130) explored EC measures in terms of possible neurobiological markers in groups of subjects with schizophrenia with high or low suicide risk and contrasted these with healthy controls during the verbal "N-back" task. The large-scale networks were defined by unidirectional and bidirectional connections between the two regions of the medial PFC and PC as well as working memory effects on these regions. This pilot study presented novel findings in subjects with schizophrenia at suicidal risk in terms of increased intrinsic connectivity from the PC to the MFG in both groups with FES (in comparison to healthy controls). This finding was interpreted as a possible association to schizophrenia, in which increased intrinsic connectivity from the MFG to the PC in the subjects with high risk of suicide could reflect vulnerability of suicide. However, the results are not directly comparable to the other DCM studies because of the study population, which focused on the issue of suicide. The findings were also not interpreted in context of the dopamine or glutamate hypotheses.

We highlight main experimental and methodological limitations in the four DCM studies, which impede the comparability of findings (please see **Table 3** for details). The main experimental limitation focuses on the discrepancies between the different patient subpopulations. Two studies analyzed working memory fMRI data of subjects with ARMS and FES in comparison to healthy controls (127, 129),whereas one study modeled scans from subjects with EST (128). Zhang et al. (130) reported findings of a unique patient population of FES with high and low suicidal risk. In terms of methodological issues, one limitation lies in different definitions of model spaces for the large-scale networks, despite equivalence in the experimental tasks. Another limitation is that the reviewed DCM studies employed deterministic DCM for the comparison of the models. Deterministic models can predict processes perfectly if all inputs are known (131). However, at this early stage of employing biophysical modeling approaches to human brain function, we do not have a full understanding of the brain responses to working memory. Future studies may employ stochastic DCM as an extension (117, 118, 132). A further limitation is that different DCM versions were applied across the four studies, which impede the comparability of the findings. The priors are differently defined in the used DCM versions, which give rise to a variation in model evidence between the studies (117). Thus, it is possible that discrepancies in EC findings could be due to the prior definition and may not be solely due to differences in performance, brain function, or clinical aspects of subjects with schizophrenia. Lastly, a general limitation of DCM for fMRI is that maximally 10 regions within a large-scale network can be modeled. This simplification results in difficulties of biophysical modeling of tasks, which are likely to encompass more than ten regions. Furthermore, not only the definition of different regions and different numbers of regions but also different modulatory inputs result in further extensions to the model space. Such model spaces are difficult to validate and analyze.

The four DCM studies presented evidence for increased taskdependent EC and increased task-independent EC findings during verbal/numeric working memory in subjects with schizophrenia. We discuss these EC findings in context of (i) the dopamine and glutamate hypothesis and (ii) FC findings during verbal/numeric working memory in subjects with schizophrenia.

The four studies modeled large-scale networks during the "Nback" task in subjects with schizophrenia. However, only two out of these four studies consider their DCM results in the light of biophysical processes (128, 129). The findings of reduced EC (namely, the effect of task-modulation) of the prefrontal–parietal connection in subjects with schizophrenia in contrast to healthy controls were interpreted biophysically and linked to the NMDA receptor hypofunction model and the dysconnection hypothesis (128, 129). Both studies reported reduced EC findings of the prefrontal– parietal connection during working memory, however, these findings need to be considered carefully due to different experimental designs (i.e., patient subpopulations, antipsychotic medication treatment of FGA and SGA) and methodological implementation (i.e., model space, DCM settings, and inference techniques).

Three of the DCM studies reported altered EC findings of the prefrontal–parietal and parieto-prefrontal connections during the "N-back" task in subjects with schizophrenia in contrast to healthy controls. Deserno et al. (128) and Schmidt et al. (129) presented reduced EC (effect of task-modulation) of the prefrontal– parietal connection in subjects with schizophrenia in contrast to healthy controls, whereas Zhang et al. (130) found increased EC (intrinsic connectivity) of the parietal–prefrontal connection. The reduced task-dependent EC findings are in keeping with reduced FC findings of these connections, although increased FC between a different prefrontal subregion and the PC was reported (100).

The study by Crossley et al. (127) reported increased EC (intrinsic connectivity) of the prefrontal–temporal connection in subjects at HR and FES (in contrast to healthy controls). Reduced FC of the prefrontal–temporal connection during the "N-back" task in subjects with schizophrenia has been previously reported in PET studies (105, 106). However, the regions within the PFC and temporal region differ between the studies.

*Findings of altered effective connectivity during verbal fluency in subjects with schizophrenia.* Here, we discuss bilinear and non-linear DCM studies, which have assessed large-scale networks during verbal fluency [namely, the Hayling sentence completion task (HSCT)] in subjects with schizophrenia and healthy controls (**Table 4**). One bilinear DCM study in healthy controls investigated the task-dependent modulation of response initiation and response suppression in EC between left hemispheric temporal and prefrontal regions (133). The main finding was a difference in connection strength of the modulatory effect in response initiation and response suppression.

Two clinical bilinear DCM studies have investigated EC measures during the HSCT in HR subjects and healthy controls: (i) Subjects at high clinical risk of schizophrenia [ARMS; Ref. (134)] and (ii) subjects at high familial risk of schizophrenia (135). Allen et al. (134) investigated increased fronto-temporal EC (intrinsic connectivity) as a potential measure of vulnerability of developing schizophrenia. Two main findings were reported: firstly, no significant effect of task-dependent modulation on the frontotemporal connection between ARMS subjects and healthy controls was revealed. Secondly, ARMS subjects displayed increased intrinsic connectivity between the ACC and the MTG in comparison to healthy controls. Furthermore, the Bayesian model selection (BMS) approach revealed that the same network was equally likely


**4|SchizophreniacognitivebrainnetworkdisorderIIofmainfindingsinverbalfluencyneuroimagingandbiophysical**

inferiorfrontalgyrus;IPS,intraparietalsulcus;MDmediodorsal;MFG,middlefrontalgyrus;MTG,middletemporalgyrus;

 

aSubjects at high clinical risk of schizophrenia.

bSubjects at high familial risk of schizophrenia.

cAt the time of scanning.

dBMS at the group-level. eBMSatthe modelfamilylevel.

to explain the given HSCT fMRI data in both the ARMS subjects and the healthy controls. No reference to the glutamate hypothesis was made.

Dauvermann et al. (135) modeled EC measures in a similar version of the HSCT that was used by Allen et al. (134). This study was conducted in subjects at high familial risk of schizophrenia and healthy subjects. The results reported by Allen et al. (134) of a similar large-scale network in both HR subjects and healthy controls was replicated<sup>7</sup> . This finding was also confirmed by Dauvermann et al. (135), when the group of HR subjects was subdivided into high risk subjects without transient psychotic symptoms (referred to as HR−), high risk subjects with transient psychotic symptoms (referred to as HR+) and high risk subjects who subsequent to scanning developed schizophrenia [referred to as HRill; please see Ref. (136, 137)]. Comparability between these two studies is limited due to differences in the model space. The model space in Dauvermann et al. (135) includes the IPS and the mediodorsal thalamus, which are not incorporated in the model space by Allen et al. (134). In addition, endogenous connections and task-dependent modulations were accordingly changed [Ref. (135);**Table 4**]. There was no reference to the glutamate hypothesis of schizophrenia.

Limitations of bilinear DCM have been addressed through the development of non-linear DCM for fMRI (119). This method was applied in the genetic high risk study reported by Dauvermann et al. (135). The progress from the bilinear DCM to the non-linear DCM as reported by Dauvermann is based on the biophysical modeling of connection strength with non-linear modulation during the HSCT response. The authors show that relative to healthy controls there is reduced connection strength with non-linear modulation of the thalamo-cortical connection during the HSCT in HR+ subjects and a further reduction in this connection strength in HRill subjects (135). The authors suggest that reduced gain control may underlie the reduced strength in the thalamo-cortical connection. Furthermore, the findings of reduced connection strength with non-linear modulation of the thalamo-cortical connection could reflect altered glutamatergic neurotransmission, which may underlie a disruption of synaptic plasticity in this thalamo-cortical connection [Ref. (135); **Table 4**]. Thus, the findings were interpreted in context of the NMDA receptor hypofunction model and the dysconnection hypothesis.

### **Summary of studies modeling functional large-scale networks – dynamic causal modeling for fMRI**

Evidence from brain function in working memory in subjects with schizophrenia at the level of functional large-scale networks (i.e., clinical and cognitive neurosciences) and neurobiological mechanisms in working memory in animal models of schizophrenia (preclinical neurobiological research) in combination with computational neuroscientific approaches has informed and enabled research in computational neuropsychiatry.

Exemplary DCM studies in subjects with schizophrenia have reported both increased and reduced EC findings during cognition in subjects with schizophrenia in contrast to healthy controls. These studies applied DCM as a biophysical modeling approach to functional large-scale networks, which enabled the interpretation of EC findings on the basis of the glutamate hypothesis of schizophrenia, namely the NMDA receptor hypofunction model and the dysconnection hypothesis (128, 129, 135). We emphasize that the findings support not only the glutamate hypothesis but also the dopamine hypothesis. Dopamine is a neuromodulator that may crucially affect glutamate-induced synaptic plasticity. Synaptic plasticity may be involved in a regulation of dopamine synthesis and release via other neurotransmitter systems. Specifically for non-linear effects, it has been shown that dopamine acts as a neuromodulator mediating postsynaptic gain (74, 138).

In a recent study, it has been reported that the combination of the DCM analysis of numerical "N-back" task in EST (128) and generative embedding resulted in the dissection of three subgroups of EST based on the mechanistically inferred DCM findings (139). This exemplary study showed that DCM can be applied as a generative model of large-scale networks in individuals with schizophrenia. In summary, DCM is a promising approach for modeling synaptic plasticity; nevertheless in its current form it cannot reflect the full complexity in the processing required for the implementation of tasks such as working memory (**Figure 3**).

### **UNDERSTANDING OF SCHIZOPHRENIA IN DEVELOPMENT**

Our understanding of schizophrenia is in continuous development and with more preclinical and clinical findings being published this understanding will advance further. A critical aspect of this understanding is the facilitation of multidisciplinary approach between preclinical and clinical research in schizophrenia.

The original understanding of schizophrenia as a brain disorder stems from observational clinical work, which led onto preclinical

<sup>7</sup> It is noted, however, that the large-scale networks differed slightly from the previous study.

investigation. Over time, the knowledge of alterations of cellular, chemical, and molecular mechanisms has increased: (i) findings of dopaminergic and glutamatergic modulation of working memory (and clinical features) in animal models of schizophrenia contributed to form the understanding of schizophrenia as a cognitive brain disorder; (ii) findings of neurotransmitter circuit systems, mainly dopaminergic and glutamatergic systems, were found to modulate working memory in animal models of schizophrenia in combination with computational studies (140), which plays a role in shaping the understanding of schizophrenia as a cognitive network disorder.

Understanding of schizophrenia has not only been shaped by preclinical research but also by clinical research in subjects with schizophrenia, which has been and continues to be illuminated by preclinical neurobiological and computational work. The field of clinical and cognitive neurosciences has contributed to forming our understanding of schizophrenia as a cognitive brain disorder. Importantly, the multidisciplinary field of computational neuropsychiatry (preclinical neurobiology, clinical and cognitive neurosciences, and computational psychiatry) has allowed for progress in our understanding of schizophrenia as a cognitive brain network disorder.

### **SCHIZOPHRENIA AS COGNITIVE BRAIN NETWORK DISORDER**

The use of computational neuropsychiatric research in developing our understanding of schizophrenia as a cognitive brain network disorder is at an early stage. Here, we focused on FC and EC studies (DCM studies) during the verbal/numeric "N-back" task in subjects with schizophrenia and healthy controls. We discuss these FC and EC findings in context of two key research questions. Consideration of these questions was seen as a means to inform future schizophrenia research in the fields of clinical and cognitive neurosciences and/or computational neuropsychiatry:

## **To what extent do these sets of findings support the dopamine hypothesis and/or the glutamate hypothesis in subjects with schizophrenia?**

Studies reported both increased and reduced FC during the "Nback" task in subjects with schizophrenia in contrast to healthy controls. These findings have introduced the notion of human large-scale networks underlying brain function during working memory. The FC correlational analyses do not allow for the inference of directions or weights of in functional connections. Thus, from FC findings it is not practical to draw inferences on neurobiological causal processing.

Studies, which applied DCM as a biophysical modeling approach to functional large-scale networks, showed that reduced EC findings could be interpreted in context of the NMDA receptor hypofunction model and the dysconnection hypothesis.

In summary, FC findings cannot be interpreted in context of the dopamine or glutamate hypothesis. For EC findings, the computational neuropsychiatric approach of modeling largescale networks requires biophysically plausible networks, which are hypothesis-driven from neurobiological and cognitive neuroscience in subjects with schizophrenia. EC findings have been interpreted in the context of the glutamate hypothesis and the dopamine hypothesis.

## **Do the findings from computational neuropsychiatry lead to a gain in understanding of schizophrenia in comparison to the findings from clinical and cognitive neurosciences?**

Functional connectivity findings from cognitive and clinical neuroscience have contributed to the understanding of schizophrenia as a cognitive brain disorder. The analysis of altered working memory at the level of large-scale networks has advanced our knowledge of cognitive function in humans. However, it is not wholly understood what altered FC during cognition neurobiologically means in schizophrenia. EC findings from computational neuropsychiatry, here specifically modeling functional large-scale networks with DCM, have shown indications of linkage between clinical network-based working memory (large-scale networks) and preclinical neurotransmitter modulation of cognitive function. Altered synaptic plasticity during working memory can be interpreted with dopaminergic and glutamatergic mechanisms. We emphasize that the interpretation of altered neurotransmitter circuits should be considered carefully because the DCM method is likely to underestimate the processing complexity in neurobiological circuits. Nonetheless, a strength of DCM lies in interpretation of altered synaptic plasticity based on the inference of mechanistic information.

The consideration of schizophrenia as a cognitive brain network disorder from computational neuropsychiatry offers a holistic view of schizophrenia. Computational neuropsychiatry seeks to bridge the gap between neurobiology and cognitive and clinical neurosciences in subjects with schizophrenia. It is hoped that this research will enhance our understanding of schizophrenia, clinical treatment, and improve outcome in people with schizophrenia.

## **FUTURE OUTLOOK AND OPEN QUESTIONS**

The reviewed findings in biophysical modeling of functional largescale networks are promising. In order to reach the objective of predicting and improving clinical treatment in subjects with schizophrenia, longitudinal study designs, and the combination of subfields within computational neuropsychiatry should be pursued.

We consider computational neuropsychiatric research areas for the combination of biophysical modeling of functional large-scale networks and other computational (neuro)psychiatric approaches, which are of clinical relevance for subjects with schizophrenia, for example:


We suggest specific study designs, which may increase our understanding for developing clinical treatment for subjects with schizophrenia:

	- (a) Brain function and brain circuit model (12);
	- (b) Brain function and behavior (141);
	- (c) Brain function and effect of antipsychotic medication:
	- (a) FMRI and EEG/magnetoencephalography study designs;
	- (b) FMRI and transcranial magnetic stimulation study designs (142);
	- (c) FMRI and MRS study designs;
	- (d) FMRI and PET study designs;
	- (a) Associative learning (143, 144);
	- (b) Machine learning approach (139, 145);
	- (c) Reinforcement learning (109).

Findings of modeling functional large-scale networks contribute to shaping the understanding of schizophrenia as a cognitive brain network disorder. The combination of computational neuropsychiatric areas may bring researchers closer to the common long-term objectives of developing a diagnostic tool for schizophrenia along with the development of more effective treatments.

## **ACKNOWLEDGMENTS**

We acknowledge Vincent Valton for his help. Maria R. Dauvermann and Thomas W. J. Moorhead are supported by Dr Mortimer and Theresa Sackler Foundation. Heather C. Whalley is supported by Royal Society Dorothy Hodgkin Fellowship.

## **REFERENCES**


or down. *Am J Psychiatry* (2003) **160**:2209–15. doi:10.1176/appi.ajp.160.12. 2209


**Conflict of Interest Statement:** Maria R. Dauvermann, Neil Roberts, Stephen M. Lawrie, and Thomas W. J. Moorhead have received financial support from Pfizer (formerly Wyeth) in relation to imaging studies of people with schizophrenia. Stephen M. Lawrie has done consultancy work for Roche Pharmaceuticals in connection with a possible new treatment for schizophrenia. Stephen M. Lawrie has also received honoraria for lectures, chairing meetings, and consultancy work from Janssen in connection with brain imaging and therapeutic initiatives for psychosis.

*Received: 04 September 2013; accepted: 10 March 2014; published online: 25 March 2014.*

*Citation: Dauvermann MR, Whalley HC, Schmidt A, Lee GL, Romaniuk L, Roberts N, Johnstone EC, Lawrie SM and Moorhead TWJ (2014) Computational neuropsychiatry – schizophrenia as a cognitive brain network disorder. Front. Psychiatry 5:30. doi: 10.3389/fpsyt.2014.00030*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2014 Dauvermann, Whalley, Schmidt, Lee, Romaniuk, Roberts, Johnstone, Lawrie and Moorhead. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

## **APPENDIX**

#### **DYNAMIC CAUSAL MODELING**

Dynamic causal modeling is general framework for modelbased assessment of competing theories about neuronal circuits (18, 146). In particular, DCM is a generic Bayesian system identification technique, which allows for inference on "hidden" neurophysiological mechanisms that generated observed measures, such as blood-oxygen-level-dependent signal in functional magnetic resonance imaging (fMRI) or evoked responses measured with electroencephalography (EEG). The principle idea thereby is to formulate a simplified model of neuronal population responses (*z*) and combine this with a modalityspecific forward model (λ) such that one can predict the measurement (*y*) that would arise from any particular neuronal circuit (18).

In DCM for fMRI, the dynamics of the neural states underlying regional BOLD response are modeled by a bilinear differential equation that describes how the neural states (*x*) change over time (*t*) as a function of endogenous inter-regional connections (matrix A), modulatory effects on these connections (matrix B), and direct (driving) inputs (matrix C) (Eq. A1) (18, 122). The endogenous connections represent coupling strengths in the absence of input u<sup>j</sup> to the system, whereas the modulatory effects represent task-specific alterations in this connectivity.

$$f(\mathbf{x}, u) = \frac{d\mathbf{x}}{\mathbf{x}t} = \left(A + \sum\_{i=1}^{m} \mu B^{(j)}\right) \mathbf{x} + Cu \tag{A1}$$

The bilinear state equation has subsequently been extended by a non-linear term, where the modulation of connection strengths by experimental inputs is supplemented by direct modulation with neural activity in one or more regions (119). In other words, non-linear DCMs allow addressing how the connection between two neuronal units is gated by activity in other units. This is of particular interest as gating processes represent a key mechanism for many neurobiological processes and thus increasing the biological realism of non-linear compared to bilinear DCMs. To this end, compared with the bilinear state equation, the new term in the non-linear equations are the *D* matrices (Eq. A2), which encode how the *n* regions gate connections in the system.

$$f(\mathbf{x}, u) = \frac{d\mathbf{x}}{dt} = \left(A + \sum\_{i=1}^{m} \imath B^{(j)} + \sum\_{j=1}^{n} \mathbf{x}\_{j} D^{(j)}\right) \mathbf{x} + Cu \quad \text{(A2)}$$

#### **BAYESIAN MODEL SELECTION AND BAYESIAN MODEL AVERAGING**

Bayesian model selection is an essential procedure of DCM studies as it can be used to test competing hypotheses (different DCMs) about the neural mechanisms generating the data. BMS rests on comparing the evidence of a predefined set of models (the model space). The model evidence is the probability of observing the empirical data, given a model, and represents a principled measure of model quality, derived from probability theory (147, 148). Concretely, it represents the mean predicted data under random sampling from the model's priors or, alternatively, a principled measure of the balance between model fit and model complexity. A random-effects BMS approach has been suggested for group studies, which is capable of quantifying the degree of heterogeneity in a population while being extremely robust to potential outliers (20, 67). The probability that one model is more likely than any other model, given the group data, can be expressed by the exceedance probability (ϕ*k*) of each model:

$$\begin{aligned} \exists k \in \{1 \ldots k\}, \forall j \in \{1 \ldots k|j \neq k\}: \\ \varphi\_k = p(r\_k > r\_j|\jmath; a) \end{aligned}$$

After inference on the most likely network architecture underlying a specific neural process, one can compare the parameter estimates of the most likely model obtained from BMS (winning model) for between-group inferences. However, statistical comparison of model parameter estimates across groups is only valid if those estimates stem from the same model. Given that different models may be found to be optimal across groups, Bayesian model averaging (BMA) has been recommended as standard approach for clinical DCM studies (146). BMA averages posterior parameter estimates over models, weighted by the posterior model probabilities (148). Thus, models with a low posterior probability contribute little to the estimation of the marginal posterior. In brief, BMS and BMA are central components of DCM studies to infer on neural mechanisms at the neural system level and on specific model parameters across groups, respectively (146).

In non-linear DCM analysis, the connection strengths between selected nodes are assessed for activity-dependent modulation of the reciprocal neuronal projections by the introduction of gating mechanisms. Non-linear DCM is applied to the models identified as winning models from the application of bilinear state equation. The bilinear model and the non-linear models differ only in the introduction of gating mechanisms such as a parametric response in the tested functional task. Such gating mechanisms are applied to nodal connections, which are expected to explain the variation in subject response to the functional task. The appropriate placement of the gating input is assessed through the application of model space partitioning and family inference. The exceedance probabilities of the models are compared and the non-linear models,which provide higher exceedance probabilities than the bilinear models are identified as winning models.

# Dysconnectivity in the frontoparietal attention network in schizophrenia

#### **Jonathan P. Roiser <sup>1</sup>\* † , RebekahWigton<sup>2</sup>† , James M. Kilner 3,4, Maria A. Mendez <sup>5</sup> , Nicholas Hon<sup>6</sup> , Karl J. Friston<sup>3</sup> and Eileen M. Joyce<sup>4</sup>**

1 Institute of Cognitive Neuroscience, University College London, London, UK

<sup>2</sup> Psychosis Studies, Cognition and Schizophrenia Imaging Lab, Institute of Psychiatry, King's College London, London, UK

<sup>3</sup> Wellcome Trust Centre for Neuroimaging, University College London, London, UK

4 Institute of Neurology, University College London, London, UK

<sup>5</sup> Department of Forensic and Neurodevelopmental Sciences, Institute of Psychiatry, King's College London, London, UK

<sup>6</sup> Department of Psychology, National University of Singapore, Singapore

#### **Edited by:**

Stefan Borgwardt, University of Basel, Switzerland

#### **Reviewed by:**

Maria R. Dauvermann, Harvard Medical School, USA André Schmidt, University of Basel, Switzerland

#### **\*Correspondence:**

Jonathan P. Roiser, Institute of Cognitive Neuroscience, University College London, 17 Queen Square, London WC1N 3AR, UK e-mail: j.roiser@ucl.ac.uk

†Jonathan P. Roiser and Rebekah Wigton have contributed equally to this work.

Cognitive impairment is common in patients with schizophrenia, and even those with relatively preserved function perform worse than healthy volunteers (HVs) on attentional tasks. This is consistent with the hypothesis that connectivity – in the frontoparietal network (FPN) activated during attention – is disrupted in schizophrenia.We examined attentional effects on connectivity in the FPN, in schizophrenia, using magnetoencephalography (MEG).Twenty-three HVs and 19 first-episode schizophrenia patients were scanned during a simple visual change test, known to activate the FPN, in which attention was monitored and directed with an orthogonal flicker-detection task. Dynamic causal modeling (DCM) of evoked responses was used to assess effective connectivity – and its modulation by changes in the attended stimulus dimension – in the following network: higher visual area; temporoparietal junction (TPJ); intraparietal sulcus (IPS); dorsal anterior cingulate cortex; and ventrolateral prefrontal cortex. The final MEG analysis included 18 HVs and 14 schizophrenia patients. While all participants were able to maintain attention, HVs responded slightly, but non-significantly, more accurately than schizophrenia patients. HVs, but not schizophrenia patients, exhibited greater cortical responses to attended visual changes. Bayesian model comparison revealed that a DCM with attention dependent changes in both top-down and bottom-up connections best explained responses by patients with schizophrenia, while in HVs the best model required only bottom-up changes. Quantitative comparison of connectivity estimates revealed a significant group difference in changes in the right IPS-TPJ connection: schizophrenia patients showed relative reductions in connectivity during attended stimulus changes. Crucially, this reduction predicted lower intelligence.These data are consistent with the hypothesis that functional dysconnections in the FPN contribute to cognitive impairment in schizophrenia.

**Keywords: schizophrenia, frontoparietal, magnetoencephalography, dynamic causal modeling, dysconnectivity, DCM**

## **INTRODUCTION**

Patients with schizophrenia exhibit reliable impairments on almost all cognitive tests (1). They exhibit particular impairments on tests that depend on attention (2, 3) and working memory (4–6). This is a pattern observed even in patients whose cognitive performance is otherwise in the normal range. Generalized cognitive impairment, indexed by intelligence quotient (IQ), precedes the onset of psychosis and is detectable as far back as infancy (7). In addition, the severity of impairment is linearly related to the risk of developing psychosis and predicts functional outcome following illness onset (7, 8). Furthermore, supporting the notion that cognitive impairment represents a core feature of the syndrome in addition to positive and negative symptoms, it has also been reported in unaffected first degree relatives of patients with schizophrenia (9–15). This suggests that the neural abnormalities underlying impaired cognition reflect the neurodevelopmental susceptibility to schizophrenia. Cognitive impairment is thus a core feature of schizophrenia that significantly impacts on the course of illness; warranting a more detailed understanding of its generalized nature and neurobiological basis.

Functional neuroimaging has highlighted a core network of prefrontal and parietal regions that is activated during the performance of a variety of disparate cognitive tasks. For example, this frontoparietal network (FPN) is activated in tasks ranging from set-shifting (16) to feature selection (17) and object orientation (18) as well as recognition memory (19), working memory (20) and attention [for reviews see Ref. (21–23)]. One hypothesis is that the FPN serves as a "multiple demand system" which undertakes the information processing requirements common to different cognitive tasks, for example by updating and maintaining

changes to attended stimuli (24). This system is supported by various attention networks such as the orientation network (25, 26) as well as global connectivity within this network (27). Furthermore, the development, and thus differences, of these specific attention systems – particularly in the FPN – are driven by interactions between cognition and our environment (28). This is further supported by the findings of an fMRI study by Hon et al. (29) using a task in which subjects were instructed to attend to a subset of visual stimuli and press a button when they detected a brief "flicker." This task robustly activated the FPN simply when attended visual stimuli changed, independent of any flickers or responses. Importantly, interposed unattended stimulus changes did not elicit FPN activation, thus implicating involvement of FPN specifically in task-relevant information processing.

The "multiple demand" hypothesis predicts that dysfunction of the FPN will adversely impact on the performance of a wide range of cognitive tasks. In schizophrenia, fMRI studies have shown that nodes in this network respond abnormally across different tasks such as attention, executive control, and working memory (19, 30, 31). However, no studies in schizophrenia have examined FPN integrity in the context of a task that recruits a fundamental attentional process – a process that is essential for the performance of many higher order tasks but which itself requires "minimal decision and control" (29, 32). We therefore investigated FPN function in schizophrenia using the task described by Hon et al. (29).

Regarding possible pathophysiological mechanisms underlying schizophrenia, the "dysconnection" hypothesis is one of the most influential (33–35). The central tenet of this model is that the symptoms observed in schizophrenia arise from abnormal regulation of *N*-methyl-d-aspartate (NMDA) receptor-dependent synaptic efficacy by modulatory transmitters, such as dopamine or acetylcholine (35). This is proposed to result in abnormal integration of neural processes, which can be measured in terms of effective connectivity, or the influence that one neural system has over another (36). Using dynamic causal modeling (DCM) – which estimates effective connectivity – prior studies have found that aberrant perceptual processes in schizophrenia are associated with such dysconnectivity (37, 38). These differences in effective connectivity are thought to reflect aberrant NMDA receptor-dependent regulation of synaptic efficacy; similar dysfunction may also occur in the FPN, resulting in dysfunctional processing in the multiple demand system and thus impairment on several different types of cognitive tests. Furthermore, disrupted connectivity assessed using DCM has also been reported during tasks that examined working memory and verbal fluency in individuals at risk for psychosis (39–41). However, most previous studies assessing differences in connectivity employed functional magnetic resonance imaging (fMRI), which precludes the use of detailed and neurophysiological plausible (neural mass) models, due to its low temporal resolution.

In this study we therefore measured neural responses in patients with schizophrenia and healthy volunteers (HVs) with magnetoencephalography (MEG) during a simple visual change task known to engage the FPN (29); we assessed effective connectivity in the FPN with DCM. MEG, in particular, provides more temporally precise data than fMRI. This detailed time course information enables the efficient estimation of much more realistic and detailed neural mass models – particularly models with different types of connection (i.e., forward; backward; and lateral) among distinct neuronal populations. Cortical areas previously reported to be activated by the task employed (using fMRI) were used as regions of interest (ROIs) when specifying the different neural network models for DCM [see Figure 2 and Table 2 of Ref. (29)]. DCM uses these co-ordinates as spatial priors when specifying the location of the electromagnetic sources that generate sensor-space responses. With these models, we were able to estimate how effective connectivity between nodes in the FPN – and its modulation by attended visual changes – differs between patients with schizophrenia and HVs. We predicted that patients with schizophrenia would show abnormal effective connectivity (and its modulation by attention), as reflected in the coupling estimates from DCM, and – crucially – that these would predict cognitive impairment (i.e., low IQ).

## **MATERIALS AND METHODS**

#### **PARTICIPANTS**

Nineteen patients with schizophrenia were recruited from a longitudinal study of cognitive impairment in first-episode psychosis (6). The diagnosis was ascertained using a structured interview, the diagnostic module of the Diagnostic Interview for Psychosis (42), which includes items from the Operational Criteria Checklist for Psychosis (43) and the World Health Organisation Schedules for Clinical Assessment in Neuropsychiatry (44). Diagnoses were made at study entry and reviewed 1 year later. Twenty-three HVs were recruited via advertisement.

Exclusion criteria for all participants included: medical disorders likely to cause cognitive impairment; IQ less than 70; lefthandedness; and recent illicit substance use. Additional exclusion criteria for HVs included: past or present psychiatric or neurological disorders; and any first degree relatives with a psychotic illness. All HVs were screened for Axis-I psychopathology using the Mini International Neuropsychiatric Interview (45). One healthy volunteer was excluded because he had a brother with psychosis; another was excluded for being left-handed. Three HVs and two patients with schizophrenia were excluded because we were unable to identify clear visual evoked responses. Two patients with schizophrenia were excluded due to failure to follow task instructions. One patient with schizophrenia was excluded because her diagnosis was subsequently changed to depression with psychosis. After exclusions, the final MEG analysis included 18 HVs (9 male, 9 female) and 14 patients with schizophrenia (10 male, 4 female), all right-handed.

Demographic and clinical data are presented in **Table 1**. The groups were similar in terms of age, premorbid IQ [estimated using the National Reading Test: (46)], and current IQ [estimated using the Wechsler scale of adult intelligence (WAIS)-III] (47). All patients but one were taking antipsychotic medication; four were also taking antidepressant medication; and one was only taking antidepressant medication. None of the HVs were taking any psychotropic medication. Symptom severity scores for the patients were measured using the Scale for Assessment of Negative Symptoms (SANS) (48) and the Scale for Assessment of Positive Symptoms (SAPS) (49). All participants provided written informed consent before the start of the experiment and were



IQ, intelligence quotient; M, male; F, female.

WAIS,Wechsler adult intelligence scale; NART, national adult reading test; CPZ, chlorpromazine equivalent (72, 73); SANS, scale for assessment of negative symptoms; SAPS, scale for assessment of positive symptoms.

compensated for participation. The study was approved by Ealing and West London Local Research Ethics Committee.

## **COGNITIVE ACTIVATION PARADIGM**

The paradigm was presented as described in Hon et al. (29), with the exception that all stimuli were black and white. Participants were presented with a series of complex, nonsense shapes, similar to those shown in **Figure 1**. The images were displayed to the left, right, top and bottom of a central box. Each stimulus on either the vertical or horizontal axis was a mirror reflection of the opposite image. The shapes were hand drawn to ensure that they did not resemble any familiar patterns or objects. Each stimulus was presented for a period of 1000 ms before an off period of 500 ms, during which only the central box appeared. All the images and lines appeared and disappeared at the same time. This task comprised two blocks, each consisting of 180 trials, and lasted four and a half minutes per block. For each trial, either: (1) the shapes on the horizontal axis changed, in relation to the previous trial but the shapes on the vertical axis remained the same (60 occurrences); or (2) the shapes on the vertical axis changed and the shapes on the horizontal axis remained the same (60 occurrences); or (3) no shape change occurred on either axis (60 occurrences). On onethird of trials (20 in each condition) either the horizontal or the vertical axis flickered briefly. Trial types were presented in a random order. In order to maintain stimulus novelty, no shape was reused in each block.

This task was chosen to be sufficiently straightforward so that patients with schizophrenia would be able to perform it to a high level of accuracy. Participants were instructed to maintain fixation

on the box in the center of the screen, which was confirmed using concurrent eye-tracking. During each run, participants were instructed to watch either the horizontal axis or the vertical axis, and press a button on an MEG-compatible button-box with their right index finger only when a flicker occurred on the axis they were attending to. The order was counterbalanced across participants. Two runs were conducted for each participant, each lasting 4.5 min.

## **MAGNETOENCEPHALOGRAPHY**

#### **MEG data acquisition**

Magnetoencephalography data were recorded using 275 first order axial gradiometers with the Omega275 CTF MEG system (VSMmedtech, Vancouver, Canada) at a 480 Hz sampling frequency. Data were recorded in a magnetically shielded room. Each participant's head was supported by a padded headrest to restrict head movement during recording. To monitor head motion, sensors were placed near the ears and nose. None of the participants included in the final analysis exceeded a threshold of 5 mm for head movement within a single run. Eye-tracking was also used to ensure that all participants attended continually to the visual stimuli.

### **MEG data analysis**

Data analysis was performed using Statistical Parametric Mapping 12 (SPM12) (Wellcome Trust Centre for Neuroimaging, London, UK; www.fil.ion.ucl.ac.uk/spm) in MATLAB 12.1 (MathWorks Inc., Sherbon, MA, USA). Blocks with attended flicker-detection (hit) rates lower than 60% or incorrect response (false alarm) rates greater than 15%, were excluded from the analysis. The remaining data were epoched with a peristimulus window from −100 to 1300 ms. Data were processed using artifact detection routines to identify and exclude eye blinks or movement before bandpass filtering between 0.5 and 16 Hz. The data were then down sampled to 200 Hz and baseline corrected using the −100–0 ms period. One bad MEG channel was excluded from the analysis across all participants.

All flicker and false alarm trials (i.e., if the participant made a response when the attended axis had not flickered) were removed from further analysis. Robust averages were calculated for each event type (attended change, unattended change and no change) across all valid trials for each session – note that no trial included in the analysis featured a flicker stimulus or any response by the participant. Grand average responses were created for each participant for all sessions that met the performance criteria specified above. These average responses were interpolated into images and visually inspected for artifacts and the presence of visual evoked fields (VEFs). Data were excluded if a clear VEF could not be identified over occipital sensors. From these grand average images, contrast images were created for the following comparisons: (attended change minus unattended change); (attended change minus no change); and (unattended change minus no change). A smoothing kernel of (2 mm × 2 mm × 2 mm) was applied to each contrast image.

These sensor-space contrast images were combined at the group-level for random-effects analysis, which was performed across all sensors and all time points. For group comparisons, significant clusters were defined as those surviving an uncorrected voxel-level threshold of *P* < 0.001 with a cluster-level threshold corrected across the whole of sensor-space and peristimulus time, controlling the family-wise error (FWE) rate at *P* < 0.05. For within-group contrasts, which produced extensive activation in HVs, we used an uncorrected voxel-level threshold of *P* < 0.0005, again controlling the FWE rate at *P* < 0.05 at the cluster-level.

## **Dynamic causal modeling**

Dynamic causal modeling was applied to assess effective connectivity. DCM uses the concept of effective connectivity, or the influence that one neural system has over another,to create a model of coupled neuronal populations that is used to explain evoked responses (50). The parameters (effective connectivity and other synaptic constants) are optimized by fitting responses generated by the model – in response to stimuli – to observed responses using standard Bayesian model inversion techniques (51). In addition, the evidence for a particular model (irrespective of the particular parameters) is evaluated in terms of model evidence through Bayesian model comparison (BMC) (52, 53). Crucially, DCM estimates not just the effective connectivity or coupling between sources of electromagnetic responses but also sets of experimental changes in coupling. This allows one to use BMC to assess the evidence for context dependent changes in connectivity – such as changes induced by the nature of the stimulus (attended versus unattended). This method incorporates an exceedance probability to determine the best fitting model or the likelihood that one model fit the data better than any of the other models.

The network architecture used for the DCM comprised sources that were previously identified as being activated in an fMRI study employing the same task (29). These sources were consistent with the most robust responses in the (attended change minus unattended change) analysis of sensor-space responses [see Figure 2 and Table 2 of Ref. (29)]. They included: higher visual area (HVA) ([48, −66, −4] and [−48, −66 −16]); temporoparietal junction (TPJ) ([64, −38, 6] and [−64, −38, 6]); intraparietal sulcus (IPS) ([26, −62, 42] and [−24, −66, 50]); ventrolateral prefrontal cortex (vlPFC) ([36, 35, −4] and [−44, 34, 6]); and dorsal anterior cingulate (dACC) ([14, 26, 44] and [−6, 14, 48]). All models had a fixed model architecture within each hemisphere (both forward and backward connections) as follows: between HVA and TPJ; between TPJ and IPS; between IPS and dACC; and between IPS and vlPFC. Furthermore, all models had fixed lateral connections between vlPFC and dACC, and inter-hemispherically between all corresponding regions (i.e., between left and right HVA, TPJ, IPS, vlPFC, and dACC) (see **Figure 2**). Using the above co-ordinates as spatial location priors, three DCMs were constructed, varying in relation to modulatory effects – the modulation attributable to a stimulus change on the attended dimension or axis – on forward connections (reflecting bottom-up effects), on backward connections (reflecting top-down effects), or on both forward and backward connections (see **Figure 2**). This approach is consistent with prior studies using DCM for MEG (38, 54). All models had driving inputs into the HVA, modeling subcortical visual input.

Each DCM was estimated at the subject level, but BMC was conducted at the group-level, in each group separately, to determine the most parsimonious explanation for the data in terms of attentional modulation of connectivity (models with the most evidence represent an accurate explanation of observed data with minimal complexity). To make quantitative inferences about differences in connection strengths, effective connectivity, and its modulation were compared using classical statistics at the group-level. Fixed and modulatory coupling estimate parameters (indicating the

strength of each fixed connection and its modulation by attended visual change, respectively) were computed using random-effects Bayesian model averaging (BMA) (55), and submitted to analysis in SPSS 17 (SPSS Inc., Chicago, IL, USA). After BMA, we obtained connectivity parameters for all fixed and modulatory connections (only forward and backward) for each hemisphere as well as the intrahemispheric connections: total 46 parameters. However, we did not examine any of the intrahemispheric connections so this left a total of 36 connections that were examined for group differences.

Group differences were assessed with *t*-tests and relationships with demographic, clinical, and cognitive variables with Pearson's*r* correlation coefficients. These tests were not corrected for multiple comparisons. Since this was an exploratory analysis, and because our primary inference related to the BMC, an alpha level of 0.05 was adopted for all group comparisons. We refer to trend significance at a threshold of 0.05 < *P* < 0.1. However, to reduce the number of correlations performed, only those coupling parameters that differed between the groups at least at a trend level were entered into correlation analyses.

## **RESULTS**

#### **BEHAVIORAL PERFORMANCE**

Patients with schizophrenia responded slightly, but nonsignificantly, less accurately than HVs, with a lower hit rate (**Table 2**). There were no significant differences between the groups in terms of reaction times, false alarms or *d* 0 . Concurrent eyetracking confirmed that all participants attended to the visual stimuli throughout the task.

#### **MEG SENSOR-SPACE RESPONSES**

The MEG grand average image in the healthy controls (**Figures 3A–C**) showed a consistent and sustained deviation in cortical responses between attended and unattended stimulus changes from approximately 200 ms after stimulus onset,primarily in right lateral frontal sensors. This increase was evident for almost 1 s. This deviation becomes statistically significant (cluster-level correction) at approximately 375 ms over right frontal sensors [peak at 395 ms: *t*(17) = 4.87, *P*FWE = 0.039]. The response at this peristimulus time is consistent with the P300 response often observed in oddball paradigms (56, 57), of which our task is a variant (with an attentional manipulation). Similar effects were observed later in peristimulus time over the same sensors, and over more anterior frontal sensors (see **Table 3**).

By contrast, the MEG responses from the patients with schizophrenia showed no difference between the attended change and unattended change conditions (**Figures 3D–F**). This finding was paralleled in the statistical analyses, where no differences in sensorspace responses were observed between any of the conditions, even when the threshold was lowered to *P* < 0.001 (uncorrected), extent threshold 10 voxels. However, patients with schizophrenia did, nonetheless, show a clear VEF across all conditions (**Figure 3F**), suggesting that they did engage with the task and have detectable electromagnetic responses.

Although patients with schizophrenia exhibited diminished cortical responses to attended visual change at frontal sensors in comparison to the controls (**Figures 3B,E**), no group differences survived stringent correction for multiple comparisons across the whole of sensor-space and peristimulus time at any time point (**Table 3**). However, at a more liberal threshold of *P* < 0.001 (uncorrected), extent threshold 10 voxels, diminished responses in the schizophrenia patients were detected over right frontal sensors at 260 ms [*t*(30) = 3.78, *P* < 0.001, uncorrected], 320 ms [*t*(30) = 3.50, *P* < 0.001, uncorrected], 360 ms [*t*(30) = 3.55, *P* < 0.001, uncorrected], and 375 ms [*t*(30) = 3.60, *P* < 0.001, uncorrected], and over left parietal sensors 755 ms


#### **Table 2 | Behavioral data from the visual change task.**

SD, standard deviation; ms, milliseconds.

[*t*(30) = 3.85, *P* < 0.001, uncorrected] (see **Table 3**). We report these results descriptively, noting that they require replication.

Neither group showed any activation in the (unattended change minus no change) contrast at a threshold of *P* < 0.0005 uncorrected, extent threshold 40 voxels. The (attended change minus no change) contrast produced similar results to the (attended change minus unattended change) contrast.

#### **DYNAMIC CAUSAL MODELING**

#### **Bayesian model comparison**

Using BMC, we were able to determine which model provided the most parsimonious explanation for the effects of attended stimulus changes in both groups. Different DCMs provided the best explanation in patients with schizophrenia and HVs, providing *prima facie* evidence for dysconnectivity during attentional processing in schizophrenia. In HVs, there was greatest evidence for the model with only forward modulations, with an exceedance probability of 99% over other models. In patients with schizophrenia there was greatest evidence for the model with both forward and backward modulations, with an exceedance probability of 75% over the other models (**Figure 4**).

#### **Quantitative connectivity estimates**

To understand better why the groups differed in terms of the most parsimonious architecture, we quantified the connectivity and its attentional modulation, using BMA (**Table 4**). This effectively weights the coupling estimates, under each model, according to the probability of that model and accommodates uncertainty about the underlying architectures.

*Fixed connections.* For the fixed connections, the left IPS-dACC fixed forward connection was marginally stronger in the HVs [*t*(30) = 2.019, *P* = 0.052]. This means that HVs show a trend toward being relatively more sensitive to parietal afferents to the dACC than schizophrenia patients, irrespective of whether the stimulus change was on the attended dimension or not.

*Modulatory connections.* For the modulatory connections (i.e., altered coupling elicited by attended visual change), only the right IPS-TPJ modulation differed significantly between the groups [*t*(30) = 2.428, *P* = 0.021]. This modulator was negative in schizophrenia and slightly positive in healthy subjects. In other words, the top-down afferents from the parietal source to the TPJ were reduced in their strength in schizophrenia, relative to HVs, only when the stimulus changed in the attended dimension. Note that

a negative modulation corresponds to a reduction in connectivity (because connection strengths in DCM for electromagnetic responses are always positive – targeting excitatory and inhibitory neuronal populations).

There was also a trend toward a group difference in the right IPS-dACC modulation [*t*(30) = 1.701, *P* = 0.099], with a negative modulator in HVs but a positive modulator in schizophrenia patients.

#### **Correlations with demographic, clinical, and cognitive measures**

No correlations approaching significance were identified between the left IPS-dACC fixed connection, and behavior, current IQ, premorbid IQ, age, SAPS, SANS, or chlorpromazine equivalent dose. However, changes in the IPS-TPJ connectivity correlated positively with premorbid IQ score in the patient group (*r* = 0.547, *P* = 0.043) but not in the HVs (*r* = −0.297, *P* = 0.231; difference in correlation coefficients: Fisher's *Z* = 2.32, *P* = 0.02) (**Figure 5**). In other words, patients with the most (abnormal) decrease in this connection during attended stimulus change had the lowest premorbid IQ. Importantly, this modulation did not correlate significantly with the percentage of correct flicker-detections (*r* = 0.127, *P* = 0.665, in patients with schizophrenia and *r* = -0.250,*P* = 0.318 in HVs), suggesting that this difference does not simply reflect poor engagement in the task.

## **DISCUSSION**

Using DCM, we found that patients with schizophrenia and HVs differ in their recruitment of the FPN during the processing of salient (attended) visual changes. Both inference about the architectures – and the connectivity of those architectures – pointed to an abnormality in top-down modulation of stimulus evoked responses, when stimulus changes were in the attended dimension or axis. In patients with schizophrenia the winning model featured modulation of both backwards (top-down) and forwards (bottom-up) connections. Specifically, patients with schizophrenia showed a relative failure (decreased connection strength) of the backward IPS-TPJ connection when processing attended visual changes.

Interestingly, this quantitative reduction was associated with lower premorbid IQ in schizophrenia patients only. In previous studies we showed that lower premorbid and current IQ (markers of general cognitive ability) at illness onset predicted poorer social function after 4 years (8). Another study also found that lower IQ and higher psychotic symptoms correlated with a loss in gray matter tissue (58). Previous studies also found that

global connectivity within the FPN correlated negatively with IQ (27). We have also reported that in patients with schizophrenia, but not healthy controls, lower premorbid and current IQ was related to reduced frontotemporal cortical area and predicted progressive parietal cortical thinning over the following 3 years (59, Gutiérrez-Galve, unpublished observations). Although in the present study we did not find relationships with current IQ, taken together our findings support the hypothesis that FPN dysfunction is the basis of generalized cognitive impairment in schizophrenia (60).

#### **Table 3 | Results for contrast images for MEG sensor-space responses.**


Unattended > attended change for schizophrenia patients > healthy volunteers<sup>b</sup> No clusters survived threshold

<sup>a</sup>Analyses within groups thresholded at P < 0.0005, cluster size 40 voxels.

<sup>b</sup>Analyses between groups thresholded at P < 0.001, cluster size 10 voxels.

\*P < 0.05 FWE whole-brain cluster-level corrected.

k = number of contiguous voxels.

Differences in connectivity were observed between the groups in several parts of the FPN, all involving parietal cortex connections. The left forward (bottom-up) IPS-dACC fixed connection showed a trend toward being stronger in HVs. This would be consistent with a relatively inefficient transmission of information from parietal to frontal regions in schizophrenia, and could explain the lack of frontal cortical activity elicited by attended visual changes in patients. We also identified a trend difference in the modulation of the right IPS-dACC forward (bottom-up) connection, which was negative only in HVs. However, we note that neither of these differences achieved statistical significance. The only effect to differ significantly between the groups was in the right IPS-TPJ backward (topdown) connection: this connection was lower in patients with schizophrenia.

These results were consistent with the sensor-space results which indicated reduced cortical responses to attended visual changes in patients with schizophrenia, consistent with prior reports of attenuated P300 responses in this group (61). Our DCM suggests that these attenuated responses are caused by selective differences in coupling between the TPJ, parietal cortex, and anterior cingulate. Importantly, the MEG results we observed are unlikely to reflect solely a lack of engagement with the task, as all patients included in the analysis could perform at a high level and, as a group, did not differ from the HVs in terms of reaction times or false alarms; although, we did observe a slight impairment in percentage correct flicker-detection. However, this measure did not correlate with any of the DCM parameters. Together, these results are consistent with the dysconnection hypothesis of schizophrenia – which proposes that aberrant cortical coupling arises due to aberrant regulation of NMDA-dependent synaptic plasticity. Within the FPN we speculate that this may lead to differences in the efficiency in which regions of the FPN are able to interact, resulting in altered effective connectivity, as assessed with MEG. Furthermore, these fit findings comfortably with recent computational accounts of dysconnectivity as a failure to adjust or contextualize

the synaptic efficacy that underlies the routing of salient or precise sensory information (62).

It is interesting to interpret our results in relation to computational formulations of the dysconnection hypothesis – in particular, predictive coding models of hierarchical Bayesian inference. In these formulations, psychotic symptoms are understood in terms of the false inference that arises when the salience or precision afforded sensory information is aberrant (62–64). Formally, top-down or descending predictions in cortical hierarchies try to explain representations at lower levels by forming prediction errors at each level of the hierarchy. Ascending prediction errors are then passed forward to improve the representations at higher levels – and thereby minimize prediction error at every hierarchical level. Crucially, the influence prediction errors have on high level representations depends upon their precision (inverse variance or reliability), which itself has to be optimized (62, 65). Physiologically, precision is thought to be encoded by the postsynaptic neuromodulatory gain of superficial pyramidal cells reporting prediction error (66). Psychologically, the optimization of precision provides a simple explanation for attentional gain; in other words, the selection of ascending prediction errors that are considered salient of precise in any given context (65).

This formulation is particularly relevant here, for two reasons: first, it provides a computational account of dysconnection – that can be reduced to an abnormality of message passing during perceptual inference that rests explicitly on the aberrant modulation of synaptic efficacy (62). Second, it speaks to the important role of attentional deficits in disclosing this aberrant modulation.

Although speculative, it is tempting to interpret our results from this computational perspective. The results of the DCM in HVs are entirely consistent with predictive coding; in that attention endows ascending prediction errors from lower (temporal) regions with greater precision and preferential access to higher (parietal) regions – resulting in an increase in forward effective connectivity (67). However, in schizophrenia it appears that the descending predictions – that are informed by ascending prediction errors – "fall on deaf ears," when descending to the temporal region. This is evidenced by a reduction in the backwards connection. Although this account is somewhat heuristic, it is consistent with the fact that backward connections target superficial cortical layers that are rich in NMDA receptors – receptors that are crucial for neuromodulatory effects and play a central role in the dysconnection hypothesis (35).

Prior studies have highlighted the importance of the interaction between the IPS and the TPJ in the detection of behaviorally salient events (23). The interaction between these two areas can be split between two proposed FPN systems: the dorsal FPN, which is responsible for the orientation and maintenance of selective attention (26, 68); and the ventral FPN, which acts to direct attention to salient events (23) – and aids in the application of attentional set (69). The dorsal FPN also plays an important role in the synchronization of activity between the visual cortex and other areas of the dorsal FPN as a means of mediating top-down visual attention (69) to exert control over tasks from bottom-up sensory signals (26). It is thought that these networks are co-activated during stimulus-driven reorientation when a salient and behaviorally relevant event occurs, and is highly right-lateralized (69). Differences in integration between these systems of the FPN are evident in the connection between the IPS and TPJ; where the IPS plays a greater role in the dorsal FPN and the TPJ is more involved in the ventral FPN (23, 69). The reduction in the strength in this connection in patients with schizophrenia during attended stimulus change could be interpreted as a relative failure of functional integration between these two regions. Furthermore, a failure of this backward (top-down modulation) in attentional processing of salient events may thus explain why patients with schizophrenia recruit a more complicated connectivity architecture during the processing of visual change. Due to this connection's role in both the dorsal and ventral FPN, we speculate this relative failure in top-down connectivity might contribute to the difficulty that patients with schizophrenia face in modifying behavior in response to salient stimuli, though this requires testing in future studies.

The groups included in this study were well matched in terms of their task performance and demographic variables. Although patients with schizophrenia made slightly fewer correct responses when detecting flickers, the false alarm rate and reaction times were similar to those of the healthy volunteer group, and the average performance in schizophrenia patients exceeded 75%. Despite performing well above chance and close to controls in being able to detect attended stimulus flickers, the activation of the FPN in schizophrenia patients to attended visual change was diminished. The implication of this finding is that the updating of attended information necessary to perform more complex tasks would be compromised due to dysconnectivity within this network, which could account for poor cognitive function more generally.

#### **Table 4 | Connectivity estimates from Bayesian model averaging.**


HVA, higher visual area; TPJ, temporoparietal junction; IPS, intraparietal sulcus; dACC, dorsal anterior cingulate cortex; vlPFC, ventrolateral prefrontal cortex. \*Trend toward significance (P < 0.1).

\*\*P < 0.05.

Several limitations of our study merit comment. The first is the low number of participants that we were able to include in the final analysis. This affects the statistical sensitivity (Type-II error) of our analyses. It is possible that there are other differences between the groups that we did not have sufficient sensitivity to detect. Second, the relatively preserved task performance we observed is not

common to most studies of attention and working memory in schizophrenia (70), and indicates that the patients included in the present study were high-functioning. This subgroup of patients may not exhibit as extreme FPN network dysconnectivity as other groups, which would also raise the chance of Type-II error, and means that our results may not generalize to other schizophrenia patients. Third, patients with schizophrenia did exhibit a slight, non-significant impairment in their ability to detect flickers. It is therefore possible that the results that we report reflect a lack of attentional engagement during the task. However, the presence of a clear VEF in each MEG – as well as concurrent monitoring of eye-gaze – suggests that patients with schizophrenia did indeed engage with the task. Moreover, there was no relationship between DCM parameters and sensitivity to detect flickers. Fourth, all but one of the schizophrenia patients were taking antipsychotic medication, raising the possibility that group differences were either caused by or even masked by medication. A recent paper using DCM supports the latter possibility (71). In that study, both individuals at risk for psychosis and in a firstepisode exhibited FPN dysconnectivity (assessed using fMRI); but this was normalized by antipsychotic medication. However, in the present study we did not detect any correlation between chlorpromazine equivalent dosage and behavioral performance or DCM coupling estimates. Finally, it is important to note that these results require replication, especially as the results for the between-group comparison in sensor-space did survive correction for multiple comparisons.

In summary, these data support the notion of FPN dysconnectivity in schizophrenia. This exists despite patients with schizophrenia being able to engage in the task and perform at a high standard on average. This dysconnectivity was mainly reflected in a reduction in top-down connectivity between the right IPS and TPJ in patients when processing attended stimuli. This reduction was associated with low premorbid IQ in the schizophrenia group, and may indicate aberrant integration between the dorsal and ventral components of the FPN. Future work should investigate the association between FPN dysconnection and impairment on specific neurocognitive measures, and assess the impact of antipsychotic medication.

## **ACKNOWLEDGMENTS**

The authors thank John Duncan for helpful discussions. This study was funded by the Wellcome Trust (programme grant 064607/Z/01/Z to EMJ and Principal Research Fellowship 088130/Z/09/Z to KJF), and EMJ was funded by the UCLH Biomedical Research Centre. We are grateful to Isobel Harrison and Libby Matheson from the West London team for help with patient recruitment.

## **REFERENCES**


from the bipolar-schizophrenia network on intermediate phenotypes (B-SNIP) study.*Am J Psychiatry* (2013) **170**:1275–84. doi:10.1176/appi.ajp.2013.12101298


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 October 2013; accepted: 09 December 2013; published online: 24 December 2013.*

*Citation: Roiser JP, Wigton R, Kilner JM, Mendez MA, Hon N, Friston KJ and Joyce EM (2013) Dysconnectivity in the frontoparietal attention network in schizophrenia. Front. Psychiatry 4:176. doi: 10.3389/fpsyt.2013.00176*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Roiser, Wigton, Kilner, Mendez, Hon, Friston and Joyce. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Connectivity, pharmacology, and computation: toward a mechanistic understanding of neural system dysfunction in schizophrenia

#### **Alan Anticevic 1,2,3,4,5,6\*, MichaelW. Cole<sup>7</sup> , Grega Repovs <sup>8</sup> , Aleksandar Savic <sup>9</sup> , Naomi R. Driesen<sup>1</sup> , GenevieveYang1,4,Youngsun T. Cho<sup>1</sup> , John D. Murray <sup>10</sup>, David C. Glahn1,5, Xiao-JingWang<sup>10</sup> and John H. Krystal 1,2,3,11**

<sup>1</sup> Department of Psychiatry, Yale University School of Medicine, New Haven, CT, USA


#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Martin Walter, Otto-von-Guericke-Universität Magdeburg, Germany Bernhard J. Mitterauer, Voltronics-Institute for Basic Research, Psychopathology and Brain Philosophy, Austria Andrei Manoliu, Technische Universität München, Germany

#### **\*Correspondence:**

Alan Anticevic, Department of Psychiatry, Yale University, 34 Park Street, New Haven, CT 06519, USA e-mail: alan.anticevic@yale.edu

Neuropsychiatric diseases such as schizophrenia and bipolar illness alter the structure and function of distributed neural networks. Functional neuroimaging tools have evolved sufficiently to reliably detect system-level disturbances in neural networks.This review focuses on recent findings in schizophrenia and bipolar illness using resting-state neuroimaging, an advantageous approach for biomarker development given its ease of data collection and lack of task-based confounds. These benefits notwithstanding, neuroimaging does not yet allow the evaluation of individual neurons within local circuits, where pharmacological treatments ultimately exert their effects. This limitation constitutes an important obstacle in translating findings from animal research to humans and from healthy humans to patient populations. Integrating new neuroscientific tools may help to bridge some of these gaps. We specifically discuss two complementary approaches. The first is pharmacological manipulations in healthy volunteers, which transiently mimic some cardinal features of psychiatric conditions. We specifically focus on recent neuroimaging studies using the NMDA receptor antagonist, ketamine, to probe glutamate synaptic dysfunction associated with schizophrenia. Second, we discuss the combination of human pharmacological imaging with biophysically informed computational models developed to guide the interpretation of functional imaging studies and to inform the development of pathophysiologic hypotheses. To illustrate this approach, we review clinical investigations in addition to recent findings of how computational modeling has guided inferences drawn from our studies involving ketamine administration to healthy subjects.Thus, this review asserts that linking experimental studies in humans with computational models will advance to effort to bridge cellular, systems, and clinical neuroscience approaches to psychiatric disorders.

**Keywords: schizophrenia, pharmacology, functional connectivity, computational modeling, thalamus, NMDA receptors, glutamate**

## **INTRODUCTION**

The human brain is a complex, dynamic system with computations occurring at several levels of organization, from individual synapses to networks that span multiple brain regions. These large-scale neural systems ultimately produce complex behaviors that are profoundly altered in the context of psychotropic drug administration or neuropsychiatric disease.

One example is schizophrenia – a common, multi-faceted, and heterogeneous neuropsychiatric syndrome (1) associated with disturbances in perception (2), belief (3), emotion (4), and cognition (5). A number of theoretical models of schizophrenia suggest that the clinical features of this disorder emerge from disturbances in neural connectivity and deficits in synaptic plasticity (6). Progress in explicating the pathophysiology of schizophrenia has been slowed by our limited understanding of the neurobiology of schizophrenia, shortcomings of animal models for this disorder, and the challenge of translating basic and clinical research approaches to this disorder. This knowledge gap has constrained our ability to develop new and more effective pharmacotherapies for schizophrenia (7), accounting for little change in the public health impact of this disease over the past two decades (8).

Disturbances in the structural and functional connectivity of the cerebral cortex are thought to be central to the neurobiology of schizophrenia (6) and are thought to impair the function of large-scale neural systems (9–13). Efforts have been made to reconcile these system-level observations with the cellular neuropathology of schizophrenia. One leading mechanistic model proposes that glutamate synaptic abnormalities associated with schizophrenia, mimicked in part by the effects of drugs that block the *N*-methyl-d-aspartate glutamate receptor (NMDAR) (7), disturb the local balance of excitation and inhibition and thereby contribute to alterations in large-scale neural system functional connectivity (14). This influential hypothesis is based on a key observation: sub-anesthetic doses of non-competitive NMDAR antagonists such as ketamine produce symptoms resembling those of schizophrenia in healthy humans (15). There is also growing evidence from pre-clinical (16), post-mortem (17), neuroimaging (18), and pharmacological experiments (15) illustrating that abnormal NMDA receptor function may be one pathophysiological mechanism occurring in people with schizophrenia. Alterations in synaptic function of gamma-aminobutyric acid (GABA) (19, 20) and dopamine (21) have also been implicated in schizophrenia, which likely affect local circuit computations (22). Collectively, this work provides insight into how disturbances in synaptic glutamate signaling (among other neurotransmitters) could contribute to producing schizophrenia symptoms (7). Nevertheless, there remains an important explanatory gap between mechanistic cellular-level hypotheses of schizophrenia and non-invasive neuroimaging studies that characterize the function of neural systems.

Functional neuroimaging has consistently revealed both region-specific and network alterations in schizophrenia across a number of cognitive measures (23). For instance, there is now substantial evidence suggesting profound alterations in networks supporting complex cognition in schizophrenia [e.g., working memory (WM) (24–26)]. This work has been complemented by a parallel focus on characterizing functional connectivity alterations in schizophrenia (27). Resting-state functional connectivity is based on the analysis of low-frequency fluctuations present in the blood-oxygenation-level-dependent (BOLD) signal (28, 29). These low-frequency fluctuations have been shown to be temporally correlated within spatially distinct but functionally related networks (30), establishing an intrinsic functional architecture (31) across species (32). The functional networks identified at rest also are correlated with other measures of structural and functional connectivity in healthy populations (33) and allow for characterization of distributed circuit abnormalities in neuropsychiatric illness (34, 35). Such approaches have been successfully applied to the study of schizophrenia and have increasingly revealed neuroimaging markers of this complex illness (12, 27, 36– 38), which may be consistent with established theoretical models of this disease (39).

This review first focuses on resting-state functional connectivity MRI (rs-fcMRI) studies of schizophrenia that are providing distinct insights into cortical dysfunction associated with this disorder (37, 38). We highlight emerging connectivity strategies that deal with the complexity and heterogeneity of schizophrenia anatomy, physiology, and behavior (40). For instance, we articulate how data-driven tools that capture distributed connectivity abnormalities – such as global brain connectivity (GBC) – have the potential to avoid biases and identify network disturbances that traditional seed-based approaches may fail to detect. We next detail recent specific efforts to assay thalamo-cortical dysfunction in schizophrenia (38), long thought to be important to the clinical features of this disorder (41–43). We also discuss the potential utility of such data-driven approaches to discover variations in large-scale systems that may be altered across diagnostic categories (e.g., bipolar illness with psychosis and schizophrenia) and that may inform symptom-based or circuit-based diagnostic systems, such as the NIMH Research Domain Criteria (RDoC) (44–46).

The encouraging progress in understanding functional connectivity alterations in schizophrenia creates new opportunity to reconcile system-level findings with hypotheses emerging from molecular and cellular studies of this disorder. Failure to integrate these multiple levels of research undermines our understanding of the pathophysiology of this disorder and it will impede medications development. To this end, we next turn to studies using pharmacological models (47), such as the NMDAR antagonist ketamine, to test hypotheses related to the causes of circuit dysfunction in schizophrenia (18). Here we review a focused set of pharmacological neuroimaging (ph-fMRI) studies that utilized NMDAR antagonists to alter behavior and connectivity in healthy volunteers as a way to better understand schizophrenia (18, 48–54). These ph-fMRI investigations provide insight into how specific manipulations of synaptic function have effects that scale to produce both behavioral and system-level alterations that may be observed in patients. We articulate a set of directions for future studies that can capitalize on ph-fMRI as a tool to provide insight into specific illness-related mechanisms, especially when combined with advanced functional connectivity approaches.

Finally, we briefly articulate the utility of neuroscience theory and computational models for iteratively guiding our pharmacological and clinical experiments (55, 56). We focus on one type of computational modeling that may hold promise in this regard – namely, biophysically realistic computational models that contain cellular-level detail necessary to characterize specific synaptic disturbances that may occur in disease states (56, 57). This level of biophysical detail can provide a vital opportunity to test both hypothesized synaptic alterations in schizophrenia (55) and also possible pharmacotherapies that may attenuate such disturbances (58). Despite possible advantages, we note some key limitations of these modeling approaches, pertaining to the constrained behavioral repertoire and neural architectures that are currently effectively modeled in this way, largely owing to gaps in our basic understanding of neurobiology that can constrain such models. Therefore, we articulate a key objective for the future of schizophrenia research: biophysically realistic computational models need to be systematically developed and scaled to the level of neural systems (59), to inform fMRI-level observations in schizophrenia as well as those observed following pharmacological manipulations in healthy humans (**Figure 1**). We argue that this approach could be especially productive for the study of functional connectivity in psychiatric conditions, with the

and Johnson (60).

specific mechanisms across neural systems may produce complex behavioral alterations seen in serious mental illness. This challenge is also exemplified by

ultimate aim of developing mechanistically derived biomarker predictions.

Building on these insights, we emphasize the need for efforts to translate our basic discoveries in neuroscience and cellularlevel hypotheses in schizophrenia with system-level observations that may directly relate to the complex behavioral abnormalities observed in this illness. We highlight multiple complementary neuroscientific approaches, including recent clinical studies (37), pharmacological neuroimaging experiments (54), and theoretical/computational neuroscience approaches (18, 58) that can be synergistically harnessed to inform our understanding of underlying mechanisms and guide development of better pharmacotherapies for schizophrenia.

## **DEVELOPMENTS IN FUNCTIONAL CONNECTIVITY APPLICATIONS TO NEUROPSYCHIATRIC DISEASE**

Nearly two decades ago Biswal and colleagues (61) demonstrated that coherent fluctuations in the BOLD signal exist across time and space, demarcating a functional network architecture in the human brain. Such early studies highlighted this phenomenon by focusing on the fluctuations between the left and right motor cortex. This key property of the BOLD signal has in turn generated a paradigm shift in non-invasive human neuroimaging and allowed for characterization of distributed neural networks across species in the absence of task-mediated effects (28–30, 32, 62). Importantly, a close correspondence between resting-state networks and task-based networks has been established within individual subjects and using meta-analytic techniques (31). The value of this approach has been enhanced by the subsequent emergence of analytic techniques that have provided insights into the component structure and regulation of distributed cortical networks (62). For instance, recent advances in graph-theoretical approaches have shown that cortical and subcortical networks can be segregated into unique community structures, providing a comprehensive data-driven mapping of human functional networks (63). Another set of studies has demonstrated the temporal non-stationary properties of the large-scale functional networks, delineating temporal functional modes of the human brain (64). In other words, there seem to be distinct and independent patterns of connectivity over both space and time. Although a comprehensive review of connectivity method developments is beyond the scope of this manuscript, we refer the reader to recent detailed reports on this topic (62). Here we specifically focus on select clinical applications in schizophrenia and bipolar illness.

Note: top left receptor figure was adapted with permission from Kotermanski

Resting-state functional connectivity MRI approaches have been increasingly applied to neuropsychiatric illness (65). Use of this technique is built upon the hypothesis that specific neuropsychiatric conditions are brain disorders that affect computations across large-scale networks of regions or specific circuits in such a way that these alterations can be identified with functional connectivity measures. This hypothesis suggests that such disturbances in neural network function may reflect alterations in more basic cellular-level mechanisms. Thus, disturbances in molecular signaling or synaptic function would be hypothesized to scale and produce disturbances in large-scale neural systems (66). Such network-level disturbances in a given functional system may then reflect specific psychiatric symptoms (12, 27). This framework has guided the application of functional connectivity with the objective of identifying putative biomarkers via BOLD fMRI that could reveal neural system-level disturbances in neuropsychiatric conditions even in the absence of specific tasks. Such an approach may hold promise if its sensitivity and specificity is ultimately refined to the point of diagnostic utility.

Specific advantages of rs-fcMRI include: (i) speed, as the functional network architecture can be reliably assayed within 10 min of data acquisition (67), and perhaps even faster with recent advances in multi-band imaging technology (68, 69); (ii) possible cost-effectiveness. Task-based studies may require longer time for data acquisition, depending on the precise paradigm being imaged. (iii) Lack of performance confounds, which likely affects nearly every cognitive activation task when applied to a psychiatric population (70). (iv) In addition, when coupled with techniques such as GBC or independent component analyses (ICA), rs-fcMRI studies allow the unbiased, simultaneous examination of all neural networks (see below for more discussion). All of these features suggest that rs-fcMRI may be a feasible method for eventually guiding and tracking symptoms, illness progression and possibly response to pharmacotherapies and/or behavioral treatments.

The approach itself, however, is not without limitations. For instance, the size of the correlation coefficient is often used to examine the strength of coupling between different network nodes (71). While this is an analysis-level consideration, the majority of published studies use this technique. This approach may not be ideal for examining shared versus non-shared variance between regions and may obscure more complex differences that could occur in psychiatric conditions (71). For instance, it may be important to uniquely identify functional connectivity alterations that truly reflect altered neuronal communication between two network nodes, as opposed to the influence of a third region on both (a correlation-based approach could not disambiguate these possibilities). It is also important to emphasize that functional connectivity, as measured at rest, is purely correlational and therefore can only be used to make tentative causal inferences. That is, if there is evidence for an alteration in a given circuit, this alteration could be predictive of disease onset or could be a consequence of some upstream physiological alteration. Moreover, there are additional complications involved in the study of neuropsychiatric illness concerning what is a primary deficit, as opposed to a compensation or treatment effects. These potential confounds underscore the complementary value of causal experimental designs involving pharmacologic, neuro-stimulation, or cognitive manipulations of circuit activity. Finally, there are still notable differences in the methods used to analyze rs-fcMRI data. For instance, one ongoing controversial issue relates to removal of the global signal (i.e., global signal regression, GSR) from functional connectivity data (72, 73). This step effectively regresses out variance associated with global brain fluctuations, which can be large in magnitude and can perhaps obscure certain real functional relationships (74). Nonetheless, this step effectively shifts the mean of the connectivity distribution, resulting in a portion of correlations being moved into the negative range – thus spuriously inducing at least some anti-correlations. There is convincing evidence from electrophysiology experiments using animals that

anti-correlations indeed exist (75, 76). However, GSR could still complicate some aspects of between-group comparisons and the interpretations of clinical studies (73), especially when examining system-level relationships across groups. Therefore, future studies should carefully continue to consider the possible impact of this step on between-group differences in clinical rs-fcMRI studies.

Another methodological issue pertains to how networks are identified. Many early functional connectivity studies both in healthy adults and in clinical populations used a seed-based approach, whereby the correlation from a given region of interest is examined with all other voxels in the brain. This approach has also been complemented by more complex graph theoretical and network-based methods (77). Seed-based approaches inherently assume a consistent difference across a set of regions in a clinical condition. However, the pattern of dysconnectivity for a single region could indeed be more variable both across subjects and within regions. Such spatial variability in functional networks across subjects may present a limitation of seed-based approaches when applied in studies of clinical populations (11, 40). This may be the case in complex mental illnesses such as schizophrenia. Therefore, detecting more complex patterns of dysconnectivity requires new approaches taking into account these confounds (78–80).

New methods are needed to detect more complex patterns of disorder-related disturbances in connectivity (80), illustrated by the GBC approach (11, 35, 40, 81). The GBC technique is specifically designed to consider connectivity from a given voxel or (area) to all other voxels (or areas) simultaneously by computing either average connectivity strength (weighted GBC) or by counting the number of connections above a given connection strength (unweighted GBC). Thus, this approach is data-driven and unbiased as to the location of connectivity disruption. That is, unlike typical seed-based approaches, GBC requires only the seed region to be relatively consistently located across subjects, while the target regions can vary substantially across subjects (**Figure 2**). For example, if a given area is perturbed in its functional connectivity consistently, irrespective of the overall network spatial configuration where the perturbation is, GBC will remain sensitive to this alteration. Further, unlike typical seed-based approaches, GBC involves one statistical test per voxel (or ROI) rather than one test per voxel-to-voxel pairing – substantially reducing multiple comparisons (e.g., 30,000 rather than ~900 million tests). These two improvements over typical seed-based approaches can dramatically increase the chance of identifying group differences in connectivity, or individual differences in connectivity correlated with behavioral symptoms (11, 35, 40, 81). Indeed, consistent with these proposed advantages, GBC has now been applied to identifying regions with large-scale disruptions in functional connectivity across a variety of mental illnesses, such as schizophrenia (11), bipolar disorder (35), and obsessive-compulsive disorder (OCD) (82). GBC is explicitly designed to address questions about brain connectivity that are qualitatively distinct from traditional seedbased functional connectivity analyses. For instance, areas of high GBC are "hubs" of connectivity in the brain that are maximally functionally connected with other areas and may play a role in coordinating large-scale patterns of brain activity (40). In that sense, a group difference in GBC may reflect areas and networks

differences in connectivity patterns within the region. Seed or ICA approaches would be unable to identify dysconnectivity in such a region because of

in which the large-scale coordination of information process-

properties of GBC – detection of regions with large-scale functional connectivity disruptions along with tolerance for individual differences – makes this method particularly powerful for identifying regions with consistently large and distributed disruptions that have functional consequences reflected in individual differences in psychiatric symptoms. Continued refinement of graph-theoretical data-driven metrics such as GBC may allow a powerful path forward for delineating biomarkers within and across diagnostic categories.

with permission from Cole and colleagues (11).

Another data-driven method that has increasing applications in psychiatric neuroimaging involves ICA of BOLD signal fluctuations at rest. Here ICA provides a tool to identify spatially (78, 83) or temporally (64) independent modes of brain function. ICA has been successfully used to identify distributed network

ing is affected in the disease state. For instance, decreased GBC may reflect decreased participation of a brain region in broader networks. Conversely, increased GBC may suggest a pathological broadening or synchronization of functional networks. Related to this point, we have now published a manuscript examining distributed networks in OCD using GBC where specific striatal and orbitofrontal circuits have been implicated. Yet, GBC remains sensitive to alterations in neuropsychiatric conditions that may actually be associated with more focal deficits (which may be the case in OCD versus schizophrenia). In that sense, GBC is not necessarily more useful for "global" versus "restricted" deficits, but rather dysconnectivity to a given "hub" region that may be affected in its participation in wide-spread neural networks. These

abnormalities in both schizophrenia and bipolar illness (84). In particular, temporal ICA (64) may provide a novel and powerful method to assay the non-stationary properties of distributed neural systems across different clinical conditions. It remains to be systematically tested if recently defined temporal functional modes exhibit alterations in neuropsychiatric illness. One possibility, given wide-spread neurotransmitter disruptions in schizophrenia, is that the temporal functional modes are substantially more nonstationary in this illness. Powerful new acquisition sequences [e.g., multi-band imaging (68)], that allow much finer temporal sampling, will provide critical technological innovations at the level of neuroimaging acquisition with direct clinical applications, in particular when using ICA-type measures.

Collectively, the continued success of these methods suggests that ongoing refinement of tools to examine functional connectivity remains vital for development of biomarkers for psychiatric illness. Next, we turn to specific findings in the field of schizophrenia and bipolar research. We detail recent methodological advances in connectivity work that have been directly applied to better understand network disruption in these psychiatric conditions.

## **EMERGING EVIDENCE SUGGESTS LARGE-SCALE DYSCONNECTIVITY IN SCHIZOPHRENIA**

There are now a number of influential hypotheses suggesting that, at its core, schizophrenia may be a disorder of large-scale neural connectivity (6, 85), i.e., "dysconnectivity." Here we refer to dysconnectivity as an alteration in neural communication that specifically produces behavioral pathology, as opposed to a simple alteration in functional connectivity that deviates from the norm. The models proposing such dysconnectivity range from theoretical to formal computational hypotheses (42, 43, 66, 86–88). This work has inspired the search for connectivity biomarkers for schizophrenia and it broadly includes: (i) studies that use rs-fcMRI, that is investigations that study the intrinsic properties of BOLD signal fluctuations to delineate system-level alterations in psychiatric illness (27); (ii) studies that examine task-based dysconnectivity, that is alterations in connectivity during specific task contexts; (iii) pharmacological studies of individuals at rest or performing tasks, which we discuss in the upcoming section. While the literature on task-based connectivity has offered important clues regarding network alterations during specific cognitive processes (e.g., WM) (27, 89, 90), here we specifically detail emerging efforts to use rs-fcMRI to define biomarkers for schizophrenia. This is not to say that characterizing functional connectivity during cognitive processing is not important. In fact, such studies represent a vital effort to move our understanding beyond activationbased hypotheses regarding cognitive deficits in schizophrenia or regional hypotheses of specific symptoms. In the future, task-based functional connectivity studies will ultimately aid our understanding of network dysfunction with respect to specific symptoms or specific cognitive processes known to be profoundly affected in schizophrenia (e.g., WM) (26). Additionally, task-based studies of functional connectivity provide an important assurance that observed differences in resting-state functional connectivity do not merely reflect differences in the dynamics and content of mental processes subjects might engage in during rest, but rather stable disruptions in functional connectivity that persist across mental

states (90). However, rs-fcMRI can be used to identify neuralsystem alterations even more broadly. With that in mind, we argue that rs-fcMRI offers a unique tool to identify distributed neural system alterations, extending beyond a given task context. Moreover, as noted above, this approach in particular bypasses task performance confounds which plague the task-based activation literature (70) and such concerns will likely extend to task-based connectivity work.

Despite its promise, a major obstacle in delineating successful neural markers for schizophrenia (or any other severe neuropsychiatric disease) using connectivity (or any other approach for that matter) has been the complexity of this illness and the vast range of behaviors it affects (91), as well as the complex temporal dynamics of its progression. That is, the apparent clinical complexity of schizophrenia is a major obstacle to biomarker development that may apply to all patients. In addition, the associated clinical heterogeneity (i.e., differences in symptoms across patients), the presence of numerous comorbidities and environmental modifiers, the ubiquitous confound of long-term antipsychotic treatment, and the wide range of affected behaviors collectively reduce the likelihood that a single biomarker would be applicable to the entire syndrome. Moreover, many researchers and clinicians would argue that schizophrenia is a theoretical construct used to label co-occurring symptoms whereby it may be challenging (or perhaps impossible) to map out a reliable and replicable marker of the entire syndrome in a parsimonious way [although theoretical models of the disease often consider it as one entity (66)]. We argue that functional connectivity tools offer a useful path forward in this regard by providing methods to test specific large-scale dysconnectivity patterns in relation to this heterogeneity that may be otherwise difficult to capture.

Certain studies using rs-fcMRI have, however, attempted to deal with clinical heterogeneity by focusing on the relationship between altered functional connectivity in specific pathways and linking those pathways to particular features of schizophrenia. For instance, studies by Hoffmann and colleagues have focused on better understanding alterations in connectivity specifically associated with auditory hallucinations (92), which many patients experience (1). They found compelling evidence for disruptions between regions associated with auditory processing. Specifically, they found that when they seeded Wernicke's area, there was significantly greater functional connectivity to Brodmann's areas 45/46 among hallucinating patients compared with non-hallucinating patients. In a subsequent analysis, they reported that the functional connectivity within a functional loop including the Wernicke's area, inferior frontal gyrus, and putamen was robustly greater for hallucinating patients compared with non-hallucinating patients. Vercammen and colleagues also found that patients with schizophrenia evidenced attenuated functional connectivity between the left TPJ (temporo-parietal junction) and the right Broca's area (93). These are examples where targeted seed-based approaches may identify alterations in circumscribed circuits associated with specific symptoms.

Another set of emerging studies have studied the "salience" and "control" systems, focused on striatal and insular dysconnectivity in schizophrenia (94, 95), particularly in relation to the "aberrant salience" hypothesis (96, 97). Briefly, the "aberrant salience" hypothesis has been linked to abnormal striatal dopamine function, suggesting that during psychotic states patients have a higher likelihood of forming inappropriate associations and respond excessively to random neutral events. Related to this issue a study by Tu and colleagues examined whether schizophrenia is associated with functional connectivity alterations within the cinguloopercular (CO) network specifically (98). They identified significantly reduced functional connectivity in the bilateral putamen for patients with schizophrenia, which was related to cognitive performance in patients. The authors concluded that schizophrenia is associated with disconnection within cortico-striatal circuits. A complementary study by Moran and colleagues (99) focused on the anterior insula as a key node involved in modulation of distributed neural systems (as part of the CO system). The authors tested for disruptions in the functional relationships between the insula and control/default networks in patients with schizophrenia. Similarly, they examined a relationship between these network disturbances and cognitive deficits. Consistent with *a priori* predictions, Moran and colleagues found strong support for the disrupted relationship between the anterior insula and the control-executive and default-mode networks in schizophrenia, which again was predictive of cognitive performance. Most recently, a study by Palaniyappan and colleagues (100) also focused on the relationship between the salience (insular) system and the executive [lateral prefrontal cortex (PFC)] networks in schizophrenia. They explicitly tested for evidence of disrupted directional influence across the networks. In other words, similar to Moran and colleagues, they used an auto-regressive technique (i.e., Granger causality) to examine both feed-forward and reciprocal connectivity between the aforementioned networks. The authors reported significant differences in patients with regard to time-lagged functional relationships between executive and insular systems. The authors conclude that this "breakdown" in directional functional relationships may reflect aberrant processing of novel salient information. These studies provide compelling emerging evidence that there may indeed exist causal breakdowns across functional networks in schizophrenia. Nevertheless, it is important to note that lag-based causality measures in the context of BOLD signal analyses have not been without controversy (101–103), primarily because of systematic difference in the hemodynamic response function lags across areas. Given these concerns, the use of Granger causality in rs-fcMRI studies needs to be interpreted with caution [see Ref. (104) for a more detailed treatment of using auto-regression techniques with the BOLD signal]. Future studies using concurrent electrophysiology/fMRI protocols could provide convergent evidence to address issues regarding temporal dependencies across cortical networks when analyzing the BOLD signal.

We discussed studies that focused on either specific regions or networks that may be abnormal in schizophrenia. An alternative tactic, precisely because of the complexity of this illness, is to use data-driven methods to study large-scale connectivity alterations that may be difficult to pinpoint *a priori*. A number of prominent and well-established models of schizophrenia neuropathology implicate profound disruptions in PFC function (105), likely stemming from a confluence of glutamate, GABA, and dopamine alterations that could jointly affect PFC function (7, 19, 106–108) (see **Box 1**). One area that has been repeatedly implicated in schizophrenia neuropathology is the dorso-lateral PFC (23). However, this evidence has largely been marshaled through taskbased studies such as those examining WM deficits in this illness (70). A deficit in task-evoked computations of a given region does not necessarily guarantee that the same node will show other forms of functional "dysconnectivity." Moreover, such evidence does not guarantee that this same area may be disrupted in its functional connectivity in the absence of a task (i.e., during resting-state). Nevertheless, PFC functional deficits have been considered a hallmark feature of the illness. Therefore, one promising approach is to examine global PFC dysconnectivity in a data-driven way.

As articulated above, the GBC functional connectivity approach is specifically designed to test the hypothesis that a given functional brain region has altered coupling with the rest of the brain (or a large anatomical portion of the brain, such as the PFC). To test the efficacy of this approach, Cole and colleagues (11) used a restricted GBC (rGBC) approach focused on PFC in particular in a sample of patients with chronic schizophrenia relative to demographically matched healthy comparison subjects. Cole and colleagues reported that bilateral regions centered on the lateral PFC showed reductions in their PFC rGBC, suggesting that these regions have profound alterations in their connectivity patterns with the rest of PFC. To further test the relationship between identified PFC rGBC alterations and cognitive deficits, Cole and colleagues quantified the relationship between IQ and the PFC global connectivity alterations. The authors found that greater connectivity between the identified lateral PFC and the rest of PFC predicted better cognitive performance, suggesting that a lower index of global PFC coupling may in part relate to cognitive deficits in schizophrenia. Moreover, the authors additionally examined the pattern of whole-brain coupling of the discovered right lateral PFC region using seed-based techniques. The authors found that patients diagnosed with schizophrenia showed increased coupling with sensory cortices (**Figure 3B**, posterior regions, shown in yellow-red). Patients also showed reduced connectivity with prefrontal and other higher-order temporal regions (**Figure 3B**, shown in blue). Collectively, this initial report demonstrates that data-driven functional connectivity approaches could identify regions previously implicated in schizophrenia using task-based methods. Moreover, these identified areas showed robust and complex alterations of connectivity with the rest of the brain and importantly, related to observed symptoms and other cognitive measures.

Similar data-driven efforts have been applied to study global connectivity alterations in other psychiatric conditions. Specifically, Anticevic and colleagues extended the approach to study bipolar patients with and without a history of psychotic symptoms (35). While examining neural system-level disturbances in schizophrenia is a vital objective, psychosis occurs across a number of diagnostic categories. For instance, many bipolar patients experience frank psychosis (121). Given this clinical observation, some studies divide bipolar patients based on the presence or absence of psychotic symptoms (122, 123). In that sense, there may be somewhat limited utility in predefined diagnostic boundaries for understanding variation in neural circuits that could possibly be affected across neuropsychiatric conditions (45). Recognizing this

### **Box 1 Glutamate versus dopamine: upstream versus down-stream mechanisms in schizophrenia**.

It is increasingly acknowledged that schizophrenia is associated with disturbances in multiple neurotransmitter systems, including alterations at the cortical microcircuit level in glutamate and γ-aminobutyric acid (GABA), as well as disturbances in dopaminergic neurotransmission along striatal-thalamic-cortical pathways (7, 106, 109). Additional studies have also implicated the glial system as possibly compromised in schizophrenia (110, 111) and other neuropsychiatric conditions (112). However, it remains unknown whether dopamine or glutamate alterations are the upstream causes or down-stream consequences of the disease process (109). Addressing this question has important implications for appropriately constraining computational models, both at the microcircuit level (58) and expanding those mechanisms to the system-level (18).There are two broad possibilities to consider: (i) there may be dissociable groups of patients each associated with a primary abnormality in one of the broad neurotransmitter systems, but as a consequence there may be secondary abnormalities in the other system due to shared pathways and functional loops (43, 88). (ii) One alternative possibility is that there are always primary alterations in only one of the two systems, followed by alterations in the other. Disambiguating between these causal possibilities can have important implications for developing optimized targeted pharmacotherapies for specific patient groups, which is a likely possibility given the heterogeneity of the illness. By extension, resolving these issues may have implications for optimized pharmacotherapies for a given phase of illness [if initial stages of schizophrenia are primary associated with hyper-glutamatergic neurotransmission (113)]. Also, these two possibilities have important implications for the utility of pharmacological models of psychosis (e.g., amphetamine versus the NMDAR antagonist challenge studies in healthy volunteers).

Here it is important to consider some key differences between the dopamine and glutamate hypotheses in schizophrenia in relation to pharmacological findings and their therapeutic effects: (i) dopaminergic medication has not proved successful at ameliorating the full range of impairing symptoms in schizophrenia, particularly cognitive deficits (26); (ii) pharmacological models of schizophrenia targeting the dopamine system (e.g., amphetamine challenge) have typically produced a clinical profile marked by acute psychosis, as opposed to a broader range of symptoms produced by pharmacological agents targeting the glutamate system (15, 114, 115). Therefore, while it is certainly important to acknowledge dopamine as a key component of disrupted neurotransmission in schizophrenia (21, 116, 117), it remains to be determined if dopamine is indeed a down-stream cause or a consequence of primary disruptions in glutamate (118). Relatedly, irrespective of cause or consequence arguments it will vital to consider the diversity of DA receptors (D1 versus D2) and their respective sites of influence in cortex (22, 119, 120) versus striatum (108), which in turn generates important constraints for computational modeling studies that incorporate dopaminergic and glutamatergic signaling mechanisms.

in 90 patients diagnosed with schizophrenia relative to 90 matched healthy controls (38). Anticevic and colleagues found robust alterations in thalamo-cortical information flow in schizophrenia, whereby sensory-motor cortical regions showed over-connectivity in schizophrenia (regions shown in yellow-red), but prefrontal-striatal-cerebellar regions showed under connectivity in schizophrenia relative to controls (regions shown in blue). Anticevic and colleagues fully replicated this pattern in an independent

region identified via GBC; as with the thalamic seed, there was increased coupling with sensory (posterior regions, shown in yellow-red) but reduced connectivity with prefrontal and other higher-order temporal regions (shown in blue). This over/under pattern recapitulated qualitatively the observations found for the thalamic analysis in **(A)**, as shown in the distribution plots on the bottom of each panel. Note: figures adapted with permission from Anticevic and colleagues (38) and Cole and colleagues (11).

limitation, there are emerging efforts to map common aspects of neural dysfunction across diagnostic categories.

Focusing on bipolar illness, Anticevic and colleagues examined the possibility that there are similar abnormalities in bipolar illness with a history of psychosis to those found in schizophrenia. Consistent with this hypothesis, Anticevic and colleagues found that bipolar illness is indeed associated with altered PFC rGBC specifically in a medial prefrontal/ventral cingulate region implicated in regulation of emotion (35). Importantly, this effect was primarily driven by a presence of psychosis history, and the magnitude of the observed PFC rGBC disruption was correlated with the severity of prior psychosis. This effect, however, was obtained in euthymic bipolar individuals (as opposed to symptomatic schizophrenia patients above), which suggests that at least some alterations in global connectivity may be a stable (trait) feature of the illness and may not necessarily only manifest in overtly symptomatic individuals. In that sense, data-driven functional connectivity may provide a tool to examine individual variability in both current and/or lifetime symptoms. Such data-driven approaches can also be used to establish alterations in functional neural architecture over time – namely, whether the observed disruptions in schizophrenia and bipolar illness change longitudinally as the illness progresses. This possibility is consistent with prominent neurodevelopmental models of severe psychiatric illness (14, 124–126), which suggest that there may be profound functional changes along the illness progression, to which these tools may be sensitive.

While promising, a key challenge facing data-driven approaches will be to establish whether identified effects consistently replicate across sites and samples. It may be possible that illness sample heterogeneity, illness stage, symptom severity, anatomical heterogeneity, or other factors will critically impact the patterns of data-driven effects across studies and samples. That is, it may be possible that different foci are detected as showing disruptions in different samples. Therefore, it will be critical to determine which of the identified data-driven global connectivity alterations are stable and replicable (perhaps capturing some core disturbances) and which alter as a function of other variables (perhaps capturing state effects). Such ongoing data-driven efforts should also continue to capitalize on recent advances in graph theory (80) as well as neuroimaging acquisition-level improvements (64, 69, 127) which could jointly improve the sensitivity of resting-state connectivity-derived metrics.

As described above, the PFC has been repeatedly implicated in schizophrenia neuropathology. However, despite data-driven efforts to map PFC dysfunction, the PFC remains a challenge in the study of schizophrenia, owing largely to the complex individual differences in function and anatomy of the PFC. Therefore, to ultimately establish a viable large-scale, brain-wide marker of neural alterations in schizophrenia, an alternative approach can be taken. One possible path is to start from the thalamus as a key region of interest (42). That is, while the PFC has classically been implicated in schizophrenia neuropathology, recent studies highlight disturbances in additional neural foci. In particular, the thalamus has emerged as an important disrupted locus in schizophrenia. While we presented some arguments against using seed-based approaches, the thalamus may present a unique opportunity in this case. Here we argue for the utility of specifying a seed when there is a specific theoretically guided hypothesis implicating a given area in the disease process. In this case, examining thalamic connectivity using rs-fcMRI in schizophrenia capitalizes on several key aspects of this subcortical region: (i) the thalamus is topographically connected to the entire cortex (128, 129), and may therefore represent a node particularly sensitive to network-level disturbances (42); and (ii) the thalamus contains anatomically and functionally segregated nuclei readily identifiable via neuroimaging (130) thus providing a lens for examining parallel yet distributed large-scale connectivity disturbances in neuropsychiatric disease.

Consistent with this view, a large body of evidence implicates significant thalamo-cortical communication disturbances in schizophrenia neuropathology (41, 42, 131–134). In fact, a fundamental aspect of large-scale brain organization preserved across mammalian species is thalamo-cortico-striatal sub-circuits that are thought to integrate various functions such as emotion processing and motor output (135–138). Such circuits may become profoundly dysregulated in schizophrenia, and the thalamus, as an organized hub of cortical and subcortical connections, may be especially sensitive to such dysregulation. Essentially, the thalamus serves as a nexus for parallel circuits through which diverse cortical and subcortical functions are integrated and distributed throughout the cortical mantle (128, 139, 140). These thalamic circuits have been implicated in schizophrenia pathophysiology on the basis of neuropathology studies (88, 141–143), pre-clinical lesion models (144, 145), structural imaging studies (146, 147), and computational models (43). Moreover, thalamic abnormalities are repeatedly implicated in sensory gating (148, 149) and filtering disruptions (150, 151) associated with this disorder (42). Indeed, one prominent model of schizophrenia neuropathology is centered on the thalamus as a key hub of disrupted computations in this illness. This "cognitive dysmetria" hypothesis articulates a distributed disruption in information processing across widespread cortical and subcortical nodes (39). In their seminal theoretical piece, Andreasen and colleagues argue that in order to explain the full range of schizophrenia symptoms, the field has to move away from region-specific models, but rather consider a distributed processing deficit, which could also parsimoniously explain a computational abnormality in a given node of a distributed complex system. In that sense, the thalamus is a uniquely positioned set of nuclei that communicates with virtually every cortical territory and is likely to be profoundly affected in schizophrenia.

Building on theoretical, pre-clinical, and anatomical work, recent studies of functional connectivity have begun to map thalamo-cortical alterations in schizophrenia. The first study to do so, by Welsh and colleagues (152), focused on the medio-dorsal nucleus of the thalamus as a seed region. This focused approach is justified given that specific thalamic nuclei in schizophrenia may show particularly profound functional connectivity disruptions. The medio-dorsal nucleus projects heavily to PFC regions (153, 154), and is thought to be compromised in schizophrenia (142, 144, 155). Welsh and colleagues found lower connectivity between the medio-dorsal nucleus and the PFC in patients with schizophrenia relative to healthy comparison subjects. However, this early

investigation was based on a very small sample and could therefore be limited in its ability to provide conclusions regarding more subtle disruptions elsewhere. Moreover, Welsh and colleagues could not provide information regarding additional thalamic nuclei given the explicit focus on the medio-dorsal nucleus. In a subsequent study, Woodward and colleagues employed a substantially more powered sample (*N* = 62) and extended the approach to other thalamic nuclei (37). The authors used a parcellation scheme of thalamo-cortical connections at rest provided by Zhang and colleagues (129). In the study by Zhang and colleagues, cortical areas were clustered into subdivisions that exhibited unique functional connectivity with distinct thalamic nuclei based on the similarities in resting-state BOLD signal. Woodward and colleagues harnessed this segmentation scheme to test the hypothesis that unique thalamo-cortical circuits may show different patterns of disturbances in schizophrenia. Strikingly, Woodward and colleagues found that compared to healthy controls, the thalamic segmentations associated with the PFC showed reduced connectivity in schizophrenia. In contrast, cortical territories centered on sensory-motor regions showed increases in thalamic coupling in schizophrenia. This evidence suggests that there exists a profound alteration in thalamo-cortical information flow in schizophrenia but one that seems to follow an anatomical dissociation between sensory and higher order association regions. Building on this robust evidence for thalamo-cortical alterations in schizophrenia it is important to provide information about the specific connections being affected. Specifically, the cortical parcellation scheme, while a powerful initial demonstration, did not allow for examination of the cerebellum for instance. Cerebellum is a structure that, like the striatum, has projections to the cortex by way of the thalamus (156), and has been implicated in schizophrenia pathophysiology (39). Lastly, perhaps due to restricted power, the authors could not examine subtle relationships between symptoms and identified dysconnectivity.

A subsequent report by Anticevic and colleagues examined thalamo-cortical dysconnectivity in 90 patients diagnosed with schizophrenia relative to 90 matched healthy comparison subjects (38). The key objective of this investigation was to determine if thalamo-cortical disturbances span across diagnostic boundaries that share similar symptoms. This cross-diagnostic extension directly informs the objectives articulated by the NIMH RDoC initiative, which aims to develop biomarker-driven diagnostic systems (45). First, the authors replicated the core findings by Woodward and colleagues, demonstrating that schizophrenia is associated with increased coupling between the thalamus and all sensory-motor cortices. In contrast, frontal-striatal-cerebellar nodes showed reduced coupling with the thalamus in schizophrenia relative to healthy comparison subjects (**Figure 3A**). Both patterns were fully replicated in an independent and smaller sample of patients. Critically, further analyses demonstrated that these two sources of disturbance were functionally related. That is, those patients with the highest sensory-motor-thalamic coupling also showed the lowest prefrontal-striatal-cerebellar-thalamic coupling. This effect was most prominent for thalamic clusters centered on the medio-dorsal nucleus with known dense projections to the PFC, ruling out the possibility of pan-thalamic dysconnectivity that is uniform. Furthermore, the magnitude of the

sensory-motor-thalamic over-connectivity was correlated with PANSS symptom severity across patients, confirming its functional relevance. The magnitude of this correlation, however, was small (*r* = 0.23) – indicating that the observed pattern explains only a small portion of the variance in symptom variation across subjects. An alternative possibility, given that the majority of patients were quite symptomatic, is that the small magnitude reflects a restricted range in symptoms whereby there was little variability in symptom severity across the patient sample.

The identified thalamic dysconnectivity was successfully used for diagnostic classification via multivariate pattern analysis (MVPA) (157) with high levels of sensitivity and specificity across both the discovery and replication samples. This implies that the identified dysconnectivity patterns, while not yet qualifying for a robust biomarker, may be refined and used to predict risk and assess treatment response. In particular, there are major ongoing improvements in neuroimaging acquisition (127) and processing (69) technology as a direct consequence of the Human Connectome Project (158) that can enable future studies to iteratively refine neuroimaging approaches. Such studies can focus on the identified patterns to ultimately improve methods and fine-tune the identified patterns for biomarker use.

Lastly, the authors found that the bipolar illness sample exhibited an"intermediate"pattern of disturbance such that the patterns of thalamo-cortical connectivity were "shifted" relative to healthy comparison subjects but not as severely altered as those identified in schizophrenia. This finding in particular offers promise for using neuroimaging markers to inform our understanding of shared disturbances in the underlying biology that cut across diagnostic categories (123). The next step will be to understand how such shared neural system-level "endophenotypes" (159, 160) map onto co-occurring behavioral disturbances (e.g., psychosis) as well as onto possibly shared alterations at the level of cortical microcircuits (which we discuss in the last section).

Most recently, Klingner and colleagues provided convergent results from a sample of 22 patients diagnosed with schizophrenia and 22 matched comparison subjects (161). In their study they separately examined the left and right thalamic seeds. They found robust evidence for increased thalamic connectivity with bilateral sensory-motor and auditory cortices. The authors conclude that their results suggest a possible "lack of thalamic control on motor/sensory information processing resulting in increased (and less filtered) forwarding of information to the prefrontal cortex." This hypothesis is consistent with findings from the two earlier and larger studies (37, 38). Interestingly, Klingner and colleagues did not observe notable reductions in thalamic coupling with frontal-striatal-cerebellar nodes in schizophrenia, reported by both aforementioned groups. There are at least two possible interpretations for this difference: (i) the sample size in the Klingner study was much smaller and possibly underpowered to find both sets of patterns (although the effect size analysis and replication analyses by Anticevic and colleagues argues against this possibility); and (ii) Klingner and colleagues may have employed techniques, such as using GSR as a preprocessing step, that could have led to different results (73). It remains to be systematically determined what the true contribution of the global signal is in these analyses, especially in the

patient groups. One possibility is that the global signal carries biologically meaningful information regarding cortical-thalamic disruptions in schizophrenia that needs to be carefully characterized. Methodological issues notwithstanding, this emerging body of work strongly and consistently implicates disruptions in thalamo-cortical information flow as a neural system marker in schizophrenia.

While this initial progress in mapping thalamo-cortical disturbances in schizophrenia represents a promising advance in psychiatric neuroimaging research,there are still fundamental gaps in our understanding of how such findings relate to neuropathological mechanisms of this illness. There are a number of future directions that the field should pursue to better understand these observations. For instance, it remains unknown why there are dissociable disturbances across thalamic nuclei in schizophrenia. Future studies focused more exclusively on the medio-dorsal nucleus versus, say, the pulvinar (known to be more involved in visual processing) could begin to explain mechanism behind these differences. Although the original study by Welsh and colleagues provides clues here, follow-up studies with more power that focus on the mediodorsal nucleus could provide finer-grained information regarding its patterns of dysconnectivity with the PFC.

Another complex issue that is not adequately resolved by any of the noted investigations relates to medication effects. For instance, there are considerations of medication dose, type of medications (given that patients are often treated with multiple drugs from different medication classes), and possible systematic differences in the medications received by patients carrying the diagnoses of schizophrenia, schizoaffective, and bipolar (e.g., mood stabilizers and anti-depressants). While all studies address this issue statistically to a certain extent (by computing chlorpromazine equivalents and then co-varying for the medication dose), future studies in un-medicated patients are needed. It is possible, however, that medication may not necessarily be a confound in this case – instead, antipsychotic medication could actually stabilize thalamo-cortical dysconnectivity. Studies explicitly aimed at testing medication effects on connectivity could provide more detailed insight into this issue. It also remains unknown if the identified patterns of large-scale thalamo-cortical dysconnectivity are characteristic only of chronic stages of schizophrenia or whether they already appear in the prodromal or early stages of the illness. Establishing the link between identified thalamo-cortical dysconnectivity and illness progression remains a vital effort to inform the viability of this marker for predicting and/or tracking risk and progression of the disease. While one of the studies noted above provides a functional link between sensory-motor-thalamic overconnectivity and PANSS symptoms, it remains unknown if these patterns relate to cognitive and executive functional deficits characteristic of the schizophrenia syndrome (26). It may be possible that alterations in thalamo-cortical function (especially the PFC component) in part relate to cognitive deficits observed in this illness.

Additional questions still exist pertaining to the crossdiagnostic relevance of these observations. As noted above, it was shown by Anticevic and colleagues that qualitatively similar patterns of thalamic dysconnectivity are apparent in bipolar illness (although smaller in magnitude than those found in schizophrenia). It remains to be determined if those bipolar patients with a history of co-occurring psychosis are quantitatively more similar to alterations identified in schizophrenia (123). Such finer-grained cross-diagnostic investigations have further potential to inform and refine the clinical relevance of the identified marker. These studies should be complemented with targeted efforts to improve the classification provided by Anticevic and colleagues and allow for even more precision in harnessing this putative biomarker. Recently Fox and Greicius discussed progress in neuropsychiatric studies using resting-state connectivity. They appropriately concluded at the time that in schizophrenia there has been remarkably little progress in producing replicable results (34), perhaps owing to the complexity and heterogeneity of this neuropsychiatric illness noted above. These recent studies reviewed here, which collectively focused on identifying patterns of thalamo-cortical disruption in schizophrenia, may be converging on a parsimonious final common pathway of this complex disease, at least at the neural systems level (43). In that sense, these effects may be one of the better-replicated findings in the schizophrenia connectivity literature to date, offering promise for biomarker development and refinement.

A longer-term goal will entail bridging this neural system-level marker of schizophrenia with evolving cellular-level hypotheses of schizophrenia neuropathology. It remains unknown how the identified neural system-level markers relate to hypotheses that propose disruptions at the cellular level in schizophrenia. For instance, Anticevic and colleagues articulate a possible role of the disruption in cortical excitation (E) and inhibition (I) balance within the cortical microcircuit in producing system-wide disruptions, which may occur in schizophrenia (see next section) and in turn affect cognition (162). It remains unknown, however, how such alterations can scale to produce the presently observed pattern of aberrant thalamo-cortical connectivity. A corollary of this hypothesis relates to observations in bipolar illness, which was associated with an intermediate pattern of thalamo-cortical alterations. In bipolar illness different cellular-level hypotheses have been proposed from those hypothesized to occur in schizophrenia (112, 163, 164) [although some authors have articulated shared disturbances in GABA interneuron function (165)]. Therefore, either distinct mechanisms operate in these different neuropsychiatric conditions that converge on the same alterations or there may be, at least in part, a shared alteration in some of the same mechanisms across the two conditions. That is, it is possible that some patients with bipolar illness share some of the features of cellular neuropathology that affect patients with schizophrenia (165), especially those bipolar patients who present with cooccurring psychosis (166). Moreover, alterations along a number of distinct neurotransmitter pathways, involving a confluence of glutamate (7, 167), GABA (19), and dopamine disturbance (106), could jointly converge on a profound disturbance in thalamocortical function. In the upcoming sections, we discuss additional neuroscientific tools, namely pharmacological neuroimaging and computational modeling, which can be combined to help elucidate the role of specific cellular and synaptic mechanisms in observed system-level disruptions that occur in schizophrenia.

## **PHARMACOLOGICAL NEUROIMAGING – TOWARD A MECHANISTIC UNDERSTANDING OF SYSTEM-LEVEL DISRUPTIONS IN PSYCHIATRIC ILLNESS**

We presented evidence, supported by several emerging studies, for profound alterations in thalamo-cortical information flow in schizophrenia, as well as evidence for alterations in PFC connectivity. Yet, the mechanisms that could inform rationally guided pharmacotherapy for these disturbances in schizophrenia remain unknown. One leading mechanistic hypothesis proposes possible disruptions in the E/I balance in the cortical micro-circuitry resulting from hypo-function of the NMDAR (7), which might affect cortical computations, leading to large-scale dysconnectivity (14). A way to link such pharmacological mechanisms to neural system-level observations is to directly compare clinical patient studies and results following pharmacological manipulations, or to separately test the effects of pharmacological manipulations in healthy volunteers.

A powerful candidate approach is to use ketamine, a noncompetitive NMDAR antagonist and a leading schizophrenia pharmacological model, which transiently, reversibly, and safely induces characteristic schizophrenia symptoms in healthy volunteers (7). Here pharmacological manipulations provide a method with which researchers can test the effects of a given neurotransmitter perturbation in a constrained, causal, and hypothesisdriven way. Moreover, such synaptic hypotheses can be implemented directly into computational models to generate experimental predictions (discussed in the last section). A prevailing hypothesis regarding ketamine's effects on cortical micro-circuitry proposes preferential antagonism of interneurons with subsequent disinhibition of pyramidal cells (168), a mechanism we implemented in a recent computational modeling investigation (58) (see below for a discussion). As an extension of this hypothesis, a cortex-wide disruption in E/I balance might de-stabilize thalamocortical information flow in ways observed in schizophrenia (or other neuropsychiatric conditions). It should be noted that it still remains unclear what the relative contribution is of NMDARs on pyramidal cells (169) versus interneurons (58) may be in relation to the hypothesized alterations in E/I balance. Understanding the relative contribution of such cell-specific receptor alterations remains an important future direction, which could inform targeted pharmacotherapies. Such detailed studies of how cellular-level alterations could give rise to thalamo-cortical alterations following pharmacological manipulations remain to be done. Nevertheless, there is emerging evidence from a few focused investigations detailing the effects of ketamine on large-scale cortical connectivity. These investigations provide preliminary clues for how ketamine's effects on large-scale systems may resemble effects seen in schizophrenia.

For instance, a recent resting-state connectivity study by Driesen and colleagues investigated the effects of acute ketamine administration to healthy volunteers on large-scale global cortical connectivity (**Figure 4A**). The authors use a resting-state connectivity approach similar to that applied by Cole and colleagues in chronic schizophrenia (40), extended to include the entire brain (i.e., all voxels without imposing a PFC restriction). A number of studies have demonstrated that NMDAR antagonist administration is associated with excessive pyramidal cell activity, increases extracellular glutamate levels (16), and increases in perfusion and cortical metabolism (48–51, 53). A logical extension of this hypothesis is that administration of ketamine may profoundly affect large-scale cortical connectivity. Consistent with pre-clinical studies, Driesen and colleagues found that NMDAR antagonist administration resulted in a global elevation of functional connectivity (i.e., everywhere in the brain). This observation is broadly consistent with cellular-level hypotheses of ketamine's effects on glutamate release, which may increase coupling of cortical circuits at rest by increasing the E/I ratio (i.e., decreasing cortical microcircuit inhibition). It is, however, critical to point out that chronic schizophrenia has typically been associated with reductions in cortical connectivity (170) and activation (70), especially in the PFC (11), as described above. Therefore, there are at least some evident discrepancies between the effects of pharmacological models such as ketamine and the actual illness (54). This important discrepancy between ketamine's effects on PFC circuits and observations in schizophrenia should be reconciled in prospective studies that directly compare pharmacological and clinical effects using resting-state connectivity measures. One possible factor that could explain differences between NMDAR antagonist effects and chronic schizophrenia is that such pharmacological models may be relevant only to specific patient subgroups or illness stages. One possibility is that the increased connectivity under ketamine is similar to the early stages of psychotic illness. This hypothesis is consistent with elevated glutamate levels reported early in the illness course (113, 171). It is also consistent with the observation that ketamine tends to produce symptoms associated with incipient illness stages, rather than auditory hallucinations that occur in frank psychosis and chronic schizophrenia (15, 49, 172–174). Moreover, significant functional dynamical changes may occur during schizophrenia progression (46) that could profoundly affect PFC function, structure, and integrity. This hypothesis is supported by recent meta-analytic findings reporting decreases in glutamate across the illness progression (113). Whether such alterations across the schizophrenia illness course are reflected in PFC connectivity changes remains unknown, as does ketamine's impact on PFC functional network architecture. Future pharmacological studies as well as cross-sectional and longitudinal clinical investigation are needed to test this hypothesis. In addition, careful pre-clinical experiments could possibly inform such hypotheses (118).

A complementary area of research has investigated the effects of ketamine on functional connectivity (177) in relation to its potential anti-depressant effects (178). While a comprehensive treatment of anti-depressant effects of ketamine is beyond the scope of this review, it is important to note that studies examining its effects on glutamateric pathways in the context of mood symptoms (178) may be highly informative for developing our understanding of its relevance to schizophrenia (111). Briefly, emerging models in this area postulate that ketamine may act as anti-depressant by promoting synaptic plasticity via intra-cellular signaling pathways, ultimately promoting brain-derived neurotrophic factor expression via synaptic potentiation (179) and in turns synaptic growth (178). In that sense, acute NMDAR antagonism may promote synaptic plasticity along specific pathways impacted in mood disorders, such as ventral medial PFC (180, 181, p. 916).

Conversely, when administered to patients diagnosed with schizophrenia, NMDAR antagonists seem to worsen their symptom profile (182), perhaps by "pushing" an already aberrantly elevated glutamatergic signaling profile upward. Collectively such dissociable effects of ketamine may imply that along distinct circuits there may be an inverted-U relationship between ketamine's effects and symptoms: depressed patients may be positioned on the low end of the inverted-U (178) and schizophrenia patents may be positioned on the higher end (183). Both task-based and resting-state functional connectivity techniques are well positioned to interrogate such system-level effects of NMDAR antagonists in humans.

As an example of such an approach, another study examining the effects of ketamine by Anticevic and colleagues focused on understanding the functional impact of NMDAR antagonism on the organization of the large-scale, anti-correlated neural systems (**Figure 4B**). This pharmacological neuroimaging investigation was explicitly focused on understanding ketamine effects on WM. However, in the context of this cognitive question, Anticevic and colleagues also assessed whether the task-based functional connectivity of large-scale neural systems is affected by ketamine administration. As we noted above, such task-based connectivity approaches can be useful to pinpoint how large-scale systems are affected during specific cognitive operations. The authors found that an acute administration of a low ketamine dose profoundly altered the typically observed anti-correlated structure of the largescale neural systems – namely the task-positive fronto-parietal regions (184) and the task-negative regions (typically termed the default-mode network, DMN) (28). In line with this observed decrease in functional connectivity, Driesen and colleagues (185) also found a reduction in WM task-based functional connectivity along fronto-parietal areas (when using a DLPFC seed) following ketamine administration. The investigators reported *increased* functional connectivity of the DLPFC during rest using the exact same seed. This set of observations in particular sheds light on how a pharmacological manipulation such as NMDAR antagonism may have profoundly different effects in the context of a cognitive task (e.g.,WM) and during rest. The mechanisms behind this observed difference are beyond the scope of this review and will be discussed in forthcoming studies.

Briefly, it may be possible that large-scale, network-level synchrony during WM is critically dependent on appropriate contentspecific signals between neural subpopulations (186). In contrast, the BOLD fluctuations during rest likely relate to coupling between regions at the infra-slow-frequency ranges (187). It may be possible that a reduction in appropriate task-evoked synchrony by NMDAR antagonism reduces the ability of a network to form a coherent and optimal level of functional connectivity during WM. This may be exacerbated by an amplification of shared "noise" in the system, which is reflected in the apparent hyper-synchrony at rest (54). Collectively, therefore, an NMDAR antagonist may "disinhibit" the system, which gives rise to infra-slow, spatially distributed fluctuations across large areas of cortex that manifest in aberrant hyper-connectivity at rest. Precisely due to this elevated background noise, in combination with disrupted capacity

for appropriate task-evoked synchrony, the net effect during WM may be a reduction in functional connectivity (which contrasts with observations at rest). Moreover, task-evoked activity can suppress the slow fluctuations associated with rest (188); if this suppression were weakened in schizophrenia, the"signal-to-noise" ratio would be degraded and task-based functional connectivity could be reduced. Because task-based and resting-state synchrony may be separated by timescale, EEG studies, combined with fMRI, could potentially address such hypotheses (189–191). Moreover, precisely because of such important differences between task and rest in certain contexts (e.g., ketamine manipulation) it remains important for task-based and resting-state investigations to inform one another. In the upcoming section, we explicitly discuss the recent developments in computational modeling that can provide a platform for formal integration of neuroscience theory both in the context of resting-state studies as well as formal task-based experiments.

## **BRIDGING LEVELS OF ANALYSIS VIA COMPUTATIONAL MODELING**

Above we discussed recent clinical and pharmacological neuroimaging findings that shed light on the nature of large-scale neural system alterations in schizophrenia. In particular, the pharmacological experiments provide a causal method to explicitly manipulate specific neurotransmitter mechanisms that may be involved in schizophrenia. In that sense, these studies can begin to address given neurotransmitter contributions to neural systemlevel and behavioral alterations observed in schizophrenia. Still, these studies cannot measure synaptic and cellular-level phenomena alone. Therefore, one possible methodological integration involves a formal link between such pharmacological/clinical experiments and computational models that contain this level of functional detail. One branch of computational model that provides a particularly productive platform involves biophysically based models that contain the relevant synaptic mechanisms (57, 192) thought to be disrupted in neuropsychiatric illness (55). Such microcircuit models have been already harnessed to make predictions in the context of ketamine experimental manipulations of WM (58) (discussed more below). There is, however, an ongoing need to scale such models to the level of neural systems to provide relevant predictions for both resting-state and task-based clinical and pharmacological studies.

Recent computational models have been developed to explicitly capture how the global pattern of resting-state functional connectivity arises through cortico-cortical interactions (193). In particular, modeling studies in this area have focused on the extent to which functional connectivity can be predicted by long-range anatomical connectivity (**Figure 5A**). The dynamic interactions between neural populations will also shape functional connectivity. The starting point for these models is an anatomical coupling matrix reflecting long-range connections between cortical regions, derived either from tracer studies in macaque monkeys or from diffusion tractography in humans. The activity of a local region (a node in the large-scale network) follows some dynamics and is shaped by input from other areas, propagating via the long-range connections. Ongoing activity, either due to chaos or noise, produces fluctuations in the activity across the network. The functional connectivity of the model can then be calculated and compared to functional connectivity observed in experiments. Honey and colleagues (194) studied long-range chaotic synchronization when local nodes follow oscillatory dynamics, using connectivity from human diffusion tractography. They found that the global dynamics of the network could partially explain the presence of strong functional connections between regions that lack direct anatomical connection. Cabral and colleagues (195) used a similar oscillatory model with connectivity data from healthy subjects, and parametrically varied the overall strength of long-range connections. They found that decreasing long-range connection strength altered functional connectivity patterns in a manner similar to those observed in schizophrenia (9), with reduced overall functional connectivity strength and changes in certain graph-theoretic measures of the functional connectivity matrix.

Deco and colleagues (59, 193) extended this approach, deriving long-range connectivity from human diffusion tractography and implementing the local node dynamics with a biophysically based model of a cortical microcircuit. In particular, the local microcircuit incorporates recurrent excitation with realistic synaptic dynamics, and the strength of recurrent excitation enables multistable dynamics that can subserve cognitive computations such as WM (197, 198). Noisy background inputs to nodes induce fluctuations in activity that are shaped into correlated patterns by long-range coupling between nodes through the anatomically derived connectivity. Functional connectivity in the model is given by the pattern of these correlated fluctuations. The authors parametrically varied two global parameters with biophysical relevance: the strength of recurrent excitation within local nodes, and the strength of long-range connections between nodes. The strengths of local and long-range connections combine to provide recurrent excitation in the network. They found that the similarity between model and experimental functional connectivity patterns was maximized when the network's baseline state is near the boundary in parameter space between stability and instability induced by excess excitation. These studies reveal that the pattern of functional connectivity in the model is sensitive to both the pattern of anatomical connectivity and the physiological parameters that scale local and long-range connections. The biophysical basis of these models makes them directly applicable to address the dynamical consequences of anatomical and physiological changes induced by disease processes or pharmacological manipulation. Changes in the strengths of local and long-range connections may induce differential effects in the patterns of functional connectivity. Therefore, fitting models to functionally connectivity in patients could distinguish among distinct synaptic alterations. This approach could also potentially reveal the effects of complex drug actions on local circuit tuning and long-range interactions.

The models described above were explicitly constructed to simulate resting-state fluctuations, rather than to implement a particular function such as WM. Nevertheless, functional models can still make predictions that can be tested using functional connectivity, especially task-based functional connectivity (81). To this end, Anticevic and colleagues extended a microcircuit model of WM to study interactions between large-scale networks and

microcircuit models have made an impact on our understanding of cortical dynamics (56), the challenge remains to scale such models to incorporate dynamic interactions across large-scale neural systems, which are likely profoundly affected in schizophrenia (and other severe neuropsychiatric conditions). **(A)** A recently published elegant study by Deco and colleagues (59) illustrates an approach where a biophysically realistic model of cortical computations has been applied to understand the generation of slow-frequency fluctuations in the BOLD signal. The authors used diffusion-spectrum imaging to anatomically constrain the model and in turn fitted the modeling results to empirically-derived resting-state functional connectivity data. The result illustrates that coherent fluctuations in the BOLD signal (i.e., resting functional connectivity) may emerge from a system that is at the edge of chaos, allowing linear but transient departures

their disruption by synaptic perturbation (**Figure 5B**) (18). The biophysically based model consists of two modules of spiking microcircuits: one that is task-activated and capable of WM computations, and one that is task-deactivated from a high-activity baseline state (hypothesized to model the activity pattern of the default-mode network in the context of cognitive activation). The interactions between these modules are mutually suppressive, via long-range projections onto inhibitory interneurons, a feature founded on findings of anti-correlated fluctuations between task-positive and default-mode networks (30). Within the model authors specifically studied the functional impact of disinhibition via NMDA hypo-function on interneurons, to relate such microcircuit hypotheses to neural changes observed under ketamine administration. The authors found that disinhibition of the entire network results in a failure to shut off the default-mode module, impairing the pattern of activation and deactivation during WM tasks. This modeling study provides one example of how a microcircuit model of a specific cognitive process (i.e., WM) can be scaled to incorporate system-level interactions and make predictions relevant for task-based functional connectivity. In addition, one strength of computational models is the ability to systematically explore different operating regimes in the space of model parameters. For example, schizophrenia (26), to better understand the role of NMDA receptor function in the interaction of large-scale anti-correlated neural systems. Specifically they studied the functional antagonism present during a cognitive task between the task-activated (fronto-parietal module) and task-deactivated (default-mode module) networks (196). Following a complete parameter sweep (left), the authors found that a small perturbation of the NMDARs on inhibitory interneurons within each cortical microcircuit captured the firing that was observed experimentally following ketamine administration in healthy volunteers (18). Collectively, these studies offer examples for how biologically constrained modeling approaches can be applied to understand large-scale neural system physiology in both resting-state (non-functional) and task-based (functional) settings. Note: **(A)** of the figure was adapted with permission from Deco and colleagues (59).

Anticevic and colleagues contrasted reductions in recurrent excitation onto interneurons versus pyramidal cells, generating testable predictions for elevated versus reduced excitation/inhibition balance. As a test of the model architecture and operation, the authors analyzed task-based functional connectivity between taskactivated and task-deactivated networks, computed across trials during the delay period of a WM task. Under control conditions, the two networks exhibited robust negative correlation, in line with effective antagonistic interactions in the model. In contrast, under ketamine administration this negative correlation disappears (**Figure 4B**), in line with the model prediction that the task-positive module cannot effectively suppress the hyperactive default-mode module. Collectively, this study provides preliminary evidence that functional models can be directly related to functional connectivity predictions in the context of WM. Future computational/experimental studies should be designed to extend this framework to more complex processes and symptoms that may be disrupted in schizophrenia. We argue that such ongoing efforts for the integration of theory, pharmacological experiments and clinical work will be a vital path for the field of clinical neuroscience to provide testable and rationally guided advances for understanding disease mechanisms and putative treatments.

## **CONCLUDING REMARKS**

Collectively, we articulated recent focused developments in three areas of clinical neuroscience of schizophrenia: (i) we reviewed methodological advances in resting-state functional connectivity that were directly translated to understand neural system-level disturbances in schizophrenia. We specifically focused on datadriven techniques that offer a promising way to detect disrupted connectivity while bypassing the likely complexity and regional heterogeneity of network alterations that are present in schizophrenia. We also discussed ongoing developments in studies of thalamo-cortical dysconnectivity in schizophrenia that are directly informed by influential theoretical models of the illness. (ii) We highlighted select pharmacological studies of the NMDARs that offer a causal way to understand neural system alterations in psychiatric illness. (iii) We articulated the developments in biophysically based computational modeling studies that provide a platform for testing specific synaptic alterations. In turn, we demonstrate that such models can potentially be scaled to the level of neural systems to make relevant predictions for both restingstate or task-based connectivity experiments. We argue that the ongoing blend of these three approaches can provide advances in the field of clinical neuroscience that create a final output that is much greater than the sum of its parts.

## **ACKNOWLEDGMENTS**

Financial support for this study was provided by NIH grant DP5OD012109-01 (PI: Alan Anticevic), NIAAA grant 2P50AA012870-11 (PI: John H. Krystal), NIH grant MH096801 (PI: Michael W. Cole), NIH grant R01 MH062349 (PI: Xiao-Jing Wang) the National Alliance for Research on Schizophrenia and Depression (PIs: Alan Anticevic), the Fulbright Foundation (Aleksandar Savic), and the Yale Center for Clinical Investigation (YCCI, PI: Alan Anticevic). We thank Sharif Youssef for assistance with this manuscript.

## **REFERENCES**


[18F]fluorodeoxyglucose (FDG). *Eur Neuropsychopharmacol* (1997) **7**:9–24. doi:10.1016/S0924-977X(96)00039-9


new evidence. *Annu Rev Pharmacol Toxicol* (2001) **41**:237–60. doi:10.1146/ annurev.pharmtox.41.1.237


blockade to dopaminergic and cognitive disruptions associated with the prefrontal cortex. *J Neurosci* (1997) **17**:2921–7.


**Conflict of Interest Statement:** John H. Krystal consults for several pharmaceutical and biotechnology companies with compensation less than \$10,000 per year. All other authors declare that they have no conflict of interest.

*Received: 15 September 2013; paper pending published: 12 October 2013; accepted: 04 December 2013; published online: 24 December 2013.*

*Citation: Anticevic A, Cole MW, Repovs G, Savic A, Driesen NR, Yang G, Cho YT, Murray JD, Glahn DC, Wang X-J and Krystal JH (2013) Connectivity, pharmacology, and computation: toward a mechanistic understanding of neural system dysfunction in schizophrenia. Front. Psychiatry 4:169. doi: 10.3389/fpsyt.2013.00169*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2013 Anticevic, Cole, Repovs, Savic, Driesen, Yang , Cho, Murray, Glahn, Wang and Krystal. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*

# Dysfunction and dysconnection in cortical–striatal networks during sustained attention: genetic risk for schizophrenia or bipolar disorder and its impact on brain network function

#### **Vaibhav A. Diwadkar <sup>1</sup>\*, Neil Bakshi <sup>1</sup>† , Gita Gupta<sup>1</sup> , Patrick Pruitt <sup>1</sup>† , RichardWhite<sup>1</sup> and Simon B. Eickhoff 2,3**

<sup>1</sup> Department of Psychiatry and Behavioral Neurosciences, Wayne State University, Detroit, MI, USA

2 Institute of Clinical Neuroscience and Medical Psychology, Heinrich-Heine University Düsseldorf, Düsseldorf, Germany

3 Institute of Neuroscience and Medicine (INM-1), Research Center Jülich, Jülich, Germany

#### **Edited by:**

André Schmidt, University of Basel, Switzerland

#### **Reviewed by:**

Karl Friston, University College London, UK Paolo Brambilla, University of Udine, Italy

#### **\*Correspondence:**

Vaibhav A. Diwadkar, Department of Psychiatry and Behavioral Neurosciences, Wayne State University School of Medicine, Suite 5B, Tolan Park Medical Building, 3901 Chrysler Drive, Detroit, MI 48201 USA e-mail: vdiwadka@med.wayne.edu

#### **†Present address:**

Neil Bakshi, School of Medicine, University of Michigan, Ann Arbor, MI, USA; Patrick Pruitt, Department of Neuroscience, University of Michigan, Ann Arbor, MI, USA

Abnormalities in the brain's attention network may represent early identifiable neurobiological impairments in individuals at increased risk for schizophrenia or bipolar disorder. Here, we provide evidence of dysfunctional regional and network function in adolescents at higher genetic risk for schizophrenia or bipolar disorder [henceforth higher risk (HGR)]. During fMRI, participants engaged in a sustained attention task with variable demands. The task alternated between attention (120 s), visual control (passive viewing; 120 s), and rest (20 s) epochs. Low and high demand attention conditions were created using the rapid presentation of two- or three-digit numbers. Subjects were required to detect repeated presentation of numbers. We demonstrate that the recruitment of cortical and striatal regions are disordered in HGR: relative to typical controls (TC), HGR showed lower recruitment of the dorsal prefrontal cortex, but higher recruitment of the superior parietal cortex. This imbalance was more dramatic in the basal ganglia.There, a group by task demand interaction was observed, such that increased attention demand led to increased engagement in TC, but disengagement in HGR. These activation studies were complemented by network analyses using dynamic causal modeling. Competing model architectures were assessed across a network of cortical–striatal regions, distinguished at a second level using randomeffects Bayesian model selection. In the winning architecture, HGR were characterized by significant reductions in coupling across both frontal–striatal and frontal–parietal pathways. The effective connectivity analyses indicate emergent network dysconnection, consistent with findings in patients with schizophrenia. Emergent patterns of regional dysfunction and dysconnection in cortical–striatal pathways may provide functional biological signatures in the adolescent risk-state for psychiatric illness.

**Keywords: attention, brain networks, schizophrenia, bipolar disorder, dynamic causal modeling abstract**

## **INTRODUCTION**

Sustained attention or the ability to remain consistently focused on an ongoing task is one of the most basic of cognitive domains (1, 2) and serves as a fundamental process underlying mechanisms of memory and control (3). Attention competence in childhood and adolescence increases through emergence of functional integration within cortical–striatal circuits. The engagement of frontal regions has been documented in children as young as 4–6 years of age (4) and the maturation of the circuit (including the basal ganglia and the parietal lobe) extends through adolescence (3, 5). This multi-node attention network (6) includes executive regions of the frontal lobe (the dorsal prefrontal cortex and the dorsal anterior cingulate), regions such as the basal ganglia (including the caudate and the putamen) that presumably play central roles in relaying information between and linking signals across brain networks (7, 8), and the parietal lobe that is essential for mechanisms of spatial orientation (9). The ascent of attention competence in adolescence

corresponds with linear progression in the development of and anatomical connectivity between these key brain structures in the attention network (10, 11).

Deficits in sustained attention deficits are widely implicated in several psychiatric disorders that are adolescent onset or the origins of which lie in adolescence. These include not only core attention-related disorders such as attention deficit hyperactivity disorder (6, 12), but also bipolar disorder (13) and schizophrenia (14). The evidence regarding schizophrenia and bipolar disorder is compelling as studies now suggest attention deficits serve as a prelude in adolescence to the emergence of these late adolescent or adult-onset phenotypes. In this framework, adolescents with known risk-factors for psychiatric illness may present with neuropsychological deficits, which in turn are expressions of emergent dysfunction in critical brain networks (15). Adolescent children of parents with psychiatric diagnoses (mood disorders or schizophrenia) are an important risk group in whom familial risk may impact the integrity of function in attention networks, in turn decreasing the integrity of attention-related processing and subsequently leading to an increase in expressed attention deficits. In fact, adolescent children of parents with major depressive disorder, bipolar disorder, or schizophrenia all show deficits in neuropsychological tasks of attention including continuous performance tasks (CPT) (14, 16, 17) and other tasks with significant attention components (17). These groups are at significantly higher risk (HGR) for the emergence of psychiatric disorders (18–21). Consequently, a better understanding of the neurobiological impairments of attentional networks may provide important insight and potential biomarkers for the emergence of these disorders.

However, understanding of these biological bases remains obscure. Volumetric studies imply cortical–striatal reductions in brain structure (22, 23) that may be associated with impaired attention function (24, 25). However, the relationship between brain structure and function (as measured with structural and functional MRI, respectively) is not straightforward (26). This limits insight into disordered brain function in adolescence and its implication for psychiatric illness. In turn, understanding disordered effective connectivity between brain regions using causal modeling of brain network interactions assumes particular importance for understanding dysregulated networks in psychiatric disorders (27–29).

Effective connectivity mediates the integration of information between brain regions and refers to the "the influence that one neural system exerts over another, either at a synaptic (i.e., synaptic efficacy) or population level" (27, 30). Assessing brain activations and effective connectivity respectively permit exploration of relative specialization and functional integration of information in the brain (27). The temporal properties of the BOLD response, and the relationship of this to biophysical forward models of the neuronal response (31) permit the modeling of and inference on parameters of effective connectivity estimated from fMRI (32). While different methods for analyzing effective connectivity exist, dynamic causal modeling (DCM) is the currently best evaluated and most widely used approach toward this endeavor (32–35).

Our aims in this investigation were twofold. First, we assessed differences in regional responses across the extended cortical– striatal attention network (36, 37) including frontal, striatal, and parietal cortices. These differences in part constitute differences in the regional specialization of function between groups. Next, using DCM (33), we investigated differences between cortical–striatal network interactions using a competitive network identification framework based on Bayesian model selection (BMS) (38) and comparisons of Bayesian parameter averages (34, 39). fMRI data were collected in children and adolescents (8 years ≤Age < 20 years) with a family history of psychiatric illness (bipolar disorder or schizophrenia) (henceforth HGR) and controls free of such history to the second degree [henceforth typical controls (TC)]. During the fMRI task, extended attention blocks (120 s) were employed using a variant of the wellestablished CPT (identical pairs version, CPT-IP) (40) in which subjects must monitor rapidly presented stimuli (in the current context numbers were used) and indicate repetitions in the sequence.

## **MATERIALS AND METHODS**

### **SUBJECTS**

A total of 46 children and adolescents provided informed consent or assent for the fMRI studies approved by the institutional review board at Wayne State University. Of these 46, 24 were TC, with no family history of schizophrenia or mood disorder to the second degree and remaining 22 had a parent with schizophrenia or bipolar disorder and hence at HGR. Subjects were recruited from the greater Detroit area through advertisements and through in patient services at Wayne State University School of Medicine. Screening questionnaires administered using both telephone and personal interviews were used for both rule-outs and to ascertain if subjects had a history of psychotic illness in first-degree relatives. Diagnoses for parents of HGR were reached using the Structured Clinical Interview for DSM-IV schizophrenia (41). Subjects younger than 15 years were clinically characterized using the Schedule for Affective Disorders and Schizophrenia-Child Version (K-SADS) (42); those aged 15 years or above were assessed using the SCID. **Table 1** provides information on subject demographics and characteristics.

## **fMRI**

Functional data were acquired using a full body Bruker MedSpec 4.0 T system running the Siemens Syngo console. Gradient echo planar images (EPI) were collected using an eight-channel head coil (TR = 2000 ms; TE = 30 ms; matrix size = 64 × 64; field of view (FOV) = 240 mm; voxel size = 3.75 mm × 3.75 mm × 4 mm). Images were axially acquired in 24 continuous 4 mm slices positioned parallel to the anterior commissure/posterior commissure (AC–PC) line.

### **TASK**

During fMRI, all subjects performed a modified version of a CPT (Identical Pairs version) previously employed in studying illnesses including schizophrenia and bipolar disorder, and children and adolescents at risk for psychiatric illness (14, 16, 40). Numbers were presented in rapid sequence (50 ms, 250 ms SOA in each condition) and subjects were required to detect the repeated presentation of a number. Attention demands of the task were maintained by manipulating figure-ground contrast (white characters, RGB: 255, 255, 255; Off-white background, RGB: 225, 225, 225) in order to preempt attention gain under maximal contrast (43). Attention load was manipulated across epochs utilizing sequences of two-digit numbers ("low" load) or three-digit numbers ("high" load), motivated by evidence suggesting that access to numerosity

### **Table 1 | Demographic information for the investigate sample is shown**.


HGR were healthy apart from the following co-morbidities: separation anxiety (n = 1), attention deficit hyperactivity disorder (n = 3), and social phobia (n = 1). TC by definition were healthy and free of diagnosis.

though rapid (44) interacts with attention systems in the frontal, striatal, and parietal regions (45–47). A goal of manipulating load was to investigate separable load-related effects on region-specific interactions in each experimental group, particularly as parametric variations in load have proven useful in assessing differential regional specialization in risk and disease (48, 49).

To ensure large effect sizes of continuous or sustained attention, we used very long blocks of 120 s (therefore in removing low frequency drifts and fluctuations in subsequent analyses, we used a lenient high pass filter to preserve attention-related responses in the fMRI signals further noted in the fMRI analyses section below). Target frequency during the 120 s experimental epochs was 25%. In addition to experimental epochs, we also employed corresponding two- or three-digit control epochs (for each corresponding level of demand). During these epochs subjects passively observed two- or three-digit strings ("00" or "11"; "000" or "111"). Pure rest epochs (20 s) were also interspersed throughout the experimental run. Subjects signaled responses by button press on a standard response box. A schematic of the task is presented in **Figure 1**.

#### **fMRI PROCESSING (ACTIVATION ANALYSES)**

Data were processed with Statistical Parametric Mapping (SPM8). Realignment was performed to correct for head motion artifact during the scan. Realigned images were normalized to the Montreal Neurological Institute (MNI) EPI template and voxels resliced (2 mm × 2 mm × 2 mm). Normalized images were smoothed using an 8 mm FWHM Gaussian kernel. Images where estimated motion exceeded 4 mm were discarded from the analyses (<1% of all images).

In the first (within-subject) level analyses, rest, control, and attention epochs were modeled with boxcar stimulus functions that were convolved with a canonical hemodynamic response function to form regressors. Serial correlations were modeled with an auto-regressive process and low frequency fluctuations were removed with a high pass filter (using a discrete cosine set covering

frequencies of 1/256 s or lower). Note that we did not model phasic or event related responses to targets. This was because we were primarily interested in the responses associated with sustained attention.

First level contrasts for each level of demand relative to the corresponding control condition (Attention > Control) were computed for each individual subject. That is, we contrasted the beta-estimates for the low-attention condition with those for passively viewing two-digit strings and those for the high attention condition with those for passively viewing three-digit strings. This was performed to identify responses to attention-related (as opposed simply to visual) processing. First level maps were submitted to second level analyses of covariance with Group (HGR, TC) as the independent factor, demand (two-digit vs. three-digit) as non-independent factor, and age, gender, and task performance (assessed with *d* 0 )(50) as covariates. Clusters of activation (*p* < 0.05, cluster level corrected for multiple comparisons) (51) were employed to identify significant brain regions for each of the effects.

#### **fMRI PROCESSING (LINEAR DCM ANALYSES)**

More formal coverage of DCM can be found elsewhere (33, 52, 53). Briefly, DCM allows the interpretation of causal interaction between hidden state variables (32). The brain is viewed as a bilinear input (experimental conditions) – output (fMRI measured hemodynamic response) system. Changes in the neural responses are modeled using the following state differential equation:

$$\frac{d\mathbf{x}}{dt} = \left(A + \sum\_{j=1}^{m} \mu\_j B^{(j)}\right) \mathbf{x} + C\boldsymbol{\mu}.$$

where, *A* represents task-independent endogenous coupling between regions, *B* (*j*) represents putative modulation of endogenous connections by experimental manipulations (e.g., Attention, *uj*), and *C* represents sensorimotor driving inputs on (typically) unimodal cortical regions.

A goal of DCM is to identify model(s) with the highest evidence given the observed fMRI data by testing competing hypotheses on a model space (54). Therefore, assessment of effective connectivity using DCM requires evaluation and comparison of neurobiologically plausible competing models, each representing hypotheses on the connective-architecture of the investigated neural system.

The *a priori* attention network of interest included regions both within the executive network (dACC, dPFC, and caudate nucleus) and sensory and spatial attention-related regions (parietal cortex and visual cortex) (6, 36, 37, 55). The particular focus of the modeling space (competing hypothesis) was the role of the dorsal anterior cingulate cortex and the contextual modulation of its efferent connections to other regions of the attention network. This approach was motivated in large part by the significant role played by the dACC in cognitive and resource control as it relates to attention and conflict (56), and its particular place in the control-related hierarchy of the forebrain (57, 58). Notably, disordered cognitive control has emerged as a general framework for understanding the schizophrenia and bipolar diathesis which the at-risk participants in our sample fall under (29, 59–61). DCM

was implemented using DCM8 in SPM8. An *a priori* network of nodes was derived using regions of interests in stereotactic space (62). Within each structurally defined node of this network, we summarized regional activity on a subject-specific basis employing the principal eigenvariate of voxels within a 5 mm radius of the peak. **Figure 2** shows the resulting architecture.

#### **MODEL ESTIMATION**

Prior to modeling, time series were extracted from each region of interest (ROI) according to established procedures (63, 64) using spheres (5 mm radius) centered on the peak of the "effects of interest" *F*-contrast (*p*FWE < 0.05, adjusted for "effects of no interest"). Each of the 72 models was estimated across subjects. To determine the most likely generative model, a random-effects (RFX) BMS procedure was applied. In particular, we used the variational Bayes method to estimate posterior probabilities of competing models. Bayesian parameter averages of coupling estimates (with a focus on modulatory coupling) were analyzed to determine potential differences in modulation as a function of attention and statistical significance was assessed using Bonferroni correction (*p* < 0.05)(32, 34).

## **RESULTS**

#### **BEHAVIORAL RESULTS**

Behavioral performance, which is the sensitivity to distinguish targets from distracters was assessed using *d* 0 (65), an established metric in Signal Detection theory (50,66). The metric incorporates the hit-rate (e.g., the rate of responding "same" to successively presented stimuli in the same valence category) and the false alarm rate (e.g., the rate of responding "different" to successively presented stimuli in difference valence categories), and is based on the difference between the inverse function of the cumulative Gaussian distribution applied to each, with a higher *d* 0 reflecting greater sensitivity to the task.

Behavioral data were analyzed in a repeated measures analysis of variance with Group (HGR vs. TC) as between subjects' factor and attention demand (two-digit vs. three-digit) as within-subjects factor. The main effect of load was significant indicating that attention load reduced the sensitivity of observers, *F*1,45 = 11.67, *p* < 0.001, MSe = 0.25. A main effect of group was marginally significant, *F*1,45 = 2.86, *p* < 0.05, onetailed, MSe = 1.71 suggesting that subjects with a family history of psychiatric illness were marginally less sensitive than controls. No other effects reached significance. **Figure 3** depicts performance data across conditions and groups.

## **ACTIVATION ANALYSES WITH fMRI (DIFFERENCES IN REGIONAL SPECIALIZATION OF FUNCTION)**

A significant main effect of Group (HGR 6= TC) and Group × Demand interaction was investigated in the constituent regions across the network of interest. Significant clusters under the main effect were observed in both the dorsal prefrontal cortex and the parietal lobe (*p* < 0.05, cluster level) and significant clusters under the interaction term were observed in the basal ganglia. Directionality (HGR 6= TC) of the statistical effects and the interaction terms were inferred based on estimates of the modeled responses extracted under the overall peak within the cluster of significance.

First, relative to TC, HGR subjects evinced reduced engagement of the dorsal prefrontal cortex irrespective of the degree of attention demand. **Figure 4** depicts significant clusters rendered on lateral and medial surfaces of the cortex. By comparison, HGR evinced increased engagement of the parietal cortex irrespective of the degree of attention demand (**Figure 5**).

In addition to the main effect of group a significant Group × Demand interaction in the basal ganglia (**Figure 6**). As seen in the accompanying graph of the modeled responses, the interaction resulted from an increase in BG engagement with increases in load with a corresponding decrease in engagement in HGR. These activation results suggest that genetic risk confers an imbalance in the patterns of *relative specialization* of attentionrelated function in adolescence, in particular with diminished engagement on executive regions of the network including the dPFC and the BG, but aberrantly increased reliance on the parietal cortex. The DCM results provide a notable complement for these activation-based analyses by demonstrating the effects of genetic risk in adolescence on the *functional integration* of information across regional networks for attention.

## **DCM ANALYSES OF fMRI DATA (DIFFERENCES IN FUNCTIONAL INTEGRATION)**

Random-effects analyses and BMS revealed a single winning model in each of TC and HGR. **Figure 7** depicts model structure (specifically the pathways modulated by attention) and the observed exceedance probabilities for each of the TC and HGR groups. Notably, these results suggest that the likeliest generative models of the data did not differ across groups, with attention modulating the dACC efferents to the BG and the Parietal cortex. This convergence of model structure implies that any differences in effective connectivity between TC and HGR were to be expected in the parameter estimates of endogenous coupling, or contextual modulation of that coupling by attention (32).

To test for group differences we used the Bayesian parameter average over subjects within each group. This is appropriate because the best model was the same for both groups and

therefore a comparison of the group-specific Bayesian parameter averages is unbiased by differences in Bayesian selection. This procedure provides posterior densities over the effective connectivity parameters for both groups, enabling one to estimate the difference between group means and posterior confidence in those differences (shown in terms of a posterior standard error in the figures). Group differences significant at a corrected level of *p* < 0.05, Bonferroni corrected (constituting, *p* < 0.003 for each of the 13 tests) are indicated (\*). These *P* values were based upon the posterior differences in group-specific Bayesian parameter averages – and their significance can be visualized in terms of posterior standard errors in the **Figures 8** and **9** below.

**Figure 8** depicts observed estimates of endogenous coupling for each of the pairwise connections modeled across the endogenous network. As seen, the results provide an admixture of excitatory and inhibitory coupling across network pairs across the task. The most notable and symmetric finding was the bi-directional hypo-connectivity in the dACC↔BG pathway observed in HGR compared to TC (matched shaded insets). Notably, relative to TC, in HGR virtually every dACC efferent pathway was characterized by hypo-connectivity, suggesting convergence with hypothesis on the dysfunctional role of the dACC in schizophrenia and mood disorders. In addition, we also observed a difference between TC

and HGR on dPFC↔dACC and the dACC↔Parietal pathways, with TC characterized by inhibitory coupling but HGR characterized by excitatory coupling (former) and decreased inhibitory coupling (latter).

We also observed notably differences in the attention-related contextual modulation of the efferent pathways from the dACC to the BG and the parietal lobe (**Figure 9**). In both cases, HGR were characterized by attention-related dysmodulation, albeit differing in character. Firstly, during attention epochs the dACC↔BG pathway was inhibited in HGR but increased in TC. Secondly, the dACC↔Parietal pathway was increase during attention in both

groups, but the degree of modulation was reduced in HGR. In the remainder of the paper, we discuss the import of these results in inferring the role of genetic risk on brain networks for attention, and the interpretation of the relationships between the analyses of relative specialization differences (activation) and functional integration differences (effective connectivity). We also reflect on the import role of network analyses of fMRI data in inferring accurate profiles of psychiatric risk in brain networks.

## **DISCUSSION**

Assessing activation and effective connectivity differences between TC and HGR revealed striking differences in (a) the regional brain responses and interactive effects of attention demand, and (b) patterns of estimated endogenous and contextual effective connectivity between specific sub-networks, particularly related to dACC efferents. Activation analyses revealed an imbalance in regional brain function in HGR: the degree of dPFC engagement was reduced, with an apparent shifting in the relative degree of engagement to the parietal cortex. Furthermore, the BG in TC was responsive to variations in attention load, but disengaged in HGR. These activation-derived imbalances in regional recruitment in HGR suggest a relative shift away from relying on the dPFC and the BG core regions of the executive attention network (37, 67, 68), and toward regions such as the parietal lobe that may be more associated with spatial attention and orientation (69). These activation-based analyses provide a degree of convergence with fMRI patterns observed in adult patients with frank phenotypes of psychosis or mood disorders. For example, forebrain areas in

schizophrenia appear hypo-active during conscious and rapid (as opposed to deliberative) processing tasks that engage attention (40, 70, 71). Moreover, in stimulus-response integration tasks with significant attentional demand, regional profiles of engagement are shifted in schizophrenia toward the parietal cortex (72). Similarly, adult bipolar patients are characterized by decreased basal ganglia activity during sustained attention and thalamus during a sustained attention task (73). Given the unique and integrative role of regions such as the dPFC and the basal ganglia in corticostriatal loops subserving attention and memory (74), this disengagement may reflect the role of genetic risk in creating a latent functional deficit in adolescence that alters the relative specialization of function of attention-related regions. This latent deficit is an important vulnerability marker of predisposition to disorders of psychosis or mood (75–77).

Whereas the observed activation-related in HGR appear to foreshadow adult studies in schizophrenia and bipolar disorder, evidence of disordered effective connectivity in HGR constitutes an entirely novel line of inquiry into the functional network biology attention networks and their relationship to risk (29). In general, applications of effective connectivity analyses of fMRI data in adult schizophrenia or bipolar patients has been fruitful in revealing disordered connectivity in tasks of learning (78), sentence completion tasks (79), emotion processing (80), and working

memory (81). Perhaps the closest predecessor related to the current set of results is recent work investigating working memory related disordered effective connectivity in young individuals in the prodromal state for schizophrenia (35). In these subjects, reduced frontal–parietal connectivity during working memory (and intermediate between TC and schizophrenia patients) and the implied reduction in functional integration within these critical brain circuits may be predictive of the eventual transition to psychosis. We note that the prodromal state is itself a unique risk-state, constituting an advanced stage of (non-specific) clinical symptoms, and therefore distinct from the HGR group assessed herein. Nevertheless a significant proportion of HGR subjects are likely to transition toward frank phenotypes by way of prodromal symptoms. In this clinical/sub-clinical context, we highlight two points of plausible convergence. Firstly, reduced dACC↔BG endogenous connectivity in HGR may reflect a latent dyscoupling in the dormant risk-state that may impair the scaling up of cortical networks to implement higher order tasks. Given the important role of this sub-network in tasks as diverse as memory, attention, motor and cognitive control, and skill learning (82–86), it is likely that a connectivity deficit in this sub-circuit will foreshadow likely deficits in a slew of psychological domains. Indeed, it is unsurprising that in general large neuropsychological assessments of HGR indicate widespread impairments in

neurocognitive domains, most of which rely in some form on attention processing (21, 23, 87, 88).

Disordered contextual modulation of dACC↔BG and dACC↔Parietal efferent pathways provides a parallel and likewise intriguing aspect of network-related dysfunction, particularly as assessments of contextual modulation provide a highly unique contribution of DCM to the study of brain network dynamics and systems theory (33, 89). In this regard, reduced (positive) modulation of the dACC↔Parietal pathway and strong inhibition of the dACC↔BG pathway in HGR are suggestive of differences of attention-related implementation in the same network. Given that these parameters represent an increase or decrease in connection strength as a function of the implemented task, the inhibition of the dACC↔BG pathway indicates the disengagement of this interaction in response to attention processing. As this disengagement is contrary to the expected excitation in TC, it suggests that in addition to being hypo-connected, the dACC↔BG sub-network is also "turned down" during attention processing. This turning down (and the disordered response to load) suggests that frontal–striatal network function is suboptimal in the risk-state. This is consistent with the relationship between dopamine and frontal–striatal function (90), the developmental tuning of the dopamine response, and the relevance of frontal–striatal dopamine dysfunction for schizophrenia and risk for schizophrenia (91).

## **RELATIVE SPECIALIZATION AND FUNCTIONAL INTEGRATION IN HGR**

Both the more conventional assessment of regional activation strength and the more advanced analysis of effective connectivity revealed impaired cortical–striatal signals in HGR. Both of them, however, provide unique insight. Activation-based approaches with a general linear model framework do not explicit distinguish between network and/or task constituents (e.g., endogenous connections, modulation by task, sensory inputs) and from the perspective of system's theory, these approaches are slightly incomplete (89). Thus, the observed disordered activations in HGR are neutral in revealing network-based dysfunction underlying genetic risk and provide more general assessments of differences in the relative engagement of brain regions. We also note that activation-based approaches did not identify HGR-TC differences in regions such as the dACC, perhaps reflecting a limitation in classical statistical approaches to fMRI.

By comparison, DCM is limited by *a priori* assumptions in the assessed network and the structure of the model space. It nevertheless has proven to be more sensitive in identifying abnormal biological signatures in risk-groups, where activation analyses were not. For example, using DCM we recently documented disordered cortical–limbic endogenous connectivity and contextual modulation during an emotional appraisal task in children of schizophrenia parents (34). Notably, this finding emerged despite widespread overlap in activation networks across risk and control groups. DCM thus proved to be highly sensitive in uncovering emergent impairments in functional brain organization, not apparent in regional brain activation patterns or behavioral. It has hence repeatedly been proposed that effective connectivity is an important aspect of research on high-risk samples, not only because of its more realistic interactional model but also because of its reliance on Bayesian statistics (38, 92). By contrast, the absence or restriction of significant regional effects in classical statistics may partly be related to its premises of minimizing the type I error (while Bayesian statistics rely on the highest posterior probability) (28, 32).

#### **LIMITATIONS AND PROSPECTUS**

We conclude with a brief note on the limitations of the study, and a brief note on the prospective role of fMRI in high-risk research. The present study is limited by the relatively small sample size, though this limitation is slightly offset by the robustness of the results, particularly the effective connectivity analyses. Moreover, we acknowledge that HGR is a heterogeneous group, which may explain the more heterogeneous pattern of model evidence for HGR (**Figure 7**) compared to TC. This heterogeneity is well known and has been characterized before with MRI (25, 36, 93). Moreover, even though children of schizophrenia and bipolar patients are not distinct in terms of attention impairment (assessed with neuropsychological measures) (14), it is plausible that implementation of attention in brain networks may differ. We acknowledge that these differences are not knowable in the current analyses on account of sample size limitations. Also, in assessing effective connectivity, in this first approach, we did not explicitly model effects of attention load (providing a point of asymmetry with the activation-based analyses), though we are currently augmenting our analyses to investigate these effects.

The study of adolescents at genetic risk for schizophrenia or bipolar disorder offers opportunities and challenges. As indicated previously, HGR constitute a unique risk group, distinct from prodromal or clinical high-risk samples (94–97). Studying HGR in the medication naïve state can provide interesting insights into the intersection of genetic risk and abnormal neurodevelopment (98, 99). By focusing on a *profile of cumulative genetic risk*, rather than on *individual genes,* such approaches are important given the polygenic and non-specific genetic bases of psychiatric disorders (100, 101). However, the emergence of frank phenotypes (typically in early adulthood) is mediated by a host of unknown and uncontrolled factors (102), and neurobiological signatures in HGR may be non-specific and carry uncertain predictive value. Nevertheless, we suggest that a focus on critical domains such as sustained attention, and understanding of brain network dysfunction underlying these domains in HGR may provide a particularly fruitful path forward in understanding how genetically mediated vulnerability is encoded in disordered brain network interactions.

#### **AUTHOR CONTRIBUTIONS**

Vaibhav A. Diwadkar designed and supervised the project, data acquisition and all analyses and interpretation. Neil Bakshi conducted DCM analysis. Gita Gupta conducted activation bases analysis. Patrick Pruitt and Richard White assisted in data

collection and analyses. Simon B. Eickhoff collaborated with Vaibhav A. Diwadkar in design, analyses, and interpretation.

## **ACKNOWLEDGMENTS**

Vaibhav A. Diwadkar acknowledges support from the National Institute of Mental Health (MH6860), the National Alliance for Research on Schizophrenia and Depression (NARSAD), the Children's Research Center of Michigan (CRCM), the Children's Hospital Foundation, the Prechter Pediatric Bipolar Program World Heritage Foundation, and the Lyckaki-Young Fund from the State of Michigan. Additional support was provided by a Career Development Chair from the Office of the President, Wayne State University. We thank Jacqueline Radwan and Eric Murphy for assistance in data analyses and collection, Dalal Khatib for assistance in data collection, and R. Rajarathinam, Al Pizzuti, Caroline Zajac-Benitez, Usha Rajan, and M. S. Keshavan for assistance in recruitment and characterization.

## **REFERENCES**


*Prog Neuropsychopharmacol Biol Psychiatry* (2011) **35**:1349–54. doi:10.1016/j. pnpbp.2011.04.009


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

*Received: 01 November 2013; accepted: 28 April 2014; published online: 09 May 2014. Citation: Diwadkar VA, Bakshi N, Gupta G, Pruitt P, White R and Eickhoff SB (2014) Dysfunction and dysconnection in cortical–striatal networks during sustained attention: genetic risk for schizophrenia or bipolar disorder and its impact on brain network function. Front. Psychiatry 5:50. doi: 10.3389/fpsyt.2014.00050*

*This article was submitted to Schizophrenia, a section of the journal Frontiers in Psychiatry.*

*Copyright © 2014 Diwadkar, Bakshi, Gupta, Pruitt , White and Eickhoff. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.*