AUTHOR=Buhule Olive D. , Minster Ryan L. , Hawley Nicola L. , Medvedovic Mario , Sun Guangyun , Viali Satupaitea , Deka Ranjan , McGarvey Stephen T. , Weeks Daniel E.
TITLE=Stratified randomization controls better for batch effects in 450K methylation analysis: a cautionary tale
JOURNAL=Frontiers in Genetics
VOLUME=5
YEAR=2014
URL=https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2014.00354
DOI=10.3389/fgene.2014.00354
ISSN=1664-8021
ABSTRACT=Background: Batch effects in DNA methylation microarray experiments can lead to spurious results if not properly handled during the plating of samples.
Methods: Two pilot studies examining the association of DNA methylation patterns across the genome with obesity in Samoan men were investigated for chip- and row-specific batch effects. For each study, the DNA of 46 obese men and 46 lean men were assayed using Illumina's Infinium HumanMethylation450 BeadChip. In the first study (Sample One), samples from obese and lean subjects were examined on separate chips. In the second study (Sample Two), the samples were balanced on the chips by lean/obese status, age group, and census region. We used methylumi, watermelon, and limma R packages, as well as ComBat, to analyze the data. Principal component analysis and linear regression were, respectively, employed to identify the top principal components and to test for their association with the batches and lean/obese status. To identify differentially methylated positions (DMPs) between obese and lean males at each locus, we used a moderated t-test.
Results: Chip effects were effectively removed from Sample Two but not Sample One. In addition, dramatic differences were observed between the two sets of DMP results. After “removing” batch effects with ComBat, Sample One had 94,191 probes differentially methylated at a q-value threshold of 0.05 while Sample Two had zero differentially methylated probes. The disparate results from Sample One and Sample Two likely arise due to the confounding of lean/obese status with chip and row batch effects.
Conclusion: Even the best possible statistical adjustments for batch effects may not completely remove them. Proper study design is vital for guarding against spurious findings due to such effects.