Empirical bayes model comparisons for differential methylation analysis

Abstract

A number of empirical Bayes models were developed to investigate the differential methylation analysis. However, it is not clear which empirical Bayes model performs best in differential methylation analysis. In this paper, five empirical Bayes models were constructed and applied to the differential methylation analysis of A2780 cells between before and after 1, 3, and 5 round of cisplatin treatment. The log-normal model with the background variance showed the lowest minimized negative log-likelihood. It inferred increasing number of differentially methylated loci from 1 to 3 to 5 rounds of cisplatin treatment on the A2780 cells, which was consistent to cisplatin resistant IC50 data. Among differentially methylated loci selected from each empirical model, three time dependent methylation patterns were defined: stochastic hypomethylation, stochastic hypermethylation, and random methylation. If the empirical Bayes model of the DNA methylation assumed log-normal distribution, both stochastically hypomethylated loci and stochastically hypermethylated loci were enriched with a number of transcription factor binding sites. Almost no TFBS enrichment was observed if the gamma distribution was assumed in the empirical Bayes model. In summary, by comparing the performances of the differential methylation analysis and the TFBS enrichment analysis, log-normal distribution is a better statistical assumption than the gamma distribution in the empirical Bayes model.