Detection of Differentially Methylated Regions Using Bayesian Curve Credible Bands

Abstract

DNA methylation is one of the most crucial epigenetic modifications involved in regulating gene transcription, cellular differentiation, development, and disease. In recent years, aided by fast parallel sequencing technology, a number of genome-wide bisulfite sequencing platforms have been developed to provide high-throughput DNA methylation data. These are essentially short reads that can be classified as methylated or unmethylated for a particular CpG site. Numerous sophisticated statistical methods have been developed to analyze such a massive amount of correlated data, but they are mainly for detecting differentially methylated loci (DMLs). To detect differentially methylated regions (DMRs), which are often more relevant biologically, a post-processing step from the identified DMLs is needed. In this paper, we address this shortcoming and other issues by proposing a latent variable Bayesian smoothing Curve (BCurve) method for detecting DMRs directly by means of constructing Bayesian credible intervals and bands. In addition to direct detection of DMRs, BCurve differs from existing methods in several other aspects, including its ability to accommodate between-sample variability, taking correlation of methylation levels among nearby loci into account when detecting DMRs, and the construction of credible bands (not just point estimates). We carried out an extensive simulation study to evaluate the performance of BCurve and to compare it with an existing method, BSmooth. The results show that BCurve outperforms BSmooth in all scenarios considered. Finally, we applied BCurve to a dataset to illustrate its utility in real data applications.