Evaluation of Recent Statistical Methods for Detecting Differential Methylation Using BS-seq Data

Abstract

Whole genome profiling of differential DNA methylation between diseased and normal samples has significant implications in research to understand the role of epigenetic regulations of cells. In recent years, the development of bisulfite sequencing (BS-seq)-based molecular technology has enabled the measurement of DNA methylation at a nucleotide resolution throughout the genome. Given the availability of this new type of DNA methylation data, certain features challenge traditional analytical methods such as the Fisher exact test, and a number of new statistical methods have recently been proposed for analyzing them. However, the relative performances of many of these methods are still not fully understood, despite a couple of recent review papers. In order to arm researchers with knowledge so that they can select the best-suited method and software for analyzing their data, we selected and evaluated four methods that are able to account for between-sample variability, which is a prominent feature of DNA methylation data. The commonality and differences of these methods in terms of several important characteristics, including modeling, data filtering, statistical testing, and covariate handling were reviewed. Simulation studies as well as applications to two datasets were carried out to compare and contrast the methods. To be as unbiased as possible, our simulation modeling took into account characteristics of real data but did not use any of the specific models in these four methods.