A Nonparametric Method for Detecting Differential DNA Methylation Regions

DNA methylation has long been known as an epigenetic gene silencing mechanism. For a motivating example, the methylomes of cancer and non-cancer cells show a number of methylation differences, indicating that some of cancer cells aggressive properties could be a result of certain methylation features. Robust methods for detecting differentially methylated regions (DMRs) could help scientists narrow down biologically important regions in the genome. Despite the number of statistical methods developed for detecting DMR, there is no default or strongest method. Fisher's exact test is direct, but is inadequate for data with multiple replications, while regression-based methods often come with a large number of assumptions. More complicated methods have been proposed, but those are often difficult to interpret. In this paper, we propose a three-step nonparametric kernel smoothing method that is both flexible and straightforward to implement and interpret. The proposed method relies on local quadratic fitting method to find the set of equilibrium points (points at which the first derivative is 0) and the corresponding set of confidence windows. The potential regions are further refined using biological criteria, and selected based on a Bonferroni adjusted t-test cutoff. Using a comparison of three senescent and three proliferating cell lines to illustrate our method, we were able to identify a total of 1,077 DMRs on chromosome 21.