ABSTRACT
Motivation DNA methylation is an important epigenetic modification related to a variety of diseases including cancers. One of the key issues of methylation analysis is to detect the differential methylation sites between case and control groups. Previous approaches describe data with simple summary statistics and kernel functions, and then use statistical tests to determine the difference. However, a summary statistics-based approach cannot capture complicated underlying structure, and a kernel functions-based approach lacks interpretability of results.
Results We propose a novel method D3M, for detection of differential distribution of methylation, based on distribution-valued data. Our method can detect high-order moments, such as shapes of underlying distributions in methylation profiles, based on the Wasserstein metric. We test the significance of the difference between case and control groups and provide an interpretable summary of the results. The simulation results show that the proposed method achieves promising accuracy and outperforms previous methods. Glioblastoma multiforme and lower grade glioma data from The Cancer Genome Atlas and show that our method supports recent biological advances and suggests new insights.
Availability R implemented code is freely available from https://cran.r-project.org/web/packages/D3M/ https://github.com/cran/D3M.
Contact ymatsui{at}med.nagoya-u.ac.jp
