A Bayesian hierarchical model to detect differentially methylated loci from single nucleotide resolution sequencing data

Abstract

DNA methylation is an important epigenetic modification that has essential roles in cellular processes including gene regulation, development and disease and is widely dysregulated in most types of cancer. Recent advances in sequencing technology have enabled the measurement of DNA methylation at single nucleotide resolution through methods such as whole-genome bisulfite sequencing and reduced representation bisulfite sequencing. In DNA methylation studies, a key task is to identify differences under distinct biological contexts, for example, between tumor and normal tissue. A challenge in sequencing studies is that the number of biological replicates is often limited by the costs of sequencing. The small number of replicates leads to unstable variance estimation, which can reduce accuracy to detect differentially methylated loci (DML). Here we propose a novel statistical method to detect DML when comparing two treatment groups. The sequencing counts are described by a lognormal-beta-binomial hierarchical model, which provides a basis for information sharing across different CpG sites. A Wald test is developed for hypothesis testing at each CpG site. Simulation results show that the proposed method yields improved DML detection compared to existing methods, particularly when the number of replicates is low. The proposed method is implemented in the Bioconductor package DSS.

Effector CD8 T cells dedifferentiate into long-lived memory cells

“…The random set of regions was then annotated the same way as the original DMRs. Differential methylation analysis of CpG methylation among the datasets were further assessed using a Bayesian hierarchical model to detect differences among methylation at 3 CpG sites 36 .…”

A flexible, efficient binomial mixed model for identifying differential DNA methylation in bisulfite sequencing data

“…Effectively, our model improves and generalizes the beta-binomial model by introducing this extra g term to model individual relatedness due to population structure or stratification. In the absence of g, our model becomes similar to other beta-binomial models previously developed for modeling count data [31,33,47,62].…”

“…As a result, DNA methylation levels will frequently covary with kinship or population structure, and failure to account for this covariance could lead to spurious associations or reduced power to detect true effects. This phenomenon has been extensively documented for genotypephenotype association studies [35,36,[40][41][42], and controlling for genetic covariance between No DSS [31], MOABS [32], RadMeth [33] Linear mixed model No Yes Yes GEMMA [34], EMMA [35], EMMAX [36], FaST-LMM [37] Binomial mixed model Yes Yes Yes MACAU samples is now a basic requirement for genome-wide association studies. Similar logic applies to analyses of gene regulatory phenotypes and studies of gene expression variation often do take genetic structure into account by using mixed model approaches [43][44][45].…”

“…To address this problem, several recently introduced methods for differential DNA methylation analysis implement a beta-binomial model (e.g., 'DSS: Dispersion Shrinkage for Sequencing data' [31], 'RADMeth: Regression Analysis of Differential Methylation' [33], and 'MOABS: Model Based Analysis of Bisulfite Sequencing data' [32]). These methods model the binomial nature of bisulfite sequencing data, while taking into account the well-known problem of overdispersion in sequencing reads.…”

“…Because these methods work directly on count data, they can reliably account for variation in read coverage across sites and individuals. Consequently, betabinomial methods consistently provide increased power to detect true associations between genetic or environmental sources of variance and DNA methylation levels [31][32][33].…”