ME-Class2 reveals context dependent regulatory roles for 5-hydroxymethylcytosine.

Since the discovery of 5-hydroxymethylcytosine (5hmC) as a prominent DNA modification found in mammalian genomes, an emergent question has been what role this mark plays in gene regulation. 5hmC is hypothesized to function as an intermediate in the demethylation of 5-methylcytosine (5mC) and in the reactivation of silenced promoters and enhancers. Further, weak positive correlations are observed between gene body 5hmC and gene expression. We previously demonstrated that ME-Class is an effective tool to understand relationships between whole-genome bisulfite sequencing data and expression. In this work, we present ME-Class2, a machine-learning based tool to perform integrative 5mCG, 5hmCG and expression analysis. Using ME-Class2 we analyze whole-genome single-base resolution 5mCG and 5hmCG datasets from 20 primary tissue and cell samples to reveal relationships between 5hmCG and expression. Our analysis indicates that conversion of 5mCG to 5hmCG within 2 kb of the transcription start site associates with distinct functions depending on the summed level of 5mCG + 5hmCG. Unchanged levels of 5mCG + 5hmCG (conversion from 5mCG to stable 5hmCG) associate with repression. Meanwhile, decreases in 5mCG + 5hmCG (5hmCG-mediated demethylation) associate with gene activation. Our results demonstrate that ME-Class2 will prove invaluable to interpret genome-wide 5mC and 5hmC datasets and guide mechanistic studies into the function of 5hmCG.

Modeling complex patterns of differential DNA methylation that associate with gene expression changes.

Numerous genomic studies are underway to determine which genes are abnormally regulated by DNA methylation in disease. However, we have a poor understanding of how disease-specific methylation changes affect expression. We thus developed an integrative analysis tool, Methylation-based Gene Expression Classification (ME-Class), to explain specific variation in methylation that associates with expression change. This model captures the complexity of methylation changes around a gene promoter. Using 17 whole-genome bisulfite sequencing and RNA-seq datasets from different tissues from the Roadmap Epigenomics Project, ME-Class significantly outperforms standard methods using methylation to predict differential gene expression change. To demonstrate its utility, we used ME-Class to analyze 32 datasets from different hematopoietic cell types from the Blueprint Epigenome project. Expression-associated methylation changes were predominantly found when comparing cells from distantly related lineages, implying that changes in the cell's transcriptional program precede associated methylation changes. Training ME-Class on normal-tumor pairs from The Cancer Genome Atlas indicated that cancer-specific expression-associated methylation changes differ from tissue-specific changes. We further show that ME-Class can detect functionally relevant cancer-specific, expression-associated methylation changes that are reversed upon the removal of methylation. ME-Class is thus a powerful tool to identify genes that are dysregulated by DNA methylation in disease.

Application of Bayesian network structure learning to identify causal variant SNPs from resequencing data.

Using single-nucleotide polymorphism (SNP) genotypes from the 1000 Genomes Project pilot3 data provided for Genetic Analysis Workshop 17 (GAW17), we applied Bayesian network structure learning (BNSL) to identify potential causal SNPs associated with the Affected phenotype. We focus on the setting in which target genes that harbor causal variants have already been chosen for resequencing; the goal was to detect true causal SNPs from among the measured variants in these genes. Examining all available SNPs in the known causal genes, BNSL produced a Bayesian network from which subsets of SNPs connected to the Affected outcome were identified and measured for statistical significance using the hypergeometric distribution. The exploratory phase of analysis for pooled replicates sometimes identified a set of involved SNPs that contained more true causal SNPs than expected by chance in the Asian population. Analyses of single replicates gave inconsistent results. No nominally significant results were found in analyses of African or European populations. Overall, the method was not able to identify sets of involved SNPs that included a higher proportion of true causal SNPs than expected by chance alone. We conclude that this method, as currently applied, is not effective for identifying causal SNPs that follow the simulation model for the GAW17 data set, which includes many rare causal SNPs.