Interactions between core histone marks and DNA methyltransferases predict DNA methylation patterns observed in human cells and tissues

Abstract

DNA methylation and histone modifications are two major epigenetic marks in mammalian cells. Previous studies have revealed that these two mechanisms interact although a quantitative model of these is still lacking in mammalian cells. Here we sought to develop such a model by systematically evaluating the quantitative relationship between DNA methylation and the core histone modification marks in human epigenomes. This model reflects the interactions of ADD and PWWP domains of DNA methyltransferase (DNMTs) with histone 3 lysine tails. Our analysis integrated 35 whole genome bisulphite sequencing data sets (about 800 million CpG sites), 35 chromatin states and 175 ChIP-Seq histone modification profiles across 35 human cell types. The logistic regression model we built shows that more than half of the variance across DNA methylomes can be explained by the five-core histone modification across varied types of human cell and tissue samples. Importantly, we find that H3K4me3 has a dramatic effect in DNA methylation patterning, highlighting the essential interaction between ADD domain of DNMTs and histone 3 lysine 4 in human. Moreover, our model suggests DNA methylation is generally inhibited by the presence of H3K4me3, H3K4me1 and H3K27me3, while increased levels are found in regions that are marked by H3K9me3 and H3K36me3. In summary, our results provide a comprehensive evaluation of the crosstalk between DNA methylation and histone modification in a variety of human cell types, and shows that DNA methylation patterns can be largely explained by interactions between histone 3 lysine tails and specific domains of DNA methyltransferases.

Keywords: DNA methylation imputation; Epigenetic interaction; epigenetic modelling.

Figure 1.

Schematic illustration of the integrative approach.

Figure 2.

Relationships among DNA methylation and histone modifications in the integrated dataset. (a). Distribution of DNA methylation values in the integrated dataset. (b). Distribution of H3K4me3 normalized log fold change values in the integrated matrix. (c). Genome-wide Pearson correlations between the five histone modifications and DNA methylation values. (d). Boxplot of DNA methylation values in each learnt chromatin states (y axis). The dots within boxplots represent the medium methylation values.

Figure 3.

Evaluations of model performance. (a). Model performance based on different combinations of histone modifications as input variables. The x-axis represents the model performance evaluated by R_McFadden. (b). Scatterplot of observed DNA methylation and predicted DNA methylation values. The colour represents density counts of the dots. (c). An example genomic region of the predicated DNA methylation, observed DNA methylation and its corresponding histone modification signal. (d). Model performance in each 35 cell types. The y-axis represents the names of the cell types. The legend represents the corresponding tissue for each cell type.

Figure 4.

Characteristics of residual DNA methylation values. (a). Distribution of residual DNA methylation values (Predicted minus observed). The two squares represent the regions where residual values have variations larger than 0.5. (b). Receive operation curve (ROC) analysis of the power for the model in predicting hypo- and hypermethylated CpG sites. (c). Density plot of residual DNA methylation values in each chromatin states. (d). Gene ontology enrichment analysis for mis-predicted (high in predictions) CpG sites for repressed polycomb chromatin state regions. (e). DNA motif enrichment analysis for mis-predicted (high in predictions) CpG sites for bivalent enhancers and repressed polycomb chromatin state regions.