Nonparametric Bayesian clustering to detect bipolar methylated genomic loci

Abstract

Background: With recent development in sequencing technology, a large number of genome-wide DNA methylation studies have generated massive amounts of bisulfite sequencing data. The analysis of DNA methylation patterns helps researchers understand epigenetic regulatory mechanisms. Highly variable methylation patterns reflect stochastic fluctuations in DNA methylation, whereas well-structured methylation patterns imply deterministic methylation events. Among these methylation patterns, bipolar patterns are important as they may originate from allele-specific methylation (ASM) or cell-specific methylation (CSM).

Results: Utilizing nonparametric Bayesian clustering followed by hypothesis testing, we have developed a novel statistical approach to identify bipolar methylated genomic regions in bisulfite sequencing data. Simulation studies demonstrate that the proposed method achieves good performance in terms of specificity and sensitivity. We used the method to analyze data from mouse brain and human blood methylomes. The bipolar methylated segments detected are found highly consistent with the differentially methylated regions identified by using purified cell subsets.

Conclusions: Bipolar DNA methylation often indicates epigenetic heterogeneity caused by ASM or CSM. With allele-specific events filtered out or appropriately taken into account, our proposed approach sheds light on the identification of cell-specific genes/pathways under strong epigenetic control in a heterogeneous cell population.

Figure 1

Promoter methylation pattern of prodynorphin gene in human brain [ 27 ]. This figure shows a genomic region of 95 bp with 9 CpG dinucleotides. Most reads are completely unmethylated (open circles) or methylated (filled circles), but only a few reads are with both unmethylated and methylated cytosines.

Figure 2

Comparison between DPM search, k -means and Bayesian mixture clustering in Simulation III. A. Average mis-classification rates. B. Average power. These results are obtained from 1,000 simulations under the alternative hypothesis, for different number of reads and for different cell-type proportions, using DPM search, k-means and Bayesian mixture clustering.

Figure 3

Analysis of mouse brain methylomes. A. Venn diagram shows the relationships between (a) bipolar segments identified from pooled dataset, (b) DMR identified between neuron and glia, (c) bipolar segments identified from glia, and (d) bipolar segments identified from neuron. B. Gene ontology analysis of genes associated with bipolar methylated segments in neuron and glia datasets, respectively. P-values for GO enrichment were adjusted with Bonferroni correction.

Figure 4

Analysis of human blood methylomes. A. Venn diagram shows the relationships between (a) bipolar segments identified from pooled dataset, (b) DMR identified between B cell and neutrophil, (c) bipolar segments identified from B cell, and (d) bipolar segments identified from neutrophil. B. Gene ontology analysis of genes associated with bipolar methylated segments in B cell and neutrophil datasets, respectively. P-values for GO enrichment were adjusted with Bonferroni correction.