Hierarchical Bayesian Models for ChIP-seq Data

Histone modifications (HMs) are post-translational modifications of the nucleosome. Studying the presence or absence of these modifications in genomic regions is a central topic in modern epigenetics. HMs regulate various biological processes by overwriting the DNA-inscribed code. Experimental evidence suggests that they perform this task through a complex biological network. In other words, HMs combinatorially influence gene expression. We present two model-based approaches to decode this mechanism using ChIP-seq data. Both approaches are based on hierarchical Bayesian models. The first model derives a conditional independence structure among the HMs through a graphical model. The challenge here is to model the unobserved binary (presence/absence) status of HMs on the basis of read counts. The other critical aspect is to model the dependence between these latent binaries in a way that allows tractable posterior inference. The second model relates HMs and functional genomics through a local bi-clustering approach. Here HMs are clustered and each HM cluster gives rise to a (nested) partition of genomic loactions, with respect to that subset of HMs. These models are, to the best of our knowledge, the first model-based fully Bayesian approaches to discovering epigenetic associations. Validation with known experimental findings suggests the importance and usefulness of these approaches in our understanding of gene regulation.