A novel Bayesian segmentation model for ChIP-seq data analysis and its practical applications

Chromatin immunoprecipitation followed by next-generation sequencing (ChIP-seq) is quickly gaining widespread adoption in studies of DNA-binding protein interactions. This technique is commonly used to investigate transcription factor binding sites, as well as profiling histone modifications. ChIP-seq generates many millions of short reads. Mapping these reads to the genome reveals regions of enrichment, which correspond to sites of occupancy for the immunoprecipitated protein. We present a novel Bayesian segmentation model and an associated estimation procedure that aims to identify these enriched areas. In contrast to existing methods, our model does not depend on arbitrary windows or parameter estimation through simulation methods like MCMC. Instead, our segments are data driven and we determine posterior means using explicit formulas. This can be performed in a reasonable time complexity due to our approximation algorithm, BCMIX, without significant loss in accuracy compared to the Bayes estimator. Evaluation of the novel method showed it performed in step with other algorithms in identifying TFBS but with higher resolution--successfully smoothing noise and self-selecting segments. Furthermore, it was capable of identifying broad, diffuse regions of enriched histone modifications. BCP, thus, represents a unified algorithm for a wide variety of ChIP-seq data analysis tasks. We highlighted this by applying BCP in an integrated analysis of various data types to determine the role of FOXO1 in Treg and naïve CD4+ T cells. This included peak-calling in transcription factor ChIP-seq data and island detection on histone modification ChIP-seq data and DNase-I hypersensitivity sequencing data, where BCP's unique versatility shined. We found that FOXO1 controlled a novel Treg program involved in mediating immune suppression. Additionally, we identified a set of FOXO1 target genes in naïve CD4+ T cells and a group of putative cofactors. Candidate binding sites showed a significant correlation with H3K4me3 and DNase I hypersensitivity, which will be interesting to investigate in more detail in future studies.