BIDCHIPS: bias decomposition and removal from ChIP-seq data clarifies true binding signal and its functional correlates.

Background: Unraveling transcriptional regulatory networks is a central problem in molecular biology and, in this quest, chromatin immunoprecipitation and sequencing (ChIP-seq) technology has given us the unprecedented ability to identify sites of protein-DNA binding and histone modification genome wide. However, multiple systemic and procedural biases hinder harnessing the full potential of this technology. Previous studies have addressed this problem, but a thorough characterization of different, interacting biases on ChIP-seq signals is still lacking.

Results: Here, we present a novel framework where the genome-wide ChIP-seq signal is viewed as being quantifiably influenced by different, measurable sources of bias, which can then be computationally subtracted away. We use a compendium of 123 human ENCODE ChIP-seq datasets to build regression models that tell us how much of a ChIP-seq signal can be attributed to mappability, GC-content, chromatin accessibility, and factors represented in input DNA and IgG controls. When we use the model to separate out these non-binding influences from the ChIP-seq signal, we obtain a purified signal that associates better to TF-DNA-binding motifs than do other measures of peak significance. We also carry out a multiscale analysis that reveals how ChIP-seq signal biases differ across different scales. Finally, we investigate previously reported associations between gene expression and ChIP-seq signals at transcription start sites. We show that our model can be used to discriminate ChIP-seq signals that are truly related to gene expression from those that are merely correlated by virtue of bias-in particular, chromatin accessibility bias, which shows up in ChIP-seq signals and also relates to gene expression.

Conclusions: Our study provides new insights into the behavior of ChIP-seq signal biases and proposes a novel mitigation framework that improves results compared to existing techniques. With ChIP-seq now being the central technology for studying transcriptional regulation, it is most crucial to accurately characterize, quantify, and adjust for the genome-wide effects of biases affecting ChIP-seq. Our study also emphasizes that properly accounting for confounders in ChIP-seq data is of paramount importance for obtaining biologically accurate insights into the workings of the complex regulatory mechanisms in living organisms. R and MATLAB packages implementing the framework can be obtained from http://www.perkinslab.ca/Software.html.