Identifying differential transcription factor binding in ChIP-seq

ChIP seq is a widely used assay to measure genome-wide protein binding. The decrease in costs associated with sequencing has led to a rise in the number of studies that investigate protein binding across treatment conditions or cell lines. In addition to the identification of binding sites, new studies evaluate the variation in protein binding between conditions. A number of approaches to study differential transcription factor binding have recently been developed. Several of these methods build upon established methods from RNA-seq to quantify differences in read counts. We compare how these new approaches perform on different data sets from the ENCODE project to illustrate the impact of data processing pipelines under different study designs. The performance of normalization methods for differential ChIP-seq depends strongly on the variation in total amount of protein bound between conditions, with total read count outperforming effective library size, or variants thereof, when a large variation in binding was studied. Use of input subtraction to correct for non-specific binding showed a relatively modest impact on the number of differential peaks found and the fold change accuracy to biological validation, however a larger impact might be expected for samples with more extreme copy number variations between them. Still, it did identify a small subset of novel differential regions while excluding some differential peaks in regions with high background signal. These results highlight proper scaling for between-sample data normalization as critical for differential transcription factor binding analysis and suggest bioinformaticians need to know about the variation in level of total protein binding between conditions to select the best analysis method. At the same time, validation using fold-change estimates from qRT-PCR suggests there is still room for further method improvement.