Characterizing batch effects and binding site-specific variability in ChIP-seq data

Multiple sources of variability can bias ChIP-seq data toward inferring transcription factor (TF) binding profiles. As ChIP-seq datasets increase in public repositories, it is now possible and necessary to account for complex sources of variability in ChIP-seq data analysis. We find that two types of variability, the batch effects by sequencing laboratories and differences between biological replicates, not associated with changes in condition or state, vary across genomic sites. This implies that observed differences between samples from different conditions or states, such as cell-type, must be assessed statistically, with an understanding of the distribution of obscuring noise. We present a statistical approach that characterizes both differences of interests and these source of variability through the parameters of a mixed effects model. We demonstrate the utility of our approach on a CTCF binding dataset composed of 211 samples representing 90 different cell-types measured across three different laboratories. The results revealed that sites exhibiting large variability were associated with sequence characteristics such as GC-content and low complexity. Finally, we identified TFs associated with high-variance CTCF sites using TF motifs documented in public databases, pointing the possibility of these being false positives if the sources of variability are not properly accounted for.