A comparison of peak callers used for DNase-Seq data

Genome-wide profiling of open chromatin regions using DNase I and high-throughput sequencing (DNase-seq) is an increasingly popular approach for finding and studying regulatory elements. A variety of algorithms have been developed to identify regions of open chromatin from raw sequence-tag data, which has motivated us to assess and compare their performance. In this study, four published, publicly available peak calling algorithms used for DNase-seq data analysis (F-seq, Hotspot, MACS and ZINBA) are assessed at a range of signal thresholds on two published DNase-seq datasets for three cell types. The results were benchmarked against an independent dataset of regulatory regions derived from ENCODE in vivo transcription factor binding data for each particular cell type. The level of overlap between peak regions reported by each algorithm and this ENCODE-derived reference set was used to assess sensitivity and specificity of the algorithms. Our study suggests that F-seq has a slightly higher sensitivity than the next best algorithms. Hotspot and the ChIP-seq oriented method, MACS, both perform competitively when used with their default parameters. However the generic peak finder ZINBA appears to be less sensitive than the other three. We also assess accuracy of each algorithm over a range of signal thresholds. In particular, we show that the accuracy of F-Seq can be considerably improved by using a threshold setting that is different from the default value.