Features of ChIP-seq data peak calling algorithms with good operating characteristics

Abstract

Author descriptionReuben Thomas is a Staff Research Scientist in the Bioinformatics Core at Gladstone Institutes\n\nSean Thomas is a Staff Research Scientist in the Bioinformatics Core at Gladstone Institutes\n\nAlisha K Holloway is the Director of Bioinformatics at Phylos Biosciences, visiting scientist at Gladstone Institutes and Adjunct Assistant Professor in Biostatistics at the University of California, San Francisco.\n\nKatherine S Pollard is a Senior Investigator at Gladstone Institutes and Professor of Biostatistics at University of California, San Francisco.\n\nKey PointsO_LIPeak-calling using Chip-seq data consists of two sub-problems: identifying candidate peaks and testing candidate peaks for statistical significance.\nC_LIO_LITwelve features of the two sub-problems of peak-calling methods are identified.\nC_LIO_LIMethods that explicitly combine the signals from ChIP and input samples are less powerful than methods that do not.\nC_LIO_LIMethods that use windows of different sizes to scan the genome for potential peaks are more powerful than ones that do not.\nC_LIO_LIMethods that use a Poisson test to rank their candidate peaks are more powerful than those that use a Binomial test.\nC_LI\n\nAbstractChromatin immunoprecipitation followed by sequencing (ChIP-seq) is an important tool for studying gene regulatory proteins, such as transcription factors and histones. Peak calling is one of the first steps in analysis of these data. Peak-calling consists of two sub-problems: identifying candidate peaks and testing candidate peaks for statistical significance. We surveyed 30 methods and identified 12 features of the two sub-problems that distinguish methods from each other. We picked six methods (GEM, MACS2, MUSIC, BCP, TM and ZINBA) that span this feature space and used a combination of 300 simulated ChIP-seq data sets, 3 real data sets and mathematical analyses to identify features of methods that allow some to perform better than others. We prove that methods that explicitly combine the signals from ChIP and input samples are less powerful than methods that do not. Methods that use windows of different sizes are more powerful than ones that do not. For statistical testing of candidate peaks, methods that use a Poisson test to rank their candidate peaks are more powerful than those that use a Binomial test. BCP and MACS2 have the best operating characteristics on simulated transcription factor binding data. GEM has the highest fraction of the top 500 peaks containing the binding motif of the immunoprecipitated factor, with 50% of its peaks within 10 base pairs (bp) of a motif. BCP and MUSIC perform best on histone data. These findings provide guidance and rationale for selecting the best peak caller for a given application.