Assessing Computational Methods for Transcription Factor Target Gene Identification Based on ChIP-seq Data

Chromatin immunoprecipitation coupled with deep sequencing (ChIP-seq) has great potential for elucidating transcriptional networks, by measuring genome-wide binding of transcription factors (TFs) at high resolution. Despite the precision of these experiments, identification of genes directly regulated by a TF (target genes) is not trivial. Numerous target gene scoring methods have been used in the past. However, their suitability for the task and their performance remain unclear, because a thorough comparative assessment of these methods is still lacking. Here we present a systematic evaluation of computational methods for defining TF targets based on ChIP-seq data. We validated predictions based on 68 ChIP-seq studies using a wide range of genomic expression data and functional information. We demonstrate that peak-to-gene assignment is the most crucial step for correct target gene prediction and propose a parameter-free method performing most consistently across the evaluation tests.Author Summary: Transcription factors (TFs) are the main regulators of gene transcription. Thus, knowing the genes that are targeted by a specific TF is of utmost importance for understanding developmental processes, cellular stress response, or disease etiology. Chromatin immunoprecipitation coupled with deep sequencing (ChIP-seq) allows for measuring the genome-wide binding of TFs. Several computational methods have been used for inferring the genes that are targeted by TFs using this binding information, but a thorough evaluation of their performance has not been performed so far. Here we present an assessment of a range of TF-target-calling methods using 68 ChIP-seq datasets. It turns out that the first step of the scoring, the assignment of binding events to genes, is the most important for correctly calling target genes. Our evaluation revealed important performance differences between the target-calling methods, with some simplistic methods exhibiting a particularly poor performance compared to more elaborate scorings. One of the methods is particularly attractive, because it does not require the a priori definition of any parameter — all parameters are ‘learned’ from the data. This and other methods tested were implemented in a freely available software package for future testing and application to other ChIP-seq datasets.