Enhancer prediction in the human genome by probabilistic modelling of the chromatin feature patterns

Abstract

Background: The human genome is far from completely annotated. Specifically, the locations ofgenedistal regulatory enhancers are difficult to locate. Enhancers are binding sites of transcription factors and occupied by nucleosomes with modified histones. The binding sites of transcription factors (TFs) and the localization of histone modifications can be quantified by the chromatin immunoprecipitation assay coupled with next generation sequencing (ChIP-seq). The resulting data has been successfully adopted for genomewide enhancer identification by several unsupervised and supervised machine learning methods. However, the current methods predict different numbers and different sets of enhancers for the same cell type, and they do not utilize the pattern of the ChIP-seq coverage profiles efficiently. It is also difficult to estimate the accuracy and specificity of the genome-wide enhancer predictions. Results:We developed PREPRINT, a PRobabilistic Enhancer PRedictIoN Tool. We considered the pattern of, for example, the ChIP-seq coverage profile around the enhancers. The data at the positive and negative examples of enhancers was utilized to probabilistically model the enhancer coverage pattern and to train a kernel-based classifier. We demonstrated the performance of the method using ENCODE data from two cell lines. The predicted enhancers were computationally validated based on the TFs and co-regulatory factor binding sites. We compared our enhancer predictions to the ones obtained by other methods. The effects of different parameter choices during training, testing and validation were studied, and finally, the approach to validate the genome-wide predictions was investigated. Conclusion: PREPRINT performed comparably to the state-of-the-art methods and provided probabilistic interpretation (i.e. uncertainty) for the predictions. PREPRINT generalized to data from cell type not utilized for training, and often the performance of PREPRINT was superior to RFECS. We observed that the choice of training data and the choice of parameter values at different steps of the enhancer prediction and validation influenced on the final set of predictions. PREPRINT identifed biologically validated enhancers not predicted by the competing methods. The enhancers predicted by PREPRINT can aid the genome interpretation in functional genomics and clinical studies.