A Deep Learning Model for Predicting Transcription Factor Binding Location at Single Nucleotide Resolution

Abstract

Transcriptional regulation by transcription factors (TFs) plays a pivotal role in controlling the gene expression. However, understanding the mechanism through which the transcription factors regulate the gene expression is a challenging task. This is primarily hindered by the low specificity in identifying transcription factor binding sites (TFBS). The emergence of the ChIP-exonuclease (ChIP-exo) method enables the detection of TFBS at single nucleotide sensitivity, providing us an opportunity to study the detailed mechanisms of TF regulation. Nevertheless, there is still a lack of computational tools that can also provide single base pair (bp) resolution prediction of TFBS. In this paper, we propose DeepSNR, a Deep Learning algorithm for Single Nucleotide Resolution prediction of transcription factor binding site. Our proposed method is inspired by the similarity between predicting the specific binding location from input nucleotide sequence and image segmentation. Particularly, we adopted the deconvolution network (deconvNet); a deep learning model designed for image segmentation, and developed a TFBS specific deconvNet architecture constructed on top of 'DeepBind'. We trained a deconvNet for predicting CTCF binding sites using the data from ChIP-exo experiments. The proposed algorithm achieved median precision and recall of 87% and 77% respectively, significantly outperforming motif search based algorithms such as MatInspector.