GraphPred: An approach to predict multiple DNA motifs from ATAC-seq data using graph neural network and coexisting probability

Abstract

Assay for Transposase-Accessible Chromatin sequencing (ATAC-seq) utilizes hyperactive Tn5 transposase to cut open chromatin and reveal chromatin accessibility at a genome-wide level. ATAC-seq can reveal more kinds of transcription factor binding regions than Chromatin immunoprecipitation sequencing (ChIP-seq) and DNase I hypersensitive sites sequencing (DNase-seq). Transcription factor binding sites (TFBSs) prediction is a crucial step to reveal the functions of TFs from the high throughput sequencing data. TFBSs of the same TF tend to be conserved in the sequence level, which is named motif. Several deep learning models based on the convolutional neural networks are used to find motifs from ATAC-seq data. However, these methods didnt take into account that multiple TFs bind to a given sequence and the probability that a fragment of a given sequence is a TFBS. To find binding sites of multiple TFs, we developed a novel GNN model named GraphPred for TFBSs prediction and finding multiple motifs using the coexisting probability of k-mers. In the light of the experiment results, GraphPred can find more and higher quality motifs from 88 ATAC-seq datasets than comparison tools. Meanwhile, GraphPred achieved an area of eight metrics radar (AEMR) score of 2.31.