Characterizing collaborative transcription regulation with a graph-based deep learning approach

Abstract

Human epigenome and transcription activities have been characterized by a number of sequence-based deep learning approaches which only utilize the DNA sequences. However, transcription factors interact with each other, and their collaborative regulatory activities go beyond the linear DNA sequence. Therefore leveraging the informative 3D chromatin organization to investigate the collaborations among transcription factors is critical. We developed ECHO, a graph-based neural network, to predict chromatin features and characterize the collaboration among them by incorporating 3D chromatin organization from 200-bp high-resolution Micro-C contact maps. ECHO predicted 2,583 chromatin features with significantly higher average AUROC and AUPR than the best sequence-based model. We observed that chromatin contacts of different distances affected different types of chromatin features' prediction in diverse ways, suggesting complex and divergent collaborative regulatory mechanisms. Moreover, ECHO was interpretable via gradient-based attribution methods. The attributions on chromatin contacts identify important contacts relevant to chromatin features. The attributions on DNA sequences identify TF binding motifs and TF collaborative binding. Furthermore, combining the attributions on contacts and sequences reveals important sequence patterns in the neighborhood which are relevant to target sequence's chromatin feature prediction. The attribution results that reveal TF collaboration activities are provided on a website https://echo.dcmb.med.umich.edu/echo/.

A generalizable framework to comprehensively predict epigenome, chromatin organization, and transcriptome

“…Furthermore, sequence patterns generated from attribution scores reflect TF binding patterns including binding motifs and co-binding patterns. The attribution scores of genomic sequences toward TF binding reflect the important regions that contribute to TF binding prediction, which can be used to generate sequence binding patterns using TF-MoDISco [30], and these sequence patterns are able to recover TFs’ known binding motifs or reflect other sequence binding patterns such as co-binding patterns [31, 32]. From sequence patterns generated for each TF using EPCOT, we first observed that sequence patterns of some TFs recovered their known binding motifs, for example SPI1 (Fig.2e).…”

A generalizable framework to comprehensively predict epigenome, chromatin organization, and transcriptome

“…Furthermore, sequence patterns generated from attribution scores reflect TF binding patterns including binding motifs and co-binding patterns. The attribution scores of DNA sequences toward TF binding reflect the important regions that contribute to the TF binding prediction, which can be used to generate sequence binding patterns, and these sequence patterns are able to recover TFs' known binding motifs or reflect other sequence binding patterns such as co-binding patterns [26,27,28]. From sequence patterns generated for each TF using EPCOT, we first observed that sequence patterns of some TFs recovered their known binding motifs, for example SPI1 (Fig.…”