Predicting genomic interactions using deep learning

Abstract

Physical promoter-enhancer and CTCF-CTCF interactions organize the human genome in 3-dimensions, and contribute to the regulation of gene expression. Hi-C and related approaches have enabled profiling of these interactions, though how the instructions for these interactions are encoded in the genome is still largely not understood. We develop a deep learning model, Deep3DGenome, to predict genomic interactions using both genomic sequence data and chromatin features. We find that a machine learning model that has anchor specific modules and uses rich chromatin features outperforms previous approaches at predicting 3D interactions.