Recent years have seen rapid strides in 3D chromosomal conformation capture technologies, with the development of improved Hi-C protocols and techniques like HiChIP to enrich for specific types of 3D interactions, all enabling the generation of high resolution interaction data sets. However, the computational analysis of these data presents many challenges, and resolving how 3D genomic architecture influences gene regulation remains an important open problem.
We will present recent work in the lab (1) to improve the computational analysis of 3D interaction data and (2) to incorporate 3D genomic architecture into predictive models of gene regulation using deep learning. First, we will describe HiC-DC+, our new computational tool for identifying significant and differential 3D interactions from Hi-C and HiChIP data sets, and we will show how more rigorous statistical analysis empowers biological interpretation. Second, we will present a graph attention network framework to incorporate 3D interactions in predictive models of gene regulation, using 1D epigenomic data with or without genomic DNA sequence together with 3D connectivity to predict gene expression. We use feature attribution on these models to infer functional enhancer-promoter interactions, outperforming the state-of-the-art approach to this problem.