UniversalEPI: Harnessing Attention Mechanisms to Decode Chromatin Interactions in Rare and Unexplored Cell Types

Abstract

Chromatin interactions between cis-regulatory elements (CREs) and specific gene promoters offer insights into the mechanisms of gene activation and repression. This knowledge is crucial for evaluating the effects of non-coding variants and chromatin alterations in cell differentiation or disease. Experimental techniques Hi-C and single-cell Hi-C provide information on CRE-promoter interactions, but the high costs and labor requirements limit their widespread use. We introduce UniversalEPI, an attention-based neural network designed to predict CRE-promoter interactions up to 2 Mb apart across unseen cell types without model retraining and using only DNA sequence and chromatin accessibility data as input. In UniversalEPI, chromatin accessibility profiles with masked non-accessible regions and DNA sequence data get passed through several convolution layers to identify binding motifs of three key transcription factors ubiquitously involved in promoter activation and CRE-promoter looping; a transformer block then takes the information from the convolutional levels passed through stochastic gating to predict chromatin interactions. While being a relatively lightweight model because of its architecture, UniversalEPI outperforms the state-of-the-art methods in inferring CRE-promoter interactions in unseen cell types across human cancerous and non-cancerous cell lines (Spearman’s Rho > 0.92). The model can also incorporate pseudo-bulk chromatin accessibility data from single-cell ATAC-seq experiments to pinpoint cell subtypespecific interactions. We illustrate the ability of UniversalEPI to track the dynamics of CRE-promoter interactions during macrophage reprogramming and the transition between cancer transcriptional states in esophageal adenocarcinoma. UniversalEPI thus advances in-silico 3D chromatin modeling, supporting studies of chromatin dynamics in cell differentiation and impacts of non-coding variants.