UniversalEPI: a generalized attention-based deep ensemble model to accurately predict enhancer-promoter interactions across diverse cell types and conditions

Interactions between enhancers and gene promoters provide insights into gene regulation. Experimental techniques, including Hi-C, that map these enhancer-promoter interactions (EPIs), have high costs and labor requirements, which limits their use. Therefore, in silico methods have been developed to predict EPIs computationally, but there are challenges with the generalizability and accuracy of existing methods. Here, we introduce UniversalEPI, an attention-based deep ensemble model designed to provide uncertainty-aware predictions of EPIs up to 2 Mb apart, which can generalize across unseen cell types using only DNA sequence and chromatin accessibility (ATAC-seq) data. Benchmarking shows that UniversalEPI significantly outperforms existing approaches in accuracy and efficiency, even though it is a lightweight model that only assesses interactions between accessible chromatin regions. UniversalEPI enables statistical comparison of predicted interactions across conditions, which we demonstrated by tracking the dynamics of EPIs during human macrophage activation. We also used UniversalEPI to assess chromatin dynamics between different cancer cell states in human esophageal adenocarcinoma. Thus, UniversalEPI advances the accuracy and applicability of in silico 3D chromatin modeling to investigate chromatin dynamics in development and disease.