eScreen: a deep learning framework for functionally decoding the regulatory genome at single-nucleotide resolution

The human genome is densely populated with cis-regulatory elements (CREs), yet deciphering their functional regulatory syntax and combinational logic remains a fundamental challenge. Here, we integrate 379 genome-scale CRISPR screen experiments, encompassing 21 million perturbations across 23 cell types, to construct a compendium of 41,239 high-confidence functional CREs from 530,527 candidates. Leveraging this resource, we develop eScreen, a deep learning model built on the StripedHyena2 architecture to functionally decode the regulatory genome at single-nucleotide resolution. eScreen achieves three primary functions: (1) predicts genome-wide cell-type-specific CRE functional activity with high accuracy, outperforming existing models; (2) provides mechanistic interpretation of regulatory syntax at single-nucleotide resolution; (3) dissects the functional organization of enhancer clusters through in silico perturbation analysis. We perform multiple independent CRISPR knockout, CRISPR interference (CRISPRi), and base editing screens to validate these functions of eScreen both at scale and on individual cases. Furthermore, we provide an interactive web server (https://escreen.huanglabxmu.com/) for the community to access the integrated CRISPR screen resources and eScreen functions. Collectively, our work establishes a highly precise and convenient tool to decode the causal effects of the regulatory genome.