Interpretable deep learning framework for mapping E3–substrate binding interfaces
Dianke Li, Yuting Zhang, Yuan Liu, Zihao Zhang, Yingjie Qu, Jiajun Li, Linyang Jiang, Lihong Diao, Ziding Zhang, Lingqiang Zhang, Chun-Ping Cui, Dong Li · Nature Communications · 2026
E3 ubiquitin ligases recognize substrates through specific interfaces. Accurate delineation of these interfaces is essential, as mutations disrupting them impair protein ubiquitination and drive cancer progression. However, available E3-substrate interface data are sparse and systematic prediction methods remain lacking. Here, we propose MetaESI, a deep learning framework that simultaneously predicts E3-substrate interactions and leverages its interpretable architecture to infer binding interfaces de novo. With a two-stage meta-learning strategy, MetaESI generalizes across diverse E3s and achieves state-of-the-art performance in both interaction and interface prediction. We applied MetaESI at the proteome scale to generate MetaESI-Atlas, which comprises 68,056 annotated interactions across eight species. Integrating multi-omics data, we identified mutations at MetaESI-predicted interfaces that disrupt E3-substrate binding, and experimentally validated representative examples including JunB Q244E and SPOP F102C as oncogenic drivers. By combining interpretable AI with mechanistic insight, MetaESI establishes a methodological paradigm for interpretable model design and a foundational resource for precision oncology and targeted protein degradation.