Interpretable deep learning framework for mapping E3-substrate binding interfaces

Area of Science:

• Biochemistry and Molecular Biology
• Artificial Intelligence in Bioinformatics
• Genomics and Proteomics

Background:

• E3 ubiquitin ligases are crucial for protein ubiquitination, with mutations in their substrate interfaces driving cancer.
• Existing E3-substrate interface data are limited, hindering systematic prediction and analysis.
• There is a need for computational methods to accurately predict E3-substrate interactions and identify critical binding interfaces.

Purpose of the Study:

• To develop a deep learning framework (MetaESI) for simultaneous prediction of E3-substrate interactions and binding interfaces.
• To leverage MetaESI's interpretable architecture for de novo inference of E3-substrate binding interfaces.
• To create a comprehensive resource (MetaESI-Atlas) of E3-substrate interactions across multiple species.

Main Methods:

• A two-stage meta-learning strategy was employed within the MetaESI deep learning framework.
• MetaESI was applied at the proteome scale to predict interactions and interfaces.
• Multi-omics data integration was used to identify and validate mutations at predicted interfaces.

Main Results:

• MetaESI achieved state-of-the-art performance in predicting both E3-substrate interactions and binding interfaces.
• The MetaESI-Atlas database contains 68,056 annotated E3-substrate interactions across eight species.
• Mutations at predicted interfaces were identified as drivers of cancer, with specific examples (JunB Q244E, SPOP F102C) experimentally validated.

Conclusions:

• MetaESI provides a powerful, interpretable AI method for predicting E3-ubiquitin ligase-substrate interactions and binding interfaces.
• The MetaESI-Atlas serves as a foundational resource for precision oncology and targeted protein degradation.
• This work establishes a paradigm for interpretable AI model design in biological applications.