ScanNet: an interpretable geometric deep learning model for structure-based protein binding site prediction.

Area of Science:

• Structural bioinformatics
• Computational biology
• Machine learning in structural biology

Background:

• Predicting protein functional sites (e.g., binding sites) is crucial for understanding protein function.
• Current methods like feature-based machine learning and comparative modeling have limitations in feature expressivity and data availability.

Purpose of the Study:

• Introduce ScanNet, an end-to-end geometric deep learning model for predicting protein functional sites directly from 3D structures.
• Evaluate ScanNet's performance in identifying protein-protein and protein-antibody binding sites.
• Apply ScanNet to predict functional sites on the SARS-CoV-2 spike protein.

Main Methods:

• Developed ScanNet, an interpretable geometric deep learning model that learns features from 3D protein structures.
• Represented atoms and amino acids based on the spatio-chemical arrangement of their neighbors.
• Trained and validated ScanNet on binding site prediction tasks, including proteins with novel folds.

Main Results:

• ScanNet demonstrated high accuracy in predicting protein-protein and protein-antibody binding sites.
• The model showed strong performance even on protein folds not seen during training.
• Interpretable filters learned by ScanNet provide insights into its predictions.
• Predicted epitopes on the SARS-CoV-2 spike protein, confirming known antigenic regions and identifying novel ones.

Conclusions:

• ScanNet is a versatile, powerful, and interpretable tool for functional site prediction in proteins.
• The model advances the field by learning features directly from 3D structural data.
• ScanNet has potential applications in drug discovery and understanding viral mechanisms, as exemplified by its use on SARS-CoV-2.