ScanNet

ScanNet predicts functional binding sites on protein 3D structures using interpretable geometric deep learning to identify small-molecule, protein, and antibody interaction sites relevant to drug discovery and immunology.

Key Features:

• End-to-End Geometric Deep Learning: Processes 3D structural data directly with an end-to-end geometric deep learning model without relying on handcrafted features.
• Interpretability: Produces interpretable filters that reveal spatio-chemical arrangements of atoms and amino acids that drive binding site predictions.
• Versatility Across Protein Folds: Detects protein–protein and protein–antibody binding sites with high accuracy, including on proteins with previously unseen folds.

Scientific Applications:

• Binding Site Prediction: Predicts locations on proteins where small molecules, other proteins, or antibodies are likely to bind by learning from 3D structures.
• Epitope Mapping: Applied to predict epitopes on the SARS-CoV-2 spike protein, validating known antigenic regions and identifying previously uncharacterized ones.

Methodology:

Constructs representations of atoms and amino acids from their spatial and chemical context in 3D protein structures and leverages the geometric arrangement of neighboring elements and learned filters within a geometric deep learning model to predict binding sites.