PInet

PInet predicts protein-protein interaction (PPI) interfaces from pairs of protein point-cloud structures to identify atomic-level interaction sites for studying molecular recognition and supporting therapeutic design.

Key Features:

• Unified Geometric Deep Neural Network (GDNN): Integrates data-driven and physics-based methodologies to combine machine learning with physical modeling for interface prediction.
• Point Cloud Representation: Processes pairs of point clouds representing 3D structures of interacting protein partners to capture structural details required for interface prediction.
• Patch-wise Attention Mechanism: Applies patch-wise attention to focus on specific protein surface regions likely to mediate interactions.
• Geometrical and Physicochemical Complementarity Modeling: Learns geometrical and physicochemical complementarities between molecular surfaces to ground predictions in molecular recognition principles.
• Joint Segmentation for Interface Prediction: Performs joint segmentation on protein surface representations to simultaneously predict interface regions on both interacting proteins.

Scientific Applications:

• Large-scale high-throughput PPI mapping: Predicts interaction interfaces across many protein pairs when experimental determination of complex structures is impractical.
• Atomic-level interface annotation for drug design: Provides atomic-level interface predictions to inform design of therapeutics targeting specific protein-protein interactions.
• Mechanistic interpretation of molecular recognition: Identifies interface regions and complementarities to generate hypotheses about interaction mechanisms.

Methodology:

Consumes point cloud data of protein pairs and applies a Geometric Deep Neural Network integrating data-driven and physics-based modeling with a patch-wise attention mechanism and joint segmentation to learn geometrical and physicochemical complementarities; performance has been benchmarked against state-of-the-art predictors.