ProtGeoNet-Pocket: A Binding Site Prediction Approach Integrating Sequence, Geometry, and Graph Structure.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Protein binding site recognition is crucial for understanding biological functions and drug discovery.
• Existing methods often struggle with the complex 3D structures and diverse shapes of protein binding pockets.
• Integrating multimodal data (sequence, geometry, graph) offers a promising avenue for improved prediction accuracy.

Purpose of the Study:

• To develop an innovative multimodal prediction framework, ProtGeoNet-Pocket, for enhanced protein binding site recognition.
• To leverage multiscale structural information to address protein structural complexity and binding pocket shape diversity.
• To achieve high-precision binding site prediction through effective feature integration and advanced network architectures.

Main Methods:

• Utilized a PointNet module for extracting geometric features from residue coordinates, enhanced by an attention mechanism.
• Fused geometric features with encoded sequence information and graph edge features.
• Employed a Graph Isomorphism Network (GIN) with a message-passing mechanism to capture topological relationships.

Main Results:

• Achieved an F1 score of 72.87% on the scPDB training set.
• Demonstrated strong predictive performance across five independent benchmark datasets (COACH420, HOLO4K, SC6K, PDBbind, ApoHolo).
• Visualization confirmed high spatial overlap between predicted and actual binding sites, indicating superior performance.

Conclusions:

• ProtGeoNet-Pocket effectively integrates multimodal data for accurate protein binding site prediction.
• The framework demonstrates superior performance compared to existing methods.
• This approach holds significant potential for advancing drug discovery and protein function studies.