PLiSAGE: enhancing protein-ligand interaction prediction with multimodal surface and geometry encoding.

Area of Science:

• Computational biology
• Structural biology
• Drug discovery

Background:

• Predicting protein-ligand interactions is crucial for understanding molecular recognition in drug discovery and biological processes.
• Current methods often rely on limited structural or sequence data, hindering feature learning and neglecting surface geometric/chemical details.
• This limits interpretability and mechanistic insights into binding events.

Purpose of the Study:

• To develop a multimodal framework for accurate protein-ligand interaction prediction.
• To integrate 3D structural and surface geometric information for enhanced predictive power.
• To improve interpretability and mechanistic understanding of binding interactions.

Main Methods:

• Introduced PLiSAGE, a framework combining 3D structural and surface geometric embeddings.
• Utilized joint pretraining of encoders via unsupervised contrastive learning and point cloud reconstruction.
• Represented protein surfaces as segmented point cloud patches and employed a Transformer encoder for spatial dependencies.

Main Results:

• PLiSAGE achieved superior performance in binding affinity prediction and interaction classification compared to baselines.
• The model effectively captures fine-grained geometric and chemical cues from protein surfaces.
• Ablation studies confirmed the importance of surface features and the pretraining strategy for generalization.

Conclusions:

• PLiSAGE offers a robust and accurate method for predicting protein-ligand interactions by integrating structural and surface data.
• The framework enhances the expressive capacity of protein representations, improving predictive performance.
• This approach holds promise for advancing drug discovery and understanding molecular recognition.