Multimodal graph, surface, and language-based model for protein protein interaction prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein-protein interactions (PPIs) are vital for cellular functions and disease.
• Experimental PPI determination is costly and time-consuming.
• Current deep learning models for PPI prediction often neglect crucial protein surface information.

Purpose of the Study:

• To develop a novel multimodal deep learning framework for accurate PPI prediction.
• To integrate protein surface features, 3D structural information, and sequence embeddings.
• To improve upon existing PPI prediction methods by addressing limitations in feature representation and fusion.

Main Methods:

• Developed GSMFormer-PPI, a multimodal framework integrating geometric deep learning (MaSIF) for surface descriptors, graph convolutional networks for structural context, and a transformer encoder for cross-modal interactions.
• Utilized physicochemical surface descriptors, 3D structural graphs, and residue-level sequence embeddings.
• Employed advanced feature fusion techniques beyond simple concatenation.

Main Results:

• GSMFormer-PPI demonstrated superior performance compared to traditional graph-based models on the PINDER dataset.
• Cross-dataset evaluation showed performance comparable or superior to other leading PPI prediction models.
• Ablation studies confirmed the significant contribution of protein surface features and the advanced fusion strategy.

Conclusions:

• Integrative analysis of surface, structure, and sequence data is a highly effective strategy for advancing PPI prediction.
• GSMFormer-PPI represents a significant advancement in computational approaches for understanding biological processes and disease mechanisms.
• The proposed multimodal framework offers a promising direction for future research in protein interaction prediction.