GraphPBSP: Protein binding site prediction based on Graph Attention Network and pre-trained model ProstT5

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein-protein/peptide interactions are vital for biological processes.
• Accurate prediction of binding sites is crucial but challenging.
• Existing methods require improvement for enhanced prediction accuracy.

Purpose of the Study:

• To develop an effective computational model for predicting protein-protein/peptide binding sites.
• To introduce novel features and methodologies for improved binding site prediction.
• To provide a valuable tool for protein engineering and drug design.

Main Methods:

• Developed GraphPBSP, a model utilizing Graph Attention Network, Convolutional Neural Network, and Multilayer Perceptron.
• Integrated diverse features: interface residue pairwise propensity, ProstT5 sequence embeddings, physicochemical properties, and structural features.
• Implemented a spatial neighbor-based feature statistic method and a multi-scale objective function for enhanced learning.

Main Results:

• GraphPBSP significantly outperforms existing state-of-the-art methods on protein-protein/peptide binding site prediction.
• The model demonstrates strong generalization ability on protein-DNA/RNA binding site prediction tasks.
• Novel features like ProstT5 embeddings and residue pairwise propensity contribute to improved performance.

Conclusions:

• GraphPBSP is a highly effective and promising method for binding site prediction.
• The model's performance highlights the utility of integrated sequence and structural features.
• GraphPBSP offers valuable insights for advancing protein engineering and drug discovery efforts.