PocketStruct: Integrating Protein Pocket Structural Features for Protein-Peptide Binding Prediction.

Area of Science:

• Computational biology
• Drug discovery
• Structural bioinformatics

Background:

• Peptides offer low toxicity and small binding interfaces, making them promising drug candidates.
• Predicting peptide-protein binding is crucial for drug development.
• Current methods often rely solely on sequence information, potentially missing key structural insights.

Purpose of the Study:

• To evaluate the impact of integrating protein pocket structural information into peptide-protein binding prediction models.
• To compare the performance of different structural encoders within a unified framework.
• To assess the robustness of structure-aware models on novel data.

Main Methods:

• Developed a transformer-based framework with mutual attention to integrate protein pocket structural data.
• Systematically compared three pocket-structure encoders: SE(3)-Transformer, ProtGVP (geometric graph neural network), and ESM-IF1 (large-scale structure-based pretrained model).
• Employed rigorous data partitioning with strict separation of training and test sets.

Main Results:

• Incorporating pocket structural information consistently improved binding prediction accuracy over sequence-only models.
• ProtGVP demonstrated particularly effective pocket representations.
• Structure-based model variants showed superior robustness when predicting binding on previously unseen data.

Conclusions:

• Protein pocket structure is a vital feature for enhancing peptide-protein binding prediction accuracy.
• Geometric deep learning models like ProtGVP show significant promise for representing structural information.
• Structure-aware models offer improved generalization and robustness for drug discovery applications.