DeepBindPPI: Protein-Protein Binding Site Prediction Using Attention Based Graph Convolutional Network.

Area of Science:

• Computational biology
• Immunoinformatics
• Drug discovery

Background:

• Protein-protein interactions are crucial for biological defense mechanisms.
• Current deep learning models for binding site prediction lack precision, leading to high false positives.
• Accurate prediction of binding sites is essential for laboratory experiments in drug discovery.

Purpose of the Study:

• To develop an enhanced strategy for predicting protein-protein interaction binding regions, specifically for antigen-antibody interactions.
• To improve the precision of computational methods for identifying interacting amino acids.
• To validate the predicted binding sites using a docking environment.

Main Methods:

• Employed a deep learning technique, DeepBindPPI, integrating graph convolutional networks with an attention mechanism.
• The model was trained on a general protein pool and fine-tuned using antigen-antibody data.
• A separate spatial network was incorporated to enhance precision.

Main Results:

• The integrated model predicts interacting amino acids with improved precision.
• The developed model demonstrates comparable performance to existing techniques.
• The incorporation of a spatial network increased prediction precision from 0.4 to 0.5.
• Utilizing predicted interface information in the HDOCK server yielded promising docking results with high-quality structures in top ranks.

Conclusions:

• DeepBindPPI offers a more precise method for predicting protein-protein interaction binding sites, particularly in antigen-antibody systems.
• The integration of graph convolutional networks and attention mechanisms, along with a spatial network, significantly enhances prediction accuracy.
• The validated binding site predictions show potential for improving drug discovery pipelines and structural biology research.