Ligand Binding Prediction Using Protein Structure Graphs and Residual Graph Attention Networks

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Computational prediction of ligand-target interactions accelerates drug discovery by reducing experimental screening costs.
• Deep learning (DL) models offer enhanced predictive power with accumulating bioactivity data.
• Existing DL models often use simplified protein representations or inefficient 3D structure voxelization.

Purpose of the Study:

• To develop an advanced deep learning approach for accurate prediction of ligand-target binding affinity.
• To integrate 3D protein structures and 2D ligand graphs for improved interaction prediction.
• To identify critical protein regions involved in ligand binding using attention mechanisms.

Main Methods:

• Developed the Protein Structure Graph-Binding Affinity Regression (PSG-BAR) model.
• Utilized 3D protein structural information and 2D ligand graph representations.
• Incorporated attention scores to weigh important protein regions for ligand binding.

Main Results:

• Achieved state-of-the-art performance on binding affinity benchmarking datasets.
• Demonstrated that attention-based pooling identifies critical surface residues for ligand binding.
• Validated model predictions against experimental assays for SARS-CoV-2 Mpro.

Conclusions:

• The PSG-BAR approach offers a powerful new method for predicting ligand-target interactions.
• Attention mechanisms enhance model interpretability by highlighting key binding site residues.
• This method has significant implications for accelerating the discovery of novel therapeutics, including antivirals.