Pose ensemble graph neural networks to improve docking performances.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Predicting small molecule-protein interactions is crucial but challenging due to complexity.
• Existing machine learning (ML) methods often map 3D pose features to experimental structures or binding affinities.
• Physics-based tools like molecular docking have limitations in accuracy and scope.

Purpose of the Study:

• Introduce Dockbox2 (DBX2), a novel ML approach for modeling small molecule-protein interactions.
• Enhance the prediction of binding pose likelihood and binding affinity.
• Improve drug discovery pipelines through accurate interaction modeling.

Main Methods:

• Developed Dockbox2 (DBX2), a graph neural network (GNN) framework.
• Encoded ensembles of computational poses using energy-based features from molecular docking.
• Jointly trained the GNN model for node-level (pose likelihood) and graph-level (binding affinity) prediction tasks.
• Utilized the PDBbind dataset for training and validation.

Main Results:

• DBX2 demonstrated significant performance in retrospective docking and virtual screening experiments.
• Achieved superior results compared to state-of-the-art physics-based and ML-based tools.
• Validated the effectiveness of learning from conformational ensembles for interaction prediction.

Conclusions:

• DBX2 offers a powerful new approach for modeling small molecule-protein interactions.
• Encourages further research into ML models that leverage conformational ensembles.
• Provides a valuable tool for advancing drug discovery and understanding molecular thermodynamics.