Combining Docking Pose Rank and Structure with Deep Learning Improves Protein-Ligand Binding Mode Prediction over a...

Area of Science:

• Computational chemistry and structural biology
• Machine learning in drug discovery
• Deep learning for molecular interactions

Background:

• Predicting protein-ligand activity and binding modes is crucial for drug discovery.
• Deep learning combined with docking shows promise but can be affected by dataset bias.
• Distinguishing true interaction effects from dataset biases is a significant challenge.

Purpose of the Study:

• To present a modular, graph-based convolutional neural network for protein-ligand interaction prediction.
• To develop a dual-graph architecture capable of separating ligand identity biases from interaction effects.
• To create a deep learning model for binding mode prediction that leverages docking structures and rankings.

Main Methods:

• Utilized a dual-graph neural network architecture processing ligand topology and protein-ligand contact maps.
• Generated protein-ligand complex structures using standard docking procedures.
• Developed a binding mode prediction model incorporating docking ranking and structural data.

Main Results:

• The dual-graph network successfully learned from protein structural information when using unbiased datasets.
• The model demonstrated the ability to differentiate dataset biases from genuine protein-ligand interaction effects.
• The binding mode prediction model outperformed AutoDock Vina on various tests, including cross-docking scenarios.

Conclusions:

• The proposed graph-based neural network effectively models protein-ligand interactions, addressing dataset bias issues.
• The network predictions provide reliable measures of model confidence, enhancing usability in drug discovery.
• This approach offers a robust strategy for improving the accuracy and reliability of molecular docking and activity prediction.