Open-ComBind: harnessing unlabeled data for improved binding pose prediction.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Structural Biology

Background:

• Accurate prediction of ligand binding poses is crucial for in silico drug discovery.
• Traditional molecular docking often analyzes single ligands, neglecting shared binding interactions among multiple ligands for the same receptor.
• Existing methods lack efficient ways to leverage information from multiple ligands to refine pose selection.

Purpose of the Study:

• To introduce Open-ComBind, an open-source molecular docking pipeline.
• To enhance ligand pose selection by integrating information from multiple ligands.
• To improve the accuracy of predicting non-covalent protein-ligand binding interactions.

Main Methods:

• Developed Open-ComBind, an accessible version of the ComBind pipeline.
• Created distributions of feature similarities between ligand pose pairs.
• Compared near-native poses with all sampled docked poses to capture feature likelihoods.
• Combined similarity distributions with per-ligand docking scores for pose selection.

Main Results:

• Enhanced overall pose selection by 5% for high-affinity ligands and 4.5% for congeneric series.
• Reduced the average Root Mean Square Deviation (RMSD) of ligands by 9.0% in the benchmark dataset.
• Demonstrated improved pose prediction performance through leveraging multi-ligand data.

Conclusions:

• Open-ComBind effectively improves ligand pose prediction accuracy in molecular docking.
• The open-source tool provides a valuable enhancement for in silico drug discovery pipelines.
• Leveraging multi-ligand information offers a significant advantage in predicting protein-ligand binding poses.