Benchmarking Cross-Docking Strategies in Kinase Drug Discovery.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning in drug discovery

Background:

• Machine learning (ML) is revolutionizing drug discovery, particularly small molecule design.
• Predicting bioactivity requires accurate protein-ligand complex structures, which is a current limitation.
• Structural information can enhance ML scoring but relies on reliable complex structure prediction.

Purpose of the Study:

• To develop practical methods for generating useful kinase-inhibitor complex geometries for ML scoring.
• To create a kinase-centric docking benchmark to evaluate docking and pose selection strategies.
• To assess the recapitulation of experimentally observed binding modes in a realistic cross-docking scenario.

Main Methods:

• Assembled a benchmark dataset of 589 protein kinase structures with 423 ATP-competitive ligands.
• Evaluated various docking and pose selection strategies, including physics-based docking, shape overlap, and maximum common substructure (MCS) matching.
• Utilized the KinoML framework and OpenEye Toolkits for automated complex generation.

Main Results:

• Docking methods biased by the cocrystallized ligand (shape overlap with/without MCS) outperformed standard physics-based docking.
• Docking into multiple protein structures significantly increased the likelihood of generating accurate (low RMSD) poses.
• A combined approach (Posit) using MCS to select similar ligands and structures achieved a 70.4% success rate in reproducing binding poses.

Conclusions:

• Ligand-biased docking strategies and multi-structure docking enhance protein-ligand complex prediction accuracy.
• The Posit approach offers an efficient method for generating reliable poses for ML applications.
• Findings, though focused on kinases, are potentially transferable to other protein families for improved drug discovery pipelines.