Assessing the Robustness and Scalability of Machine Learning Methods to Accelerate Ultralarge High-Throughput Docking...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Structure-based virtual screening is crucial for computational drug discovery.
• Machine learning (ML) protocols accelerate high-throughput screening of large chemical libraries.
• Previous ML validation studies used limited targets and small molecule libraries.

Purpose of the Study:

• To extend the validation of ML protocols for accelerated virtual screening.
• To assess ML performance on large-scale (∼100M) and diverse (10 protein targets) chemical libraries.
• To demonstrate the efficiency of ML in retrieving top-scoring molecules from ultralarge datasets.

Main Methods:

• Utilized two standard publicly available ∼100M molecule libraries.
• Employed a comprehensive benchmark set of 10M molecules across 10 protein targets.
• Used PLANTS and AutoDock Vina docking programs for molecular docking scores.
• Validated ML protocols for retrieving virtual hits and assessed performance metrics.

Main Results:

• ML protocols retrieved >60% and >70% of top 10k and 1k molecules, respectively, from a 10M molecule set.
• Achieved >97% reduction in docking evaluations on average.
• Demonstrated that increased training set size proportionally enhances ML performance on larger libraries.
• Confirmed robust performance of ML protocols across diverse targets and large datasets.

Conclusions:

• ML methods are highly effective for retrieving top-scoring molecules from chemical libraries of hundreds of millions to billions of compounds.
• ML significantly reduces the computational cost and time required for virtual screening.
• The role of ML models is critical for exploring vast chemical spaces where brute-force docking is infeasible.