Accelerating high-throughput virtual screening through molecular pool-based active learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Structure-based virtual screening (SBVS) is crucial for early-stage drug discovery, evaluating protein-ligand interactions.
• The exponential growth of virtual libraries (over 10^8 molecules) escalates computational demands for exhaustive screening.
• Bayesian optimization offers a solution by training surrogate models to predict ligand affinities, enabling efficient library exploration.

Purpose of the Study:

• To investigate the application of Bayesian optimization techniques to computational docking datasets.
• To assess the impact of surrogate model architecture, acquisition function, and batch size on optimization performance.
• To demonstrate significant reductions in computational costs for large-scale virtual screening.

Main Methods:

• Employed Bayesian optimization with surrogate structure-property relationship models.
• Utilized directed-message passing neural networks as a surrogate model architecture.
• Evaluated upper confidence bound and greedy acquisition strategies with varying batch sizes.
• Applied methods to a 100 million-member virtual library dataset.

Main Results:

• Achieved substantial reductions in computational screening costs.
• Identified 94.8% of top ligands using upper confidence bound strategy after testing only 2.4% of candidates.
• Identified 89.3% of top ligands using greedy acquisition strategy after testing only 2.4% of candidates.
• Demonstrated the efficacy of model-guided searches in large-scale virtual screening.

Conclusions:

• Bayesian optimization effectively mitigates the rising computational costs associated with screening massive virtual libraries.
• This model-guided approach accelerates high-throughput virtual screening campaigns.
• The methodology has broad applications beyond computational docking in drug discovery and related fields.