Query Matters: How Selection Strategies Influence Active Learning in Drug Discovery

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning in drug discovery

Background:

• The Design-Make-Test-Analyze (DMTA) cycle is crucial for preclinical drug discovery but is limited by time and cost.
• Simulating the DMTA cycle allows for efficient exploration of factors influencing its effectiveness.

Purpose of the Study:

• To present SimDMTA, an in silico framework for simulating the DMTA cycle.
• To evaluate different sampling strategies for hit discovery and model generalization within the DMTA cycle.

Main Methods:

• Developed an in silico framework (SimDMTA) simulating the DMTA cycle.
• Utilized a machine learning model to predict docking scores, acting as a proxy for biological assays.
• Employed various query strategies, including uncertainty-based sampling, for compound selection and iterative model retraining.
• Focused on molecules derived from a 3,5-dimethyl-4-phenylisoxazole scaffold targeting the Bromodomain 4 (BRD4) BD1 binding site.

Main Results:

• Uncertainty-based sampling significantly outperformed greedy and hybrid approaches in hit discovery.
• Uncertainty-based sampling enhanced the generalization ability of the predictive model.
• By the final iteration, 37 of the top 50 ranked compounds were in the top 1% of the evaluated chemical space.
• Strategies incorporating random selection improved bias correction but were less effective for identifying top molecules.

Conclusions:

• Incorporating molecular diversity and uncertainty into active learning design strategies accelerates model refinement and improves hit identification.
• Uncertainty-based sampling is a superior strategy for efficient preclinical drug discovery simulations.
• SimDMTA provides a feasible approach to explore DMTA cycle efficiencies beyond traditional experimental limitations.