Improving the Hit Rates of Virtual Screening by Active Learning from Bioactivity Feedback

Area of Science:

• Computational Chemistry
• Drug Discovery
• Bioinformatics

Background:

• Virtual screening methods in drug discovery often yield low hit rates due to simplified scoring functions.
• The high cost of laboratory validation limits the exploration of potential drug candidates.
• Current virtual screening neglects valuable bioactivity feedback from wet-lab experiments.

Purpose of the Study:

• To introduce an Active Learning from Bioactivity Feedback (ALBF) framework to enhance virtual screening hit rates.
• To improve the accuracy and cost-effectiveness of identifying potential drug candidates.
• To leverage target-specific bioactivity insights for refining screening results.

Main Methods:

• Developed a novel query strategy considering evaluation quality and influence on other molecules.
• Implemented an efficient score optimization strategy to propagate bioactivity feedback to similar molecules.
• Evaluated the ALBF framework on DUD-E and LIT-PCBA benchmarks using iterative wet-lab experiment budgets.

Main Results:

• The ALBF protocol achieved an average enhancement of top-100 hit rates by 60% on DUD-E and 30% on LIT-PCBA.
• These improvements were observed with 50 to 200 bioactivity queries across ten experimental rounds.
• The framework demonstrated consistent performance across diverse benchmark subsets.

Conclusions:

• The Active Learning from Bioactivity Feedback (ALBF) framework significantly improves virtual screening accuracy.
• ALBF enhances the cost-effectiveness of laboratory testing in drug discovery pipelines.
• This approach holds substantial potential for optimizing the identification of bioactive molecules.