Improving Generalizability in Whole-Cell Antibiotic Discovery Through Active Learning.

Area of Science:

• Drug discovery and development
• Computational chemistry
• Machine learning in biology

Background:

• Machine learning accelerates molecular discovery but struggles with out-of-distribution (OOD) chemical spaces.
• Antibiotic discovery via whole-cell phenotypic high throughput screening (HTS) is costly and resource-intensive.
• Active Learning (AL) offers an efficient ML-guided approach to navigate chemical spaces in drug discovery.

Purpose of the Study:

• To systematically evaluate and benchmark three AL strategies for whole-cell bacterial bioactivity prediction.
• To resolve algorithmic tradeoffs between novelty, bioactivity exploitation, and OOD generalizability in noisy biological systems.
• To identify an optimal AL strategy for efficient antibiotic discovery.

Main Methods:

• Retrospective simulations using *Mycobacterium tuberculosis* HTS data to evaluate AL strategies.
• Benchmarking AL strategies based on model accuracy, hit rate, and OOD performance.
• Integration of the optimal AL strategy into a closed-loop *Borrelia burgdorferi* antibiotic discovery HTS campaign.

Main Results:

• An optimal AL strategy was identified that balances novelty and hit rate.
• The AL-guided approach increased the experimental screening hit rate five-fold in *Borrelia burgdorferi* discovery.
• Prospective *in silico* selection demonstrated a 53-fold enrichment over investigator-directed screening, with 100% of validated hits showing narrow-spectrum activity.

Conclusions:

• Calibrated AL strategies can overcome data acquisition bottlenecks in drug discovery.
• AL enables the training of generalizable property predictors capable of extrapolating to OOD molecules.
• This approach significantly enhances the efficiency and success rate of antibiotic discovery campaigns.