A machine learning model trained on a high-throughput antibacterial screen increases the hit rate of drug discovery
Problem
Screening small molecule libraries for novel antibacterial compounds has a very low success rate (0.87% hit rate against antibiotic-resistant Burkholderia cenocepacia).
Why it matters
Antibiotic-resistant bacteria urgently need new antibacterial compounds, and low hit rates make discovery slow and costly.
What the AI did
A machine learning model trained on 29,537 binarized screened compounds performed virtual screening of FDA-approved and natural product libraries, and top-ranked predictions were experimentally tested for growth inhibitory activity.
Results
ROC score 0.823 on test set; experimental hit rates of 26% (FDA-approved compounds) and 12% (natural products) versus 0.87% baseline.
Impact
ML virtual screening substantially raised experimentally validated hit rates over the baseline screen.
Record
| Venue | PLoS Computational Biology |
|---|---|
| Published | 2022 |
| Work type | article |
| Authors | A. S. M. Zisanur Rahman, Chengyou Liu, Hunter Sturm, Andrew M. Hogan, Rebecca L. Davis, Pingzhao Hu, Silvia T. Cardona |
| Citations | 47 (OpenAlex, at import time) |
| References | 59 |
| Model or method | Directed-message passing neural network (D-MPNN) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1371/journal.pcbi.1010613 |
| Open access | gold |
| Keywords | machine learning, virtual screening, antibacterial, D-MPNN, hit rate, Burkholderia cenocepacia |
Importance rating
significant — thousands to millions affected
New antibacterials against antibiotic-resistant bacteria; validated hit rates rose 14-fold and hits extended to ESKAPE pathogens, but abstract states no disease-burden scale.