Leveraging Uncertainty in Machine Learning Accelerates Biological Discovery and Design
Problem
Machine learning models fail when predicting beyond their training distribution, hindering biological discovery such as finding active compounds from very small datasets.
Why it matters
Faster AI-guided identification of drug candidates could cut the cost and time of discovering medicines, including for tuberculosis.
What the AI did
Gaussian process-based uncertainty prediction on pre-trained features guided screening of a 10,833-compound library after training on only 72 compounds; predicted hits were experimentally validated.
Results
Trained on 72 compounds; predictions over 10,833-compound library; validated compounds with nanomolar affinity for diverse kinases and whole-cell growth inhibition of Mycobacterium tuberculosis
Impact
Experimentally validated AI-selected bioactive compounds, supporting uncertainty-guided discovery across biological domains.
Record
| Venue | Cell Systems |
|---|---|
| Published | 2020 |
| Work type | article |
| Authors | Brian Hie, Bryan D. Bryson, Bonnie Berger |
| Citations | 153 (OpenAlex, at import time) |
| References | 148 |
| Model or method | Gaussian process-based uncertainty prediction on pre-trained features |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1016/j.cels.2020.09.007 |
| Open access | hybrid |
Importance rating
significant — thousands to millions affected
ML failure outside training distributions broadly hinders biological discovery; abstract demonstrates experimentally validated hits including Mycobacterium tuberculosis growth inhibition from tiny training data, a new capability across domains.