Examples of AI/ML for Antibiotic Discovery

[mattodd]

A place to gather good papers involving AI/ML for the discovery of antibacterials. Papers can be summarised in comments below, and digested in this upper panel.

One of the interesting things that struck me in preparing for this talk was the repeated value of the Enamine Real dataset in helping training ML models. i.e. the open dataset of available structures was in itself a useful dataset for the research.

[mattodd]

Deep-learning-based virtual screening of antibacterial compounds, G. Scalia et al. Nat. Biotech. 2025, (DOI: 10.1038/s41587-025-02814-6)

Genentech group.
Experimental high throughput screen of two million small molecules (proprietary library) at 5 µM against a sensitized Escherichia coli strain. Gave 5161 antibacterial compounds, with a hit rate of 0.26%.
A deep learning model was trained on these data.
This model was validated.
This then virtually screens 1.4 billion molecules that are synthetically accessible.
This gave 44,437 "virtual hits". Clustered and reduced to 503 compounds (including budget).. 345 compounds ultimately selected and ordered.
82 exhibit activity on the same strain.
This hit rate is 90-fold better than the HTS that the training data are based on (0.26% vs 23.8%).
Scores slightly correlated with potency.
The MoAs of two compounds was determined - known mechanisms.
No compounds exhibited MICs vs wild type E. coli.
Actives shown not to be generically lytic.
Compounds similar to the actives could be suggested by the model as part of hit expansion.
The model, when suitably trained, demonstrated some ability to predict accumulation behaviour.

[mattodd]

Generative AI for designing and validating easily synthesizable and structurally novel antibiotics, K. Swanson et al. Nature Machine Intelligence volume 6, pages 338–353 (2024) DOI: 10.1038/s42256-024-00809-7

Monte Carlo tree search assembles novel compounds using ~132,000 molecular building blocks via and 13 well-validated chemical reactions.
Experimental dataset were three owned/commercial libraries totalling 14,427 molecules.
470 actives
Applied pre-existing models to the data.
Then applied generative model to work within those constraints, including synthesisability.
Enamine again - 40 Bn.
58 made, six actives