Why do we need antibiotics with new structures?
Bacteria slowly become resistant to the drugs that used to kill them. The authors describe a crisis of antimicrobial resistance, which calls for antibiotics with a truly different structure from the ones we know.
Until now, AI mostly searched libraries of existing molecules for candidates. That limits novelty: you only find what has already been made.
A structurally distinct antibiotic is one built differently from the known ones, so bacteria have had less chance to evolve ways around it.
What did the researchers do?
They used two generative AI approaches, meaning programs that invent new molecules instead of only sorting existing ones. The first starts from chemical fragments: it screened a very large number of them on a computer, aiming at the gonorrhoea bacterium (Neisseria gonorrhoeae) or at Staphylococcus aureus, then expanded the most promising fragments.
The second invents molecules with no starting constraint. Both use genetic algorithms and variational autoencoders (methods that evolve and recombine proposals, a bit like natural selection).
In the fragment-based method, promising fragments were expanded into full molecules. In the unconstrained method, the AI generated whole compounds with no template. The best candidates were then synthesised, meaning actually made in a chemistry lab, and tested on bacteria.
What were the results?
Of 24 molecules actually made in the lab, seven showed selective antibacterial activity. Two lead molecules killed multidrug-resistant bacteria (resistant to several antibiotics), through different mechanisms.
In mice, they reduced the amount of bacteria in a model of vaginal gonorrhoea infection and in a model of skin infection with methicillin-resistant Staphylococcus aureus. The researchers also validated closely related molecules as antibacterial. They present their method as a platform for mapping uncharted regions of chemical space, meaning the set of possible molecules that no one had explored. But most of the molecules made did not show this selective activity: the AI's proposals still always have to be checked in the lab.
"Bactericidal" means the molecules killed the bacteria rather than just slowing them. The two lead compounds also worked through distinct mechanisms of action, meaning they did not kill bacteria in the same way.
Key figures
- 24molecules designed by AI and then made in the lab
- 7 of 24showed selective antibacterial activity
- 2lead molecules reduced infections in mice (gonorrhoea, resistant Staphylococcus aureus)
What does this change for patients?
For now, nothing for patients: these are lab molecules, still far from a medicine. But the study shows that an AI can propose truly new structures, which is exactly what antibiotic research needs.
If these leads hold up, they could one day add to the arsenal against hard-to-treat infections.
The authors add that the approach could be used to explore chemical space more widely than searching existing libraries allows.
What are the limits?
Tests were limited to bacteria in the lab and two mouse models. Nothing yet says these molecules will be safe and effective in humans.
Their toxicity and how the body handles them must be studied, and they must go through clinical trials, a long road on which many candidates fail.
Mouse infections also differ from human ones, and only two infection types were tested.
Tomorrow's antibiotics, and you today, with Mima
These molecules exist only in the lab, and none is a medicine yet. The study gives hope for the future fight against resistant bacteria.
If you take antibiotics, Mima can keep your prescriptions and note your symptoms day by day, and explain side effects in plain words. A question to ask: "Should I finish the whole box, even if I feel better?"
Frequently asked questions
Can an AI invent an antibiotic?
In this study, generative AIs designed new molecules, and of 24 made, seven showed selective antibacterial activity. Testing stopped at mice.
Are these antibiotics available?
No. They are lab molecules tested in mice. They would need clinical trials before any use.
What is generative AI?
It is a program that creates new things (here, molecules) instead of only classifying or sorting existing examples.
Which bacteria did they act against?
The two lead molecules reduced mouse infections caused by Neisseria gonorrhoeae and methicillin-resistant Staphylococcus aureus.
The original study
Krishnan A, Anahtar MN, Valeri JA, Coley CW, Wong F, Collins JJ, et al. “A generative deep learning approach to de novo antibiotic design.” Cell, 14 August 2025. DOI : 10.1016/j.cell.2025.07.033
Read the full study in CellThis article summarises a scientific study and is not medical advice.