AI invented new antibiotics from scratch: 7 of 24 killed bacteria, 2 worked in mice
MIT researchers had AI design new antibiotic molecules. Two cleared drug-resistant infections in mice; none has been tried in people yet.
Bacteria that shrug off existing antibiotics are one of medicine’s hardest problems, and new antibiotics are rare because the underlying chemistry has barely changed in decades. Drug hunters normally search through libraries of molecules that already exist. A team led by James Collins at MIT did the opposite: they asked AI to invent molecules nobody had made before.
The work was published in Cell in August 2025. Across both of their methods, the models proposed more than 36 million candidate molecules. The catch is that a molecule drawn by a computer still has to be built by chemists, and most could not be. Of the designs sent to the lab, 24 could actually be made, and seven of those killed bacteria without harming human cells in the dish.
Two of the seven, named NG1 and DN1, went further. Both killed bacterial samples that resist multiple drugs, and both reduced the infection in mice: one in drug-resistant gonorrhoea, the other in an MRSA skin infection. The non-profit Phare Bio is now trying to improve them. Funders include the US Defense Threat Reduction Agency and the National Institutes of Health.
What it means
This is early-stage science, not a new medicine. It does show that AI can propose antibiotics that work in a dish and in an animal, which is the step drug hunters have struggled to reach by searching existing molecules. For patients with resistant infections, nothing changes yet: the two candidates still need years of safety and human testing, and most drugs at this stage never arrive.
Caveats
- This is lab dishes and mice. Neither molecule has been given to a person, and most antibiotics that work in mice never reach patients.
- Making the molecules, not designing them, is the bottleneck: the great majority of AI proposals could not be built by the chemists who tried. In one of the two tracks, 80 were picked and only two could be made.
- The counts of molecules generated come from MIT’s news release; the paper’s abstract describes the screening scale differently, as more than 10 million fragments.
- Mouse tests cover one infection in one body site. They say little about dosing, safety or resistance developing in people.
- For the wider pattern, a 2024 analysis of AI-discovered drugs found about 80 to 90% passed Phase 1 trials but only about 40% passed Phase 2, close to the industry’s historical average, from a small sample.