Drug-resistant infections are one of the slowest-moving emergencies in medicine. The World Health Organization’s fact sheet on antimicrobial resistance says bacterial resistance was “estimated to be associated with more than 4.7 million deaths globally in 2021”. New antibiotics are rare, and finding them by trial and error is slow.
In 2023, researchers at McMaster University in Canada and the Massachusetts Institute of Technology used a machine-learning model to speed up that search. It helped them find a compound they named abaucin, which killed a dangerous hospital bacterium in the lab and treated infected wounds in mice. Our 44-second explainer above shows how. This article sets out what the study found, what a later study added and what is still unknown.
In brief
- What happened: an AI model ranked 6,680 compounds, scientists tested 240, and one, abaucin, treated wound infections in mice.
- Why it matters: abaucin targets Acinetobacter baumannii, a hospital bacterium that often resists almost every existing antibiotic.
- What’s unknown: whether abaucin is safe and effective in people. It is not an approved medicine, and no human trial has been reported.
The model ranked 6,680 compounds in under two hours
The team first needed training data. According to MIT News, they grew Acinetobacter baumannii in lab dishes and exposed it to about 7,500 chemical compounds to see which ones stopped its growth. The paper in Nature Chemical Biology gives the exact number as 7,684. They fed each molecule’s structure, and whether it worked, into a neural network so it could learn which chemical features go with killing the bacterium.
The trained model then scored 6,680 compounds it had not seen before, from the Broad Institute’s Drug Repurposing Hub, a library of known drugs and drug candidates. MIT News says that analysis took less than two hours and produced a few hundred top hits. The researchers chose 240 to test, favouring structures unlike existing antibiotics. In the lab, nine of them strongly inhibited the bacterium’s growth.
Abaucin targets one hospital bacterium
The team narrowed those nine to the most promising. The one they picked was an existing compound, a blocker of a receptor called CCR2, which MIT News says was originally explored as a potential diabetes drug. They renamed it abaucin, after A. baumannii.
The bacterium matters. McMaster’s release notes that the WHO counts it among the world’s most dangerous antibiotic-resistant bacteria. It can cause pneumonia, meningitis and wound infections, and it survives on hospital surfaces for long periods.
Abaucin is narrow-spectrum. MIT News reports it had no effect on other species tested, including Pseudomonas aeruginosa, Staphylococcus aureus and carbapenem-resistant Enterobacteriaceae. The researchers see that as a strength: broad-spectrum antibiotics also kill helpful gut bacteria, which can open the way to infections such as C. difficile. McMaster says the narrow targeting means the bacterium is less likely to develop resistance quickly, which is the researchers’ expectation rather than a tested result.
In mice, a cream containing abaucin controlled A. baumannii infections in wounds. In lab tests it also worked against drug-resistant strains taken from patients.
A 2026 study tested abaucin against lung infections in mice
A separate team at West China Hospital, Sichuan University, has since taken abaucin further. In a paper in Nature Communications published on 11 July 2026, they reported potent activity against carbapenem-resistant strains of A. baumannii, known as CRAB, taken from patients, both in the lab and in a mouse pneumonia model.
The mice had their immune defences weakened before infection, and were given four doses over 36 hours. The authors report that lung damage was markedly reduced, to a degree comparable to meropenem, a last-line antibiotic, and that bacterial counts fell. They also used cryo-electron microscopy to show how two abaucin molecules lock a lipoprotein transport protein called LolDF, which they say explains why it hits only A. baumannii. These are still animal results from a short experiment.
The AI narrowed the search, and people did the lab work
“AI found a new antibiotic” is a tempting headline, and it overstates what happened. The model did one job: it ranked thousands of molecules by how likely they were to stop A. baumannii growing. It did not design abaucin, which already existed as a research compound. Scientists built the training data by testing about 7,500 compounds by hand, chose which predictions to test, ran the experiments, worked out how the drug acts and treated the mice.
What the AI changed was speed and focus. Instead of testing all 6,680 library compounds, the team tested 240 and found nine that worked. The paper reports that when the team also tested the 240 molecules the model ranked lowest, none worked. That is evidence the ranking was useful, within this one library and against this one bacterium.
The “nine antibiotics” also need care. They were nine molecules that strongly inhibited the bacterium in a dish, not nine medicines. Most candidates that look good at this stage never reach patients. MIT News says Stokes’s lab at McMaster is working to improve the compound’s medicinal properties “in hopes of developing it for eventual use in patients”. No human trial has been reported.
The method is spreading. TSN has covered an MIT team using AI to design a compound against drug-resistant gonorrhoea, Google DeepMind’s agent finding an antimicrobial peptide candidate that still needs lab testing, and Radical Numerics’ claim that AI-designed phages were given to a patient. Each is at a different stage, and none is yet a proven treatment.
Britain’s AMR plan targets the business problem
Discovery is only half the problem. The UK’s national action plan, Confronting antimicrobial resistance 2024 to 2029, says investment in new antimicrobials “is widely seen as commercially unattractive by industry”, because new antibiotics are held in reserve and used very little in their first years. Its answer is subscription-style payments that “de-link the price paid for antimicrobials from the volumes sold”, and it aims for resistance to be controlled and contained by 2040.
Faster discovery tools could feed more candidates into that pipeline. Turning one into a medicine still takes safety testing, human trials and someone willing to pay for a drug designed to be used sparingly.
Sources
- Liu, G. et al., “Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii”, Nature Chemical Biology 19, 1342-1350, 25 May 2023
- MIT News, “Using AI, scientists find a drug that could combat drug-resistant infections”, 25 May 2023
- McMaster University, “Scientists use AI to find promising new antibiotic to fight evasive hospital superbug”, 25 May 2023
- World Health Organization, “Antimicrobial resistance” fact sheet, 16 July 2026
- Pang, J. et al., “Conformational hijacking of lipoprotein transporter LolDF enables precision antimicrobial activity against Acinetobacter baumannii”, Nature Communications 17, 8554, 11 July 2026
- GOV.UK, “Confronting antimicrobial resistance 2024 to 2029”, 8 May 2024
Related stories
- AI Just Designed a Drug That Kills Superbugs Without Harming Good Bacteria
- Google DeepMind Shows Its Antigravity Agent Finding a New Antimicrobial Peptide Candidate, Which Still Needs Lab Testing
- Radical Numerics Claims First AI-Designed Phage Given to a Patient
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
