Related: A Rogue DNA Repair Gene Could Be a Hidden Cancer Weakness
Artificial intelligence has helped researchers find two new antibiotic candidates against drug resistant gonorrhea, a sexually transmitted infection that public health officials have warned could one day become untreatable. A team at Harvard’s Wyss Institute used deep learning to sift through roughly 6 million molecules in search of compounds that could kill the bacterium behind the disease, and the two that emerged cleared the infection in both animal and laboratory tests without triggering the resistance that has steadily eroded existing treatments.
The problem the work targets is one of the more alarming corners of medicine. Neisseria gonorrhoeae, the bacterium that causes gonorrhea, has developed resistance to nearly every drug once used against it, leaving doctors with a shrinking set of options and the World Health Organization flagging it as a priority threat. Developing new antibiotics the traditional way is slow, expensive, and commercially unattractive, which is part of why so few reach the market, and that gap between rising resistance and stalled discovery is exactly where the researchers aimed their AI.
The approach paired machine learning with old fashioned screening. The team first trained deep learning models on activity data from 38,650 molecules, teaching the software to recognize the chemical features associated with killing the bacterium. They then turned that trained model loose on a virtual library of about 6 million compounds, a scale no human team could test by hand, and used it to narrow the field to 213 promising candidates. From that shortlist, two leads stood out, named MP20 and A1, both with structures distinct from existing antibiotics, which matters because a genuinely new chemical shape is less likely to be defeated by the resistance mechanisms bacteria have already evolved.
What separates this result from a promising computer prediction is that the compounds were then tested in living systems. The researchers evaluated the leads in mice and in a so called vagina on a chip, an organ on a chip device that recreates the conditions of a human vaginal infection on a small engineered platform. In both models the two molecules rapidly and selectively killed the bacteria, and crucially they did so without inducing resistance over the course of the testing, a sign that they may hold up better than drugs that bacteria quickly learn to evade.
The work is also a marker of how far AI driven drug discovery has come. For years the field promised that machine learning could compress the slow early stages of finding a drug, and results like this one show the approach producing concrete candidates against a real and urgent target rather than hypothetical molecules. By letting software explore millions of compounds and flag the few worth making and testing, researchers can spend their limited time and money on the most promising leads, a shift that is especially valuable for diseases like drug resistant infections where the economics have long discouraged investment.
The usual and important caveats apply. This is early research, validated in mice and an engineered lab model rather than in people, and the path from a promising compound to an approved medicine is long, costly, and frequently ends in failure. MP20 and A1 will need extensive further testing for safety and effectiveness before anyone could receive them, and many candidates that look strong at this stage do not survive clinical trials. The result is a lead, not a cure, and it should be read as a meaningful step rather than a finished treatment.
Even so, the study points to a hopeful pattern in the fight against drug resistant infections, one of the slow moving health crises of the era. Antimicrobial resistance has been outpacing the discovery of new drugs for years, and a method that can rapidly surface novel compounds against a hardened target offers a way to start closing that gap. If the approach holds up against gonorrhea and extends to other resistant bugs, AI may become one of the more important tools in keeping ahead of bacteria that have spent decades learning to defeat our best medicines.
Related on Entrelligence: Anthropic’s drug discovery program for neglected diseases, Midjourney’s full-body ultrasound scanner, and how the brain responds to fructose and glucose.