A generative AI model trained on protein structures proposed millions of novel molecules — many showing potent activity against drug-resistant bacteria that current antibiotics cannot touch.
Antibiotic resistance is one of the most serious threats to global health. The World Health Organisation estimates that by 2050, drug-resistant infections could kill more people annually than cancer. Finding new antibiotics using traditional methods is painfully slow — it typically takes over a decade and costs more than a billion dollars to bring a single new antibiotic to market. A new study from MIT and the Broad Institute suggests AI could compress that timeline dramatically.
The model uses a diffusion-based architecture — similar in principle to the image generation models that produce AI art, but operating in the space of molecular geometry rather than pixel values. It learns the statistical patterns of how molecular shapes interact with bacterial proteins, then generates new molecules optimised to bind to those proteins in ways that disrupt bacterial function. Crucially, it is not constrained to the chemical families of existing antibiotics, allowing it to explore entirely novel structural territory.
Not everyone in the drug discovery community is ready to declare a revolution. Critics note that in vitro activity — the kind measured in this study — frequently fails to translate into clinical efficacy. A molecule that kills bacteria in a dish may be toxic to human cells, unable to reach the infection site in the body, or rapidly broken down before it can act. The real test will be the mouse model results, expected later in 2026. Still, even sceptical voices acknowledge that AI has meaningfully increased the diversity and novelty of candidate molecules worth investigating.
"We are not saying AI has solved antibiotic resistance. We are saying it has given us more doors to open than we have ever had before."
— Senior author, Broad Institute, 2026