Halicin: an AI found a new kind of antibiotic by exploring millions of molecules

Why do we need new antibiotics?

Bacteria end up resisting the antibiotics that have been used for a long time: this is called resistance. Infections caused by such bacteria become harder to treat, and the authors stress a growing need for new antibiotics.

The trouble is that finding a new antibiotic is slow and expensive. Testing molecules one by one in the laboratory only lets researchers explore a tiny part of all the possible molecules.

Meanwhile, few new families of antibiotics are arriving. Any tool that helps explore the possible molecules faster and more widely is therefore welcomed by researchers.

How was the AI trained?

The team trained a deep neural network, a type of AI that learns to recognise patterns, to flag molecules that stop bacteria from growing. Once trained, the AI could examine very large libraries of molecules without each one having to be tested in a tube.

It was first used on the Drug Repurposing Hub, a collection of molecules already known for other uses. It then went through a bank of more than 107 million molecules called ZINC15.

This approach changes the usual logic: instead of testing first and understanding later, the AI suggests a short list of candidates, which researchers then check in the laboratory.

What did it find?

In the first collection, the AI pointed to a molecule named halicin. Its structure is very different from usual antibiotics. In the laboratory it kills (it is called bactericidal) a wide variety of bacteria, including Mycobacterium tuberculosis, which causes tuberculosis, and carbapenem-resistant Enterobacteriaceae, bacteria that resist last-resort antibiotics.

In mice, halicin effectively treated infections with Clostridioides difficile and Acinetobacter baumannii, a pan-resistant bacterium, meaning one that resists all standard antibiotics.

In the large ZINC15 bank, the team tested in the laboratory 23 molecules picked by the AI out of more than 107 million. Eight turned out to be antibacterial, with structures far from known antibiotics.

These findings matter because tuberculosis and carbapenem-resistant bacteria are among the infections that are hardest to treat.

What does this change for patients?

For now, nothing concrete: halicin is not a medicine and has not been given to people. The value is in showing that an AI can help explore families of molecules that researchers might not have thought of testing.

In time, this approach could help widen the range of antibiotics, notably against bacteria that resist current treatments.

It also shows how computers can widen the search, so that scientists spend their lab time on the most promising molecules.

What are the limits of this study?

The results come from the laboratory and from mouse models. They say nothing about effectiveness or safety in humans. Toxicity studies and then clinical trials will be needed before a medicine can be considered, and many candidates fail at those stages.

Finally, of the 23 molecules tested from the large bank, only eight were antibacterial: the AI is not infallible.

It is also worth remembering that a mouse is not a human being: a product that is active in an animal can prove ineffective or poorly tolerated in people.