AI-designed viruses pass a first human test

A US startup administered an AI-designed bacteriophage cocktail to a patient against an antibiotic-resistant infection. The trial remains compassionate use but opens a concrete path against antibiotic resistance.

Facing antibiotic resistance, Radical Numerics has reached a rare milestone: the US startup says it administered bacteriophages designed with its AI model Omnii-1 to a patient on October 20, 2026. The goal is simple to state but difficult to achieve: find viruses capable of targeting antibiotic-resistant bacteria without harming human cells. This act does not yet prove generalized clinical efficacy, but it places AI at the heart of a new therapeutic approach.

Quand l’IA s’attaque à la précision chirurgicale

Bacteriophages are viruses that infect bacteria. They have obvious medical interest because they do not attack human cells, but their spectrum is narrow: a phage recognizes only a limited number of strains. This is precisely the lock Radical Numerics seeks to unlock with Omnii-1, a model trained on genetic and biological data to propose custom phages.

To illustrate this capability, the company worked on phage T7, which targets Escherichia coli. Omnii-1 modified certain sequences of its DNA, notably those related to bacterial adhesion. In other words, the model does not just search a catalog; it proposes variations that classical methods would have struggled to explore so quickly. Molecular biology loves details. Microbes, much less so.

Un usage compassionnel, pas encore une preuve à grande échelle

The patient received this treatment under compassionate use, a procedure allowing exceptional access to experimental therapy when standard options have failed or are insufficient. In lab tests, the phage cocktail had already eliminated bacteria taken from the patient, according to the source. The sequence is therefore encouraging but does not replace robust clinical validation.

Radical Numerics also reports solid experimental results on its variants: out of 6,694 variants whose activity was measurable, 2,161 outperformed the original phage. This type of result suggests that an AI model can accelerate the search for useful combinations in a vast genetic space, where human trial-and-error quickly hits its limits.

Ce que cette première chez l’humain change vraiment

This first human administration opens a path for personalized phage therapy, that is, the use of phages tailored to a precise infection. It could be significant in cases where antibiotics no longer work, a major public health issue as a study published in The Lancet mentions nearly 40 million deaths linked to antibiotic resistance between 2025 and 2050. At this stage, however, clinical trials, peer-reviewed publication, and safety data are needed before discussing scaling up.

If the method delivers on its promises, AI will not replace antibiotics. It could mainly help design targeted weapons faster against infections medicine already knows too well but struggles to catch up with.

Key points

  • Radical Numerics administered treatment to a patient on October 20, 2026.
  • Omnii-1 designs custom phages from biological data.
  • 6,694 variants tested, 2,161 better than the original phage.
  • The treatment was under compassionate use.
  • The Lancet mentions nearly 40 million deaths by 2050.

By the numbers

  • October 20, 2026 — date of administration to a patient, according to Radical Numerics.
  • 6,694 variants — phages experimentally evaluated, source Futura.
  • 2,161 variants — variants judged superior to the original phage.
  • Nearly 40 million — potential deaths linked to antibiotic resistance between 2025 and 2050, according to The Lancet.

To read