- San Francisco lab Radical Numerics says its Omnii-1 model designed bacteriophages that were given to a patient with a drug-resistant infection, a world first
- The phages were built by rewriting the gene for T7’s tail fiber, the protein the virus uses to latch onto bacteria
- Nobody outside the company knows whether it worked, because the results have not been published, peer-reviewed, or disclosed yet
Radical Numerics has announced its AI-designed viruses (bacteriophages) were given to a sick person, but has not provided much further insight about what happened next.
The San Francisco AI lab <a href="https://www.rutlandherald.com/news/business/radical-numerics-launches-omnii-1-a-new-generation-of-biological-ai-as-first-ai-designed/article_c26dbdf6-2eb7-5797-9163-a65657929f9b.html” rel=”nofollow noopener” target=”_blank”>announced in October 2026 that bacteriophages designed by its new genome model had been administered to a human patient with an antibiotic-resistant bacterial infection, under a physician-led compassionate-use pathway authorized by the FDA.
If the claim holds, it is the first time an AI-designed organism of any kind has been put into a person as medicine.
Not a new phenomenon, but one that could do with AI assistance
Bacteriophages are viruses that infect bacteria and ignore human cells entirely. They have been used as therapy for more than a century, predating penicillin, and have enjoyed a quiet revival as antibiotics have begun to fail.
The World Health Organization considers phage therapy broadly safe, and unlike broad-spectrum antibiotics, a phage can target one bacterial strain while leaving the rest of your microbiome alone.
That precision is also the problem: each phage recognizes a narrow set of bacterial surface receptors, so treating a patient means finding one that matches their specific infection. Clinicians screen existing phage libraries, then environmental samples, then give up and try engineering as a last resort. For someone with sepsis, that search is not an academic exercise but an important life-or-death situation.
The scale of the underlying problem is not in dispute. The Global Burden of Disease project estimated that antimicrobial resistance was associated with roughly 4.7 million deaths in 2021, with projections indicating that between 2025 and 2050, close to 40 million deaths could be attributable to it.
Radical Numerics was founded by members of the team behind Evo and Evo 2, the Arc Institute genome models that generated the first entirely AI-written phage genomes. That earlier work produced 700,000 candidate genomes; 285 were synthesized, and 16 produced viable viruses. It was a proof of concept with a punishing, yet promising hit rate.
Omnii-1 is the follow-up, and the company’s technical write-up describes a narrower, more surgical approach. Rather than inventing whole genomes, the model edits the targeting component. The team worked with T7, a well-characterized 40-kilobase E. coli phage, and rewrote gp17, the receptor-binding tail fiber.
The numbers the company is willing to share are genuinely interesting. At 11 to 15 amino acid changes from wild type, 69.9% of designed variants were recovered as viable phages, against 4.8% for random controls. Of 6,694 variants with measurable activity, 2,161 beat the original, including some carrying 28 amino acid substitutions.
A promising, but unproven and unpublished play
The company tested its designed cocktail against bacteria isolated from a real patient’s infection, confirmed it eliminated that isolate in vitro, and then administered it. The patient reportedly had an E. coli infection that had resisted not only antibiotics but every available phage, which is effectively the worst position one can be in when it comes to infectious diseases.
Radical Numerics has since declined to describe the patient’s clinical course, outcome, hospital, or physicians, citing patient privacy and the fact that even limited clinical detail could identify the patient.
The company self-published the technical report on its website and has not had it peer-reviewed, but has separately indicated the treatment is proprietary and something it hopes to commercialize.
That silence held as of October 29, nine days after the announcement, when trade coverage for infectious disease clinicians still reported no disclosed clinical course, no outcome, and no peer-reviewed publication. The technical write-up remains the only account of the case.
In fairness to Radical Numerics, its own write-up states plainly that single-patient observations cannot establish safety or efficacy, which is more cautionary than most product launches manage. CEO Eric Nguyen has also said the company intends to publish through peer review.
Outside experts have landed somewhere between intrigued and unconvinced, which is the correct place to land. Harvard microbiologist Michael Baym, who was not involved, said the result would be incredible if it works, but added that his instinct is skepticism until the treatment finishes and the viruses’ novelty can actually be examined.
The commercial logic is much easier to establish; phage hunting is slow, and Radical Numerics believes it can compress design to days rather than months.
Ottawa Hospital infectious disease clinician Marisa Azad, who is collaborating with the company, has described the conventional search as remarkably time-consuming, and she is not wrong. If AI can reliably shorten that loop, it changes who gets treated in time, allowing for a more cost-effective and broader audience for said treatments.
There is cause for some concern too: a model that can design viruses to order is, structurally, a model that can design viruses to order. Radical Numerics acknowledges this directly by committing to biosafety, containment, and data-sharing measures scaled to the model’s capability.
It is the same dual-use problem that has dogged every capable biological design tool, and one that AI labs have been publicly wrestling with as their systems get better at biology.
Engineered phages have reached patients before, most famously a 2019 case in which modified phages treated a disseminated drug-resistant Mycobacterium abscessus infection. What is new here is not the therapeutic concept but the design process, and the governance questions attach to the AI-driven process, not the concept itself.
AI has moved from reading biology to writing it and handing the result to a doctor. That is a different category of claim than an AI model proposing a cancer hypothesis or outperforming specialists on rare-disease diagnosis, both of which keep a human firmly in the loop before anything enters a bloodstream.
For now, however, a very clever experiment was run on a very sick person, and we still don’t know how it went or when somebody decides to tell us. If it does pan out however, it will have massive rammifications for an industry that is only expected to grow across the next few decades.
