Radical Numerics, the startup built by the team behind Stanford's Evo genome models, announced on October 20, 2026 that bacteriophages designed by its new Omnii-1 AI model have been administered to a human patient — the world's first clinical use of AI-generated viruses. The New York Times first reported the claim, and the company published a detailed technical report on its website the same day. For readers following the latest AI developments, it marks the moment AI-designed biology moved from the lab bench into a patient's bloodstream.
The phages were given under compassionate use, the regulatory pathway that allows patients with serious conditions to receive investigational therapies outside clinical trials when no satisfactory alternatives exist. The company says the treatment targeted an antibiotic-resistant bacterial infection, though it has declined to disclose the patient's condition, location, or outcome, citing confidentiality and privacy requirements.
From Genome Language Models to Living Viruses
Omnii-1 is a multimodal genome language model — an AI system trained on vast quantities of DNA sequences and multiomics data, then fine-tuned specifically on phage genomes. Phages, formally bacteriophages, are viruses that infect and kill bacteria without touching human cells, which has made them a decades-old candidate for treating infections that antibiotics can no longer clear.
The Radical Numerics team, whose members created the Evo and Evo 2 genome models before founding the company, used Omnii-1 to do three things: generate viable phage variants, prioritize the most promising designs, and predict which phages will successfully infect which bacterial strains. Rather than designing genomes from scratch, the model engineers specific regions — such as the receptor-binding tail fiber genes that determine which bacteria a phage can recognize — while preserving the machinery the virus needs to replicate.
What the Technical Report Shows
The company's report includes concrete benchmarks from its design campaigns. In one experiment on the well-characterized E. coli phage T7, 69.9 percent of Omnii-1-designed variants containing 11 to 15 amino acid changes were successfully recovered as viable phages, compared with just 4.8 percent of random controls. Of 6,694 variants with measurable activity, 2,161 exceeded the activity of the wild-type phage — including designs with up to 28 amino acid changes, while no random control exceeding wild type had more than five.
The model also proved effective at strain-specific prioritization. For two bacterial strains, 97 and 96 of Omnii-1's top 100 ranked candidates cleared the activity threshold. On a harder third strain where only 4.5 percent of the full library was active, 24 percent of the model's top 100 picks cleared the bar — more than five times the base rate. On a held-out clade of 37 E. coli strains, predictions from just two phages covered 91.4 percent of eligible strains, and the model outperformed specialized phage-host prediction workflows like GenoPHI when transfer learning from E. coli data to Pseudomonas aeruginosa, a dangerous pathogen behind wound, burn, and cystic fibrosis infections.
Why Phage Therapy Matters Now
The clinical milestone lands amid a worsening antibiotic resistance crisis. According to the figures cited in Radical Numerics' report, antibiotic resistance was associated with an estimated 4.7 million deaths worldwide in 2021, and deaths attributable to resistance are projected to total nearly 40 million between 2025 and 2050.
Phage therapy offers a way around that crisis because phages can target specific bacterial strains with a specificity broad-spectrum antibiotics cannot match, sparing beneficial gut bacteria in the process. The bottleneck has always been matching: today's personalized phage therapy depends on laboriously screening natural phage collections against each patient's infection, and delays can be life-threatening for patients who are out of options. An AI system that can design and rank phages on demand attacks that bottleneck directly — a development with obvious relevance to AI research coverage well beyond medicine.
Big Claims, Few Details
There are important caveats. The headline result is a company claim: the New York Times framed it as a start-up "claiming" to treat a patient, and the company's technical report has not undergone peer review. Radical Numerics declined to provide information about the patient or the doctors involved, and it cannot disclose clinical outcomes, so no one outside the company can yet verify whether the therapy actually helped. Earlier high-profile phage cases — treatments for individual patients under emergency programs — have a mixed history of replication.
Still, the milestone is real in one verifiable sense: AI-generated phages passed the regulatory and safety bar required for administration to a human being under a government-supervised compassionate use program. The Times also reports that Radical Numerics is applying Omnii-1 to other fronts, including designing cancer vaccines and assessing whether viruses could cause pandemics — a reminder that the same design capability cuts in both directions, a tension AI and biosecurity coverage will keep tracking.
Stay Ahead of AI
The boundary between AI research and real-world consequence keeps moving. Whether the next breakthrough arrives in a lab, a corporate boardroom, or a regulatory hearing, AI Buzz Wire has the context you need.
Read more AI news →
