
AI-Designed Viruses Given to Human Patient for First Time
Radical Numerics, a San Francisco AI lab, announced this week that a cocktail of bacteriophages designed by its Omnii-1 model was administered to a patient with an antibiotic-resistant bacterial infection, which the company calls the first time AI-designed phages have been given to a human patient. The treatment was delivered under compassionate use, a regulatory pathway for patients with no approved alternatives, not as an approved drug.
The implementing institution is Radical Numerics, an AI research lab building general biological intelligence, co-founded and led by CEO Eric Nguyen. The lab publicly unveiled Omnii-1 this week as its most powerful genome language model. Administration took place at a clinical site under a physician's compassionate-use access.
The viruses are bacteriophages — they infect and kill bacteria, not people. They were combined into a therapeutic cocktail targeting the patient's persistent drug-resistant infection. The company said the cocktail containing Omnii-1-designed phage variants eliminated the patient's antibiotic-resistant isolate in vitro. Phage therapy predates this work and has long been studied as an alternative as antimicrobial resistance makes some infections harder to treat with antibiotics.
Omnii-1 is a multimodal genome language model that learns across DNA, RNA, proteins and other molecular data. Researchers used it to design phages targeting the patient's infection within weeks. The model redesigned the virus's leg-like structures so it could attach to bacterial strains existing phages could not infect. The work built on T7, a phage that infects only E. coli and looks like a microscopic lunar lander — a ball-shaped protein shell holding DNA atop six leglike protrusions. In host-prediction benchmarks, Omnii-1's top two recommendations included a biologically active phage for 92.1% of eligible, randomly held-out E. coli strains.
"For most of medicine's history, we've searched for drugs that might work against a disease," Nguyen said. "We're entering a future where we can design medicines for the exact biology of a patient." But this is a company's claimed first, with no peer-reviewed data or clinical outcomes published; a single compassionate-use case is not proof of efficacy. The ability to write new viral genomes also raises biosecurity and governance questions about the boundaries and misuse potential of such models.
What follows must be clinical data — patient outcomes, safety signals, resistance changes — plus independent replication. Whether AI-designed biologics move from single compassionate-use cases to routine care will depend on regulators' approval pathways, biosafety review, and access frameworks.