Radical Numerics' AI-phage "first" is real as a claim, but the FDA and outcome details are unverified

What we checked

• Can the claimed FDA-authorized expanded-access use and the AI-designed role in the administered phage cocktail be independently verified? partly confirmed. The AI-design capability rests on peer-reviewed generative-genomics work (Evo/Science, Aug 2026) and supporting studies, but the specific administered cocktail and its in-vitro result are only in an unreviewed company report. The FDA single-patient expanded-access claim appears only in press coverage quoting a company spokesperson; the company's own blog and press release never name the FDA, and such authorizations are not publicly listed, so no independent confirmation exists. [1][2][3][4]
• Do the company's own written materials name the FDA or a specific US regulatory pathway? refuted. Both the Omnii-1 blog and the Business Wire press release describe only a 'physician compassionate use' or 'government-regulated compassionate use program' and 'oversight by the relevant government regulatory authorities.' The FDA and the phrase 'single-patient expanded access' appear only in secondary reporting attributing them to a spokesperson. [5][2]
• Has any patient clinical outcome been disclosed or independently assessed? refuted. No in-vivo or clinical result is reported anywhere; the only efficacy data is that the cocktail killed the patient's isolate in vitro. The company explicitly withholds the clinical course, and an independent microbiologist quoted in reporting withheld judgment pending completion and outside review. [1][4]

What we found

• Radical Numerics' own blog post and the Business Wire press release describe only 'physician compassionate use access' and oversight by 'the relevant government regulatory authorities.' Neither primary document names the FDA, 'single-patient expanded access,' or an emergency IND. [1][2]
• The specific 'FDA single-patient expanded access' detail surfaces only in secondary coverage that attributes it to a company spokesperson (via New York Times reporting). The company declined to give the authorization date or name the physicians, citing patient privacy and the treatment's proprietary, commercial nature. FDA does not publicly list individual expanded-access authorizations, so this cannot be independently confirmed. [3][4]
• No clinical outcome has been disclosed. The only efficacy evidence is that the cocktail 'killed the patient's bacterial isolate in laboratory testing' (in vitro); the company says observations from a single patient cannot establish safety or efficacy and has not said whether the infection cleared. [1][2]
• The clinical case itself is not peer-reviewed. The Omnii-1 model page cites no peer-reviewed validation and states 'prospective screening studies are still needed to establish clinical benefit.' Independent reporting notes the technical findings have not been published or reviewed, and a Harvard microbiologist not involved said he would withhold judgment until the treatment completes and outsiders assess how novel the viruses were. [5][4]
• The underlying AI capability is more substantiated than the clinical event. Reporting describes T7—an E. coli phage—whose tail-fiber (receptor-binding) genes Omnii was used to redesign so the phage could attach to resistant strains, and the blog reports supporting figures (e.g., 2,161 of 6,694 measured variants exceeding wild-type activity) plus peer-reviewed generative-virus work from the Evo lineage first reported in August 2026. [3][1]
• Radical Numerics is the ex-Evo/Liquid AI team (CEO Eric Nguyen) that raised $50M seed led by Emergence Capital in June 2026, positioning Omnii across human health and biodefense; the June coverage made no mention of phage therapy or clinical plans, so the clinical move is a rapid new step for the company. [6]

What it means

Treat this as a credible but unverified milestone, not a validated clinical result. The scientifically defensible part is narrow and plausible: using a genome model to mutate a phage's receptor-binding proteins to retarget a resistant strain, validated in vitro—an approach consistent with peer-reviewed generative-genomics work. Everything load-bearing for a patient—regulatory authorization, who administered it, and whether it helped—is either absent from the primary materials or attributed only to the company. Builders should note that the FDA framing lives in press quotes, not in Radical Numerics' own written record, and that single-patient expanded-access events are legitimately private, meaning "independent verification" may never come through official channels; the real test will be a peer-reviewed case report.\n\nThe practical lesson for AI-systems builders: a model producing a lab-active biological design is a genuine capability signal, but the gap between "kills the isolate in a dish" and "safe, effective therapy" is enormous and unaddressed here. The announcement also fuses a product launch with a human-subject claim, so weight the self-reported benchmarks (SOTA over AlphaGenome/Borzoi, etc.) as unaudited. If you are evaluating Omnii-1 or similar models, demand the peer-reviewed case and the model/benchmark methodology rather than the launch-day narrative, and watch the biosecurity dimension: the same retargeting capability is dual-use.

Still open

• Will the FDA pathway (e.g., single-patient expanded access/eIND) ever be documented beyond a company spokesperson quote?
• What was the patient's clinical outcome, and will a peer-reviewed case report appear?
• Which bacterium and exact cocktail composition were used (reporting says resistant E. coli; the company's written materials don't name it)?
• How much of the administered cocktail was AI-designed versus natural phage, and who manufactured and administered it?