News · SI
Radical Numerics, the lab behind the Evo genome models, has launched Omnii-1, a multimodal genome language model, and says phages it designed were given to a patient with an antibiotic-resistant E. coli infection under an FDA-authorised compassionate-use pathway. The company is not disclosing how the patient fared.
Radical Numerics says phages designed by its new AI model have been given to a human patient, which it calls the first clinical use of AI-generated phages[1]. The San Francisco lab, built by the team behind the Evo and Evo 2 genome models, announced the case on 20 October alongside the launch of Omnii-1, a multimodal genome language model that learns across DNA, RNA, proteins and other molecular data[4].
01 / 04
Omnii-1 designed variants of T7, a well-studied phage that infects E. coli. Those variants went into a phage cocktail for a patient with an antibiotic-resistant E. coli infection, given under an FDA-authorised single-patient Investigational New Drug application through expanded access, better known as compassionate use[2]. In laboratory testing, the cocktail wiped out the bacteria isolated from the patient before it was administered[1].
The company says it designed the phages for this infection within weeks, where finding a matching natural phage can take months and often fails[4].
What it has not said is whether the patient got better. Radical Numerics says it will not disclose the clinical course or outcome because even limited details could identify the patient, and that "observations from a single patient cannot establish general safety or efficacy"[1]. The therapy is not approved for wider use.
02 / 04
Phages are viruses that kill bacteria, and they are picky: one that destroys a patient's strain may ignore a close relative. Rather than writing a whole genome, Omnii-1 rewrites targeted regions. For T7 it redesigned gp17, the tail-fibre protein that decides which bacteria the phage latches onto, while reading the DNA on both sides so the rest of the roughly 40-kilobase genome still works[1].
The company's own numbers are strong. At 11 to 15 amino acid changes from the natural protein, 69.9% of Omnii-1's designs produced recoverable phages, against 4.8% of random mutants. Of 6,694 designed variants with measurable activity, 2,161 outperformed the natural phage, some with up to 28 changes[1]. Asked to pick phages for E. coli strains it had not seen, its top two recommendations included an active phage for 92.1% of randomly held-out strains[3].
03 / 04
Phage design is one of eight applications Radical Numerics showed for Omnii-1. The others include interpreting genetic variants in rare disease, detecting early-stage cancer from a blood biomarker, designing antibodies, choosing cancer vaccine targets, screening for engineered pathogens and, with Lawrence Livermore National Laboratory, designing proteins that bind rare-earth elements for extraction and recycling[3].
04 / 04
Drug-resistant infections were linked to an estimated 4.95 million deaths in 2019[4], and phage therapy has long been held back by the slow hunt for a match. A model that can design a phage for one patient's bacteria in weeks would turn that hunt into an engineering problem. But this is one undisclosed case, with lab results reported by the company itself. The proof will be outcome data, more patients and independent replication, which Radical Numerics is seeking through clinical partners who can send it bacterial isolates to screen[1].
Sources
FAQ
Radical Numerics says variants of the T7 phage, a virus that kills E. coli bacteria, designed by its Omnii-1 model were part of a phage cocktail given to a patient with an antibiotic-resistant E. coli infection. In the lab, the cocktail killed the bacteria isolated from the patient. Treatment ran under an FDA-authorised single-patient Investigational New Drug application through expanded access, also called compassionate use.
We don't know. Radical Numerics says it will not disclose the patient's clinical course or outcome to protect their privacy, and stresses that the therapy is not approved and that a single patient cannot establish safety or efficacy.
Omnii-1 is a multimodal genome language model from Radical Numerics, the team behind the Evo and Evo 2 models. It learns across DNA, RNA, proteins and other molecular data, and the company has post-trained it for tasks including phage design, genetic variant interpretation, cancer detection, antibody design and pathogen screening.
In Radical Numerics' own tests on T7's tail-fibre protein, 69.9% of Omnii-1 designs with 11 to 15 amino acid changes produced recoverable phages, against 4.8% of random mutants. Of 6,694 designed variants with measurable activity, 2,161 beat the natural phage. Its top two phage picks included an active phage for 92.1% of randomly held-out E. coli strains.
Phages are very specific: one that kills a patient's bacterial strain may miss another. Today labs screen collections of natural phages against a patient's sample, which Radical Numerics says can take months and often fails. It says Omnii-1 designed phages for this patient's infection within weeks.
That’s the story
Spectrum Web Co ships SI and Web3 products, places talent and builds social narratives as monthly services.