Radical Numerics says a patient received AI-designed viruses, but withholds outcomes

AI-designed viruses have reached a human patient, according to Radical Numerics—but whether the treatment helped remains undisclosed. In an October 20 research account, the startup said viruses designed with its Omnii-1 model were administered against an antibiotic-resistant bacterial infection. The company describes a physician-led compassionate-use treatment, not a clinical trial demonstrating that the approach works.

A route to treatment outside a trial

Radical Numerics says the administration took place under an authorized compassionate-use pathway. That route allows patients with serious conditions to receive investigational therapies outside clinical trials when satisfactory alternatives are unavailable. It remains subject to oversight and safety requirements set by the relevant government authorities.

The treatment uses phages: viruses that infect bacteria. Phage therapy itself is not new, nor is the use of engineered phages in patients. The company's proposed advance is using AI to help design and select viruses suited to a particular infection, rather than relying only on a search through existing phage collections.

Designing a virus for a particular bacterium

That matching problem begins with a sample of the bacteria causing the infection. Researchers screen phages against it, looking for candidates that can kill the bacteria. A phage's narrow targeting can be useful, but it also makes finding a suitable treatment difficult: a virus that infects one bacterial strain may fail against another.

Omnii-1 is a genome language model, trained on genetic sequences and further trained on phage genomes. Radical Numerics says it can generate replacement DNA for selected regions while accounting for the surrounding sequence. This does not require inventing an entire viral genome: the company describes changing regions that help determine which bacteria a phage can recognize.

The company also trains the model on measured interactions between phages and bacteria to predict compatibility. Its workflow then returns to the laboratory: researchers make the designed variants, measure their activity and use those results to inform further design rounds. The model guides which candidates to test; physical experiments supply the feedback.

Laboratory activity is not a patient outcome

In an illustrative experiment with T7, a phage that infects E. coli, Radical Numerics modified a protein involved in recognizing bacterial hosts. Among designs with 11–15 amino acid changes, it recovered 69.9% of AI-designed variants, compared with 4.8% of random controls. Recovery is a proxy for whether a design can produce viable phages—not a measure of treatment success.

For the patient-specific treatment, Radical Numerics says its phage cocktail eliminated the patient's bacterial isolate in laboratory testing before administration. That is a result against a sample outside the body, not evidence that the patient's infection cleared.

The company's account is self-published and has not been peer reviewed. Futurism reports that Radical Numerics declined to share clinical information about the patient or the doctors with the Times, citing the proprietary nature of a treatment it hopes to monetize.

Radical Numerics says it is withholding the patient's clinical course and outcomes to protect privacy. CEO Eric Nguyen told the Times that the company intends to publish its findings through peer-reviewed research, with future disclosure of clinical details subject to appropriate consent.