AI-designed phages reach a patient in a world first

What happened

On October 20, Radical Numerics announced (Radical Numerics) that phages designed with Omnii-1, its multimodal genome language model, reached a patient. The company calls it the world's first clinical use of AI-designed phages. The designs were administered to a patient with an antibiotic-resistant bacterial infection through an authorized, physician-led compassionate-use pathway.

In laboratory testing, the cocktail containing the designed phage variants eliminated the patient's resistant isolate in vitro. The company is not disclosing the clinical course or the outcome, because even limited detail could identify the patient. Radical Numerics also notes that observations from a single patient cannot establish general safety or efficacy.

How Omnii-1 designs phages

Phage therapy is a long-standing hope against resistant infections, but finding a suitable phage still relies mostly on empirical screening. Omnii-1 does not write a whole genome. It rewrites selected regions of a phage genome using the surrounding sequence as context on both sides. The worked example is gp17, the receptor-binding tail fiber protein of the E. coli phage T7. The T7 genome is around 40 kilobases, more than seven times the size of the ΦX174 genome used in earlier Evo work.

The model was also trained on public phage-host interaction data, so it can rank which variants are likely to be active against a given bacterial strain. Synthesized variants are measured in the lab, and those measurements feed the next design round.

The numbers

• Viability: at 11-15 amino acid changes, 69.9% of designed variants were recovered, against 4.8% of random controls.
• Activity: of 6,694 variants with measurable activity, 2,161 exceeded wild type, including variants with up to 28 amino acid changes. Random controls beating wild type had at most five changes.
• Prioritization: for two bacterial strains, 97 and 96 of the top 100 candidates cleared the activity threshold. On a harder third strain, only 4.5% of the library cleared it, compared with 24% of the model's top 100.
• Host prediction: holding out an entire E. coli clade of 37 strains, Omnii-1 reached 91.4% of eligible strains with just two phages, ahead of task-specific methods.
• Pseudomonas aeruginosa: adding E. coli interaction data lifted average precision from 0.570 to 0.608, above GenoPHI's 0.581.

Why it matters

Antibiotic resistance is projected to cause nearly 40 million deaths between 2025 and 2050. The main obstacle to personalized phage therapy is time, because a patient with a severe infection cannot wait months while a matching phage is found. A genome model that predicts which variant is worth making could shorten that loop.

For now these are the company's own measurements, without peer review, and the clinical part covers a single patient with no reported outcome. Radical Numerics is looking for clinical partners who can supply bacterial isolates, and it promises to keep developing its biosafety and biosecurity protocols. If you work on AI-driven biological design, the post is a useful marker of how far the field has moved from the lab toward the bedside.