Radical Numerics released Omnii-1 on Tuesday, October 20, 2026, a multimodal genome language model it calls "the world's most advanced," and says phages designed with it were administered to a patient with an antibiotic-resistant bacterial infection under a government-regulated compassionate use program. The company, founded by members of the team behind the Evo and Evo 2 models, describes this as the first AI-designed phage given to a human patient. That is the company's claim. No clinical outcome, peer-reviewed paper or independent confirmation has appeared in the materials we reviewed.
The launch is wide: eight applications spanning genome regulation, phage design, rare-disease genetics, cancer diagnostics, biodefense, rare-earth binders, antibodies and cancer vaccines. This post summarizes what Radical Numerics published in its Omnii-1 announcement and its Business Wire release, marks which numbers are company-reported, and explains what a skeptical reader should check. It sits alongside our coverage of other AI-for-science work, including Claude designing scorpion-toxin protein binders and Anthropic's novel enzyme discovery.
TL;DR: claimed vs verified
| Claim | Source | Status |
|---|---|---|
| Omnii-1 trained on 28 biological modalities, trillions of tokens | Radical Numerics | Company description; no model card or weights reviewed |
| AI-designed phages given to a patient under compassionate use | Company blog and press release | Unverified; no outcome or peer-reviewed report |
| A phage cocktail killed the patient's isolate in lab testing | Press release | Company statement |
| Top-two host recommendations include an active phage for 92.1% of eligible held-out E. coli strains | Company blog | Company benchmark |
| Beats AlphaGenome and Borzoi on regulatory and variant-effect benchmarks | Company blog | Not independently reproduced |
| Pathogenicity-related protein recall rises from 78.6% to 95.8% with Omnii added to DIAMOND | Company blog | Early internal test |
| Prospective screening studies for cancer detection | Company says still needed | Not shown |
What Omnii-1 is
Omnii-1 is a genome language model: the same family of technology as large language models, trained on biological data instead of text. Radical Numerics says it developed a way to convert 28 types of biological data into a common representation, so one model can reason across DNA sequence and the measurements aligned to it, such as chromatin accessibility, DNA methylation, histone marks, transcription factor binding and gene expression, conditioned on cell type and tissue.
The pitch mirrors the shift in text AI, where general models began to beat narrow ones. The company writes that "generalist models can now outperform specialized models on tasks they were individually designed to solve." Whether that holds in biology is exactly what independent replication will test. For a comparison with another genomics model integrated into agent workflows, see our post on AlphaGenome as a skill for Antigravity.
The phage case, step by step
Phages are viruses that kill bacteria. They are a long-discussed alternative for infections that no longer respond to antibiotics, but they are highly specific, so finding one that works against a given patient's strain can take months and often fails. The press release cites an estimated 4.95 million deaths associated with drug-resistant infections worldwide in 2019, 1.27 million directly attributable.
The company's account of the case:
- Researchers used Omnii to design phages targeting a patient's persistent drug-resistant infection "within weeks."
- A cocktail containing an AI-designed phage killed the patient's bacterial isolate in laboratory testing.
- The cocktail was administered at a clinical site under physician compassionate use access.
An infectious-disease physician at The Ottawa Hospital, Dr. Marisa Azad, is quoted in the release praising the approach: "instead of guessing and testing, we can point it at a resistant strain and get back designs built to match it." One endorsement is not clinical evidence. Missing from the public materials: how many patients, what organism strain, the dose and schedule, whether the patient recovered, whether the AI-designed phage or another cocktail component did the work, and whether any adverse events occurred.
Compassionate use means an unapproved treatment given to a patient with no good options, usually after a regulator and ethics board allow it. It is not a trial and does not measure efficacy. Reports of success in single cases are encouraging but cannot show a therapy works.
The model-side number is the host-prediction benchmark: on randomly held-out E. coli strains, Omnii-1's top two recommendations included an experimentally active phage for 92.1% of eligible strains. That tells you the model is good at shortlisting phages for strains like those it was evaluated on. It does not tell you about rare or newly emerging strains.
The other seven applications
- Omnii Genome. The company says it sets a new state of the art on regulatory-track and variant-effect benchmarks, ahead of AlphaGenome and Borzoi, and predicts single-CpG DNA methylation (Pearson r of 0.879 on its evaluation set) and reconstructs missing assays in unseen cellular contexts.
- Omnii Genetics. It nominates candidate causal variants in four of five unsolved congenital myopathy cases and ranks a functional variant in its top four picks in 90% of nearly 1,800 evaluated GWAS credible sets. Applied prospectively to 153 Parkinson's GWAS signals, it proposes mechanisms, including a predicted weakening of the NUCKS1 promoter. Those prospective calls are hypotheses until lab-tested.
- Omnii Diagnostics. Post-trained to predict cell-free 5hmC, a mark it never saw in pretraining, it reaches 0.854 correlation on predicted variant effects and improves early-stage cancer detection at 98% specificity in an evaluated cohort. The company states that prospective screening studies are still needed.
- Omnii Defense. Adding Omnii-1 to a conventional pipeline lifts recall of pathogenicity-related proteins from 78.6% to 95.8% on a set of 47,754 sequences, with precision of 64.9% versus 63.8%. With whole genera held out, the combined system recovers 44.6% of simulated dengue read pairs and 22.0% of Bundibugyo, versus under 0.1% for the conventional cascade. These are simulated reads in early tests.
- Omnii Rare Earths. With Lawrence Livermore National Laboratory, it ranks which metals a protein prefers to bind: median Spearman of 0.340 versus 0.318 for ESM-C 6B across 105 element pairs and 609 proteins. A small gap, with experimental validation of generated binders still to come.
- Omnii Antibody. Over three experimental rounds, designs improved the affinity-expression tradeoff, and two bound more tightly than the assay could resolve.
- Omnii Cancer Vaccine. It reaches 0.750 allele-averaged AUROC on class I immunogenicity versus 0.558 for BigMHC-IM.
Why a single model across 28 modalities is the interesting part
Most biological AI tools solve one problem: predict a protein's structure, score a variant, rank a peptide. Omnii-1's bet is transfer. The diagnostics result is the cleanest illustration in the launch materials: the model was post-trained to predict a blood-based epigenetic mark, 5hmC, that was absent from pretraining, and it reconstructed patient-specific profiles from inherited variation and sequencing data. The rare-earth result makes the same point from a different direction, since the model had no rare-earth binding data in pretraining yet ranked metal selectivity above the strongest public comparator the company tested.
If that transfer holds up under independent testing, it would mean a lab with a new assay or a new target does not have to build a model from scratch; it can adapt a general one. That would lower the cost of experiments in diagnostics, therapeutics and defense alike, and it explains why the company frames Omnii-1 as an engine rather than a product. The risk is the usual one with generalist models: broad claims across many benchmarks are easy to publish and hard to audit, and every comparison the company chose (which baselines, which cohorts, which held-out sets) shapes the headline. Several comparisons are also against narrow baselines on narrow tests, and some effect sizes are modest, like the rare-earth gap of 0.340 against 0.318.
The biosecurity angle
The company states plainly that "the same advances that make it possible to design medicines can also be used to engineer pathogens that existing detection methods may miss." Its response is Omnii Defense, offered as an API and edge deployments, plus an advisory group the release names: Andrew C. Weber, retired Maj. Gen. Paul Friedrichs, Jill Hruby and Eric Horvitz.
That framing matters. A model that can write phage genomes is a dual-use capability, and the quality of safeguards is not something a press release can establish. Compare how frontier labs handle biology risk in general in our look at Fable 5's biology safeguards update and Anthropic's Bay Area wet lab. Ask of any such model: who can access design capabilities, what screening applies to outputs, and whether red-team results are published.
How to read this as a non-biologist
- Separate model benchmarks from medical outcomes. Spearman and AUROC scores measure prediction, not patient benefit.
- Look for held-out and prospective tests. Several results are retrospective. Prospective ones, such as the Parkinson's signals, are proposals.
- Wait for a paper. The Evo phage work was reported in August 2026 and reviewed in the open; Omnii will need the same. The company's posts per application are a start.
- Watch the access model. The launch invites partners and offers a Defense API; it is not described as an open release.
What to watch next
- A preprint or paper with the patient case details and phage characterization.
- Independent benchmarking against AlphaGenome, Borzoi and ESM models.
- Any regulator or hospital statement on the compassionate use case.
- Whether weights, model cards or evaluation code are released.
Related background on AI for biology: Biohub's virtual biology initiative and the same company's earlier Halcyon image model.
Everything above reflects Radical Numerics' own October 20, 2026 materials. We have not reviewed an independent paper, model card, or clinical record. This is not medical advice.
Related reading
- Claude designed protein binders against scorpion toxins
- Claude discovers a novel enzyme system
- AlphaGenome skill for Antigravity
- Fable 5 biology safeguards
- Anthropic's Bay Area wet lab
- Biohub virtual biology initiative
- Radical Numerics Halcyon
- Sources: Radical Numerics: Introducing Omnii-1, Business Wire release via FinancialContent