A language model trained on genomes designed a bacteriophage from scratch. Somewhere in the world, a person received it. What we know, what we do not, and why the gap should bother you.
The Sentence That Stops You
On October 20, 2026, a company called Radical Numerics published a technical blog post. Most of it is the sort of thing only a computational biologist reads for pleasure: recovery rates, mutation counts, held-out clades, average precision scores. Then, partway down, comes a sentence that does not belong in a blog post at all.
A phage cocktail designed with their model, the company writes, was administered to a patient with a drug-resistant infection through “an authorized, physician-led compassionate-use pathway.” They call it the “world’s first clinical use of AI-generated phages.” 1
Stop on that. A machine learning system wrote a genome. Someone built it out of DNA. A doctor put it in a human being.
I have spent two days with this blog post, its citations, and the literature behind it, and I want to walk you through exactly what happened, because the story contains both a genuinely impressive piece of engineering and a hole in the record the size of a hospital ward.
First, the Enemy
To understand why anyone would attempt this, you have to understand what antibiotic resistance has actually become. Not a looming problem. A current one, with a body count.
The Global Burden of Disease antimicrobial resistance collaborators published their most comprehensive accounting in The Lancet in 2024. Their estimate: in 2021, bacterial antimicrobial resistance was directly responsible for roughly 1.14 million deaths worldwide and associated with about 4.71 million. Their forecast to 2050 is worse, with resistance projected to be directly responsible for around 1.91 million deaths annually and associated with 8.22 million. 2
Those numbers describe a slow failure of the twentieth century’s best idea. Antibiotics are small molecules; bacteria evolve around them; we have not replaced them fast enough.
Which is why a much older idea keeps coming back. Bacteriophages, usually just called phages, are viruses that infect bacteria and nothing else. They were used as medicine in the 1920s and 1930s, fell out of Western favor when penicillin arrived, and persisted in Georgia and the former Soviet Union. They are exquisitely specific, which is both their great virtue and their central problem. A phage that destroys one strain of a bacterium may do nothing at all to the strain infecting your patient.
The case that put modern phage therapy on the map is worth telling properly. In 2019, Rebekah Dedrick and colleagues reported in Nature Medicine the treatment of a 15-year-old with cystic fibrosis who had undergone a double lung transplant and then developed a disseminated Mycobacterium abscessus infection. The team assembled a three-phage cocktail, two members of which had been genetically engineered to improve their killing, and gave it intravenously. The patient improved. 3
That case is the template: a desperate situation, no remaining antibiotic options, a bespoke biological agent built for one person. The bottleneck has always been the bespoke part. Finding or engineering a phage that kills the exact strain in front of you can take weeks or months, and patients in that situation do not reliably have weeks.
That bottleneck is the problem Radical Numerics says it is attacking.
The Model That Reads Genomes
Radical Numerics calls its system Omnii-1. In their description it is a genome language model, which is a phrase worth unpacking, because the analogy to chatbots is both illuminating and misleading.
A text language model learns the statistics of written language by predicting what comes next. A genome language model does the same thing to DNA, learning which sequences of nucleotides plausibly follow which others. What it absorbs, if it works, is something like the grammar of viable biology: which protein sequences fold, which genes function, which arrangements a real organism would tolerate. It is not reasoning about biology. It is a very sophisticated sense of what looks right.
The immediate predecessor here is EvoPhage, published in Science this year by King and colleagues, in which a genome language model designed variants of ΦX174, a bacteriophage with a genome of about 5.4 kilobases. 14 Small, elegant, well understood, and a reasonable first target.
Radical Numerics went after T7 instead. T7’s genome runs to roughly 40 kilobases, more than seven times larger. 1 That jump is the technical heart of the story. Designing a working variant of a small genome is hard. Designing one for a genome of that size, where a single bad choice in a single gene can produce an inert strand of DNA, is considerably harder.
Their specific target within T7 was gp17, the tail fiber protein. 1 If you think of a phage as a lunar lander (and the electron micrographs invite this: a head, legs, a landing posture), gp17 is the landing gear. It is the part that recognizes and grips the surface of a bacterium. Change the tail fiber and you change which bacteria the phage can attack. It is the natural place to aim if what you want is a phage tuned to the strain in a particular patient.
Does the Thing Even Work?
Here is the first hurdle, and it is brutally unforgiving: a designed genome either boots or it does not. Assemble the DNA, put it into bacteria, and either phage particles come out or nothing does. Biologists call this recovery, and it is as binary as biology gets.
Radical Numerics reports that at 11 to 15 amino acid changes from the wild-type sequence, designs from their model were recovered 69.9% of the time, against 4.8% for a comparison set. 1 That is a roughly fourteen-fold difference at a level of modification where random tinkering almost always produces a dud.
Working is not the same as doing anything useful, so the next test was activity. The company reports measuring 6,694 variants, of which 2,161 outperformed the wild-type phage, with functional designs carrying as many as 28 amino acid changes. Randomly mutated controls topped out at about 5. 1
That last comparison is the one I would put on a slide. Random mutation can get you five changes before the protein stops working. The model’s designs survived nearly six times that. It has, apparently, learned something real about which changes a tail fiber will tolerate, which is a form of knowledge no one wrote down anywhere for it to copy.
The Expensive Part Is Guessing Right
Laboratory screening is the real cost in phage therapy. Every candidate you test against every bacterial strain costs reagents, plates, and above all time. So the practical question is not “can the model design good phages” but “can it tell you which ones to test first.”
Radical Numerics ran that test. For two bacterial strains, 97 and 96 of the model’s top 100 ranked designs turned out to be active, and 57 of those were shared between the two lists. 1 On a strain the company characterizes as hard, where only about 4.5% of designs were active at baseline, 24% of the model’s top 100 were active. 1
A roughly fivefold enrichment on a hard target is not magic. It is something better: it is a screening budget that goes five times further. In a clinical setting where the clock is the enemy, that is the kind of improvement that changes what is possible.
The Number That Matters Most
Now the result I find most genuinely impressive, and it is a test of generalization rather than of design.
The hardest question you can ask a model trained on biological data is whether it has learned a principle or memorized a neighborhood. Phages and bacteria come in families; a model that has seen close relatives of your test case can look brilliant while knowing nothing transferable.
So the team held out an entire clade of 37 bacterial strains. Not individual strains scattered through the data, but a whole related branch, removed from training. Then they asked the model to pick phages for it.
The top two phages it selected covered 91.4% of the strains in that withheld clade. 1
Two phages. Thirty-seven strains the model had never been shown relatives of. Over nine in ten covered. If that result holds up under independent scrutiny, it suggests the model has internalized something about how tail fibers and bacterial surfaces fit together, rather than merely recalling which phage goes with which bug.
I want to be careful with my enthusiasm here, so: that number comes from the company, about the company’s model, in a blog post. It has not been through peer review. Hold it loosely.
Where the Model Struggles, Which the Company Says Out Loud
To its credit, the post does not only report wins. The team applied the same approach to Pseudomonas aeruginosa, one of the nastiest organisms in any hospital, and the results are notably thinner.
The reason is data. Their E. coli work rested on 32,681 measured phage-bacterium interactions. For P. aeruginosa they had 7,308. 1 Less than a quarter as much.
And the performance shows it. Average precision rose from 0.570 to 0.608, compared with 0.581 for an existing tool called GenoPHI. 1 A gain of under four hundredths. On E. coli, the model’s advantage is clear; on Pseudomonas, it is a modest nudge past the existing state of the art.
I find this the most honest paragraph in the whole post, and the most informative. It tells you the limiting reagent in AI biology is not model architecture. It is measurements. The model knows what it has been shown, and for the pathogens we most need help with, we have shown it relatively little.
And Then a Human Being
Which brings us back to the sentence.
Here is the entirety of what the public record contains about the patient. A person had a drug-resistant infection. A cocktail of phages designed with this system eliminated the clinical isolate in laboratory culture. The cocktail was then administered through what the company describes as an “authorized, physician-led compassionate-use pathway.” 1
And that is where the information stops. The company does not say what happened. Their stated reason, quoted exactly: “even limited clinical information could make the patient identifiable.” 1
I have turned this over repeatedly and I think it is both legitimate and insufficient, at the same time.
Legitimate, because it is probably true. Think about how few people on Earth have received an AI-designed phage: one. Any detail at all (country, organism, age bracket, hospital) narrows an already singular case. Medical privacy is not a formality, and the first patient in any new modality is maximally exposed. Patient confidentiality obligations are real and they bind the physician regardless of what the public would like to read.
Insufficient, because the announcement was still made. The company chose to publish the fact of a world first while withholding every piece of information that would let anyone evaluate it. The result is an unfalsifiable milestone. We are told something unprecedented happened in medicine and given no way to assess whether it helped, harmed, or did nothing.
And notice what the silence does rhetorically. “World’s first clinical use of AI-generated phages” reads as a triumph. But a first use is not a result. The sentence would be equally true if the patient had recovered, and equally true if they had not.
To be fair, and the post is fair on this point, Radical Numerics states plainly that “observations from a single patient cannot establish general safety or efficacy.” 1 They are not claiming a cure. They are claiming a first, and that claim is accurate as far as it goes.
What “Compassionate Use” Actually Means
A quick clarification, because the term travels badly.
Compassionate use, broadly speaking, refers to pathways by which a patient with a serious condition and no remaining approved options may receive an unapproved treatment outside a clinical trial. The specifics vary by country, and the blog post does not say which jurisdiction this occurred in, under what authorization, or from which regulator or ethics body. 1 Those details matter enormously and none are public.
What I can say about the category in general is this: it exists because sometimes the alternative to an untested treatment is no treatment at all, and because a dying patient’s calculus is legitimately different from a healthy volunteer’s. It is not a loophole. It is also not evidence. A compassionate use case tells you a treatment was permitted for one desperate person; it tells you nothing about whether it works.
The scientific record will not be enriched by this case unless the physicians involved publish it. I hope they do. A case report, with the patient’s consent and appropriate de-identification, is the normal and honorable way this gets into the literature. The Dedrick case did exactly that. 3
The Work It Builds On
A word about lineage, because this result did not come from nowhere.
The immediate precursor is a 2026 Science paper by King and colleagues, “Generative design of bacteriophages with genome language models.” 4 That is the EvoPhage work: designing whole phage genomes with a model of the same general type, demonstrated on ΦX174 at about 5.4 kilobases. Radical Numerics cites it as the state of the art it set out to exceed, and the comparison it draws is explicitly one of scale: T7’s roughly 40 kilobases against that 5.4. 1
So the honest framing of the engineering claim is not “AI can design phages,” which was already published in Science this year by someone else. It is “AI can design phages seven times bigger than the last ones, and rank them well enough to save a lab weeks.”
That is a narrower claim and a better one, because it is checkable.
It also leaves untouched the half of phage therapy that has nothing to do with design. A phage that kills bacteria in a dish must still survive in a human body long enough to reach the infection, and the human immune system has opinions about viruses regardless of whose side they are on. Antibody responses are a known, serious obstacle in the field: in one compassionate-use series of 20 patients, neutralizing antibodies appeared in 8 who received phages intravenously, possibly contributing to the lack of response in 4. 5 Nothing in this post engages with delivery, persistence, immune evasion, or dosing. Design is the part the model solves. The rest of the therapy remains exactly as hard as it was.
What I Think Is Actually True Here
Let me separate the three claims tangled together in this announcement, because they deserve very different levels of confidence.
The first claim is that a genome language model can design functional variants of a 40-kilobase phage at rates far above random, with activity beyond what mutagenesis reaches, and can rank candidates well enough to cut screening cost substantially, including on a held-out clade. This is a substantial engineering result, reported in detail, with internal comparisons and acknowledged failure cases. I believe it is probably broadly correct and I want to see it in a journal.
The second claim is that this constitutes the world’s first clinical use of AI-generated phages. Probably true, and essentially unverifiable from outside.
The third claim, the one nobody actually made but everyone will hear, is that AI-designed phages treat drug-resistant infections. That is not established by anything in this post. One patient, no outcome, no peer review, no independent replication.
Those three things will be collapsed into one in the retelling. They should not be.
The Ending
What strikes me, finally, is the asymmetry of the record.
We know that at 11 to 15 amino acid changes, 69.9% of designs were recovered. We know 2,161 of 6,694 variants beat the wild type. We know average precision on Pseudomonas moved from 0.570 to 0.608. We know two phages covered 91.4% of a withheld clade of 37 strains.
We do not know whether the patient got better.
There is something almost absurd about that ledger, and I do not think the absurdity is anyone’s villainy. The privacy reasoning is sound. The physician’s duty runs to the patient, not to my curiosity. But it leaves the most important experiment in the entire story as the only one without a reported result, and it leaves a company announcing a medical first on terms where no one can check it.
Somewhere, a person with an infection that antibiotics could not touch received a virus that no evolutionary process produced, whose genome was composed by a statistical model of life, assembled in a laboratory, and dripped into a human body under a physician’s supervision. That happened. It is one of the strangest things to have happened in medicine this decade.
And the only part of it that matters most to that person remains, for now, between them and their doctor. Which is probably where it belongs. But then the rest of us should be told rather less confidently that history was made.
References
- Radical Numerics. (2026, Oct 20). Omnii-1 and the design of therapeutic bacteriophages [Company technical blog]. https://www.radicalnumerics.ai/blog/omnii-phage (Authors listed: Iyengar, Siranosian, Sarkar, Noonan, Poli, Nguyen. Not peer reviewed.) ↩
- GBD 2021 Antimicrobial Resistance Collaborators. (2024). Global burden of bacterial antimicrobial resistance 1990-2021: a systematic analysis with forecasts to 2050. The Lancet, 404(10459), 1199-1226. https://doi.org/10.1016/S0140-6736(24)01867-1 ↩
- Dedrick, R. M., Guerrero-Bustamante, C. A., Garlena, R. A., et al. (2019). Engineered bacteriophages for treatment of a patient with a disseminated drug-resistant Mycobacterium abscessus. Nature Medicine, 25, 730-733. https://doi.org/10.1038/s41591-019-0437-z ↩
- King, S. H., Driscoll, C. L., Li, D. B., Guo, D., Merchant, A. T., Brixi, G., Wilkinson, M. E., & Hie, B. L. (2026). Generative design of bacteriophages with genome language models. Science, 393, eaec2657. https://doi.org/10.1126/science.aec2657 ↩
- Dedrick, R. M., Smith, B. E., Cristinziano, M., Freeman, K. G., Jacobs-Sera, D., Belessis, Y., et al. (2023). Phage therapy of Mycobacterium infections: Compassionate use of phages in 20 patients with drug-resistant mycobacterial disease. Clinical Infectious Diseases, 76(1), 103-112. https://doi.org/10.1093/cid/ciac453 ↩
