Stanford Researchers Used AI to Design 16 Working Viruses — and the Pipeline Should Look Familiar to Developers
Two hundred eighty-five synthetic genomes were chemically manufactured and pushed into E. coli, and 16 of them came out working — a 5.6 percent hit rate, as Ars Technica science editor John Timmer reported on August 6, 2026. It is the yield from what the researchers describe in their preprint as "the first generative design of viable bacteriophage genomes" — an entire viral genome written by an AI model rather than engineered one gene at a time. The ratio is the part worth reading twice, because almost none of the engineering that produced it lived inside the model.The work comes from a team at Stanford University, published in Science (DOI 10.1126/science.aec2657) and covered the same day by C&EN and Inside Precision Medicine. It is the peer-reviewed version of a result the group first posted as a preprint on bioRxiv on September 17, 2025, so the underlying experiment is roughly a year old; what changed this week is that it cleared review.
What the pipeline actually did
The generator was Evo 1 and Evo 2 — genome language models trained the way an LLM is trained on text, except the alphabet is A, T, C, and G. Evo 2 was released by Arc Institute in collaboration with Nvidia and trained on over 9.3 trillion nucleotides drawn from across the tree of life; the study's senior author, Brian Hie, is an innovation investigator at Arc Institute as well as an assistant professor of chemical engineering at Stanford. The design template was ΦX174, a bacteriophage that infects E. coli and, at 5,386 nucleotides encoding 11 genes, is small enough to synthesize affordably and well-understood enough to grade. The team fine-tuned both models on roughly fourteen thousand Microviridae sequences — the viral family ΦX174 belongs to — then prompted them with a short fragment of the genome's start sequence.
That prompt turned out to be a tuning knob with a narrow sweet spot. Feed the model too much of the opening sequence and it simply reproduces the rest of the real genome; feed it too little and it generates unrelated sequences. The window that worked was four to nine bases of that opening sequence — a narrow band between plagiarism and noise.
Then came the part that did the real work. Raw output was run through a stack of cheap, mechanical rejection rules before anything got built: discard any design whose spike protein — the structure the virus uses to latch onto a bacterium — fell below 60 percent identity to the real one; discard anything shorter than 4,000 or longer than 6,000 bases; discard runs of the same base longer than 10; discard sequences with unusual GC and AT ratios. The researchers separately built a custom gene-annotation pipeline, because off-the-shelf gene finders could locate at most 7 of ΦX174's 11 genes — its genes overlap, which breaks standard tools — and required at least 7 protein hits before a design advanced.
Thousands of generated sequences went in. Three hundred and two came out the far side as testable proposals; 285 of those were synthesized and inserted into bacteria, and 16 inhibited E. coli growth, meaning they functioned as viruses. Nine of those 16 were exactly what the models wrote; the other seven only became viable after picking up further mutations inside the bacteria.Why the shape of this should look familiar
Strip out the biology and this is an architecture most developers already run: a broad generative model proposes, and a stack of narrow, deterministic verifiers rejects. It is the same shape as a coding agent that drafts a patch, then hits a type checker, then a linter, then a test suite, then a human reviewer — each gate cheaper than the one after it, each one throwing away work the model was confident about.
Two design choices in the phage pipeline transfer directly. The first is ordering by cost. The sequence-level filters are pure computation and can be run over every candidate; DNA synthesis and laboratory assays cost real money and real weeks. The expensive verifier only ever saw 285 candidates because the cheap ones had already killed the rest — the same reason you run tsc before you run an integration suite, not after.
The second is that the acceptance rate is allowed to be terrible. A 5.6 percent yield would be an alarming pass rate for a human engineer and a perfectly fine one for a generator whose proposals are nearly free. The interesting figure buried in the results is what happens when you tighten the filter: among outputs with 98 percent or greater sequence similarity to the original ΦX174, viability rose to 46 percent. Constrain the generator harder and it succeeds far more often — while producing far less that is new. That trade-off will be familiar to anyone who has tuned an agent's temperature against its merge rate.
And the model was not merely reshuffling. Past work found that a single amino-acid change in one of ΦX174's proteins carries roughly a one-in-five risk, on average, of killing the virus outright — as Timmer puts it, "ΦX174 doesn't just fear change; it's typically murdered by it." Run that math forward, and above 25 individual amino-acid changes the chance of producing a viable virus is essentially zero. Nearly a quarter of the AI's designs in that range were viable, including two carrying more than 50 alterations. Hie framed the underlying capability to C&EN this way — the space of nucleotide combinations that could form a gene or a genome is enormous, and almost all of it is dead ends. "But only a much smaller subset of those sequences are biologically plausible or fit into the biological world." The model had learned where that subset is.
The practical payoff
The applied result is phage therapy. Resistance to whole families of phages, ΦX174 included, is something bacteria have already evolved on their own, and the team's answer was to test a cocktail. Pitted against resistant hosts, a cocktail assembled from wild phages that ordinarily attack E. coli failed outright, while the AI-generated set rapidly evolved its way into those hosts; the team's own account puts it at resistance overcome in all three strains within one to five passages, where ΦX174 alone failed completely. Samuel King, the paper's first author, told C&EN the work is "a proof of concept of what's possible for genome design guided by AI" — and proof of concept is the right ceiling here. As Ars Technica notes, phage therapies have been under consideration for years without reaching widespread public use, despite worsening antibiotic resistance.
The biosecurity caveat
The dual-use question is not an afterthought bolted on by reporters; the authors put it in their own paper. As a precaution while training the models, the researchers withheld all sequences from viruses that target complex cells, and in their own technical writeup King and Hie state that "Evo cannot generate human viral sequences due to deliberate training data exclusions, preventing both accidental and intentional misuse for pathogen design."
But that is a property of one training run, not of the method, and Timmer puts the consequence plainly: although the team excluded viruses that infect vertebrates from its training data, "anyone with access to sufficient computing resources could repeat the process with those viruses included."
The authors write that generating new phage genomes with AI "raises important biosafety, biocontainment, and biosecurity considerations necessitating discourse on the strengths and limitations of governance, policies, and misuse mitigation strategies," adding: "Groups conducting future whole-genome design work should consult both safety and security professionals throughout the project lifecycle." Both passages are quoted from the paper by Inside Precision Medicine, which also reports that an accompanying Science editorial by Thomas Inglesby and Moritz Hanke of Johns Hopkins University goes further, arguing that screening of synthetic DNA orders should be legally mandatory rather than the current voluntary practice. Their warning: "Because AI-generated genomes can be very different from previously characterized nucleic acid sequences, screening methods that can flag new concerning designs need to be urgently developed." And their framing of the choice: "The question is no longer whether generative viral genome design will exist. It is whether society can build oversight that allows its benefits to unfold while preventing it from enabling serious harm."
That is the uncomfortable symmetry with the generate-then-filter pattern. The filters in this pipeline were what made the output useful — and in a different lab, with a different training set, the filters are also the only thing standing between the generator and something nobody wants built. Neither of those depends on the model.
Sources
Written by the vybecoding.ai editorial team
Published on August 6, 2026