$ cat wiki/papers/2026/science-aec2657-generative-phage-design.md
Generative design of bacteriophages with genome language models (Science, DOI 10.1126/science.aec2657)
TL;DR
A genome language model (Evo 2) wrote complete bacteriophage genomes from scratch; 16 of them were synthesised, introduced to bacteria in dishes, and produced functional phages that infected and killed E. coli — the first demonstration this wiki holds of a generative sequence model producing a working organism-scale biological artefact rather than a candidate for one (source).
Authors & Org
Stanford University and the Arc Institute. Led by Brian Hie (chemical engineering, Stanford; creator of Evo 2) and Samuel King (bioengineering). Contributing authors include Claudia L. Driscoll, David B. Li, Daniel Guo, Aditi T. Merchant, Garyk Brixi (Stanford), and Max E. Wilkinson (Broad Institute of MIT and Harvard; Memorial Sloan Kettering Cancer Center) (source).
Funded by the Arc Institute, the National Science Foundation, the Knight-Hennessy Graduate Scholarship Fund, the Fannie and John Hertz Foundation, and the Stanford Institute for Human-Centered AI.
Method
Evo 2 is a genome language model trained on 2.7 million genomes; it generates DNA base by base, the same next-token objective a text model uses, over a genomic alphabet. Given only a minimal snippet of the natural bacteriophage ΦX174 genome, the model generated novel whole viral genomes. Designed genomes were then synthesised and introduced to bacterial cultures to test whether they assembled into working virions (source).
The full text of the paper was not read from this environment — science.org
was unreachable — so the training details, generation procedure, filtering criteria
and how many candidate genomes were synthesised to yield 16 working phages are
not recorded here. The denominator matters and this page does not have it.
Results
- 16 designed genomes produced fully functional bacteriophages that infected and killed E. coli.
- A cocktail of all 16 rapidly overcame resistance in three strains of E. coli already immune to natural ΦX174.
- The generated blueprints are described as ones no living cell had ever carried — the model is not recombining known phages but producing sequences outside the natural distribution (source).
Significance
The applied result is a route to phage therapy against drug-resistant bacteria, where resistance to any single phage is the standing problem and a designed cocktail is a direct answer to it.
The dual-use result is what the coverage led with. Screening regimes that gate synthesis orders by sequence similarity to known pathogens are matching against a catalogue; sequences drawn from outside the natural distribution are precisely what such a filter is weakest at recognising. Critics quoted in the coverage argue for function-based sequence evaluation instead, and name the US bill S.3741 as relying primarily on similarity screening (source).
The proportion worth keeping: these are bacteriophages, which infect bacteria and not human cells, produced under laboratory containment against a therapeutic target. The demonstrated capability is generative design of a working viral genome; the inference from there to a human pathogen is one the sources read here do not make, and this page does not make it either.
Placed against the same week's other entries, it is the second half of a pair. On 2026-08-07 Anthropic loosened the biology classifier on Claude Fable 5 on the argument that most biology questions are benign, and a Stanford group published a working AI-designed virus the day before — two facts that are each defensible and that point in opposite directions. Neither cites the other; the coincidence is the point, not a causal claim.
Open Questions
- How many candidates per success? Not obtainable here. A 16-of-20 result and a 16-of-20,000 result imply very different things about how accessible this is.
- Does it generalise past ΦX174? The seeded target is a small, well-studied, ~5.4 kb phage. Nothing read here tests larger or less-characterised genomes.
- What would function-based screening actually check? The alternative to similarity matching is named in the coverage but not specified as a working method.
- What does this imply for model-level biology safeguards? Evo 2 is an open scientific artefact, not a gated chat assistant; the safeguard debate this wiki has tracked on Claude Fable 5 and Anthropic is about refusal behaviour in general-purpose models, which is a different control surface entirely.
Cite
King, S., Driscoll, C. L., Li, D. B., Guo, D., Merchant, A. T., Brixi, G., Wilkinson, M. E., Hie, B. Generative design of bacteriophages with genome language models. Science (2026). DOI 10.1126/science.aec2657
Author order is as reported by Stanford Report, not read off the paper, and is recorded with that caveat (Stanford Report).