"Generative design of bacteriophages with genome language models" is a research paper published in Science 393(6811) on August 6, 2026 (DOI 10.1126/science.aec2657) by Samuel H. King, Claudia L. Driscoll, David B. Li, Daniel Guo, Aditi T. Merchant, Garyk Brixi, Max E. Wilkinson and Brian L. Hie. It reports what its abstract calls "the first generative design of complete bacteriophage genomes using genome language models," using Evo 2 and its predecessor Evo 1. Science published it alongside a Perspective by the Johns Hopkins Center for Health Security (AI-designed viral genomes).
Method and results
The design target was a bacteriophage — a virus that infects bacteria — with specificity for the bacterial host Escherichia coli C. The natural phage ΦX174 served as the design template. The authors describe establishing "a framework for generating and evaluating thousands of AI-generated genomes, nearly 300 of which we chemically synthesized and tested in laboratory conditions, yielding 16 viable phages."
The 16 viable phages are reported to show strong host specificity and diverse fitness profiles, including competitive infection kinetics, and to differ from any known natural phage — "exhibiting de novo mutations, divergent genes and regulatory elements, and variable genome lengths." Cryo-electron microscopy confirmed that one generated phage uses a DNA packaging protein from an evolutionarily distant phage in its capsid structure.
The application the paper advances is phage therapy against bacteria that evolve resistance. A cocktail of the generated phages "rapidly overcomes ΦX174-resistant Escherichia coli strains," where the paper reports a comparable mixture of naturally sourced ΦX174-like phages could not. The authors frame the contribution as expanding "what synthetic genomics can achieve alongside methods such as directed evolution and rational engineering," and as a foundation for "the generative design of larger, more complex genomes."
The number synthesized is stated in the structured abstract as "nearly 300." Reporting published the same day gave the figure as 302 (Source: theguardian.com), while an ABC News (Australia) account published the following day put it at 285 selected for synthesis out of roughly 700,000 candidate designs generated (Source: abc.net.au). The paper body, which would settle the exact count, is paywalled and was not read; all three figures are recorded rather than reconciled.
The ABC account also quotes study author Brian Hie, an assistant professor at Stanford, saying the model generated the entire genome end-to-end in a single left-to-right pass with nothing added, and records that the researchers have made Evo 2 freely available on the argument that the risk is offset by the benefit of applying such tools to existing natural pathogens. Moritz Hanke of the Johns Hopkins Center for Health Security, who was not involved in the work, described to the New York Times a "huge disconnect" between the pace of the science and the pace of accompanying regulation (Source: abc.net.au); Hanke is a co-author of the accompanying Science Perspective (AI-designed viral genomes).
Scope of the claim
The paper's own claim is confined to bacteriophage genomes designed with genome language models. Contemporaneous coverage described the result more broadly as the first complete genomes designed by AI, and as the first AI-designed viruses (Source: theguardian.com; bbc.com). Both framings appear in the record; the narrower one is the paper's.
The organisms designed infect bacteria, not humans, animals or plants. The relevant training-side control is stated not in this paper but in the Evo 2 technical paper: genomic sequences from viruses that infect eukaryotic hosts were excluded from the Evo training corpus, and the authors of that paper note that task-specific post-training may circumvent the exclusion (Genome modelling and design across all domains of life with Evo 2).
Reception
Thomas V. Inglesby and Moritz S. Hanke of the Johns Hopkins Center for Health Security published a Perspective in the same issue arguing that the capability to compose viral genomes with generative AI now exists while the governance to steer it safely does not, and — as quoted in reporting of the same date — that new viruses with the potential to cause disease "should not be pursued" (AI-designed viral genomes).
Brian Hie has defended open availability, arguing that existing pathogens present a greater risk than potential AI designs because they are easier to access and produce, that safety checks can be built into an AI tool but not into natural evolution, and that such tools improve defensive options against both natural pandemics and engineered biological threats (Source: news.stanford.edu).
Provenance
AAAS paywalls the full text. The editor's summary, the structured abstract, the graphical-abstract caption and the article abstract were retrieved verbatim from the publisher's DOI landing page; the results and discussion body, methods, figures and references were not retrievable and no claim here rests on them. Publisher metadata confirms the title, the eight-author list in order, the date of August 6, 2026, DOI 10.1126/science.aec2657, issue 393:6811, and the related-article DOI 10.1126/science.aej8512. Confidence is held at medium because the body was not read.
Relationships
- depends-on: Genome modelling and design across all domains of life with Evo 2, Evo 2 (Arc Institute / NVIDIA)
- supports: AI for Science, AI Biosecurity
- instance-of: Dual-Use Frontier AI
- contradicts: AI-designed viral genomes — on whether this class of work should proceed
- related: Open-Weight Frontier Models, CBRN Uplift, Healthcare — AI Deployment