Key Takeaways
- Scientific result: The artificial intelligence Evo, developed by the Arc Institute and Stanford, designed complete viral genomes: 16 synthetic bacteriophages replicated and killed bacteria in laboratory conditions.
- Technology used: The model, trained on 9.3 trillion nucleotides, generated 700,000 genomic designs starting from the phage ΦX174; one variant, Evo-Φ69, multiplied 65 times over, outperforming the natural virus.
- Critical issue: NIH regulations do not cover the computational design of viral DNA, while the Evo models remain open source and accessible on HuggingFace.
A viral genome designed from scratch and turned into a living virus
The Arc Institute, in collaboration with Stanford, published findings on August 6, 2026 in Science describing an experiment in which a generative artificial intelligence designed complete viral genomes, which were then synthesized and turned into functioning viruses. The study, led by professor Brian Hie and doctoral candidate Samuel King, marks the first experimental verification of a genome designed entirely by an AI model.
The system used, called Evo, processes DNA sequences rather than text. It was trained on roughly two million bacteriophage genomes and a total of 9.3 trillion nucleotides drawn from 128,000 organisms.
From a natural bacteriophage to 16 working variants
The starting point was ΦX174, an 11-gene bacteriophage with 5,386 base pairs. From this template, Evo 1 and Evo 2 produced 700,000 genomic designs, narrowed down through computational filtering to 302 candidates for chemical synthesis. Of these, 285 were physically assembled as DNA strands and introduced into E. coli cultures. Sixteen sequences produced living bacteriophages, capable of replicating and destroying their host bacteria.
In a competitive test against the natural virus, the variant Evo-Φ69 multiplied its population 65 times over, consistently outperforming wild-type ΦX174. Another variant, Evo-Φ36, incorporated a gene from an evolutionarily distant phage — a modification that, based on established biological knowledge, should have rendered the virus non-functional. Biochemist Oliver Crook, of the University of Oxford, described these organisms as "genetic architectures a human scientist would never have thought to design."
An application against bacterial resistance
A cocktail containing the 16 synthetic phages eliminated two distinct strains of E. coli resistant to a natural bacteriophage, offering a possible path toward phage therapy against antibiotic-resistant infections. The study's authors describe the result as "a blueprint for the design of synthetic bacteriophages" and a foundation for "generative design of living systems at genomic scale."
The regulatory gap on biosecurity
The National Institutes of Health policy, published in late July 2026, prohibits experiments that increase the danger posed by known pathogens, but does not regulate the purely computational design of viral DNA. Moritz Hanke, of the Johns Hopkins Center for Health Security, pointed to the absence of criteria for classifying the risk of a virus that has never existed before.
The research team excluded from Evo's training data any genetic material related to viruses capable of infecting humans, animals, or plants. J. Craig Venter, a pioneer of synthetic DNA, called the concerns "serious" regarding a potential application of the method to pathogens such as smallpox or anthrax, emphasizing the risk of unpredictable and random outcomes.
The Evo models are available as open-source code on platforms such as HuggingFace, along with their parameters and training data. The HIV genome contains roughly 10,000 bases, while the coronavirus genome contains about 30,000 — sizes already comparable to those the system can handle.
Outlook
Patrick Cai, a synthetic biologist at the University of Manchester, called the work a "milestone." Jason Kelly, CEO of Ginkgo Bioworks, proposed making AI-driven cellular design a national research priority in the United States, supported by automated laboratories capable of continuously testing new genomic designs generated by the models. The gap between technical capability and regulatory framework remains the open question surrounding the development.
