About 700,000 potential genome designs were produced, and 285 of those were chosen for synthesis and testing.
How two AI models were used to write complete bacteriophage genomes
Researchers instructed two models to generate complete genomes for a viable bacteriophage — a virus that infects and replicates inside bacteria. They used an existing bacteriophage, ΦX174 (pronounced “fie-ex-1-7-4”), known for its ability to infect and destroy E. coli, as an example for the models to follow. From roughly 700,000 candidate designs produced by the models, the researchers selected 285 that appeared most promising and moved those designs into the laboratory for synthesis.
From digital designs to living viruses: the laboratory steps
The chosen designs were turned into new DNA molecules by synthesis and then inserted into E. coli bacteria. After insertion, the researchers observed the cultures to see whether any viable bacteriophages would emerge. Shortly afterwards, 16 Petri dishes began to show clear spots as the viruses began to attack and replicate themselves inside the E. coli, demonstrating their viability. The experiment therefore produced multiple synthetic genomes that yielded living, self-replicating bacteriophages.
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleComparative performance: some AI-designed viruses outperformed ΦX174
Among the viable viruses recovered from the experiment, some proved more effective at attacking E. coli than the original ΦX174 bacteriophage used as the example. The result is concrete: AI-produced genome designs, once synthesized and introduced into bacteria, generated viable viruses and in some cases exceeded the laboratory effect of the natural template.
What this means for researchers, regulators, and the public
- Researchers: The work demonstrates a pipeline — model-generated complete genomes, selection of candidate designs, DNA synthesis, and biological testing — that yielded multiple viable bacteriophages and produced specimens that outperformed a known template. Researchers will note that the approach can produce many candidates rapidly and that laboratory synthesis and testing can convert those candidates into living entities.
- Regulators and policy makers: The experiment shows that synthesized genomes derived from model outputs can lead to viable viruses when inserted into bacteria, highlighting the intersection of computational design and wet-lab production that regulators must consider when assessing oversight and controls.
- The public: The source described the work as both exciting and terrifying, noting it as a positive use of a synthetic virus while also saying that negative uses are imaginable. The plainly stated contrast — beneficial research outcomes versus conceivable misuse — is central to public concern.
The reported sequence of events is straightforward and stark: two models produced roughly 700,000 candidate genomes; researchers chose 285 to synthesise; synthesis and bacterial insertion yielded viable bacteriophages in at least 16 cultures; and some of those synthetic viruses were more effective against E. coli than the original ΦX174 example. The source frames these facts succinctly and leaves the reader with a pointed observation about the dual nature of the advance — useful in some hands, dangerous in others.
Link to the original story: https://www.schneier.com/blog/archives/2026/08/ai-is-learning-to-write-genetic-code.html
