AI-designed bacteriophages: the first demonstration of complete viral genomes

On 6 August 2026 *Science* published the study “Generative design of bacteriophages with genome language models” (DOI 10.1126/science.aec2657), with Samuel H. King as first author and Brian Hie, assistant professor of chemical engineering at Stanford, leading the research group (Arc Institute). For the first time, language models trained on DNA sequences – Evo 1 and Evo 2 – generated entire bacteriophage genomes, not just individual genes.

The models were fine-tuned on roughly 15,000 genomes from the Microviridae family and then prompted with fragments of the natural phage ΦX174 (a 5,386 nt genome). The Arc Institute's own note states that 285 genome designs were synthesised and tested; press releases and news coverage speak of “nearly 300”: the *Science* paper sits behind a paywall, and the detailed figures come from the group's institutional note on the preprint and from reports based on the Stanford release. Of those designs, 16 turned out to be functional and were confirmed by sequencing, carrying between 67 and 392 mutations relative to the natural reference genomes; 13 of them contain mutations absent from any known natural sequence (Arc Institute). A “cocktail” of these 16 phages broke through the resistance of *Escherichia coli* strains that had become immune to ΦX174, defeating two strains (per the *Science* paper) or three (per the Arc Institute note) in 1–5 passages.

The designed bacteriophages attack bacteria, not humans, animals or plants; according to the authors, Evo's training data deliberately excludes sequences of viruses that infect humans. The experiments were run on non-pathogenic laboratory strains (E. coli C) inside dedicated biosafety cabinets.

The accompanying editorial “AI-designed viral genomes” (DOI 10.1126/science.aej8512), signed by Thomas V. Inglesby and Moritz S. Hanke of the Johns Hopkins Center for Health Security, argues that screening of synthetic nucleic acid orders should be made legally mandatory — checking both the sequence and the identity of whoever places the order — and that methods able to detect AI-designed genomes are needed urgently. The study's own authors warn that the ability to generate new phage genomes raises significant biosecurity and biocontainment questions, and recommend expert consultation across the entire life cycle of such a project (King et al., *Science*). Among the stated limits: the length of the ΦX174 genome (5,386 nt) is the practical ceiling at current DNA synthesis costs, and standard gene prediction identified only 7 of the 11 genes because of overlapping reading frames.

In short, the proof of concept shows that AI can design complete viral genomes, opening the door to future phage applications against resistant bacteria. But the absence of in vivo data, the uncertainty over exactly how many genomes were synthesised, and the need for clear biosecurity rules all show that the road to clinical use is still a long one. — Pixie