Stanford AI designs new viruses with 50+ amino acid changes

On August 6, 2026, Stanford University researchers reported using large genome models to design 16 synthetic viruses that inhibit the growth of E. coli. The models, Evo 1 and Evo 2, were trained on DNA from bacteriophages and fine-tuned on the Microviridae family, which includes the test virus ΦX174. Prompted with short sequences from the start of the ΦX174 genome, they generated 302 proposed sequences. Filtering out outputs with damaged essential genes, extreme lengths, or unusual base compositions, the team chemically synthesized 285 sequences and inserted them into bacteria. Sixteen suppressed bacterial growth.

Most viable viruses resembled ΦX174, with viability climbing to 46 percent among those with at least 98 percent sequence similarity. Yet individual viruses showed notable differences: one lost a viral protein entirely, another gained a new gene, and one replaced a gene with a sequence from a distantly related virus. The AI-generated changes preserved function much more often than random mutations. Nearly a quarter of viruses carrying more than 25 amino acid changes remained viable, whereas random alteration would almost certainly inactivate them.

A cocktail of the 16 AI-designed viruses quickly evolved to infect E. coli strains resistant to natural bacteriophages, possibly through DNA swapping and new mutations. This points to a potential application in phage therapy. The researchers excluded vertebrate-infecting viruses from the training data but caution that anyone with sufficient computing resources could replicate the work with those viruses. The paper calls for better governance of large genome models and custom DNA synthesis, noting that AI regulation has largely failed to keep pace with the field’s rapid advancement.