AI genome models design new viruses that infect bacteria

Researchers at Stanford University used large genome models to design entirely new viruses that infect bacteria, demonstrating both the promise and potential risks of AI applied to genomic design.

The team trained two large genome models—Evo 1 and Evo 2—on over 2 million bases of DNA sequences from bacteriophages, then fine-tuned them on sequences specific to the Microviridae family. They chose ΦX174, a bacteria-killing virus that infects E. coli, as their test case. ΦX174 is simple and well-characterized: it has 11 genes spread over about 5,400 bases, each with an identified function. Crucially, it begins with a short, consistent sequence of bases—a feature that allowed the researchers to use that sequence as a prompt to guide the AI model. After experimenting with different prompt lengths, they found that four to nine bases of the starting sequence worked best to generate novel but related viral genomes.

To filter out implausible outputs, the researchers imposed several constraints. Any virus missing or severely damaged spike proteins (the genes encoding proteins that allow the virus to infect bacteria) was discarded, as were sequences shorter than 4,000 bases or longer than 6,000, those containing runs of the same base longer than 10, and those with unusual frequencies of GC and AT base pairings. These pre-conditions left 302 candidate sequences; 285 were chemically synthesized and tested in bacteria.

In most cases, the synthesized sequences produced nothing. But 16 of the 285 generated functional viruses capable of inhibiting E. coli growth—nine from direct AI output and seven that acquired mutations after insertion. The viable viruses showed a striking pattern: those most similar to the original ΦX174 (98 percent sequence similarity or higher) had a 46 percent viability rate, whereas the overall rate was only 5.6 percent. Yet individually, the viruses were quite diverse. One lost an entire viral protein and compensated elsewhere in the genome; another added a new gene entirely. Many had genes that were longer or shorter than the original, and one replaced a ΦX174 gene with a gene from a distantly related virus.

The researchers performed a revealing statistical analysis. Random mutation has roughly a 20 percent chance of inactivating a virus per amino acid change, so any virus with fewer than 25 changes should have only a 2.3 percent chance of viability. Yet nearly a quarter of the AI-generated viruses with more than 25 amino acid changes were viable, including two with over 50 alterations. This demonstrates that the AI learned to make changes that preserve viral function far more effectively than random mutation could achieve.

The potential medical value became clear in a head-to-head test. Many bacterial strains have evolved resistance to natural bacteriophage families, including ΦX174. A cocktail of natural bacteriophages failed to infect resistant E. coli strains in the laboratory. By contrast, a cocktail of the 16 AI-designed viruses succeeded—the researchers suspect through DNA-swapping and the appearance of additional mutations among the AI viruses. While phage therapy has been explored for years as a treatment for antibiotic-resistant infections, it has not seen widespread public use despite growing drug resistance.

However, the authors underscore the risks. To mitigate danger, they deliberately excluded viruses that infect complex cells (vertebrates) from their training data, reasoning that if the models cannot learn those patterns, they cannot easily generate them. Yet they note that anyone with sufficient computing resources could repeat the work with vertebrate virus sequences included—a scenario that has prompted them to call for better governance of genome models and custom DNA synthesis, though AI regulation so far has struggled to keep pace with the field's development.