Stanford researchers used the Evo 2 genomic language model to design 16 functional bacteriophages from scratch—the first complete viral genomes generated by AI. The phages infected E. coli strains resistant to natural viruses, demonstrating potential for bespoke phage therapies against antibiotic-resistant bacteria. Brian Hie, a computational biologist at Stanford University and the Arc Institute, led the study published in Science.

The AI was trained on 2 million phage genomes and fine-tuned on 15,000 related sequences. From 302 candidates, only 16 produced working phages with 5,400 DNA letters and 11 genes. Oliver Crook at University of Oxford found these viruses were 97% identical to the ΦX174 template, mainly rearranging existing genetic material. However, some showed novel protein structures and infection dynamics.

Biosecurity experts at Johns Hopkins Center for Health Security warn the same methods could lower barriers to creating harmful biological agents. Hie notes significant hurdles remain: $100,000–$200,000 synthesis costs, domain expertise requirements, and limited understanding of how AI-proposed changes affect pathogenicity. Twist Bioscience provided discounted synthesis.

Chase Beisel of Locus Biosciences sees promise in AI's ability to explore genetic combinations evolution never produced. Samuel King, first author, notes one designed virus incorporated a truncated protein from a distant phage—a configuration previously nonviable through conventional engineering.