Overview
- The Stanford and Arc Institute study, which published in Science on Thursday, trained genome‑language models called Evo to generate about 700,000 candidate viral genomes and synthesized roughly 285–302 designs for lab testing.
- Sixteen of those synthesized designs produced viable bacteriophages that infect only E. coli, and a cocktail of the new phages outperformed a comparable mixture of natural phages against some antibiotic‑resistant strains.
- The researchers excluded viruses that infect animals and humans from their training data, carried out experiments in secure facilities, and publicly released the Evo2 model to help advance research on natural pathogens.
- Commentary from the Johns Hopkins Center for Health Security said the technical capability now exists but that governance, screening and biosecurity rules for computational genome design lag behind and should restrict attempts to create eukaryote‑infecting pathogens.
- The phage genomes used are very small—about 5,400 base pairs—so experts say scaling this approach to larger viruses or cellular genomes will be substantially harder, a fact that sharpens the immediate policy focus on lab rules, sequence screening and oversight.