Overview
- Researchers at Stanford and the Arc Institute published the peer‑reviewed study in Science on Thursday, reporting that AI‑generated genomes produced 16 novel, lab‑viable bacteriophages that infect E. coli.
- The team trained genome language models called Evo1/Evo2 to generate about 700,000 candidate sequences, synthesized roughly 285–302 designs, tested about 300 in the lab, and found 16 successful phages.
- Laboratory tests showed a cocktail of the AI‑designed phages overcame resistance in some E. coli strains better than a comparable mix of naturally sourced phages, signaling potential for targeted phage therapy against antibiotic‑resistant infections.
- The authors say they excluded human‑ and animal‑infecting viruses from training data, worked in secure labs, and reported safety consultations, and media reports say the Evo2 model was publicly released, a choice that has intensified debate about openness and risk.
- Biosecurity experts warn governance lags behind the capability to compose full viral genomes, and researchers note the low yield (16 viable from hundreds built) and the much greater technical challenge of scaling this approach to larger, more complex genomes.