Overview
- The peer‑reviewed study published Thursday in Science reports researchers used genome‑language models to generate roughly 700,000 candidate sequences, synthesized about 300 designs, and validated 16 viable bacteriophages in the lab.
- The engineered phages infect only bacteria, not humans, and the team excluded complex‑organism viruses from training data and worked in secure facilities to limit risk.
- Laboratory tests found a cocktail of the AI‑designed phages could overcome resistance in some E. coli strains where comparable natural phage mixes failed, pointing to potential new treatments for antibiotic‑resistant infections.
- Biosecurity experts including authors of a Science commentary warned the capability raises urgent biosafety and governance questions and urged tighter oversight of AI‑driven genome design.
- U.S. guidance now restricts certain gain‑of‑function wet‑lab work but leaves many purely computational designs unregulated, and scientists note major technical hurdles remain to scale this method to larger, human‑infecting genomes.