Overview
- Researchers at Stanford and the Arc Institute used large genomic language models called Evo 1 and Evo 2 to generate 302 candidate genomes, synthesized 285, and validated 16 synthetic bacteriophages that replicate in E. coli.
- One synthetic phage, Evo‑Φ69, outcompeted the natural reference ΦX174 in lab tests, replicating far more efficiently in direct growth competitions.
- The authors made the models, training and inference code, and model weights publicly available and explicitly warned that the tools could be misused or retrained on other datasets.
- Experts stress practical limits to dangerous misuse because the validated phages were 93–99% similar to the natural virus, the work used a simple, harmless lab phage, and converting designs into complex human pathogens still requires scarce data, specialist labs and skills.
- Policy responses are accelerating: a July 2026 U.S. guideline broadened review to some in‑silico work and scientists are urging tighter synthetic DNA screening, risk‑tiered model access, and traceable genetic markers while calling for clearer international rules.