Overview
- The study published in Science on Thursday, Aug. 6, 2026, reports that Stanford-led genome language models called Evo1 and Evo2 generated roughly 700,000 candidate sequences, researchers synthesized about 300 designs and 16 produced viable bacteriophages.
- The models were fine-tuned on 14,266 Microviridae genomes so they could learn evolutionary constraints and design complete phage genomes rather than just individual genes.
- The designed phages infect Escherichia coli only because the team excluded human-pathogen data from training, and several AI-made variants worked together as a cocktail to overcome bacterial resistance to wild-type ΦX174.
- The work shows AI can create coadapted genomic contexts that enable novel functions, such as a truncated protein that became functional in an AI-designed genome, but the pipeline had a low hit rate that required extensive filtering and lab testing.
- An accompanying editorial from Johns Hopkins and multiple commentators warn current rules lag this capability and urge mandatory improvements to synthetic-DNA screening, legal duties for providers, and new oversight as researchers develop the platform further.