A genome language model designed working viruses from scratch

Brian Hie’s group at Stanford used Evo 2, a generative model trained on genomes, to write new bacteriophage sequences rather than to predict existing ones. Of nearly 300 designs synthesised and tested, 16 infected and killed E. coli, and some showed higher fitness than ΦX174, the natural phage they were modelled on. A cocktail of the 16 defeated E. coli strains that had already evolved resistance to ΦX174. The work, first-authored by Samuel King, appears in Science (DOI 10.1126/science.aec2657).

The result belongs on this beat for a reason that has little to do with antibiotics. A model that generates a functioning genome is not assisting biology, it is writing it — the same shift that moves a language model from autocomplete to authorship, applied to the substrate a body is made of.

The dual-use question follows immediately and the paper does not settle it. ΦX174 infects bacteria, not people, and the authors confined themselves to it. The capability demonstrated is generative design of a viral genome; the restraint is a choice of target.