AI-generated genomes yielded viable phages with substantial evolutionary novelty (www.biorxiv.org)

Researchers used genome-scale language models (Evo 1 and Evo 2) to generate whole bacteriophage genomes and experimentally validated that the approach can produce functioning viruses. Using the lytic phage ΦX174 as a design template, the models produced full-length genomes with realistic gene architectures and predicted host tropism; laboratory testing yielded 16 viable, AI-generated phages that show substantial evolutionary novelty. Cryo-electron microscopy revealed one designed phage incorporated a DNA-packaging protein that is evolutionarily distant from ΦX174’s native protein, and several generated phages outcompeted ΦX174 in growth competitions and exhibited faster lysis kinetics. The work is significant because it demonstrates, for the first time, that generative genome language models can design entire genomes that produce living, fit organisms — not just single proteins — opening a new frontier for synthetic biology and evolutionary engineering. Practically, a cocktail of the generated phages rapidly overcame ΦX174 resistance in three E. coli strains, highlighting immediate potential for designing bespoke phage therapies against fast-evolving bacterial pathogens. More broadly, the study provides a blueprint for genome-scale generative design, suggests new routes to explore evolutionary novelty, and raises important opportunities and challenges for safe deployment and regulation of AI-designed living systems.