AI generated its first working genome: a tiny bacteria killer (www.sciencenews.org)

Researchers report that two AI sequence models, Evo 1 and Evo 2, generated complete, working bacteriophage genomes that produced viable viruses in the lab. Trained on billions of nucleotide pairs (A, C, G, T) from phage genomes and guided by the well-studied ΦX174 genome, the models proposed roughly 300 candidate genomes; 16 yielded phages that infected Escherichia coli in dishes. Some AI-designed phages lysed E. coli faster than ΦX174, and mixtures of AI phages rapidly overcame three ΦX174‑resistant E. coli strains. The paper was posted to bioRxiv and has not yet been peer reviewed. This is notable as the first demonstration of AI designing entire genomes (not just single genes) that function in vitro, with direct implications for phage therapy against antibiotic-resistant bacteria: AI could quickly generate tailored phages to match a patient’s infecting strain. Technically, the work shows sequence-modeling approaches can capture multigene interactions well enough to produce viable organisms at small genome scales, but scaling to complex genomes (the human genome is orders of magnitude larger) remains a major challenge. The team limited risk by excluding human pathogens from training data and using non‑human-infecting phages, but experts emphasize strict controls and extensive safety testing before any therapeutic use. Potential downstream uses include adaptive phage cocktails, microbial manufacturing, and tools to interpret and design larger genomes.