AI-Designed Viruses Are Replicating and Killing Bacteria (singularityhub.com)

Researchers led by Stanford's Brian Hie used generative AI models (Evo 1 and Evo 2) to design complete bacteriophage genomes and showed for the first time that AI-written viral genomes can function in living cells. The team fine‑tuned models trained on a massive corpus of genomic data (Evo 2 was trained on roughly 128,000 genomes — about 9.3 trillion base pairs) on phage sequences and structural/protein constraints, then generated thousands of candidate genomes. From 302 synthesized designs they tested, 16 acted like real phages: they infected E. coli, replicated, lysed host cells, and in some cases unexpectedly infected other E. coli strains. Many AI designs retained ~40% sequence similarity to the model phage phiX174 (a small, single‑stranded DNA virus with 11 genes and 7 regulatory regions) while introducing novel sequence variants. This work is significant because it demonstrates that genome-scale language models can internalize multi-gene interactions and produce viable genetic systems, potentially accelerating phage therapy development and synthetic biology (including minimal genomes and engineered microbes). It also raises urgent biosafety and ethical questions: the team employed safeguards (excluding eukaryotic-virus data, using well-characterized lab organisms, and supervised guidance), but the approach could, if misused or scaled to larger genomes, enable harmful enhancements. The paper therefore marks both a technical milestone for AI-driven genome design and a call to expand governance, risk assessment, and containment practices as generative biology advances.