Overview
- A team from Stanford and the Arc Institute reported in Science on Thursday that genome-language models called Evo generated hundreds of thousands of candidate sequences and that 16 AI-designed genomes produced viable bacteriophages in lab tests.
- The researchers trained Evo on very large DNA corpora but excluded human-pathogen data, and they synthesized roughly 285–302 selected candidates from about 700,000 proposals before finding 16 that were functional.
- Several of the AI-made phages multiplied faster than their natural template and a cocktail of the new phages overcame resistance in E. coli strains, showing potential for faster, adaptable phage therapies against antibiotic-resistant bacteria.
- Biosecurity and biosafety experts, including authors of an accompanying Science commentary from Johns Hopkins, warned that AI-designed genomes can evade current sequence-screening tools and urged mandatory screening by nucleic-acid providers plus new governance measures.
- The study also highlights limits and safeguards: the team focused on small, tractable phage genomes, required chemical DNA synthesis and lab validation, and experts say those technical steps slow misuse but do not remove the need for updated policies and screening technology.