A.I. Learns to Write in DNA

Scientists at the Arc Institute in Palo Alto have used artificial intelligence to design and create new viruses for the first time, according to a study published Thursday in the journal Science. The researchers trained an A.I. model called Evo on genetic sequences from millions of organisms, roughly nine trillion nucleotides, teaching it to recognize the "grammar" of DNA the way large language models like ChatGPT learn patterns in text. After an initial round of training, they gave Evo a second, focused round on the bacteriophage Phi X-174 and about 15,000 related viruses, then prompted the model to generate new versions. Evo produced 700,000 candidate genomes; researchers synthesized 285 of the most promising as DNA molecules and inserted them into E. coli bacteria. Sixteen produced viable, functioning viruses, some of which multiplied faster than the natural virus they were modeled on.

The viruses created pose no threat to humans since they only infect bacteria, and the researchers deliberately excluded any training data on viruses that infect humans, animals, plants, or fungi, specifically to prevent the model from being able to generate anything hazardous. Outside scientists called the result a meaningful milestone with potential upside for medicine and biotechnology, since viruses are already used as delivery vehicles for gene therapy. But the study also sharpens a warning that's been building in biosecurity circles: the same technique could in principle be pointed at more dangerous targets, such as prompting a model to generate a more transmissible or lethal flu genome. Sources in the piece note that oversight hasn't kept pace. A new NIH policy restricting high-risk life science research, rolled out just last week, does not cover this kind of computational, A.I.-driven genome generation unless it involves a specifically designated "entity of concern," and one biosecurity researcher pointed out there's no established way yet to assess the risk of a virus that has never existed before.

Full article at NYT