AI creates 16 functioning viruses from scratch in genomic study

Researchers at Stanford University and the Arc Institute trained a genomic AI model called Evo to learn the statistical patterns embedded in biological DNA. The model was trained on over 9 trillion nucleotides—the individual building blocks of DNA—drawn from 128,000 genetic sequences spanning millions of animals, plants, microbes, and viruses. This training allowed Evo to learn which combinations of nucleotides tend to produce biologically meaningful sequences, much as a language model learns which word combinations make sense.

To test whether the model could design functional new organisms, the researchers focused on the Phi X-174 virus, a simple organism with only 11 genes and 5,386 nucleotides that has been studied for over a century and naturally infects only E. coli bacteria. After training Evo specifically on Phi X-174 and about 15,000 of its closest relatives, the model proposed 700,000 potential new viral genomes. Scientists then narrowed this down to 285 candidates and chemically synthesized them. Of those 285, 16 successfully assembled into functioning viruses capable of infecting E. coli bacteria and reproducing. The new viruses had not existed in nature before. The researchers noted that the new viruses are quite similar to their source organism—the building blocks remain the same, with variations in how they are arranged.

In one experiment demonstrating practical utility, researchers first generated three strains of E. coli that had evolved resistance to the natural Phi X-174 virus. They then exposed those resistant bacteria to mixtures ("cocktails") of the AI-designed phage variants. The AI-designed phages evolved during the experiment and ultimately overcame resistance in all three bacterial strains, proving that the synthetic viruses could adapt and function in real biological scenarios.

The findings appear in the journal Science. The researchers emphasized that all new viruses were completely harmless to humans and that they deliberately avoided training the AI on any genetic data from human pathogens. However, the study inadvertently demonstrates that highly dangerous applications are theoretically possible. Experts have raised concerns that such research is advancing faster than the regulatory and safety guardrails needed to govern it. The researchers themselves noted that creating similar results in human pathogens would be "a different ballgame."