Researchers at Stanford University and the Arc Institute have used an artificial intelligence model called Evo 2 to design new viruses that do not exist in nature.
Evo 2 works in a similar way to a large language model such as ChatGPT, except that it predicts the next letter in a strand of DNA rather than the next word in a sentence.
DNA is built from just four letters, adenine, thymine, cytosine and guanine, commonly shortened to A, T, C and G.
To learn the patterns behind functional genomes, Evo 2 was trained on roughly nine trillion nucleotides, the individual chemical units that make up DNA, drawn from 50 years of sequencing data across plants, animals, bacteria and fungi.
Designing a virus from a template
The scientists focused the model on bacteriophages, viruses that infect and destroy bacteria rather than human cells.
Starting from a short snippet of an existing bacteriophage called Phi X174, Evo 2 generated roughly 700,000 candidate genome sequences.
The team selected 285 of these designs to synthesise physically in the laboratory.
From code to living virus
Sixteen of the synthesised sequences turned out to be viable, meaning they functioned as real, replicating viruses once introduced into bacterial cells.
Six of those sixteen proved particularly effective at killing bacteria, with some outperforming Phi X174, the natural virus they had been modelled on.
The researchers say the AI-generated viruses pose no threat to humans, because Evo 2 was not trained on viruses that infect people.
Why scientists are excited
Bacteriophages have drawn renewed interest as a possible complement to antibiotics, as antimicrobial resistance contributes to nearly five million deaths worldwide each year, according to the World Health Organization.
The Stanford team says combinations of their AI-designed phages were able to overcome bacterial resistance that had defeated the natural virus alone.
Evo 2 has been made freely available to other researchers, and the scientists behind it say the next step is extending the model to longer and more complex genomes, including small bacterial genomes.
A five-year leap, and a new question
Researchers involved in the work describe it as a marker of how far generative AI has come, moving from having no commercial large language models to designing functional living systems within five years.
The study's authors acknowledge that because Evo 2's code and training data are openly available, the same techniques could, in principle, be misused to work toward more dangerous pathogens.
They argue the real barriers to doing so remain significant, since designing larger, more complex organisms demands both far more data and far greater laboratory resources than a small bacteriophage.
