AI genome models design new viruses that infect bacteria

Large genome models apply the same machine-learning approach as large language models, but to DNA sequences instead of text. They are trained to predict the next base in a sequence by learning from millions of existing genomes. Because DNA uses only four letters (A, T, C, and G) but varies widely in how strictly the next base is constrained—some regions tolerate any base, others require precise sequences—the models must learn biological context that humans have not yet fully mapped. The researchers trained Evo 1 and Evo 2 on over 2 million bases of bacteriophage sequences and then fine-tuned them on the Microviridae family, to which ΦX174 belongs.

The Stanford team discovered that the AI could generate viable viral genomes by working within carefully set constraints: they filtered outputs to retain only those with spike proteins at least 60 percent identical to the original, between 4,000 and 6,000 bases in length, and free of runs of identical bases longer than 10. This filtering reduced 302 candidate sequences to 285 worth synthesizing and testing. Notably, the AI proved more effective than random mutation at producing functional variants; while random mutation has roughly a 20 percent chance of inactivating a virus per amino acid change, the AI generated viable viruses with over 50 amino acid alterations—a feat that random mutation would render essentially impossible. This suggests the AI learned implicit rules about viral fitness that allow it to explore sequence space more intelligently than chance.