AI-designed viruses offer new possibilities for phage therapy but raise governance concerns
Researchers at Stanford University used large genome models to generate new bacteriophages capable of infecting E. coli, with potential applications in treating antibiotic-resistant infections. The study also highlights risks and calls for better oversight.
Scientists at Stanford University have applied large genome models, known as Evo 1 and Evo 2, to design complete genomes of viruses that attack bacteria. This marks a step beyond protein design, as the models work directly with DNA sequences and can generate functional viral genomes.
Using the well-studied ΦX174 virus, which infects E. coli and has 11 genes in about 5,400 base pairs, the team fine-tuned the models with additional bacteriophage sequences and prompts. After filtering outputs for viability criteria, they synthesized 285 of 302 proposed viral sequences. Only 16 inhibited bacterial growth: nine from direct AI output and seven after picking up extra mutations.
The most effective viruses closely resembled the original ΦX174. While overall viability was 5.6%, viruses with at least 98% sequence similarity were viable 46% of the time. However, some designed viruses showed unusual features: one lost a protein, another gained a new gene, and one swapped a gene from a distant virus. Mathematical modeling suggested that random mutations causing more than 25 amino acid changes would almost always inactivate the virus, yet a quarter of AI-generated viruses with such changes remained viable.
The researchers also tested a cocktail of the AI-generated viruses against resistant E. coli. It successfully evolved to overcome resistance, while a natural bacteriophage cocktail failed. This suggests potential for phage therapy, though such treatments are not yet widely used.
The study raises biosecurity concerns. Although viruses that infect vertebrates were excluded from training data, the authors warn that others could recreate the process with those sequences included. They call for improved governance of AI and custom DNA order screening. The research appears in Science with DOI 10.1126/science.aec2657.
