Scientists have spent decades studying the genetic sequences that allow viruses to infect, replicate, and evolve. Stanford researchers have now used artificial intelligence to go a step further, designing complete viral genomes that did not previously exist and testing whether they could work.
The team created 16 functional bacteriophages, viruses that infect bacteria, using genome language models developed with the Arc Institute. And the result marks the first reported demonstration that generative AI can design entire viral genomes that produce viable viruses.
The research further gives scientists a new way to explore bacteriophages that could eventually help fight bacterial infections. It also gives biosecurity researchers a concrete example of AI moving from predicting biological sequences to generating functional ones.
AI Designed The Viral Genomes
The researchers used Evo 1 and Evo 2, genome language models that learn patterns in DNA sequences in a way that is broadly comparable to how language models learn patterns in text. For this experiment, the team focused on ΦX174, a well-studied bacteriophage that infects E. coli.
The models generated large numbers of candidate genomes, and researchers then selected hundreds for laboratory testing. Of 302 designs that were synthesized, 285 were successfully built and assembled, and 16 produced viable bacteriophages capable of infecting E. coli.
Several of the AI-designed phages performed better than the natural ΦX174 in laboratory tests. A mixture of the generated phages also overcame ΦX174 resistance in three E. coli strains, pointing to a possible use in phage therapy, where viruses are used to target bacteria that may be difficult to treat with antibiotics.
The AI models were trained on genetic codes from viruses, bacteria, plants and people. Photo Credit: Getty Images
The Medical Potential Is Significant
The research could matter as antibiotic resistance continues to make some bacterial infections harder to treat. Designing phages with AI could eventually help researchers produce candidates tailored to specific bacterial strains and respond more quickly as bacteria develop resistance.
The Stanford team describes the work as a foundation for designing useful biological systems at the genome level. Stanford has also highlighted the 16 bacteriophages as an example of how DNA models are moving beyond analysis and into biological design.
While the technology is still far from producing a ready-to-use treatment, these experiments were conducted on bacteriophages that target bacteria and further testing would be required before any medical application.
It Also Creates A New Biosecurity Problem
The same capability that makes the research valuable has also raised concerns among biosecurity experts. The researchers deliberately limited the work to bacteriophages and excluded viruses that infect humans, animals, and plants from the relevant training data.
Still, experts from the Johns Hopkins Center for Health Security warned in an accompanying Science commentary that the ability to generate functional viral genomes creates urgent biosafety and biosecurity questions. Separate research from the same group has argued that biological AI systems require risk reviews before development, rather than relying only on safeguards after a model has already been built.
This makes Stanford’s experiment important beyond the 16 viruses themselves, as AI has now demonstrated that it can help produce complete viral genomes that function in the real world.
Now the next challenge is making sure the [AI] systems built to accelerate biological discovery develop alongside the safety measures needed to control their misuse.
