Scientists have used artificial intelligence to design 16 previously unknown viruses from scratch, marking a major development in synthetic biology and opening new possibilities for fighting antibiotic-resistant bacteria.
Researchers from Stanford University and the Arc Institute used generative AI models called Evo 1 and Evo 2 to generate novel genomes for bacteriophages — viruses that infect bacteria. The resulting viruses were synthesized and tested in a laboratory, where 16 of nearly 300 candidates proved functional.
The research, published in the journal Science, demonstrates that AI can move beyond analyzing biological information and generate genetic sequences capable of producing functioning organisms.
Evo 2 is a biological foundation model designed to work with DNA sequences. Like a language model predicts patterns in words and sentences, Evo learns patterns in genetic sequences and can use those patterns to generate new DNA.
The model was trained on genome sequences spanning different forms of life. The researchers then applied it to bacteriophage design, focusing on ΦX174, a small and well-studied virus that infects E. coli.
Rather than simply copying ΦX174, the researchers used it as a starting point for generating new phage genomes. The objective was to see whether AI could produce genetic sequences that retained the biological functions required for a virus to infect bacteria while becoming substantially different from naturally occurring viruses.
Stanford researchers synthesized nearly 300 AI-generated phages and tested them against E. coli. Sixteen ultimately produced functional bacteriophages capable of infecting the target bacteria and inhibiting their growth.
Despite the alarming headlines surrounding the breakthrough, the viruses created in the experiment were bacteriophages, meaning they target bacteria.
The researchers designed the project around E. coli, rather than viruses that infect humans. Arc Institute says the functional phages demonstrated restricted host ranges, infecting the targeted E. coli strains while showing no growth on several unrelated strains tested in the experiments.
The experiment does not mean scientists created 16 new viruses capable of causing human disease. Instead, it demonstrated that an AI system could generate genetic instructions that ultimately produced functional bacterial viruses.
The researchers also excluded eukaryotic viruses from Evo 2’s training data as a safety measure. Arc Institute says testing indicated that this restriction substantially weakened the model’s ability to model human viruses, while red-team evaluations found its generated sequences for pathogenic viral proteins were effectively random.
The primary motivation was to explore a potential new weapon against antibiotic-resistant bacteria.
Antibiotics have transformed modern medicine, but bacteria can evolve resistance to them. As resistance increases, infections that were once straightforward to treat can become increasingly difficult to manage.
Bacteriophages offer a different approach. They naturally infect bacteria and can destroy bacterial cells, making them an area of interest for researchers developing alternatives or complements to conventional antibiotics.
The problem is that bacteria can also develop resistance to bacteriophages.
Instead of relying exclusively on viruses found in nature, researchers could potentially use generative models to design new bacteriophages capable of targeting bacterial strains that have developed resistance to existing phages.
In the study, combinations of the AI-designed phages were tested against E. coli strains that had developed resistance to the natural ΦX174 phage. The researchers found that the synthetic phages could overcome that resistance and establish infections.
Stanford described the work as a potential path toward a new generation of bacteria-fighting treatments.
The experiment also showed that the AI was not simply reproducing familiar viral sequences.
According to Arc Institute, the 16 functional genomes contained between 67 and 392 novel mutations compared with their closest natural genomes. Thirteen of the genomes contained mutations that researchers could not find in known natural sequences.
One of the synthetic phages also contained a protein component from a distantly related phage. Researchers used cryo-electron microscopy to study how the unusual combination worked.
These findings suggested that Evo could coordinate genetic changes that had not previously appeared together in nature while still producing a functioning virus.
The researchers were not merely asking AI to search through an existing biological library. They were testing whether it could learn enough about biological constraints to create something that nature had not already produced.
The researchers believe the technology could eventually contribute to personalized phage therapies.
Bacterial infections can evolve rapidly, meaning a treatment that works against one strain may become less effective as the bacteria change. AI-generated phages could, in principle, expand the pool of possible bacterial-targeting viruses and help researchers respond more quickly to emerging resistance.
Arc Institute has emphasized that safety was incorporated into the project, including excluding eukaryotic viruses from Evo 2’s training data. The institute also acknowledges that future biological foundation models will require continued work on safety and alignment as their capabilities increase.
