Scientists at Stanford University have used an artificial intelligence model trained on millions of genetic sequences to design new viral genomes, with 16 of the resulting viruses proving capable of infecting E. coli in laboratory tests.
The study, published in Science, found that the AI model, called Evo, could learn patterns in DNA from millions of genomes and use those patterns to generate viral sequences that were different from naturally occurring ones.
Researchers subsequently synthesized AI-generated DNA sequences and tested them in the laboratory. Sixteen of the resulting bacteriophages were functional and successfully infected E. coli, with some also overcoming resistance mechanisms used by the bacteria.
The researchers did not include human pathogen data in the model’s training material, and the viruses generated in the experiment were not capable of infecting humans.
The findings could have practical applications in medicine, particularly for developing treatments against antibiotic-resistant bacteria. Researchers could potentially use similar AI systems to design bacteriophages tailored to attack specific bacterial infections.
At the same time, the ability of AI to generate functional biological sequences has raised concerns among biosecurity experts about potential misuse. Scientists have warned that increasingly capable biological AI could eventually make it easier to design harmful pathogens, creating challenges for regulators and researchers.
In a commentary published alongside the study, biosecurity experts Thomas Inglesby and Moritz Hanke of the Johns Hopkins Center for Health Security called for stronger safeguards around AI-assisted biological research. They argued that similar technologies should face strict restrictions when applied to pathogens affecting humans, animals or crops.
Other experts have cautioned against overstating the immediate danger. Tom Ellis, a professor of synthetic genome engineering at Imperial College London, said the viruses produced in the study have relatively simple genomes and that creating a virus capable of causing significant harm to humans would be substantially more difficult.
The study nevertheless demonstrates a significant advance in AI-assisted biology: a generative model was able to move beyond analysing existing genetic material and produce novel viral genomes that worked in real laboratory experiments.
