Researchers from Stanford University and the Arc Institute used the genomic language model Evo to generate approximately 700,000 candidate viral genome sequences, from which they synthesized 285, and ultimately 16 of them became bacteriophages capable of replication, infection, and killing E. coli in the laboratory.
The related study was published in the journal Science on August 6, marking a leap in AI-generated biology from designing individual proteins or genes to de novo design of entire functional viral genomes. The model's output were not finished viruses, but rather DNA sequence strings, which researchers screened, synthesized, and tested in E. coli. Most experiments showed no reaction, while some plates displayed clear spots indicating that bacteria had been lysed by new viruses.
The research team chose the bacteriophage Phi X-174 as a template, as it only infects bacteria, has a short genome, and a rich research history, making it suitable for verifying the entire design process. Evo first read about 9 trillion nucleotides, then specifically trained on around 15,000 small virus genomes similar to Phi X-174. The 16 surviving viruses were not barely functioning; some even outperformed natural Phi X-174 in competition experiments and lysis speed. Researchers also mixed AI-designed bacteriophages through generations, and after just 1 to 5 rounds, the viruses broke through three resistance barriers of E. coli, opening up a new path for bacteriophage therapy against drug-resistant bacteria.
Safety Valves and Unknown Risks
However, "not found in nature" does not mean "never seen in biology." Researchers at the University of Oxford pointed out that these viruses are still very similar to natural species, relying on the same biological mechanisms. An independent analysis in June also found that the sequences favored by Evo were usually closer to natural viruses, with the model mainly improving search efficiency within the range of natural variation. The research team removed viruses that infect humans and their close relatives when training Evo, and used non-pathogenic E. coli in the experiment, claiming this version could not generate human virus sequences.
But what if a different training dataset is used? The answer remains unknown. Researchers at Johns Hopkins University have expressed concerns that future genome models might be used to rewrite the transmissibility or lethality of viruses like influenza. In late July, the U.S. government announced a policy on high-risk life science research, prohibiting federally funded experiments that make biological agents more dangerous. However, there is currently no consensus on how reviewers can assess the risks of an AI-generated genome that has never appeared in nature. Although there are multiple steps between generating a sequence and creating a dangerous pathogen, including synthesis, assembly, and host adaptation, this research undoubtedly sounds a proactive alarm for biosecurity governance.
