Researchers at Stanford University and the Arc Institute used a genomic artificial intelligence model to design 16 new, viable viruses that do not exist in nature, Tom's Hardware reported. The AI-generated viruses successfully infected E. coli bacteria and reproduced inside them.
Published in the journal Science, the research relied on an AI model called Evo. Scientists trained Evo on over 9 trillion nucleotides spanning 128,000 genetic sequences drawn from millions of plants, animals, microbes, and viruses. Much like large language models learn word combinations, Evo learned how combinations of nucleotides form biological instructions for proteins.
To create the new organisms, researchers trained Evo on the 11 genes and 5,386 nucleotides of the Phi X-174 virus, along with 15,000 of its close relatives. Phi X-174 is a simple virus known to infect only E. coli bacteria. Evo generated 700,000 proposed genomes, which the team narrowed down to 285 for chemical synthesis. Testing yielded 16 viable viruses that functioned similarly to Phi X-174, featuring small mutations in their genetic layout.
Bacterial resistance testing
During a resistance experiment, scientists created three strains of E. coli that had developed immunity to natural Phi X-174. Exposing these resistant strains to cocktails of the AI-designed phages caused the viral populations to evolve, ultimately overcoming resistance in all three bacterial strains.
The team stated that the newly created viruses are harmless to humans, noting that they deliberately omitted genetic data from human pathogens during training. Nevertheless, experts warned that regulatory guardrails remain behind the speed of AI development. Separate AI systems have also shown unpredictable behavior, including an OpenAI agent that recently hacked Hugging Face.
