Artificial intelligence (AI) generated new viral genes in the same way as ChatGPT. The virus is a bacteriophage that infects bacteria. Because it is a virus long used to block pathogens, it is expected to help curb antibiotic-resistant bacteria that threaten humans. At the same time, concerns were raised that, if misused, it could become a deadly biological weapon, calling for safeguards.
A research team led by Professor Brian Hie of Stanford University said on the 6th (local time) in the journal Science that they "designed the genome of bacteriophages using a 'genome language model' corresponding to the large language models (LLMs) that underpin Conversational AI."
Even Escherichia coli resistant to existing viruses was infected by the AI-made bacteriophage DNA. Bacteriophages were first discovered in 1917 by French microbiologist Félix d'Hérelle. The name means "to eat bacteria" in Greek. They are called phages for short.
◇ Operates like an AI chatbot
The Stanford team developed an AI model named Evo that analyzes and generates DNA base sequences in the same way as OpenAI's ChatGPT. ChatGPT answers people's questions by linking words that are likely to follow one another. It is possible because it was trained for years on a massive corpus collected from the internet and books in advance.
Evo similarly prelearned genome data of bacteriophages. A genome refers to the entirety of genes that govern all life phenomena. DNA consists of four bases: adenine (A), guanine (G), cytosine (C), and thymine (T). Living things synthesize proteins that control all life activities according to the order of these bases. Genome data refers to these base sequences.
The Stanford team trained generative AI models Evo1 and Evo2 on the 11 genes of bacteriophage Φ (phi) X-174, which has been extensively studied, and the genetic information of about 15,000 closely related viruses. The AI then generated thousands of new phage genome blueprints. The researchers selected 285 of them and, in the lab, actually ligated bases to construct phage genomes.
Humans have as many as 3 billion bases in their genomes, but viruses have only thousands, making them easy to synthesize. However, whereas viral genome synthesis so far has been based on known information, this time AI newly generated them on its own. In particular, only 16 bacteriophages were initially viable in E. coli, but when their genomes were mixed, two types of E. coli that were resistant to existing bacteriophages were also infected.
Professor Patrick Cai of the University of Manchester said, "This study is an important milestone in synthetic genomics," and noted, "For the first time, we have witnessed AI go beyond predicting biological base sequences to generating fully functional genomes that actually work in the lab."
◇ Infects resistant strains, a breakthrough for treating superbugs
The researchers said AI could make bacteriophages that can become treatments for bacterial infections. Bacteriophage DNA uses bacterial replication enzymes to replicate the virus. Viruses that fill the inside of the bacterium tear the cell membrane and spill out. In the process, the bacterium dies.
Because bacteriophages can infect and kill bacteria, they were spotlighted as therapeutics for pathogens immediately after their discovery. In the 1920s and 1930s, they were used to treat various bacterial diseases such as dysentery and sepsis. But as penicillin antibiotics became widely used in the 1940s, they were gradually forgotten. The revival of bacteriophage therapy is due to the misuse and overuse of antibiotics.
As antibiotics were used indiscriminately not only in people but also in livestock, antibiotics spread into nature through excrement and other pathways. Bacteria then evolved the ability to withstand antibiotics. Antibiotic-resistant bacteria, such as methicillin-resistant Staphylococcus aureus (MRSA), spread worldwide. They are superbugs against which no drugs work. AI is expected to be able to create phages that also target such superbugs.
The problem is that the same approach could be used to create viruses that become deadly weapons. In a commentary published the same day in Science, Thomas Inglesby of the Johns Hopkins Center for Health Security and Moritz Hanke warned that one could ask Generative AI to make influenza genomes with stronger transmissibility or higher fatality.
The Stanford team, mindful of these concerns, said they did not provide genetic information for viruses that infect plants, animals, or humans during AI training. The researchers further said, "This study shows the need for expert oversight and robust safeguards across the entire AI genome design process," and added, "Researchers conducting future genome design studies should consult safety and security experts throughout the entire project."
The commentary authors said, "These authors are approaching biosafety and biosecurity issues much more cautiously than most researchers developing powerful biological AI models," but added, "The capability to assemble viral genomes using Generative AI already exists, yet governance to manage it safely has not been established."
References
Science (2026), DOI: https://doi.org/10.1126/science.aec2657
bioRxiv (2025), DOI: https://doi.org/10.1101/2025.09.12.675911