Article on # Scientists Used AI to Create...

For the first time, an artificial intelligence system has designed a series of previously unknown viruses that can infect and kill bacteria. Researchers at Stanford University and the Arc Institute created 16 functional bacteriophages using the AI models Evo 1 and Evo 2, with results published this week in the journal Science. The work points to a new route for treating antibiotic-resistant infections - and raises fresh questions about whether oversight can keep pace with the technology. Bacteriophages are viruses that infect only bacteria. Their small genomes make them relatively easy to synthesize and manipulate in the lab, and they are already seen as a promising alternative to antibiotics. The Stanford and Arc team used the AI models to generate thousands of completely new genomes, using the well-studied bacteriophage Phi X-174 - which infects E. coli - as a guide. The goal wasn't to reproduce Phi X-174. Instead, the algorithms drew on patterns learned from millions of genomes across all domains of life to produce genetic sequences with the architecture needed to recognize E. coli, insert viral DNA, replicate, and assemble new viral particles. The resulting sequences differed considerably from anything found in nature. From the AI-generated genomes, the researchers selected 300 candidates based on gene organization, regulatory elements, and other biological criteria. They synthesized each genome molecule by molecule and introduced them into E. coli bacteria. Sixteen produced fully functional viruses with previously unpublished sequences, new regulatory elements, and varying genome sizes. Some phages infected bacteria faster; others replicated differently. The team also tested whether AI-designed phages could overcome resistant bacteria. When exposed to E. coli strains that had already developed resistance to natural phages, the AI-generated viruses were able to establish infection. The authors said the results demonstrate "a path toward artificial intelligence-generated phage therapies against rapidly evolving bacterial pathogens." The study demonstrates how AI models can generate functional biological outputs, a topic explored in AI for Science & Research.

Biosecurity concerns

The same capability that could yield new treatments also raises the possibility of misuse. Moritz Hanke, a researcher at the Johns Hopkins Center for Health Security, told The New York Times there is "a huge disconnect" between how fast the technology is advancing and the development of effective regulation. The concern isn't new. Three years ago, the Rand Corporation warned that advanced AI systems could refine the planning and execution of biological weapons attacks. As the models improve, that risk is likely to grow faster than governments' ability to oversee it.

Why this matters for scientists and researchers

For researchers working on phage therapy or antimicrobial resistance, the study offers a concrete method for generating novel phages on demand - and a reminder that AI models trained on genomic data can produce functional biological agents, not just predictions. Microbiologists looking to apply AI in their own work can follow the AI Learning Path for Microbiologists.