AI designs 16 functional viruses that kill drug-resistant bacteria

Researchers led by Stanford University have used generative artificial intelligence to design 16 viable bacteriophages — viruses that infect bacteria — that do not exist in nature. The study, published in Science, marks the first reported instance of AI generating complete, functional viral genomes.

The team trained specialized DNA language models on roughly 15,000 genomes from the Microviridae family of small bacterial viruses. After fine-tuning, the models produced thousands of theoretical genome blueprints based on the natural phage PhiX174.

From these digital designs, the scientists synthesized 285 candidate genomes as physical DNA sequences and introduced them into E. coli host cells. Sixteen of the candidates successfully assembled into reproducing viruses that infected and killed the target bacteria.

In follow-up tests, a cocktail of the synthetic phages eliminated bacterial strains that had evolved resistance to the natural PhiX174 virus. The authors state the results show generative AI can capture the evolutionary design space of viral genomes with sufficient fidelity to produce viable organisms.

The research demonstrates a potential pathway for rapidly designing customized viruses to target specific drug-resistant pathogens, a major challenge in modern medicine. Current phage therapy relies on isolating naturally occurring viruses, a slow and limited process.

Safety measures included removing human-infecting viruses from the AI training data. The researchers and an accompanying Perspective by Thomas Inglesby and Moritz S. Hanke emphasize that clinical use will require extensive animal and human testing, and that biosecurity oversight must keep pace with advancing genome-design tools.

The work remains at the experimental stage; the 16 viable phages represent a small fraction of the 285 designs tested, and their efficacy in complex biological environments is unproven. Further research is needed to assess stability, host range, and immune responses.

The study was led by Samuel H. King and colleagues, with the primary paper appearing in Science under DOI 10.1126/science.aec2657 and the Perspective under DOI 10.1126/science.aej8512.