Scientists have successfully designed and created the first functional viruses using artificial intelligence, a development that promises new avenues for treating bacterial infections while simultaneously introducing significant biosecurity questions. The research, published in the journal Science, details how AI models were used to generate complete viral genomes, which were then synthesized and tested in the laboratory.
The viruses created are bacteriophages, which specifically infect bacteria and do not harm human, animal, or plant cells. This specificity makes them a promising tool in the fight against antibiotic resistance, a growing global health crisis. In laboratory tests, a mixture of these AI-designed phages effectively killed E. coli bacteria that had developed resistance to natural bacteriophages. Some of the AI-generated phages even outperformed a natural phage, phiX174, in their ability to kill bacteria.
The research team, led by Brian Hie, an assistant professor at Stanford University and an investigator at the Arc Institute, utilized two genome language models, Evo 1 and Evo 2. These models were trained on a vast dataset of over 2 million bacteriophage genomes, enabling them to learn patterns in DNA sequences akin to how large language models learn human language. The AI was tasked with generating entirely new genomic sequences for bacteriophages targeting Escherichia coli C.
The process involved generating thousands of potential viral genomes. Researchers then filtered these designs, selecting approximately 300 high-potential sequences for laboratory synthesis. Of these, 285 were successfully assembled in E. coli cells, and 16 yielded viable bacteriophages that could reproduce and kill bacteria. Notably, these AI-generated phages often differed from known natural phages, featuring novel gene combinations, regulatory elements, and variable genome lengths.
The ability to rapidly design and tune viruses for specific bacterial targets, especially to overcome existing resistance, could "transform phage therapy," according to the researchers. Phage therapy, using viruses to combat bacterial infections, has long been explored as an alternative to antibiotics, but finding suitable phages has been a slow and challenging process. AI offers a way to accelerate this discovery and potentially design phages with enhanced efficacy or novel properties.
However, this scientific advancement also brings forth significant safety and security considerations. Experts have voiced concerns about the "dual-use" nature of this technology, where the same capabilities used for beneficial medical research could potentially be weaponized. Thomas Inglesby and Moritz Hanke of Johns Hopkins University, in a commentary accompanying the study, highlighted these issues. They called for legally mandated screening of synthetic genetic material orders to identify potentially dangerous sequences, suggesting that voluntary safeguards are insufficient.
The researchers themselves acknowledged these concerns, noting the importance of "biosafety, biocontainment, and biosecurity considerations." They intentionally excluded genetic data from viruses that infect plants, humans, or animals from the AI's training set to mitigate the risk of designing dangerous pathogens. Despite these precautions, the ability to "compose viral genomes using generative AI now exists; the governance to safely steer it does not," according to Johns Hopkins health security experts.
The current study focused on bacteriophages targeting E. coli, a non-pathogenic bacterium, and used well-studied viruses and laboratory strains. Experts suggest that designing larger or more complex viruses, reliably controlling their host range, and ensuring safety and effectiveness in clinical settings remain significant future challenges. The development underscores the need for robust ethical guidelines and regulatory frameworks to manage the growing power of generative AI in biological research.