AI Writes Complete Bacteriophage Genomes: 16 Designs Survive, Bringing Biosafety to the Fore

Antibiotic resistance is forcing scientists to seek new tools for killing bacteria, and bacteriophages, which specifically infect bacteria, have long been important candidates. Now, researchers have gone a step further, using artificial intelligence not merely to modify a few segments of DNA, but to write complete viral genomes and produce bacteriophages in the laboratory that can reproduce, infect, and kill E. coli.

The study used the Evo family of genomic models. According to the research team, the base model learned from approximately 2 million bacteriophage sequences and was then fine-tuned on 14,466 Microviridae sequences. The researchers selected 285 candidate genomes generated by the model for synthesis, and 16 ultimately formed viable bacteriophages. The success rate of less than 10% also shows that the model remains some distance from becoming a reliable, predictable viral-engineering tool.

The experiments used the natural bacteriophage ΦX174 as a design reference. Some AI-generated bacteriophages could infect and kill nonpathogenic laboratory E. coli C or the closely related E. coli W. Their host range remained quite limited, with no evidence of an ability to infect arbitrarily across species. This is regarded as among the first cases in which AI wrote complete viral genomes that were then synthesized into bacteriophages capable of reproduction. Its significance lies primarily in demonstrating the feasibility of biological design at the genomic scale, rather than delivering a medical product ready for immediate use.

The research team also combined the generated bacteriophages into cocktails and tested them against three E. coli strains that had evolved experimentally to resist ΦX174. These combinations overcame the bacteria’s resistance within one to five passages, whereas ΦX174 alone failed. However, the bacteria in this experiment were resistant to a natural bacteriophage, not to antibiotics. Describing the results directly as demonstrating the ability to eliminate “antibiotic-resistant bacteria” would therefore go beyond the experimental evidence. For use against clinical infections, their efficacy against pathogenic bacteria, stability, immune responses, and in vivo safety would still need to be demonstrated.

Another boundary of the research stems from ΦX174 itself: its genome is unusually short and its structure relatively simple. The ability to generate functional designs for viruses of this kind does not mean the model has mastered large or structurally complex viruses, nor can it be inferred that the model can safely design viruses that infect humans, animals, or plants. The researchers said that eukaryotic viral sequences were excluded during training, and the experiments used only nonpathogenic laboratory bacteria.

However, when AI can generate complete and functional viral genomes, the risks no longer exist only in the petri dish. Biosafety experts argue that safeguards must simultaneously cover model access, research review, screening of DNA synthesis orders, and laboratory containment measures. For now, this result is more like a technical milestone and a stress test for governance: it expands the design space for bacteriophage therapy while also forcing regulators to consider in advance which capabilities should be opened, tracked, and restricted as “writing genomes” gradually becomes a scalable computational task.