AI Writes Functional Complete Bacteriophage Genomes, Bringing a New Tool for Drug-Resistant Infections—and Regulatory Questions

As antibiotics gradually lose their effectiveness, bacteriophage therapy offers a different path: instead of using drugs to suppress bacteria, it employs viruses that specifically infect and destroy them. Researchers are now advancing artificial intelligence from designing individual proteins to designing entire viral genomes, so the search for new bacteriophages no longer has to be entirely limited to samples already found in nature. This capability could expand the toolkit for combating drug-resistant infections, while also presenting biosafety rules with new detection challenges.

Using the ΦX174 bacteriophage, which infects only bacteria, as a template, the research team further fine-tuned the Evo genome model on 14,466 microviridae sequences. Designing a complete genome is far more difficult than generating a single gene: ΦX174 has approximately 5,386 nucleotides and 11 overlapping genes, and the model must simultaneously preserve functions including replication, packaging, regulation, and host recognition. A change in any one location could affect multiple components.

Of the 285 designs sent to the laboratory, the researchers confirmed that 16 bacteriophages were functional. They could replicate in nonpathogenic *Escherichia coli* strain C and the closely related strain W, but did not grow in six other tested strains; during the design stage, the team retained spike protein features associated with host range to limit which organisms could be infected. Some of the resulting bacteriophages differed from the closest natural genome by hundreds of mutations, and one could even constitute a new species under some classification thresholds.

A test more closely aligned with medical needs was whether they could keep pace with bacterial evolution. The team first cultured three strains resistant to natural ΦX174, then subjected mixed populations composed of multiple AI-designed bacteriophages to serial passage. These populations overcame the resistance of all three strains within one to five passages, whereas natural ΦX174 could not; the successful bacteriophages had mosaic genomes formed through recombination among multiple designs, indicating that AI-generated diversity can provide a broader starting point for subsequent evolution.

However, this is not yet a product capable of treating drug-resistant infections. The study was conducted only in a small number of laboratory *E. coli* strains, and overcoming resistance still depended on passage and recombination. It has not yet answered questions about bacteriophage safety, immune responses, pharmacokinetics, or stable manufacturing in animals or humans. For now, the results represent complete-genome design and in vitro functional validation, not clinical efficacy.

The researchers deliberately excluded training data from viruses that infect humans and other eukaryotes, used nonpathogenic hosts, and strengthened laboratory containment and waste disposal. These measures reduced the risks of this work, but cannot prove that all future models will observe the same boundaries. The principal investigator also noted that the methods at the time still required considerable expertise, time, and experimental capabilities, and had not yet clearly lowered the barrier to producing dangerous viruses.

The truly difficult regulatory issue arises after a sequence leaves the model and enters the DNA synthesis process. Some suppliers’ biosafety screening remains voluntary, while conventional systems largely rely on similarity to known pathogen sequences; AI-generated genomes that differ more substantially from natural viruses may be harder for such methods to identify. If complete-genome design continues to advance, governance cannot focus solely on model training data. It must also cover order screening, customer verification, functional risk assessment, and laboratory controls to establish an enforceable defense between therapeutic potential and the risk of misuse.