AI designs working viruses from scratch; researchers flag biosecurity risks

Using AI language models called Evo that they had previously built, Samuel King and colleagues developed a computational approach to design functional bacteriophage genomes from scratch. Their method combined these genomic language models with computational biology and experimental validation. Bacteriophages—viruses that infect bacteria—have substantial utility as biotechnologies and therapeutic relevance as treatments for bacterial infections, making them an ideal test case for whole-genome design.

The researchers used the well-studied ΦX174 bacteriophage as a model system. They computationally designed hundreds of candidate genomes and then tested them experimentally in the laboratory. This screening process identified 16 functional phages—phages that actually worked and could infect and replicate in bacterial cells. Notably, the genetic sequences and structures of these 16 engineered phages differed substantially from one another, showing that the generative approach could produce diverse solutions rather than converging on a single design. Some of the engineered phages performed comparably to naturally occurring relatives in terms of functionality.

The most significant finding was that some combinations of the new engineered phages were able to overcome resistance in two strains of E. coli that had evolved resistance to ΦX174-like phages. This demonstrates a key potential advantage: AI-guided design could eventually enable the creation of more durable phage-based therapies that remain effective even as bacteria develop resistance to existing treatments. This addresses a longstanding clinical challenge in phage therapy development.

However, King and colleagues emphasize that this same capability introduces important biosafety and biosecurity concerns. The ability to design and synthesize complete, functional AI-generated viral genomes means that sensitive viral sequences could potentially be engineered by malicious actors. The researchers argue that existing safety frameworks should be adapted to generative genomics, and propose that model-level protections—such as excluding sensitive viral sequences from the training data used to build AI models—may provide an additional layer of risk mitigation. They call for groups conducting future whole-genome design work to consult both safety and security professionals throughout the project lifecycle. Commentators Inglesby and Hanke frame the challenge as a critical governance question: the technology for generative viral genome design now exists, so the focus must shift to building robust oversight that allows therapeutic benefits while preventing misuse.