For the first time in history, generative artificial intelligence has designed a complete bacteriophage genome—a virus that infects bacterial cells—that proved to be viable. This is not just a laboratory curiosity, but a fundamental shift in our ability to engineer biological systems.

For a long time, generative models in biology were limited to working at the level of individual genes or proteins. Creating an entire genome was considered an unattainable milestone: even the simplest genomes are extremely complex, and a single mutation can render the entire construct non-viable. However, a research team led by Brian Hie and first author Samuel King managed to overcome this barrier.

Sampling from thousands of genome variants

In their work, the scientists used the Evo 1 and Evo 2 models, trained on DNA datasets from millions of genomes across all domains of life. By analogy with text language models, these systems learn the evolutionary constraints that DNA sequences obey in the wild. The natural phage ΦX174 served as the basis, acting as a template for design with specified specificity to the bacterium Escherichia coli C.

The models generated thousands of genome variants. From this pool, nearly 300 samples were chemically synthesized and tested under laboratory conditions. Sixteen variants proved viable—and none of them matches known natural phages.

The resulting phages fully retained specificity for the target host, but diverged significantly from natural counterparts in several key parameters: new mutations, divergent genes, regulatory elements, and different genome lengths. Cryo-electron microscopy confirmed an important fact—one of the viruses incorporated a DNA packaging protein into its capsid, borrowed from an evolutionarily distant phage. The model was by no means limited to a simple search of mutations around the base template.

Overcoming bacterial resistance

The main result of the work concerns the comparison with the wild type, rather than the mere fact of generation. Bacterial resistance to phages remains a central problem in phage therapy—a field now viewed as the primary response to the antibiotic crisis.

The scientists tested the generated viruses on special strains of E. coli that had developed resistance to ΦX174. A mixture of the designed phages overcame this defense very quickly. At the same time, a comparable mixture of natural phages related to ΦX174 failed to accomplish the same task.

This opens up an important applied perspective: in the future, phage preparations could be generated for a specific rapidly evolving pathogen, instead of a lengthy search for suitable variants in natural collections.

My view on this breakthrough: we are witnessing a transition from editing to generative design of biological functions at the level of the entire genome. This is not just a step forward for phage therapy—it is a prelude to creating complex synthetic systems where AI becomes a full-fledged partner in life engineering. Questions of ethics and biosecurity now arise with renewed urgency, but the potential for medicine and biotechnology is truly colossal.