Neural network generates viable virus that does not exist in nature for the first time

Researchers from the Arc Institute in the United States have created a neural network capable of independently designing functional bacteriophages—viruses that infect bacteria. Some of the generated variants had never been seen in nature before. The results were published in the journal Science.

How the neural network worked

At the core of the development is the Evo model, pre‑trained on vast amounts of genetic data. For the specific task, the team, led by Brian Hie, additionally fine‑tuned the algorithm on the genomes of viruses from the Microviridae family.

A well‑known bacteriophage, ΦX174, was used as the base prototype. The AI proposed hundreds of structural variants. Of these, 16 were successfully synthesised in the laboratory and confirmed to be capable of neutralising antibiotic‑resistant strains of Escherichia coli.

Why this matters

Bacteriophages are seen as a promising tool against antimicrobial resistance—one of the most serious challenges in modern medicine. Interest in phage therapy, which was historically developed extensively in the Soviet scientific school, has grown considerably worldwide in recent years.

The new viruses created by AI proved effective against pathogenic bacteria resistant to standard treatments.

Safety concerns

The authors emphasise that the neural network was trained only on data from viruses that infect bacteria. Genomes of humans or pathogens dangerous to humans were not included in the training. Nevertheless, the developers call for the establishment of ethical and regulatory frameworks for the use of AI in biological research to mitigate potential risks.

In brief

The Evo neural network has, for the first time, fully designed viable bacteriophages that do not exist in nature. Sixteen variants were synthesised in the laboratory and successfully destroyed antibiotic‑resistant bacteria. The work opens new possibilities for phage therapy, but scientists stress the need for strict ethical and regulatory safeguards. The study was published in the journal Science.