AI Creates 16 Brand-New Viral Genomes Never Seen in Nature, Scientists Call It a Turning Point

Artificial intelligence has been used to design 16 viral genomes that had never existed in nature, with the resulting bacteriophages able to infect and kill E. coli in laboratory tests and prompting scientists to describe the work as a turning point for AI-driven biological design.

Researchers at Stanford University and the Arc Institute generated about 300 candidate genomes for laboratory testing. Sixteen produced functional viruses, while several showed greater fitness than the natural virus used as the template and overcame bacterial resistance.

The study, published in Science, marks a step from using generative AI to analyse genetic sequences towards using it to design complete genomes that can function inside living cells.

Evo Models Used To Design Bacteriophages

The researchers used Evo 1 and Evo 2, AI models trained to identify patterns in genetic sequences. The models were used to design bacteriophages, viruses that infect bacteria, rather than viruses that infect humans.

The team focused on ΦX174, a well-studied bacteriophage that infects E. coli. The models generated new genomes based on the structure of the natural virus, after which selected designs were chemically synthesised and tested.

Of roughly 300 genomes synthesised in the laboratory, 16 produced viable bacteriophages. The viruses were able to replicate, infect E. coli and kill the bacteria.

Brian Hie, a Stanford University researcher involved in the study, shared that it was the first time generative AI had been used to design a complete genome capable of replication and other cellular functions.

AI-Designed Phages Beat Bacterial Resistance

Several of the AI-designed phages showed greater fitness than the natural ΦX174 virus used as the starting point. Some also infected bacterial strains that had developed resistance to the original phage.

A combination of AI-designed phages overcame resistance in three E. coli strains during laboratory testing. The result could support further research into phage therapy, which uses bacteriophages to target bacteria.

Phage therapy is being studied as a potential treatment for infections that no longer respond to antibiotics. The study does not show that the AI-designed viruses can treat patients, and no clinical testing was reported.

Researchers Limit Work To Bacterial Viruses

The experiment was limited to bacteriophages that infect bacteria. The researchers excluded viruses capable of infecting complex organisms from the relevant training data and carried out the work in a controlled laboratory environment.

The AI-generated phages were designed to target E. coli and were not human viruses. The findings therefore do not demonstrate that generative AI can create a human pathogen.

Biosecurity Concerns Around Novel Genomes

The Science paper was accompanied by a commentary from Thomas Inglesby and colleagues at the Johns Hopkins Center for Health Security. They warned of urgent safety and security questions as generative AI becomes capable of designing viral genomes.

The concern extends beyond the 16 phages created in the experiment. More capable systems could eventually generate biological sequences with properties that are difficult to assess against existing genetic databases.

The researchers' work also raises questions about how DNA synthesis screening should handle sequences with no natural counterpart. Current screening systems can compare orders against databases of known biological threats, but novel sequences require additional methods to assess potential risks.

That does not establish that existing screening systems would fail to identify dangerous AI-generated DNA. It highlights the need to assess how screening methods perform as biological design moves further from naturally occurring sequences.

Evo Marks Shift From Analysis to Design

Evo 2 was developed as a genome-scale model that can analyse and generate DNA sequences from a broad range of organisms. Its developers have compared the approach with language models such as ChatGPT, except that the models work with genetic sequences rather than written language.

The latest experiment tested whether those predictions could produce complete biological systems. Researchers generated candidate genomes, selected designs for synthesis and then tested whether the resulting phages functioned.

Most of the designs did not produce viable viruses. The 16 successful genomes nevertheless showed that AI can generate previously unseen genetic sequences that work inside living cells.

AI Biology Reaches New Experimental Milestone

The research adds to a growing field of generative AI applications in biology, including protein design, enzyme engineering and genome analysis. Much of the technology's earlier use focused on understanding existing biological structures or predicting how they behave.

This study took the process into biological design. The researchers started with genetic patterns learned by AI, produced new complete viral genomes and tested whether those designs could function.

The immediate work remains confined to bacterial viruses and laboratory experiments. Its longer-term significance will depend on how genome-generation models develop and how researchers, laboratories and regulators address the safety questions that come with them. For now, the result is specific: 16 AI-designed viral genomes that had not previously existed in nature produced functional bacteriophages capable of infecting and killing E. coli in laboratory tests.