AI-designed bacteriophages overcome bacterial resistance

Artificial intelligence (AI) has been used to design functional viruses from scratch for the first time.

Previous studies have used AI to design individual genes, proteins and other genetic sequences, but designing a complete functional viral genome has remained a challenge. Evo 1 (see BioNews 1266) and Evo 2 (see BioNews 1331), developed by researchers at the Arc Institute and Stanford University in California, are genomic language models capable of analysing and generating biological sequences. Like the large language models underpinning ChatGPT, genomic language models are trained to predict the next part of a sequence. Instead of learning patterns in written language, they learn patterns in DNA sequences that can be associated with biological function.

'This is a next step in the complexity that's designable by generative AI, this is the first time generative AI has been used to design a complete genome. It's something that can replicate and have other functions inside cells... this was new territory for us,' Dr Brian Hie, senior author of the findings published in Science and assistant professor at Stanford University, told the BBC.

The researchers chose to design bacteriophages related to Phi-X-174, a virus that infects E. coli bacteria and has been studied for almost 100 years. This was a deliberate choice to avoid producing any viruses that could pose a threat to humans, animals, plants or fungi. Specifically, the researchers further trained Evo 1 and Evo 2 on 14,266 genomes from the Microviridae family, which includes Phi-X-174.

The models generated around 700,000 potential genomes for new bacteriophages. From these, 302 candidates were selected for experimental testing, and DNA was synthesised for 285. Of these, 16 produced viable bacteriophages when tested in E. coli cultures. Some of the AI-designed phages also had higher fitness or replicated faster than the natural Phi-X-174 strain they were based on.

A cocktail of the AI-designed bacteriophages was then tested against E. coli that had developed resistance to the natural Phi-X-174 phage. The cocktail was able to overcome this resistance, suggesting that AI-designed phages could potentially be used alongside naturally occurring phages to target bacteria that have developed resistance.

The researchers said the findings could have applications in phage therapy, in which bacteriophages are used to target disease-causing bacteria. However, they also raise concerns about the potential misuse of AI to design biological agents.

'Although this is promising for life sciences applications, it also raises urgent biosafety and biosecurity questions. The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not,' said Professor Thomas Inglesby and Dr Moritz Hanke, from the Centre for Health Security at Johns Hopkins University, Maryland, in an accompanying Science article.

The study does not show that Evo can design any virus. The researchers focused on a small, well-characterised group of bacteriophages and used related viral genomes to further train the models. It therefore remains unclear whether the same approach could be used to design functional genomes for more complex or unfamiliar viruses.

Whether these advances can be safely extended to other viruses is unclear, and the findings have highlighted questions about whether existing scientific policies and regulations are sufficient to oversee AI-assisted genome design.

---

The impact of AI on genomics will be explored at this year's PET Annual Conference, AI and Automation in Fertility and Genomics: Pipeline? Or Pipe Dream?, taking place in central London on Wednesday 9 December 2026.

Find out more and register here.