AI Model Evo 2 Designed Viruses That Kill Drug-Resistant Bacteria

A team from Stanford University and the Arc Institute in California has shown that generative artificial intelligence can design a working virus genome from scratch. The Evo 2 model, built by researcher Brian Hie, generated nearly 300 new variants of a bacteriophage, a virus that attacks bacteria, of which 16 turned out to be fully functional and successfully killed strains of E. coli that had become resistant to the natural ancestor virus.

The starting point was PhiX174, one of the simplest known bacteriophages, with a genome of fewer than 6,000 base pairs. Instead of modifying individual genes of this virus, the researchers had Evo 2 generate entire genomes from scratch, in a single pass, with no manual editing of the genetic code. The authors stress that the model did not copy existing sequences but learned the rules governing the evolution of this family of viruses and used them to create variants that never occur in nature.

How Evo 2 works

Evo 2 is a language model trained on millions of DNA sequences, further fine-tuned on roughly 15,000 genomes of viruses from the Microviridae family, to which PhiX174 belongs. Rather than predicting words like conventional language models, Evo 2 predicts the next nucleotides, building a complete viral genome step by step. Of the nearly 300 sequences designed this way, the researchers physically synthesized 285 candidates and tested which ones actually functioned as living viruses.

The design success rate was far from perfect. Only 16 of the synthesized genomes, about 6 percent, produced functional viruses capable of replicating and infecting bacteria. Some of them, according to the authors, performed better than the natural PhiX174 the model had trained on.

A phage cocktail against resistant bacteria

The key result of the experiment does not concern a single virus but a mixture of all 16 designed phages. Bacteria that had previously developed resistance to natural PhiX174 and no longer responded to treatment with a single virus turned out to be defenseless against the cocktail of genetically diverse AI phages. The mechanism resembles combining several antibiotics at once: to survive, a bacterium would have to simultaneously develop resistance to many different, unrelated virus variants, which is far harder than becoming resistant to a single strain.

This is another step up in the complexity that can now be designed with generative AI. This is the first time this kind of technology has been used to design a complete genome, something that can replicate and carry out other functions inside cells, this is a completely new area for us - Brian Hie, Stanford University, creator of the Evo 2 model

The team sees this as a potential answer to the growing problem of antibiotic-resistant bacteria, against which conventional pharmacology has fewer and fewer effective tools. Phage therapy, treating bacterial infections with viruses that specifically target a given strain, has been known for decades, but it has been limited by the fact that natural phages matched to a specific resistant bacterium are often simply unavailable. The ability to design such a virus to order would change that.

Safety experts' concerns

This same scientific result, however, is raising concern among biosecurity experts. If a model can design a functioning genome for a bacteriophage harmless to humans from scratch, the question arises of how far that capability could be pushed toward viruses dangerous to people. The study's authors stress that sequences of viruses that infect humans were deliberately removed from Evo 2's training data as an added safeguard.

It's no longer a question of whether generative AI design of virus genomes will exist, but whether it can be harnessed without enabling serious harm - Thomas Inglesby and Moritz Hanke, Center for Health Security, Johns Hopkins University

Brian Hie nonetheless decided to release Evo 2 as an open, free tool. He argues that naturally occurring pathogens pose a greater threat today than AI-designed viruses, and that open models of this kind can give scientists an edge in defending against natural pandemics rather than leaving the technology solely to large, closed labs.

What's next for phage therapy

Clinical use is still a long way off. The study was limited to a simple, well-understood laboratory model, E. coli and the PhiX174 virus, rather than bacteria responsible for dangerous hospital infections, such as resistant strains of Klebsiella or Pseudomonas. The authors say the next step will be testing the approach on more complex, clinically relevant phages and on bacterial strains that cause real infections in patients.

For Polish hospitals and patients, where infections from multidrug-resistant bacteria are a growing problem in intensive care units, the practical significance of this work will only emerge once similar methods can be applied to the pathogens actually responsible for hard-to-treat hospital infections. For now, the experiment is more a proof of concept than a ready therapy.

The work fits into a broader trend of using AI models in synthetic biology, alongside tools for protein design and molecular structure prediction. Unlike those projects, creating entire, replication-capable organisms pushes the boundary of what generative AI can do in living matter, and it is precisely that boundary that raises the most questions among biosecurity researchers.