Artificial intelligence has opened new paths in the fight against antimicrobial resistance by designing viruses that attack bacteria from scratch.
Researchers at Stanford University and the Arc Institute used generative AI models to design complete genomes for bacteriophages, viruses that infect bacteria. In a study published in Science, researchers used Evo 1 and Evo 2 to generate new bacteriophage genomes based on ΦX174, a small phage that infects E. coli.
The researchers emphasize that the results are a limited proof of concept and do not establish that the models can reliably design more complex viruses. They also note that Evo 2 performed poorly when generating proteins from human viruses, which they said prevented “unconstrained or accidental generation of human viral proteins.”
Antimicrobial Resistance A High Priority Crisis
The scientific community has warned that bacterial infections are becoming harder to treat as antibiotic resistance grows. The World Health Organization attributes this trend largely to the misuse and overuse of antibiotics in clinical care and food production.
Scientists have been exploring bacteriophages as an alternative to antibiotics for treating bacterial infections. Bacteriophages are viruses that naturally infect and destroy bacteria. However, their limitation is that certain strains can target only specific bacteria. A bacterium can also develop resistance to phages over time, raising concerns about the limited repository of phages currently identified.
These limitations have driven researchers to seek out a more diverse pool of phages to test against different bacterial strains. The Stanford and Arc Institute study suggests that AI could help researchers expand the pool of available phages by designing new candidates rather than relying solely on those found in nature.
This innovation from Stanford and Arc Institute suggests that AI could add another tool to that effort, generating multiple phages designed to target specific bacteria. Instead of relying solely on phages found in nature, researchers could use AI to design new candidates with different genetic characteristics and then test whether they can target bacteria that existing phages cannot.
How Evo 2 Designed A Virus That Targets Bacteria
Evo 2 is an open-source DNA language model developed by researchers at the Arc Institute and NVIDIA, with collaborators at Stanford University, UC Berkeley and UC San Francisco. The group says they trained the model on 9.3 trillion DNA base pairs from more than 128,000 genomes spanning bacteria, phages, plants, animals and other organisms.
Like a language model that learns patterns in text, Evo 2 learns patterns in DNA sequences. Researchers say the model can predict how genetic sequences function, identify disease-causing mutations and generate new DNA sequences.
When Stanford and Arc researchers used Evo 2 to design bacteriophages, they based the model on a phage that infects E. coli. The model generated thousands of possible sequences and new phage genome designs, which researchers then screened computationally to select candidates for lab testing. Viable candidates were then synthesize as DNA and physically tested on if they could infect and inhibit the growth of E. coli.
Only 16 of the 285 AI-generated designs produced functional phages that infected the target E. coli strains and inhibited their growth, but researchers found that a cocktail of the 16 phages could overcome resistance in E. coli to the natural ΦX174 phage.
Other researchers are also exploring AI models that can predict which phages are most likely to infect specific bacteria. In 2024 the National Institutes of Health has increasingly funded computational and AI-enabled phage research, including a recent $6 million funding opportunity year to develop computational and in silico tools, synthetic or bioengineered phages, and approaches for optimizing phage cocktails.
Researchers Highlight Built-In Safety Measures
The Stanford team says they are already looking further ahead. Lead researcher Brian Hie said future work will explore longer and more complex DNA, including the possibility of designing small bacterial genomes that could produce chemicals, medicines or fuels.
Evo 2 researchers have said the model has limitations and safety measures, including poor performance on human viral proteins. According to researchers, they excluded viruses that infect humans and other complex organisms from the model’s training data. The model’s exposure to information about those viruses were also reportedly limited.
According to the study, Evo 2 performed poorly when researchers tested its ability to generate proteins from human viruses. When researchers prompted the model to generate viral proteins, its sequence recovery was essentially random, which the researchers said effectively prevented human viral generation.
AI Genome Design Faces Limits And Safety Concerns
Some biosecurity experts raise concerns that people can misuse AI too for bioterrorism. The ability to generate novel viral genomes has raised debate on how AI-enabled biological design should be governed.
These calls for action focus on Evo 2’s open availability and the risk that bad actors could modify the tool to design biological threats and viruses. The Center for Strategic and International Studies warns that AI biological design tools could “develop more harmful or even novel epidemic- or pandemic-scale pathogens.”
The Washington-DC based NGO writes “rapid advancements in state-of-the-art BDTs—illustrated by the foundation model Evo 2—point to a world in which more capable models could help develop new or enhanced pathogens and evade existing safeguards.”
There are currently no federal restrictions specifically limiting the availability or use of Evo 2. Some U.S. policymakers have also proposed legislation that would establish federal screening requirements for synthetic DNA, reflecting concerns that AI could make it easier to design sequences that evade existing screening systems.
Similarly, CSIS recommends increased federal funding for the National Institute of Standards and Technology and the U.S. Center for AI Standards and Innovation. The advisory group says funding for the federal agencies could help strengthen the department’s work at the intersection of AI and biosecurity.
The group also recommends greater federal involvement in evaluating and testing advanced biological design tools, including noncommercial model, including an standardized, AI-enabled system for screening synthetic DNA sequences.
For now, the federal government has not taken action to limit access to biological AI models. But as these models advance and become more capable, federal oversight and safeguards are likely to evolve.
