Stanford University and Arc Institute Use Evo 2 to Build Viable Phages

Key Takeaways

• Evo 2 generated functional variants of the Phi X-174 bacteriophage, demonstrating that genomic AI can design complete viral genomes.
• The research could accelerate phage therapies for antimicrobial-resistant bacterial infections.
• The same capabilities are intensifying scrutiny of model access, DNA screening, laboratory controls, and dual-use governance.

Researchers led by Stanford University and the Arc Institute have demonstrated that a generative AI model can produce complete viral genomes that become functional after synthesis. The work moves genomic AI beyond predicting biological properties and into designing viable organisms, albeit exceptionally small ones with tightly limited hosts.

The researchers used Evo 2, a genomic foundation model that interprets DNA sequences somewhat like a large language model interprets text. Trained on more than two million bacteriophage genomes, Evo 2 generated thousands of candidate variants of Phi X-174, a small and extensively studied virus that infects bacteria. A refined dataset included 15,000 close relatives of that phage.

From the generated candidates, the team selected 302 genomes for synthesis. It successfully built and tested 285, and 16 infected an E. coli bacterium. Those results establish an important technical threshold: AI-generated genomic sequences can translate into functioning biological entities rather than merely plausible strings of genetic code.

The scope remains narrow. Phi X-174 has 5,386 base pairs, compared with three billion in the human genome, and the generated phages were designed to infect E. coli rather than people. Testing involved non-pathogenic E. coli strains under enhanced containment. Genetic sequences belonging to viruses that infect humans, animals, or plants had also been excluded from Evo 2's training data.

Still, this is not just a laboratory curiosity. Bacteriophages could help address antimicrobial resistance by attacking bacteria that no longer respond reliably to conventional antibiotics. Because phages tend to target particular bacterial hosts, however, developing suitable therapies can require considerable searching, testing, and iteration. Generative models may compress that design cycle by proposing many candidates for laboratory evaluation.

"New doors in science are now open because of what we can do with these models," the study's lead author said in a press statement quoted by Popular Mechanics. For pharmaceutical and biotechnology businesses, those doors could lead to faster phage discovery, new antimicrobial products, and more programmable approaches to biological engineering.

Commercial investment is already following the broader field. Research and Markets tracks a rapidly expanding AI drug-discovery sector, while MarketsandMarkets projects growth from about $1.86 billion in 2024 to $6.89 billion by 2029, representing a 29.9% compound annual growth rate. Companies including Insilico Medicine, Atomwise, and Recursion are applying related computational techniques to molecular and protein design.

The boundary between therapeutic design and biological misuse is not embedded neatly in the software. A model capable of exploring useful proteins or genetic sequences may also lower parts of the expertise barrier for harmful work. How should access change when a system progresses from recommending molecules to generating viable genomes?

NIST has warned that AI applied to protein engineering and genome editing could support the design of more virulent viruses or toxins, including sequences that might evade screening based on known genetic matches. That possibility puts pressure on safeguards at several points: model release, user screening, sequence ordering, synthesis-provider checks, laboratory access, and post-deployment monitoring.

Evo 2 adds another wrinkle because Arc Institute and NVIDIA released its weights, code, and training data as open source in February 2025. Open access can broaden scientific participation and improve independent scrutiny. It can also make centralized usage controls harder to enforce. The practical response is likely to involve layered governance rather than one restriction, drawing on tools such as the NIST AI Risk Management Framework and established biosecurity practices.

For business leaders, the immediate issue is not whether genomic AI is inherently beneficial or dangerous. It is how quickly governance can mature alongside capability. Firms developing biological models may increasingly face questions about training-data exclusions, red-team testing, customer verification, synthesis partnerships, audit trails, and incident reporting. Stanford University and the Arc Institute have shown what the technology can now do. The next test is whether industry oversight can keep pace without choking off credible medical research.