AI Genome Models Create Viable Viruses, Raising Biosecurity Questions
Stanford researchers used Evo genome models to create viable E. coli phages, revealing potential tools against drug resistance and new biosecurity risks.
Summary
Verified findings: Stanford University researchers used the large genome models Evo 1 and Evo 2 to design complete genomes for bacteriophages, viruses that infect bacteria. As reported by Ars Technica on August 6, 2026, the models produced variants related to ΦX174, a well-studied virus that targets E. coli. The work, published in Science under DOI 10.1126/science.aec2657, did not create an unprecedented type of pathogen: Every generated virus remained closely related to an existing bacteriophage, and the study did not involve viruses capable of infecting humans or other vertebrates.
Large genome models apply a next-token prediction approach to DNA’s four bases, A, T, C, and G, rather than to human language. The researchers had intentionally excluded vertebrate-infecting virus sequences when training Evo 1 and Evo 2. For this experiment, however, the models received more than 2 million additional DNA bases from bacteriophages and were then fine-tuned on Microviridae, the viral family containing ΦX174. That virus offered a controlled test because its roughly 5,400-base genome contains 11 genes with known functions and has a characteristic starting sequence. Prompts containing four to nine bases produced the most useful results.
The team filtered the model outputs before laboratory testing. Candidates had to contain between 4,000 and 6,000 bases, avoid runs of the same base longer than 10 positions, have plausible GC-to-AT composition, and encode a host-binding spike protein at least 60 percent identical to ΦX174’s version. This left 302 candidate genomes, of which 285 could be chemically synthesized and introduced into bacteria. Sixteen, or 5.6 percent, inhibited E. coli growth. Nine worked as originally generated, while seven became functional after acquiring further mutations inside bacteria.
The viable designs generally resembled ΦX174, but some contained changes that would be difficult to obtain through unguided mutation. Among candidates at least 98 percent identical to the original virus, viability reached 46 percent. Most retained the standard gene set and all preserved the region where genome replication begins. Nevertheless, one viable design eliminated a viral protein and compensated elsewhere, another added a new gene, and a third replaced a ΦX174 gene with one from a distantly related virus. The authors’ probability analysis suggested random amino-acid changes should usually destroy functionality, yet several AI designs remained viable despite more than 25 alterations, including two with over 50.
Interpretation and implications: The findings indicate that genome models can identify coordinated changes that preserve viral function more effectively than random mutation. That could aid phage therapy, which uses bacterial viruses against antibiotic-resistant infections. In the study, a cocktail of natural E. coli phages failed against resistant hosts, while the 16 AI-derived viruses collectively evolved the ability to infect them, possibly through DNA exchange and additional mutations. However, phage therapy is not yet widely used, so the clinical value remains uncertain. The more immediate concern is dual use: A sufficiently resourced group could retrain a related system on vertebrate-infecting viruses. The researchers therefore called for stronger governance, including oversight of custom DNA orders, but the article notes that AI regulation has not kept pace. Whether these techniques can safely produce useful treatments, or be extended to more dangerous pathogens, remains unresolved.
Positives
- Evo 1 and Evo 2 produced 16 viable bacteriophage variants among the 285 synthesized genomes tested against E. coli.
- Viability reached 46 percent among generated sequences that were at least 98 percent similar to the original ΦX174 genome.
- Several viable viruses tolerated extensive coordinated changes, including two designs with more than 50 amino-acid alterations.
- A cocktail of the 16 AI-derived phages overcame resistant E. coli in testing, whereas the comparison cocktail of natural phages failed.
- The researchers excluded viruses that infect vertebrates from the models’ training data as a precaution against generating potentially dangerous sequences.
Risks & concerns
- Only 16 of the 285 synthesized candidate genomes inhibited E. coli growth, giving the overall experiment a viability rate of 5.6 percent.
- Seven of the 16 functional viruses required additional mutations after being inserted into bacteria, showing that the original AI outputs were not independently viable in those cases.
- The AI-derived phage cocktail appears to have overcome bacterial resistance through further evolution and possibly DNA exchange, introducing behavior that was not fully controlled by the initial designs.
- A sufficiently equipped actor could potentially train a similar genome model on vertebrate-infecting viruses because the researchers’ training-data restriction is a voluntary safeguard rather than a technical barrier.
- The researchers warned that governance of genome models and custom DNA ordering has not kept pace with the technology’s potential biosecurity risks.
- The therapeutic significance remains uncertain because bacteriophage treatments have been studied for years without achieving widespread clinical use.