Evo 2 designs viable viruses from scratch for the first time

What this is about

Researchers at Stanford University and the Arc Institute have created entirely new, viable viruses from genomes designed by the AI models Evo 1 and Evo 2. The work, published in Science on August 6, 2026, concerns bacteriophages: viruses that infect bacteria, not humans.

The experiment is still a turning point. The models did not merely propose individual proteins; they wrote complete genomes. From roughly 700,000 designs, the team selected 285 candidates. After synthesis, 16 variants were able to reproduce inside E. coli.

What Evo 2 actually does

Evo 2 is a language model for genetic sequences. Instead of predicting the next word, it predicts plausible arrangements of the DNA bases A, C, G, and T. For this study, it was asked to produce complete genomes modeled on the well-studied ΦX174 phage.

The setup was deliberately constrained. Training data from viruses that infect humans or animals was excluded. ΦX174 infects E. coli and has only about 5,400 base pairs. The team had the genomes synthesized and introduced them into bacteria. Clear areas in 16 Petri dishes showed that new phages had emerged and destroyed bacteria. Some variants worked better against certain E. coli strains than the natural reference.

Why it matters

Phages could one day be used selectively against antibiotic-resistant bacteria. A model capable of designing complete, functional genomes could accelerate the search for suitable candidates. The study offers concrete proof of feasibility, not a finished therapy.

At the same time, the security question changes. Generative biology has mainly assisted with proteins or partial sequences. A complete, viable virus is a different level of capability. In an accompanying Science commentary, Thomas Inglesby and Moritz Hanke warn that the ability to compose viral genomes already exists while suitable governance does not.

For research institutions, model providers, and regulators, this creates a practical task: access controls, training data, synthesis orders, and laboratory experiments must be considered together. A barrier at only one point is insufficient if the other steps remain open.

In plain language

A protein model is a little like a system that designs individual bicycle parts. In this experiment, Evo 2 wrote a complete blueprint that resulted in a bicycle that actually moved. The fact that it only ran on a fenced test track does not make the blueprint unimportant; it shows that the whole chain works.

A practical example

Suppose a hospital identifies an antibiotic-resistant E. coli strain. A research team could generate 100,000 phage genomes digitally, select 200 using safety and efficacy criteria, and test only those in the laboratory. If ten viable candidates emerged, the search would be far more targeted than broad screening of natural samples.

This is a realistic scenario, but it is not clinical routine. Before use in humans, further safety tests, manufacturing standards, approvals, and clinical trials would be required. The current work shows only that AI can design complete phage genomes that are viable in a laboratory.

Scope and limits

• The generated phages infect E. coli. The study does not show that Evo 2 designed a virus for humans.
• Sixteen viable variants from 285 synthesized candidates demonstrate feasibility, not reliable biological design on demand.
• The safeguards were chosen voluntarily. There is no uniform global standard that makes such controls mandatory.
• Long-term effects, mutations, and interactions of new phages cannot be inferred from the digital design alone.
• Open model access can accelerate research, but it increases the need for DNA-synthesis screening and laboratory oversight.

SEO & GEO keywords

Evo 2, Stanford University, Arc Institute, AI-designed viruses, bacteriophages, E. coli, synthetic biology, genome language model, phage therapy, biosafety, biosecurity, Science