AI-designed phage genomes: the first working virus and the push to beat superbugs

Brian Hie wanted to know whether a language model could write a virus. Not a protein, not a single gene — a complete genome, the full instruction set for something that had never existed and would only count as a success if it could still kill a bacterium.

Hie runs a lab at Stanford and works as an innovation investigator at the Arc Institute. He handed the problem to Evo 1 and Evo 2, two genome language models trained on millions of bacterial and viral sequences. The models generated candidate genomes for phages — viruses that infect bacteria — aimed at Escherichia coli.

His team ordered 285 of those designs from a DNA synthesis company and put them in front of the bacteria. Sixteen assembled into working viruses.

That third point is the reason this research landed in Science with an accompanying editorial warning rather than a victory lap. The therapeutic pipeline and the biosecurity problem share one model, one training run, and one synthesis order form.

The commercial context is moving on the same schedule. Phage therapy spent a century as a compassionate-use rescue for patients out of options. It is turning into a prediction and manufacturing problem, which is a far more investable shape.

The century before the algorithm

Phages were discovered in 1917, commercialised in the 1920s, then quietly abandoned once penicillin proved cheaper to make and easier to standardise. For eighty years the field survived on individual case reports.

TIMELINE: Phage therapy, 1917-2026
─────────────────────────────────────────────────────────────
  1917 ────── 1940s ────── 2017 ────── 2024 ────── 2026
  🧪          ⚠️           🔬          📊          ◉ NOW
  Félix       Antibiotics  Personalized 100-case   AI-designed
  d'Hérelle   make phages  phage rescue series     genomes from
  isolates    obsolete     saves a      published  Evo; NIAID
  phages      in the West  patient      in Nature  funds 3 centers

Each step solved a different problem. Discovery solved biology. Antibiotics solved manufacturing. Personalized therapy solved matching. What none of them solved was prediction — knowing, before you treat, which phage will work on which strain.

Evo learns to write a genome

The Stanford and Arc team built on ΦX174, a small lytic phage with a 5,386-nucleotide genome and eleven genes. That is a favourable test case: small enough to synthesize cheaply, well enough studied to compare against a natural benchmark.

Evo 1 and Evo 2 generated thousands of candidate genomes with plausible genetic architecture. The researchers filtered the set down to the designs most likely to assemble, then chemically synthesized roughly 285 of them and screened each one against E. coli C in the lab.

Sixteen produced functional phages. Their sequences and structures differed substantially from one another, which matters more than the headline number: the models were not memorising one template and mutating it. Cryo-electron microscopy showed that one generated phage uses an evolutionarily distant DNA packaging protein inside its capsid — a structural choice the researchers had not specified.

In direct competition, some of the designed phages outperformed the natural ΦX174 at killing bacteria. A mixture of the designed phages also worked against E. coli strains that had evolved resistance to ΦX174. That is the result with commercial weight: it suggests the design process can iterate faster than bacterial resistance.

Three centers and a prediction problem

The National Institute of Allergy and Infectious Diseases (NIAID) funded three Centers for Accelerating Phage Therapy to Combat ESKAPE Pathogens (CAPT-CEP) in 2026 — the first coordinated U.S. research network for phage therapeutics. The named targets are the ESKAPE pathogens: Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacter species.

CAPT-CEP network, funded by NIAID under RFA-AI-24-069, 2026

Gladstone's piece of that network is the Center for PhAIge Therapy, directed by Seth Shipman. The grant starts at $2 million with up to $10 million available across five years. Its stated job is to generate enough experimental data to make phage-matching models accurate, rather than merely plausible.

Phages have the potential to treat drug-resistant infections, but for patients to benefit from that potential, we need to be able to predict which phage to use for which patient, and design phages that are more effective than what we have today.— Seth Shipman, investigator, Gladstone Institutes

What the clinic already has

Engineered phages are not starting from zero. SNIPR Biome's SNIPR001, a CRISPR-armed phage that targets E. coli in stem-cell transplant patients, enrolled its first patient in the Phase 2 portion of a Phase 1b/2a trial on 17 September 2026, expanding the study to 66 patients with Memorial Sloan Kettering and Dana-Farber participating.

Locus Biosciences has moved further down the regulatory path. Its CRISPR-Cas3 phage candidate LBP-EC01 is in a controlled Phase 2 trial for urinary tract infections caused by drug-resistant E. coli, backed by a BARDA contract worth up to $85 million.

What those trials test is a fixed product against a fixed pathogen. What the AI work promises is a product family generated against a moving target. The gap between the two is the entire thesis.

As we wrote in July, the first engineered phages were already reaching pneumonia patients. The difference now is that the design step itself can be automated.

Where this breaks: the oversight gap

Thomas Inglesby and Moritz Hanke, writing the companion editorial in Science, put the problem plainly.

The generation of functional viral genomes has urgent biosafety and biosecurity implications.— Thomas V. Inglesby and Moritz S. Hanke, Science, 2026

The biology is dual-use by construction. A model that writes a phage genome can, in principle, write any small viral genome, and DNA synthesis is a commercial service. The paper's own authors flag the need for expert oversight and safeguards across the design pipeline, and the editorial argues that the field is generating capability faster than it is generating governance.

For investors this is a real diligence item, not a public-relations one. A therapeutic platform whose core model output is regulated as a potential biosecurity tool inherits a compliance surface that most biotech companies have never had to build.

What has to be true by 2030

Three things decide whether AI-designed phages become a therapeutic category or a striking laboratory result.

The 1917 discovery took decades to industrialise. The 2026 version has a model, a synthesis vendor, a regulator, and a national network in the same loop. Whoever closes the prediction gap first gets to define the manufacturing standard for the category.