Stanford's Evo 2 AI Model Designs Phages That Kill E. coli

Antibiotic-resistant bacteria kill hundreds of thousands every year, and no new drugs have come along in a decade. The only realistic backups currently on the horizon are bacteriophages, or phages, the viruses that prey on bacteria-and now they can be designed by computers. Researchers at Stanford have designed, chemically synthesized and identified sixteen that killed E coli in the lab using the Evo 2 AI model for designing almost 300 phage genomes from scratch. 

This post details what the researchers really developed, how the Evo 2 AI model produced working viral genomes in a single pass, why the researchers used a 16-phage cocktail versus a single treatment and what the open-sourced nature of the work represents for AI-designed biology round 2.

Quick Answer: Brian Hie and Samuel King have used the DNA-trained generative AI model Evo 2 to write nearly 300 candidate bacteriophage genomes from scratch; sixteen were chemically synthesized and successfully killed E coli strains-including resistant ones-in the lab. The work, published in Science and confirmed by Stanford, is the first peer-reviewed demonstration of generative AI designing complete, functional viral genomes. Evo 2 has been released open-source, which has also prompted independent biosecurity researchers to flag gaps in how synthetic DNA orders are currently screened.

This article covers:

• What Stanford built with the Evo 2 AI model?

• How Evo 2 writes a full phage genome in one pass?

• Why ΦX174 was the right test case?

• Why the team chose a 16-phage cocktail against E. coli?

• Where Evo 2 goes next: MRSA, Pseudomonas, and engineered microbes?

• The open-source release and the biosecurity question

What Stanford Built With the Evo 2 AI Model?

Stanford's team asked a generative model to write viral DNA from scratch, then tested whether the output actually worked. According to Stanford's own account of the research, the researchers synthesised close to 300 phages designed by the Evo 2 AI model and identified 16 that showed strong E. coli-killing activity in the lab.

The project was led by Brian Hie, an assistant professor of chemical engineering at Stanford and a Dieter Schwarz Foundation Stanford Data Science Faculty Fellow, with bioengineering graduate student Samuel King running the experimental side. Hie built Evo 2. King built the computational framework that decided which of the thousands of AI-generated genomes were worth spending synthesis money on.

The target organism was ΦX174, a bacteriophage that infects E. coli and has been studied for decades. What matters here is not the phage itself but the claim behind the experiment: the Evo 2 AI model created entire viable viral genomes, not just small local edits to an existing sequence.

How Evo 2 Writes a Full Phage Genome in One Pass?

Evo 2 generates DNA the way a large language model generates text: it starts from a short snippet and predicts the next base, then the next, until a full sequence exists. The Stanford team pushed that behaviour to its limit. They asked the model to produce an entire ΦX174 genome end-to-end in a single left-to-right pass.

"In this case, we wanted the model to generate the entire genome end-to-end in a single left-to-right pass. We didn't add anything," Hie said, according to Stanford's report. The model produced thousands of candidate genomes before the team narrowed them down for chemical synthesis.

That narrowing step matters, because DNA synthesis is expensive. King's framework scored each candidate against traits drawn from ΦX174 and its close relatives, then selected the most promising ones for the lab.

The Design Loop, Step by Step

1. Generate thousands of candidate genomes with the Evo 2 AI model.

2. Score each candidate against biological criteria derived from ΦX174 and related phages.

3. Select the top-ranked genomes.

4. Chemically synthesise the selected DNA.

5. Test the synthesised phages in the lab against E. coli.

Hie said the framework kept synthesis costs manageable by focusing spending on the candidates his team judged most viable. Even so, computational screening, chemical synthesis, and laboratory assays were still needed to translate AI output into a working phage. The Evo 2 AI model shortens the front end of that pipeline; it does not replace the wet lab.

Why ΦX174 Was the Right Test Case?

ΦX174, pronounced "FYE-ex-1-7-4," is one of the smallest and best-characterised phages in biology. Its genome contains fewer than 6,000 base pairs. The human genome, by comparison, runs to roughly 3 billion. That size gap is why ΦX174 works as a stress test for a generative model: it is small enough to synthesise cheaply and study thoroughly, but large enough that end-to-end generation is a real result.

Even at that scale, interpretation is hard. Hie said researchers still face a difficult task reading a 5,400-character DNA sequence gene by gene. The headline finding is that some of the AI-designed phages showed higher fitness in the lab than native ΦX174. In other words, the Evo 2 AI model did not just copy a known virus. It produced variants that outperformed the reference organism on at least one measurable trait.

Why the Team Chose a 16-Phage Cocktail Against E. coli

A single phage is a fragile treatment. Bacteria mutate quickly, and resistance to one virus can spread through a population within hours. Stanford's researchers designed around that problem by picking 16 different E. coli-targeting phages rather than one.

"If the bacteria gains resistance to a single phage, it's game over for the medication," Hie said. "But if you have multiple genetically distinct phages in a mixture, it would be harder for the bacteria to develop resistance to the entire cocktail," Stanford reports. The 16-phage cocktail rapidly overcame resistance in E. coli that was already immune to native ΦX174.

Cocktail design is standard practice in phage therapy research today. What is new here is the source of the diversity: the members of this cocktail came from a generative model, not from environmental sampling.

Where Evo 2 Goes Next: MRSA, Pseudomonas, and Engineered Microbes?

Hie said similar work could target methicillin-resistant Staphylococcus aureus, better known as MRSA, and Pseudomonas aeruginosa, which Stanford describes as a leading cause of drug-resistant infections acquired in hospitals. Both are on almost every priority list for new antibacterial approaches.

The roadmap does not stop at phages. According to Stanford, small bacterial genomes could become targets for the model, opening the path to engineered microbes designed to produce chemicals, medicines, or fuels. Hie framed the remaining research as two open questions: "How do we get greater genetic novelty and how do we get greater controllability of the outcomes?" The next phase will extend Evo 2 to longer and more complex DNA.

Generative models are already reshaping adjacent parts of healthcare in ways that raise similar questions about trust and appropriate use. MIT and Columbia researchers recently found that AI explainability tools in health settings can help a non-expert and mislead an experienced clinician with the same output, a reminder that how an AI tool's reasoning is presented to different users matters as much as the tool's raw accuracy. Evo 2 sits in a related but distinct category: it's a design tool for expert researchers rather than a decision-support tool for frontline users, which shifts the relevant safeguards toward lab protocol and synthesis oversight rather than interface design.

Generative biology also fits a wider industry pattern of AI labs releasing powerful tools as open-weight rather than keeping them closed — a strategy playing out right now in large language models too, where vendors like Alibaba and Moonshot are testing revenue-sharing licences instead of giving everything away for free. The commercial logic differs from Evo 2's fully open release, but the underlying tension — open access accelerating research while widening who can use a powerful tool — is the same one biosecurity researchers are now raising about generative genome design.

King put the researcher-side view more simply: "One of the most rewarding parts of this project is the creativity Evo 2 allows. New doors in science are now open because of what we can do with these models."

The Open-Source Release and the Biosecurity Question

Hie has released Evo 2 as open-source software. Any researcher can download the model and use it to design genomes. That decision matters both for the pace of the science and for its risk profile.

Stanford's account acknowledges that the release has raised safety and security discussions. Hie said bad actors could modify versions of the tool, but argued that existing pathogens present a greater risk because they are easier to access and produce than a model-designed organism. He also said AI-enabled systems can help respond to naturally occurring pandemics and provide defence options against engineered biological threats. Those are his assessments, not settled policy.

Independent biosecurity researchers have pushed back on parts of that framing. Thomas Inglesby and Moritz Hanke of the Johns Hopkins Centre for Health Security, writing in a commentary published alongside the Stanford team's paper in Science, made a more specific, practical point: current DNA-synthesis screening systems work by matching a submitted sequence against databases of known dangerous genomes — and an AI-generated sequence that resembles nothing in nature won't necessarily trigger a match. They also note that, in the United States, no law currently requires DNA synthesis companies to screen orders at all; the main existing framework applies chiefly as a condition of federal research funding, not as a universal legal requirement. It's worth noting the researchers' own design choice here too: Hie and King trained this system on bacteriophage genomes only and explicitly excluded data on viruses that infect humans, animals, or plants, which is precisely the category the Johns Hopkins commentary argues needs stronger safeguards before any similar approach is extended to it.

What to Watch Next?

The near-term signal to track is whether Evo 2 can generate viable genomes for pathogens that matter clinically, not just for a well-behaved model phage. MRSA and Pseudomonas are the announced targets. If the same design loop generates, scores, synthesises, tests, and produces working phage cocktails against either one, generative biology moves from proof-of-concept into a serious contribution to the antibiotic-resistance response. In parallel, watch the policy side: whether the Johns Hopkins call for mandatory synthetic-DNA screening gains traction will likely determine how this technology is governed well before the science itself becomes the limiting factor.

FAQs

What is the Evo 2 AI model?

Evo 2 is a generative AI model built by Brian Hie's group at Stanford that produces DNA sequences from a short starting snippet. It can generate entire viral genomes in a single left-to-right pass, and it has been released as open-source software, so other researchers can use it.

Did the AI-designed phages actually kill E. coli?

Yes. Stanford synthesised close to 300 phages designed by Evo 2 and narrowed them to 16 that showed strong E. coli-killing activity in the lab. A cocktail of those 16 phages also overcame E. coli that was resistant to the native ΦX174 phage the model was trained around.

Why did the researchers use ΦX174 as the test case?

ΦX174 has a very small genome, fewer than 6,000 base pairs, and is one of the most-studied phages in biology. That combination makes it cheap to synthesise and easy to compare against a known reference, which is why it works well as a first end-to-end test for a generative DNA model.

Could Evo 2 be used to design dangerous pathogens?

Hie has acknowledged that concern publicly and argued that existing pathogens are a larger practical risk because they are more accessible than tools that require expertise, computation, and DNA synthesis to use. Independent biosecurity researchers at Johns Hopkins have taken a more cautious position, pointing out that current synthetic-DNA screening infrastructure isn't designed to catch novel, AI-generated sequences, and that no US law currently mandates such screening.

What comes after phages for Evo 2?

Stanford says the next phase will extend the Evo 2 AI model to longer and more complex DNA. Hie and collaborators are working on additional bacteriophage designs, with MRSA and Pseudomonas aeruginosa named as candidate targets, and small bacterial genomes as a longer-term goal for engineered microbes.