Stanford’s Evo 2 AI Model Designs New Phages to Fight E. coli

Artificial intelligence is moving beyond software and data analysis and into the field of biological research. A new study from Stanford researchers demonstrates how generative AI could help scientists design entirely new bacteriophages—viruses that infect and destroy bacteria.

Researchers at Stanford have synthesized nearly 300 phages from DNA sequences generated by the Evo 2 generative AI model. After laboratory testing, 16 of those candidates demonstrated particularly strong activity against Escherichia coli (E. coli).

The research focuses on bacteriophage ΦX174, pronounced “FYE-ex-1-7-4.” The project was led by Brian Hie, an assistant professor of chemical engineering and Dieter Schwarz Foundation Stanford Data Science Faculty Fellow. Bioengineering graduate student Samuel King led the experimental work described in the research paper.

Evo 2 Generates Complete Phage Genomes

The Stanford research explores a significant question in AI-assisted biology: Can a generative AI system create an entire viable viral genome rather than simply suggesting small changes to an existing DNA sequence?

For the experiment, researchers used Evo 2 to generate DNA sequences based on a short starting segment of a phage genome. Rather than asking the model to modify individual sections, the team instructed Evo 2 to generate an entire ΦX174 genome in one continuous, left-to-right process.

“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 explained.

The model produced thousands of possible genome sequences. Researchers then evaluated these candidates and selected a smaller group for chemical DNA synthesis and laboratory experiments.

ΦX174 was particularly useful as an experimental system because its genome is relatively small. It contains fewer than 6,000 base pairs, while the human genome contains approximately 3 billion base pairs.

Despite its compact size, interpreting a DNA sequence remains challenging. Hie noted that researchers still face substantial difficulties when trying to understand a sequence gene by gene, even when it consists of only around 5,400 characters.

Interestingly, some of the AI-generated phages demonstrated higher fitness than naturally occurring ΦX174 during laboratory testing. This finding provides evidence that generative models can potentially move beyond analyzing biological sequences and begin creating novel, functional genomes.

Computational Screening Reduced the Number of Candidates

Generating thousands of possible genomes creates another problem: Scientists cannot practically synthesize and test every sequence.

To address this challenge, Samuel King developed a computational framework designed to identify the most promising candidates before DNA synthesis.

The framework examined characteristics associated with ΦX174 and related phages. Researchers first generated genomes using Evo 2, evaluated them according to predefined design criteria, selected the strongest candidates, chemically synthesized the corresponding DNA, and finally tested the resulting phages in laboratory experiments.

“One of the main parts of the design framework was figuring out what traits the genomes should have based on ΦX174 and related phages,” King said.

This filtering process was important not only scientifically but also economically. Chemical DNA synthesis and laboratory testing can be expensive, so narrowing thousands of AI-generated possibilities to a smaller set of promising candidates allowed the researchers to focus resources where they were most likely to produce useful results.

The overall workflow demonstrates that AI-generated biological designs still require substantial human and laboratory involvement. Evo 2 produced the possibilities, but computational analysis, DNA synthesis, and experimental testing were necessary to determine which designs actually worked.

A 16-Phage Cocktail Helps Address Bacterial Resistance

One of the major challenges in using bacteriophages as antibacterial treatments is bacterial resistance.

If bacteria become resistant to a single phage, that phage may no longer be effective. The Stanford researchers therefore explored the use of multiple genetically distinct phages rather than relying on only one.

The final group included 16 phages targeting E. coli. The concept behind such a mixture is straightforward: If bacteria develop resistance against one phage, they may still remain vulnerable to other members of the cocktail.

“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.”

According to Stanford, the 16-phage cocktail rapidly overcame resistance in E. coli that had become immune to native ΦX174.

This result highlights one possible advantage of AI-designed phages. Generative models could potentially expand the range of genetically distinct phages available for developing targeted antibacterial mixtures.

Potential Applications Beyond E. coli

Although the current research focuses on E. coli and ΦX174, the researchers see opportunities to apply similar approaches to other bacterial pathogens.

Hie has suggested that future research could investigate phages designed to target methicillin-resistant Staphylococcus aureus (MRSA).

Another potential target is Pseudomonas aeruginosa, which Stanford identifies as a leading cause of medically resistant infections acquired in hospitals.

These applications are particularly significant because antibiotic resistance continues to create challenges for modern healthcare. Bacteriophages offer a fundamentally different approach by using viruses that naturally target bacteria.

However, the Stanford research does not suggest that AI-designed phages are already ready for widespread medical use. Instead, it demonstrates a research pathway in which AI-generated biological designs can be experimentally evaluated and refined.

Evo 2 Is Available as Open-Source Software

Another important aspect of the project is the availability of Evo 2 as open-source software.

Researchers can download the model and use it to generate and investigate biological sequences. Making the technology openly available could accelerate research by allowing scientists from different institutions to experiment with AI-driven genome design.

At the same time, the release has prompted discussions about biological safety and security.

Hie acknowledged that malicious users could potentially modify AI systems, but he also argued that existing pathogens can present greater risks because they are already accessible and can be produced more easily.

He believes AI-enabled biological systems could eventually help researchers respond to naturally occurring pandemics and develop defensive strategies against biological threats created by humans.

For King, the value of Evo 2 extends beyond its immediate experimental results. He emphasized the creative possibilities offered by generative AI and the new directions it can open for scientific research.

What Comes Next for AI-Designed Genomes?

Stanford researchers plan to push Evo 2 beyond relatively small phage genomes and explore longer and more complicated DNA sequences.

Hie is already collaborating with researchers at Stanford and other institutions on additional bacteriophage designs.

The researchers are also interested in applying AI genome generation to small bacterial genomes. Such systems could potentially be used to create engineered microbes capable of producing chemicals, medicines, or fuels.

However, major scientific challenges remain. Generating a DNA sequence is only one part of the problem. Researchers must also understand how genetic changes affect biological behavior and develop reliable ways to control the resulting organisms.

Hie identified two central questions for the future: how researchers can achieve greater genetic novelty and how they can gain greater controllability over biological outcomes.

The Future of Generative AI in Biology

The Stanford Evo 2 research represents an important step toward using generative AI for biological design. Instead of simply analyzing existing genomes, AI models are beginning to generate entirely new biological sequences that can then be tested in the laboratory.

The work involving nearly 300 synthesized phages and the eventual identification of 16 particularly effective E. coli-targeting candidates shows both the promise and limitations of this technology. AI can dramatically expand the number of biological possibilities researchers can explore, but computational predictions still need to be validated through real-world experiments.

As models such as Evo 2 become more capable, their applications could extend from bacteriophages to engineered microbes and other biological systems. The Stanford team’s work suggests that the combination of generative AI, computational screening, DNA synthesis, and laboratory testing could become an increasingly important approach to biological discovery.