Stanford Researchers Use Evo 2 AI to Design Virus Genomes That Kill E. coli

Stanford researchers have used the generative AI model Evo 2 to design entirely new bacteriophage genomes, creating laboratory-tested viruses capable of killing E. coli and overcoming bacterial resistance.

Researchers at Stanford University have synthesised nearly 300 bacteriophages from DNA sequences generated by Evo 2, a generative artificial intelligence model designed for biological sequences.

After laboratory testing, the researchers identified 16 AI-designed phages that demonstrated particularly strong activity against Escherichia coli, or E. coli.

The research focuses on the bacteriophage ΦX174, pronounced “FYE-ex-one-seven-four,” a small virus that infects bacteria.

Evo 2 was developed by a team led by Brian Hie, assistant professor of chemical engineering at Stanford and Dieter Schwarz Foundation Stanford Data Science Faculty Fellow. Bioengineering graduate student Samuel King led much of the experimental work.

AI generates an entire viral genome

Instead of using AI to suggest individual genetic mutations, the researchers challenged Evo 2 to generate an entire ΦX174 genome from beginning to end.

The researchers provided the model with a small section of a phage genome and asked it to generate the remaining DNA sequence in a single continuous process.

Evo 2 produced thousands of possible genomes.

The team then used computational screening to determine which candidates were most likely to function before choosing approximately 300 sequences for chemical DNA synthesis and laboratory testing.

ΦX174 was selected partly because of its relatively small genome.

The virus contains fewer than 6,000 DNA base pairs, compared with approximately 3 billion base pairs in the human genome.

Even with a genome of only around 5,400 characters, researchers still face significant challenges understanding how each genetic change affects the behaviour of the virus.

According to Hie, some of the AI-generated phages demonstrated higher biological fitness than naturally occurring ΦX174 during laboratory testing.

The experiment suggests that generative AI could move beyond recommending individual genetic edits and potentially design complete biological genomes capable of functioning in the real world.

AI generation still requires laboratory validation

The researchers emphasised that Evo 2 did not automatically produce successful viruses without additional scientific work.

King developed a computational framework to evaluate the thousands of genomes produced by the model.

The screening process analysed characteristics found in ΦX174 and related bacteriophages, allowing researchers to identify candidates that appeared most promising.

The workflow involved several stages:

Evo 2 generated potential genomes, researchers computationally evaluated the sequences, promising candidates were selected, the DNA was chemically synthesised, and the resulting phages were tested in laboratory experiments.

The filtering process also helped reduce DNA synthesis costs by limiting laboratory production to candidates considered most likely to succeed.

The research therefore demonstrates a combined AI and laboratory workflow rather than a fully automated biological design system.

16 AI-designed phages overcome bacterial resistance

One potential application of bacteriophages is phage therapy, where viruses that specifically infect bacteria are used to fight bacterial infections.

One major challenge is resistance.

Just as bacteria can become resistant to antibiotics, bacteria may also evolve resistance against individual bacteriophages.

To address this problem, the Stanford researchers tested a mixture containing 16 genetically different AI-generated phages.

Stanford reported that the phage cocktail rapidly overcame resistance in E. coli strains that had become immune to the natural ΦX174 virus.

Using multiple genetically distinct phages could make it more difficult for bacteria to develop resistance against an entire treatment.

The same strategy could eventually be explored against more dangerous antibiotic-resistant bacteria.

Hie identified methicillin-resistant Staphylococcus aureus, or MRSA, and Pseudomonas aeruginosa as possible future targets.

Both are associated with serious infections, particularly in healthcare environments where antimicrobial resistance is an increasing global concern.

Evo 2 is available as open-source software

Evo 2 has also been released as open-source software, allowing researchers to download the model and experiment with genome design.

The availability of powerful biological AI systems has created discussions around biosafety and biosecurity.

Hie acknowledged that malicious users could potentially modify biological AI tools, but argued that existing pathogens currently represent a more immediate threat because many are already accessible without advanced AI systems.

He also believes AI-powered biological design could strengthen humanity's ability to respond to naturally occurring pandemics and potentially help develop countermeasures against engineered biological threats.

For researchers, however, the immediate opportunity is scientific discovery.

Generative biological models could allow scientists to explore genetic designs that would be extremely difficult to identify through conventional experimentation alone.

From AI-generated phages to engineered bacteria

The Stanford team now plans to explore longer and more complex genomes.

Future research could include additional bacteriophages as well as relatively small bacterial genomes.

AI-designed microbes could eventually be engineered to produce useful materials such as medicines, industrial chemicals, biofuels, and other biological products.

Two major challenges remain: generating greater genetic novelty while maintaining control over what the resulting organisms actually do.

The Stanford research represents another significant step in the rapidly emerging field of generative biology, where artificial intelligence is moving from analysing biological information to proposing entirely new genetic systems that can be physically created and tested.

For AI research, the breakthrough is not simply that a model generated DNA.

It is that some of those AI-generated sequences became functional biological viruses in the laboratory.