Researchers at Stanford University have demonstrated how generative AI could move beyond analyzing biological data and begin designing entirely new viral genomes.
Using the Evo 2 AI model, the team generated thousands of candidate DNA sequences based on the bacteriophage ΦX174, a virus that infects bacteria. Nearly 300 of the AI-generated genomes were ultimately synthesized and tested, with 16 showing particularly strong activity against E. coli.
The research offers an early example of AI being used to design complete biological systems rather than simply suggesting individual genetic modifications.
From AI-generated DNA to real-world testing
The research focused on ΦX174, a relatively small bacteriophage with a genome containing fewer than 6,000 base pairs. Its compact size made it a practical system for testing whether an AI model could generate an entire viable viral genome.
Researchers led by Stanford chemical engineering professor Brian Hie asked Evo 2 to generate a complete ΦX174 genome from a small starting sequence. Rather than making targeted changes to an existing genome, the model was instructed to produce the sequence from beginning to end.
The system generated thousands of possible genomes. Researchers then evaluated those candidates and selected a smaller group for chemical synthesis and laboratory testing.
According to Stanford, some of the resulting phages demonstrated greater fitness in laboratory experiments than naturally occurring ΦX174.
Screening helped narrow thousands of candidates
Generating DNA sequences with AI is only one part of the process. The researchers needed a way to determine which proposed genomes were most likely to function before spending resources synthesizing them.
Graduate researcher Samuel King developed a computational framework to evaluate the candidates against characteristics associated with ΦX174 and related phages.
The process involved generating sequences with Evo 2, evaluating them according to predefined criteria, selecting promising candidates, synthesizing those genomes, and then testing their biological performance.
This screening process was particularly important because synthesizing every sequence produced by the model would have been impractical and expensive.
The approach demonstrates an important distinction between AI-generated biological designs and functioning biological systems: computational predictions still need to be validated experimentally.
A combination of phages could help address resistance
The researchers also investigated whether multiple phages could be used together to make it more difficult for E. coli to develop resistance.
A treatment relying on a single phage can potentially become ineffective if bacteria develop resistance to that particular virus. A mixture containing genetically distinct phages could create multiple targets for the bacteria to overcome.
Stanford reported that a cocktail containing the 16 selected phages was able to overcome resistance in E. coli that had become resistant to naturally occurring ΦX174.
The findings suggest that AI-designed phage combinations could eventually become an area of research for combating bacterial infections.
Potential applications beyond E. coli
The same approach could potentially be applied to other bacteria.
The Stanford researchers have pointed to organisms such as methicillin-resistant Staphylococcus aureus, commonly known as MRSA, as a possible future target. Pseudomonas aeruginosa is another potential area of interest because of its role in difficult-to-treat infections.
However, the current work represents an early research demonstration rather than a ready-to-use medical treatment. Significant additional testing would be required before AI-designed phages could be considered for clinical applications.
Open-source AI brings opportunities and risks
Evo 2 has also been released as open-source software, allowing researchers to access the model and explore its capabilities for biological sequence design.
That openness introduces a broader discussion around the risks associated with AI-generated biological systems. Making advanced models widely accessible could accelerate legitimate scientific research while also creating concerns about potential misuse.
Hie has argued that AI could ultimately help researchers respond to naturally occurring biological threats and potentially strengthen defenses against engineered ones. At the same time, the availability of biological design tools raises questions about how such systems should be governed and monitored.
The debate reflects a broader challenge emerging as AI becomes increasingly capable of working with biological information: researchers must balance scientific accessibility with appropriate safeguards.
The next frontier is more complex biology
The Stanford team plans to explore whether Evo 2 can generate longer and more complex DNA sequences.
Researchers are also interested in applying these techniques to bacterial genomes. In the longer term, AI-designed microbes could potentially be developed for applications involving the production of chemicals, medicines, fuels, and other useful materials.
For now, one of the central challenges is increasing the genetic novelty of AI-generated designs while maintaining control over their resulting characteristics.
The research illustrates how quickly generative AI is expanding beyond traditional text, image, and software applications. Models such as Evo 2 are beginning to operate in a domain where their outputs can become physical biological systems, creating new possibilities for biotechnology while also raising difficult questions about safety, validation, and oversight.
