Researchers at Stanford University have used generative AI to design complete bacteriophage genomes and then brought those designs into the laboratory for synthesis and testing, demonstrating how genome-scale AI design can move from computational output to experimentally validated biological systems.
The study, published in Science, used genome language models to generate complete genomes based on the bacteriophage ΦX174, which infects Escherichia coli. Stanford researchers synthesized and tested nearly 300 generated phages and identified 16 that performed particularly well against E. coli. The work builds on a broader shift toward using AI to design biological sequences rather than using machine learning only to analyze existing biological data.
Moving AI-generated genomes into the lab
The researchers used Evo 2, a generative AI model developed in part at Stanford, to produce novel phage genomes from a small amount of ΦX174 DNA. The process generated thousands of possible designs, creating another challenge: deciding which candidates were worth the cost and effort of physical synthesis and testing.
Samuel King, first author of the study, developed a computational framework that evaluated the generated genomes against design criteria and narrowed the candidates before chemical synthesis. The selected genomes were then transferred into bacteria and experimentally tested. The approach illustrates how AI can be integrated into research workflows to prioritize experiments rather than replace bench validation.
“With the cost of DNA synthesis still quite high,” senior author Brian Hie said, the framework allowed the team to concentrate on the most viable candidates. Several generated designs performed particularly well during testing, and 16 viable phages emerged from the experimental process.
For lab managers overseeing similar computational-to-bench projects, candidate prioritization can directly affect synthesis spending, instrument time, staffing, and throughput. The study also reinforces the importance of maintaining experimental validation as AI takes on a larger role in biological design.
Designing around bacterial resistance
The researchers also explored whether genetically diverse phages could make it more difficult for bacteria to develop resistance. They combined the 16 generated phages into a cocktail and found that it rapidly overcame resistance in E. coli strains resistant to the native ΦX174 phage.
The distinction is important: the experiment demonstrated the ability to overcome phage resistance, not antibiotic resistance in the tested E. coli. However, the work could contribute to longer-term efforts to develop phage-based approaches for difficult-to-treat bacterial infections.
What whole-genome AI design means for labs
The researchers have made Evo 2 openly available, while acknowledging that increasingly capable genome-design tools introduce safety and biosecurity considerations. Stanford researchers are now investigating more genetically novel phages and longer, more complex DNA sequences.
For research lab leaders, the advance points toward workflows in which computational models generate biological candidates, algorithms help prioritize them, and wet-lab teams determine whether those designs function as intended. As AI-driven laboratory workflows expand, managing the handoff between computational design and reproducible experimental validation will become an increasingly important operational consideration.
This article was created with the assistance of Generative AI and has undergone editorial review before publishing.
