Stanford and Arc Institute researchers have used genome language models to design functional bacteriophages in the lab, turning AI from a tool for reading biology into one that can help write it.
A new paper in Science puts a clear marker down for synthetic biology. Researchers using Evo 1 and Evo 2, genome-scale AI models trained on vast libraries of microbial and viral DNA, generated hundreds of candidate versions of phiX174, a small bacteriophage that infects E. coli. Out of 302 AI-designed genomes, 16 were synthesised, tested, and shown to infect bacteria, replicate inside them, and kill their hosts. That is not a small software demo. It is a working biological system designed substantially by a model and then validated at the bench.
The work was led by Brian Hie and collaborators at Stanford and the Arc Institute, and it matters because the models are not simply predicting whether a known sequence will work. Evo 1 was trained on 2.7 million prokaryotic and phage genomes, while Evo 2 extends the approach to a much larger scale, with training across trillions of nucleotides. These systems learn patterns that sit across DNA, RNA, and proteins, which lets them reason across the central dogma rather than one molecule at a time. In practical terms, that means an AI model can propose changes that affect genome structure, protein function, and viral fitness together.
The resulting phages were not cosmetic edits of nature. Several were distinct enough to qualify as new species under International Committee on Taxonomy of Viruses rules, with altered sequences and rearranged gene orders. Some also performed better than their natural reference strain. They lysed bacterial cells faster, won direct competition experiments, and worked in combinations that overcame resistance across three E. coli strains. Cryo-electron microscopy added another layer of evidence, showing that one packing protein was evolutionarily distant from known structures while still functioning inside the viral particle. That is exactly the sort of result that makes investors, founders, and regulators pay attention.
PhiX174 is a good place to start because it is small, well studied, and historically important. The virus has only about 5,000 bases and 11 genes, and scientists have been working with it since the early days of molecular biology. But its simplicity should not make the achievement feel trivial. Designing a complete viral genome that can assemble, replicate, and compete is different from making a single protein variant or tuning a promoter. It asks the model to keep a whole system coherent. According to the Science paper, these were complete genome designs, not a few hand-picked mutations pasted onto a familiar scaffold.
Biosecurity Sounds Alarm
This is where the story stops being only about scientific capability and becomes a policy problem. Lab-designed life is no longer a distant thought experiment when a model can generate functional viral genomes that survive experimental testing. The upside is obvious: better phage therapies, faster antimicrobial tools, and new ways to fight bacteria that no longer respond to traditional antibiotics. The risk is just as plain. The same design logic that can produce a useful therapeutic phage could, in the wrong setting, help someone design a harmful biological agent.
That dual-use tension is why biosecurity groups are watching this field closely. Red teaming, publication review, screening of DNA synthesis orders, and restricted access to frontier biology models are all part of the conversation, but none of them is a complete answer. OpenAI and other AI labs have leaned on external testing for dangerous capabilities in language and code models, yet biology moves differently because the output can eventually become physical. Stanford's result shows that the gap between sequence generation and working organism is narrowing. The question is not whether this research should stop. It is whether the surrounding guardrails can mature quickly enough to match the pace of the tools.
Builders Face Reckoning
For startups, the message is direct. Genome models are becoming design infrastructure. A young company working on synthetic biology, phage therapy, agricultural microbes, diagnostics, or protein engineering can now imagine a development cycle where models produce candidate systems and labs test the most promising ones. That could reduce cost, shorten discovery timelines, and let small teams explore biological design spaces that once required years of specialist work. It also means the old pitch, that AI can help scientists search biology faster, is giving way to a stronger claim: AI can propose biology that did not previously exist.
Investors will like the speed, but they will have to price the risk. Biology companies already face long validation cycles, strict regulatory review, and public trust issues. Add open-weight models, cheaper DNA synthesis, and global access to advanced computational tools, and the market becomes both more exciting and harder to govern. The United States, European Union, and other jurisdictions are already discussing dual-use AI and biotechnology controls, but rulemaking is still behind the frontier. The companies that win here will not be the ones that move fastest without limits. They will be the ones that can show technical progress, containment discipline, and a credible path from model output to safe deployment.
Also read: GPT-Image 2 redraws the creative AI map with entity persistence and native watermarking • GPT-Image 2.0 drops 8K photorealism, slashing creative production costs • IRS AI audits go live, forcing tax tech to build defenses
