AI designed a new gene editor that edits human DNA, and it is free to use
An AI trained on over a million natural CRISPR systems invented it. It works on human DNA in the lab and is free to use, even commercially.
CRISPR gene editors are borrowed from bacteria, and they often behave worse inside human cells than in the microbes they came from. Profluent, a California company, asked whether a model could design better ones directly.
The team gathered more than 1 million CRISPR gene clusters from public genome data and trained language models on them, the same kind of model that predicts the next word in a sentence, but predicting protein sequences instead. The models then wrote new editors from scratch, along with the guide molecules each one needs to find its target.
The paper, published in Nature on 30 July 2025, reports that several of the generated editors matched or beat SpCas9, the standard natural editor, on how well they cut and on how precisely they hit the right spot. One of them, OpenCRISPR-1, edited human DNA in cells and also worked for base editing, the finer version that changes a single DNA letter.
What makes it notable is how far from nature it sits: about 400 mutations away from SpCas9, so it is not a tweak of a known protein. Profluent released it free to license for research and commercial use, and published the dataset too.
What it means
Gene editors no longer have to be found in nature and adapted; they can be designed, which widens the pool of tools labs can try. For patients this changes nothing yet: the work stops at cells in a dish, and no trial of an AI-designed editor has been registered. Releasing it free lowers the cost of experimenting with it, though the licence belongs to a company and could change.
Caveats
- No clinical use. The work shows editing in cells in a lab, not treatment of a patient, and no trial of an AI-designed editor has been registered.
- The company’s press release says off-target cutting was lower and immune reactivity weaker than with SpCas9, but gives no figures for either. Treat those as self-reported.
- “Designed by AI” still rests on nature: the models were trained on real CRISPR systems, and every candidate was validated in the lab by people.
- Specificity in a genome-wide assay is not the same as safety in a person, where off-target edits matter most.
- The licence is set by a private company and can change, even though the sequence is published.