Researchers at UC Berkeley used artificial intelligence to design new versions of a small CRISPR enzyme called TnpB. Several of the resulting proteins, called SynTnpBs, cut DNA at least as well as the natural enzyme. Many even outperformed it, across bacterial, plant, and human cells alike. This held even though their amino acid sequences diverged substantially from the original protein. Biochemist and Nobel laureate Jennifer Doudna led the team, which published its findings on 16 July in the journal Science.
Gene-editing nucleases such as Cas9 and Cas12 are the workhorses of CRISPR technology. They rely on a guide RNA to find a DNA target, then carry out several coordinated steps to cut it. According to Doudna, altering even a small part of that sequence usually breaks the protein’s function. Evolution has already fine-tuned each nuclease’s structure to its task, leaving little room for change. That constraint has made it difficult for scientists to engineer meaningfully different nucleases without losing activity altogether.
TnpB is a compact nuclease thought to be an evolutionary ancestor of Cas12. Cas12 is one of the two main nuclease families used in CRISPR gene editing today. Its small size makes it attractive for applications with limited delivery capacity. Certain plant gene-editing methods are one example, according to the study’s published summary. Other research groups had already applied AI language models to design new nucleases. However, those methods tended to reproduce the natural protein’s DNA-binding regions almost exactly. Designs from earlier models typically kept more than 99 percent sequence identity to the original enzyme. That left little room to test whether AI tools could generate more divergent enzymes that still work.
To design new nucleases, the team used a model called ESM Inverse Folding, or ESM-IF1. It works in the opposite direction of a structure-prediction tool such as AlphaFold. Instead of predicting a protein’s shape from its sequence, the model proposes sequences that should fold into a given shape.
First, the researchers gave the model the three-dimensional structure of natural TnpB. They then asked it to generate new sequences that could fold into that same shape. This step alone produced thousands of possible sequences. Many of those sequences, though, altered the regions responsible for binding DNA and RNA, risking the enzyme’s function.
To address this, the team layered on a second set of constraints based on evolutionary data. Positions in the protein that had stayed highly conserved across related enzymes were fixed to match the natural sequence. So were positions that appeared to coordinate closely with the nucleic acids. The model was then allowed to redesign the remaining positions freely.
Full-length designs built this way still showed limited activity. In response, the researchers split TnpB into two structural regions and designed each one separately before recombining the best-performing pieces. Roughly 50 top candidates from each region were combined in every pairing and tested for activity in bacteria.
In the initial bacterial screen, 466 of 1,980 designed protein combinations showed detectable cutting activity. About 8 percent of those active designs outperformed the natural TnpB enzyme in this bacterial assay.
The researchers then tested several of the strongest candidates in plant and human cells. In human cells, two variants edited a test gene more efficiently than the natural enzyme, at rates of 46 percent and 50 percent. The natural enzyme edited the same gene in 28 percent of cells. At some DNA sites in these cells, the best-performing designs edited nearly four times more efficiently than natural TnpB. Other new variants performed at levels comparable to the natural enzyme rather than exceeding it.
The team also tested variants in plant cells. There, too, some designs matched or exceeded the natural enzyme’s activity, consistent with the pattern seen in bacteria and human cells. Overall, the most active variant shared only 77 percent of its amino acid sequence with natural TnpB. The researchers said that level of divergence exceeded what earlier AI design methods had achieved while still preserving function.
To understand why some of the redesigned proteins worked so well, the researchers turned to cryo-electron microscopy. They used it to determine the structures of the most divergent variants. Comparing these structures with natural TnpB, the researchers found that the engineered protein had gained new electrostatic and hydrogen-bond interactions. These formed at the interface between the guide RNA and the target DNA. According to the study’s published summary, these are the first experimentally determined structures of AI-designed RNA-guided nucleases. The interactions appeared across several different conformational states that the enzyme passes through while it binds and cuts DNA.
Doudna and her coauthors say the platform shows that generative AI, guided by evolutionary information, can design functional gene-editing enzymes. These enzymes differ substantially from anything found in nature. Jennifer Doudna, the study’s senior author, described the results in an interview with Fierce Biotech. Her team, she said, was “able to develop non-natural nucleases that were active in human, plant and bacterial cells.” She and her colleagues suggested the approach could eventually let scientists design enzymes for specific applications. Possible examples include treating a genetic disease or helping crops tolerate a changing climate. The paper itself, however, does not test any such application directly.
Isabel Esain-Garcia, a postdoctoral researcher in Doudna’s lab at UC Berkeley and co-first author of the study, noted that testing remains a bottleneck. The design model “could be designing millions of sequences,” she said. But laboratory testing can only keep pace with a fraction of that output. Petr Skopintsev, a structural biologist and postdoctoral researcher in Doudna’s lab and the study’s other co-first author, said the broader value of the work lies in the pipeline itself. “People can take this and apply this method for other systems,” he said.
Outside scientists offered a mostly positive assessment. Soeren Lienkamp, a molecular biologist at the University of Zurich, was not involved in the research. He said the paper “marries two transformative fields” — AI-guided protein design and CRISPR biology. Benjamin Kleinstiver, a gene-editing enzyme engineer at Mass General Brigham and Harvard Medical School, was not involved in the work. His assessment was more cautious. He said the approach itself, not the creation of one more small nuclease, is what makes the paper noteworthy. Kleinstiver also noted that the study does not yet demonstrate new editing capabilities beyond what existing tools can already do.
Bacterial screening produced strong results, and the enzymes also showed some success in plant and human cells. Even so, the validated pool of working enzymes stayed small relative to what the AI model proposed. Of the several variants tested in depth in human cells, only two showed clearly higher editing activity than natural TnpB. The rest performed at roughly similar levels. That pattern suggests the design pipeline can occasionally find real improvements, but it does not yet guarantee them.
The human-cell experiments relied on a standard laboratory cell line rather than animal models or human patients. These results describe how the enzymes behave in cultured cells, not in a living organism or a clinical setting.
So far, researchers have demonstrated the design strategy on only one protein family, TnpB. The published study does not report testing the same pipeline on other CRISPR nuclease families, such as Cas9 or Cas12a. It remains unclear how well the approach would transfer to those systems.
