Berkeley, California, USA
July 16, 2026
Since the discovery of CRISPR-Cas9’s remarkable ability to precisely edit DNA in 2012, scientists have scoured the genomes of microbes in search of other similar DNA-cutting nuclease enzymes. This process of finding enzymes in nature and reengineering them as gene-editing tools has gradually expanded the array of tools in the genome editor’s toolbox, but the process is slow, limited by the variation generated through evolution, and reliant on chance discoveries.
In a paper out today in Science titled “Structure and evolution-guided design of minimal RNA-guided nucleases,” researchers at the Innovative Genomics Institute describe a new AI-assisted method for designing functional nucleases beyond those found in nature, opening the possibility of designing custom editors with specific properties. In their experiments, AI-generated variants of a model nuclease showed similar or improved editing activity across bacterial, plant, and human cells, relative to a known nuclease.
The study was led by first authors Petr Skopintsev and Isabel Esain-Garcia, postdoctoral researchers, together with former lab member Evan deTurk in Jennifer Doudna’s lab at the IGI. The full team was a collaboration between multiple IGI Investigators’ labs including the Doudna lab, Steve Jacobsen’s lab at UC Los Angeles, as well as Jamie Cate and Jill Banfield’s labs at UC Berkeley, combining expertise in biochemical wet-lab experimentation with cutting-edge computational science, a hallmark of modern genomics research.
Co-authors Petr Skopintzev and Isabel Esain Garcia in the Doudna Lab at the Innovative Genomics Institute, UC Berkeley (photo: Glenn Ramit)
“Usually, when we find new systems — there have been extensive searches of Cas12s, for example — people find them in nature and then they do protein engineering to further enhance them,” says Skopintsev.
This process has been fruitful but has limits. Evolution doesn’t explore the full possible design space; the evolutionary process is not only slow, it’s constrained by what came before. This is where Skopintsev, Esain-Garcia, and DeTurk saw a role for generative AI: to expand beyond what nature has already done with the aim of being able to rationally design new proteins.
“We proposed that we should be able to extend even further beyond what nature has designed for these proteins, and potentially even program in a controlled way for properties that we want to have in terms of activity, specificity, et cetera,“ says Skopintsev.
Building a better gene editor
RNA-guided nucleases like Cas9 have two properties that make them challenging for a would-be designer. First, these proteins interact with both an RNA sequence (the guide RNA that targets a DNA sequence) and the targeted DNA sequence, and these interactions have to be coordinated. Second, the structures are flexible, using different conformations to recognize and cut DNA.
Earlier attempts using AI design faced challenges recreating these properties. Language models trained on protein sequence data tended to generate active nuclease sequences too similar to the ones they were trained on. Their recipe for success was not tinkering with the part of the natural molecule that is essential to recognition of nucleic acids, which is especially challenging for design. Skopintsev, Esain-Garca, and deTurk reasoned that they could improve on past attempts and push the boundaries of generated sequences by incorporating both structural and evolutionary information into an AI model.
To do this, they used what is called an “inverse folding model” developed by Chloe Hsu and Alex Rives formerly at Meta AI. The Nobel Prize–winning technology AlphaFold can take a sequence of a protein and predict its three-dimensional structure. Inverse folding models do the opposite: give the model a structure, and it generates sequences that are predicted to fold into that shape. The team additionally modeled the nucleic acid interface of these generated sequences using evolutionary constraints, inspired by the work of Caleb Weinreb in the Marks lab at Harvard, Martin Weigt at Sorbonne Université, and Sergey Ovchinnikov and Hetunandan Kamisetty in David Baker’s lab at the University of Washington.
To test their dual approach in the lab, the team worked with TnpB, a hypercompact nuclease similar to CRISPR-Cas12, as a model. Their new hybrid AI method not only designed highly divergent enzymes compared to previous models, but when the enzymes were produced in the lab, they had a high rate of success. Nearly 1 in 4 were found to be active nucleases in lab experiments, testing roughly 2000 different proteins over the course of the study.
“We generated variants that outperformed the activity of the wild type TnpB nuclease, yet maintained the specificity,” says Skopintsev.
The team tested the newly designed gene editors in bacterial, plant, and human cells, finding activity across all three.
“We wanted to show activity in multiple genomes,” says Esain-Garcia. “So we tested it against the human genome and we showed that many of these variants either outperform the wild type while still maintaining high divergence, and then in the plant genome to show that we can have applications not only for potential human disease treatment, but also for agriculture.”
For one of the most divergent variants, they detailed the full three-dimensional structure of the molecule, making it the first ever solved structure for an AI-designed functional RNA-guided nuclease.
“Beyond this one protein, more importantly we established the pipeline, the set of methods to generate proteins at scale,” says Skopintsev. “People can take this and apply this method for other systems.”
Custom gene editors, on-demand
Cas12 has evolved to play a number of diverse roles in different lineages, that is, nature has already repurposed Cas12s for different applications. This design approach could allow researchers to harness that flexibility and optimize for specific use cases, designing a molecule to fit the immediate need.
“I think it’s very exciting because it opens the possibility to make tailored properties on-demand for enzymes,” says Esain-Garcia.
Potential properties include activity, speed, specificity to different nucleic acids, or variants that are more stable at different temperatures.
“In the future, when we think about personalized medicine and how we have to rapidly generate new genome-editing enzymes tailored to different diseases, this type of approach where there are a lot of custom properties that can be designed quickly would be beneficial,” says Esain-Garcia.
Read more: Skopintsev P, Esain-Garcia I, DeTurk EC, Yoon PH, Zhou Z, Weiss T, Kamalu M, Chamraj A, Loi KJ, Langeberg CJ, Boger R, Nisonoff H, Karp HM, Chen L, Shi H, Vohra K, Banfield JF, Cate JHD, Jacobsen SE, Doudna JA. Structure and evolution-guided design of minimal RNA-guided nucleases. Science. https://doi.org/10.1126/science.aed6123
This work was supported by an NSF Plant Genome Research Program grant (2334027) to S.E.J., J.A.D. and J.F.B. J.A.D. and S.E.J. are Investigators of the Howard Hughes Medical Institute. P.S. was supported by the Swiss National Science Foundation Mobility fellowship (P500PB_214418). A.C. was supported as a summer research student by HHMI.
Topics
Species
