AI Designs Working CRISPR-Like Enzymes From Scratch, Expanding the Gene-Editing Toolbox

A Doudna-led team at UC Berkeley has built RNA-guided nucleases, enzymes that cut DNA at locations specified by an RNA guide, that drift up to 28% away from any natural protein sequence while still working in living cells. Many of the designed variants cut DNA at or above wild-type levels in bacteria, plants, and human cells, the team reports in a peer-reviewed Science paper covering the work.

Four jobs have to fit into one protein. RNA-guided nucleases such as CRISPR-Cas9 and Cas12 do them at once: recognize a guide RNA, recognize a DNA target, activate, and cut. Coordinating those jobs across a single multi-domain protein is finicky; small changes in sequence often scramble one job without touching the others. Models that propose new sequences by reading the evolutionary record alone tend to stay close to known proteins, because natural sequence patterns are the easiest signal to imitate. The Berkeley group asked a different question: how far can a designed protein wander from nature and still work?

In the team's test case, the answer is far. They started with TnpB, a small transposon-encoded nuclease that helps transposable genetic elements cut their way around a genome and is a stripped-down ancestor of Cas12, the same kind of enzyme that CRISPR tools rely on for DNA cutting. Their pipeline paired ESM-IF1, a structure-guided inverse-folding model that proposes sequences for a given protein shape, with residue constraints drawn from evolutionary analysis of the TnpB family. The combination pushes the model away from any single natural sequence while keeping the geometry that lets the enzyme hold its RNA guide and clamp onto DNA.

The result is a set of "SynTnpBs." The two halves of the designed proteins that touch DNA and RNA sit at 83% and 72% sequence identity, respectively, to their closest natural counterparts. By comparison, sequences proposed by biological language models that read evolution alone stay above 99% identical to existing TnpBs. Structure-and-evolution guidance breaks the ceiling that sequence-only models cannot.

The team ran high-throughput screens in E. coli, in plant cells, and in human cells. Many SynTnpBs retained or exceeded wild-type TnpB activity across all three settings, the GEN summary reports, and cryo-EM structures of the most divergent active variant showed new electrostatic and hydrogen-bonding networks holding the RNA and DNA in place across the enzyme's different working states.

Protein design is moving from observation and minor variant tweaking toward genuine creation of new biological tools. The paper, titled "Structure and evolution-guided design of minimal RNA-guided nucleases" (bioRxiv preprint DOI 10.64898/2025.12.08.692503; published in Science, DOI 10.1126/science.aed6123), frames this as an expansion of the designable protein space beyond what evolution has produced. Senior author Jennifer Doudna co-discovered CRISPR-Cas9 in 2012 and runs the Innovative Genomics Institute and the California Institute for Quantitative Bioscience (QB3) at UC Berkeley.

The paper's own designable set still includes some reference-like variants that look much more like natural TnpBs, so "from scratch" is a directional claim about the most divergent active designs rather than a statement that every designed protein is novel. The result that holds across all the data is the activity one: AI-designed, evolution-divergent nucleases can cut DNA in living cells at or above wild-type levels.

Several co-authors, including Doudna, have filed a patent on aspects of the work. Funding came from the NSF Plant Genome Research Program, the Howard Hughes Medical Institute, and a Swiss National Science Foundation Mobility fellowship.

Whether the same structure-plus-evolution recipe generalizes to other multi-domain enzyme families beyond TnpB is the open question. If it does, the next round of custom editors, base editors, and nucleic-acid tools may be designed rather than discovered.