Gene Editing · global
Rewriting a Gene-Editing Scissor from a Natural Scaffold: AI-Designed TnpB Retains Editing Activity Across Species
A research team used protein structure prediction in reverse, then applied evolutionary information to preserve key contact surfaces, creating hundreds of synthetic nucleases that remained functional. Results in bacteria, plants, and human cells show that extensively rewriting sequences may not require sacrificing performance, but safety and generalizability remain to be verified.
Gene-editing tools have long relied on prototypes supplied by nature, which researchers then modify incrementally. Scientists are now attempting to cross that boundary: rather than changing only a few amino acids in natural enzymes, they are using artificial intelligence to rewrite protein sequences while preserving the ability to recognize guide RNA, target DNA, and complete cleavage. Published in *Science*, the study shows that synthetic nucleases with substantial structural differences can also function in multiple types of cells.
The team selected TnpB, a compact, RNA-guided nuclease and an evolutionary relative of Cas12-class CRISPR enzymes. Small tools are easier to fit into delivery vectors with limited capacity, making them attractive for plant engineering and some gene therapy strategies. However, TnpB must switch between different conformations, and its protein, RNA, and DNA contacts influence one another, so arbitrary sequence substitutions can easily disable the entire mechanism.
The researchers used the ESM inverse-folding model to infer new sequences capable of forming the same structure from an existing three-dimensional scaffold. They also incorporated evolutionarily conserved information from the natural TnpB family, fixing amino acids essential for RNA or DNA recognition. The DNA-binding and guide RNA-binding surfaces were designed separately, after which approximately 50 leading candidates from each were combined and subjected to large-scale screening in *Escherichia coli*. This approach allowed the model to extensively rewrite noncore regions without completely severing the molecular contacts on which the enzyme depends.
Among 1,980 design combinations, 466 showed measurable activity, and approximately 8% outperformed the natural reference enzyme. Selected candidates were then tested in human cells and *Arabidopsis thaliana* protoplasts. At one test gene in human cells, two SynTnpB variants achieved editing rates of 46% and 50%, respectively, compared with 28% for the natural version; at some DNA targets, the improvement approached fourfold. The most active variant shared only 77% sequence identity with natural TnpB, showing that function is not necessarily tied to a highly similar natural sequence.
Structural analysis also provided a mechanistic explanation. A publicly available protein database record shows that one 408-amino-acid synthetic nuclease formed an R-loop complex with RNA and DNA, resolved by single-particle cryo-electron microscopy at 2.80 angstroms. This allowed the research team to observe a TnpB conformational state that had previously been proposed but not directly captured, and to find that new contact networks introduced by AI could stabilize the RNA–DNA interface. The findings therefore represent more than an activity screen; they also reveal why the model’s designs may work.
Several hurdles remain before these nucleases can become practical tools. The study examined only one family of small nucleases, so the same strategy cannot be assumed to apply to all CRISPR editors. The cellular experiments also did not address genome-wide off-target effects, immune responses, long-term toxicity, or in vivo delivery. For now, the more cautious conclusion is that combining structural models with evolutionary constraints can expand the explorable space of gene-editing scissors, but any medical use will still require comprehensive safety, manufacturing, and regulatory validation.