Gene Editing · global
AI Goes Beyond Modifying Nature’s Scissors: Synthetic TnpB Achieves Gene Editing in Three Cell Types
The research team combined protein structure models with evolutionary constraints to design RNA-guided nucleases whose sequences are far removed from their natural parents yet remain functional; the results span bacterial, plant, and human cells, but safety and delivery hurdles remain before medical or agricultural applications.
Gene-editing tools have mostly been discovered in nature and then gradually improved through experimentation. Now, artificial intelligence is beginning to venture beyond this boundary: a research team has designed a group of RNA-guided nucleases never found in nature that can not only recognize target DNA, but also retain—and even exceed—the editing activity of natural proteins in bacterial, plant, and human cells.
The study, published in *Science*, started with TnpB. TnpB is a compact nuclease evolutionarily related to CRISPR-Cas12 that can cleave specific DNA under RNA guidance. The team combined an “inverse folding” model, which infers amino acid sequences from three-dimensional structures, with constraints on key residues identified through evolutionary information to generate a series of synthetic variants called SynTnpB.
The key to the design was not simply copying natural proteins. The researchers deliberately explored regions that differed substantially from known TnpB sequences, then used high-throughput experiments to screen for candidates that were genuinely functional. The results showed that some AI-generated editors maintained or surpassed wild-type activity across different biological systems, indicating that the model may be able to find viable protein solutions beyond natural sequences, rather than merely making localized modifications to existing enzymes.
Structural data also provided mechanistic support for these findings. Using cryo-electron microscopy, the research team resolved the structure of the most divergent variant and observed that, in different conformations, it formed contacts at the RNA and DNA interfaces that helped stabilize the complex. The publicly available 9YYG structural record shows that this 408-amino-acid synthetic nuclease forms a complex with RNA and DNA, with a single-particle electron microscopy resolution of 2.80 angstroms.
TnpB’s compact size makes it a potential starting point for a new type of editing platform. If its recognition range, efficiency, and precision can be further adjusted, it could potentially be used in the future for disease research, the development of gene therapy tools, or crop breeding. More broadly, AI may expand the range of nucleases available for engineering and reduce the limitation of developers relying solely on ready-made proteins from nature.
However, activity inside cells does not mean that the conditions for practical application have been met. Existing data are not yet sufficient to determine these synthetic enzymes’ genome-wide off-target risks, immunogenicity, long-term stability, or capacity for effective delivery, nor has safety evidence at the animal or clinical level been established. Although greater differences between artificial sequences and natural proteins may bring new functions, they also make it more difficult to apply existing experience to toxicology, immune response, and regulatory assessments.