Rewriting Smaller CRISPR Scissors: AI-Designed TnpB Works Across Bacteria, Plants, and Human Cells

Gene-editing tools do not necessarily have to be borrowed from nature. Researchers have now used artificial intelligence to rearrange the amino acid sequence of TnpB, a compact RNA-guided nuclease, creating a group of synthetic versions called SynTnpB that have never been found in nature. These molecules not only remained functional in bacteria, but some also matched or exceeded the editing efficiency of the natural enzyme in plant and human cells, providing experimental evidence for the design of more compact CRISPR tools.

TnpB is considered an evolutionary relative of the CRISPR-Cas12 system and can recognize and cut specific DNA under the guidance of guide RNA. Its more compact size theoretically makes it better suited for delivery vectors with limited capacity and may also facilitate plant genetic modification. However, such enzymes must simultaneously bind RNA, recognize DNA, and undergo structural transitions, so extensive sequence rewriting often disrupts one of these functions.

The research team therefore adopted a constrained design strategy. Starting from a known structure, it used the ESM inverse protein-folding model to propose new sequences, then used sequence conservation and coevolutionary information from the natural TnpB family to preserve amino acids that might be critical to function. The DNA-binding and guide RNA-binding regions were designed separately, and the leading candidate components were then recombined. This expanded sequence diversity while preventing the model from freely altering key contact surfaces.

In high-throughput screening in *Escherichia coli*, 466 of 1,980 designed combinations showed detectable activity, and about 8% outperformed the natural reference enzyme. The better-performing candidates were subsequently tested in human HEK293T cells and *Arabidopsis thaliana* protoplasts. According to the study results, two synthetic versions achieved editing efficiencies of 46% and 50%, respectively, at a test gene in human cells, compared with 28% for natural TnpB. At some human genome targets, the strongest design produced nearly four times as much editing as the natural enzyme while retaining the expected target-recognition sequence requirements.

The significance of these findings extends beyond fine-tuning a natural enzyme to work faster. The most active SynTnpB shared only 77% sequence identity with the natural reference protein. When calculated separately, the redesigned DNA- and RNA-interacting regions shared only 83% and 72% identity, respectively, with their closest natural counterparts. This indicates that the model found different molecular contact strategies within an evolutionary framework rather than simply copying existing proteins.

Cryo-electron microscopy provided further structural validation. The synthetic nuclease deposited in the Protein Data Bank contains 408 amino acids, and the structure of its RNA–DNA complex was resolved at 2.80 angstroms. The researchers observed that the designed amino acids formed new contact networks that helped stabilize the RNA and DNA interface. They also captured a TAM-binding conformation that had previously been proposed but had not been directly observed experimentally, explaining how these extensively rewritten sequences could still undergo the dynamic changes required for cleavage.

However, the current evidence remains an early-stage validation of tool development. The study began with only one TnpB family, the plant experiments used protoplasts, and the human tests were limited to cultured cells. The results do not yet demonstrate that this method can be broadly transferred to other CRISPR enzymes, nor do they address in vivo delivery in animals, off-target editing, immune responses, long-term safety, or efficacy across different tissues. AI has written new molecules that work in this case, but becoming therapeutic or agricultural products will still require more comprehensive evaluation of efficacy and risk.