Letting AI Rewrite Miniature Gene Scissors: Doudna’s Team Creates Highly Active SynTnpB

Gene-editing tools do not necessarily have to be borrowed solely from nature. A research team led by Nobel laureate Jennifer Doudna used artificial intelligence to redesign the miniature RNA-guided nuclease TnpB, creating a series of synthetic variants called SynTnpB. Some designs retained or even surpassed the editing activity of the original enzyme despite amino acid sequences that diverged substantially from the natural protein. The study was published in *Science*.

TnpB is considered an evolutionary ancestor of some CRISPR-Cas12 systems. Its small size could be advantageous for gene-editing work constrained by delivery-vector capacity. The challenge is that these enzymes must simultaneously recognize guide RNA and target DNA while switching between different conformations; even a slightly ill-considered sequence change could disrupt the entire set of molecular interactions.

The team used ESM-IF1, an inverse-folding model developed by Meta, to infer amino acid sequences that could form a specified three-dimensional structure, then used evolutionary information from the TnpB family to preserve residues critical to function. The researchers separately designed the RNA- and DNA-binding regions, then recombined them and conducted experimental screening, constraining the model’s exploratory capacity with biological rules.

Among 1,980 proteins and component combinations tested in bacteria, 466 showed detectable activity, and about 8% outperformed the natural reference enzyme. In subsequent human-cell tests, two variants achieved editing rates of 46% and 50%, respectively, at one test site, compared with 28% for the natural version. At some human DNA targets, the leading design was nearly four times as efficient as the natural enzyme. Candidate enzymes were also validated in Arabidopsis thaliana cells.

These findings reflect not only increased activity but also AI’s ability to move beyond the narrow region surrounding natural sequences. The most active SynTnpB shared only 77% sequence identity with the wild type. The researchers also used cryo-electron microscopy to determine the structure of one designed protein. The publicly available structure 9YYG shows this 408-amino-acid synthetic nuclease forming a complex with RNA and DNA at a resolution of 2.80 angstroms, and also captures a TAM-binding conformation that had previously only been hypothesized.

However, this remains a proof of concept focused on a single family of miniature nucleases, and it cannot be assumed that the same method applies to all gene editors. The available data also do not yet address therapeutic-development questions such as in vivo delivery, off-target cleavage, immune responses, and long-term safety. For now, SynTnpB’s more immediate value lies in demonstrating that structural models, evolutionary knowledge, and large-scale experimental screening can work together to create functional non-natural enzymes, rather than representing an editing product ready to enter the clinic.