The gene-editing revolution may be about to get a powerful upgrade — not from discovering a new natural enzyme, but from designing one from scratch. Researchers at the Innovative Genomics Institute at UC Berkeley, led by Nobel laureate Jennifer Doudna, have used artificial intelligence to create synthetic versions of CRISPR enzymes that are more efficient than anything nature has produced.
The study, published in Science, focused on TnpB proteins, the ancient evolutionary ancestors of the CRISPR-Cas12 family. These are hypercompact nucleases, roughly 400 amino acids long, that are small enough to be delivered by AAV vectors for gene therapy. But natural TnpBs have a problem: they are inefficient in human cells, typically achieving editing rates well below 30 percent.
The AI approach
The team used ESM Inverse Folding (ESM-IF1), a structure-guided protein design model originally developed at Meta AI. Given a target 3D protein structure, the model predicts amino acid sequences that should fold into that shape. But the team added a crucial constraint: they masked functionally critical residues at the nucleic acid interface, keeping them fixed based on evolutionary conservation data from a Potts model trained on 16,335 TnpB sequences.
“Without those constraints, the AI might generate a perfectly folded protein that simply cannot bind DNA or RNA,” said Petr Skopintsev, the study’s co-lead author.
From the ISDra2 TnpB scaffold of Deinococcus radiodurans, the AI generated thousands of candidate sequences. Of the 1,980 designs tested in E. coli, 466 (24 percent) were functional nucleases, and roughly 8 percent exceeded wild-type activity. The top nine variants — SynTnpB v1 through v9 — were tested in human and plant cells.
Performance in human cells
The best variants, SynTnpB v1 and v5, showed dramatic improvements in human embryonic kidney (HEK293T) cells. At the BFP reporter locus, wild-type ISDra2 TnpB achieved roughly 28 percent editing. SynTnpB v5 reached 50 percent. At the endogenous EMX1 locus, the improvement was 3.8-fold. At multiple other loci — RUNX1, NIBAN1, AGBL1 — the synthetic variants consistently outperformed the natural enzyme, achieving 20 to 35 percent editing versus single-digit rates for wild-type.
The synthetic proteins are remarkably different from their natural counterparts. Sequence identity between SynTnpB variants and natural ISDra2 TnpB is only 50 to 60 percent, making them essentially a new protein family. Cryo-electron microscopy confirmed that the AI-designed proteins fold into the predicted structure, forming novel stabilizing contacts at the RNA-DNA interface.
The variants also worked in plant cells — Arabidopsis thaliana protoplasts — where they outperformed wild-type TnpB at most target sites.
Still behind Cas9
The SynTnpB variants remain less efficient than optimized Cas9 systems, which routinely exceed 80 percent editing in human cells. But their small size (roughly 45 kilodaltons, compared to Cas9’s 160 kDa) gives them a delivery advantage for gene therapy, where AAV vectors have limited cargo capacity.
The PAM (or TAM, in TnpB terminology) remains a limitation: the synthetic variants still recognize the 5-nucleotide TTGAT motif, which restricts the range of targetable sequences. The researchers note that future work could engineer the PAM specificity. Off-target effects also varied, with v5 and v7 showing more off-target sites than wild-type.
Sources
- Skopintsev P, Esain-Garcia I, DeTurk EC, et al. “Structure and evolution-guided design of minimal RNA-guided nucleases.” Science 393(6808):313-318, 2026. DOI: 10.1126/science.aed6123
- Nature News: CRISPR gets a power boost from AI-designed molecular scissors