## Are AI-Designed CRISPR Enzymes Finally Better Than Nature's Own?

Two AI-designed variants of the compact TnpB gene-editing enzyme have achieved **46% and 50% editing efficiency** in human cells, compared to 28% for the natural reference enzyme — and at some genomic targets, the best designs delivered nearly **fourfold higher editing** than wild-type TnpB. The results, published July 2026 in *Science* (DOI: 10.1126/science.aed6123) by Petr Skopintsev and colleagues, represent a meaningful advance in [computational protein design](https://synbiointel.com/glossary/computational-protein-design): a structure- and evolution-guided AI pipeline that generates compact nucleases divergent enough from natural sequences to be genuinely novel, yet functional enough to outperform them.

The study used [EvolutionaryScale](https://synbiointel.com/companies/evolutionaryscale)'s ESM Inverse Folding model (ESM-IF1) to propose new amino acid sequences built around the known TnpB structure, while holding evolutionarily conserved residues — those responsible for RNA and DNA recognition — fixed. Of 1,980 designed protein-part combinations screened in bacteria, 466 showed detectable activity. Approximately 8% outperformed the natural reference enzyme in that bacterial screen. The best candidates were then validated in human cells and *Arabidopsis* plant cells, earning the designation **SynTnpBs**.

This is a credible result from a peer-reviewed source, and the efficiency numbers are specific enough to anchor real comparisons. That said, the study covers only one compact nuclease family — TnpB — and the authors are appropriately cautious about generalizability.

---

## Why TnpB, and Why Now?

TnpB is a compact evolutionary ancestor of [CRISPR-Cas12](https://synbiointel.com/glossary/crispr-cas12) enzymes. Its small size is its primary selling point: delivery vehicles like [AAV](https://synbiointel.com/glossary/aav) have strict cargo limits, and compact editors that retain high activity are among the most commercially and clinically valuable tools in gene editing. The problem has been that TnpB's architecture — multiple functional domains and conformational states — makes it unusually difficult to redesign without disrupting DNA-editing activity.

Previous AI-based protein generation efforts largely produced enzymes that stayed very close to their natural templates. The *Science* study quantifies this gap directly: earlier language-model approaches generated proteins retaining **99% identity** to natural DNA-binding domains. The SynTnpBs, by contrast, achieved **83% identity** in the DNA-interacting lobe and **72% identity** in the RNA-interacting lobe relative to their closest natural counterparts. That divergence is the headline for protein engineers — it means the AI is not just interpolating within natural sequence space, it is beginning to explore genuinely novel territory while maintaining function.

---

## The Pipeline: Structure-First, Then Evolution Filtering

The methodological core of the paper is worth unpacking for practitioners. The team did not simply prompt a generative model and screen outputs. They:

1. **Started with known TnpB structure** as a scaffold for ESM-IF1 sequence generation
2. **Applied evolutionary constraints** — amino acids critical for RNA or DNA recognition were pinned, limiting the search to sequences that preserved essential contacts
3. **Screened in bacteria first** — a fast, cheap filter before committing to mammalian cell validation
4. **Validated top performers in human cells and *Arabidopsis*** — two very different cellular contexts, which adds confidence to the generality of the results

This tiered approach is practically important. The bacterial screen processed nearly 2,000 designs; only the most active variants consumed the more expensive human and plant cell validation bandwidth. For any team building an [enzyme engineering](https://synbiointel.com/glossary/enzyme-engineering) pipeline, that funnel design is the transferable lesson.

---

## What This Means for the Broader Field

The compact nuclease space is crowded with commercial interest. [Mammoth Biosciences](https://synbiointel.com/companies/mammoth-biosciences) has built its platform around ultra-compact CRISPR systems. Agricultural biotech companies targeting plant delivery constraints — where size limitations are acute — will be watching SynTnpB performance in *Arabidopsis* closely.

More broadly, this paper is a data point in an ongoing methodological debate: **structure-guided inverse folding vs. sequence-space language models** for enzyme design. The authors' framing — that existing language models generated proteins with 99% identity to natural homologs — is a pointed critique. ESM-IF1 is itself an [EvolutionaryScale](https://synbiointel.com/companies/evolutionaryscale) product, so there is a commercial dimension to that comparison worth noting. Independent replication of the divergence claims, particularly at the 72%/83% identity figures, will matter for anyone looking to build on this work.

The authors gesture toward eventual **de novo design** of RNA-guided systems — editors built from scratch rather than derived from natural scaffolds. The SynTnpB results are a necessary but not sufficient step toward that goal. The study establishes that the structure-plus-evolution approach can produce functional, divergent variants of *one* compact nuclease. Whether the same pipeline scales to other enzyme families, or to more complex multi-domain editors, remains an open experimental question.

For agricultural applications specifically, the *Arabidopsis* validation is encouraging — but crop-relevant species present their own delivery and regulatory challenges well beyond editing efficiency metrics.

---

## Key Takeaways

- **Two SynTnpB variants achieved 46% and 50% editing efficiency** in human cells, versus 28% for the natural TnpB enzyme
- **Nearly fourfold higher editing** was observed at some human genomic targets with the best-performing designs
- Of 1,980 protein-part combinations screened in bacteria, **466 showed detectable activity** and approximately **8% outperformed the natural reference**
- The ESM-IF1 structure-guided approach produced variants with **72–83% identity** to natural counterparts — substantially more divergent than prior language-model approaches (~99% identity)
- The study covers only the TnpB family; generalizability to other compact nucleases is not yet established
- Validation in both human cells and *Arabidopsis* opens a path toward compact agricultural editors where delivery constraints are significant

---

## Frequently Asked Questions

**What is TnpB and why does it matter for gene editing?**
TnpB is a compact RNA-guided nuclease that is an evolutionary ancestor of CRISPR-Cas12 enzymes. Its small size makes it attractive for delivery contexts with strict size limits, such as AAV-mediated gene therapy and plant cell editing, where larger editors like Cas9 may not fit.

**What editing efficiency did the AI-designed SynTnpBs achieve?**
Two SynTnpB variants reached 46% and 50% editing efficiency in human cells. The natural TnpB reference enzyme achieved 28% under the same conditions. At some targets, the best designs delivered nearly fourfold higher editing than wild-type TnpB, according to the *Science* paper.

**How is ESM-IF1 different from standard protein language models for enzyme design?**
ESM-IF1 is an inverse folding model — it proposes amino acid sequences compatible with a given 3D protein structure, rather than generating sequences based on sequence patterns alone. The study authors report that standard language models produced proteins retaining ~99% identity to natural DNA-binding domains, while their structure-plus-evolution approach achieved 72–83% identity, indicating greater novelty.

**Does this work apply to other CRISPR enzymes beyond TnpB?**
Not yet demonstrated. The authors explicitly note the study focused on one compact nuclease family, and the pipeline's performance on other RNA-guided systems remains to be tested experimentally.

**What are the agricultural implications of this research?**
The team validated SynTnpB activity in *Arabidopsis* plant cells, which is a relevant proof-of-concept for agricultural gene editing. Compact nucleases are particularly valuable in plants because delivery constraints often limit the size of editing payloads. However, moving from model plant to commercial crops involves additional regulatory and delivery challenges beyond editing efficiency.