AI Is Now Designing CRISPR Enzymes That Nature Never Made

Every CRISPR enzyme you have ever pipetted was found, not built.

SpCas9 came from Streptococcus pyogenes. Cas12a came from Acidaminococcus and Lachnospiraceae. Cas13 came from Leptotrichia. The entire genome editing toolbox is a catalog of things bacteria evolved to fight off viruses, which we then borrowed. For thirteen years, expanding that toolbox meant one thing: sequence more microbes and hope something useful turns up.

That era is ending. In July 2025, a team at Profluent published a Cas9-like editor in Nature that no organism has ever made — generated by a protein language model, 403 mutations from SpCas9, and released for free as OpenCRISPR-1. In July 2026, Jennifer Doudna's group at the Innovative Genomics Institute published SynTnpBs in Science: synthetic versions of a compact nuclease, up to a quarter of their residues written by an AI model, several of them editing human cells better than the wild-type enzyme.

This is the biggest shift in how we get gene editors since the field started. It is also, as usual, being oversold. The press coverage quoted numbers that do not quite match the papers, and both papers contain caveats that make the "AI beats evolution" headline more complicated than it looks.

So let's go through what was actually measured.

What you'll learn

• Why "mining for editors" was always a bottleneck, and what replaces it
• What OpenCRISPR-1 and SynTnpB are, with the real numbers from the papers
• Why protein size is the hidden constraint behind all of this (and which of these actually helps)
• Where AI is changing editing outcomes without designing any protein at all
• Whether you can use any of this today, and how you'd validate it
• The caveats that the headlines skipped

1. The problem with mining nature for enzymes

When you need a Cas enzyme with a particular property — a smaller protein, a different PAM, better activity at 37 °C, lower immunogenicity — your options have historically been:

1. Go find one. Search metagenomes for a natural ortholog that happens to have it.
2. Engineer one. Directed evolution or structure-guided mutagenesis on an enzyme you already have.

Both work. Both are slow, and both are constrained. The Nature paper puts the problem plainly: natural CRISPR systems "typically exhibit tradeoffs in critical attributes such as basal activity in target cells, protospacer-adjacent motif (PAM) selectivity, thermal optima or in vitro biochemical properties." Evolution optimized these proteins to defend a bacterium. It did not optimize them to work in a HEK293T cell, dodge the human immune system, or fit inside an AAV capsid.

Directed evolution has its own ceiling. Fitness landscapes for multi-domain nucleases are rugged, and selection-based screens are hard to run in human cells. Structure-guided design needs an explicit structural hypothesis, which is difficult when the function you want depends on a protein cycling through several conformational states — which is exactly what an RNA-guided nuclease does.

The AI approach sidesteps both. Instead of searching what exists or nudging what you have, you learn the statistical rules that make a Cas protein a Cas protein, and then generate new ones.

If you want the mechanistic background on what these enzymes are doing in the first place, start with how CRISPR-Cas9 works.

2. OpenCRISPR-1: a Cas9 written by a language model

Profluent's approach was to treat protein sequences like text.

They first built what they call the CRISPR-Cas Atlas: 26.2 terabases of assembled genomes and metagenomes, mined down to 1,246,088 CRISPR-Cas operons. That is 4.1× more Cas9 protein clusters than UniProt holds. Then they fine-tuned a protein language model (ProGen2) on that atlas and generated 4 million sequences — later, a Cas9-specific model that produced a further million.

The generated proteins were genuinely novel. Clustered against natural Cas9s, they made up 94.1% of total phylogenetic diversity, with an average of 56.8% identity to any natural sequence.

Novel is easy. Functional is the hard part. They tested 209 Cas9-like proteins in HEK293T cells, narrowed to 48, and profiled those across five on-target and fifteen known SpCas9 off-target sites by amplicon NGS.

The winner, internally PF-CAS-182, became OpenCRISPR-1:

• 1,380 residues, 403 mutations from SpCas9, and 182 mutations from the nearest natural protein in their atlas
• On-target: median indel rate 56.4% vs SpCas9's 47.1% across the tested sites
• Off-target: median indel rate 0.32% vs SpCas9's 6.1% — a 95% reduction
• Genome-wide (SITE-Seq): a higher proportion of cleavage events at on-target sites than SpCas9 at all four RNP concentrations tested
• Critically, OpenCRISPR-1's off-targets were a subset of SpCas9's. It is not cutting in new places.

Two details make this more interesting than a better Cas9.

It works with base editors. Fused as a D10A nickase to ABE8.20, OpenCRISPR-1 gave 35–60% A-to-G conversion across three loci, comparable to SpCas9 nickase, without indel formation. If you work with base editors, this is a drop-in scaffold.

It may be less immunogenic. OpenCRISPR-1 lacks the known immunodominant and subdominant SpCas9 T-cell epitopes for HLA-A*02:01. An iELISA across serum from 40 healthy donors showed lower human antibody binding than SpCas9. This is a real argument for synthetic editors: SpCas9 is a protein from a human pathogen that most of us have been exposed to. A generated protein has no such history. The usual worry about de novo proteins is that unfamiliar epitopes might provoke a response — here the preliminary data points the other way, though 40 donors and an ELISA is a long way from a clinical immunogenicity package.

One more finding worth noting for anyone following the protein design field: the authors benchmarked against structure-based design with LigandMPNN, and those sequences showed no activity at all. Sequence-based language models beat structure-based design badly on this problem.

3. SynTnpB: designing the small one

Doudna's group went after a different target — and a different method.

TnpB is the transposon-encoded ancestor of Cas12. It is tiny: about 408 amino acids, a 46 kDa protein, against SpCas9's 1,368. It is RNA-guided, it cleaves DNA, and it uses a TAM (transposon-associated motif, TTGAT for the ISDra2 enzyme) the way Cas9 uses a PAM. Size is the whole appeal, and we'll get to why in a moment.

The method here is not a language model. They used ESM-IF1, an inverse folding model — you give it a 3D backbone, it proposes sequences that would fold into it. That alone was not enough: ESM-IF1 preserved the fold and the catalytic DED triad, but scrambled residues that recognize the TAM and contact the guide RNA, because it had no knowledge of the nucleic acids the protein has to bind.

The fix was elegant. They derived positional conservation from alignments of natural TnpBs and residue-coupling signals from a Potts model (GREMLIN) trained on paired TnpB-RNA and TnpB-DNA sequences, then froze the evolutionarily critical residues and let ESM-IF1 redesign everything else. Structure from the folding model, functional constraints from evolution.

They also found the protein does not tolerate redesign evenly. Splitting it into REC (DNA-recognition) and NUC (RNA-binding, catalytic) lobes revealed that the NUC lobe tolerates far more divergence than the REC lobe — 13 of 16 generated NUC lobes stayed active, against 1 of 16 for REC. Screening all 1,980 pairwise lobe combinations in a bacterial selection assay, 24% showed detectable activity and about 8% of those beat wild-type.

Nine variants went into human and plant cells. The results:

• BFP knockout reporter assay, HEK293T: wild-type ISDra2 averaged 28% editing. Variants v1 at 46% and v5 at 50% (both p < 0.001). This 50% figure, and the "1.8× wild-type" number in the press coverage, both come from this reporter assay.
• Endogenous loci (four genes, NGS-quantified): more modest. v1 gave 26%, 26%, 28%, 21%; v5 gave 21%, 20%, 24%, 35%; wild-type gave 7%, 18%, 25%, 30%. The standout was a 3.8-fold improvement by v1 at the locus where wild-type only managed 7%.
• The most divergent variant, v7, shares just 77% identity with wild-type — 83% in the REC lobe, 72% in the NUC lobe — with 85 residues newly generated. It hit 44% editing at one locus, and they solved its cryo-EM structure at 2.8 Å.

That structure is the part that should impress you. It captured a TAM-bound conformational intermediate never before observed in a TnpB. The AI-introduced residues built new electrostatic networks stabilizing the RNA-DNA interface, and an almost entirely AI-generated helical segment preserved a kink motion required for heteroduplex formation. The model reproduced a dynamic mechanism it was never explicitly told about.

The caveat the headlines dropped: specificity was not uniformly good. By genome-wide Tn5 tagmentation profiling, v1 was comparable to wild-type, but v5 and v7 had more detectable off-target sites. v5 is one of the two best-performing variants. Higher activity came with a specificity cost, which is the oldest tradeoff in this field and evidently not one AI has repealed.

4. Side by side

5. Why size is the constraint nobody talks about

Here is the thing that connects all of this to actual therapeutic work.

Adeno-associated virus is the workhorse vector for in vivo delivery, and its packaging limit is about 4.7 kb of total cargo. SpCas9's coding sequence alone is roughly 4.1 kb. Add a promoter, a poly(A) signal, a U6 cassette and the sgRNA, and you are over budget. The field's workarounds — split-intein dual-AAV systems, smaller orthologs like SaCas9, non-viral LNP delivery — all carry real costs in efficiency, complexity, or tissue tropism.

TnpB at ~408 amino acids is roughly 1.2 kb. An all-in-one AAV carrying a TnpB, its promoter, and its guide RNA fits comfortably.

This is precisely where OpenCRISPR-1 does not help you. At 1,380 residues it is marginally larger than SpCas9. It is a better Cas9 — more specific, likely less immunogenic, freely licensed — but it does not touch the delivery problem. Those are two different advances, and conflating them is the most common error in coverage of this area.

If compact nucleases are your problem, the Cas12a vs Cas9 comparison covers where the natural compact options currently sit.

6. The other kind of AI: predicting repair, not designing protein

Worth separating clearly, because it often gets lumped in: Pythia, from UZH, Ghent, and ETH Zurich, published in Nature Biotechnology in August 2025.

Pythia designs no enzyme. It attacks the other half of the problem — what the cell does after the cut.

You already know this pain. You deliver a perfectly good guide and a donor template, and you get a mess: partial integrations, deletions at the junction, frameshifted cassettes. The cut was fine; the repair was not.

The team found that repair at the genome-cargo interface is predictable by deep learning models and follows sequence-context-specific rules. From those predictions they designed repair templates using tandem-repeat homology arms matching microhomologies at the break. The results: precise integrations at 32 loci in HEK293T cells, germline-transmissible integration in Xenopus, and endogenous protein tagging in adult mouse brain — non-dividing, differentiated cells, where HDR normally fails outright.

For a bench scientist, this may be the most immediately useful of the three. It needs no new enzyme and no new license. It changes how you design the oligo you were going to order anyway.

7. Can you actually use any of this?

OpenCRISPR-1: yes. Profluent released it on GitHub and Addgene under a free license covering both research and commercial use — no upfront fees, milestones, or royalties. The license carries ethical-use terms, including a prohibition on human germline editing. Since it uses the SpCas9 sgRNA scaffold and an NGG PAM, it functions as a drop-in replacement in an existing SpCas9 workflow. Your guides do not change.

Pythia: yes. It is published as a design tool, so you can generate templates and order them.

SynTnpB: not yet, practically. These are recent academic constructs. Expect to contact the authors or wait for Addgene deposits, and expect the usual MTA friction. Watch this space rather than planning experiments around it.

How you'd validate an AI-designed editor is no different from validating any new nuclease, and your existing methods all apply:

• Editing efficiency at the target: T7E1/Surveyor for a quick yes/no, TIDE or ICE from Sanger traces for a fast quantitative read, amplicon NGS when you need real precision
• Off-target assessment: in silico prediction first, then an empirical genome-wide method — GUIDE-seq, CIRCLE-seq, or Digenome-seq

One warning specific to this situation. In silico off-target prediction tools were trained and benchmarked on SpCas9. Their scoring assumptions about mismatch tolerance may not transfer cleanly to a protein 403 mutations away, and they certainly do not transfer to a TnpB with a different targeting motif. For an AI-designed editor, lean harder on empirical off-target methods than on prediction. This is exactly the sort of thing that is easy to miss when a tool reports a confident-looking score.

8. The skeptic's corner

Five things to hold onto:

1. "95% fewer off-targets" is assay-specific. That figure comes from 15 previously characterized SpCas9 off-target sites — sites selected because SpCas9 hits them. It is a real and meaningful result, backed up by genome-wide SITE-Seq, but it is not a universal constant. Your locus may behave differently.

2. The best SynTnpB variants were less specific, not more. v5 and v7 showed more off-target sites than wild-type. Activity gains and specificity gains did not arrive together.

3. Reporter assays flatter everyone. SynTnpB's headline 50% came from a BFP knockout reporter. At endogenous loci the same variant ranged from 20% to 35%, sometimes below wild-type. Always ask which assay produced the number.

4. The authors themselves are cautious. Profluent's discussion states it will "be important to examine OpenCRISPR-1 activity across a range of experimental conditions, cell types and delivery methods to more thoroughly characterize robustness." They also note their measured SpCas9 activity ran below published levels because of lipofection efficiency — an honest disclosure that also means the comparison was run on a particular footing.

5. Nothing here is in the clinic. These are 2025 and 2026 papers. No AI-designed editor has entered a clinical trial. Immunogenicity in humans, long-term genotoxicity, and manufacturing at scale are all open questions.

9. So should you switch?

A straight answer, since this is what most people actually want to know:

• Routine knockouts in cell lines? Stay with SpCas9. It works, everyone knows its failure modes, and your reagents are in the freezer.
• Therapeutically-oriented work where specificity is the bottleneck? OpenCRISPR-1 is worth benchmarking head-to-head against your SpCas9 at your loci. It is free, it drops into your existing guides, and a 95% off-target reduction — if it holds at your sites — is a serious argument.
• Fighting AAV cargo limits for in vivo delivery? Compact nucleases are the right direction, and SynTnpB is the one to track. Not yet a reagent you can order.
• Struggling with knock-in precision or integration junk? Look at Pythia now. Lowest barrier, most immediate payoff, no new enzyme required.

Resources

My take

The result that stayed with me is not either efficiency number. It is the cryo-EM structure of SynTnpB v7.

An inverse folding model, trained only on protein backbones and never shown a nucleic acid, proposed 85 new residues. Those residues built coherent electrostatic networks across an RNA-DNA interface, preserved a helix kink that the enzyme needs in order to form its heteroduplex, and stabilized a conformational state nobody had previously observed in a TnpB. The model did not just produce something that folds. It produced something that moves correctly, and it did so without being told what the mechanism was.

That is a different claim from "AI made a better enzyme." A better enzyme is incremental. What this suggests is that these models have internalized constraints about protein dynamics that we have not explicitly written down — and that is the thing that generalizes beyond nucleases.

The honest framing for the bench, though, is this: we have gone from "search nature's catalog" to "generate candidates, then screen them." The screening did not disappear. Profluent tested 209 proteins to find one. The Berkeley group screened 1,980 lobe combinations to find nine. AI moved the bottleneck from discovery to validation — which means the assays you already run, the T7E1 gels and the TIDE traces and the GUIDE-seq libraries, matter more now, not less. Somebody has to check whether the generated protein actually works, and that somebody is at a bench.

If you want the full path from mechanism through guide design to NGS validation, that's what the CRISPR from Bench to Analysis series covers, and the expanded version is available as the book.

Would you run an AI-designed editor in your own experiments, or does it need clinical data first? I'm curious where people draw that line — drop a comment below.