There are many ways to tinker with a protein. Making one genuinely better, more active or more useful, is the hard part, and it has usually meant years of trial and error. Researchers at UC Berkeley shortened that loop. They used an AI protein design model, guided by the evolutionary record of thousands of natural relatives, to redesign an enzyme instead of stumbling toward improvements one mutation at a time. The enzyme was TnpB, the primitive ancestor of CRISPR, so what they built is a working gene editing system that does not exist in nature and matches the original, with a couple of versions that edit better. The caveat: this is a proof of concept in cultured cells, not a therapy, and it leans on an unusually well studied enzyme. But it is a small, real pointer to where this is heading. AI is compressing the time from question to insight in biology. Read the paper in Science.
Authors & Institutions
Petr Skopintsev, Isabel Esain-Garcia, Evan DeTurk and colleagues, with senior authors Steven E. Jacobsen and Jennifer A. Doudna (corresponding). Based primarily at UC Berkeley's Innovative Genomics Institute and the Departments of Molecular & Cell Biology, Chemistry, and Earth & Planetary Science, with collaborators at UCLA, the Gladstone Institutes, Lawrence Berkeley National Lab, HHMI, Monash, and Yale. Doudna shared the 2020 Nobel Prize in Chemistry for CRISPR.
Funding
Largely public and nonprofit: an NSF Plant Genome Research Program grant, HHMI (Doudna and Cate are HHMI Investigators), the Swiss National Science Foundation, and the NIH.
Conflicts worth noting
Substantial. Doudna and co-author Cate are inventors on CRISPR patents, and Doudna co-founded or advises a long list of gene-editing companies (Caribou, Intellia, Mammoth, Scribe, Editas, Evercrisp and others). None of this invalidates the work, but the senior author has deep commercial stakes in the field this paper advances.
What they did
Start with the problem. Gene editors like Cas9 and Cas12 are big, multi-part machines that have to grip DNA and RNA at the same time and change shape as they cut. That makes them hard to redesign. Change one thing and you usually just break it. Nature has explored only a sliver of the possible designs.
The team's insight was to let two different kinds of AI argue with each other. One model, an "inverse folding" network called ESM-IF1, looks at the 3D shape of the enzyme and proposes new amino-acid sequences that should still fold into that shape. The second input is evolution itself: by lining up thousands of natural TnpB relatives, the team measured which positions nature never lets change (because they are load-bearing) and locked those down. Everything else was free to drift. Turn a dial, and you get sequences that are a little different, or a lot different, from the original.
They then did something clever with the architecture. TnpB has two halves, a lobe that binds DNA (REC) and a lobe that binds the guide RNA (NUC). Rather than redesign the whole thing at once, they redesigned each lobe separately and mixed and matched, testing nearly 2,000 combinations. To find the ones that actually worked, they used a bacterial survival trick: a variant only lives if its editor successfully cuts a toxic plasmid. Winners were then tested in human kidney cells and in Arabidopsis plant cells, and the best variant was frozen and imaged by cryo-electron microscopy to see exactly what the AI had done.
What they found
The AI-designed enzymes worked, and not marginally. In human HEK293T cells the natural editor knocked out its target gene about 28% of the time. The synthetic variants landed in the same 23 to 32% range, and two of them (v1 and v5) climbed to 46% and 50% at one gene, clearly beating the original. Across four human genes and several plant targets, the best synthetic editors matched or exceeded the natural enzyme.
The headline variant, v7, is the striking one. Its sequence differs from the natural enzyme at roughly a quarter of its positions, yet it was the single most active editor in the human experiments and just as thermally stable as the original. When they imaged it, the cryo-EM structure showed the AI had invented an entirely new network of contacts holding the RNA-DNA interface together, and even captured the enzyme in a shape that had never been seen in the natural version. The AI did not just paraphrase evolution. It found a different way to build the same machine.
TnpB, in one breath. TnpB is a compact, roughly 400-amino-acid RNA-guided DNA-cutting enzyme found in bacterial transposons. It is thought to be the evolutionary ancestor of CRISPR-Cas12. Its small size makes it attractive for delivery into cells, which is why it is a favorite target for engineering.
Strengths
- Real functional validation, at scale. This is not an in-silico paper. They screened thousands of designs and nearly 2,000 lobe combinations in living bacteria, then confirmed hits in two independent organisms.
- Independent replication across systems. Activity held up in bacteria, human cells, and plant cells. Three different chassis is strong evidence the effect is real.
- Structural proof, not just a black box. Cryo-EM at 2.8 Å on the most divergent variant showed physically what the AI changed, and reversal mutagenesis (breaking the AI's changes on purpose) confirmed those residues actually matter.
- Honest characterization. They measured thermal stability, in-vitro cutting kinetics, and genome-wide off-target activity rather than reporting only the flattering numbers.
- Genuine novelty. The variants are meaningfully different from any natural sequence while staying functional, which is the hard part. Most divergent designs simply die.
Weaknesses
- It is a tool advance, not a therapy. Editing rates of 23 to 50% in cell lines and plant protoplasts are a real result, but this is a proof of concept. No animals, no in-vivo delivery, no clinical claim.
- The best variant has trade-offs. v7 showed more off-target sites than the natural enzyme and cut somewhat more slowly in the test tube. "Better editing" and "cleaner editing" are not the same thing, and for therapeutic use the off-target profile is what matters most.
- The method may not generalize easily. It leans on a scaffold that is unusually well served by data: solved structures plus thousands of well-catalogued natural relatives with their partner DNA and RNA. Enzymes without that evolutionary depth may not be redesignable this way.
- "Comparable to WT" is the honest framing for most variants. The two standouts aside, the majority of synthetic editors matched the original rather than beating it, so the practical win here is diversity and designability, not raw performance.
- Commercial gravity. The senior author's extensive company ties do not change the data, but they are the reason a "we can now design novel editors" result deserves a second read.
The Doudna group is not alone. "Use AI to invent better gene editors" has quietly become a crowded race, and the entrants are coming at it from different directions.
Direct rivals — AI that designs the editor itself
Profluent Bio. The closest analog and the obvious head-to-head. The Berkeley startup (founded 2023, CEO Ali Madani) used protein language models, essentially ChatGPT trained on protein sequences, to generate OpenCRISPR-1, a Cas9-like editor hundreds of mutations removed from anything in nature. They published in Nature and open-sourced it in 2024. Same destination as this paper, different route: Profluent generates from sequence, the Doudna team designs from structure.
Arc Institute (Evo and Evo 2). A nonprofit lab rather than a company, but arguably the most ambitious effort. Patrick Hsu and Brian Hie's genome-scale foundation model reads and writes DNA directly and can generate entire CRISPR systems, the protein and its guide RNA together, from scratch.
EvolutionaryScale. Less a gene-editing company than the toolmaker. Its ESM family of protein language models, including the ESM-IF1 model this very paper relied on, is the shovel much of the field is digging with. The team joined the Chan Zuckerberg Biohub in 2026.
Platform players who could pivot to editors
Generate Biomedicines, Cradle Bio, Google DeepMind (AlphaFold 3, AlphaProteo) and Microsoft Research all build general-purpose AI protein design. None is focused on gene editors today, but the same machinery that designs an antibody or an enzyme can, in principle, be aimed at a nuclease.
Incumbents working the old-fashioned way
Metagenomi and Mammoth Biosciences (the latter co-founded by Doudna) find and engineer novel compact editors the traditional way, by mining nature's diversity rather than generating new sequences. Mammoth's ultracompact NanoCas, plus its Regeneron and Vertex partnerships, shows where the commercial value sits. AI design is simply a newer path to the same prize: editors that are smaller, more precise, and easier to deliver.
Sources
Primary: Skopintsev P, Esain-Garcia I, DeTurk EC, et al. "Structure and evolution-guided design of minimal RNA-guided nucleases." Science 393(6808), 16 July 2026. DOI: 10.1126/science.aed6123. science.org/doi/10.1126/science.aed6123
Editor's summary: Funk MA. Science 393(6808), 2026.
Landscape: Profluent OpenCRISPR-1 (Nature, 2024). profluent.bio. Arc Institute Evo / Evo 2 genomic foundation models. arcinstitute.org/tools/evo. EvolutionaryScale ESM3. evolutionaryscale.ai. Mammoth Biosciences NanoCas; Metagenomi compact nucleases.