AI designed gene-editing enzymes that outperform everything evolution built

Researchers from the Innovative Genomics Institute and the Doudna lab used Meta's ESM Inverse Folding model to design 1,980 synthetic TnpB variants, with the best reaching 46% and 50% editing efficiency in human cells compared to 28% for wild-type TnpB, and up to nearly fourfold higher editing at some targets, according to a Science paper published July 16.

A Science paper from Jennifer Doudna's group shows AI-designed gene editors can beat natural TnpB enzymes in human cells, with the best variants reaching nearly four times the editing efficiency at some targets. Evolution had roughly a billion years to tune TnpB. The new point is sharper than that sounds: researchers have now shown they can design versions of this tiny gene-editing enzyme that nature didn't make, then watch some of them work better than the original in human cells. In a Science paper published on July 16, a team from the Innovative Genomics Institute and the Doudna lab used Meta's ESM Inverse Folding model to generate 1,980 synthetic TnpB variants. TnpB sits near the ancestral root of the CRISPR-Cas12 family, which is why gene-editing researchers care about it. That matters. It is small and programmable - and awkward enough that simply tinkering with it can break the whole machine. The screen was not a fishing trip through bacterial genomes. It was a design run. According to the IGI's release on the study, 466 of the 1,980 designed protein-part combinations showed detectable activity in bacterial tests. About 8% outperformed the natural reference enzyme. That's a real hit rate. The human-cell data are the part you should sit with. The best synthetic variants reached 46% and 50% editing efficiency, compared with 28% for wild-type TnpB. Nature reported that, at some human DNA targets, the strongest designs delivered nearly fourfold higher editing than the natural enzyme. That isn't a neat lab curiosity. It's AI producing working molecular hardware that evolution didn't hand over. Why size matters TnpB's size is not a minor technical note. If you want to edit genes inside the body, you have to deliver the editor into cells, and one of the most used delivery vehicles is adeno-associated virus, or AAV. AAV has a packaging ceiling of roughly 4.7 kilobases. Full-size Cas9 is already large before you add the guide RNA and regulatory pieces needed to make the system work. That is where compact editors become more than academic toys. A smaller nuclease gives you more room for the parts that make delivery possible - expression, targeting, all of it. TnpB is only around 400 amino acids, compared with roughly 1,300 for Cas9, as IGI noted in a separate plant-editing report earlier this year. You don't need to be a molecular biologist to see the practical edge there. More space in the vehicle means more room to build an actual therapy. There is a catch. Small editors often come with weaker performance. Earlier work has shown TnpB can edit human and plant cells, but activity has been uneven across targets. The Science paper doesn't erase that problem. It shows a way through it: design variants that keep the enzyme's shape and function while changing enough of the sequence to find activity nature never selected for. The design leap ESM-IF1 works in the opposite direction from the protein-prediction story most people know. AlphaFold predicts structure from sequence. ESM inverse folding starts with a known structure and proposes amino acid sequences that should fold into it. Doudna's team then layered in evolutionary conservation data and co-evolution signals between TnpB and its RNA and DNA partners, giving the model more than geometry to work with. That extra context mattered. The resulting SynTnpBs diverged to 72% and 83% sequence identity in key DNA- and RNA-interacting regions, according to the Science study details summarized by Phys.org. Prior language-model protein designs for related DNA-binding work had often stayed above 99% identity to natural templates. These were not lightly polished natural enzymes. They were new proteins with enough of the old architecture left to function. Frankly, that is the line between automation and invention. If an AI model only produces near-copies, it can still be useful. But if it can design divergent enzymes that survive contact with bacteria, Arabidopsis cells, and human cells, then biotech companies get a different kind of starting point. You are no longer waiting for a microbe in some database to have already solved your problem. Capital is following the same trail The Doudna lab is not working in a vacuum. Fierce Biotech reported in November 2025 that Azalea Therapeutics, a spinout from Doudna's lab, launched with $82 million in total funding, including a $65 million Series A led by Third Rock Ventures. Azalea is focused on in vivo cell therapy, not SynTnpB specifically, but the investment tells you where money is moving: toward tools that can make genome editing work inside the body rather than only in a dish. Earendil Labs is another signpost. The company said on March 20, 2026, that it had raised $787 million to scale AI-driven biologics discovery and development, including antibody and biologic programs. That is not the same science as TnpB gene editing. It is the same bet. Protein design is becoming a computational discipline before it becomes a wet-lab campaign. The harder work starts now. The Science paper tested activity across model systems, but medicine demands specificity and safety data that a bacterial screen cannot provide - off-target behaviour, cell type, manufacturing. A fourfold gain at one target is exciting, but a therapy is built across thousands of unforgiving details. Still, the direction is clear. If compact AI-designed editors keep improving, the first serious pressure will fall on areas where AAV delivery is already established and cargo size is a bottleneck, including eye and neurological programmes. The old search method asked what evolution happened to make. This one asks what you need the enzyme to do next. Also read: Sam Altman's World raises $52.5 million selling iris-scan tokens as AI deepfakes rewrite the identity problem https://startupfortune.com/sam-altmans-world-raises-525-million-selling-iris-scan-tokens-as-ai-deepfakes-rewrite-the-identity-problem/ • Meta and Microsoft walk into earnings week with $145 billion question marks hanging over them https://startupfortune.com/meta-and-microsoft-walk-into-earnings-week-with-145-billion-question-marks-hanging-over-them/ • Prentis AI lab co-founded by Reid Hoffman and Marc Pincus is in talks to raise $100 million https://startupfortune.com/prentis-ai-lab-co-founded-by-reid-hoffman-and-marc-pincus-is-in-talks-to-raise-100-million/