AI-Designed Synthetic CRISPR Enzymes Outperform Natural Gene-Editing Tools, Doudna Lab Reports in Science
Jennifer Doudna's team used AI protein design to create synthetic TnpB nucleases that edited human genes more efficiently than the natural enzyme.
Overview
A team led by biochemist Jennifer Doudna at the University of California, Berkeley used artificial intelligence to design synthetic versions of a compact gene-editing enzyme, and several of the resulting proteins outperformed the natural enzyme they were built from, according to a study published July 16 in the journal Science and reported by Nature. The work centers on TnpB, which Nature describes as “a group of tiny nucleases called TnpBs, which are evolutionary precursors to the commonly used Cas12,” the enzyme family used in many CRISPR gene-editing systems.
What We Know
The research team, which included scientists from the Innovative Genomics Institute and the California Institute for Quantitative Bioscience — both at UC Berkeley — along with outside collaborators, used an AI model called ESM Inverse Folding (ESM-IF1) combined with evolution-informed constraints to generate new versions of TnpB, according to Genetic Engineering & Biotechnology News (GEN). Rather than starting from a gene sequence, Nature reports: “The team began by providing an AI model with the final conformation of a type of TnpB and asking it to reverse-engineer changes to the underlying DNA templates that would nevertheless maintain the protein’s final shape.” The resulting synthetic enzymes are designated SynTnpBs.
In an initial bacterial screen, “466 of 1,980 designed protein-part combinations showed detectable activity, with about 8% outperforming the natural reference enzyme,” according to Phys.org. When the designs advanced to human cells, Phys.org reports that “two variants edited a test gene more efficiently than the natural enzyme, reaching 46% and 50%, versus 28% for the original enzyme.” Phys.org further reports: “At some human DNA targets, the best designs delivered nearly fourfold higher editing than the natural TnpB.”
The synthetic proteins also diverged from their natural template far more than earlier AI-designed nucleases. The study authors wrote, as quoted by Phys.org: “Unlike LMs that generated proteins that retained wild-type DNA binding domains with >99% identity to natural homologs, our structure- and evolution-guided design approach created DNA- and RNA-interacting lobes with AI-generated contacts that had 83% and 72% identity to their closest counterparts in nature, respectively.” GEN reports that cryo-electron microscopy of the most divergent variants showed the engineered proteins had formed novel electrostatic and hydrogen-bonding networks stabilizing their RNA-DNA interactions, and that the study authors concluded their approach helps “establish a strategy for creating non-natural RNA-guided nucleases and conformationally active nucleic acid binders, enlarging the designable protein space.”
Doudna, who won the 2020 Nobel Prize in Chemistry for her foundational CRISPR work and is identified by The Scientist as a UC Berkeley biochemist, described the iterative design process to Nature: “Once you start tweaking things, you realize pretty quickly that while you can make changes, they ultimately produce something that isn’t functional.” Soeren Lienkamp, a molecular biologist at the University of Zurich who was not involved in the research, told Nature: “Much like CRISPR democratized the ability to edit DNA at will, AI-based protein design promises to allow anyone to create totally novel properties in the protein space.”
The advance follows a broader push to use AI to expand the gene-editing toolkit beyond naturally occurring enzymes, including Eli Lilly’s investment in Profluent’s AI-designed recombinases, as previously reported.
What We Don’t Know
The reporting reviewed does not specify a timeline for moving SynTnpBs toward therapeutic or agricultural applications, nor does it detail off-target editing rates for the synthetic enzymes compared with natural TnpB. It also does not name every institution involved beyond the Innovative Genomics Institute and the California Institute for Quantitative Bioscience.
Analysis
The result is notable less for the specific enzyme than for the design philosophy behind it. Most prior AI-generated gene-editing proteins have hewed closely to natural sequences, preserving over 99 percent identity to known homologs to maximize the odds of a functional protein. The Berkeley team’s willingness to accept enzymes with only 83 percent and 72 percent identity to nature — and still find variants that edit DNA more efficiently than the original — suggests AI protein design is starting to move from imitating natural enzymes to genuinely redesigning them, expanding the pool of gene-editing tools available for both research and, eventually, clinical or agricultural use.