Researchers have used Google DeepMind's AlphaFold to redesign CRISPR-Cas9, dropping the gene editor's off-target activity from 28% to 5% while preserving its ability to hit the intended sequence. The team, based at institutions in China, published the work in Nature on July 24, 2026, and released a companion analysis pipeline called ContactSeek. The approach identifies exactly which amino acids inside Cas9 flex to accommodate mismatched DNA, then swaps them out.
The safety problem in gene editing is a numbers game. Guide RNAs pair with roughly 18 bases of genomic DNA, a sequence that should occur randomly only about once in 70 billion bases. The human genome runs 3 billion bases, so in principle each guide RNA should hit exactly one target. In practice, Cas9 tolerates a small number of mispaired bases, and any therapy that edits millions of cells will produce off-target edits somewhere.
The Chinese team built a library of real off-target sites by running a modified CRISPR system with 10 different guide RNAs and sequencing every location the enzyme touched. That gave them a broad, empirical picture of where and how Cas9 goes wrong, rather than relying on computational predictions of where mismatches might be tolerated.
Key facts
- 01Researchers cut CRISPR-Cas9 off-target edits from 28% to 5% after using AlphaFold to identify 10 key amino acid positions.
- 02The team, based in China, tested 23 amino acid swaps and published the results in Nature on July 24, 2026.
- 03Their tool, ContactSeek, uses AlphaFold's 8-Angstrom contact probability metric to compare on- and off-target Cas9 structures.
- 04Off-target sites caused Cas9 to shift structure in two-thirds of cases and altered RNA-contacting amino acids in over 95%.
- 05The method also worked on a second gene-editing system built around Cas12, suggesting the approach generalizes.
They then fed AlphaFold the target DNA, the guide RNA, and the Cas9 protein sequence and asked it to model the resulting complex. An initial attempt that also included the base-modifying enzyme failed — AlphaFold placed one of the proteins in the wrong location. Stripping the model down to just DNA, RNA, and Cas9 produced structures that matched experimental data.
“A perfectly matched DNA-RNA hybrid will have one structure, while one with one or more mispaired bases will have a slightly different structure.”— Research team, Authors of the Nature study
Comparing on-target and off-target complexes revealed a consistent pattern. About two-thirds of off-target sites caused Cas9 to adopt a slightly different overall structure. But more than 95% of them altered which amino acids inside Cas9 actually contacted the guide RNA, even when the protein's backbone stayed similar. The mistakes weren't happening because Cas9 was contorting — they were happening because individual amino acids were flexing to grip mismatched sequences.
That is where AlphaFold's contact probability metric became useful. The tool already scores the likelihood that any two atoms sit within 8 Angstroms of each other. The team, running that analysis across on- and off-target complexes, could point to specific amino acids in Cas9 whose contacts shifted only when a mismatch was present. They named the pipeline ContactSeek.
“Evolution has optimized the structure of the Cas9 to stick to the former.”— Research team, Authors of the Nature study
ContactSeek initially produced a long list of candidate residues, so the researchers filtered for clusters — patches of the protein where multiple flexing amino acids sat close together. They then tested 23 different swaps across 10 key positions. One variant matched wild-type Cas9's activity on intended targets while cutting off-target edits from 28% to 5%. The same approach produced comparable gains with different guide RNAs and worked on a second gene-editing system built around Cas12.
The work slots into a longer effort to fix Cas9's specificity problem. Earlier teams used directed evolution — mutating Cas9 at random and selecting variants with better fidelity — to produce their own improved versions. The AlphaFold-designed variants performed similarly or slightly better in head-to-head tests. The more interesting difference is that ContactSeek's fixes are tailored to specific guide RNA and mismatch combinations, rather than being general-purpose. In principle, the two approaches could be stacked.
The caveats are real. The team tested 10 guide RNAs, not the tens of thousands used across the field, and the specificity gains still leave residual off-target activity that would matter for a therapy editing billions of cells. Whether ContactSeek-designed variants translate to primary human cells and in vivo delivery, rather than the biochemical assays used in the paper, is untested. And AlphaFold's initial failure on the full four-component complex is a reminder that its accuracy drops as the system gets more crowded.
For the AI industry, this is one of the cleaner examples of a foundation model being repurposed as a design tool rather than a chatbot. AlphaFold was trained to predict static protein structures; the Chinese team used its contact-probability output — a byproduct of that training — to solve a problem the original model was never built for. That pattern, where the most valuable use of a large model turns out to be an internal signal exposed as an API, is going to keep showing up. The gene-editing companies pursuing CRISPR therapies now have a computational route to fidelity gains that used to require years of directed evolution, and the broader lesson is that protein-folding models have quietly become a design substrate for anyone working at the DNA-protein interface.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
