AlphaFold Redesigns CRISPR Proteins to Slash Off-Target Gene Edits From 28% to 5%

AlphaFold Redesigns CRISPR Proteins to Slash Off-Target Gene Edits From 28% to 5%

AlphaFold has spent years showing researchers what proteins look like. A research team based at institutions in China just used it to show how to rebuild one.

Working from AI-predicted protein structures, the team pinpointed the specific amino acids in CRISPR's Cas9 enzyme that let the gene-editing system tolerate mismatched DNA, then engineered variants that cut off-target editing from 28 percent to 5 percent while preserving precision at the intended target. The results, published in Nature, demonstrate that structural predictions can serve as engineering blueprints rather than just scientific maps 1.

CRISPR's off-target problem is rooted in a numbers paradox. The guide RNA that directs Cas9 to its target pairs with roughly 18 bases of DNA, a sequence long enough that it should appear only once in 70 billion bases. The human genome contains about 3 billion. By that arithmetic, editing should never miss. But Cas9 tolerates a small number of mispaired bases without losing its grip, and the locations where mismatches are tolerated vary enough that predicting which guide RNAs are dangerous remains difficult 1. The practical consequence compounds quickly: therapies must edit large numbers of cells, and even low-probability errors across millions of edits become inevitable 1.

The method the team developed, called ContactSeek, turns AlphaFold's structural output into a diagnostic tool. A perfectly matched guide RNA and DNA target produce one Cas9 conformation. A mismatched pair produces a slightly different one. ContactSeek compares AlphaFold's predicted structures at matched and mismatched sites, then identifies which amino acids in Cas9 shift their contact with the RNA when the DNA is wrong. AlphaFold already reports what it calls contact probability, meaning the likelihood that any two molecules sit within eight Angstroms of each other. ContactSeek diffs those probabilities between correct and incorrect targets to locate the residues enabling errant editing 1.

The structural comparison exposed a clear pattern. About two-thirds of off-target sites caused Cas9 to adopt a slightly different overall shape, but more than 95 percent of them altered which amino acids contacted the RNA. Cas9 sometimes keeps its structure while rearranging internal contacts to accommodate mismatched DNA. Those flexible residues are the engineering targets 1.

From that map, the team tested 23 amino acid substitutions across 10 positions, focusing on regions where flexible residues clustered as a sign they were adapting to mismatched bases. The best variant matched normal Cas9's activity at correct targets while dropping off-target editing from 28 percent to 5 percent. Results held across different guide RNAs 1.

Existing high-fidelity Cas9 variants were already produced through directed evolution, a process of generating thousands of random mutations and screening for the ones that happen to improve specificity. That approach works but cannot explain why any given variant is safer. ContactSeek's rational design starts from a structural account of the problem and engineers against it directly. When the two approaches were compared, the rationally designed variants matched or slightly exceeded the evolved ones in both activity and specificity 1.

The method also generalized. Applied to Cas12, a different CRISPR system that recognizes DNA through a distinct mechanism, ContactSeek produced comparable improvements. That cross-system result suggests the approach is not specific to one enzyme but applicable to other programmable DNA-targeting proteins 1.

For clinical-stage gene therapies, the shift from evolutionary screening to computational design compresses a critical bottleneck. Every off-target edit in a patient is a potential safety event. A method that engineers specificity from structural first principles, generalizes across editing systems, and matches years of trial-and-error screening gives therapy developers a faster and more transparent route to the safety profiles regulators require.

ContactSeek's value is not any single redesigned protein. It is the demonstration that comparing AI-predicted structures at correct and incorrect targets can reveal, residue by residue, exactly where a protein's tolerance for error lives. Apply that structural diff to other enzymes that recognize specific molecular targets, and the same engineering logic could guide design decisions far beyond CRISPR.