AlphaFold Finds the Weak Spots That Make CRISPR Miss

A China-based team used AI protein-folding software to map — and then engineer away — the amino acids that let Cas9 edit the wrong DNA.

By The space-quantum Desk

[1]

[1]

The problem is structural. A perfectly matched DNA-RNA hybrid will have one structure, while one with one or more mispaired bases will have a slightly different structure. [1] Evolution has optimized the structure of Cas9 to stick to the former, but it apparently hasn't prevented Cas9 from adopting slightly different conformations that can interact with mispaired structures.

To map those conformations systematically, the team first built a large library of off-target editing sites using a modified CRISPR system that converts the DNA base adenine to a related chemical, inosine, isolating any DNA fragments that contain it. [1] They repeated the process with 10 different guide RNAs to get a broad picture of the types of off-target sequences present.

Then they fed those sequences into AlphaFold — but not all at once. When the team initially supplied AlphaFold with a target DNA sequence, a guide RNA, the Cas9 sequence, and an enzyme that chemically modifies bases and can stick to Cas9, it choked, placing one of the proteins in what was clearly the wrong location. [1] Stripping the input back to just the DNA, RNA, and Cas9 protein produced structures that agreed with experimental data.

The simplified runs revealed a consistent pattern. [1] The team called their analysis pipeline ContactSeek, built around AlphaFold's built-in contact-probability scores — the likelihood that any two items, such as amino acids or nucleotides, sit within eight Angstroms of each other.

What makes this approach worth watching is less the individual mutations found and more the logic behind finding them — a shift from trial-and-error protein engineering toward something closer to reading a blueprint before picking up the tools.

ContactSeek initially returned long lists of candidate amino acids, so the team focused on clusters — regions where multiple flagged residues sat close together, suggesting coordinated adaptation to mismatched bases. [1] In all, they tested 23 different amino acid swaps across 10 key positions.

The best variant maintained activity at correct target sites while its off-target activity dropped from 28 percent to 5 percent.

Compared against Cas9 variants produced by directed evolution — the previous benchmark for off-target reduction — the new designs produced similar or slightly better specificity. [1] The key distinction: the changes identified by ContactSeek may be somewhat more specific to a given guide RNA and mismatch combination rather than more generally effective.

That specificity cuts both ways. Tailored fixes are harder to generalize, but they may also be harder to break — a trade-off the field will need to weigh as the first wave of gene-editing therapies moves through clinical use.

The researchers also suggest the ContactSeek framework extends beyond CRISPR entirely — any protein-DNA interaction that needs fine-tuning could, in principle, be analyzed through the same lens. [1] For gene therapy developers already trying to clear regulatory bars on off-target safety, a computational shortcut to engineering cleaner proteins is a practical addition to the toolkit.

AlphaFold was built to predict how proteins fold from scratch. Using it to find the exact amino acids that let a molecular editor go rogue — and then design those amino acids away — is a narrower, quieter application. It may also be the more consequential one.

edit enough cells—and therapies generally have to edit many—and errors become inevitable.Ars Technica