AlphaFold helps redesign gene-editing proteins to reduce off-target edits

Introduction

Gene-editing therapies are beginning to reach patients, but safety remains a central concern. Even when guide RNAs are carefully chosen, the human genome is large enough that similar sequences can appear by chance. If a Cas protein tolerates a few mismatches, it may still bind and enable an edit at the wrong location.

A recent Nature study described a new way to tackle that problem: use AlphaFold to compare how a CRISPR complex behaves at correct and incorrect sites, then redesign the protein regions that appear to accommodate mismatched DNA.

Key points

• Specificity is not controlled by the guide RNA alone. Guide RNAs help direct the system, but Cas proteins also enforce—or fail to enforce—sequence discrimination. Cas9 can sometimes remain bound even when the RNA/DNA pairing is imperfect.
• The team built a broad off-target dataset. Researchers used a modified editing system that converts adenine into inosine, then isolated DNA fragments carrying those changes. They repeated the process with 10 guide RNAs to gather a diverse set of off-target sequences.
• AlphaFold was used as a structural comparison tool. An initial attempt to model the full complex did not work cleanly, so the team simplified the input to DNA, guide RNA, and Cas9. That produced structures more consistent with experimentally determined complexes.
• ContactSeek highlighted the moving parts. By comparing AlphaFold’s contact probability outputs for on-target and off-target sites, the researchers identified Cas9 amino acids whose contacts shifted in the presence of mismatches. Many off-target sites slightly changed Cas9’s conformation, while over 95 percent altered which amino acids contacted the RNA.
• Protein redesign reduced unwanted activity. The team tested 23 amino-acid substitutions across 10 key positions. One Cas9 variant retained similar on-target activity while reducing off-target activity from 28 percent to 5 percent. The strategy was also shown to work with a related Cas12 system.

Why it matters

The main contribution is not a universal cure for off-target editing, but a more rational design workflow. Existing approaches include careful guide RNA selection, directed evolution, and previously engineered high-fidelity Cas variants. This work adds a structure-guided method that asks a more specific question: which parts of the Cas protein help it tolerate a known mismatch, and can those parts be altered?

That could matter most when a therapy has a promising target but runs into a known off-target risk. Instead of abandoning the guide or relying only on broad screening, developers may be able to tailor the editing protein to that mismatch profile.

There are limits. The changes identified by this method may be specific to a given guide RNA and mismatch combination rather than broadly applicable to all editing contexts. Still, the study shows how AI structure tools can move beyond passive prediction and become part of an engineering loop for safer biological systems.

Source: Ars Technica AI