AlphaFold AI redesigns CRISPR proteins to cut off-target edits

AlphaFold AI redesigns CRISPR proteins to cut off-target edits

Researchers have developed a method to make gene-editing systems safer by using AI to identify and modify the specific parts of CRISPR proteins responsible for unwanted edits. The approach reduced off-target editing activity from 28 percent to 5 percent in laboratory tests.

A team based at institutions across China adapted AlphaFold—Google's protein-folding AI—to analyze how the Cas9 protein interacts with DNA sequences it shouldn't be editing. By comparing structures of on-target versus off-target binding, they pinpointed exactly which amino acids in Cas9 flex to accommodate mismatched base pairs between guide RNA and DNA.

The off-target problem

Gene-editing therapies face a persistent safety challenge: even highly specific systems occasionally edit the wrong DNA sequences. While guide RNAs can be designed to target unique 18-base sequences that statistically should appear only once in the human genome, Cas9 proteins tolerate small numbers of mismatched bases without losing their ability to bind DNA.

When therapies must edit millions of cells to be effective, these low-probability errors become inevitable. Previous efforts to improve safety have focused on optimizing guide RNA sequences and evolving better Cas proteins through trial-and-error approaches.

How the AI analysis worked

The researchers first built a library of off-target editing sites by running modified CRISPR systems with 10 different guide RNAs and cataloging where unwanted edits occurred. They then fed AlphaFold the sequences of target DNA, guide RNA, and Cas9 protein for both correct and incorrect binding sites.

AlphaFold's contact probability analysis revealed that over 95 percent of off-target sites altered which amino acids in Cas9 contacted the RNA, even when the protein's overall structure remained similar. The team developed a computational method they called "ContactSeek" to systematically compare these contact patterns and identify amino acid positions that adapt to accommodate mismatched bases.

Testing redesigned proteins

Focusing on regions where problematic amino acids clustered, the researchers made 23 different substitutions across 10 key positions in Cas9. The optimized variant maintained normal activity at intended target sites while dramatically reducing off-target editing. The approach also worked with Cas12, a different CRISPR protein, suggesting broad applicability.

When compared against Cas9 variants developed through directed evolution by other research teams, the AI-designed versions showed similar or slightly better performance. The researchers noted their changes may be more tailored to specific guide RNA and mismatch combinations, potentially allowing combination with other improvements for even greater specificity.

Why it matters

This work provides a systematic method for customizing gene-editing systems to prevent known off-target effects, which could accelerate the development of safer therapies. Beyond CRISPR applications, the approach offers a general framework for fine-tuning protein-DNA interactions across biotechnology. As gene-editing treatments move from laboratory to clinic, computational tools that can predict and prevent editing errors before they occur in patients become increasingly valuable.

The findings were first reported by John Timmer at Ars Technica, based on research published in Nature.