AlphaFold used to redesign CRISPR proteins, cutting off-target edits from 28% to 5%

Gene-editing therapies represent a major advance in medicine, but their development has been constrained by a persistent safety problem: off-target effects. Because the human genome contains about 3 billion bases, even guide RNAs designed to match only one target sequence can sometimes bind to unintended locations that happen to share similarity. The Cas9 protein, which recognizes both the guide RNA and the target DNA and enforces specificity through base-pairing rules, can tolerate mismatched bases—a feature that limits the precision of current systems. While therapies typically need to edit many cells, this tolerance means that off-target edits become inevitable at scale.

A team of researchers, based at various institutions in China, set out to use artificial intelligence to identify which parts of the Cas9 protein enable these problematic off-target interactions. They began by creating a large library of off-target editing sites using a modified CRISPR system that converts the DNA base adenine to inosine, then isolated and analyzed DNA fragments containing these modifications across 10 different guide RNAs. Next, they used AlphaFold—an AI protein-folding software updated to handle protein-nucleic acid complexes—to model how the CRISPR machinery interacted with these sequences. When they initially tried to model the full complex (DNA, RNA, Cas9, and a modifying enzyme), AlphaFold placed one protein incorrectly. They simplified the input to focus on just the DNA, RNA, and Cas9, the primary factor determining sequence specificity. This produced structures that matched experimentally determined ones.

By comparing structures from on-target and off-target sites, the researchers identified a key pattern: about two-thirds of off-target sites caused Cas9 to adopt a slightly different overall shape, while over 95 percent altered which amino acids made contact with the RNA. Using AlphaFold's "contact probability" metric—the likelihood that two molecular components are within eight Angstroms of each other—they developed a computational tool called ContactSeek to identify exactly which amino acids in Cas9 shift their interactions when mismatches are present. They then focused on regions where these shifted amino acids clustered, reasoning that these areas were adapting to accommodate mismatched bases.

The researchers tested 23 different amino acid swaps at 10 key positions identified by this analysis. One resulting variant maintained normal activity at intended targets while reducing off-target activity from 28 percent to 5 percent. Similar improvements were achieved with different guide RNAs, and the approach also worked for Cas12, a different gene-editing protein. When compared to Cas9 variants developed through directed evolution, the new designs showed similar or slightly better activity and specificity. The researchers noted that their approach may produce changes more tailored to specific guide RNA and mismatch combinations rather than universally effective across all scenarios. They also suggested that individual modifications identified through their method could potentially be combined with previously developed Cas9 variants for even greater safety improvements, though this was not tested. The work describes a general method for designing gene-editing systems to prevent known off-target events, potentially addressing a bottleneck in therapy development, and may have broader applications for optimizing protein-DNA interactions beyond gene editing.

Gene-editing therapies have faced a fundamental safety challenge since CRISPR systems were first adapted for human use. Although guide RNAs are designed to match only the intended target sequence, the sheer size of the human genome (about 3 billion bases) means that even rare sequences can appear multiple times by chance. More problematically, the Cas9 protein can tolerate small numbers of mismatched base pairs without losing its ability to bind DNA, making it difficult to predict which off-target sequences will be cut. Researchers have pursued various strategies to address this—optimizing guide RNA selection, improving the Cas proteins themselves, and modifying the effector proteins that make the actual DNA change. This new work takes a computational approach, leveraging AlphaFold's ability to model protein-nucleic acid interactions to identify the exact structural features of Cas9 that enable off-target binding.

What makes this approach particularly valuable is its specificity and generality. By analyzing how Cas9's structure changes when binding mismatch-containing sequences, the researchers found that amino acids shift their contact patterns in predictable ways. This allowed them to design targeted modifications—swapping amino acids at identified positions—that reduce off-target activity dramatically without losing on-target performance. The fact that the method worked for both Cas9 and Cas12 suggests it could be applied to other gene-editing systems. While other research teams have achieved similar or slightly better results using directed evolution, this computational method offers the potential to design systems tailored to specific guide RNAs and known mismatch patterns, opening a path toward personalized safety optimization in gene therapy development.

No comments yet. Be the first to share your thoughts!