AlphaFold AIを活用して遺伝子編集タンパク質を再設計、安全性を向上Team uses AlphaFold AI to redesign gene-editing proteins to make them safer

Researchers have reportedly used DeepMind's AlphaFold protein-structure prediction system to redesign the proteins at the heart of CRISPR gene editing, aiming to reduce so-called off-target effects that have long complicated efforts to bring genome editing safely into the clinic. The work matters because unintended edits at the wrong sites in the genome remain one of the central safety concerns for gene therapies, and computational protein design offers a route to engineering more precise tools without the slow, expensive trial and error of purely laboratory-based approaches.

CRISPR systems typically pair a guide RNA, which matches a target DNA sequence, with a nuclease protein such as Cas9 that cuts the DNA at that location. The precision of an edit depends heavily on how tightly the protein and guide discriminate between the intended sequence and similar-looking sequences elsewhere in the genome. When the machinery tolerates mismatches, it can cut at unintended locations, potentially disrupting healthy genes or introducing harmful mutations. Reengineering the nuclease so that it binds and cleaves only when the match is exact is therefore a direct way to improve safety.

According to the reported work, the team used AlphaFold to model how candidate protein variants would fold and interact with DNA, then used those predictions to guide which amino acid changes to introduce. Because AlphaFold can generate structural predictions far faster than experimental methods such as X-ray crystallography or cryo-electron microscopy, it allows researchers to screen many hypothetical designs computationally before committing to bench experiments. The approach appears to combine structural prediction with iterative testing, using the model to narrow a vast space of possible sequences down to a smaller set of promising candidates that are then validated in the lab.

It is worth noting the distinction between structure prediction and de novo protein design. AlphaFold, first released by DeepMind in 2020 and dramatically expanded with AlphaFold2 and later AlphaFold3, was built primarily to predict the three-dimensional shape of a protein from its amino acid sequence. Designing entirely new or heavily modified proteins is a related but separate challenge, and tools such as the University of Washington's RoseTTAFold and RFdiffusion, developed in David Baker's lab, are aimed more explicitly at generating novel structures. In practice, researchers increasingly combine predictive and generative models, and using AlphaFold to evaluate or refine engineered variants of an existing protein like Cas9 fits that emerging pattern.

The broader context is a wave of interest in what is sometimes called AI-driven biology. The 2024 Nobel Prize in Chemistry was awarded in part for AlphaFold and for computational protein design, signaling how central these methods have become to the field. Gene editing itself reached a regulatory milestone in late 2023, when Casgevy, a CRISPR-based therapy for sickle cell disease and beta thalassemia, was approved in the United States and the United Kingdom. That approval demonstrated the clinical promise of the technology while also underscoring why safety refinements remain valuable, particularly as developers pursue in vivo editing, where the tool is delivered directly into the body rather than applied to cells outside it.

Reducing off-target activity is an active area beyond this specific study. Earlier engineered variants such as high-fidelity Cas9 and enhanced-specificity Cas9 were produced through more conventional protein engineering, and base editors and prime editors were developed to make more controlled changes with fewer double-strand breaks. Computational design tools could complement these efforts by making it faster to explore modifications and to understand why certain changes improve specificity.

Several caveats apply. Structural predictions, however accurate, do not fully capture the dynamic behavior of proteins as they search and bind DNA, so laboratory and eventually animal testing remain essential to confirm that a redesigned enzyme is both safer and still effective. Improvements measured in cell-based assays do not automatically translate into clinical benefit, and any therapeutic use would require extensive further validation and regulatory review. The reported reduction in off-target effects is likely an encouraging proof of concept rather than a finished therapeutic product.

Even so, the work illustrates a growing convergence between machine learning and molecular biology, in which predictive models increasingly inform the design of the biological tools themselves. If the approach holds up under wider scrutiny, it could accelerate the development of more precise gene-editing systems and offer a template for applying structure-prediction AI to other protein-engineering problems.