AI-Driven Inverse Design Yields Non-Natural Molecular Scissors, Enhancing CRISPR Gene Editing Efficiency
Background
The CRISPR gene editing system originated from the bacterial defense mechanism against viral attacks. Currently, it relies on identifying and utilizing nucleases like Cas9 or Cas12 from natural microbial genomes. However, these natural enzymes suffer from limitations such as low cellular delivery efficiency and off-target effects. Attempts to improve their performance through protein sequence modification often result in complete loss of activity due to minor changes. The intricate mechanism of molecular scissors makes it challenging to achieve novel, customized enzyme designs using conventional methods.
Key Findings
A team led by Jennifer Doudna at the University of California, Berkeley, developed an AI-based approach to design non-natural gene editing enzymes. They targeted the nuclease 'TnpB,' which has a small molecular size, making it advantageous for intracellular delivery. They employed an AI model called 'ESM-IF1,' which reverse-engineers protein 3D structures to infer amino acid sequences. By integrating evolutionarily conserved sequence conditions essential for maintaining function into the algorithm, they significantly improved the success rate of the design process.
Out of approximately 2,000 designed candidates, experiments confirmed that about 25% possessed functional gene editing activity. The synthesized artificial protein variants are named 'SynTnpB.' SynTnpB variants exhibited superior gene editing performance compared to wild-type enzymes in human, plant, and bacterial cells. Subsequently, using cryo-EM, they precisely determined the 3D structure of highly active variants with significant structural differences from the wild-type sequence. Analysis revealed that the artificial scissors uniquely established a novel stabilizing interaction network at the RNA-DNA junction, which is not present in natural enzymes. This demonstrates that AI can autonomously create non-natural molecular structures optimized for function.
Significance and Prospects
In the commercialization of gene therapies, a major challenge has been the lack of delivery vehicles capable of safely transporting large gene editing enzymes into cells. The widely used Cas9 enzyme is too large to be loaded entirely into adeno-associated virus (AAV) vectors, which have limited genomic capacity. In contrast, SynTnpB is reduced to less than half the size of Cas9, making it easier to load into a single AAV vector along with the therapeutic gene. This provides a suitable alternative to overcome delivery limitations and maximize efficiency.
However, before applying this technology clinically, it is necessary to proactively address potential immunogenicity issues. The human immune system may recognize non-natural amino acid sequences as foreign and mount an immune response. Therefore, extensive validation steps, including toxicity testing and in vivo stability assessment in various animal models, such as non-human primates, are essential.
Nature, Published online: 16 July 2026; doi:10.1038/d41586-026-02217-wResearchers used artificial intelligence to design functional CRISPR enzymes not seen in nature.
This research provides a practical means to accelerate the clinical translation of gene editing technology, which has been confined to the laboratory setting. A prominent application area is the treatment of inherited retinal degeneration and rare neurological disorders. In these diseases, the ability to safely deliver gene editing enzymes to specific target cells in a localized area is crucial for therapeutic success. Previously, a complex technique involving splitting a large gene editing enzyme into two parts and delivering them separately before reassembling them inside the cell was used. However, the introduction of the ultra-small SynTnpB allows for the precise targeting and correction of desired gene sequences using a single AAV vector.
From an industrial perspective, it enables the establishment of a system for the large-scale design of customized gene editing enzymes. Pharmaceutical companies can leverage AI models to instantly derive amino acid sequences that meet optimal binding structures and operating conditions when targeting specific disease genes, paving the way for a customized design service. This is expected to create an efficient research and development environment, reducing the candidate discovery period from several years to just a few months.