AI-Designed Synthetic CRISPR Enzymes: Evolving Molecular Scissors Beyond Natural Constraints

Abstract

Therapeutic gene editing has historically relied on naturally evolved bacterial nucleases (SpCas9, Cas12a), imposing constraints: rigid protospacer-adjacent motif (PAM) requirements, protein sizes challenging adeno-associated virus packaging, and pre-existing human immunity. This narrative review assesses the paradigm shift from natural-product discovery toward generative artificial intelligence (AI) for engineering non-natural CRISPR-Cas nucleases with user-defined properties. This study synthesizes the design–build–test–learn pipeline encompassing protein language models (ESM3), diffusion-based structure generators (RFdiffusion, ProteinMPNN), and property-specific architectures (Protein2PAM). High-throughput platforms like Sequence Display generate >10 million sequence–activity data points per experiment for closed-loop validation. AI-generated OpenCRISPR-1 (>400 mutations from SpCas9) matches natural on-target efficiency while achieving sub-1% median off-target indels and up to 95% reduction at specific sites. Protein2PAM-evolved Nme1Cas9 variants show broadened PAM tolerance and up to 50-fold increased activity without iterative evolution. Generative AI addresses PAM restriction, packaging, and immunogenicity simultaneously. However, in vivo validation, long-term genotoxicity data, and biosecurity governance gaps remain unresolved.