AI designs functional CRISPR-like nucleases surpassing natural models

Artificial intelligence is beginning to redesign the molecular machinery that powers genome editing. A new Science report describes how researchers created synthetic, RNA-guided nucleases that rival—or outperform—natural enzymes. The work extends the CRISPR toolbox by showing that structure-guided protein design can yield genome-editing proteins with substantially different sequences while preserving function.

CRISPR-Cas systems work by pairing an RNA guide with a DNA target and then using a nuclease domain to cut. Most widely used technologies rely on natural Cas proteins shaped by evolution. But whether AI can generate functional equivalents with novel sequence features, rather than near replicas of known enzymes, has been a difficult question.

Petr Skopintsev and colleagues tackled that challenge using a strategy built around ESM Inverse Folding (ESM-IF1). Instead of allowing the model to drift toward sequences close to training references, the team introduced evolution-informed residue constraints designed to keep key functional elements in place while still permitting large sequence divergence.

Their starting point was TnpB, a minimal CRISPR-Cas12-like nuclease. The researchers designed new variants they call SynTnpBs, which are engineered to remain RNA-guided and catalytically active despite being non-natural. This “minimal nuclease” context is especially important because multi-domain architectures can be fragile—small changes can destroy activity.

After design, the synthetic nucleases were screened across bacterial cells, plant cells, and human cells. Many of the AI-generated enzymes retained or exceeded the performance of the natural TnpB across these distinct environments, supporting the idea that the designs were not merely theoretical but biochemically robust.

To understand how these divergent proteins work, the authors turned to cryo-electron microscopy (cryo-EM). They determined structures for the most divergent SynTnpB variants, providing the first experimentally resolved pictures of AI-designed RNA-guided nucleases.

Across multiple conformations, cryo-EM revealed new stabilizing interactions at the RNA-DNA interface. Rather than matching the natural enzyme’s exact contact patterns, the engineered nucleases appear to solve the same recognition problem through alternative molecular geometry—an outcome consistent with the promise of protein design guided by structure.

Together, the study demonstrates a practical path for creating active genome-editing enzymes that do not simply imitate nature’s sequences. By combining inverse folding with evolution-informed constraints and validating function in real cellular systems, the work suggests that AI-driven design can generate new classes of editing tools.