AI-Engineered Synthetic Enzymes Match or Outperform Natural Nucleases

Engineering Synthetic Nucleases with ESM-IF1

Scientists have long relied on natural Cas proteins for genome editing, but these biological tools are often complex and difficult to modify without losing functionality. CRISPR systems are based on the machinery bacteria use to defend themselves against viruses, using a “guide RNA” to direct a nuclease to a target DNA sequence. The nuclease acts like molecular scissors to snip the targeted material. However, the multi-domain architecture of enzymes like Cas9 and Cas12 makes them fragile; the scientists noted that seemingly small changes can disrupt enzyme activity because the proteins rely on coordinated RNA and DNA recognition, activation, and cleavage by distinct conformational states.

To overcome these limitations, a research team including members from the Innovative Genomics Institute and the California Institute for Quantitative Bioscience—both at the University of California, Berkeley—along with collaborators at other institutions, employed artificial intelligence to design synthetic variants of TnpB, a minimal CRISPR-Cas12-like nuclease. While sequence-based biological language models have been used previously, they often produce proteins that closely resemble the reference sequences used to train them. To avoid this, Petr Skopintsev and colleagues utilized the ESM Inverse Folding (ESM-IF1) model to reverse-engineer protein sequences from desired three-dimensional structures.

By introducing evolution-informed residue constraints, the researchers ensured that the AI-generated variants maintained critical functional elements while allowing for significant sequence divergence from natural counterparts. This approach allowed the team to sample protein sequences that do not exist in nature, effectively enlarging the designable protein space. The scientists wrote that the results establish a strategy for creating non-natural RNA-guided nucleases and conformationally active nucleic acid binders.

Validation Across Bacterial, Plant, and Human Cells

The resulting synthetic nucleases, dubbed SynTnpBs, were engineered to remain RNA-guided and catalytically active despite being non-natural. These synthetic nucleases were screened across bacterial cells, plant cells, and human cells to determine if their theoretical designs translated into biochemical activity. The results indicated that many of these AI-designed enzymes retained or surpassed the activity of natural TnpB in multiple cell types, confirming their utility for practical genome-editing tasks. By showing that structure-guided protein design can yield genome-editing proteins with substantially different sequences while preserving function, the study extends the CRISPR toolbox.

Implications for the CRISPR Toolbox

The ability to design functional, non-natural nucleases marks a departure from traditional gene-editing development, which typically involves tweaking existing bacterial systems. Soeren Lienkamp, a molecular biologist at the University of Zurich in Switzerland who was not involved in the research, noted that the paper marries two transformative fields: AI-guided design and RNA-guided nucleases. He stated, Much like CRISPR democratized the ability to edit DNA at will, AI-based protein design promises to allow anyone to create totally novel properties in the protein space.

Current enzymes still produce off-target effects, and researchers have been exploring ways to extend the capabilities of these systems for more efficient gene editing. The new AI-based model was used to design a synthetic nuclease with less off-target activity. As the field advances, the success of SynTnpBs suggests that AI models could eventually allow for the design of enzymes with novel properties that go far beyond the defensive mechanisms currently co-opted from bacteria. The research team’s findings establish a strategy, offering a new pathway for scientists to engineer proteins that are better suited for medical and agricultural gene-editing needs.