Scientists at UC Berkeley have developed a new AI-assisted method to design custom genome-editing enzymes that go beyond what evolution has created in nature. In a study published in Science, researchers showed that AI-generated variants of a model nuclease performed as well as or better than naturally occurring enzymes across bacterial, plant, and human cells, opening the door to personalized gene-editing tools tailored for specific diseases.

Why Can't Scientists Just Use Enzymes Found in Nature?

Since the discovery of CRISPR-Cas9 in 2012, scientists have hunted through microbial genomes searching for new DNA-cutting enzymes to expand their gene-editing toolkit. This approach has worked, but it comes with significant limitations. Evolution doesn't explore the full range of possible designs; the process is slow, constrained by what came before, and relies on chance discoveries. Researchers at the Innovative Genomics Institute reasoned that artificial intelligence could break through these natural boundaries.

"Usually, when we find new systems, people find them in nature and then they do protein engineering to further enhance them," explained Petr Skopintsev, a postdoctoral researcher leading the work. "We proposed that we should be able to extend even further beyond what nature has designed for these proteins, and potentially even program in a controlled way for properties that we want to have in terms of activity, specificity, et cetera."

Petr Skopintsev, Postdoctoral Researcher, Innovative Genomics Institute

How Did Researchers Train AI to Design Better Enzymes?

The challenge with RNA-guided nucleases like Cas9 is their complexity. These proteins must simultaneously interact with both a guide RNA that targets a specific DNA sequence and the DNA itself, while using flexible structures to recognize and cut their targets. Earlier AI attempts failed because language models trained on protein sequences tended to generate enzymes too similar to their training data.

The research team took a different approach by combining two complementary AI techniques. They used an "inverse folding model," technology developed by researchers at Meta AI that works like the reverse of AlphaFold. While AlphaFold predicts a protein's three-dimensional structure from its genetic sequence, inverse folding does the opposite: given a desired structure, it generates sequences predicted to fold into that shape. The team then added evolutionary constraints to model how the generated sequences would interact with nucleic acids.

What Were the Results of This Hybrid AI Approach?

The team tested their method using TnpB, a hypercompact nuclease similar to CRISPR-Cas12. The results were striking: nearly 1 in 4 of the roughly 2,000 different proteins they designed turned out to be active nucleases when produced in the laboratory. This success rate far exceeded previous AI-design attempts. More importantly, the AI-generated variants showed high divergence from natural enzymes while maintaining or improving their editing activity.

When tested in living cells, the newly designed enzymes worked across multiple organisms. The team demonstrated activity in bacterial cells, plant cells, and human cells, proving the approach could have applications beyond human medicine. For one of the most divergent variants, researchers even solved its complete three-dimensional structure, making it the first ever solved structure for an AI-designed functional RNA-guided nuclease.

Steps to Customize Gene Editors for Specific Applications

• Define Target Properties: Researchers can now specify desired characteristics such as editing speed, specificity for particular DNA sequences, or stability at different temperatures, rather than accepting whatever nature provides.
• Use Inverse Folding Models: AI systems generate protein sequences predicted to fold into structures optimized for the desired properties, incorporating evolutionary constraints to ensure functionality.
• Test at Scale: The pipeline allows researchers to produce and test hundreds or thousands of variants rapidly, identifying the most promising candidates for further development and real-world use.
• Validate Across Cell Types: Successful designs must be tested in multiple biological systems, bacterial, plant, and human cells, to confirm broad applicability and safety.

"Beyond this one protein, more importantly we established the pipeline, the set of methods to generate proteins at scale," said Skopintsev. "People can take this and apply this method for other systems."

Petr Skopintsev, Postdoctoral Researcher, Innovative Genomics Institute

What Could This Mean for Medicine and Agriculture?

The ability to design custom enzymes on demand could transform how gene therapies are developed. Instead of waiting to discover new enzymes in nature or spending months engineering existing ones, researchers could rapidly generate tailored tools for specific diseases. Isabel Esain-Garcia, another lead researcher on the project, emphasized the potential for personalized approaches.

"In the future, when we think about personalized medicine and how we have to rapidly generate new genome-editing enzymes tailored to different diseases, this type of approach where there are a lot of custom properties that can be designed quickly would be beneficial," said Esain-Garcia.

Isabel Esain-Garcia, Postdoctoral Researcher, Innovative Genomics Institute

The implications extend beyond human health. The same approach could optimize gene editors for agricultural applications, allowing scientists to design enzymes that work best in crop plants or livestock. Potential customizable properties include activity levels, cutting speed, specificity to different nucleic acids, and thermal stability for use in different environments.

The research represents a fundamental shift in how scientists approach protein design. Rather than being limited by what evolution has already explored, AI now enables researchers to rationally design new proteins with properties tailored to immediate needs. As this pipeline becomes more widely adopted, the pace of gene-editing tool development could accelerate dramatically, bringing personalized genetic medicine closer to clinical reality.