Scientists have leveraged artificial intelligence models to create synthetic CRISPR proteins that edit the genome more efficiently than their natural counterparts. These synthetic CRISPR systems could one day drive discoveries in fields ranging from medicine to agriculture.

The results¹ were published on July 16 in Science.

“Just as CRISPR democratized the ability to edit DNA at will, AI-based protein design promises to allow anyone to create entirely new properties in the protein space,” says Soeren Lienkamp, ​​a molecular biologist at the University of Zurich in Switzerland, who was not involved in the research. He adds that the paper “combines two transformative fields”: AI-guided design and enzymes called RNA-guided nucleases, which can cut strands of DNA and RNA.

Cut, cut

These nucleases form the backbone of the gene editing system known as CRISPR, which uses a “guide RNA” to direct the nuclease to a target DNA sequence. The nuclease then acts like molecular scissors and cuts the target material, allowing scientists to edit, delete or add genetic information. CRISPR systems are based on the machinery that bacteria use to defend themselves against viruses. The most common CRISPR nucleases, such as Cas9 and Cas12, are obtained from bacteria.

As powerful a tool as gene editing is, the process is complex. Nucleases must complete a series of carefully orchestrated steps, making it difficult to go beyond what evolution has already produced, says Jennifer Doudna, a biochemist at the University of California, Berkeley, and lead author of the paper, who shared the Nobel Prize in Chemistry 2020 for his work on CRISPR systems. “Once you start modifying things, you quickly realize that while you can make changes, they end up producing something that is not functional.”

Former Meta scientists present gigantic AI protein design model

Artificial intelligence tools offer the opportunity to enhance the process of identifying promising candidates for new functional nucleases. Instead of running hundreds or even thousands of exploratory experiments, researchers could, in theory, ask machine learning to do it for them. To test this idea, Doudna and his collaborators focused on creating synthetic versions of a group of small nucleases called TnpB, which are evolutionary precursors to the commonly used Cas12. The scientists wanted to know to what extent they could alter protein sequences while preserving the proteins’ ability to edit genes.

For a protein to function, it often needs to assume a certain shape or conformation. The team began by providing an AI model with the final conformation of one type of TnpB and asking it to reverse engineer changes to the underlying DNA templates that would nonetheless maintain the final shape of the protein. This approach produced thousands of potential changes, but said nothing about whether the resulting protein would be active.