Artificial intelligence is already great at predicting how proteins fold. And soon, AI might become great at designing new proteins from scratch.
A group of researchers led by Nobel laureate Jennifer Doudna at the Innovative Genomics Institute and the University of California, Berkeley, has now inched us a step forward by using AI to design new RNA-guided nucleases based on the TnpB family of CRISPR-Cas12-like proteins (Science 2026, DOI: 10.1126/science.aed6123).
“We were curious whether modern generative models are capable of designing as complex systems as RNA-guided nucleases,” says structural biologist Petr Skopintsev, a member of the Doudna Lab and one of the lead authors of the new paper. This isn’t the first time AI has been used to create novel gene editors, but the approach is different.
Skopintsev says the research group used a hybrid AI approach when designing its TnpB variants, dubbed SynTnpBs. To design the SynTnpBs, the team defined which parts of the protein sequence need to remain fixed according to evolutionary data from the TnpB family. The researchers combined the fixed sequence information with inverse protein-folding models, in which AI tries to generate sequences that result in specific protein structures. And they split the design process in two, separately creating AI variants of the DNA-binding interface and the guide RNA–binding interface.
Coming up with novel protein sequences is one challenge, but testing them is another. “Computationally you could be designing millions of sequences, but you’re always going to be limited by what experimental method you have to test them at scale,” says Isabel Esaín-Garcia, a biochemist in the Doudna Lab and a lead author of the paper.
Roughly 50 of the best candidates for each nucleic acid–binding interface were combined in all permutations in Escherichia coli and screened for editing activity. Esaín-Garcia says the top candidates were further assessed for editing activity in human and plant genomes. Some candidates were then characterized structurally, using methods including cryo-electron microscopy.
The results show that some SynTnpBs have editing capabilities equal to and even greater than that of wild type TnpBs. The SynTnpB with the most editing activity shared only 77% sequence identity with wild type TnpB. And the research team was even able to identify a conformational state of TnpBs that had previously been proposed but never observed.
“Attempting to marry the strengths of structural and evolutionary data is a growing trend in AI for protein engineering,” Eli Bixby, cofounder and head of machine learning at Cradle, an AI protein design company, tells C&EN in an email. “In particular this work attempts to address a central weakness of both structural and sequence models within this space—that they tell you what changes to make but not where to make them.”
The SynTnpBs themselves may not have broad applicability in gene-editing research, but the authors say this paper brings us closer to bespoke, AI-generated proteins. “We live in a world where we’re moving towards personalized medicine, and we think the possibility of being able to create [enzymes with] your own tailored properties are very important for that,” Esaín-Garcia says.
Max Barnhart is an assistant editor and life sciences reporter at C&EN.