AI model streamlines prime editing

When scientists use prime editing to develop new gene-editing medicines or for research, the prime editing guide RNA, or pegRNA, they use can greatly impact editing performance. The pegRNA is an important component of the prime editing system — it instructs the prime editor both where to edit and what edit to make at the target site. But there are often hundreds or thousands of possible pegRNA configurations to choose from for any given edit.

Evaluating them all in the lab can be cost-prohibitive, so researchers at the Broad Institute have come up with a better solution using AI. Members of the David Liu lab have developed OptiPrime, a machine learning model that can predict the performance of pegRNAs based on their sequence and prioritize those predicted to work best, saving the costly and time-consuming step of generating and testing hundreds of pegRNAs in the lab.

The scientists demonstrated that OptiPrime could identify pegRNAs that improved prime editing efficiency, and used it to rapidly design a prime editor that precisely corrected a pathogenic mutation in mice. Described in Nature Biotechnology and available free for non-commercial use online, OptiPrime allows researchers to more effectively use prime editing in their work.

The tool will also support efforts, such as the recently launched Center for Therapeutic Genetics, to more rapidly develop new gene-editing medicines for rare diseases.

“Our goal with OptiPrime is to make prime editing as easy to use as possible for other researchers, who might not be gene-editing experts and who want to do fewer experiments to develop prime editing solutions that meet their needs,” said David Liu, core institute member, Richard Merkin Professor, vice chair of the faculty, and director of the Merkin Institute of Transformative Technologies in Healthcare at the Broad Institute of MIT and Harvard.

The work was also led by co-first authors Alvin Hsu, Peter Chen, and Angus Li.

Building in the biology

The Liu lab isn’t the first to use machine learning to predict how well various pegRNAs will work. However, OptiPrime is the first AI to incorporate knowledge they and others gleaned about the biological mechanisms of prime editing — the key steps that determine prime editing outcomes.

With OptiPrime, the team embedded these insights into the mathematical structure of the model itself, so that it predicts the rates of individual biochemical steps involved with each pegRNA. OptiPrime integrates those rates to score thousands of pegRNAs based on their sequences and reveal the most promising ones.

The researchers trained OptiPrime using hundreds of thousands of data points from experiments in their lab and other labs, then tested the model on new data, demonstrating its ability to identify the best pegRNAs for making a desired edit.

Surprisingly, OptiPrime was also able to make accurate predictions about prime editing variants it had never been trained on. “This finding is encouraging because it supports our hope that the model benefits from knowledge of the mechanism of prime editing,” said Liu.

To test the model’s potential for therapeutic applications, they used OptiPrime to rapidly nominate pegRNAs that could efficiently edit pathogenic mutations back to healthy DNA sequences, both in cells and in a mouse model of a severe human neurological disorder. In mice suffering neurological disease from mutations in Kif1a, treatment with prime editing systems that took only 4 weeks to optimize resulted in 40% bulk brain cortex correction of Kif1a, in vivo editing efficiencies the team anticipates to be therapeutically relevant. “After a few weeks of optimization from starting pegRNAs proposed by OptiPrime, we achieved efficient editing levels that normally take months or longer to achieve,” Liu said.

The team is hopeful that OptiPrime and the versatility of prime editing, which can correct the vast majority of known mutations that cause genetic disorders, will similarly accelerate the development of strategies to efficiently treat other genetic misspellings that cause thousands of diseases — a critical need for customized, on-demand gene-editing medicines.