Development of novel C:G-to-G:C base editors enabled by CRISPRi DNA repair screens and machine learning

Base editors are genome editing tools capable of installing targeted point mutations in the genomes of living cells. Two general classes of base editors have been described to date, cytosine base editors (CBEs) and adenine base editors, which together can install all four transition point mutations. Here we describe the development of engineered cytosine-to-guanine transversion base editors (CGBEs) that convert C:G base pairs to G:C base pairs. To develop these tools, we studied the mechanisms that lead to rare transversion byproducts of CBEs. We investigated the effects of various deaminase domains, Cas binding domains, and other CBE components, and performed a CRISPRi screen for the molecular determinants of C:G-to-G:C base editing outcomes to uncover the DNA repair proteins involved in generating these unusual mutations. Integrating these findings, we engineered a panel of CGBEs consisting of selected cytidine deaminase, Cas protein, and DNA repair protein components in a single fusion protein that show distinct editing characteristics, with some offering efficient and high-purity C:G-to-G:C editing at several tested genomic DNA sites in human cells. To maximize effective use of these reagents, we characterized the ability of six CGBE candidates to process 10,638 genomically integrated target sites in mouse embryonic stem (mES) cells, then trained machine learning models on the resulting data sets to enable researchers to predict whether their target site is amenable to editing by CGBEs and, if so, which CGBE should be used to maximize editing efficiency and product purity. We demonstrate that CGBEs are compatible with Cas9-NG, broadening their genome targeting scope, and compare CGBEs to prime editors on the same target sites, finding that CGBEs and prime editors offer complementary strengths and weaknesses. These efforts provide molecular and machine learning tools that expand the types of nucleotide conversions that can be achieved with base editing, and advance our understanding of cellular determinants of base editing outcomes.