A comprehensive benchmark for multiple highly efficient base editors with broad targeting scope.

As the toolbox of base editors (BEs) expands, selecting appropriate BE and guide RNA (gRNA) to achieve optimal editing efficiency and outcome for a given target becomes challenging. Here, we construct a set of 10 adenine and cytosine BEs with high activity and broad targeting scope, and comprehensively evaluate their editing profiles and properties head-to-head with 34,040 BE-gRNA-target combinations using genomically integrated long targets and tiling gRNA strategies. Interestingly, we observe widespread non-canonical protospacer adjacent motifs (PAMs) for these BEs. Using this large-scale benchmark data, we build a deep learning model, named BEEP (Base Editing Efficiency Predictor), for predicting the editing efficiency and outcome of these BEs. Guided by BEEP, we experimentally test and validate the installment of 3,558 disease-associated single nucleotide variants (SNVs) via BEs, including 20.1% of target sites that would be generally considered as "uneditable", due to the lack of canonical PAMs. We further predict candidate BE-gRNA-target combinations for modeling 1,752,651 ClinVar SNVs. We also identify several cancer-associated SNVs that drive the resistance to BRAF inhibitors in melanoma. These efforts benchmark the performance and illuminate the capabilities of multiple highly useful BEs for interrogating functional SNVs. A practical webserver (http://beep.weililab.org/) is freely accessible to guide the selection of optimal BEs and gRNAs for a given target.