Optimal protospacer sequences recommended by ensemble deep learning for high-efficiency base editing

Identification of an optimal protospacer at a disease-associated gene is critically important for the design of a high-efficiency base editor that does not have much bystander effect to edit this gene. Current machine learning methods are prone to overestimating the editing efficiencies of those low-efficiency editors, leading to wrong recommendations of high-efficiency protospacers; meanwhile they made separate predictions for editing efficiency and outcome proportion using two independent models, giving rise to performance inconsistency and confusing the identification of optimal protospacers. We propose an ensemble of paired convolutional neural networks for accurate prediction of outcome proportions and then we derive editing efficiencies directly from the proportions. Our method is able to significantly reduce the performance inconsistency between editing efficiency and editing proportion predictions caused by the two-model approach. Our method generalizes well to work on a range of different editing platforms. Furthermore, our method recommends optimal protospacers by ranking the candidates’ picking score which we newly defined as a harmonic indexing score integrating both of on-target editing efficiency and bystander editing effect.