Deep learning models simultaneously trained on multiple datasets improve base-editing activity prediction

CRISPR-derived base editors (BE) enable precise single nucleotide substitution without introducing double-stranded DNA breaks. Apart from the base editing enzymes, efficient base editing strongly depends on both the CRISPR guide RNA (gRNA) efficiency and the edited position. Here, we show that the accuracy of BE gRNA design can be significantly improved by generating more data and by introducing deep neural networks trained on multiple different datasets simultaneously. Generating ~20,000 gRNAs for A•T to G•C and C•G to T•A conversions, we present such deep learning models, which also allow users to do dataset-aware predictions. The methods are available online and as stand-alone software. CRISPR base editing enables the precise introduction of single-nucleotide mutations in the genome. Here, authors generated new adenine/cytosine base editor dataset and proposed a deep-learning model CRISPRon-BE for base editor efficiency prediction, thereby enhancing base editing applications.