shRNAI: A deep neural network for the design of highly potent shRNAs

Abstract

miRNA-mimicking short hairpin RNAs (shRNAmirs), which depend on the endogenous miRNA biogenesis pathway, have been widely used to investigate gene function and to develop therapeutic strategies due to their stable and robust knockdown of target genes. However, despite the efforts to design potent shRNAmir guide RNAs (gRNAs), relevant biological features beyond the primary sequence have not been fully explored. Here, we present shRNAI, a convolutional neural network model for predicting highly potent shRNAmir gRNAs. Even when trained solely on gRNA sequences, shRNAI outperforms previous algorithms. We further improved the model (shRNAI+) by adding features related to shRNAmir processability and target site context, resulting in superior performance across both public datasets and our own experimental tests. Although shRNAI was initially trained on datasets built with a CNNC motif-free pri-miR-30 backbone, it also displayed improved performance on the CNNC motif. Overall, our study provides a robust framework for designing potent shRNAmir gRNAs, as well as a versatile tool for developing RNAi therapeutics.