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

Abstract

AbstractmiRNA-mimicking short hairpin RNA (shRNAmir), which depends on whole miRNA biogenesis, has been used to elucidate the function of target genes and to develop therapeutic approaches due to its stable, robust downregulation. Despite efforts to design potent shRNAmir guide RNAs (gRNAs), biological features other than the sequence have not been fully considered. Here, we developed shRNAI, a convolutional neural network model for the prediction of highly potent shRNAmir gRNAs. The shRNAI model trained with gRNA sequences alone predicted more potent shRNAmir gRNAs than the existing algorithms. The shRNAI+ model, trained with the additional features of shRNAmir processibility and target site context, was further improved on both public and our experimental test datasets. Although the shRNAI models were trained on datasets generated with the CNNC motif-free shRNAmir backbone, they also exhibited better performance for the CNNC motif. Our study provides not only a rational framework for designing shRNAmir gRNAs for targets but also a means of designing optimal RNA interference drugs necessary for RNA therapeutics.