DeepsmirUD: Precise prediction of regulatory effects on miRNA expression mediated by small molecular compounds using competing deep learning frameworks

Abstract

Aberrant miRNA expression has pervasively been found to relate to a growing number of human diseases. Therefore, targeting miRNAs to regulate their expression levels has become an important therapy against diseases that stem from the dysfunction of oncogenic pathways regulated by the miRNAs. In recent years, small molecule compounds have demonstrated enormous potential as drugs to regulate miRNA expression (i.e., SM-miR). A clear understanding of the mechanism of action of small molecules on down- and up-regulating miRNA expression allows precise diagnosis and treatment of oncogenic pathways. However, outside of a slow and costly process of experimental determination, computational strategies to assist this in an ad hoc manner have still not been enabled. In this work, we develop, to the best of our knowledge, the first prediction tool, DeepsmirUD, to infer small molecule-mediated regulatory effects on miRNA expression. This method is powered by an ensemble of 12 cutting-edged deep learning frameworks and achieves state-of-the-art performance with AUC values of 0.840/0.969 and AUCPR values of 0.866/0.983 on two independent test datasets. With a complementarily constructed network inference approach based on similarity, we report a significantly improved accuracy of 0.813 in determining regulatory effects of nearly 650 SM-miR relations formed with either novel small molecules or novel miRNAs. By further integrating miRNA-cancer relations, we established a database of potential drugs associated with cancers to aid in explaining the drug mechanisms of action in diseases and offering novel insight into drug repositioning. Taken together, our method shows great promise to assist and accelerate the therapeutic development of potential miRNA targets and small molecule drugs. Furthermore, we have employed DeepsmirUD to predict the regulatory effects of a large number of high-confidence SM-miR relations curated from Psmir, which are publicly available through https://github.com/2003100127/deepsmirud and https://rujinlong.shinyapps.io/deepsmirud/ alongside the DeepsmirUD standalone package.