Cupid: simultaneous reconstruction of microRNA-target and ceRNA networks

MicroRNAs (miRNAs) regulate RNA stability and mRNA translation (Filipowicz et al. 2008) and their dysregulation has been implicated in a wide range of human diseases including cancer (Garzon et al. 2009). Consequently, establishing accurate and comprehensive repertoires of miRNA–target interactions is a necessary step toward elucidating their mechanistic role in pathophysiology. Dissecting miRNA regulation, however, has proven challenging because candidate miRNA binding sites are ubiquitous and their regulatory effects are context specific (Liu et al. 2005; Lu et al. 2005; Mukherji et al. 2011). As a result, and despite their relatively low accuracy, computational prediction methods that incorporate context-specific data are preferred for screening for miRNA–target interactions in tumor contexts (Carroll et al. 2013; Erhard et al. 2014). To address these challenges, we introduce Cupid, an integrative framework for the context-specific inference of miRNA targets. Cupid integrates sequence-based evidence and functional clues derived from RNA and miRNA expression analysis, predicting candidate miRNA binding sites and associated target genes using ensemble machine learning classifiers that are trained on validated interactions. Candidate interactions emerging from this step are then refined based on independent, context-specific clues, including their predicted ability to mediate competitive endogenous RNA (ceRNA) interactions, where mRNA compete for shared miRNA regulators (Fig. 1A; Tay et al. 2014). Thus, Cupid simultaneously infers both interaction types (ceRNA and miRNA–target interactions). In addition, we considered evidence for combinatorial regulation by multiple miRNA species (Fig. 1B; Boissonneault et al. 2009; Xu et al. 2011) and for indirect miRNA regulation through effector proteins (Fig. 1C). Taken individually, these clues are predictive of bona fide miRNA–target interactions and can significantly improve the tradeoff between precision and recall. Figure 1. Methodology. (A) Cupid first reevaluates sites predicted by TargetScan, miRanda, and PITA, selecting and rescoring each candidate site (Step I). Sites are used to select and score miRNA-target interactions (Step II), which are then examined for evidence ... We show that Cupid predictions outperform other leading algorithms, based on multiple experimental assays, including PAR-CLIP data, miRNA perturbation followed by mRNA and protein expression profiles, and 3′ luciferase activity assays. Critically, while Cupid predicts fewer interactions than other methods (Fig. 1D–F), its predictions are much more likely to be consistent with experimental evidence. This is critical since high false-positive prediction rates are a key limitation of current miRNA–target prediction methods.