Multitask regression for condition-specific prioritization of miRNA targets in transcripts
Deregulation of miRNAs is implicated in many diseases in particular cancer, where miRNAs can act as tumour suppressors or oncogenes. As sequence-based miRNA target predictions do not provide condition-specific information, many algorithms combine expression data for miRNAs and genes for prioritization of miRNA targets. However, common strategies prioritize miRNA-gene associations, although a miRNA may only target a subset of the alternative transcripts produced by a gene. Thus, current approaches are suboptimal. Here we address the problem of transcript and not gene based miRNA target prioritization. We show how to leverage methods that were developed for gene expression based miRNA-target prioritization for transcripts. In addition, we introduce a new multitasking based learning (MTL) method that uses structured-sparsity inducing regularization to improve accuracy of the learning. The new MTL approach performs especially favorable in small sample size settings, for genes with many transcripts and with noisy transcript expression level estimates as shown with simulated data. In an analysis of real liver cancer RNA-seq data we show that the MTL approach better predicts transcript expression and outperforms simpler approaches for miRNA-target prediction.