Computational prediction of microRNA targets: thermodynamic, probabilistic and evolutionary models parameterized by genome-scale experimental data

MicroRNAs, or miRNAs, post-transcriptionally repress the expression of protein-coding genes. The human genome encodes over 1000 miRNA genes that collectively target the vast majority of messenger RNAs (mRNAs). Base-pairing of the so-called miRNA "seed" region with mRNAs identifies many thousands of putative targets. Evaluating the strength of the resulting mRNA repression remains challenging, but is essential for a biologically informative ranking of potential miRNA targets. To address these challenges, predictors may employ thermodynamic, evolutionary, probabilistic, or sequence-based features. We developed an open source software library, miRmap, which for the first time comprehensively covers all four approaches using eleven predictor features, three of which are novel. This allowed us to examine feature correlations and to compare their predictive power in an unbiased way using high throughput experimental data from immunopurification, transcriptomics, proteomics and polysome fractionation experiments. Overall, target site accessibility appears to be the most predictive feature. Our novel feature based on PhyloP, which evaluates the significance of negative selection, is the best performing predictor in the evolutionary category. We combined all the features into an integrated model that almost doubles the predictive power of TargetScan