Composite MicroRNA Target Predictions and Comparisons of Several Prediction Algorithms

Abstract

Motivation: MicroRNAs (or miRNAs) are short noncoding RNAs whose primary role is to repress translations by negatively regulating gene expressions post-transcriptionally through binding to their mRNA targets. Currently, there exist a few computational algorithms for human miRNA target predictions, but their results may vary widely. Therefore, it would be useful to consolidate and filter through these discrepant results so that researchers may have a greater degree of certainty in the validity of these targets before engaging in costly experiments. Results: We studied three of the most popular target prediction algorithms, miRanda, TargetScan, and PicTar, systematically through the use of three measures of similarity/distance and a statistical test on the gene ontology categories that are enriched in the target lists.Furthermore, two composite statistics were devised to combine and rank the composite target list. We applied the methods developed to all human miRNAs that had been identified. Our results indicate that TargetScan and PicTar tend to have a greater degree of similarity, based on all measures considered. We also demonstrate that our composite statistics and the program implementing them can be useful tools for filtering through large target sets to short list genes for downstream experiments.