Survey of Computational Algorithms for MicroRNA Target Prediction

MicroRNAs (miRNAs) are 19 to 25 nucleotides non-coding RNAs known to possess important post-transcriptional regulatory functions. Identifying targeting genes that miRNAs regulate are important for understanding their specific biological functions. Usually, miRNAs down-regulate target genes through binding to the complementary sites in the 3' untranslated region (UTR) of the targets. In part, due to the large number of miRNAs and potential targets, an experimental based prediction design would be extremely laborious and economically unfavorable. However, since the bindings of the animal miRNAs are not a perfect one-to-one match with the complementary sites of their targets, it is difficult to predict targets of animal miRNAs by accessing their alignment to the 3' UTRs of potential targets. Consequently, sophisticated computational approaches for miRNA target prediction are being considered as essential methods in miRNA research. We surveyed most of the current computational miRNA target prediction algorithms in this paper. Particularly, we provided a mathematical definition and formulated the problem of target prediction under the framework of statistical classification. Moreover, we summarized the features of miRNA-target pairs in target prediction approaches and discussed these approaches according to two categories, which are the rule-based and the data-driven approaches. The rule-based approach derives the classifier mainly on biological prior knowledge and important observations from biological experiments, whereas the data driven approach builds statistic models using the training data and makes predictions based on the models. Finally, we tested a few different algorithms on a set of experimentally validated true miRNA-target pairs [1] and a set of false miRNA-target pairs, derived from miRNA overexpression experiment [2]. Receiver Operating Characteristic (ROC) curves were drawn to show the performances of these algorithms.