Discovery of miRNA-Gene regulatory networks by using an integrative data-mining approach

Introduction MicroRNAs (miRNAs) represent the largest class of small non-coding RNAs (20-24 nucleotide long) acting as post-transcriptional regulators of many genes and playing a pivotal role in important biological processes, in almost all organisms and in a large number of human diseases. Computational approaches have been proven to be fundamental in the miRNA research for both gene-specific and large-scale predictions of miRNA targets, for the formulation of new functional hypothesis on their biological role, for gene network discovery and to guide experimental validations. However, their effectiveness is negatively affected by high uncertainty of miRNA gene target predictions and by the complexity of rules governing miRNA functional targeting, whose mechanisms still remain elusive. In order to improve predictions of miRNA targets and to support the elucidation of miRNA functional role in the context of gene regulatory networks, we have recently developed a new two-stepped computational approach based on: i) a semi-supervised ensemble-based classifier for the prediction of miRNA target interactions (MTIs) [1] and, ii) a biclustering algorithm (HOCCLUS2) for the prediction of miRNA-gene regulatory networks (MGRNs) [2]. Data produced are available at ComiRNet, a user-friendly web-based system providing efficient query, retrieval, export, visualization and analysis of predicted MTIs and MGRNs. Method In the first step, a semi-supervised ensemble-based classifier is learned from both experimentally validated interactions (positively labelled examples), extracted from miRTarBase [3], and miRNA gene target predictions (MTIs), returned from several prediction algorithms (unlabelled examples) and extracted from mirDIP [4]. This classifier acts as a meta-classifier of unlabelled examples. As a result of the first step, a unique (meta-)prediction score is available for all possible interactions. In the second step, these prediction scores are used to identify miRNA-gene regulatory networks (MGRNs) through the biclustering algorithm HOCCLUS2. HOCCLUS2 exploits the large set of produced predictions, with the associated probability, to extract a set of overlapping and hierarchically organized biclusters each one representing putative MGRNs. The construction of the hierarchy is performed by an iterative merging, considering both distance and density-based criteria. Extracted biclusters are also ranked on the basis of the p-values obtained by the Student's T-Test which compares intra- and inter- functional similarity of miRNA targets, computed on the basis of the gene classification provided in Gene Ontology (GO) [5]. The ComiRNet database relies on PostgreSQL DBMS, while the web-based platform is built through the Play 2.2 Java framework and the Cytoscape library. Results The effectiveness of the computational approach has been validated on a number of alternative combinations of competitive algorithms for the first [1] and the second step [2]. Both the predicted MTIs and the MGRNs can be queried, retrieved, exported and visualized through the web-based system ComiRNet (Co-clustered miRNA Regulatory Networks). Based on the principles of our approach, genes in a bicluster are likely to function together as a network and miRNAs in the same bicluster are likely to cooperatively target groups of networked genes. The use of computational predictions in place of only experimentally validated interactions offers the possibility to detect single interactions and regulatory modules that would be otherwise impossible to reconstruct by considering only experimentally validated interactions, which are strictly dependent on the cell type and experimental conditions used. This paves the way to the systematic use of ComiRNet for i) a comprehensive anal