Detection of features predictive of microRNA targets by integration of network data

Abstract

Gene activity is controlled by multiple molecular mechanisms, for instance through transcription factors or by microRNAs (miRNAs), among others. Established bioinformatics tools for the prediction of miRNA target genes face the challenge of ensuring accuracy, due to high false positive rates. Further, these tools present poor overlap. However, we demonstrated that it is possible to filter good predictions of miRNA targets from the bulk of all predictions by using information from the gene regulatory network. Here, we take advantage of this strategy that selects a large subset of predicted microRNA binding sites as more likely to possess less false-positives because of their over-representation in RE1 silencing transcription factor (REST)-regulated genes from the background of TargetScanHuman 7.2 predictions to identify useful features for the prediction of microRNA targets. These enriched miRNA families would have silencing activity for neural transcripts overlapping the repressive activity on neural genes of REST. We analyze properties of associated microRNA binding sites and contrast the outcome to the background. We found that the selected subset presents significant differences respect to the background: (i) lower GC-content in the vicinity of the predicted miRNA binding site, (ii) more target genes with multiple identical micro- RNA binding sites and (iii) a higher density of predicted microRNA binding sites close to the 3' terminal end of the 3'-UTR. These results suggest that network selection of miRNAmRNA pairs could provide useful features to improve microRNA target prediction.