Supplementary text: Predicting microRNA target sites using a combined target gene prediction approach

A variety of strategies has been proposed for computationally predicting microRNA target genes (see [1-4] for review). One important methodology uses complete annealing of microRNA and mRNA (similar to local alignments), as in miRanda [5], DIANA-microT [6], RNAhybrid [7] and the algorithm by Watanabe et al. [8]. Another methodology uses string matching of the seed region (potentially with an extension to a full alignment) as in microInspector [9], PicTar [10], TargetScan [11], TargetScanS [12] and the algorithms by Stark et al. [13]/Brennecke et al. [14]. These algorithms use specific parameters for filtering their predicted targets, such as score value cutoffs, experimental information coupled with empirical rules (e.g. number of G:U wobbles and the predicted bulge position), or annotation data (e.g. target function and tissue specificity). Therefore, any overlap observed between the predictions of different programs is relatively small [2, 4, 15].