Comparative miRNA Target Prediction

To compare predictions from different miRNA target prediction tools, we collected the following freely downloadable predictions: AnTar (predictions from either miRNA-transfection or CLIP-seq models) (Wen et al., 2011 (link)), DIANA-microT-CDS (September 2013) (Reczko et al., 2012 (link)), ElMMo v5 (January 2011) (Gaidatzis et al., 2007 (link)), MBSTAR (all predictions) (Bandyopadhyay et al., 2015 (link)), miRanda-MicroCosm v5 (Griffiths-Jones et al., 2008 (link)), miRmap v1.1 (September 2013) (Vejnar and Zdobnov, 2012 (link)), mirSVR (August 2010) (Betel et al., 2010 (link)), miRTarget2 (from miRDB v4.0, January 2012) (Wang, 2008 (link); Wang and El Naqa, 2008 (link)), MIRZA-G (sets predicted either with or without conservation features and either with or without more stringent seed-match requirements, March 2015) (Gumienny and Zavolan, 2015 (link)), PACCMIT-CDS (sets predicted either with or without conservation features) (Marin et al., 2013 (link)), PicTar2 (from the doRiNA web resource; sets conserved to either fish, chicken, or mammals) (Krek et al., 2005 (link); Anders et al., 2012 (link)), PITA Catalog v6 (3/15 flank for either ‘All’ or ‘Top’ predictions, August 2008) (Kertesz et al., 2007 (link)), RNA22 (May 2011) (Miranda et al., 2006 (link)), SVMicrO (February 2011) (Liu et al., 2010 (link)), TargetRank (all scores from web server) (Nielsen et al., 2007 (link)), TargetSpy (all predictions) (Sturm et al., 2010 ), TargetScan v5.2 (either conserved or all predictions, June 2011) (Grimson et al., 2007 (link)), and TargetScan v6.2 (either conserved predictions ranked by the context+ model or all predictions ranked by either the context+ model or P_CT scores, June 2012) (Friedman et al., 2009 (link); Garcia et al., 2011 (link)). For algorithms providing site-level predictions (i.e., ElMMo, MBSTAR, miRSVR, PITA, and RNA22), scores were summed within genes or transcripts (if available) to acquire an aggregate score. For algorithms providing multiple transcript-level predictions (i.e., miRanda-MicroCosm, PACCMIT-CDS, and TargetSpy), the transcript with the best score was selected as the representative transcript isoform. In all cases, predictions with gene symbol or Ensembl ID formats were translated into RefSeq format. When computing r² to the test sets, mRNAs that were not predicted by the algorithm to be a target were assigned the worst score in the range of all scores generated by the algorithm.