Fig. 4

Performances of different target-prediction algorithms in flies. a The differential ability of algorithms to predict the mRNAs most responsive to miRNAs transfected into Drosophila cells. Shown for each algorithm in the key are mean mRNA fold changes observed for top-ranked predicted targets, evaluated over a sliding sensitivity threshold using the six miRNA transfection datasets. Some methods, such as PicTar, which generated relatively few predictions, could be evaluated at only a few thresholds, whereas others, such as RNA22 and TargetSpy, could be evaluated at many more. For each algorithm, predictions for each of the six miRNAs were ranked according to their scores, and the mean fold-change values were plotted at each sensitivity threshold. For example, at a threshold of 16, the 16 top predictions for each miRNA were identified (not considering predictions for mRNAs expressed too low to be accurately quantified). mRNA fold-change values for these predictions were collected from the cognate transfections, and the mean fold-change values were computed for each transfection for which the threshold did not exceed the number of reported predictions. The mean of the available mean values was then plotted. Also plotted are the mean of mean mRNA fold changes for all mRNAs with at least one cognate canonical 7–8-nt site in their 3′ UTR (dashed line), the mean of mean fold change for all mRNAs with at least one conserved cognate canonical 7–8-nt site in their 3′ UTR (dotted line) and the 95% confidence interval for the mean fold changes of randomly selected mRNAs, determined using 1000 resamplings (without replacement) at each cutoff (shading). Sites were considered conserved if their branch-length scores exceeded a cutoff with a signal:background ratio of 2:1 for the corresponding site type (cutoffs of 1.0, 1.6, and 1.6 for 8mer, 7mer-m8, and 7mer-A1 sites, respectively; Fig. 2b). Thresholds at which the distribution of fold changes for predicted targets of the context model was significantly greater than that of any other model are indicated (*, one-sided Wilcoxon rank-sum test, P value < 0.05). See also Additional file 2: Figure S4. b The differential ability of algorithms to predict the mRNAs most responsive to knocking out miRNAs in flies. Shown for each algorithm in the key are mean mRNA fold changes observed for top-ranked predicted targets, evaluated over a sliding sensitivity threshold using the three knockout datasets. Otherwise, this panel is as in a. c and d The differential ability of algorithms to predict targets that respond to the miRNA despite lacking a canonical 7–8-nt 3′ UTR site. These panels are as in a and b, except they plot results for only the predicted targets that lack a canonical 7–8-nt site in their 3′ UTR. Results for our context model and other algorithms that only predict targets with canonical 7–8-nt 3′ UTR sites are not shown. Instead, results are shown for a 6mer context model, which considers only the additive effects of 6mer, offset 6mer, and 6mer-A1 sites and their corresponding context features. e and f The difficulty of predicting mRNAs that respond to miRNA transfection or knockout despite lacking canonical 6–8-nt 3′ UTR sites. These panels are as in c and d, respectively, except they plot results for mRNAs with 3′ UTRs that lack a canonical 6–8-nt site