(a) AIN-IP Transcripts are enriched for binding sites with extensive 5’ seed pairing. The horizontal axis is ordered according to final enrichment for increasing stringency in 5’ seed matches with an indicated number of G:U wobble pairs or a single bulge on the mRNA side of the duplex. The vertical axis shows the enrichment for seed matches at the indicated stringency in AIN-IP versus all other transcripts both before (light blue) and after (dark blue) implementation of the “site filter” (see Figure 1 and Supplemental Methods). Asterisks designate significant enrichments with P < 0.05. (b) AIN-IP Transcripts are enriched for binding sites that lie within structurally accessible regions. The horizontal axis shows the calculated accessibility of local sequence windows, either: across the entire binding site (red, dashed line) or within a 25 nucleotide window upstream of the binding site (light and dark blue, solid line) or downstream (green, dotted line). After applying the site filter, enrichment was calculated for upstream windows only (dark blue, solid line). c) AIN-IP transcripts are enriched for binding sites with favorable free energies. The set of conserved binding sites in AIN-IP transcripts are more likely to have favorable (i.e., negatively-valued) total hybridization energies than their counterparts in non-AIN-IP transcripts (light and dark blue lines). Also shown is hybrid energy (purple line), which reflects the stability of the final microRNA:target duplex, and corresponds to the minimal free energy (MFE). Enrichment for ∆G_total significantly increases after applying the site filter (dark blue line).

(a,b) Since the various comparison methods differ in degree of similarity, Venn diagrams were split into two groups. MiRanda, PicTar, and TargetScanS all use seed-matching and conservation in their prediction method, so part (a) shows the degree of overlap between mirWIP, Miranda, PicTAR, and TargetScanS. mirWIP selects most of the targets for PicTAR and TargetScanS that these two methods share, and relatively few of the targets not shared by these two methods. PITA and rna22, neither of which uses conservation to identify microRNA targets are compared to mirWIP in panel (b). (c) Most AIN-IP transcripts can be accounted for by containing conserved binding sites for known microRNAs. However, 29% of the IP-ed genes do not have strong, conserved binding sites in their annotated 3’ UTRs. Conserved binding sites can be found for an additional 8% of the AIN-IP transcripts, but these sites fail to meet our minimum free energy threshold and have been termed “weak” sites. Lack of conservation and poor UTR annotation are the most likely reasons for the rest of these non-predictions. See Results and Supplemental Results for a discussion of the remaining AIN-IP transcripts rejected by mirWIP.

Analysis of predicted microRNA binding sites in the 3’ UTR sequences of AIN-IP transcripts reveals enriched contextual features. An initial set of predicted microRNA sites was obtained and analyzed for enriched features, and these enriched features were used to score individual predicted binding sites (see Methods and Supplemental Methods). Binding site scores were then combined into total microRNA family scores for each target, which estimates the likelihood that a given transcript is regulated by a particular microRNA family. Finally, the microRNA family scores were combined into a total target score for each transcript, estimating the likelihood that a given transcript is regulated by microRNAs (see Results, Methods and Supplemental Methods sections).

(a) Choosing the optimum score threshold. The AIN-IP sensitivity of the algorithms is defined as the percent of AIN-IP targets correctly identified as a target of any microRNA (shown in red, above). The specificity of the algorithm is defined as the percentage of total predicted UTRs that are in the AIN-IP list (shown in blue, above). A compromise, for balancing the trade-off between sensitivity and specificity is defined by the point where the two curves meet: a mirWIP score threshold of 18. This corresponds to a sensitivity and specificity of approximately 40%. (b) Training mirWIP for AIN-IP sensitivity also optimizes true positive identification. A Receiver Operator Characteristic (ROC) curve is shown for mirWIP and five other microRNA prediction methods. The vertical axis shows the true positive rate (TPR), here represented by the number of verified targets correctly matched to the regulating microRNA. The horizontal axis shows the maximum false positive rate (FPR), the fraction of predicted UTRs that are not in the AIN-IP list. The performance of the mirWIP algorithm as a function of scoring threshold is shown as a blue line, and the 40% sensitivity/specificity compromise point (defined in panel a) is indicated by the large blue dot. mirWIP outperforms all five other methods by nearly doubling the TPR at a lower FPR (vertical gray line). (c) mirWIP is specific enough to reject all of the known false lsy-6 targets (Listed in Supplementary Table 1).