Fig 1.
Cancer gene-miRNA interactomes.
(A) Selection of 17 cancer genes involved in multiple cancer types. For each gene the number of interactions identified in the 3’ UTR reporter screening is listed. (B) Overview of 3’ UTR reporter screening results. Average interaction scores for all probed miRNA-3’ UTR combinations. (C) The miRNA interactome of TP53. (D) The miRNA interactome of MYCN.
Fig 2.
Interaction score performance.
(A) ROC-curve analysis of different metrics for high-throughput screening data-analysis on a set of positive and negative controls in the 3’ UTR reporter screening. Interaction scores and z-scores are calculated as described in the Methods section. B-scores are obtained by applying Tukey median polish to z-scores, in order to remove plate positional bias. Knockdowns are calculated by expressing normalized reporter activities (NRAs) relative to the average NRA of four non-targeting miRNA treated controls in the same assay plate. (B) Distributions of average interaction scores for positive and negative controls are clearly distinct. Application of the interaction score cutoff retrieves positive controls with 51% sensitivity, whereas negative controls are correctly called with a specificity of 99%. Precision obtained with this cutoff (i.e. the proportion of identified interactions that are true interactions) is 88%. Reprinted from Van Peer et al. [37] under a CC-BY 4.0 license, with permission from Oxford University Press, original copyright 2016.
Fig 3.
3’ UTR reporter screening reproducibility.
(A) Correlation of interaction scores from replicate 3’ UTR reporter screenings. (B) Correlation of interaction scores with a density profile, showing that the largest fraction of interaction scores is centered around 0. (C) Hierarchical clustering of miRNAs according to their activity in the 3’ UTR reporter screening. For each miRNA pair, the Pearson correlation between average interaction scores for all 17 cancer genes was calculated. Correlation vectors for all miRNAs are subsequently clustered using Euclidean distance as the distance measure. Members of the same miRNA family, in addition to families with identical or similar seed-sequences, cluster together. For the let-7 family, 9 out of 9 members cluster together. The mir-34 (2 out of 3 members) and the mir-449 family (2 out of 2 members) also cluster together. The only member not clustering (hsa-miR-34b) is the only one having a different, 1-nucleotide offset seed sequence. The mir-302 (4 out of 5 members) and mir-515 family (6 out of 32 members) cluster together with miRNAs with identical or 1-nucleotide offset seed sequences (red underline) such as hsa-miR-20b, hsa-miR-512-3p, hsa-miR-372, hsa-miR-373 and hsa-miR-17-5p. The mir-130 family (3 out of 3 members) clusters together with hsa-miR-454-3p that has an identical seed sequence (brown underline).
Fig 4.
Predicted and established interactions.
Cumulative distributions of average interaction scores for all 7990 miRNA-3’ UTR combinations probed. (A) according to the number of models that predict them as true interactions. Interaction scores are clearly lower for combinations that are predicted by more models. All distributions are significantly different from one another (one-sided Kolmogorov-Smirnov p-values < 0.01 after Benjamini-Hochberg multiple testing correction). (B) according to prediction by individual models. MirTarget2 predictions have the lowest scores. For each model, the distribution of interaction scores for predicted interactions is significantly different from that of non-predicted interactions (one-sided Kolmogorov-Smirnov p-values < 0.01 after Benjamini-Hochberg multiple testing correction). (C) according to whether they have previously been established as true interactions or not. Previously established interactions clearly have lower interaction scores. Distributions are significantly different (one-sided Kolmogorov-Smirnov p-value < 0.001).
Fig 5.
3’ UTR reporter rescue experiment.
Rescue of 3’ UTR reporter regulation. (A) For four MYCN interactions significant down-regulation of reporter activity after miRNA expression modulation can no longer be demonstrated upon canonical binding site mutation (one-sided t-test; p < 0.001 ***; p < 0.01 **; wt = wild-type 3’ UTR; mut = mutant 3’ UTR) in two independently replicated reporter experiments. Reporter activity is expressed relative to non-targeting miRNA treated controls (NTC). Error bars represent standard deviations on three technical replicates. Successful rescue of MYCN regulation could only be achieved in one experiment for hsa-miR-494. (B) For four hsa-miR-449 interactions significant down-regulation of reporter activity after miRNA expression modulation can no longer be demonstrated upon canonical binding site mutation (one-sided t-test; p < 0.001 ***; p < 0.01 **; wt = wild-type 3’ UTR; mut = mutant 3’ UTR) in two independently replicated reporter experiments. Reporter activity is expressed relative to non-targeting miRNA treated controls (NTC). Error bars represent standard deviations on three technical replicates. Successful rescue of regulation by hsa-miR-449 could only be achieved in one experiment for MYB.
Fig 6.
Cumulative distributions of log2 relative expression levels of endogenous mRNAs measured with RT-qPCR after miRNA modulation. The distribution for interactions identified in the 3’ UTR reporter screening is significantly lower than that for miRNA-3’ UTR combinations for which no interaction was observed (one-sided Kolmogorov-Smirnov p-value < 0.001).
Fig 7.
Canonical binding site potency.
Cumulative distributions of average interaction scores for all 7990 miRNA-3’ UTR combinations probed. (A) according to the presence of canonical binding site patterns. Combinations with multiple canonical binding site patterns have lower interaction scores than combinations with a single pattern, that in their turn have lower scores than combinations without canonical binding site patterns. All distributions are significantly different from one another (one-sided Kolmogorov-Smirnov p-values < 0.001 after Benjamini-Hochberg multiple testing correction). (B) according to the presence of different types of canonical binding site patterns. Combinations with at least one 8mer pattern, have lower interaction scores than combinations with at least one 7mer-m8, one 7mer-A1 and one 6mer pattern, respectively (combinations with multiple types of binding site patterns are considered in all respective distributions). Notably, the presence of multiple 8mer patterns produces the largest shift in distribution. All distributions are significantly different from one another (one-sided Kolmogorov-Smirnov p-values < 0.01 after Benjamini-Hochberg multiple testing correction). (C) according to the presence of 3’ supplementary pairing. Combinations harboring canonical binding site patterns with 3’ supplementary pairing have lower interaction scores than those without. All distributions are significantly different from one another (one-sided Kolmogorov-Smirnov p-values < 0.001 after Benjamini-Hochberg multiple testing correction).
Fig 8.
Non-canonical binding site potency.
Cumulative distributions of average interaction scores for all 7990 miRNA-3’ UTR combinations probed. (A) according to the presence of offset 6mer binding site patterns. Combinations with at least one offset 6mer pattern, have lower interaction scores than combinations without. Distributions are significantly different (one-sided Kolmogorov-Smirnov p-value < 0.001). (B) according to the presence of seed-mismatched or G:U wobble binding site patterns. Combinations with at least one seed-mismatched or G:U wobble pattern have lower interaction scores than combinations without. Distributions are significantly different (one-sided Kolmogorov-Smirnov p-value < 0.01). (C) according to the presence of G-bulge binding site patterns. Combinations with G-bulge patterns don’t have detectably lower interaction scores than combinations without. Distributions are not significantly different (one-sided Kolmogorov-Smirnov p-value > 0.05). (D) according to the presence of offset 6mer binding site patterns with 3’ supplementary pairing. Combinations harboring offset 6mer patterns with 3’ supplementary pairing have lower interaction scores than those without. All distributions are significantly different from one another (one-sided Kolmogorov-Smirnov p-values < 0.001 after Benjamini-Hochberg multiple testing correction). (E) according to the presence of seed-mismatched or G:U wobble binding site patterns with 3’ supplementary pairing. Combinations harboring seed-mismatched or G:U wobble patterns with 3’ supplementary pairing have lower interaction scores than those without. All distributions are significantly different from one another (one-sided Kolmogorov-Smirnov p-values < 0.05 after Benjamini-Hochberg multiple testing correction). (F) according to the presence of G-bulge binding site patterns with 3’ supplementary pairing. Combinations harboring G-bulge patterns with 3’ supplementary pairing have lower interaction scores than those without. All distributions are significantly different from one another (one-sided Kolmogorov-Smirnov p-values < 0.05 after Benjamini-Hochberg multiple testing correction).