Figure 1.
miRNA biogenesis and gene silencing.
miRNA biogenesis begins with transcription of a pri-miRNA which is processed into a hairpin, exported from the nucleus, processed into a mature miRNA, and incorporated into a RISC. Minimal functional RISCs consist of an Argonaute protein and a miRNA. RISCs hybridize to mRNAs at targeting sites complementary with the miRNA, and prevent translation by either cleaving the mRNA or maintaining its position at the target site and stopping ribosomal activity.
Figure 2.
Marginal expression level comparisons of known target pairs.
Marginal correlations of m- and miRNA expression levels are typically and inappropriately positive in all datasets under analysis. Further, in the Madison and Broad datasets the amount of variation in targeted mRNA expression captured by that of targeting miRNA is extremely low. In the combined dataset, high R2 values are indicative only of the amount of variation in mRNA expression due to data origin.
Figure 3.
Marginal and system biology-compensated E2F1 and miR-17-5p expression levels.
E2F1 and miR-17-5p expression levels in the Madison data are compared marginally in the top left and after compensation for biological and idiosyncratic covariates on the top right. Analogous results for the OSU data are provided in the bottom row. In either dataset, marginal RNA expression levels are not representative of the known targeting relationship between miR-17-5p and E2F1. After compensation the relationship between these RNAs can be observed.
Table 1.
Identification performances of marginal and regression models.
Table 2.
Significances of numbers of correct identifications under randomization nulls.
Figure 4.
System biology-compensated expression level comparisons of known target pairs.
Improvements in the percentages of variation in targeted mRNA expression are observed through use of the adjusted R2 statistic. Partial correlations of proxies to targeting Ago 2 RISCs and targeted mRNAs are lower than those of m/miRNA expressions. Paired Wilcoxon tests of the hypothesis H0: μ0−μ1 = 0 vs. HA: μ0−μ1≠0 were rejected for the Madison (p = 0.0135), Broad (p = 0.0186) and combined (p<0.0001) datasets, where μ0 and μ1 refer to mean correlations of marginal m/miRNA expression and mean partial correlations respectively.
Figure 5.
p-values of observed correct identifications under no-targeting and random pair nulls.
The AIC-optimal procedure compensating for biological and idiosyncratic covariates yields numbers of correct identifications that are typically at least marginally significant in any of the datasets under study, using either randomization null. In no case is the number of correct identifications achieved by marginal correlation significantly greater than what would be expected.
Figure 6.
Estimated effects of targeting Ago 2 and Ago 1,3,4 RISC proxies.
X- and y-axes correspond to estimated effects of targeting RISC proxies composed of Ago 2 (β1) and Ago 1, 3 or 4 (β5) on targeted mRNA levels for those cases where the model successfully identifies the known targeting pair and estimates both effects in the minimum AIC submodel. Observed values are inversely related in sign with similar magnitudes of effect strength. This is consistent with the hypothesized interference of observable target cleavage by RISCs composed of Ago 2 by those composed of Ago 1,3 or 4.
Figure 7.
Validation of computationally predicted targets.
Hash marks denote the 95% confidence level of identification numbers under the no-targeting null for the miRNA and validation technique under consideration. Overall, 7.83% of computational predicted targets were verified using marginal expression level comparisons, and 4 miRNAs showed substantially larger numbers of verifications than what would be typically expected. After compensating for biological and idiosyncratic effects, 25.68% of these targets were verified, and 9 miRNAs showed large numbers of identifications.
Figure 8.
Estimated effects of marginal miR-29c expressions and miR-30d Ago 2 RISC proxies.
Distributions of miR-30d Ago 2 RISC proxy effects across all genes measured in the Madison dataset suggest two classes of genes – those with no relationship to miR-30d, and those negatively regulated by miR-30d. Analogous distributions of marginal miR-29c expression effects suggest no such dichotomy, and are negatively biased.