RIP-Chip analysis supports different roles for AGO2 and GW182 proteins in recruiting and processing microRNA targets

Abstract

Background: MicroRNAs (miRNAs) are small non-coding RNA molecules mediating the translational repression and degradation of target mRNAs in the cell. Mature miRNAs are used as a template by the RNA-induced silencing complex (RISC) to recognize the complementary mRNAs to be regulated. To discern further RISC functions, we analyzed the activities of two RISC proteins, AGO2 and GW182, in the MCF-7 human breast cancer cell line.

Methods: We performed three RIP-Chip experiments using either anti-AGO2 or anti-GW182 antibodies and compiled a data set made up of the miRNA and mRNA expression profiles of three samples for each experiment. Specifically, we analyzed the input sample, the immunoprecipitated fraction and the unbound sample resulting from the RIP experiment. We used the expression profile of the input sample to compute several variables, using formulae capable of integrating the information on miRNA binding sites, both in the 3'UTR and coding regions, with miRNA and mRNA expression level profiles. We compared immunoprecipitated vs unbound samples to determine the enriched or underrepresented genes in the immunoprecipitated fractions, independently for AGO2 and GW182 related samples.

Results: For each of the two proteins, we trained and tested several support vector machine algorithms capable of distinguishing the enriched from the underrepresented genes that were experimentally detected. The most efficient algorithm for distinguishing the enriched genes in AGO2 immunoprecipitated samples was trained by using variables involving the number of binding sites in both the 3'UTR and coding region, integrated with the miRNA expression profile, as expected for miRNA targets. On the other hand, we found that the best variable for distinguishing the enriched genes in the GW182 immunoprecipitated samples was the length of the coding region.

Conclusions: Due to the major role of GW182 in GW/P-bodies, our data suggests that the AGO2-GW182 RISC recruits genes based on miRNA binding sites in the 3'UTR and coding region, but only the longer mRNAs probably remain sequestered in GW/P-bodies, functioning as a repository for translationally silenced RNAs.

Keywords: RIP-Chip data analysis; RISC proteins AGO2 and GW182; microRNA regulatory activity; microRNA target prediction.

Fig. 1

RIP-Chip experiments overview. a and c Western Blot analysis of proteins immunoprecipitated and co-immunoprecipitated with anti-AGO2 or anti-GW182 antibody (IP). IgGs in a) are the negative controls. IN and FT made up 1% of the cytoplasmic lysate used for each IP sample. GW182 was specifically co-immunoprecipitated with AGO2 (c, left panel), and AGO2 was specifically co-immunoprecipitated with GW182 (c, right panel). b Enrichment analysis of seven highly expressed miRNAs in anti-AGO2 and anti-GW182 IP compared to IgG-IP controls. d Average Linkage Cluster analysis of mRNA and miRNA expression profiles of IP, IN and FT samples from three independent experiments; distance is computed as 1- Correlation (Pearson). AGO2-IP and GW182-IP mRNA expression profiles are highlighted in blue and green, respectively. In mRNA expression clustering, we considered all the 16,323 genes with a detected expression level in the samples considered. In miRNA expression clustering, we considered 508 miRNAs with a detected expression level in at least one sample

Fig. 2

Behavior overview of variables listed in Table 1. a Heatmap representation of the correlation block matrix of the variables computed with AGO2_IN1 miRNA and mRNA expression profiles. The reported numbers are the correlation values, expressed in the range [− 100:100]. b ROC-AUC values obtained by classifying enriched/underrepresented genes associated with the variables computed with each IN expression profile. c Wilcoxon test p-values (log10) obtained by comparing the variable values associated with the enriched/underrepresented gene sets. In both b) and c), the variables computed with the three AGO2 IN profiles were used to distinguish enriched and underrepresented genes in AGO2-IP vs FT. The variables computed with the three GW182 IN profiles were used to distinguish enriched and underrepresented genes in GW182-IP vs FT.

Fig. 3

Graphic representation of selected features values associated to enriched and underrepresented genes. Empirical cumulative distribution function (ECDF) of F6, F4d, F8, L1 and L2 variables computed for enriched (UP) and underrepresented (LOW) genes in AGO2 IP vs FT and GW182 IP vs FT analyses. The reported p-values were obtained by performing a Wilcoxon-test comparing the values assumed by the selected set of genes with the values assumed by all the genes (16,363, green lines)

Fig. 4

Graphic representation of the effect of miRNA expression profile shuffling. Each boxplot represents the AUC values obtained with 1000 simulated miRNA expression profiles. The percentage on the right of each boxplot refers to the number of times an AUC value was greater than the AUC obtained with the original miRNA expression profile (red vertical line). a Performance of simulated F6 variables in distinguishing AGO2 enriched/underrepresented genes. b Performance of simulated F6 variables in distinguishing GW182 enriched/underrepresented genes. c Performance of simulated F4d variables in distinguishing AGO2 enriched/underrepresented genes. d Performance of simulated SVM models (F6&F4d variables) in distinguishing AGO2 enriched/underrepresented genes. e Performance of simulated F8 variables in distinguishing AGO2 enriched/underrepresented genes. f Performance of simulated F8 variables in distinguishing GW182enriched/underrepresented genes

Fig. 5

Support Vector Machine models performance summary. AUC values of SVM models trained with any pair of variables defined in Table 1, used to classify enriched/underrepresented genes in AGO2-IP vs FT comparison. Variables were computed by using the AGO2_IN1 expression profiles. Values are in the range [0:100]. Values in the diagonal refer to single variable performance. The ROC plot at bottom left represents the results obtained with the best-performing SVM model (F6&F4d, black line) and with the two single variables, F6 (red line) and F4d (green line)