A new algorithm for integrated analysis of miRNA-mRNA interactions based on individual classification reveals insights into bladder cancer

Abstract

Background: MicroRNAs (miRNAs) are small non-coding RNAs that regulate gene expression. It has been proposed that miRNAs play an important role in cancer development and progression. Their ability to affect multiple gene pathways by targeting various mRNAs makes them an interesting class of regulators.

Methodology/principal findings: We have developed an algorithm, Classification based Analysis of Paired Expression data of RNA (CAPE RNA), which is capable of identifying altered miRNA-mRNA regulation between tissues samples that assigns interaction states to each sample without preexisting stratification of groups. The distribution of the assigned interaction states compared to given experimental groups is used to assess the quality of a predicted interaction. We demonstrate the applicability of our approach by analyzing urothelial carcinoma and normal bladder tissue samples derived from 24 patients. Using our approach, normal and tumor tissue samples as well as different stages of tumor progression were successfully stratified. Also, our results suggest interesting differentially regulated miRNA-mRNA interactions associated with bladder tumor progression.

Conclusions/significance: The need for tools that allow an integrative analysis of microRNA and mRNA expression data has been addressed. With this study, we provide an algorithm that emphasizes on the distribution of samples to rank differentially regulated miRNA-mRNA interactions. This is a new point of view compared to current approaches. From bootstrapping analysis, our ranking yields features that build strong classifiers. Further analysis reveals genes identified as differentially regulated by miRNAs to be enriched in cancer pathways, thus suggesting biologically interesting interactions.

Figure 1. Flow chart for identifying differentially regulated interactions.

Input data is depicted by orange rectangles. Output data is indicated by red rectangles. The ellipse refers to the set of inferred interactions. This set is independent of the input data, though it can be changed. Operations to manipulate data are depicted as diamonds.

Figure 2. Spread of samples inside the bladder cancer data set.

For the collective of all samples (A) and the collective of tumor samples (B). The first two principal components (PC) of the distance matrix based on interaction states are shown. Circles refer to samples of the control group, triangles are invasive tumor samples and crosses refer to samples of non-invasive tumors. The first two principal components explain 85.39% of the total variance (PC1 = 78.12% and PC2 = 7.27%) for (A) and 48.35% of the total variance (PC1 = 33.36% and PC2 = 14.99%) for (B).

Figure 3. Hierarchical clustering of the bladder cancer data set based on interaction states.

For the whole collective of samples (A) and the collective of tumor samples (B) using interaction states computed by algorithm. “N” refers to samples of the control group. Tumor samples were labeled by their pathological staging (non-invasive: pTa or invasive: pT1 and higher).

Figure 4. Mean specificities and sensitivities for the bladder cancer data set using our algorithm CAPE RNA.

Models for the whole collective (A) and only for tumor samples (C) were generated from training sets by selecting all interactions with a Jaccard-index equal to or higher than a threshold. A and C illustrate the mean specificities and mean sensitivities for the models to classify an unknown test set. B (generated from the whole collective) and D (only tumor samples) shows the average number of interactions included in a model.

Figure 5. Mean specificities and sensitivities for the bladder cancer data set using pamr.

Prediction Analysis of Microarrays for R was used to train models for the collective of all samples based on miRNA expression (A) and mRNA expression (C), as well as for the collective of tumor samples based on miRNA (B) and mRNA (D). Models were generated with different thresholds. A-D) illustrate the mean specificities and mean sensitivities to classify unknown test sets.

Figure 6. Distribution of Jaccard-indexes for the bladder cancer data set.

For the whole collective (normal and tumor samples) (A) and the group of tumor samples (invasive and non-invasive tumors) (B). Jaccard-indexes were calculated using our approach. Only Jaccard-indexes corresponding to interactions which show a negative correlation, i.e. ρ≤−0.4, for at least one experimerntal group between normalized miRNA and mRNA expression values are depicted.

Figure 7. FGF3R pathway: detected miRNA-mRNA interactions in normal urothelium (N), non-invasive- (nT) and invasive (iT) bladder tumor tissue samples.

Interactions between miRNAs (triangles) and mRNAs (boxes), which have a Jaccard-index ≥0.4 that exhibit a negative correlation, i.e. ρ≤−0.4, between the normalized miRNA and mRNA expression values for at least one experimental group, are shown (blue edges). The black edges indicated the general signal cascade. Mean expression status of the analysed miRNA mRNAs interactions were indicated (red indicates up regulation and green down regulation).

Figure 8. KEGG bladder cancer pathway: calculated miRNA-mRNA interactions differences between normal urothelium (N) and non-invasive- (nT) and invasive (iT) bladder tumor tissue samples.

Interactions between miRNAs (triangles) and mRNAs (boxes), which have a Jaccard-index ≥0.4 that exhibit a negative correlation, i.e. ρ≤−0.4, between the normalized miRNA and mRNA expression values for at least one experimental group, are shown (blue edges). The black edges indicated the general signal cascade adopted from KEGG bladder cancer pathway. Mean expression status of the analysed miRNA-mRNA interactions are indicated (red indicates up regulation and green down regulation).

Figure 9. Venn diagram of predicted miRNA-mRNA interactions in bladder cancer derived from five different methods of integrative analysis (TaLasso, GenMiR++, Spearman correlation, a Meta analysis approach and our new algorithm CAPE RNA).

Two different collectives were analysed: (A) the entire bladder cancer dataset of normal (n = 8) and tumor samples (n = 16) and (B) only the collective of invasive (n = 8) and non-invasive (n = 8) bladder cancer tumor samples. The Venn diagram illustrates the intersection between the top 500 predicted miRNA-mRNA interactions by each method.