Recently, microRNAs (miRNA), small noncoding RNAs, have taken center stage in the field of human molecular oncology. However, their roles in tumor biology remain largely unknown. According to the assumption that miRNAs implicated in a specific tumor phenotype will show aberrant regulation of their target genes, we introduce an approach based on the miRNA target–dysregulated network (MTDN) to prioritize novel disease miRNAs. Target genes have predicted binding sites for any miRNA. The MTDN is constructed by combining computational target prediction with miRNA and mRNA expression profiles in tumor and nontumor tissues. Application of the proposed method to prostate cancer reveals that known prostate cancer miRNAs are characterized by a greater number of dysregulations and coregulators and the tendency to coregulate with each other and that they share a higher proportion of targets with other prostate cancer miRNAs. Support vector machine classifier, based on these features and changes in miRNA expression, is constructed and gives an average overall prediction accuracy of 0.8872 in cross-validation tests. The classifier is then applied to miRNAs in the MTDN. Functions enriched by dysregulated targets of novel predicted miRNAs are closely associated with oncogenesis. In addition, predicted cancer miRNAs within families or from different families show combinatorial dysregulation of target genes, as revealed by analysis of the MTDN modular organization. Finally, 3 miRNA target regulations are verified to hold in prostate cancer cells by transfection assays. These results show that the network-centric method could prioritize novel disease miRNAs and model how oncogenic lesions are mediated by miRNAs, providing important insights into tumorigenesis. Mol Cancer Ther; 10(10); 1857–66. ©2011 AACR.

MicroRNAs (miRNA) are endogenous noncoding (∼22 nucleotides) RNAs that regulate the expression of target mRNAs at the posttranscriptional level (1). It was recently shown that miRNAs potentially regulate the majority of all human genes, and an individual miRNA is capable of regulating dozens of distinct mRNAs whereas multiple miRNAs can coregulate an individual gene (2, 3). miRNAs play critical roles in various biological processes such as cell proliferation (4), signal transduction (5), development (6), and apoptosis (7). Thus, abnormal miRNA functions can have a profound impact on multiple features of cell biology, ultimately resulting in complex pathologic events including cancers.

An increasing number of studies are exploring miRNA functions in normal physiologic processes and in disease states (8–10). miRNA roles in many types of cancers have typically been investigated using an expression profiling approach (11–13) in which miRNAs with differential expression in tumor and normal tissues are detected. This is one of the most widely used analysis methods in screens for potential tumor markers. However, limited attempts have been made to search for reliable methods for identifying gene targets and the functional significance of miRNAs in cancer. Liu and colleagues described a method to identify functional mRNA targets of altered miRNAs from expression profiles in clear cell renal cell carcinoma (14). A cancer–miRNA network was developed by mining the literature for experimentally verified cancer–miRNA relationships, and the authors found that miRNAs in the same cancer–miRNA module have a number of common predicted target genes, which suggests a combinatorial effect of miRNAs within a module on target regulation (15). Therefore, the influences of miRNAs on cell pathology and physiology are likely to be complex because of target multiplicity and miRNA cooperativity.

Most studies have focused on analyzing the functional and topological properties of miRNAs in the same or a similar disease. In a previous study, we showed that disease miRNAs have more synergism, reflecting their higher functional complexity, and that miRNAs associated with the same disease are close to each other in the miRNA–miRNA functional synergistic network and regulate targets with the same or similar functions (16). Volinia and colleagues used all expressed miRNAs as the input to build a global miRNA expression network for individual cancer types and found that different cancers involve separate and unlinked miRNA subsets and that the functions of these subnetworks are involved in many cancer-related events (17). Some important patterns between miRNAs and human diseases are also revealed by integrating published miRNA–human disease associations (18, 19).

Prostate cancer is one of the most commonly diagnosed malignant tumors. Recent studies have shown that miRNAs are significantly altered in prostate cancer (20–22). However, collation of data on known prostate cancer miRNAs validated by low-throughput methods revealed that in many studies, only a small fraction of validated miRNAs were probed for their influence on prostate cancer. Therefore, investigation of the importance of miRNAs in prostate cancer remains a challenge.

To further our understanding of miRNAs in cancers, we explored the regulatory changes from miRNAs to their direct targets rather than the differential properties of individual miRNAs by integrating miRNA and mRNA expression levels. Direct target genes are required to have predicted binding sites for any miRNA. We hypothesized that miRNAs implicated in a specific tumor phenotype will show aberrant regulations of their target genes. Biologically this is quite plausible, as various oncogenic events are manifested as gains or losses of regulatory interaction capability. Here, we introduce a network-centric approach that focuses on dysregulations of miRNAs to prioritize novel disease miRNAs. Our method can consider details of how miRNA behavior has changed and the specific mechanisms that lead to the pathologic transition.

### Data

In the text, we used the expression profiles reported by Ambs and colleagues (GSE8126 for miRNAs and GSE6956 for mRNAs; ref. 20). Samples simultaneously assayed for miRNA and mRNA expression were analyzed, and there were 60 primary prostate tumors and 15 nontumor prostate tissue samples. We normalized mRNA data using the Robust Multi-array Average (RMA) algorithm (23) and directly downloaded the normalized miRNA data. miRNA and mRNA expression intensities were then log transformed (base 2), and miRNA precursors were mapped to mature miRNAs using miRBase (24).

Candidate miRNA–target relationships were predicted by 1 or more of the following target prediction algorithms (union set): TargetScan 5.1 (25), miRanda (miRBase version 5; ref. 24), PicTar (4-way; ref. 3), and DIANA-microT (version 3.0, the default loose score threshold; ref. 26). We used Entrez gene IDs to represent corresponding genes. In total, there are 484,043 regulations between 320 miRNAs and 12,713 target genes.

### Overview of prostate cancer miRNA prioritization

The procedure comprises 3 distinct stages. First, we constructed the miRNA target–dysregulated network (MTDN) using differences in correlation within tumor and nontumor subgroups. Second, to prioritize prostate cancer miRNAs, we built gold standard data sets containing prostate cancer miRNAs and non–prostate cancer miRNAs and defined 4 topological features for miRNAs in the MTDN that significantly differ between prostate cancer and non–prostate cancer miRNAs. Finally, a support vector machine (SVM) was trained on the gold standard data sets and used to prioritize novel prostate cancer miRNAs.

### Constructing the MTDN

We analyzed each miRNA–target relationship in sequence to determine those showing aberrant behavior in prostate cancer. First, the miRNA and mRNA expression profiles were split into tumor and nontumor subgroups. For each regulation by miRNA i of target j, we calculated the change in Pearson's correlation coefficient according to the following equation:

where M and T denote expression for miRNA i and target gene j, respectively, C is the subgroup of cancer samples and A is the subgroup of normal prostate tissue samples. Summations are over the number nc or nA of samples in each subgroup. $\overline {M_{\rm c}}$⁠, $\overline {T_{\rm c}}$⁠,$\,\overline {M_{\rm A}}$⁠, and TA are the average expression levels of miRNA i and target j in the tumor and nontumor subgroups, respectively, and SMASTA and SMASTA are the products of the SDs for miRNA and target expression for the 2 subgroups, respectively. This measure gives an estimate of the degree of dysregulation by an miRNA of a target between the tumor and nontumor subgroups.

To determine whether the deviation in correlation between the 2 groups is significant, we randomly reassigned patients to the 2 groups 10,000 times and recalculated the Dys. Therefore, the P value for regulation by a miRNA of a target was given as the frequency of the absolute values of the random Dys being greater than the absolute value of the real Dys divided by 10,000. We controlled for multiple hypotheses using the false discovery rate (FDR), and only pairs passing an FDR of 0.01 were considered to be significantly dysregulated (27).

After assembling all significant miRNA–target pairs, we generated the MTDN. This network is a bipartite graph containing 2 sets of vertices (nodes) corresponding to miRNAs and target genes. A directed edge (connection) from an miRNA to 1 of its targets exists if their relationship is significantly dysregulated, and the weights of all edges are set to 1.

### Building the training set

We constructed gold standard data sets for prostate cancer and non–prostate cancer miRNAs as follows. Prostate cancer miRNAs were defined as miRNAs associated with prostate cancer as detected by low-throughput methods such as Northern blotting and quantitative reverse transcriptase PCR. Information was obtained by querying the miR2Disease database (release 2010 Jan 18; ref. 28) and manually reading studies published before May 13, 2010, as revealed by a search of the PubMed database and Google Scholar using the search term “prostate cancer and microRNA.” After mapping to miRNA expression profiles, we identified 37 prostate cancer miRNAs (Supplementary Table S1).

Compiling a list of non–prostate cancer miRNAs is currently difficult or even impossible. Several studies have referred to genes with the highest expression in a specific tissue as tissue-expressed genes (29, 30), so it is reasonable to assume that the 50 miRNAs with lowest expression in normal prostate tissue are the least involved in prostate cancer. After filtering 6 prostate cancer miRNAs from this set of 50, the remaining 44 miRNAs were defined as non–prostate cancer miRNAs.

### Defining the feature set

For each miRNA in the MTDN, we defined 4 measures for assessing topological properties (Table 1). Dout is the number of links from node i (number of dysregulated targets for miRNA i). NmiRNA is the number of its coregulators, which are miRNAs that share at least 1 target with miRNA i in the MTDN including itself. Rpc-miRNA is the proportion of prostate cancer miRNAs in its coregulator set. The measure of Rtarpc-miRNA designates the fraction of targets which are coregulated by miRNA i and other prostate cancer miRNAs. Besides these 4 topological measurements, we also considered the fold change for expression of miRNA i.

Table 1.

Topological feature set and medians of the topological features between prostate cancer miRNA set and non–prostate cancer miRNA or general background miRNA set

FeatureDescriptionMedianPaPb
Prostate cancerNon–prostate cancerBackground
Dout The number of dysregulated target genes 10 0.0865 0.0104
NmiRNA The number of its coregulators 11 5.5 7.5 0.0150 0.0587
Rtarpc-miRNA The proportion of prostate cancer miRNAs in its coregulator set 0.2381 0.0727 0.1087 1.7E-8 <0.0001
Rtarpc-miRNA The fraction of targets which are coregulated by itself and other prostate cancer miRNAs 0.1667 0.0486 0.0976 0.0303 0.0554
Fold_C Fold changes in expression 0.8942 1.0837 0.9419 0.00013 0.9309
FeatureDescriptionMedianPaPb
Prostate cancerNon–prostate cancerBackground
Dout The number of dysregulated target genes 10 0.0865 0.0104
NmiRNA The number of its coregulators 11 5.5 7.5 0.0150 0.0587
Rtarpc-miRNA The proportion of prostate cancer miRNAs in its coregulator set 0.2381 0.0727 0.1087 1.7E-8 <0.0001
Rtarpc-miRNA The fraction of targets which are coregulated by itself and other prostate cancer miRNAs 0.1667 0.0486 0.0976 0.0303 0.0554
Fold_C Fold changes in expression 0.8942 1.0837 0.9419 0.00013 0.9309

NOTE: Significances between prostate cancer miRNAs and 2 control sets are listed, respectively.

aCalculated by Wilcoxon rank-sum test by comparing feature values of prostate cancer miRNAs with those of non–prostate cancer miRNAs.

bCalculated by permutation test to compare prostate cancer miRNAs and general background miRNAs.

To assess the significance of differences in these measures between the prostate cancer and non–prostate cancer miRNA sets, the Wilcoxon rank-sum test for equal medians is applied. We also used a permutation test to determine whether prostate cancer miRNAs features significantly differ from those of general background miRNAs, defined as all miRNAs in the MTDN except for prostate cancer miRNAs. We randomly selected the same number of miRNAs as prostate cancer miRNAs from general background miRNAs and computed average values for the 4 topological features. We repeated this procedure 10,000 times. The P value for each feature is the fraction of feature values for random miRNA sets that are larger than the corresponding values for prostate cancer miRNAs.

### Classification algorithm and validation

An SVM, using a radial basis function and the libSVM package (31), was used to construct a classifier to distinguish prostate cancer from non–prostate cancer miRNAs in terms of the 5 features mentioned above. A scaling scheme was used for every feature vector by restricting all entries to the range 0 to 1 by calculating (X −Min)/(Max − Min) for every feature, where X is the feature score and Min and Max are the minimum and maximum values of X in the vector. The SVM was trained using 5-fold cross-validation on a data set of 37 prostate cancer miRNAs and 44 non–prostate cancer miRNAs. Similarly sized data sets were used to avoid bias for a larger set. Two parameters were varied to maximize the accuracy: the error penalty (C) for an incorrect prediction and the radial basis function parameter (γ) which controls the smoothness of the boundary between prostate cancer and non–prostate cancer miRNA areas. A genetic algorithm was used to optimize C and γ.

The prediction quality was evaluated in term of the overall prediction accuracy, sensitivity, and specificity as follows:

where TP, TN, FP, and FN are the true-positive, true-negative, false-positive, and false-negative rates, respectively. The receiver operating characteristic (ROC) curve is a robust approach for classifier evaluation (32) and is drawn by plotting sensitivity against the false-positive rate, which equals 1 − specificity. The area under the ROC curve (AUC) can be used as a reliable measure of classifier performance (33).

Using the optimal kernel functions obtained previously, the SVM classifier was applied to miRNAs in the MTDN. For each miRNA, the greater the posterior probability, the more closely the miRNA is related to prostate cancer.

### In vitro validation of miRNA target regulations in prostate cancer cell lines

Human prostate cancer cell lines PC3 and DU145 were obtained from the American Type Culture Collection. Cells were maintained in RPMI 1640 medium (HyClone) with 10% heat-inactivated FBS (HyClone) and 1% penicillin/streptomycin at 37°C and in a 5% CO2 atmosphere.

PC3 and DU145 cells were transfected with a negative control or hsa-miR-205 and hsa-miR-145 (Bioneer) using Lipofectamine 2000 transfection reagent (Invitrogen). Total RNA was isolated after 48 hours using TRIzol reagent (Invitrogen). Samples of 1 μg of RNA were reverse transcribed into cDNA using a high-capacity cDNA reverse transcription kit (ABI). PCR was carried out using TaKaRa Taq (TaKaRa) on an S1000 thermal cycler (Bio-Rad). The specific primers used are RDH11 forward: 5′-TGGCTGCGCCCCAAATCAGG-3′, reverse: 5′-TCTCCGGGCCAGTTCCTGGG; LRP1 forward: 5′-ATCGTGCCGCGAGTATGCCG-3′, reverse: 5′-GTGTGGCGCGTGATGGTGGA-3′; and SMAD3 forward: 5′-GTGCTGAGACTGACCCAA-3′, reverse: 5′-CCTGAGGCTAAGAATGAAAC-3′. The PCR products were analyzed on 2% agarose gel.

### Global properties of the MTDN

We first identified miRNA–target pairs with significant dysregulation in tumor versus nontumor data sets based on differences in correlation, as described in Materials and Methods. Using an FDR threshold of 0.01, we detected 3,758 dysregulations between 274 miRNAs and 2,511 target genes, 88.03% of which are significant in the normal or cancer subgroups. A dysregulation is defined as gain, only if the correlation coefficient is significant for disease samples and not normal samples or is significant in both subgroups but is higher for disease samples. A lost regulation is defined as a dysregulation in a similar manner but using a lower correlation coefficient for disease samples as a criterion. In total, there are 2,433 losses and 875 gains of regulation.

We found that 90.51% of the miRNAs dysregulate at least 2 targets, and approximately, one third of mRNAs are codysregulated by 2 or more miRNAs. As shown in Fig. 1A, most dysregulations in the MTDN are connected and form a large connecting subnetwork. These results indicate a complicated combination in terms of both target multiplicity and miRNA cooperativity. We evaluated the out-degree distribution of miRNAs (Dout) and the in-degree distribution of targets and observed a power law and an exponential distribution, respectively (Fig. 1B and C). Therefore, like many large-scale networks, the MTDN displays scale-free characteristics, indicating that the MTDN is not random but is characterized by a core set of organizing principles in its structure that distinguish it from randomly linked networks (34). Understanding prostate cancer in the context of these network principles allows us to address some fundamental properties of prostate cancer miRNAs. We further evaluated the robustness of the MTDN structure for a looser FDR threshold (FDR = 0.2) and found that the MTDN is still scale free (Supplementary Fig. S1).

Figure 1.

The layout of the MTDN and its structural features. A, the MTDN generated by the procedure described in the Materials and Methods. This network consists of 3,758 dysregulations between 274 miRNAs and 2,511 target genes. A circle node marks miRNA and a diamond node marks gene. An edge represents a dysregulation from a miRNA to 1 of its targets. Prostate cancer miRNAs are also colored by black, other miRNAs are color coded on the basis of their posterior probabilities computed by the SVM classifier, and range from dark gray (high values) to white (low values). B, out-degree distribution of the MTDN. Most of miRNAs are lowly connected and only a few are relatively highly connected. The examination of the out-degree distribution of the MTDN reveals a power law with a slope of −0.392 and R2 = 0.85. C, in-degree distribution of the MTDN. In degree is defined as the number of dysregulatory miRNAs for each target, signifying an exponential distribution with an exponent of −1.367 and R2 = 0.99.

Figure 1.

The layout of the MTDN and its structural features. A, the MTDN generated by the procedure described in the Materials and Methods. This network consists of 3,758 dysregulations between 274 miRNAs and 2,511 target genes. A circle node marks miRNA and a diamond node marks gene. An edge represents a dysregulation from a miRNA to 1 of its targets. Prostate cancer miRNAs are also colored by black, other miRNAs are color coded on the basis of their posterior probabilities computed by the SVM classifier, and range from dark gray (high values) to white (low values). B, out-degree distribution of the MTDN. Most of miRNAs are lowly connected and only a few are relatively highly connected. The examination of the out-degree distribution of the MTDN reveals a power law with a slope of −0.392 and R2 = 0.85. C, in-degree distribution of the MTDN. In degree is defined as the number of dysregulatory miRNAs for each target, signifying an exponential distribution with an exponent of −1.367 and R2 = 0.99.

Close modal

### Topological feature values differ significantly between prostate cancer and non–prostate cancer miRNAs

Next, we investigated whether prostate cancer miRNAs have specific topological patterns in the MTDN that could measure miRNAs modularity from different but related facets. We found that the prostate cancer miRNAs have significantly higher out-degree than non–prostate cancer miRNAs or general background miRNAs (Table 1), which indicates that prostate cancer miRNAs dysregulate more targets and influence more functions. Rtarpc-miRNA is approximately 3 times greater for prostate cancer miRNAs than for non–prostate cancer miRNAs, so prostate cancer miRNAs coordinately dysregulate proportionally more targets than miRNAs in the 2 control sets.

To quantify the cooperation complexity, we used the number of coregulators (NmiRNA) as a measure. The greater this value, the greater is the number of coregulatory miRNAs. As a result, prostate cancer miRNAs have significantly more miRNA coregulators than non–prostate cancer miRNAs; the median NmiRNA values of 11 for prostate cancer miRNAs and 5.5 for non–prostate cancer miRNAs indicate that prostate cancer miRNAs exhibit more cooperation, in accordance with the conclusion that disease miRNAs exhibit more synergism (16). Furthermore, prostate cancer miRNAs had a significantly greater proportion of coregulatory prostate cancer miRNA partners than non–prostate cancer miRNAs (P < 1.7E-8) and the general background miRNA set (P < 0.0001). Finally, we analyzed the fold change in miRNA expression, which is generally used to identify disease-associated miRNAs. Although the fold change significantly differs between prostate cancer and non–prostate cancer miRNAs, a permutation test reveals no significant difference between prostate cancer and general background miRNAs.

We found that topological differences between prostate cancer and non–prostate cancer miRNAs are also significant in the MTDN using a loose FDR threshold, and these features all have highly significant positive correlations in the 2 MTDNs (Supplementary Table S2). These results indicate that topological features of prostate cancer miRNAs are not influenced by the threshold used to construct the MTDN. Interestingly, there are no significant topological differences between prostate cancer and non–prostate cancer miRNAs in the original miRNA–target regulatory network (Supplementary Table S3). We conclude that prostate cancer miRNAs preferentially have a coregulatory relationship with other prostate cancer miRNAs in the specific posttranscriptional prostate cancer regulatory network rather than the whole miRNA–target regulatory network, as proved in our previous work (16).

### Satisfactory performance of SVM classifier using topological features in the MTDN

Because identification of miRNA over- or underexpression is one of the most widely used screening methods for potential tumor markers, the fold change in expression is used to build a classifier for comparison. Our new classifier also considers the 4 topological features discussed above. Our model shows higher power than the control model, achieving sensitivity of 0.8643, specificity of 0.8833, accuracy of 0.8772, and AUC of 0.9189 (Table 2 and Fig. 2). The results show that a combination of the MTDN topological features and expression could improve the prediction power for miRNAs associated with prostate cancer.

Figure 2.

The ROC curves of SVM classifier. It is created on 2 combinations of features: 4 topological features of miRNAs in the MTDN and their fold changes in expression (the bold solid line); only fold change in expression (the bold dashed line). The diagonal line is the baseline.

Figure 2.

The ROC curves of SVM classifier. It is created on 2 combinations of features: 4 topological features of miRNAs in the MTDN and their fold changes in expression (the bold solid line); only fold change in expression (the bold dashed line). The diagonal line is the baseline.

Close modal
Table 2.

Five-fold cross-validation results of the SVM classifier

Feature setsAccuracySensitivitySpecificity
Network features + fold change 0.8772 ± 0.0599 0.8643 ± 0.1015 0.8833 ± 0.1185
Fold change 0.6649 ± 0.0644 0.5357 ± 0.1002 0.7722 ± 0.0795
Feature setsAccuracySensitivitySpecificity
Network features + fold change 0.8772 ± 0.0599 0.8643 ± 0.1015 0.8833 ± 0.1185
Fold change 0.6649 ± 0.0644 0.5357 ± 0.1002 0.7722 ± 0.0795

### Prediction of novel prostate cancer miRNAs

We applied the trained classifier to calculate the posterior probability for each miRNA in the MTDN. Figure 3A summarizes the ability of our method (or fold changes in expression) to enrich prostate cancer miRNAs within top-ranked candidates. Specifically, in our model, 100%, 70%, and 73.3% of the top-ranked 5, 10, and 15 miRNAs, respectively, are known prostate cancer miRNAs; in contrast, these values are only 40%, 40%, and 33.3% when ranking miRNAs by the fold change in expression. Even when the number of ranked top miRNAs is increased, our model still outperforms the fold change method. Moreover, our model assigns high ranks to most of the known oncogenic miRNAs, with 32 of these 37 prostate cancer miRNAs ranked in the top 84 and with posterior probabilities greater than 60% whereas only 8 of those are significantly up- or downregulated according to a fold change criterion of 1.5 or above. These results imply that although prostate cancer miRNAs do not show significant changes in expression, they play important roles in the MTDN.

Figure 3.

Performance of our model on the miRNA set with high posterior probabilities and examples of predicted novel disease miRNAs. A, the percentage of recalled prostate cancer miRNAs in the ranked miRNA set. The left y-axis represents the percentage of known prostate cancer (PC) miRNAs in the top-ranked miRNA set, corresponding to the bar figure. The right y-axis represents the capability to recall prostate cancer miRNAs, corresponding to line figure. Black is corresponding to miRNAs ranked by their posterior probabilities; gray is corresponding to miRNAs ranked by the degree of expression changes. B, an example of a predicted novel miRNA, hsa-miR-203 with posterior probability of 91.55%, including its 16 dysregulated targets and 24 coregulatory miRNAs. The colors of miRNAs have the same meanings as in Fig. 1A.

Figure 3.

Performance of our model on the miRNA set with high posterior probabilities and examples of predicted novel disease miRNAs. A, the percentage of recalled prostate cancer miRNAs in the ranked miRNA set. The left y-axis represents the percentage of known prostate cancer (PC) miRNAs in the top-ranked miRNA set, corresponding to the bar figure. The right y-axis represents the capability to recall prostate cancer miRNAs, corresponding to line figure. Black is corresponding to miRNAs ranked by their posterior probabilities; gray is corresponding to miRNAs ranked by the degree of expression changes. B, an example of a predicted novel miRNA, hsa-miR-203 with posterior probability of 91.55%, including its 16 dysregulated targets and 24 coregulatory miRNAs. The colors of miRNAs have the same meanings as in Fig. 1A.

Close modal

Next, we checked whether our predicted novel candidate miRNAs are involved in prostate cancer. Although we found slightly expression changes in prostate cancer for the top 9 novel disease miRNAs with highest posterior probability, many functions enriched by dysregulated targets of these miRNAs are associated with prostate cancer according to analysis using the tool of Gene Set Analysis Toolkit (Supplementary Table S4; ref. 35). Adjusted significant P values (Padj) are obtained using the Benjamini–Hochberg procedure. For example, it has recently been reported that hsa-miR-203 (posterior probability of 91.55% here) suppresses prostate cancer progression and metastasis and influences cellular proliferation and apoptosis in prostate cancer cells (36). The topological feature values are higher for hsa-miR-203 than for non–prostate cancer miRNAs (Fig. 3B) and 2 functions are also enriched by its dysregulated targets (Padj = 0.0719 and 0.0235, respectively). Furthermore, hsa-miR-203 influences hedgehog signaling pathway (Padj = 0.0333), which is involved in cell differentiation and proliferation and plays roles in several human cancers including prostate cancer (37, 38).

### Modular organization of miRNAs associated with prostate cancer in the MTDN

As discussed above, prostate cancer miRNAs have complex dysregulatory relationships and cooperation. To reveal the detailed modular organization of miRNAs associated with prostate cancer in the MTDN, we attempted to predict miRNA–mRNA dysregulatory modules. Here, an miRNA–mRNA regulatory module is defined as a biclique, which is a complete bipartite graph such that an edge is realized from every vertex of a miRNA set to every vertex of a gene set. We focused on maximal bicliques to avoid redundancy and all modules are searched using an algorithm downloaded from the Web site of Computational Biology Laboratory, Department of Computer Science, Iowa State University, Ames, IA (39). In total, 1,118 modules are detected. As shown in Fig. 4A, as the minimum number of miRNAs in a biclique (k) increases, there is a sharp decrease in the number of modules and the proportion of miRNAs. We found that most miRNAs are within at least 1 module, and the majority of modules contain less than 4 miRNAs.

Figure 4.

The modular organization of miRNAs associated with prostate cancer in the MTDN and downregulated expression of RDH11 at the mRNA level in DU145 prostate cancer cells by transfecting hsa-miR-205. A, cumulative ratios of miRNAs in bicliques and the number of bicliques with the number of miRNAs are not smaller than k. The left y-axis represents cumulative proportions of miRNAs in bicliques, corresponding to the solid line. The right y-axis represents number of bicliques with different minimum k values, corresponding to the dashed line. B, the trend of ER of miRNAs associated with prostate cancer in modules with different minimum number of miRNAs (k). ER is the degree to which miRNAs from bicliques with different k values are under- (ER < 1) or overrepresented (ER >1) relative to miRNAs in the MTDN. C, an example of the subnetwork containing 3 densely connected modules with at least 3 miRNAs and 5 targets. The colors of miRNAs have the same meanings as in Fig. 1A. D, expression of RDH11 is downregulated by transfection of hsa-miR-205 in DU145 cell lines in vitro. hsa-miR-205 or a negative control are transfected into DU145 cells. RNA was extracted after 48 hours of transfection. The expressions of RDH11 were measured using reverse transcriptase PCR. β-Actin serves as a loading control. It was shown that RDH11 expression decreases as the concentration of transfected hsa-miR-205 increases. PP, posterior probability.

Figure 4.

The modular organization of miRNAs associated with prostate cancer in the MTDN and downregulated expression of RDH11 at the mRNA level in DU145 prostate cancer cells by transfecting hsa-miR-205. A, cumulative ratios of miRNAs in bicliques and the number of bicliques with the number of miRNAs are not smaller than k. The left y-axis represents cumulative proportions of miRNAs in bicliques, corresponding to the solid line. The right y-axis represents number of bicliques with different minimum k values, corresponding to the dashed line. B, the trend of ER of miRNAs associated with prostate cancer in modules with different minimum number of miRNAs (k). ER is the degree to which miRNAs from bicliques with different k values are under- (ER < 1) or overrepresented (ER >1) relative to miRNAs in the MTDN. C, an example of the subnetwork containing 3 densely connected modules with at least 3 miRNAs and 5 targets. The colors of miRNAs have the same meanings as in Fig. 1A. D, expression of RDH11 is downregulated by transfection of hsa-miR-205 in DU145 cell lines in vitro. hsa-miR-205 or a negative control are transfected into DU145 cells. RNA was extracted after 48 hours of transfection. The expressions of RDH11 were measured using reverse transcriptase PCR. β-Actin serves as a loading control. It was shown that RDH11 expression decreases as the concentration of transfected hsa-miR-205 increases. PP, posterior probability.

Close modal

To further probe the relationship between module size and miRNAs involved in prostate cancer, we used excess retention (ER) as a measure of the degree to which miRNAs from a particular group are represented relative to the entire miRNA set in the MTDN. After setting a posterior probability threshold to define predicted miRNAs associated with prostate cancer, we divided miRNAs into groups according to biclique k values, and the ER for predicted disease miRNAs is calculated for each group. ER has been described in detail elsewhere (40). Interestingly, ER increases with k, indicating that although the proportion of miRNAs contained in modules decreases with increasing k, the proportion of predicted disease miRNAs identified in modules increases to a much greater extent than expected for random events (Fig. 4B). The same tendency is observed when using different posterior probability thresholds to define miRNAs associated with prostate cancer. For k = 11, 48.89% of miRNAs associated with prostate cancer have a posterior probability greater than 80%, which is 3.2672 times greater than the random fraction of predicted disease miRNAs. These results indicate that miRNAs associated with prostate cancer tend to reside in modules with many other miRNAs and work coordinately in clusters. Therefore, whereas individual miRNAs may provide only modest levels of repression, combinatorial targeting by multiple miRNAs can potentially achieve a wide range of target-level modulations (3).

A subnetwork of 3 densely connected modules provides an example of combinatorial regulation among miRNAs within a family or from different families. For miRNAs in the subnetwork of mir-125, mir-181, and mir-145 miRNA families, the minimum posterior probability is 77.4%, indicating that they are associated with prostate cancer. As shown in Fig. 4C, target genes are strictly regulated, and the mir-125 family plays important roles. We found that 5 targets are transmembrane proteins (FGFR2, MET, CELSR2, NYNRIN, and SYNGR1), and FGFR2 and MET in particular are involved in cancer pathway. They may be involved in contact-mediated communication, cell adhesion, and receptor–ligand interactions. In addition, by functional enrichment of target genes codysregulated by the mir-125 family and its coregulators, we found that several functions are associated with cancers, such as pathways in prostate cancer, cell adhesion, and apoptosis.

### Three miRNA target regulations hold in prostate cancer cells by in vitro assay

Finally, to establish the accuracy of our method, we conducted an in vitro analysis of 3 dysregulatory relationships: 2 regulations between hsa-miR-205 and its targets RDH11 and LRP1 and 1 between hsa-miR-145 and SMAD3. As a result, hsa-miR-205 transfection produces significant reductions in RDH11 mRNA levels in both DU145 and PC3 cells and of LRP1 levels in PC3 cells (Fig. 4D and Supplementary Fig. S2). In particular, RDH11 expression in DU145 cells decreases as the concentration of transfected hsa-miR-205 increases. hsa-miR-145 also downregulates SMAD3 expression. These results confirm that with the miRNA–target relationships, we detected hold in prostate cancer cells.

In addition, we found that hsa-miR-205 and has-miR-145 (posterior probability 78.77% and 92.49%, respectively) are underexpressed in tumor tissues whereas their target genes are overexpressed. Several studies show that hsa-miR-205 and has-miR-145 act as oncosuppressors in prostate cancer (41, 42), and these targets are closely associated with carcinogenesis Therefore, changes in these validated miRNA–mRNA regulatory relationships could cause prostate cancer. RDH11, an oncogene, is located in a high-level amplification region in tumor cells (43) and is specifically expressed in the prostate, playing important roles in maintaining intracellular balance of steroid hormones (44). LRP1 could promote the invasion and migration ability of lung cancer cells and glioblastoma cells (45) and is directly regulated by hsa-miR-205 in lung cancer cells (46). Therefore, we propose that hsa-miR-205 may inhibit prostate cancer cell proliferation and migration by targeting RDH11 and LRP1. In addition, hsa-miR-145 might influence prostate cancer cell motility by targeting SMAD3, a major TGF-β signaling transducer (47). Overall, the analysis increases our understanding of hsa-miR-205 and hsa-miR-145 in prostate cancer.

In this study, we combined paired miRNA and mRNA expression data and computational target prediction to detect aberrant behavior between miRNAs and their direct targets. Among all the regulatory relationships considered, only a small fraction show aberrant behavior in prostate cancer, which implies that although miRNAs have many different targets—dozens in some cases, hundreds in others—it seems likely that only a small number of them have a crucial role in cancer pathogenesis in specific prostate cancer cells. A key difference from other network-based methods is that we identified dysregulated network edges (regulations) instead of dysregulated nodes (miRNAs) to assemble disease-related signatures.

The functions of most miRNAs remain unknown and even for relatively well-studied miRNAs, only a few targets have been rigorously studied. Knockout studies in model organisms have had only limited success in describing miRNA functions, possibly because of functional redundancy among miRNAs. Thus, analysis of the properties of miRNA-dysregulated targets via the MTDN is a promising approach for predicting specific miRNA functions in the context of prostate cancer. By conducting functional enrichment of their dysregulated targets, we found that the functions of known or predicted oncogenic miRNAs are closely related to prostate carcinogenesis.

To determine changes in correlation for dysregulations in an independent data set, correlations are calculated as described above for all the miRNA–mRNA relationships in the data set reported by Taylor and colleagues (described in detail in the Supplementary text; ref. 48). We found that dysregulations identified in the text are significantly enriched in the regulation subset with large changes identified in the data set of Taylor and colleagues, indicating that most dysregulations in the former data set show large changes in correlation according to the data set of Taylor and colleagues (Supplementary Fig. S3). We further used this data set to construct a new MTDN and found that 4 topological features significantly differ between prostate cancer and non–prostate cancer miRNAs. A new classifier built using these topological features shows satisfactory performance. More importantly, disease miRNAs predicted by our model are more reproducible between data sets than miRNAs selected according to expression levels (Supplementary Tables S5 and 6). In addition, candidate cancer miRNAs selected in 1 data set show good classification performance in the other. Detailed results obtained for the data set of Taylor and colleagues are in the Supplementary text.

Prostate cancer miRNA prioritization is realized in 3 stages, which has recently been proposed to stepwise control of the overall false discoveries (49, 50). A subset of miRNA mRNA regulations that pass a stringent FDR threshold for correlation differences is chosen in the first stage. In the second stage, 4 topological features for miRNAs are calculated on the basis of this filtered subset. Finally, a SVM is trained on the gold standard miRNA data sets and used to prioritize novel prostate cancer miRNAs. To determine the effects of the initial miRNA–target relationship generated by different integrating modes of miRNA target prediction algorithms, our model is applied to regulations predicted by 2 or more algorithms. Topological features significantly differ between prostate cancer and non–prostate cancer miRNAs, except for the out degree, and the classifier power is still higher than that of the expression classifier. In addition, 94.81% of dysregulations in the MTDN are also significant in the normal or cancer subgroups for an FDR threshold of 0.05. Therefore, our model is reliable.

Our method can be used to prioritize genome-wide candidate miRNAs for specific cancers. Owing to the computational complexity of MTDN construction in the present study, we only analyzed predicted direct miRNA target regulation. When the model is specifically applied to a group of miRNAs of interest, such as miRNA families, we could attempt to identify secondary effects by enumerating all combinations of the miRNAs and genes in the human genome but keeping the requirement of significant changes in regulation. Attention could be focused on miRNAs with high values for topological features or in densely connected modules. Note that our method does not identify regulation by translation inhibition because this would not affect mRNA levels; however, if protein levels are also measured, the method could easily be extended to identify such regulation.

The results present here provide new insights into posttranscriptional regulation of prostate cancer, despite limitations in the data. Our model is suitable for investigating miRNA regulation in other complex human diseases.

No potential conflicts of interest were disclosed.

The authors thank the editor and 3 anonymous referees for their constructive advice and comments to improve this work.

This work is supported in part by the National Natural Science Foundation of China (grant nos. 30871394, 61073136, 91029717, and 30600367), the National High Tech Development Project of China, the 863 Program (grant no. 2007AA02Z329), and the National Science Foundation of Heilongjiang Province (grant nos. ZD200816-01, ZJG0501, GB03C602-4, and JC200711).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
Farh
KK
,
Grimson
A
,
Jan
C
,
Lewis
BP
,
Johnston
WK
,
Lim
LP
, et al
The widespread impact of mammalian microRNAs on mRNA repression and evolution
.
Science
2005
;
310
:
1817
21
.
2.
Friedman
RC
,
Farh
KK
,
Burge
CB
,
Bartel
DP
.
Most mammalian mRNAs are conserved targets of microRNAs
.
Genome Res
2009
;
19
:
92
105
.
3.
Krek
A
,
Grun
D
,
Poy
MN
,
Wolf
R
,
Rosenberg
L
,
Epstein
EJ
, et al
Combinatorial microRNA target predictions
.
Nat Genet
2005
;
37
:
495
500
.
4.
Cheng
AM
,
Byrom
MW
,
Shelton
J
,
Ford
LP
.
Antisense inhibition of human miRNAs and indications for an involvement of miRNA in cell growth and apoptosis
.
Nucleic Acids Res
2005
;
33
:
1290
7
.
5.
Cui
Q
,
Yu
Z
,
Purisima
EO
,
Wang
E
.
Principles of microRNA regulation of a human cellular signaling network
.
Mol Syst Biol
2006
;
2
:
46
.
6.
Karp
X
,
Ambros
V
.
Developmental biology. Encountering microRNAs in cell fate signaling
.
Science
2005
;
310
:
1288
9
.
7.
Xu
P
,
Guo
M
,
Hay
BA
.
MicroRNAs and the regulation of cell death
.
Trends Genet
2004
;
20
:
617
24
.
8.
Ulitsky
I
,
Laurent
LC
,
Shamir
R
.
Towards computational prediction of microRNA function and activity
.
Nucleic Acids Res
2010
;
38
:
e160
.
9.
Tsang
JS
,
Ebert
MS
,
van Oudenaarden
A
.
Genome-wide dissection of microRNA functions and cotargeting networks using gene set signatures
.
Mol Cell
2010
;
38
:
140
53
.
10.
Backes
C
,
Meese
E
,
Lenhof
HP
,
Keller
A
.
A dictionary on microRNAs and their putative target pathways
.
Nucleic Acids Res
2010
;
38
:
4476
86
.
11.
Ferracin
M
,
Veronese
A
,
Negrini
M
.
Micromarkers: miRNAs in cancer diagnosis and prognosis
.
Expert Rev Mol Diagn
2010
;
10
:
297
308
.
12.
Barbarotto
E
,
Schmittgen
TD
,
Calin
GA
.
MicroRNAs and cancer: profile, profile, profile
.
Int J Cancer
2008
;
122
:
969
77
.
13.
Osaki
M
,
Takeshita
F
,
Ochiya
T
.
MicroRNAs as biomarkers and therapeutic drugs in human cancer
.
Biomarkers
2008
;
13
:
658
70
.
14.
Liu
H
,
Brannon
AR
,
Reddy
AR
,
Alexe
G
,
Seiler
MW
,
Arreola
A
, et al
Identifying mRNA targets of microRNA dysregulated in cancer: with application to clear cell renal cell carcinoma
.
BMC Syst Biol
2010
;
4
:
51
.
15.
S
,
Mitra
R
,
Maulik
U
,
Zhang
MQ
.
Development of the human cancer microRNA network
.
Silence
2010
;
1
:
6
.
16.
Xu
J
,
Li
CX
,
Li
YS
,
Lv
JY
,
Ma
Y
,
Shao
TT
, et al
MiRNA-miRNA synergistic network: construction via co-regulating functional modules and disease miRNA topological features
.
Nucleic Acids Res
2011
;
39
:
825
36
.
17.
Volinia
S
,
Galasso
M
,
Costinean
S
,
Tagliavini
L
,
Gamberoni
G
,
Drusco
A
, et al
Reprogramming of miRNA networks in cancer and leukemia
.
Genome Res
2010
;
20
:
589
99
.
18.
Lu
M
,
Zhang
Q
,
Deng
M
,
Miao
J
,
Guo
Y
,
Gao
W
, et al
An analysis of human microRNA and disease associations
.
PLoS One
2008
;
3
:
e3420
.
19.
Jiang
Q
,
Hao
Y
,
Wang
G
,
Juan
L
,
Zhang
T
,
Teng
M
, et al
Prioritization of disease microRNAs through a human phenome-microRNAome network
.
BMC Syst Biol
2010
;
4
Suppl 1
:
S2
.
20.
Ambs
S
,
Prueitt
RL
,
Yi
M
,
Hudson
RS
,
Howe
TM
,
Petrocca
F
, et al
Genomic profiling of microRNA and messenger RNA reveals deregulated microRNA expression in prostate cancer
.
Cancer Res
2008
;
68
:
6162
70
.
21.
Shi
XB
,
Tepper
CG
,
White
RW
.
MicroRNAs and prostate cancer
.
J Cell Mol Med
2008
;
12
:
1456
65
.
22.
Coppola
V
,
De Maria
R
,
Bonci
D
.
MicroRNAs and prostate cancer
.
Endocr Relat Cancer
2010
;
17
:
F1
17
.
23.
Gentleman
RC
,
Carey
VJ
,
Bates
DM
,
B
,
Dettling
M
,
Dudoit
S
, et al
Bioconductor: open software development for computational biology and bioinformatics
.
Genome Biol
2004
;
5
:
R80
.
24.
Griffiths-Jones
S
,
Saini
HK
,
van Dongen
S
,
Enright
AJ
.
miRBase: tools for microRNA genomics
.
Nucleic Acids Res
2008
;
36
:
D154
8
.
25.
Lewis
BP
,
Shih
IH
,
MW
,
Bartel
DP
,
Burge
CB
.
Prediction of mammalian microRNA targets
.
Cell
2003
;
115
:
787
98
.
26.
Maragkakis
M
,
Reczko
M
,
Simossis
VA
,
Alexiou
P
,
GL
,
Dalamagas
T
, et al
DIANA-microT web server: elucidating microRNA functions through target prediction
.
Nucleic Acids Res
2009
;
37
:
W273
6
.
27.
Hochberg
Y
,
Benjamini
Y
.
Controlling the false discovery rate: a practical and powerful approach to multiple testing
.
J R Stat Soc
1995
;
57
:
289
300
.
28.
Jiang
Q
,
Wang
Y
,
Hao
Y
,
Juan
L
,
Teng
M
,
Zhang
X
, et al
miR2Disease: a manually curated database for microRNA deregulation in human disease
.
Nucleic Acids Res
2009
;
37
:
D98
104
.
29.
Hu
Z
,
Gallo
SM
.
Identification of interacting transcription factors regulating tissue gene expression in human
.
BMC Genomics
2010
;
11
:
49
.
30.
Pennacchio
LA
,
Loots
GG
,
Nobrega
MA
,
Ovcharenko
I
.
Predicting tissue-specific enhancers in the human genome
.
Genome Res
2007
;
17
:
201
11
.
31.
Available from
: http://www.csie.ntu.edu.tw/~cjlin/libsvm.
[cited 2010 Sep 24]
.
32.
Swets
JA
.
Measuring the accuracy of diagnostic systems
.
Science
1988
;
240
:
1285
93
.
33.
AP
.
The use of the area under the ROC curve in the evaluation of machine learning algorithms
.
Pattern Recognition
1997
;
30
:
1145
59
.
34.
Barabasi
AL
,
Gulbahce
N
,
Loscalzo
J
.
Network medicine: a network-based approach to human disease
.
Nat Rev Genet
2011
;
12
:
56
68
.
35.
Zhang
B
,
Kirov
S
,
Snoddy
J
.
WebGestalt: an integrated system for exploring gene sets in various biological contexts
.
Nucleic Acids Res
2005
;
33
:
W741
8
.
36.
Saini
S
,
Majid
S
,
Yamamura
S
,
Tabatabai
ZL
,
Suh
SO
,
Shahryari
V
, et al
Regulatory role of miR-203 in prostate cancer progression and metastasis
.
Clin Cancer Res
2011
;
17
:
5287
98
.
37.
Sheng
T
,
Li
C
,
Zhang
X
,
Chi
S
,
He
N
,
Chen
K
, et al
Activation of the hedgehog pathway in advanced prostate cancer
.
Mol Cancer
2004
;
3
:
29
.
38.
Chen
M
,
Tanner
M
,
Levine
AC
,
Levina
E
,
Ohouo
P
,
Buttyan
R
.
Androgenic regulation of hedgehog signaling pathway components in prostate cancer cells
.
Cell Cycle
2009
;
8
:
149
57
.
40.
Wuchty
S
,
Almaas
E
.
Peeling the yeast protein network
.
Proteomics
2005
;
5
:
444
9
.
41.
Gandellini
P
,
Folini
M
,
Longoni
N
,
Pennati
M
,
Binda
M
,
Colecchia
M
, et al
miR-205 Exerts tumor-suppressive functions in human prostate through downregulation of protein kinase Cepsilon
.
Cancer Res
2009
;
69
:
2287
95
.
42.
Fuse
M
,
Nohata
N
,
Kojima
S
,
Sakamoto
S
,
Chiyomaru
T
,
Kawakami
K
, et al
Restoration of miR-145 expression suppresses cell proliferation, migration, and invasion in prostate cancer by targeting FSCN1
.
Int J Oncol
2011
;
38
:
1093
101
.
43.
Guillaud-Bataille
M
,
Brison
O
,
Danglot
G
,
Lavialle
C
,
Raynal
B
,
Lazar
V
, et al
Two populations of double minute chromosomes harbor distinct amplicons, the MYC locus at 8q24.2 and a 0.43-Mb region at 14q24.1, in the SW613-S human carcinoma cell line
.
Cytogenet Genome Res
2009
;
124
:
1
11
.
44.
Lin
B
,
White
JT
,
Ferguson
C
,
Wang
S
,
Vessella
R
,
Bumgarner
R
, et al
Prostate short-chain dehydrogenase reductase 1 (PSDR1): a new member of the short-chain steroid dehydrogenase/reductase family highly expressed in normal and neoplastic prostate epithelium
.
Cancer Res
2001
;
61
:
1611
8
.
45.
Song
H
,
Li
Y
,
Lee
J
,
Schwartz
AL
,
Bu
G
.
Low-density lipoprotein receptor-related protein 1 promotes cancer cell migration and invasion by inducing the expression of matrix metalloproteinases 2 and 9
.
Cancer Res
2009
;
69
:
879
86
.
46.
Song
H
,
Bu
G
.
MicroRNA-205 inhibits tumor cell migration through downregulating the expression of the LDL receptor-related protein 1
.
Biochem Biophys Res Commun
2009
;
388
:
400
5
.
47.
Craft
CS
,
Romero
D
,
Vary
CP
,
Bergan
RC
.
Endoglin inhibits prostate cancer motility via activation of the ALK2-Smad1 pathway
.
Oncogene
2007
;
26
:
7240
50
.
48.
Taylor
BS
,
Schultz
N
,
Hieronymus
H
,
Gopalan
A
,
Xiao
Y
,
Carver
BS
, et al
Integrative genomic profiling of human prostate cancer
.
Cancer Cell
2010
;
18
:
11
22
.
49.
Lin
DY
.
Evaluating statistical significance in two-stage genome-wide association studies
.
Am J Hum Genet
2006
;
78
:
505
9
.
50.
McCarthy
MI
,
Abecasis
GR
,
Cardon
LR
,
Goldstein
DB
,
Little
J
,
Ioannidis
JP
, et al
Genome-wide association studies for complex traits: consensus, uncertainty, and challenges
.
Nat Rev Genet
2008
;
9
:
356
69
.