Integrated transcriptional and competitive endogenous RNA networks are cross-regulated in permissive molecular environments

Abstract

Competitive endogenous (ce)RNAs cross-regulate each other through sequestration of shared microRNAs and form complex regulatory networks based on their microRNA signature. However, the molecular requirements for ceRNA cross-regulation and the extent of ceRNA networks remain unknown. Here, we present a mathematical mass-action model to determine the optimal conditions for ceRNA activity in silico. This model was validated using phosphatase and tensin homolog (PTEN) and its ceRNA VAMP (vesicle-associated membrane protein)-associated protein A (VAPA) as paradigmatic examples. A computational assessment of the complexity of ceRNA networks revealed that transcription factor and ceRNA networks are intimately intertwined. Notably, we found that ceRNA networks are responsive to transcription factor up-regulation or their aberrant expression in cancer. Thus, given optimal molecular conditions, alterations of one ceRNA can have striking effects on integrated ceRNA and transcriptional networks.

We and others have postulated that the activity of microRNAs (miRNAs) is regulated by the abundance of their targets and, as a consequence, the concentration of a target RNA may impact the activity of a miRNA toward its other targets (1–4). Thus, miRNA targets may cross-regulate each other by competing for shared miRNAs. We termed transcripts acting in this fashion “competitive endogenous RNAs” (ceRNAs) (3).

The ceRNAs crosstalk is based on an “miRNA response element (MRE) language” that may be applied to any MRE-containing RNA molecule including noncoding RNAs such as pseudogenes and long noncoding RNAs (lncRNAs) (3). As most transcripts harbor several MREs and most miRNAs target numerous transcripts, ceRNAs likely act in complex networks (ceRNETs), which can be deconvoluted on the basis of their MRE language (5, 6).

Several noncoding RNAs have been demonstrated to sequester miRNAs and thus regulate the expression of mRNAs in plants and mammalian cells (2, 7, 8). These studies have uncovered a thus far unappreciated function of noncoding RNAs in gene expression, ascribing potentially critical regulatory roles to the transcribed “junk” of the genome. Importantly, guided by the MRE language of ceRNAs, we have identified multiple protein-coding mRNAs that control expression of the tumor suppressor phosphatase and tensin homolog (PTEN) by competing for common miRNAs (5, 6). Independent studies have provided further evidence that mRNAs have transregulatory functions through miRNA sequestration (9, 10).

In this study, we define the rules and dynamics of the ceRNA language and formulate a model that explains the effect of single-factor perturbation on whole ceRNA networks and how these effects are amplified by the intrinsic coregulated nature of the network. Our analysis establishes a framework for the study of ceRNA function that will deepen our understanding of this novel aspect of RNA biology. In addition, we show that transcription factor (TF) and ceRNA networks are tightly intertwined in both physiological and pathological conditions.

Results

ceRNA Model and Its Rules.

Regulatory networks of ceRNAs are complex and consist of numerous RNA components. Indeed, using the miRNA target prediction tool TargetScan (www.targetscan.org/vert_61/vert_61_data_download/Predicted_Targets_Info.txt.zip), we found that more than 50% of miRNA families have between 1 and 400 mRNA targets, whereas miRNA families with over 1,000 target mRNAs are rare (Fig. S1A). Conversely, only a few conserved miRNA families target the majority of mRNAs (Fig. S1B). This complexity of mRNA–miRNA interactions may create a permissive molecular environment that allows for the evolution of ceRNA networks.

We hypothesized that the expression levels of individual components of a ceRNA network (i.e., ceRNAs and miRNAs) influence cross-regulation. We presume that when the total number of transcripts vastly exceeds the number of miRNA molecules, overall ceRNA activity is minimal due to the limited number of available miRNAs (Fig. 1A). Conversely, if miRNA molecules are more abundant than ceRNA molecules, cross-regulation is unlikely to occur as most transcripts are fully repressed (Fig. 1A). We therefore hypothesized that optimal ceRNA-mediated cross-regulation occurs at a near-equimolar equilibrium of all ceRNAs and miRNAs within a network (Fig. 1A).

Fig. 1.

ceRNA interaction model. (A) Schematic graph depicting expression levels of the total pool of ceRNAs and miRNAs in a theoretical ceRNET. (B) Schematic outline of the considerations upon which the ceRNA model is based. Green rectangles are DNA molecules, which are transcribed with rates k_S, k_R1, and k_R2 to create miRNA molecules S (orange stars) and mRNA molecules R₁ and R₂ (blue and red pentagons), respectively. mRNA molecules form complexes with miRNAs (C₁ or C₂) with association rates k_1,+ or k_2,+ and dissociation rates k_1,− or k_2,−. Each complex degrades with rate g₁ or g₂, respectively, and miRNAs are recycled with the probability (1 − α), whereas mRNAs are always lost. Free molecular species (R₁, R₂, and S) can degrade with rates g_i (where i = {R₁, R₂, S}). Gray shapes represent degraded molecules. mRNA translation is not taken into account. (C) Prediction of the effect of altered ceRNA expression on a ceRNA network consisting of two ceRNAs and one miRNA. (D) The Pearson coefficient for ceRNA1 and ceRNA2 in C is shown.

Given the complexity of ceRNA networks, we sought to model how perturbations of individual components affect the entire network. Previous work on protein–protein interactions (11, 12), small RNA (sRNA) regulation in bacteria (11–15), and miRNA–target threshold effects (16) suggest that ceRNA interactions underlie a titration mechanism that leads to a threshold-like behavior between the interacting molecules and ultrasensitivity in proximity to the threshold (11, 12). Such a mechanism effectively described sRNA–target and miRNA–target interactions in bacteria (13) and eukaryotes (16), respectively. In bacteria, titration entails hierarchical crosstalk between common targets of sRNA (14) and prioritized sRNA target expression (17).

To describe the ceRNA effect we propose a kinetic model of the molecular titration mechanism leading to the effective interaction between any two genes through a common set of miRNAs they are in competition for (see SI Materials and Methods and Figs. S2 and S3 for details on such a model). The minimal system of miRNA–ceRNA interaction involves one miRNA and two target mRNAs (the ceRNAs): In the model we take into account the dynamic properties of free miRNAs (S), free mRNAs (ceRNAs) for the two targets (R1 and R2), and complexes of the miRNA with its targets (C1 and C2) (Fig. 1B). The model’s parameters are transcription and degradation rates for R_1,2 and S, and association, dissociation, and degradation rates for C_1,2. Other parameters such as subcellular localization and MRE accessibility regulated by secondary structure or RNA-binding proteins may influence the crosstalk between ceRNAs. However, as these parameters remain to be firmly established we did not consider them for our current model.

Solving the system of equations describing the above interactions (Eq. S1) at steady state and in the limit of strong miRNA–target interactions, we obtain simple relations for the levels of the ceRNAs and the miRNA as a function of the kinetic constants of the model,

where k_i and g_i are transcription and degradation rates, respectively, and, for illustrative purposes, all target mRNAs have approximately the same stability (g_R1 ∼ g_R2 = g_R).

The parameter α is a measure of the catalyticity of the miRNA action on its targets and denotes the probability that R1 or R2 degradation in the miRNA complex coexists with degradation of the miRNA S (13, 14, 18). It is the least known parameter in the model and almost pure stoichiometric (α ∼ 1) or catalytic (α ∼ 0) interactions are found in the literature (19–23). When α = 0, the steady-state solution of Eq. S1 predicts that no crosstalk between ceRNAs occurs. Interestingly, however, even in a purely catalytic interaction, threshold-like behavior, ultrasensitivity near the threshold, and ceRNA crosstalk occur in out-of-equilibrium systems for timescales shorter than the time needed for the complex to reach the equilibrium (SI Materials and Methods and Figs. S2 and S3).

The simple titration mechanism so far considered gives rise to the threshold-like response of Eq. 1 (13, 14, 16). Around the threshold all species are expressed at about the same concentration and the system shows ultrasensitive responses to small changes in the miRNA and ceRNA concentrations (12, 14). The threshold is located at k_S = α(k_R1 + k_R2) and is determined by the model’s kinetic reaction rates (14): transcription, degradation, and association. The definition of these parameters presents a major challenge due to the number of biological conditions and events affecting them. Nevertheless, once these kinetic parameters are known, the model can be used for quantitative predictions.

Fig. 1C depicts an example simulation of a network consisting of two ceRNAs and one shared miRNA. Increasing the expression of ceRNA2 results in a decrease of free miRNA molecules and a concomitant elevation in the amount of unrepressed (i.e., not targeted by the common miRNA) ceRNA1. Optimal crosstalk between the two ceRNAs as determined by their Pearson correlation coefficient is near an equimolar state of miRNA and ceRNAs (Fig. 1D).

Dosage of ceRNAs and miRNAs.

We reasoned that the effect of altered expression of one ceRNA on the entire network may depend on its baseline expression levels (Fig. S1C). Using the ceRNA model, we assessed the effect of increasing expression of one ceRNA on a network that consists of three ceRNAs and one shared miRNA. Increased expression of ceRNA1 resulted in elevated levels of unrepressed ceRNA2 and ceRNA3 and a concomitant decrease of free miRNA (Fig. S2A). Interestingly, when ceRNA1 was expressed threefold higher than ceRNA2 and ceRNA3 in this system, its silencing had a pronounced effect on ceRNAs 2 and 3 (Fig. S2B). When the expression of ceRNA1 was further decreased (e.g., to 50%, 10%, and 0%), its relative effect on the other two ceRNAs was diminished (Fig. S2B). We obtained similar effects when we considered more complex systems: In a network consisting of 10 ceRNAs and 10 common miRNAs, altered expression of one ceRNA impacts the expression levels of the other 9 ceRNAs (Fig. S2C).

We then explored the effect of silencing a ceRNA in a system where two ceRNAs are expressed at increasing ratios (1:1–8:1). Importantly, silencing the higher-expressed ceRNA by 50% or 90% had a greater effect on the lower-expressed ceRNA than vice versa (Fig. 2 A and B). Thus, the effect on the network caused by the loss of a highly expressed ceRNA is greater than the effect elicited by silencing a low-abundance ceRNA. Moreover, consistent with the notion that optimal crosstalk occurs near the threshold, the greatest effect was elicited under equimolar conditions of the two ceRNAs.

Fig. 2.

Analysis of the ceRNA dosage effect in vitro. (A and B) The dosage effect predicted by the ceRNA model. ceRNA1 and -2 are expressed at different ratios (ceRNA2:ceRNA1 = 1:1–8:1 in A and ceRNA2:ceRNA1 = 1:1–1:8 in B) and the effect of silencing ceRNA2 by 50% (red bars) or 90% (blue bars) on ceRNA1 is shown. Parameters can be found in SI Materials and Methods. (C) Graph showing the VAPA:PTEN mRNA ratio in five cancer cell lines. (D) Combined expression of PTEN- and VAPA-targeting miRNAs. (E) Representative Western blot displaying the effect of VAPA silencing on PTEN expression and vice versa. P, PTEN; V, VAPA; NC, negative control. (F) Quantification of E. (G) The average percentages of ceRNA-mediated silencing of PTEN and VAPA in response to short interfering (si)RNA against VAPA and PTEN, respectively.

To validate these predictions, we explored the role of ceRNA dosage on cross-regulation of PTEN and its ceRNA VAMP (vesicle-associated membrane protein)-associated protein A (VAPA). We determined PTEN and VAPA mRNA levels as well as the expression levels of eight miRNAs that have been validated to target both PTEN and VAPA (6) in several human cell lines. VAPA is expressed at higher levels than PTEN and the VAPA:PTEN ratio varies between cell lines (Fig. S4A). Similarly, expression levels of the individual miRNAs also vary between cell lines; however, some cell lines display similar combined levels of expression for the eight miRNAs (Fig. S4 B–D).

To address the importance of the ceRNA dosage, we selected five cell lines with increasing VAPA:PTEN expression ratios (Fig. 2C), but similar combined miRNA expression (Fig. 2D) to exclude effects due to variations in miRNA expression. siRNA-mediated silencing of VAPA decreased PTEN expression in all five cell lines (Fig. 2 E and F). Notably, cell lines that displayed a lower VAPA:PTEN ratio exhibited a more robust ceRNA effect on PTEN (Fig. 2 E and F) as their VAPA:PTEN ratio is likely nearer the threshold (as described above) and thus allowing for more effective regulation of PTEN by VAPA. Conversely, silencing of PTEN only slightly reduced VAPA expression (Fig. 2 E and F). Averaging the ceRNA effect on PTEN by silencing VAPA and vice versa demonstrated that silencing the higher-expressed ceRNA (VAPA) elicits a more robust reduction in expression of the lower-expressed ceRNA (PTEN) (Fig. 2G). These in vitro findings thus corroborate our model predictions and support our theory that ceRNA dosage is critical for ceRNA activity.

The expression levels of miRNAs involved in ceRNA networks may be critical for cross-regulation (Fig. S1D). We used the ceRNA model to explore the role of miRNA concentrations on ceRNA cross-regulation in a system consisting of three ceRNAs and one common miRNA. In this prediction, transcriptions of ceRNA2 and ceRNA3 were kept constant, whereas the ceRNA1 transcription rate was increased and the miRNA expression levels were set to low, medium, and high (Fig. 3 A–C). When the miRNA was expressed at low levels, increasing the ceRNA1 transcription rate had negligible effects on ceRNA2 and ceRNA3 because they are both in an almost completely unrepressed state (Fig. 3A). In contrast, a medium or high miRNA transcription rate allowed for a significant increase of ceRNA2 and ceRNA3 in response to a higher ceRNA1 transcription rate (Fig. 3 B and C). Cross-regulation of ceRNA1 and ceRNA2—intended as the difference in the amount of free molecules between the highest and the lowest ceRNA1 expression rates—was most efficient at an intermediate miRNA transcription rate (Fig. 3D).

Fig. 3.

Analysis of the miRNA dosage in silico and in vitro. (A–C) Effect of varying miRNA expression levels on ceRNA cross-regulation. Shown are three plots representing the steady-state mean number of free molecules for ceRNA1, ceRNA2, and ceRNA3 and miRNA in a system of one miRNA and three ceRNAs as a function of ceRNA1 transcription rate. The values for the miRNA transcription rate are as follows: (B) k_s = 0.001 [1/s] (∼4 miRNA molecules per hour), (C) k_s = 0.0083 [1/s] (∼30 miRNA molecules per hour), and (D) k_s = 0.02 [1/s] (∼70 miRNA molecules per hour). The remaining parameters can be found in SI Materials and Methods. (D) Plot comparing the steady-state mean number of free molecules for ceRNA2 as a function of ceRNA1 transcription in the systems with varying miRNA expression rates shown in A–C. (E) Graph showing the VAPA:PTEN mRNA ratio in five cancer cell lines. (F) Combined expression of PTEN- and VAPA-targeting miRNAs. (G) Representative Western blot displaying the effect of VAPA silencing on PTEN expression and vice versa. P, PTEN; V, VAPA; NC, negative control. (H) Quantification of G.

We further investigated the role of miRNA concentrations in vitro in the context of PTEN and VAPA. We selected two triplets of cell lines with similar VAPA:PTEN ratios (Fig. 3E) but increasing amounts of VAPA- and PTEN-targeting miRNAs (Fig. 3F and Fig. S4 B–D). Silencing of VAPA resulted in diminished levels of PTEN in all six cell lines and, notably, the efficiency of ceRNA cross-regulation correlated with the amount of miRNA expression (Fig. 3 G and H). Taken together, our data demonstrate that ceRNA crosstalk depends on relative miRNA concentrations.

Number of MREs.

Within ceRNETs some pairs or groups of ceRNAs may share a large number of miRNAs, whereas others have only very few miRNAs in common. We determined whether a greater number of shared miRNAs intensifies cross-regulation. In this scenario, the network consists of 10 ceRNAs and 10 miRNAs; however, not every ceRNA is targeted by every miRNA (outline in Fig. S5). Increased expression of ceRNA10 had a greater effect on ceRNAs with which it has more miRNAs in common (blocks 1 and 2 in Fig. S5) and a negligible effect on ceRNAs that are targeted by nonshared miRNAs (blocks 3 and 4 in Fig. S5). This result suggests that ceRNA cross-regulation increases with the number of shared miRNAs and is weakened when a ceRNA pair is targeted by too many nonshared miRNAs.

Using TargetScan, we determined the number of miRNAs shared between PTEN and its predicted ceRNAs. Twenty-five conserved miRNA families are predicted to target PTEN, all of which could be shared with putative PTEN ceRNAs. Increasing the cutoff of shared miRNAs results in a rapid decrease in the number of putative PTEN ceRNAs (Fig. S6 A and B). Interestingly, the number of ceRNAs that share a certain number of miRNAs with PTEN depends on the miRNA combination. For instance, at a cutoff of 7 shared miRNAs, the best observed combination yields 21 PTEN ceRNAs, whereas other miRNA combinations are shared only between PTEN and one other ceRNA (Fig. S6C). This suggests that the specificity of the ceRNA interactions may be regulated by the miRNA expression profile.

Indirect ceRNA Interactions Amplify Crosstalk.

Given that each ceRNA shares multiple miRNAs with numerous other ceRNAs, the effect of losing one ceRNA may be “diluted” between several ceRNAs. However, silencing of PTEN ceRNAs has a profound effect on PTEN expression (5, 6). The existence of smaller subgroups within larger ceRNETs only partially explains such observations. We reasoned that a ceRNA might interact with other ceRNAs both directly and indirectly. We modeled such a scenario in a network consisting of three ceRNAs and three miRNAs. ceRNA1 shares miRNA1 with ceRNA2, and miRNA2 with ceRNA3, representing direct interactions (Fig. 4). ceRNA2 and ceRNA3 have miRNA3 in common, and thus ceRNA1 also interacts indirectly with ceRNA2 and ceRNA3 (Fig. 4). Silencing ceRNA1 profoundly reduces the expression of ceRNA2 and ceRNA3 (Fig. 4). However, if miRNA2 and hence the indirect interaction of ceRNA1 with ceRNA2 are eliminated from the network, the effect of silencing ceRNA1 on ceRNA2 expression is diminished (Fig. 4). These results indicate that indirect interactions critically contribute to ceRNA cross-regulation in ceRNETs.

Fig. 4.

Indirect interactions amplify ceRNA crosstalk. Silencing of ceRNA1 (black line) has a stronger effect on ceRNA2 (red line) when ceRNA1 indirectly interacts with it through ceRNA3 (blue line) (Upper) than if no indirect interaction exists (Lower). Schematic outlines of the two networks are shown in the Upper Right corner of each panel. Solid arrows depict direct interactions, whereas dashed arrows depict ceRNA1’s indirect interactions.

To investigate whether indirect interactions are common in ceRNETs we built an unbiased network of predicted ceRNA interactions by integrating the MRE analysis and conserved coexpression, with the latter based on a large collection of nearly 7,500 manually annotated samples from which many tissue-specific and multicondition conserved coexpression networks have been built and merged together (24). The obtained network contains 1,927 nodes and 1,621 edges. Its largest connected component contains 164 genes that are interconnected by sharing numerous MREs (Fig. S7A). Importantly, indirect interactions occur frequently in this ceRNET, suggesting they may contribute to ceRNA crosstalk.

Integrated Transcription Factor Networks and ceRNETs.

We next performed Gene Ontology analysis of the ceRNET to interrogate which biological processes may be subjected to ceRNA-mediated regulation. Remarkably, we found an overrepresentation of genes involved in transcription and the regulation of gene expression (Fig. S7 A and B). Intriguingly, DICER1, argonaute RISC catalytic component 1 (AGO1), and trinucleotide repeat containing 6B (TNRC6B), genes directly involved in microRNA processing were found in the largest component, suggesting that ceRNA interactions may play a role in the regulation of the microRNA machinery. Indeed, AGO1 and TNRC6B as well as several genes found in the ceRNET that are involved in regulation of gene expression are putative DICER1 ceRNAs (Fig. S7C).

On the basis of the Gene Ontology analysis, we analyzed the involvement of TFs in ceRNA networks. Specifically, we tested whether the induction of the oncogenic PAX/FKHR translocation [commonly involving paired box (PAX)3 or PAX7 and forkhead box O1 (FOXO1) in 90% of alveolar rhabdomysoarcoma (ARMS)] regulates the expression of its ceRNAs and the ceRNAs of its transcriptional targets in a dataset of RMS samples (Fig. S8A). Expression of FOXO1 and its targets MET, fibroblast growth factor receptor 4 (FGFR4), insulin-like growth factor 1 receptor (IGF1R), and anaplastic lymphoma receptor tyrosine kinase (ALK) is indeed elevated in ARMS compared with embryonal RMS (ERMS) (Fig. S8B). We next predicted putative ceRNAs for each of these five genes and excluded ceRNAs that are also putative transcriptional targets of PAX/FKHR. Notably, predicted ceRNAs of FOXO1 and its targets displayed increased expression in ARMS (Fig. 5A). Moreover, ectopic expression of the PAX3-FKHR fusion transcript in the ERMS cell line RD18 increased the expression of several predicted ceRNAs of ALK and MET (Fig. S8C). We assessed the global effect of FOXO1 overexpression on ceRNETs by analyzing FOXO1 transcriptional targets that show increased expression in ARMS compared with ERMS. For 284 FOXO1 targets we predicted 7,663 ceRNAs, and, notably, over 40% of the up-regulated genes in ARMS are ceRNAs for FOXO1 targets (Fig. 5B and Fig. S8 D and E).

Fig. 5.

TF networks underlie ceRNA regulation. (A) Graph displaying the percentage of predicted ceRNAs of FOXO1 and its targets MET, FGFR4, IGF1R, and ALK that are deregulated in ARMS. The numbers in the bars represent the number of ceRNAs in each category. No ceRNAs are down-regulated. (B) Forty-one percent of genes up-regulated in ARMS are putative ceRNAs of 284 predicted FOXO1 transcriptional targets that are up-regulated in ARMS. (C) Graphs showing the correlation between the Pearson coefficients of 37 TFs and their targets or ceRNAs. Targets and ceRNAs were ranked according to their Pearson coefficient in the high- and low-miRNA-expression subsets and the difference in average ranking between targets and ceRNAs was calculated.

To further analyze the intertwined nature of transcriptional networks and ceRNETs, we determined whether such networks are sensitive to fluctuations in miRNA expression. To this end, we studied the Pearson coefficient of coexpression between 4 TFs [androgen receptor (AR), TP53, c-Jun, and c-Fos] and their predicted ceRNAs or transcriptional targets in different subsets of human prostate cancer samples (25) defined by their average miRNA expression values.

For each TF, the ranked human prostate cancer samples were subdivided into three subsets defined by low, intermediate, and high meta_miR expression. meta_miR is defined as the combined expression of the miRNAs that target a TF. We then determined the Pearson correlation coefficient between the TFs and each target or putative ceRNA in the low- and high-meta_miR-expression subsets. We next computed the Pearson coefficient of the correlation in the low- and high-meta_miR conditions. Notably, we observed that the Pearson coefficient is greater for TFs and their targets than for TFs and their putative ceRNAs; i.e., the difference between the correlations in the low and high meta_miR is smaller for targets (Fig. S9A), suggesting that expression of TF ceRNAs is more sensitive to miRNA expression levels. These data suggest that the efficiency of ceRNA-mediated cross-regulation depends on miRNA fluctuations, whereas the activity of TFs toward their targets is more stable and independent of miRNA expression levels. We extended this analysis to 37 TFs and found in 30 cases that the correlation between the TF and its predicted ceRNAs is more sensitive to miRNA expression than that between the TF and its targets (Fig. 5C and Fig. S9B).

Discussion

By integrating mathematic, bioinformatics, and cell biology approaches we define the molecular requirements for ceRNA activity. We show that relative abundance of ceRNAs and miRNAs as well as their stoichiometry, the number of MREs, and indirect interactions all contribute to ceRNA crosstalk.

We show that the balance between miRNAs and ceRNAs is critical for ceRNA activity. Disruption of this balance affects ceRNA crosstalk and may thus promote diseases like cancer. Intriguingly, miRNA expression is globally decreased in cancers (26) and Dicer1 is a haploinsufficient tumor suppressor (27, 28) whose expression is regulated during differentiation (29). In addition, alternative polyadenylation signal utilization in cancer (30, 31) and during embryogenesis (32) results in 3′ UTRs with different lengths. Moreover, 3′ UTRs may be expressed independently of their coding sequence (33). Thus, although Dicer1 expression could serve as a critical rheostat for ceRNA interactions, aberrant expression of miRNAs and/or 3′ UTRs may critically alter regulation of ceRNETs and play important roles in normal physiology and disease.

Subnetworks contained within larger networks could explain the pronounced crosstalk alterations observed when potent ceRNAs are silenced. This notion is supported by the fact that small networks and even ceRNA pairs have evolved to regulate certain biological processes in various species (7, 8, 34, 35). Our analysis reveals that numerous indirect interactions where two ceRNAs crosstalk via a third transcript exist. Although in an unperturbed situation this complexity may ensure a tightly controlled network with minimal fluctuations, we find that indirect interactions amplify the effect elicited by loss of one ceRNA. This may explain why losing a single ceRNA in a large ceRNET can have profound effects on other components of the network, as observed previously (5, 6, 10).

We perform the mathematical analysis in simple ceRNETs but our observations are applicable to more complex ceRNETs involving numerous ceRNAs and miRNAs (36). Although our model currently allows for ceRNA identification and qualitative but not quantitative predictions, once the concentrations of all components of a ceRNET are established, as suggested by Ebert and Sharp (37), it may be suitable to predict the overall effect on a ceRNET in response to single-factor perturbations. This will enable us to rapidly screen putative disease-associated ceRNAs in silico before a thorough molecular evaluation.

Finally, we find that ceRNA and transcriptional networks are intertwined as TF mRNAs and their TF target mRNAs crosstalk with other RNA transcripts. This suggests the intriguing possibility that a transcription factor’s regulatory potency is much larger than anticipated. In turn, this has important implications not only for cellular and organismal physiology, but also for the pathogenesis of cancer and other diseases where aberrant expression of TFs occurs.

Materials and Methods

miRNA–mRNA Analysis.

The number of conserved miRNA families per target and the number of targets per conserved miRNA family were determined using TargetScan (release 6.1) (www.targetscan.org/vert_61/vert_61_data_download/Predicted_Targets_Info.txt.zip).

ceRNA Prediction.

On the basis of the miRNA–target prediction by TargetScan, we obtained a ranked list of putative ceRNAs for all genes currently featured in the TargetScan database as described previously (5, 38). Only putative ceRNAs in the first percentile were analyzed further.

miRNA Processing Network.

From the putative ceRNA lists, we defined a ceRNA network linking two genes A and B if A is within the top 40 genes in terms of distance from B and vice versa. We further refined this ceRNA network by removing all edges that are not found in the conserved coexpression networks built by Piro and colleagues (24).

Gene Silencing, Quantitative PCR, and Western Blotting.

siRNA-mediated knock-down, quantitative (q)PCR, and Western blot analyses were performed as described previously (5, 6). qPCR primer sequences for ALK and MET ceRNAs were obtained from the Harvard primer bank (http://pga.mgh.harvard.edu/primerbank/).

Virus Production and Transduction.

Lenti-PAX3-FKHR was generated by cloning PAX3-FKHR from pBabe-puro/PAX3-FKHR (B. Schafer, University Children’s Hospital, Zurich) into the pCCL.sin.PPT.hPGK.GFP.Wpre transfer vector (L. Naldini, San Raffaele-Telethon Institute for Gene Therapy, Milan). Concentrated lentiviral vector stocks were produced as previously described (39) and cells transduced using standard techniques.

Cell Culture.

PWR1E and RWPE1 cells were cultured in RPMI-1640 media, and all other cell lines were cultured in DMEM. Media were supplemented with 10% (vol/vol) FCS and L-glutamine and cells were maintained at 37 °C in a humidified atmosphere with 5% (vol/vol) CO₂.

Simulation Parameters and Transcription Factor and PAX3-FKHR Analysis.

These methods can be found in SI Materials and Methods.

Statistical Analysis.

Cell biological experiments were performed three times. Averages and SEs are shown. Values of P < 0.05 were considered statistically significant. Correlation was based on the Pearson correlation coefficient, differential expression was based on the bioconductor limma package with the statistical significance set to a Benjamini–Hochberg-corrected P value < 0.05, enrichments in FOXO analysis were evaluated by the R phyper function, and the P value was adjusted by Bonferroni correction.

Note Added in Proof.

While this paper was in press, another paper was published describing a similar mathematical model (40).