Competing endogenous RNA (ceRNA) is a regulatory mechanism where RNA transcripts compete for shared microRNA binding sites. Under the conventional view, microRNAs (miRNAs) suppress their target messenger RNAs (mRNAs) by binding to miRNA response elements (MREs). The ceRNA hypothesis adds a reversal of this logic: any RNA transcript containing MREs can effectively act as a sponge for miRNAs, thereby releasing other RNAs that share the same MREs from suppression.
Long non-coding RNAs (lncRNAs) are particularly important ceRNA sponges because they can harbor many MREs without being translated into proteins. When a lncRNA is upregulated, it absorbs miRNA molecules that would otherwise suppress target mRNAs, increasing mRNA expression levels. This lncRNA-miRNA-mRNA regulatory axis creates a three-way interaction network that can drive cancer progression when dysregulated.
Evidence for ceRNA networks in multiple cancer types had accumulated before this study, but the specific lncRNA-miRNA-mRNA regulatory network in bladder cancer (BC) had not been systematically characterized using large-scale next-generation sequencing data. Given that bladder cancer prognosis had not substantially improved over decades, identifying novel molecular regulatory mechanisms was needed to develop better prognostic biomarkers.
This study used TCGA RNA sequencing data from 414 bladder cancer and 19 normal bladder samples to construct a comprehensive ceRNA network, applying weighted gene co-expression network analysis (WGCNA) and machine learning to identify prognostically significant mRNAs, then building the full regulatory network of lncRNAs and miRNAs around those prognostic genes.
Differentially expressed mRNAs and lncRNAs were identified between 414 bladder cancer and 19 normal tissue samples from TCGA. Using the EdgeR R package with thresholds of adjusted p-value less than 0.05 and absolute log2 fold change of at least 1, the analysis identified 841 upregulated and 1,408 downregulated mRNAs, along with 788 differentially expressed lncRNAs (249 upregulated and 539 downregulated).
Weighted gene co-expression network analysis (WGCNA) was applied to identify which genes cluster together in modules that correlate with disease status. The top 60% variance mRNAs (7,836 genes) from 402 samples were analyzed using soft power 8, generating 25 co-expression modules. The yellow and lightcyan modules showed the strongest positive correlation with the bladder cancer versus normal trait and were selected as key modules containing 415 mRNAs enriched in focal adhesion, PI3K-Akt, and MAPK signaling pathways.
A parallel WGCNA analysis on the top 60% variance lncRNAs (8,604 lncRNAs) with soft power 4 generated 33 lncRNA modules, of which the pink module (357 lncRNAs) showed the strongest trait correlation. These 357 lncRNAs were submitted to MiRcode to predict which miRNAs they could bind, yielding 208 predicted miRNAs, which were cross-referenced against the top 400 expressed miRNAs in TCGA to yield 55 overlapping miRNAs.
The 55 miRNAs were used to predict target mRNAs across four databases (miRDB, TargetScan, miRTarBase, and starBase), yielding 2,266 predicted mRNAs. Overlapping these with the 2,249 differentially expressed mRNAs and the 415 WGCNA module mRNAs produced 86 candidate mRNAs -- all of which were downregulated in bladder cancer -- as the final input for prognostic model construction.
A four-gene prognostic model was constructed from the 86 candidate mRNAs using univariate Cox regression followed by LASSO regularization. From 186 TCGA bladder cancer patients with both overall survival and disease-free survival data, univariate Cox regression identified 7 genes significantly associated with OS at p less than 0.05. LASSO regression on these 7 genes further reduced the model to 4 genes: ACTC1, FAM129A, OSBPL10, and EPHA2.
The risk score for each patient was calculated as a linear combination of the 4 gene expression values weighted by their LASSO regression coefficients: risk score = (0.025202 x ACTC1 expression) + (0.114802 x FAM129A expression) + (0.271164 x OSBPL10 expression) + (0.118639 x EPHA2 expression). Lower risk scores indicate lower risk of recurrence.
Patients were divided into high-risk (n = 23) and low-risk (n = 80) groups using a risk score cutoff of 0.3073 in the training set (103 patients). Chi-square tests confirmed that training and testing sets (83 patients) were balanced across age, gender, tumor subtype, grade, T stage, TNM stage, and survival status, validating the reliability of the 5:4 random split.
For the ceRNA network construction, the 4 prognostic mRNAs were connected to 14 miRNAs predicted to target them. From 14 miRNAs, 649 lncRNAs were predicted as potential MRE-sharing sponges, and intersecting these with the 788 differentially expressed lncRNAs yielded 48 lncRNAs. The final ceRNA network contains 4 mRNAs, 14 miRNAs, and 48 lncRNAs visualized using Cytoscape.
The 4-gene risk score significantly predicted both overall survival and disease-free survival in the full cohort of 186 patients. Kaplan-Meier analysis showed that high-risk patients had substantially lower OS (p less than 0.001) and DFS (p = 0.04) compared to low-risk patients. This prognostic separation was maintained separately in both the training set (OS p less than 0.001, DFS p = 0.045) and the internal testing set.
Time-dependent ROC analysis in the training set showed AUC values of 0.665, 0.734, and 0.735 for 1-, 3-, and 5-year overall survival prediction, respectively. The increasing AUC at longer time horizons suggests the risk score captures a sustained biological difference between groups rather than a short-term treatment effect, consistent with the model capturing an underlying molecular subtype.
Chi-square analysis of the correlation between risk score groups and clinicopathological characteristics identified two significant associations: T stage (p = 0.030) and TNM stage (p = 0.017) both differed significantly between high-risk and low-risk groups, with high-risk patients showing higher proportions of T3-4 and stage III-IV disease. Age, gender, tumor subtype, and grade did not significantly differ between groups.
The significant correlation between risk score and tumor stage validates the biological meaningfulness of the model. T stage reflects the depth of tumor invasion into the bladder wall and surrounding tissue, which is the primary determinant of treatment intensity in bladder cancer. A molecular signature that captures stage-associated risk independent of direct staging adds complementary prognostic information to conventional clinical assessment.
EPHA2, a receptor tyrosine kinase, is the most highly connected mRNA node in the constructed ceRNA network. EPHA2 is simultaneously regulated by nine miRNAs in the network: miR-124, miR-141, miR-200a, miR-26a, miR-26b, miR-302b, miR-506, miR-520d, and miR-95. This broad miRNA regulation makes EPHA2 a potential hub through which multiple upstream lncRNA sponges converge to control a single oncogenic output.
EPHA2 is a receptor for progranulin and plays roles in tumor angiogenesis and cancer progression. Prior studies showed that the immunosuppressive drug leflunomide inhibited angiogenesis in a bladder carcinogenesis model by targeting the sEphrin-A1/EphA2 signaling system, suggesting EPHA2 as a therapeutically relevant pathway in bladder cancer. The network context of EPHA2 regulation through miR-26b, which has been shown to detect invasive bladder tumors with 94% specificity and 65% sensitivity in blood, further connects the ceRNA network to clinically relevant biology.
MiR-26a and miR-26b have established roles as tumor suppressors in bladder cancer, inhibiting cell proliferation and migration when overexpressed. Their inclusion as miRNA nodes that regulate EPHA2 in the constructed network is consistent with the known biology, providing partial experimental validation for the predicted network structure even without direct experimental confirmation of the full network.
FAM129A, ACTC1, and OSBPL10 -- the other three prognostic genes -- also have literature support in the ceRNA context. FAM129A/miR-195 interaction has been associated with bladder cancer development, ACTC1 is connected to miR-133a activity, and OSBPL10 is regulated by miR-124 along with EPHA2. The convergence of multiple miRNAs on both OSBPL10 and EPHA2 through miR-124 creates a particularly densely connected regulatory node in the network.
The finding that all 86 candidate mRNAs were downregulated in bladder cancer is mechanistically informative. These mRNAs were selected as target genes of the 55 miRNAs predicted from the pink module lncRNAs. Their universal downregulation suggests that the lncRNA sponging activity captured by the ceRNA model is being overwhelmed in bladder cancer -- that is, miRNA-mediated suppression of these mRNAs increases when the miRNA-sponging lncRNAs are dysregulated.
GO pathway analysis of the 86 downregulated mRNAs showed enrichment in ameboid-type cell migration, which is the mode of cell movement used by tumor cells during invasion. Several tumor cell lines show inherent amoeboid morphology, and the association of these candidate mRNAs with cell migration pathways provides a biologically plausible mechanism by which their dysregulation could promote bladder cancer invasion and progression.
The study has several important limitations. No in vitro or in vivo validation of the ceRNA network was performed -- all interactions between lncRNAs, miRNAs, and target mRNAs were computationally predicted rather than experimentally confirmed. The study is entirely retrospective and based on TCGA data without prospective clinical validation. The small size of the high-risk group (23 patients versus 80 low-risk patients) limits the statistical power for subgroup analyses.
The specific signaling pathways through which this regulatory network influences bladder cancer biology remain unknown. While the network structure was constructed and the prognostic utility demonstrated, the mechanistic step linking lncRNA sponging activity to downstream oncogenic signaling -- which pathways are affected and through what intermediate steps -- was not characterized and requires experimental follow-up.
This study established a framework connecting molecular ceRNA regulation to bladder cancer prognosis through a bioinformatics-driven discovery pipeline. By integrating differential expression analysis, co-expression network modeling, multi-database miRNA-mRNA target prediction, and survival analysis, the study constructed a lncRNA-miRNA-mRNA ceRNA regulatory network specifically tailored to the biology of bladder urothelial carcinoma.
The 4-gene prognostic model (ACTC1, FAM129A, OSBPL10, EPHA2) demonstrated consistent survival stratification across training and internal test sets, with AUCs of 0.735 for 5-year OS prediction and significant associations with tumor stage -- suggesting clinical relevance that extends beyond purely molecular characterization.
The ceRNA network of 48 lncRNAs, 14 miRNAs, and 4 mRNAs provides a resource for prioritizing experimental validation studies. Each edge in the network represents a predicted regulatory interaction that, if experimentally confirmed, could identify new therapeutic targets or diagnostic markers. The identification of EPHA2 as the most broadly regulated node, with nine targeting miRNAs, makes it a high-priority candidate for downstream functional studies.
Prospective clinical validation and experimental confirmation of key network interactions will be necessary before this framework can be translated into clinical practice. Future work should focus on confirming the lncRNA sponging activity of the most clinically relevant nodes, validating the risk score in independent multi-institutional cohorts, and identifying which regulatory interactions can be therapeutically targeted to interrupt the ceRNA-driven pathways driving bladder cancer progression.