TL;DR
This paper introduces LMSM, a novel computational framework that identifies modules of long non-coding RNAs (lncRNAs) and their target messenger RNAs (mRNAs) that collectively act as "miRNA sponges" in breast cancer. By considering the coordinated behavior of these molecules within modules, LMSM offers a more comprehensive understanding of gene regulation and identifies potential biomarkers for breast cancer subtypes.
What Problem Does This Paper Address?
Long non-coding RNAs (lncRNAs) are increasingly recognized for their crucial roles in gene regulation, particularly as "miRNA sponges." These lncRNAs can bind to microRNAs (miRNAs), effectively sequestering them and preventing them from regulating their target mRNAs. This "competing endogenous RNA" (ceRNA) mechanism creates a complex regulatory network. While previous methods focused on identifying individual lncRNA-miRNA-mRNA interactions or networks, they often overlooked the fact that genes and regulatory elements often function in coordinated groups or "modules." This paper addresses the need for a method that can identify these coordinated regulatory modules, hypothesizing that the "tug-of-war" for miRNAs occurs at a modular level rather than just between individual molecules. Understanding these modules is crucial for deciphering the complex regulatory landscape of diseases like breast cancer.
Methods
The core of the LMSM framework involves a two-stage approach:
- Identifying lncRNA-mRNA Co-expression Modules: The Weighted Gene Co-expression Network Analysis (WGCNA) method is employed to identify groups of lncRNAs and mRNAs that exhibit similar expression patterns across samples. WGCNA is a well-established technique for finding modules of co-expressed genes.
- Inferring lncRNA-related miRNA Sponge Modules (LMSM Modules): From the co-expression modules identified in stage 1, LMSM applies three key criteria to pinpoint true miRNA sponge modules:
- Significant miRNA Sharing: A hypergeometric test is used to determine if the lncRNAs and mRNAs within a module share a statistically significant number of miRNAs, indicating a potential for competition.
- High Canonical Correlation: Canonical correlation analysis (CCA) is performed to assess the collective positive correlation between the expression levels of the lncRNA group and the mRNA group within a module. This captures the "tug-of-war" at a modular level.
- Adequate Sensitivity Canonical Correlation: This criterion, adapted from sensitivity correlation analysis, evaluates the influence of shared miRNAs on the modular competition. It measures the correlation between the lncRNA and mRNA groups after accounting for the effect of shared miRNAs.
Modules that satisfy all three criteria are designated as LMSM modules. The statistical significance of these modules is then evaluated using a null model approach based on sensitivity correlation.
Datasets
The study primarily utilizes the The Cancer Genome Atlas (TCGA) breast cancer (BRCA) dataset. This dataset includes matched miRNA, lncRNA, and mRNA expression data, as well as clinical information.
For miRNA-target interactions, the researchers integrated data from several high-confidence databases:
- miRNA-lncRNA interactions: NPInter v3.0 and DIANA-LncBase v2.0.
- miRNA-mRNA interactions: miRTarBase v8.0, TarBase v7.0, and miRWalk v2.0.
BRCA-related genes were sourced from DisGeNET v5.0 and COSMIC v86, while BRCA-related lncRNAs were obtained from LncRNADisease v2.0, Lnc2Cancer v2.0, and MNDR v2.0.
Key Insights and Ideas
- Modular Competition Hypothesis: The central idea is that miRNA sponging doesn't just happen between individual molecules but rather between coordinated groups (modules) of lncRNAs and mRNAs competing for shared miRNAs. This shifts the focus from pairwise interactions to collective regulatory units.
- Mediating miRNAs as Crosslinks: The study found that most miRNAs involved in these sponge modules act as "crosslinks," connecting multiple LMSM modules. This highlights the interconnectedness of the regulatory network.
- BRCA Subtype Specificity: A significant portion of the identified LMSM modules were found to be specific to particular subtypes of breast cancer, suggesting their potential role in driving subtype-specific disease progression.
- Predictive Power: LMSM not only identifies modules but also demonstrates a capability for predicting miRNA targets, further validating its biological relevance.
Key Results
- 17 LMSM modules identified in the TCGA breast cancer dataset.
- Most LMSM modules are statistically significant (adjusted p-value = 1.00E-06).
- 10 out of 17 LMSM modules are functionally enriched in BRCA (p-value < 0.05).
- 15 out of 17 LMSM modules act as module biomarkers in BRCA (Log-rank p-value < 0.05 and HR > 2).
- Most identified LMSM modules are BRCA subtype-specific (9 unique modules identified as subtype-specific).
- LMSM modules significantly outperform a baseline method in classifying BRCA subtypes, with higher Subset accuracy (0.7547 vs. 0.3800) and lower Hamming loss (0.0892 vs. 0.2480).
- The method is robust, as demonstrated by consistent results when using an alternative co-expression module identification method (SGFA).
- Comparison with a graph clustering-based strategy (GC) showed LMSM consistently outperformed GC in terms of BRCA-related modules, module biomarkers, classification performance, and validated interactions.
Limitations and Future Work
- Experimental Validation of Modules: The authors acknowledge that validating entire modules experimentally is challenging due to their size. They suggest that biologists can select key lncRNAs and mRNAs from identified modules for targeted validation.
- Generalizability: While the method was applied to breast cancer, the authors believe it can be extended to other human cancer datasets.
- Parameter Sensitivity: LMSM is a parametric model, and the parameter settings can be adjusted for stricter or looser criteria, which could be explored further.
- Extension to Other ncRNAs: The framework is flexible and can be adapted to study other non-coding RNAs like circRNAs and pseudogenes by modifying the input data.
Summary updated