TL;DR
This paper introduces EMDN, a novel algorithm that integrates DNA methylation and gene expression data to identify "epigenetic modules" – groups of genes acting together under epigenetic control. Unlike previous methods, EMDN doesn't assume a correlation between methylation and expression, leading to more accurate identification of these modules, which can serve as biomarkers for cancer subtypes and patient survival.
What Problem Does This Paper Address?
Understanding how DNA methylation, a key epigenetic mechanism, influences gene expression and ultimately disease progression is a major challenge. While DNA methylation can inhibit gene binding and is implicated in disease, the precise relationship between methylation patterns and gene expression is complex and not always straightforward (e.g., sometimes positive correlations are observed, not just the expected negative ones). Existing computational methods often struggle with this complexity, either by relying on single data types or by making assumptions about the correlation between methylation and expression that don't always hold true. This limits their ability to accurately identify functional epigenetic modules that are crucial for understanding cancer mechanisms.
Methods
The core of this work is the Epigenetic Module based on Differential Networks (EMDN) algorithm. EMDN takes a multi-network approach, constructing two "differential" networks: one for gene co-expression and one for DNA methylation.
- Differential Network Construction:
- Coexpression/Comethylation Networks: First, standard coexpression (for gene expression) and comethylation (for DNA methylation) networks are built.
- Differential Weighting: Crucially, these networks are then weighted based on the differential expression or methylation between tumor and normal samples. This means edges connecting genes that are both coexpressed/comethylated and show significant differences between conditions are given higher importance. This step avoids pre-specifying the correlation direction between methylation and expression.
- Epigenetic Module Discovery:
- M-module Algorithm: EMDN then employs a modified version of the M-module algorithm to find common modules (groups of interconnected genes) present in both the differential coexpression and differential comethylation networks. These common modules are defined as the epigenetic modules.
- Seed Prioritization & Expansion: The algorithm uses a seed prioritization step to identify important starting genes and then iteratively expands these seeds to discover modules.
- Refinement: Candidate modules smaller than a certain size are removed, and overlapping modules are merged.
This multi-network strategy is advantageous because it integrates information from both data types simultaneously and avoids the problematic assumption of a fixed correlation direction.
Datasets
- TCGA Breast Cancer Data: The primary dataset used for training and evaluation is from The Cancer Genome Atlas (TCGA). This includes matched DNA methylation (Illumina 450k) and gene expression (RNA-seq) data for 869 breast cancer samples (785 tumor, 84 normal). Clinical information was also obtained.
- External Validation Data: For validation of subtype prediction, data from GSE60185 (Illumina 450k) was used, comprising 285 samples.
The paper does not explicitly introduce new benchmark datasets but rather uses publicly available, large-scale datasets to evaluate its method against existing ones.
Key Insights and Ideas
- Correlation-Agnostic Integration: The most significant insight is the development of a method that doesn't require pre-defining the correlation direction (positive or negative) between DNA methylation and gene expression. This is crucial because biological reality is more nuanced than a simple negative correlation.
- Multi-Network Synergy: The paper demonstrates that integrating information from multiple, distinct biological networks (coexpression and comethylation) leads to more robust and accurate identification of functional modules compared to single-network approaches.
- Epigenetic Modules as Biomarkers: The research highlights that these identified epigenetic modules are not just theoretical constructs but have tangible clinical relevance, proving useful for predicting breast cancer subtypes and patient survival.
- Both Positive and Negative Correlations Matter: The study confirms that both positively and negatively correlated epigenetic modules are biologically significant and contribute to disease characteristics.
Key Results
- Performance on Artificial Networks: EMDN outperformed other algorithms (Consensus Clustering, MolTi, Spectral Clustering) in identifying modules in simulated networks, achieving higher accuracy (NMI scores).
- Performance on TCGA Data:
- EMDN identified a higher number of modules significantly enriched in known pathways compared to other methods.
- EMDN achieved significantly higher accuracy (AUC of 0.73) in predicting breast cancer subtypes compared to other algorithms (e.g., FEM AUC 0.63, EpiMod-Exp AUC 0.64). This performance was validated on an external dataset.
- EMDN modules were significantly better correlated with gene expression than modules identified by other methods.
- Survival Analysis: A significant portion of EMDN-identified modules (7 out of 26) were significantly associated with patient survival time, demonstrating their clinical utility. Both positively and negatively correlated modules showed this association.
Limitations and Future Work
- Generalizability: While the study uses breast cancer as a proof-of-principle, the authors suggest the algorithm is a generalized framework that could be modified for other cohorts and data types (e.g., Illumina 27k methylation, Chip-seq).
- Data Integration: The authors propose that integrating additional data types, such as epigenome data and copy number variations (CNVs), could further enhance the identification of complex molecular events in heterogeneous data.
- Parameter Optimization: The paper mentions that the number of seeds is a parameter that affects performance, with 5% found to be optimal in their tests. Further exploration of parameter sensitivity could be beneficial.
- Module Definition: The current definition of epigenetic modules relies on common modules in differential networks. Future work could explore alternative definitions or integration strategies.
Summary updated