Weighted Gene Co-expression Network Analysis (WGCNA)

WGCNA transforms a correlation matrix into an adjacency matrix by raising the absolute correlation to a power β. This soft thresholding emphasizes strong correlations while penalizing weak ones, preserving the continuous nature of the co-expression information. The optimal β is chosen to achieve a scale-free topology, a network property observed in biological systems where a few hub nodes have many connections. The pickSoftThreshold function evaluates multiple powers and recommends the lowest value where the scale-free topology fit index (R²) exceeds 0.8.

WGCNA uses Topological Overlap to measure gene interconnectedness, not just direct correlation. TOM quantifies the shared neighborhood of two genes—if they are connected to the same set of other genes, their TOM is high. This metric is more robust than correlation alone for detecting functional modules. The resulting TOM dissimilarity matrix (1 - TOM) is used as input for hierarchical clustering, producing a dendrogram where branches represent distinct co-expression modules.

Each identified module is mathematically summarized by its module eigengene (ME), defined as the first principal component of the module's expression matrix. The ME captures the dominant expression pattern of the module, reducing dimensionality from potentially hundreds of genes to a single representative profile. This allows researchers to:

• Correlate entire modules with external clinical traits (e.g., tumor stage, survival time)
• Identify modules significantly associated with phenotypes of interest
• Perform eigengene network analysis to study higher-order relationships between modules

WGCNA directly links co-expression modules to external sample traits by correlating module eigengenes with clinical variables. The output is a heatmap of module-trait associations, displaying correlation coefficients and p-values for each module-trait pair. This reveals which gene clusters are significantly associated with conditions like disease status, treatment response, or patient survival. Gene Significance (GS) measures the correlation of individual genes with a trait, while Module Membership (MM) measures how well a gene represents its module—genes with high GS and MM are prime biomarker candidates.

Within each biologically relevant module, WGCNA identifies hub genes—the most highly connected nodes that are central to the network's structure. Hub genes are defined by high intramodular connectivity (kIM), meaning they have strong co-expression relationships with many other module members. These genes are often:

• Master regulators or transcription factors
• Functionally critical to the module's biological role
• High-priority therapeutic targets Hub gene status is validated by plotting Gene Significance vs. Module Membership; top-right quadrant genes are both trait-relevant and centrally connected.

WGCNA provides statistical methods to test whether a co-expression module identified in a reference dataset is preserved in an independent test dataset. The modulePreservation function computes Zsummary and medianRank statistics, which assess the conservation of network density and connectivity patterns. A Zsummary > 10 indicates strong preservation, while values < 2 suggest the module is not reproducible. This is critical for validating biomarker modules across independent cohorts before committing to expensive validation studies.

A critical parameter in WGCNA that transforms the correlation matrix into an adjacency matrix by raising absolute correlations to a power β. This soft thresholding preserves the continuous nature of co-expression information, unlike hard thresholding which dichotomizes connections. The optimal β is chosen to achieve a scale-free topology fit (R² > 0.8), ensuring the network exhibits biologically realistic hub gene structures. The pickSoftThreshold function evaluates multiple powers and plots the scale-free topology model fit against mean connectivity.

A robust measure of network interconnectedness that quantifies not just direct correlation between two genes, but also the extent to which they share common network neighbors. TOM is calculated as:

• TOM_ij = (a_ij + Σ_u a_iu * a_uj) / (min(k_i, k_j) + 1 - a_ij) This metric is superior to raw correlation for identifying functional modules because it filters out spurious isolated connections and reinforces densely connected clusters. The resulting dissimilarity (1 - TOM) serves as input for hierarchical clustering.

The first principal component of a gene module's expression matrix, representing the dominant expression pattern of the entire module. The ME serves as a summary profile that:

• Reduces dimensionality from hundreds of genes to a single vector
• Enables correlation with external clinical traits (tumor stage, survival time)
• Identifies modules significantly associated with phenotypes of interest
• Facilitates module–trait relationship heatmaps for intuitive visualization of biologically relevant modules

Two complementary metrics for prioritizing genes within modules:

• Gene Significance (GS): The absolute correlation between a gene's expression and an external trait. High GS indicates a trait-associated gene.
• Module Membership (MM): The correlation between a gene's expression profile and its module eigengene, also called kME. High MM identifies hub genes that are central to the module's structure. A strong GS–MM correlation validates that the module's internal organization is biologically relevant to the trait, not an artifact.

A statistical framework for assessing whether network modules identified in a reference dataset are also present in an independent test dataset. Key statistics include:

• Zsummary score: Values > 10 indicate strong preservation, 2–10 weak-to-moderate, < 2 no preservation
• medianRank: An aggregate ranking across multiple preservation statistics This analysis validates the reproducibility and biological robustness of co-expression modules across cohorts, platforms, or species, distinguishing true biological signals from dataset-specific noise.

An adaptive algorithm for identifying modules from the hierarchical clustering dendrogram of the TOM-based dissimilarity matrix. Unlike static cut methods that slice at a fixed height, dynamic tree cut:

• Adapts to the shape of branches, detecting nested clusters
• Uses a bottom-up merging step to combine highly similar clusters
• Requires parameters like deepSplit (0–4) controlling sensitivity and minClusterSize setting the minimum module size This method produces more biologically coherent modules than arbitrary height cutoffs.