WGCNA: A Visual Guide to Gene Co-expression Networks

What “co-expression” actually means

Imagine two genes, GENE-A and GENE-B. You have expression measurements for both across many samples — say, 40 patients, or 30 timepoints. If GENE-A tends to go up in the same samples where GENE-B goes up, and down where it goes down, the two genes are co-expressed. The correlation across samples (Pearson's r, usually) captures that pattern in a single number between −1 and 1.

Now scale that idea up. With 15,000 genes you can compute every pairwise correlation — over 100 million numbers. That matrix is a network: every gene is a node, every correlation is the weight of an edge. Strong positive correlations mean tight co-regulation. Weak correlations mean the two genes have little to say to each other.

The problem: a raw correlation network is dense, noisy, and useless to read. WGCNA — Weighted Gene Co-expression Network Analysis — is a set of transformations that turn that wall of numbers into something a biologist can actually interpret: a handful of modules (gene groups), one eigengene per module (a summary expression profile), and a ranked list of hub genes per module (the genes most representative of the module's pattern).

Why “weighted” — soft thresholding

The naive way to build a gene network is to pick a correlation cutoff (say r > 0.7) and call any pair above it “connected.” This is a hard threshold. It is also, unfortunately, fragile: a gene pair at r = 0.69 is treated identically to a pair at r = 0.01 (both unconnected), and a pair at r = 0.71 is treated identically to a pair at r = 0.99 (both fully connected). All the continuous information in the correlation is thrown away at the cutoff.

WGCNA's central trick is to keep the correlations continuous but reweight them. Each correlation is raised to a power β (the “soft threshold,” usually between 6 and 12):

a_ij = |cor(i, j)|^β

Because β is large, weak correlations get crushed toward zero (0.36 = 0.0007) while strong correlations stay almost intact (0.96 = 0.53). The network ends up dominated by the strong, biologically credible edges — without ever drawing a hard line in the sand.

How is β chosen?

Real biological networks have a characteristic shape: a few genes have lots of connections (hubs), and most genes have only a handful. Mathematically, the distribution of node degrees follows a power law — this is called scale-free topology. Random networks do not look like this. WGCNA scans candidate β values from 1 to ~30 and picks the smallest β at which the network's scale-free fit (R²) plateaus near 0.8–0.9. That is usually somewhere between 6 and 12 for typical RNA-seq data.

Modules — what they actually represent

Once correlations are reweighted (the “adjacency matrix”), WGCNA does one more transformation. It computes the topological overlap matrix (TOM): a similarity measure that says “two genes are similar if they share many of the same neighbours in the network.” TOM is more robust than raw correlation because it considers second-order structure — not just whether A and B are connected, but whether A and B both connect to C, D, E, and F.

Hierarchical clustering on TOM-based dissimilarity produces a gene dendrogram. A dynamic tree-cut algorithm then carves the branches into modules. Each module is assigned a colour name — Turquoise, Blue, Brown, Yellow, Green, and so on. The colours are sorted by module size: Turquoise is always the largest module, Blue the second largest, etc. The order is conventional and arbitrary — the colour itself has no biological meaning.

What does a module represent?

A module is a group of genes that share an expression pattern across your samples. Often (but not always) that shared pattern reflects:

• A common biological pathway — e.g. mitochondrial respiration, antigen presentation, cell cycle.
• A shared regulatory program — downstream targets of the same transcription factor or signalling axis.
• A cell-type signature — in bulk RNA-seq from heterogeneous tissue, modules often track shifting cell-type proportions.
• Sometimes, a technical artifact — batch, RIN, or sequencing depth (these are real and need to be ruled out).

Each module is summarised by its eigengene — the first principal component of the module's expression matrix. The eigengene is a single vector across samples that captures the module's overall up-and-down pattern. It is what you correlate with traits in the next step.

Hub genes — the network's spotlight

Inside any module, genes are not equal. Some sit on the periphery — loosely connected to a handful of others. A few sit at the centre: highly connected, tightly correlated with the module eigengene, and statistically the most representative members of the group. Those are the hub genes.

The standard metric is module membership (MM): the correlation between a gene's expression profile and its module eigengene. A gene with MM = 0.95 essentially is the module's pattern. A gene with MM = 0.55 is a peripheral member — it correlates somewhat, but not strongly. Sort by MM, take the top few, and you have your candidate hub list.

Hub genes often turn out to be biologically central: master regulators, key effectors, or marker genes for the cell type or process the module represents. That is why people pay attention to them. But there is a major caveat — one of the most important pieces of context for any new WGCNA user.

Module-trait correlation — connecting to phenotype

This is the step that makes WGCNA more than just a clustering algorithm. You take each module eigengene (one vector per module, length = number of samples) and correlate it with every sample-level trait you have measured: disease status, tumour grade, drug dose, age, survival, treatment response. The result is a matrix — modules along the rows, traits along the columns, correlation strength in each cell. A module that correlates strongly with a trait is a candidate molecular signature of that trait.

The classic worked example is a TCGA cancer cohort. Take TCGA-BRCA breast cancer with hundreds of patient tumours and clinical annotations — ER status, HER2 status, tumour grade, days to death. Run WGCNA on the expression matrix, then correlate every module eigengene with every trait. You usually end up with a heatmap like Figure 5.

Strong cells in that heatmap point to interpretable hypotheses: a brown module that is positively correlated with tumour grade and negatively correlated with survival is a candidate poor-prognosis signature. The hub genes inside that module are candidate prognostic markers. Pathway enrichment on the module gives you the biological theme — cell cycle, hypoxia response, EMT, etc.

What “strong” means

Module-trait correlations of |r| > 0.5 with a small p-value (corrected for multiple testing across modules and traits) are usually worth following up. |r| > 0.7 is striking. Anything below |r| ~ 0.3 is probably noise unless the sample size is very large.

When WGCNA works (and when it doesn't)

WGCNA is a correlation-based method. Everything it produces depends on the quality and quantity of the correlations it can estimate from your data. That sets some real, practical limits on when it is useful.

When in doubt: if you have a small, balanced two-group comparison, reach for DESeq2 / edgeR / limma first. Save WGCNA for cohorts where the variation across samples is itself part of the question.

Six pitfalls people fall into

Most of the disappointing WGCNA results in the wild come from a small handful of recurring mistakes. They are easy to spot once you know what to look for.

Should I run WGCNA?

A practical flowchart, derived from the “when it works” rules. Walk through the questions in order — if you cannot answer yes at any node, you probably want a different method.

How TransXplorer handles WGCNA

The full WGCNA workflow is wired into the TransXplorer app — no R, no script-tuning, no figure assembly. Here is what runs automatically when you open the Co-expression Network panel:

• Soft-threshold selection. TransXplorer sweeps β from 1 to 30 and picks the smallest value where the scale-free fit plateaus (R² ≥ 0.85). The full scale-free fit curve is shown so you can override the auto-pick if needed.
• TOM construction and module detection. Topological overlap is computed, hierarchical clustering is run, and the dynamic tree-cut produces modules. Defaults follow the WGCNA tutorial: minModuleSize = 30, mergeCutHeight = 0.25.
• Module-trait correlation panel. If you have uploaded a metadata sheet, every module eigengene is correlated against every annotated trait. The heatmap is interactive: click a cell to drill down to the underlying genes.
• Hub gene table. For each module, genes are ranked by module membership (MM). The top hubs are exported to a sortable table with gene symbol, MM, intramodular connectivity, and a link to GeneCards / Open Targets.
• Auto pathway enrichment per module. GO BP/MF/CC and KEGG enrichment runs on every module without an extra click — results are tabbed inside each module's drawer.
• Export-ready outputs. Module assignment table (CSV), eigengene matrix, hub gene lists, and high-resolution PDF of every figure.

The papers worth reading

If you want to go deeper into the WGCNA algorithm, the soft-thresholding mathematics, or eigengene network theory, these are the canonical sources. The two BMC Bioinformatics papers (2008, 2007) are the most directly useful for practitioners.