A correlation network is a hypothesis generator wearing the costume of a result. That is not a criticism — it is the correct expectation. WGCNA, ARACNe, and their descendants are excellent at collapsing ten thousand noisy measurements into a few interpretable modules, and terrible at proving that one gene regulates another. This post covers the machinery honestly.
WGCNA: co-expression modules as a first pass
Weighted correlation network analysis (Langfelder & Horvath 2008) builds a signed network from pairwise gene-gene correlations, raises them to a soft power β chosen to approximate scale-free topology, then clusters the resulting similarity matrix into modules. Genes in a module move together across samples; the module's first principal component — the eigengene — summarises its behaviour and can be correlated with traits.
- Pick β from the scale-free fit and mean connectivity diagnostic, not by habit.
- Use signed or signed-hybrid networks; unsigned networks mix up- and down-regulation into the same module.
- Set
minModuleSizeand merge threshold-e deliberately; both change the answer. - Always test module-trait association against a permutation null, not against p-values alone.
The honest limitation is that WGCNA recovers co-expression, and co-expression is not causation. Two genes correlate because one regulates the other, because both are regulated by a third, or because of a shared batch effect. Module membership is a lead, not a mechanism.
ARACNe: inferring regulatory relationships
ARACNe (Margolin et al. 2006) attacks the same problem differently. Starting from mutual information between expression profiles, it applies the data processing inequality to prune indirect edges: if gene A connects to C only through B, the A–C edge is removed unless it is directly supported. The result is a directed-ish regulatory network with transcription factors as hubs. It is a genuine inference method, not just clustering, but it is acutely sensitive to sample size and to the choice of candidate regulators. Below a few hundred samples, ARACNe networks are mostly noise with a few strong hubs.
From co-expression to causal edges: the harder tools
When you want direction rather than association, you need perturbation, genetic variation, or time. Genomic loci that control gene expression (eQTLs) give directionality from genotype to transcript and are the cleanest available instrument. Mendelian randomisation and mediation analysis can then chain an eQTL to a phenotype. Bayesian network learning and methods exploiting genetic instruments scale this further, at the cost of strong assumptions that reviewers should demand you state explicitly. Correlation is cheap; direction is expensive, and there is no statistical shortcut around that.
Epigenetic interaction networks
Epigenetic layers — DNA methylation, histone marks, chromatin accessibility — form their own networks, and the interesting nodes are the cross-layer edges. Practical constructions include:
| Layer pair | Typical signal |
|---|---|
| Methylation ↔ expression | Negative correlation at promoters, meQTL links |
| ATAC ↔ expression | Positive correlation at enhancers and promoters |
| Histone marks ↔ TF binding | H3K27ac with active regulators |
| Chromatin loops ↔ regulation | Hi-C anchors joining enhancers to targets |
The failure mode is cell-type composition. Bulk methylation and bulk accessibility are dominated by the mixture of cells in the sample, so a "network" can be entirely an artefact of differing cell proportions across individuals. Correct for composition or work in sorted or single-cell data before believing any edge.
Pathway topology analysis
Over-representation analysis counts how many of your genes fall in a pathway — it ignores structure entirely and inherits every annotation bias in the database. Topology-based methods (SPIA, Pathway-Express, and related approaches) weight genes by their position in the pathway graph: degree, betweenness, or the perturbation of downstream nodes. The gain is real: a hit on an upstream regulator propagates, and topology-aware scores can capture that where a simple count cannot. The cost is that results now depend on the pathway graph's accuracy, which is uneven.
Whichever method you use, correct for multiple testing across pathways, use the appropriate background gene universe, and never report a pathway p-value from a database whose annotations postdate your data without saying so.
What I actually trust
Trust module-trait associations with a permutation null and a replication cohort. Treat ARACNe edges as candidates for experimental validation, never as established regulation. Treat epigenetic networks as composition-sensitive until proven otherwise. And treat pathway enrichment as a labelling convenience, not evidence. Networks are a compression of your data, and compression is not confirmation.
References
- Langfelder P, Horvath S. WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics. 2008;9:559.
- Zhang B, Horvath S. A general framework for weighted gene co-expression network analysis. Stat Appl Genet Mol Biol. 2005;4:Article17.
- Margolin AA, Nemenman I, Basso K, et al. ARACNE: an algorithm for the reconstruction of gene regulatory networks in a mammalian cellular context. BMC Bioinformatics. 2006;7(Suppl 1):S7.
- Basso K, Margolin AA, Stolovitzky G, et al. Reverse engineering of regulatory networks in human B cells. Nat Genet. 2005;37(4):382–390.
- Khatri P, Sirota M, Butte AJ. Ten years of pathway analysis: current approaches and outstanding challenges. PLoS Comput Biol. 2012;8(2):e1002375.
- Tarca AL, Draghici S, Khatri P, et al. A novel signaling pathway impact analysis. Bioinformatics. 2009;25(1):75–82.
- GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science. 2020;369(6509):1318–1330.
- Buenrostro JD, Giresi PG, Zaba LC, et al. Transposition of native chromatin for fast and sensitive epigenomic profiling of open chromatin, DNA-binding proteins and nucleosome position. Nat Methods. 2013;10(12):1213–1218.
- Fraser J, Ferrai C, Chiariello AM, et al. Hierarchical folding and reorganization of chromosomes are linked to transcriptional changes in cellular differentiation. Mol Syst Biol. 2015;11(12):852.