Do Co-expression Modules Carry Shared Genetic Risk Between Diseases?

When two diseases tend to occur together, part of the reason is often genetic. Some DNA variants raise risk for both. Genome-wide summaries of this sharing can hide it: if one stretch of the genome pushes two diseases in the same direction and another pushes them in opposite directions, the average comes out near zero. Methods that scan the genome locally recover much of this hidden sharing. What they return, though, is a list of DNA segments, with no indication of which cellular programs those segments act through. Gene co-expression networks supply that missing piece. They group genes whose activity rises and falls together across tissue samples, giving data-driven "modules" that often correspond to real cellular machinery.

We propose a calibrated test of one question: does a co-expression module, built only from expression data, concentrate local cross-disease genetic covariance in and around its member genes beyond what its genomic size and layout would predict? The module is fixed before any genetic data is seen, so the test is a genuine test and not a tautology. The test weights each gene by how central it is to its module. Its null distribution accounts for the fact that co-expressed genes often sit next to each other on the chromosome. The concrete claim is that, in the right tissue, these modules explain more replicable shared genetic risk than curated pathway gene sets of the same size and footprint, and more than gene clusters derived from genetics alone.

Background and Gap

Three bodies of work frame the problem, and each plays a different role here.

The first motivates the question. Local genetic correlation methods show that much sharing is invisible globally. LOGODetect scans pairs of genome-wide association study (GWAS) results with a variable-width window that accounts for correlation between nearby variants. Across seven neuropsychiatric traits it found 227 segments with significant local genetic correlation, and 17 of 24 discovery hub regions replicated in UK Biobank. iCPAGdb links diseases through overlapping risk variants and reports that 61% of the significant trait overlaps it detected were missed by genome-wide LD score regression, the standard global method. Both stop at genomic segments or variant overlaps. Neither asks whether the genes in those segments act together in a tissue.

The second supplies the module side. Weighted gene co-expression network analysis (WGCNA) builds a gene-by-gene similarity from expression correlations raised to a power, clusters genes into modules, and summarizes each module by its first principal component, the "eigengene." The RNF181 coronary artery disease study is the closest precedent for using it with genetics. It ran WGCNA on heart tissue, ran Gene Ontology (GO) and KEGG pathway enrichment on the resulting modules, and then looked up a CAD risk variant near a hub gene. That lookup is single-disease and post hoc, with no statistical test of whether a module as a whole carries genetic signal.

Network-based drug repurposing for psychiatric disorders reports that schizophrenia, bipolar disorder and autism risk genes converge on cell-type-specific co-regulated modules in PsychENCODE single-nucleus data. Its convergence evidence rests on curated risk-gene lists and a graph neural network that ranks genes. It does not run a calibrated test of summary-statistic genetic covariance per module.

The third constrains the design. KISL improves on WGCNA by using distance correlation and semi-supervised clustering, with "must-link" constraints derived from GO enrichment. It then validates its modules partly with GO enrichment, which risks circularity. We take that as a warning: genetic information must not shape the modules if genetic enrichment is the outcome.

Standard overrepresentation analysis treats genes as independent draws. That assumption is badly wrong when module genes cluster physically, so a naive hypergeometric test of "module genes versus genes in shared segments" will be inflated.

BayesRR-RC partitions heritability across genomic annotations. It attributes at most 10% of variation to regions within 10kb upstream of genes and 22–28% to regions 10–500kb upstream. A gene-centric test that counts only gene bodies would therefore miss most of the signal. Our gene windows must reach distal regulatory sequence.

Core Claim

For a module $M$ and two traits with GWAS z-scores $z^{(1)}_s, z^{(2)}_s$ at variant $s$, the test statistic is

$$T_M = \sum_{g \in M} w_g \sum_{s \in W_g} \frac{z^{(1)}_s z^{(2)}_s - \hat\rho}{\ell_s},$$

The terms are defined as follows:

  • $W_g$ is the window from 500kb upstream of gene $g$ to the end of its gene body. Variants falling in several windows are counted once, assigned to the nearest gene.
  • $w_g$ is the gene's module membership, meaning its correlation with the module eigengene.
  • $\hat\rho$ is the expected product under sample overlap. It is estimated from null variants, as in LOGODetect and PathGPS.
  • $\ell_s$ is the variant's LD score, a measure of how many neighbors it is correlated with.

We report both the signed $T_M$, which captures concordant sharing, and a version using absolute products within LD blocks, which captures discordant local sharing that cancels globally.

Significance comes from random gene sets matched to $M$ on four properties: gene count, total window length, LD-score distribution, and the number of contiguous physical gene blocks. Matching on blocks is what prevents adjacency from masquerading as biology.

The claim is falsifiable at three points. Tissue-relevant modules should beat matched random sets. They should beat GO/KEGG sets of the same size and footprint. Their significant module–trait-pair findings should replicate in an independent GWAS at a higher rate than the baselines' findings.

Research Program

We build modules in two settings, each with a tissue chosen in advance.

Psychiatric setting. Modules are built per cell type from pseudo-bulk profiles of the PsychENCODE Phase II single-nucleus data (2.8 million nuclei, 388 donors, 23 cell classes), using both WGCNA and KISL. The genetic side uses summary statistics for the seven LOGODetect neuropsychiatric traits, giving 21 trait pairs. UK Biobank serves as the replication cohort, matching LOGODetect's own replication design.

Cardiometabolic setting. Modules come from the left-ventricle microarray series used in the RNF181 study (GSE12504, GSE22253, GSE77263, GSE49925). The genetic side uses UK Biobank traits from the disease clusters PathGPS identified: cardiovascular disease, type 2 diabetes and venous thromboembolism.

Every partition is computed without genetic input and frozen before the test is run.

Calibration is checked by simulation under realistic LD. We follow LOGODetect's approach of simulating 100,000 genotypes from 1000 Genomes haplotypes with HAPGEN2. Shared effects are planted either nowhere, in random genes, in physically clustered random genes, or in co-expressed genes. This tests two things: whether the block-matched null holds its nominal false-positive rate where gene-permutation nulls do not, and how much power the eigengene weighting adds.

For modules that pass, functional enrichment is run with a module-aware null. GO terms are tested against random sets drawn from the same expression module structure. The interpretation then says which part of a module carries the shared genetic signal, instead of re-reporting generic pathways.

Baselines

Naive reference. The naive point is the pipeline most papers would run. We take the genes inside LOGODetect's significant segments and test each module for overlap with them using a hypergeometric test, then run ORA. It uses the same modules and GWAS data. We expect it to be miscalibrated in the clustered-null simulations. Beating it means equal or greater power at a correct false-positive rate.

Curated gene sets. The strongest competitor on the scientific question is curated knowledge. We apply the identical $T_M$ statistic and matched null to GO Biological Process and KEGG sets, subsampled or merged to match each module's size and genomic footprint. If curated sets capture shared genetic risk as well as tissue modules do, expression adds nothing. The comparison uses the same windows, LD scores and permutation budget. Beating it means a higher replication rate of significant findings in UK Biobank, since replication is the only ground truth available here.

Genetics-only clusters. PathGPS builds gene–trait clusters from summary statistics alone, after subtracting environmental covariance. We run it on the discovery GWAS and score its clusters with $T_M$ on held-out trait pairs. This avoids the trivial win it would have on the pairs it was fit to.

LOGODetect itself. It is the reference for replication. No module-level replication figure exists to compare against. The only reported number is 17 of 24 hub regions replicating in UK Biobank, which serves as a calibration point for what segment-level replication looks like with these traits. We found no reported result on Lacuna for any module-level genetic covariance test on these data.

Ablations. Four ablations isolate the key ingredients. Setting $w_g = 1$ isolates the eigengene weighting. Swapping the block-matched null for plain gene permutation isolates the adjacency correction. Comparing 10kb and 500kb windows tests the distal regulatory claim from BayesRR-RC. Substituting heart modules for brain modules on psychiatric traits, and the reverse, gives a tissue-mismatch negative control.

Why not just…?

Why not just run enrichment on the genes in LOGODetect's shared segments? This discards the main thing expression adds: which genes, among many in a segment, actually work together. It also inherits ORA's independence assumption, which physically clustered genes break. The clustered-null simulation will show directly whether this naive test is inflated. If it turns out calibrated and as replicable as $T_M$, the simpler method wins and we will say so.

Why not inject the shared segments as must-link constraints, KISL-style, so modules form around genetic risk? This is the idea the research file points toward. It has two problems. The genetic signal would then shape the modules it is used to test, repeating the circularity seen in KISL. And the constraints would mostly tie together genes that are neighbors on the chromosome. We include a must-linked variant only as an exploratory comparison, scored on held-out trait pairs whose segments were not used as constraints.

Why not skip expression and use PathGPS clusters directly? This is a real threat. PathGPS separates genetic from environmental covariance, which expression modules cannot do, since tissue co-expression mixes both. The held-out replication comparison settles it. If PathGPS clusters replicate as often as tissue modules, expression-based modules are not earning their cost.

Why not use a global genetic correlation per module? Global summaries cancel opposing local effects. iCPAGdb's finding that 61% of overlaps were missed by LD score regression suggests a signed-only statistic will miss discordant sharing. Our absolute-product variant tests exactly this.

Decisive Checks

The work succeeds if four things hold:

  • The block-matched null holds nominal error in simulation where gene-permutation nulls do not.
  • Cell-type modules from the matched tissue beat footprint-matched GO/KEGG sets on UK Biobank replication rate.
  • Hub weighting adds power.
  • The tissue-swap control erases the excess signal.

It fails informatively under either of two outcomes. If curated sets or PathGPS clusters match the modules, data-driven co-expression adds nothing to shared genetic architecture beyond what curated knowledge or genetics alone already provides. If brain and heart modules perform equally on psychiatric traits, the signal reflects generic gene density and not tissue biology.