Does Co-Expression Carry Heritability Beyond Cell Composition?

Researchers who measure gene activity in tissue (RNA sequencing) often group genes that rise and fall together across samples into "modules," using weighted gene co-expression network analysis (WGCNA). They then ask whether a module matters for a disease. Today that question is usually answered in one of two weak ways. The first is to look up one top gene in a genetic association catalog. The second is to run an overlap test against pathway databases that treats every gene as an independent draw. A rigorous alternative exists: LD score regression (LDSC), a method that measures how much of a trait's inherited variation sits in a chosen part of the genome using only published association summary statistics. The obvious move is to feed module genes into it.

We think that obvious move is subtly wrong for bulk tissue, and that fixing it is the real contribution. A module built from bulk tissue often tracks how many cells of one type a sample contains, not a regulatory program running inside cells. If a disease's genetic risk concentrates in genes specific to one cell type, every module that proxies for that cell type will look "heritability-enriched" without its co-expression structure adding anything. This proposal builds and tests composition-residualized, connectivity-weighted module annotations for stratified LDSC. The goal is to answer, trait by trait, a question current pipelines cannot separate: does the network structure itself carry genetic signal once cell-type specificity is accounted for?

Core Claim

Co-expression modules should be tested for heritability enrichment in three ways at once. Each module should be built after removing the expression variance explained by estimated cell-type proportions. Each gene should be weighted continuously by how tightly it tracks the module's summary profile, rather than given a yes/no membership label. The regression should condition on cell-type-specificity and expression-level annotations. Our prediction is that a large share of the enrichment reported by the naive "module genes as a binary set" analysis disappears under this conditioning. We also predict that what survives is more stable across resampling and more reproducible across cohorts. If nothing survives, that is itself a clear finding: for that trait and tissue, bulk co-expression modules are cell-type proxies and add no genetic information.

Why the existing pieces do not already answer this

The work in the file plays three different roles.

Two papers motivate the problem. Weighted Co-Expression Network Analysis Identifies RNF181 as a Causal Gene of Coronary Artery Disease clusters 8,525 heart-tissue genes into 7 modules, runs standard pathway overlap tests, and links its hub gene to coronary artery disease through a single variant lookup (rs6738645). That is the typical workflow. It never asks whether the module as a whole carries polygenic heritability, and its authors themselves flag bulk-tissue composition as a confounder. Embracing polygenicity: a review of methods and tools for psychiatric genetics research explains why single-locus lookups are the wrong unit for traits driven by thousands of small effects.

One paper supplies the closest method and also constrains our design. Partitioning gene-based variance of complex traits by gene score regression adapts the LDSC idea to gene sets and shows that ignoring gene–gene correlation badly inflates false positives. Its simulations report precision-recall area 0.925 against 0.260 for PASCAL. It tests curated pathways through genetically predicted expression, not data-driven modules, and it has no mechanism to separate cell composition from regulation. Two further papers constrain the estimator choice. Comparison of methods for estimating genetic correlation between complex traits using GWAS summary statistics shows LDSC stays robust under reference-panel mismatch and sample overlap, where the more precise high-definition likelihood method becomes brittle. Flexible Models of Functional Annotations to Variant Effects using Accelerated Linear Algebra warns that flexible models fit through LDSC shortcuts overfit. That paper is why we keep annotations few, pre-specified and linear, and do not learn them from summary statistics.

SUPERGNOVA: local genetic correlation analysis reveals heterogeneous etiologic sharing of complex traits shows that genome-wide averages can hide opposing signals. It partitions by physical LD blocks rather than by gene networks, and we use its lesson only in a secondary signed analysis.

Research Program

We build modules from GTEx tissues matched to traits with known positive controls from the gene score regression paper: liver for HDL cholesterol, brain for schizophrenia, and whole blood for lupus. We add the five public heart datasets used for coronary artery disease, merged after batch harmonization with pyComBat from epigenelabs/inmoose.

Each tissue yields two networks. The raw network is standard WGCNA. The residual network is built after regressing each gene's expression on marker-based estimates of cell-type proportions.

Each module then becomes a continuous gene weight. We use the gene's correlation with the module's first principal component, truncated at zero. That weight is spread to nearby SNPs through a fixed window to form a SNP annotation $a_c(k)$ for module $c$ and SNP $k$. The stratified regression is

$$\mathbb{E}[\chi^2_j] = N \sum_c \tau_c\, \ell(j,c) + N a + 1, \qquad \ell(j,c)=\sum_k a_c(k)\, r^2_{jk}.$$

Here $\chi^2_j$ is SNP $j$'s association statistic, $N$ the study sample size, $r^2_{jk}$ the squared correlation between SNPs from the 1000 Genomes European panel, $\tau_c$ the per-unit heritability contribution of annotation $c$, and $Na+1$ the confounding intercept. The quantity of interest is $\tau_c$ for a residual module, estimated jointly with the following conditioning annotations:

  • a standard baseline of genomic features;
  • per-cell-type specificity annotations built from the same tissue;
  • mean expression level;
  • a module-size-matched random-gene annotation.

Stability is measured by bootstrap-resampling tissue samples, rebuilding modules, and reporting the spread of $\tau_c$ for matched modules. This targets the known fragility of irreversible hierarchical clustering.

Baselines

Naive reference point. Standard overlap testing (hypergeometric) of module genes against genes near genome-wide significant loci, plus the single-hub lookup, exactly as in the RNF181 study. This is what most papers do, and it defines the false-positive level we must beat.

Obvious strong competitor. The same stratified LDSC run on binary membership of raw WGCNA modules, with only the standard baseline annotations. This is what a careful analyst would do tomorrow, so it is the comparison that matters most. It uses identical GWAS summary statistics, LD panel, SNP windows and module count; the only differences are our three changes.

Strongest published gene-set method. Gene score regression run on the same module gene lists, using the released code and GTEx weights. Its reported simulation result is precision-recall area 0.925 Partitioning gene-based variance of complex traits by gene score regression. No reported result was found on Lacuna for any method applied to data-driven co-expression modules.

Ablations. These isolate the key ingredient by switching one component at a time: residualization on or off, continuous versus binary weights, and cell-type conditioning on or off.

What counts as beating the baselines. In simulation, our method must reach a precision-recall area at least matching 0.925 when heritability is planted in residual modules, and must hold nominal false-positive rates when heritability is planted only in cell-type-specific genes. The binary-module baseline should fail that second condition. On real data, surviving modules must replicate across independent cohorts of the same tissue more often than baseline-significant modules do.

Why not just…?

Why not just run stratified LDSC on binary WGCNA modules? This is the real competitor. Our argument that it confounds cell composition with regulation is testable, not assumed. The composition-only simulation will show whether binary modules produce false enrichment. If they do not, most of this proposal's added machinery is unnecessary, and we will report that.

Why not just use cell-type-specific expression annotations and drop networks entirely? This is a genuine threat. It may be the correct answer for many traits. The conditional $\tau_c$ of residual modules, given those annotations, is exactly the test. A null result across all four traits would argue that bulk co-expression modules add nothing genetically, which is still a useful correction to current practice.

Why not just apply gene score regression to the modules? It asks a different question, about heritability mediated through genetically predicted expression. It also depends on the tissue match of the eQTL panel and cannot remove composition. We run it side by side. Agreement between the two methods strengthens a finding; disagreement localizes where eQTL mediation matters.

Why not fit a flexible full-likelihood model instead? The DeepWAS paper shows this requires individual-level genotypes and GPU clusters, and it warns that flexible models fit through summary shortcuts overfit. Our annotations are few and fixed before the genetic data is seen, which avoids that failure.

Proposed Experiments

Simulations. Using 1000 Genomes genotypes, we simulate traits under three regimes:

  • heritability in cell-type-specific genes only;
  • heritability in residual-module genes only;
  • no gene-set structure.

For each method we score precision-recall area and false-positive rate.

Real data. We analyze HDL, schizophrenia, lupus and coronary artery disease in their matched tissues, reporting enrichment, conditional $\tau_c$ and bootstrap stability. Trait–tissue mismatches serve as negative controls; for example, HDL tested with brain modules should show no enrichment.

Decisive check. For each positive-control trait, compare the fraction of binary-module enrichments that survive residualization plus cell-type conditioning, and test whether survivors replicate in a second cohort. A secondary signed analysis asks whether modules split into opposing-direction submodules, in the spirit of SUPERGNOVA's local findings.