A single genetic variant that appears to control hundreds of genes, called a pleiotropic regulatory hotspot, is usually read as evidence of a master switch: a transcription factor, an enhancer hub, or a chromatin loop that coordinates many targets. Most of these hotspots in human data were found in bulk tissue, where the RNA of millions of cells of different types is pooled and averaged. In pooled tissue there is a duller way to get the same pattern. If a variant changes how many neurons a brain region contains, every neuron-specific gene rises or falls together, and the locus looks like a master regulator even though it may not regulate any of those genes inside any cell. We propose to tell these two cases apart for every bulk hotspot. The test rests on a geometric fact: a variant that acts only through cell composition must move its target genes along the direction set by the cell types' reference expression profiles, scaled by how much RNA each cell type makes. Whatever part of the effect falls outside that direction must come from regulation inside cells.
This matters for missing heritability, the gap between how heritable a disease is in family studies and how much of that heritability genome-wide association studies (GWAS) can assign to specific variants. Recent work reports that regulatory network hubs, built from bulk tissue, carry concentrated disease heritability. If many of those hubs are disguised cell-composition effects, the mechanism behind the recovered heritability is different from the one claimed. The heritability itself may also be partly mislocated: true within-cell regulators in rare cell types get diluted below detection, while composition loci in abundant, RNA-rich cell types get inflated. We will decompose bulk trans-hotspots into a composition part and an intracellular part, then measure how disease heritability splits between them.
Core Claim
The claim is testable. In post-mortem brain bulk RNA-seq, a substantial fraction of trans-hotspots, especially "global hubs" with very many targets, will have effect patterns that lie almost entirely in the low-dimensional span of activity-scaled cell-type signatures. These composition-explained hotspots will carry heritability enrichment for traits whose genetics act through cell abundance, such as Alzheimer's disease through microglia. For other traits they will carry less enrichment than the intracellular remainder. The claim is wrong if, after decomposition, bulk hotspots keep their target sets and their heritability enrichment almost unchanged.
The idea starts from a tension in the literature. The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution built bipartite variant–gene networks from bulk GTEx v8 in 29 tissues. It grouped them into modules with a community-detection method (CONDOR), then used stratified LD score regression (S-LDSC, which estimates how much heritability sits in a chosen set of variants after accounting for correlation between nearby variants) to show that hubs central to tissue-specific modules carry the most heritability. That paper motivates this work, but it never asks whether its hubs are cellular. Single-cell deconvolution of 3,000 post-mortem brain samples for eQTL and GWAS dissection in mental disorders supplies both the warning and the tools. Its SPLITR method estimates cell fractions from bulk brain RNA while correcting for the roughly fourfold higher mRNA output of neurons relative to glia. It shows that TMEM106B variants reduce excitatory neuron proportion. Its deQTL model finds 4,757 cell-type-specific eQTLs and 1,182 eGenes invisible in bulk. That work stays in cis, meaning variants acting on nearby genes, and never examines hotspots or partitions heritability. The sparse hotspot models Sparse regression models for unraveling group and individual associations in eQTL mapping and Graph-regularized dual Lasso for robust eQTL mapping separate group-level from single-gene associations. They absorb confounding with low-rank terms, so they would either call composition loci hotspots or remove them blindly, depending on whether the low-rank factor captures them. Neither model represents cell composition explicitly.
The Decomposition
For a candidate hotspot variant $s$, let $\beta_s$ be its vector of trans effects on $G$ bulk genes, estimated in bulk data by standard marginal regression. Let $S$ be the $G \times K$ matrix of mean expression per cell for $K$ cell types, taken from a single-nucleus reference. Let $A$ be the diagonal $K \times K$ matrix of per-cell RNA output, estimated as in SPLITR. Under a pure composition mechanism, bulk expression is a mixture, so to first order
$$\beta_s = S A \,\delta_s + \gamma_s,$$
where $\delta_s$ is the variant's effect on the $K$ cell fractions and $\gamma_s$ is the leftover intracellular effect. We estimate $\delta_s$ by generalized least squares and constrain it to sum to zero, because fractions trade off against each other. Two statistics follow. The first is the composition share, $\lVert SA\hat\delta_s\rVert^2/\lVert\hat\beta_s\rVert^2$. The second is a test of $\gamma_s=0$, calibrated by permuting genotypes across individuals so that the correlation among genes is preserved.
This decomposition does not depend on estimating per-sample cell fractions well. It uses the pattern of effects across hundreds of target genes, and that pattern is informative even for rare cell types whose fractions deconvolution methods estimate poorly. The intracellular remainders $\gamma_s$ are then re-tested within cell types using deQTL-style genotype × fraction interactions. This step asks whether intracellular hotspots are shared across cell types or confined to one.
Research Program
The main cohort is the 3,387 bulk brain samples from 1,127 individuals in ROSMAP, Mayo Clinic Brain Bank, and GTEx brain regions that SPLITR analyzed. Cell fractions and activity factors come from SPLITR's reference. Candidate hotspots are found by genome-wide trans mapping with local effects removed first, following the stepwise cis-then-residual strategy of Identifying cis- and trans-acting single-nucleotide polymorphisms controlling lymphocyte gene expression in humans. Trans power at about a thousand donors is limited, so we restrict to hotspots with many targets at a false discovery rate below 0.05 and report power explicitly.
Each hotspot is labeled by its composition share. Its variants and target genes then form three sets: composition-dominated, intracellular-shared, and intracellular-cell-restricted. Heritability is partitioned in two ways. On the variant side, the three sets enter S-LDSC as annotations alongside the bulk CONDOR hub annotations of the network-structure paper. On the gene side, Partitioning gene-based variance of complex traits by gene score regression supplies the method. Gene score regression (GSR) regresses transcriptome-wide association statistics on within-set squared gene correlations and so removes inflation from correlated nearby variants. We run it with bulk weights and with cell-type-specific weights for the same gene sets. Traits are Alzheimer's disease, schizophrenia, and the other brain-relevant traits among the ten in the network-structure paper. Height is included as a negative control with no expected brain-composition route.
Baselines
The naive reference is bulk hotspot calling with no cell-type modeling. Each variant is scored by the number of trans targets it reaches, following the global-hub definition in the network-structure paper, and partitioned with S-LDSC exactly as that paper did. All of our heritability contrasts are measured against this baseline using the same GWAS summary statistics and LD reference. That paper reports module-level findings, such as 29% of modules being trait-specific and breast-cancer variants concentrating in 12.6% of skin modules, but the file gives no single enrichment number for brain hubs that we could match. We therefore rerun it ourselves rather than quote it.
The second baseline is low-rank confounder-corrected hotspot mapping in the style of geQTL. It is the strongest hotspot-specific method in the file and might absorb composition effects through its latent factors. On yeast simulations, the related graph-regularized dual Lasso reported a test error of 0.9122, against 0.9276 for graph Lasso and 1.2153 for plain Lasso. The same work reported about 24 hours on 1,017 variants. We therefore run geQTL-ridge, the faster closed-form ridge variant, with the same candidate-variant set and the same number of latent factors that our method uses for technical covariates. It counts as beaten if its hotspot calls on our simulations have lower precision at matching recall for intracellular hotspots.
The strongest competitor on the question we actually pose is proportion-adjusted trans mapping: estimate SPLITR fractions per sample, include them as covariates, and call whatever survives intracellular. No reported result found on Lacuna for this approach applied to trans-hotspot classification. We give it the same reference, the same activity factors, and the same permutation calibration.
Two ablations isolate the key ingredient. One drops the activity matrix $A$, so the projection treats all cell types as making equal RNA. The other uses an unstructured top-$K$ principal-component subspace in place of $SA$. Our method beats these if the signature projection with activity scaling classifies simulated and pseudobulk hotspots more accurately, particularly for cell types under 5% abundance.
Why not just…?
Why not just adjust for deconvolved cell fractions as covariates? This is the real threat. If fractions were estimated perfectly, covariate adjustment would remove the composition route, much as our decomposition does. The file gives reasons to doubt that fractions are good enough. The SPLITR paper reports that CIBERSORT missed inhibitory neurons and oligodendrocyte progenitors entirely, and fraction errors leave residual composition signal that would still be called intracellular. Adjustment also discards the composition effect instead of measuring it, so it cannot answer the heritability question about which route carries risk. The pseudobulk experiment below settles the matter. If covariate adjustment matches the projection in classification accuracy, we adopt it as the classifier and the contribution narrows to the heritability decomposition.
Why not just run a GWAS on cell fractions and drop those loci? Fraction GWAS found TMEM106B, but it inherits the same fraction-estimation noise and treats a locus as all-or-nothing. A variant can shift microglial abundance and also regulate genes inside microglia. The composition share is continuous, so it represents that mixture directly. We also use fraction-GWAS hits as an external check: loci it finds should receive high composition shares from our method.
Why not just use single-cell eQTL cohorts directly? Where they exist, they are the gold standard, and we use one as ground truth. Population single-cell cohorts are small relative to bulk and rarely cover post-mortem brain at bulk sample sizes, so trans-hotspot power stays in bulk. The question is how to read bulk data correctly, not how to replace it.
Proposed Experiments
The first decisive check uses pseudobulk with known truth. Onek1k is a population single-cell cohort of about $10^6$ blood mononuclear cells used by Count Bridges enable Modeling and Deconvolving Transcriptomic Data. Provided its donor genotypes are accessible, we sum each donor's cells into a pseudobulk profile, call hotspots on the pseudobulk, and decompose them. The truth comes from the same donors' per-cell-type expression: whether each hotspot shifts cell counts or changes expression within cell types. The metrics are classification AUPRC and calibration of the $\gamma_s=0$ test. The same design can be run on brain using SPLITR's single-nucleus reference with simulated genotype effects of both kinds, varying the abundance of the affected cell type from 2% to 40%.
The second check uses positive and negative controls in brain bulk. TMEM106B should receive a high composition share in the direction of excitatory-neuron loss, and the immunostaining-validated astrocyte fraction near 18% provides a sanity check on $A$ and $S$. Hotspots driven by known transcription factors should have low composition shares.
The third check is the heritability test itself. We compare S-LDSC enrichment and GSR variance shares for composition-dominated versus intracellular hotspot sets against the bulk-hub baseline. Our hypothesis predicts specific patterns. Alzheimer's heritability should concentrate in the microglia-composition component, consistent with the reported association between Alzheimer's polygenic score and microglia fraction. Schizophrenia heritability should be shared between inhibitory-neuron composition and intracellular neuronal hotspots. Height should show neither. For context, GSR reported an AUPRC of 0.925 against 0.260 for PASCAL in its own simulations. We use GSR as the gene-level tool rather than as a competitor. If bulk local hubs keep their enrichment after the composition component is removed, the network-structure paper's interpretation stands and our central claim fails.
References
- The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution
- Single-cell deconvolution of 3,000 post-mortem brain samples for eQTL and GWAS dissection in mental disorders
- Sparse regression models for unraveling group and individual associations in eQTL mapping
- Graph-regularized dual Lasso for robust eQTL mapping
- Identifying cis- and trans-acting single-nucleotide polymorphisms controlling lymphocyte gene expression in humans
- Partitioning gene-based variance of complex traits by gene score regression
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data