Splitting Expression-Mediated Heritability into Cell Composition and Cell-Intrinsic Regulation

Most genetic risk for common disease sits in non-coding DNA, the switches that turn genes up or down. The usual way to link these switches to disease is to measure gene expression in bulk tissue, a ground-up mixture of many cell types, and ask which genes' genetically predicted expression tracks disease. A bulk measurement, however, changes for two different reasons. A variant can change how strongly a gene is expressed inside one cell type. Or it can change how many cells of that type the tissue contains, which moves every gene that type expresses at once. Current methods treat both as "the variant regulates the gene." That mistake has two costs. Real regulation in rare cell types is averaged away, and genes whose only link to disease is that they mark a cell type whose abundance is genetically controlled get credited as causal.

We propose to split the part of disease heritability that runs through expression into two channels, one through composition and one through cell-intrinsic regulation, and to carry the uncertainty of cell-type unmixing through to the final heritability estimate. The split is identifiable because a variant acting through composition has to shift all of a cell type's genes in fixed proportion to that type's expression signature. A variant acting inside cells is free to affect genes one at a time. We will fit the model on large bulk repositories and test it on single-cell pseudobulk data where the answer is known. The output is an estimate, per trait and per cell type, of how much "expression-mediated" heritability is really composition, and how much cell-intrinsic heritability bulk analysis has been missing.

Core Claim

Write the bulk expression of gene $g$ in person $i$ as $y_{ig} = \sum_k \pi_{ik} m_k x_{ikg}$. Here $\pi_{ik}$ is the fraction of cell type $k$, $m_k$ is that type's total mRNA content per cell, and $x_{ikg}$ is the gene's expression within the type. If genotype $G_i$ affects both fractions (effect $\gamma_k$) and within-type expression (effect $\beta_{kg}$), then to first order around population means (the bars) the bulk genetic effect is

$$\frac{\partial y_{ig}}{\partial G_i} \approx \underbrace{\sum_k \bar\pi_k m_k \beta_{kg}}_{\text{cell-intrinsic}} + \underbrace{\sum_k m_k \bar x_{kg}\,\gamma_k}_{\text{composition}}.$$

The composition term is low-rank across genes. Each cell type contributes a single scalar $\gamma_k$, multiplied by a signature vector $\bar x_{k\cdot}$ that is known from a reference. That constraint lets the two channels be separated from bulk data, provided the signatures are not collinear. Pushing both channels into a transcriptome-wide association study (TWAS, which tests whether genetically predicted expression associates with a trait) gives a two-part partition of trait heritability. We claim this partition changes which genes are called causal and how much heritability is attributed to minor cell types. Both claims can be checked against single-cell ground truth.

Relation to Prior Work

The closest work is Single-cell deconvolution of 3,000 post-mortem brain samples for eQTL and GWAS dissection in mental disorders. It unmixes 3,387 bulk brain samples while correcting for the roughly fourfold higher mRNA content of neurons relative to glia. It runs a GWAS on cell fractions, finding that TMEM106B variants reduce excitatory neuron fraction, and it maps cell-type expression QTLs (variants associated with a gene's expression, eQTLs). It never builds TWAS predictors or asks how much trait heritability flows through each channel. It is the direct motivation, and we reuse its $m_k$ correction and its immunohistochemistry validation. Cell-type specific inference from bulk RNA-sequencing data by integrating single cell reference profiles via EPIC-unmix recovers masked cell-type eQTLs, such as a microglial AP3B2 eQTL, from ROSMAP and MSBB using an empirical Bayes update. It supplies one of our unmixing back-ends, but it works on point estimates and does not model composition QTLs. Deep autoencoder for interpretable tissue-adaptive deconvolution and cell-type-specific gene analysis (TAPE) is an alternative back-end whose decoder adapts signatures to disease state. That matters because our identifiability argument assumes signatures are right.

On the heritability side, scPrediXcan trains cell-type expression predictors on real single-cell cohorts such as OneK1K (982 donors). It roughly doubles the causal genes found compared with bulk TWAS, but it needs a population-scale single-cell cohort that does not exist for most tissues. It also has no notion of composition. SUMMIT shows that genes with low expression heritability, down to prediction $R^2 = 0.005$, still carry reproducible trait signal once reference sample size grows. That constrains our design: unmixing reduces the effective sample size per cell type, so our predictors must borrow SUMMIT-style multi-penalty ensembling rather than discard low-heritability genes. Gene score regression supplies the method for partitioning TWAS signal while correcting for correlation between neighboring genes. We extend its gene scores to two channel-specific scores. It assumes predictor weights are error-free, which is exactly the assumption we drop.

Statistical inference for cell type deconvolution (MEAD) shows that treating estimated fractions as fixed inflates false positives, and it provides sandwich variances for them. It is the method we use for uncertainty propagation. Its limitation, that it cannot handle missing cell types, is a stated risk for us too. The developing-brain atlas in Cross-ancestry, cell-type-informed atlas of gene, isoform, and splicing regulation in the developing human brain assigns eQTLs to cell contexts by post-hoc integration rather than per-sample unmixing. Its finding that cell-informed, stage-specific regulation roughly doubles the share of prioritized psychiatric GWAS loci motivates looking for masked cell-intrinsic signal. It does not contradict our composition argument, because it never estimated composition effects.

Research Program

The pipeline has four stages.

The first stage unmixes bulk samples. It uses SPLITR-style mRNA-content and platform corrections, with EPIC-unmix or TAPE as interchangeable back-ends. MEAD sandwich variances are attached to the fractions, and a delta-method or bootstrap variance is attached to the per-sample cell-type expression.

The second stage fits a joint genetic model per locus. A composition QTL for each cell type is estimated from fractions, and cell-intrinsic effects are estimated from the unmixed expression. A rank constraint forces each variant's composition effect to act through the reference signatures. Estimation is errors-in-variables regression weighted by the stage-one variances.

The third stage trains two families of TWAS predictors. Cell-intrinsic predictors are built per gene and cell type using SUMMIT's five-penalty ensemble. Composition predictors are built per cell type, since the composition channel has only as many "genes" as there are cell types.

The fourth stage runs association with GWAS summary statistics. It partitions heritability with channel-specific gene scores in the style of gene score regression, and it inflates standard errors by the propagated predictor variance.

Beyond the method, the program answers three questions. First, what fraction of bulk TWAS hits are explained by the composition channel? Second, how much cell-intrinsic heritability in minor types (microglia in brain, rare immune types in blood) appears only after unmixing? Third, how much do the mRNA-content correction and the uncertainty propagation each shift these numbers?

Baselines

The naive reference is standard bulk TWAS, with PrediXcan-style elastic net on the same bulk samples, partitioned with gene score regression. It fixes the heritability share that any decomposition must account for. We beat it only if we attribute that share correctly, not by raising the total. The file gives scPrediXcan's comparison: bulk TWAS found 111 candidate causal genes in 108 LD blocks for type 2 diabetes versus 222 for scPrediXcan, and 54 genes in 14 blocks for lupus versus 129 in 24 blocks.

The strongest existing method for cell-intrinsic signal is scPrediXcan trained on true OneK1K single-cell data. In our blood experiment it is the ceiling rather than a rival: our method sees only pseudobulk from the same donors. For lupus, success means recovering a large majority of scPrediXcan's 24 lupus LD blocks from pseudobulk alone, while also flagging which blocks run through composition.

The strongest simple competitor is the purity-interaction model of Geeleher et al.. It regresses expression on genotype, fraction, and their product, then feeds interaction eQTLs into TWAS. We will run it on identical fractions, genotypes, and penalty ensembles, with the same tuning budget.

A fourth baseline is bulk TWAS with estimated fractions added as covariates.

For heritability partitioning, the reference number is gene score regression's reported simulation AUPRC of 0.925 against 0.260 for PASCAL. For gene prioritization, it is SUMMIT's reported AUC of 0.777 on known causal genes.

Three ablations isolate each ingredient. The first drops the rank constraint, so composition and intrinsic effects are fit freely. The second sets all $m_k = 1$. The third uses point estimates with no uncertainty propagation.

Why not just…?

Why not just add cell fractions as covariates in bulk TWAS? This throws away real, genetically driven composition heritability rather than measuring it. When fractions are themselves heritable, as the TMEM106B result shows, conditioning on an estimated mediator also biases the remaining genotype effects. It recovers nothing masked in minor cell types. The pseudobulk simulation, where both channels are known, will show the size of the bias directly.

Why not just use the genotype-by-fraction interaction model? This is the real threat. On cell-intrinsic effects in abundant cell types it may match us. It does not constrain composition effects to act through signatures, so it cannot separate a composition QTL from coordinated cell-intrinsic regulation of marker genes. It also ignores mRNA-content and platform scaling. If, on OneK1K pseudobulk, it recovers both channels as well as our model does, the contribution shrinks to uncertainty propagation and the heritability accounting. We will report that outcome plainly.

Why not just run single-cell sequencing and scPrediXcan everywhere? Where a cohort like OneK1K exists, that is better, and we treat it as ground truth. No brain single-cell cohort of comparable size appears in the file, while bulk brain repositories reach thousands of samples. EPIC-unmix found that 57.4% of its cell-type differentially expressed genes from unmixed bulk were missed by single-cell studies of similar size.

Proposed Experiments

Blood, with ground truth. We build pseudobulk from OneK1K donors, whose natural variation in fractions provides real composition QTLs. Truth for cell-intrinsic eQTLs comes from per-cell-type analysis of the single-cell data. Truth for composition QTLs comes from a GWAS on observed fractions. To mimic real bulk, we add platform perturbations: nuclear-to-whole-cell scaling and unequal mRNA content. We report precision and recall for each channel's QTLs, calibration of channel-specific TWAS tests under permuted phenotypes, and lupus LD-block recovery relative to scPrediXcan. The decisive check is the false-positive rate under the null. With propagation it should stay near nominal. Without propagation it should inflate, as MEAD predicts for fraction-based analyses.

Brain, as the application. We apply the method to the 3,387-sample ROSMAP, Mayo, and GTEx brain resource. Fractions are validated against the immunohistochemistry used by SPLITR, and the model is run against Alzheimer's disease and schizophrenia GWAS summary statistics. We report, per trait, the heritability share in each channel and cell type, and the number of bulk-TWAS genes reassigned to the composition channel. No reported brain TWAS heritability figure for these cohorts was found on Lacuna, so our bulk baseline is computed here.

We make two falsifiable predictions. Alzheimer's disease heritability should load on microglial composition and microglial cell-intrinsic channels, consistent with SPLITR's finding that polygenic risk correlates with microglia fraction. Genes near TMEM106B should move from "regulated" to "composition." If the composition channel explains a negligible share in both tissues, the core claim fails. In that case, the contribution reduces to better cell-intrinsic predictors from unmixed bulk.