A genome-wide association study (GWAS) points to a stretch of DNA, often holding dozens of genes, where some variant changes disease risk. The usual next step is to ask which gene's expression that variant changes, using bulk RNA sequencing of whole tissue. Bulk tissue, though, is a mixture of many cell types. Suppose a variant makes a tissue contain more of one cell type. Then every gene that marks that cell type will look like it is "regulated" by the variant, even though none of them is. The per-cell expression of those genes never moved; only the headcount of cells did. We propose a model for bulk RNA-seq cohorts with genotypes that splits each variant's apparent effect on a gene into two parts. The first is a composition channel: the variant changes how many cells of type $k$ are present. The second is an intrinsic channel: the variant changes how much of the gene each type-$k$ cell makes. Only the intrinsic channel is then passed into gene-to-trait causal estimation.
The identifying lever is simple and, as far as the research file shows, unexploited for this purpose. It relies on allele-specific reads, meaning RNA reads that can be traced to the maternal or paternal copy of a gene in people who carry one copy of each version (heterozygotes). A variant acting in cis inside a cell expresses its two alleles unequally. A change in cell abundance scales both alleles equally and produces no allelic imbalance. Allele-specific counts therefore carry information about the intrinsic channel only, while total counts mix both channels. Our central prediction is testable. A meaningful share of the "co-regulated clusters" that defeat transcriptome-wide fine-mapping in blood are really composition effects: many marker genes of one cell type tracking a variant that changes that cell type's abundance. Once those effects are removed, the clusters should shrink to one or a few genes.
Where this sits
Three lines of prior work set up the problem. Mediation-based prioritization, such as MRLocus, regresses GWAS effects on expression-QTL effects across several independent variant clusters. It uses a dispersion term to flag pleiotropy, but it was run on bulk tissue expression QTLs, where composition can inflate exactly that dispersion. The blood-cell TWAS study motivates our hypothesis. A transcriptome-wide association study (TWAS) tests whether genetically predicted expression tracks a trait. In that study, 93% of significant associations were driven by known sentinel GWAS hits, and 24% of loci carried complex co-regulation that FOCUS could not resolve. The traits there were blood cell counts, which are composition traits by definition. The authors do not test the composition explanation; we do.
Composition-aware expression QTL mapping supplies our method. CSeQTL models total read counts with a negative binomial and allele-specific counts with a beta-binomial. It showed that ordinary least squares interaction models on log-transformed counts leak strong effects from abundant cell types into false signals in rare ones. We reuse its count likelihood. CSeQTL, however, treats cell proportions as fixed inputs, so it cannot represent a variant that changes those proportions, and it tests variant-gene pairs rather than ranking genes at a locus. SPLITR showed that variants do change proportions. Its cell-fraction GWAS tied the TMEM106B locus to loss of excitatory neurons. It also showed that neurons make roughly four times more mRNA than glia. But SPLITR runs cell-fraction GWAS and cell-type expression QTL mapping as separate analyses and never partitions a single gene's bulk signal between them. IBSEP combines bulk and single-cell summary statistics using cohort-average proportions. By construction, it discards the between-person variation in composition that the composition channel lives on.
Deconvolution methods act as constraints and components rather than competitors. Bisque corrects platform bias between bulk and single-cell data before estimating proportions, and we use it to initialize per-person proportions. TAPE, TCA and Count Bridges recover cell-type-specific signals from mixtures. None of them conditions on genotype.
Core Claim
For gene $g$ in person $i$ with genotype dosage $x_i$ at a candidate variant, we model the expected bulk count as
$$E[y_{ig}] = s_i \sum_k \pi_{ik}(x_i)\, r_k\, \mu_{gk}(x_i).$$
Here $s_i$ is library size. $\pi_{ik}(x_i)=\mathrm{softmax}_k(a_{ik}+\beta_k x_i)$ is person $i$'s fraction of cell type $k$, with a genotype effect $\beta_k$ on abundance. $r_k$ is the total mRNA output per cell of type $k$, following SPLITR's observation that this varies across cell types. $\mu_{gk}(x_i)=\mu_{gk}\exp(\gamma_{gk}x_i)$ is per-cell expression, with intrinsic effect $\gamma_{gk}$.
To first order, the bulk association splits into a composition part, $\sum_k r_k\mu_{gk}\,\partial\pi_k/\partial x$, and an intrinsic part, $\sum_k \pi_k r_k \mu_{gk}\gamma_{gk}$. Allele-specific counts in heterozygotes are modeled as beta-binomial with an imbalance that depends only on $\gamma_{gk}$, weighted by each cell type's read share. The $\beta_k$ are shared across all genes, since one abundance shift touches every marker gene of type $k$. This sharing is what makes them estimable from the locus as a whole rather than gene by gene. Gene-to-trait mediation then follows MRLocus, with the intrinsic effects $\gamma_{gk}$ taking the place of bulk expression-QTL effects. The composition effects $\beta_k$ are reported separately as "cell-abundance mediation," not credited to any gene.
Research Program
We will build the model above on top of the CSeQTL likelihood. Proportions are initialized per person with Bisque and refined jointly. The $r_k$ are either estimated from paired single-cell references or fixed at literature ratios, as SPLITR did in brain. We will write a locus-level fitting routine: all genes within a GWAS window share $\beta$, and each gene has its own sparse $\gamma_{gk}$ under a horseshoe prior. Uncertainty in both channels is carried forward into the MRLocus-style mediation step.
The analysis has three deliverables. The first is a calibration study on pseudobulk with known ground truth. The second is a reanalysis of blood-trait loci measuring how many co-regulated clusters collapse. The third is a map of which prioritized genes change tier, in the sense of Gong et al., when composition is modeled. The main risk is identifiability when a cell type has few informative heterozygotes or low mRNA output. We will report per-locus posterior correlation between $\beta$ and $\gamma$ so that ambiguous loci are flagged rather than silently resolved.
Baselines
The naive reference is standard bulk prioritization: GTEx whole-blood expression QTLs fed into TWAS and FOCUS fine-mapping, as in the blood-cell TWAS. The reported reference point there is descriptive. It is the 24% of loci with complex co-regulation. The file gives no reported gene-level recovery accuracy for it on our benchmark.
The strongest practical competitor is proportion-covariate adjustment. Here, Bisque-estimated proportions for each person are added as covariates before expression-QTL mapping and TWAS. This is common practice, and it removes much of the composition channel directly. It will be run on identical samples, variants and loci, with the same proportion estimates as our model.
The strongest published pipeline is CSeQTL cell-type expression QTLs passed into MRLocus. CSeQTL reports more than 90% directional consistency with independent single-cell studies in CommonMind data, and MRLocus reports 80% credible intervals that contain the truth about 97% of the time in its simulations. Neither reports gene-prioritization accuracy under genotype-driven composition shifts. We found no reported result on Lacuna for that setting.
IBSEP is the representative of summary-level bulk-plus-single-cell integration. It reports a 6.5-fold gain in CD4+ T-cell gene discovery with more than 95% replication in OneK1K. We will run it on the same GTEx whole-blood data (670 individuals) and single-cell references it used.
Two ablations isolate the key ingredients. The first drops the allele-specific term, testing whether total counts plus shared $\beta$ suffice. The second fixes $\beta_k=0$, which recovers CSeQTL with a mediation layer on top.
Beating the baselines means two things on the pseudobulk benchmark. First, higher precision at fixed recall in identifying the true causal gene per locus. Second, correct attribution of spiked abundance effects, where proportion adjustment and CSeQTL→MRLocus misattribute them. On real data it means smaller credible gene sets at the co-regulated loci, with the retained genes replicating as cell-type expression QTLs in OneK1K.
Why not just…?
Why not just regress out estimated proportions before TWAS? This is the real threat. On loci where proportion estimates are accurate, it may remove the composition artifact nearly as well as our model. It has two known weaknesses. First, when done on log-transformed counts, it inherits the leakage CSeQTL documented. Second, proportions are estimated from the same marker genes whose regulation is in question, so estimation error re-enters as residual correlated signal. It also throws away the attribution. It says nothing about whether the locus acts through cell abundance, and that is itself a prioritization result, as TMEM106B shows. The pseudobulk experiment settles this directly. If covariate adjustment matches our gene-level precision, we will report that the contribution is diagnostic (attribution plus calibration) rather than a better ranking.
Why not just use single-cell expression QTLs from OneK1K, which have no composition confounding? Per-cell mapping is clean but underpowered. IBSEP's 6.5-fold gain from adding bulk cohorts shows how much single-cell-only analysis misses. The purpose here is to extract single-cell-like intrinsic effects from the far larger bulk cohorts. We use OneK1K as the truth set rather than as the method.
Why not just chain CSeQTL into MRLocus? That chain holds proportions fixed. A variant that changes abundance therefore leaks into total-count effects that CSeQTL attributes to cell-type-specific regulation of marker genes, and MRLocus then treats those as mediators. The $\beta_k=0$ ablation is exactly this pipeline, so the comparison is built in.
Proposed Experiments
Pseudobulk benchmark. We will build pseudobulk samples by summing OneK1K single-cell counts per individual. This gives exact per-person proportions and per-cell-type expression-QTL truth, including allele-specific counts from heterozygotes. We will then spike in synthetic abundance QTLs at real loci by resampling cells, and spike in synthetic trait effects that are mediated either by abundance or by one gene's intrinsic regulation. The metrics are causal-gene precision and recall per locus, calibration of channel attribution, and the false-mediator rate among marker genes.
Blood-trait reanalysis. On GTEx whole blood, we will reanalyze blood-cell-trait loci and count how many co-regulated clusters reduce to one or two genes. As a check, we will test whether the loci attributed to abundance coincide with known cell-count associations. Retained genes are validated against OneK1K and BLUEPRINT sorted-cell profiles.
Brain positive control. In GTEx brain and CommonMind cortex, with $r_k$ set to SPLITR's roughly four-fold neuron-to-glia ratio, the TMEM106B locus should load on the excitatory-neuron abundance channel. Removing $r_k$ should visibly distort attribution for glial genes.
Decisive checks. The whole proposal hinges on two outcomes. First, whether the allele-specific ablation loses attribution accuracy; if it does not, the lever is weaker than claimed. Second, whether covariate adjustment matches our precision; if it does, we will say the ranking gain is not there.
References
- MRLocus: Identifying causal genes mediating a trait through Bayesian estimation of allelic heterogeneity
- A large-scale transcriptome-wide association study (TWAS) of ten blood cell phenotypes reveals complexities of TWAS fine-mapping
- A computational method for cell type-specific expression quantitative trait loci mapping using bulk RNA-seq data
- Single-cell deconvolution of 3,000 post-mortem brain samples for eQTL and GWAS dissection in mental disorders
- A unified framework for cell-type-specific eQTLs prioritization by integrating bulk and scRNA-seq data
- Accurate estimation of cell composition in bulk expression through robust integration of single-cell information
- Deep autoencoder for interpretable tissue-adaptive deconvolution and cell-type-specific gene analysis
- Tensor Composition Analysis Detects Cell-Type Specific Associations in Epigenetic Studies
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Multi-omic insights into mitochondrial dysfunction and prostatic disease: evidence from transcriptomics, proteomics, and methylomics