Selection-Aware Cell-Type eQTL Mapping in Extreme-Phenotype Cohorts

Many genetic studies save money by sequencing only the people at the two ends of a trait distribution: the heavy smokers who never get cancer and those who get it young, or the very old who stay sharp despite brain pathology and those who decline early. This design is called extreme phenotype sampling. When such studies also measure gene activity in a tissue sample, they usually grind up the whole sample, a "bulk" measurement that averages over many cell types. Our observation is that these two habits interact badly. Selecting on the trait also selects on whatever cell makeup drives the trait, such as neuron loss in the brain or immune infiltration in a biopsy. Within the selected people, cell makeup then becomes correlated with genotype and with expression in ways it is not in the population. As a result, standard methods that unmix bulk tissue into per-cell-type signals, and then test which DNA variants change gene activity in which cell type, can report false or distorted cell-type-specific genetic effects.

We propose a mapping method for cell-type-specific expression quantitative trait loci (eQTLs: DNA variants associated with a gene's expression level) that models the selection step explicitly. It fits bulk expression, estimated cell proportions, genotype and phenotype jointly, with a likelihood that conditions on each person having been sampled from a tail. We test it by building simulated extreme-sampling designs inside a population brain cohort where single-nucleus data give us ground truth. A second, more exploratory aim pairs the corrected model with sequence-based variant scores. Its purpose is to interpret the rare non-coding variants that tail cohorts are designed to find. Put in two sentences: extreme sampling turns cell composition into a source of spurious genetic signal in bulk tissue. We correct this by writing the sampling rule into the deconvolution-plus-eQTL likelihood, and we measure how much it matters against single-cell truth.

Core Claim

Two existing lines of work never touch each other. The first is extreme-sampling statistics. MREPS showed that two-stage least squares Mendelian randomization is biased when samples come from outcome or exposure tails. It fixed this with a likelihood that conditions on the sampling rule, and it reached full-cohort conclusions with about 40% of the samples. It works only on macro-phenotypes with pre-specified instruments, so it has no tissue, cells or expression in it. The second is cell-type unmixing of bulk tissue for genetics. SPLITR/deQTL deconvolved 3,387 brain samples, ran a GWAS on cell fractions that tied TMEM106B to excitatory neuron loss, and reported 4,757 cell-type-specific eQTLs. EPIC-unmix updates single-cell priors from the target bulk data to infer per-person cell-type expression. Both were built and validated on unselected aging cohorts. EPIC-unmix's prior update is especially exposed, because it learns from the bulk distribution, and under tail selection that distribution is deliberately unrepresentative.

The claim we test is specific. When the trait depends on cell composition, tail selection biases (a) genotype-by-composition associations and (b) cell-type eQTL effect estimates from bulk unmixing. A selection-conditioned joint likelihood removes most of that bias, without the large sample-size penalty of simply discarding the design's efficiency. The motivating case is already in the literature. The S100A13 methylation study compared diabetics resistant to retinopathy against rapid progressors using bulk blood. It cannot tell whether the signal is intracellular methylation or a shift in circulating immune cells, and the same ambiguity covers every bulk-omics extreme-phenotype study.

Research Program

The model has four parts. Bulk expression of a gene in person $i$ is a weighted sum of cell-type expression. The weights are cell proportions scaled by per-cell-type mRNA content, following SPLITR's correction for neurons producing roughly four times more mRNA than glia. Cell-type expression depends on genotype through the cell-type eQTL effects we want to estimate. Cell proportions may themselves depend on genotype, which is the cell-fraction GWAS signal. The phenotype depends on proportions, on relevant cell-type expression and on covariates. The departure from prior work is the sampling correction. Writing $y_i$ for phenotype, $b_i$ for bulk expression, $g_i$ for genotype and $T$ for the union of sampled tails, we maximize

$$ L(\theta) = \prod_i \frac{p_\theta(y_i, b_i \mid g_i)}{\Pr_\theta(y_i \in T \mid g_i)}, $$

The numerator integrates over the unobserved cell proportions and cell-type expression, using single-cell reference priors. The denominator integrates the phenotype model over the tails. This is the retrospective-likelihood logic of MREPS, here applied to a latent-variable deconvolution model. The denominator is what stops a person's tail membership from being read as evidence about their composition or their eQTLs.

Inference uses variational expectation-maximization. The cell-type expression prior is initialized from snRNA-seq, and the selection integral is computed by quadrature over a one-dimensional phenotype. The non-identifiability flagged in the composition-confounding literature is handled with fixed reference priors rather than free unmixing. We report sensitivity to prior strength.

For the second aim, tail cohorts are enriched for rare large-effect variants, which have too few carriers for eQTL regression. For carriers we use the corrected model's per-cell-type expression residuals together with AlphaGenome predictions of the variant's effect on cell-type-matched accessibility and expression tracks. We ask whether the cell type with the largest predicted sequence effect matches the cell type carrying the expression shift. This aim is exploratory, and the core contribution stands or falls on the first.

Baselines

The naive reference point is a standard bulk eQTL analysis run on the tail subsample, with estimated cell fractions as linear covariates. This is what most extreme-phenotype omics studies would do today.

The strongest existing competitor for estimating cell-type expression is EPIC-unmix followed by per-cell-type eQTL regression. It reports 4.6% to 109.8% higher accuracy than TCA and bMIND on brain and PBMC benchmarks (EPIC-unmix). We found no reported result on Lacuna for EPIC-unmix or deQTL under extreme-phenotype sampling. Since that gap is the question, we establish the number ourselves.

We also run deQTL as the SPLITR authors describe it, and the purity-interaction eQTL model of Geeleher et al. extended to multiple cell fractions. That interaction model is the simplest way to separate cell-type effects in bulk.

Every method gets the same bulk RNA-seq, genotypes, snRNA-seq reference, covariates and identical tail subsamples. Each gets the same budget of 20 hyperparameter configurations, chosen on a held-out set of genes.

The key ablation is our own model with the selection denominator removed. It is identical in every other respect, so it isolates the contribution of conditioning on the sampling rule.

We count it as beating the baselines if three things hold. Type I error for genotype–composition and cell-type eQTL tests is calibrated (at most 0.06 at nominal 0.05) where the baselines are inflated. Effect-size bias relative to the full-cohort truth falls by at least half compared with EPIC-unmix. Recall of truth-set cell-type eQTLs at 10% FDR is no worse than EPIC-unmix.

Why not just…?

Why not just add estimated cell fractions as covariates? Under tail selection, composition sits on the path from genotype to phenotype, which is the very thing selection acts on. Adjusting for it does not undo the selection-induced dependence. A single covariate also cannot say which cell type carries an eQTL. The naive baseline tests this directly. If it turns out calibrated in our simulations, the confounding concern is overstated and we will report that.

Why not just run single-nucleus sequencing on the tail samples? This is a real threat. Extreme sampling makes cohorts small, which weakens the usual objection that single-cell profiling is too expensive. We do not argue it away. We run a budget-matched comparison inside ROSMAP: few tail donors with snRNA-seq against many tail donors with bulk plus our model, scored on recall of truth-set cell-type eQTLs. Two outcomes are possible. If single-nucleus wins at realistic budgets, the method's value reduces to reanalysing the many existing bulk-only extreme cohorts, such as post-mortem banks and archived biopsies. That use is narrower, but still real, given dissociation artifacts and tissue availability.

Why not run EPIC-unmix as is and apply an MREPS-style correction afterward? A post-hoc correction on outputs inherits the biased bulk-driven prior update, because the error enters before eQTL testing. The two-stage version is included as an additional baseline. If it matches the joint model, we will recommend it as the simpler tool.

Why not just score variants with AlphaGenome and skip expression? AlphaGenome improves eQTL sign prediction from 75% to 80% auROC (AlphaGenome). Its authors note weak cell-type and activation-state resolution, and MPRA work on Alzheimer's variants shows regulatory effects flipping direction between resting and activated microglia. Sequence scores alone therefore cannot settle cell context, which is why the second aim tests whether the two sources agree rather than assuming they do.

Proposed Experiments

We use ROSMAP, which has bulk DLPFC RNA-seq, genotypes, neuropathology, cognition and snRNA-seq on a subset. The phenotype is cognitive resilience: residual cognitive decline after regressing on amyloid and tau burden. We draw tail subsamples at the top and bottom 10%, 20% and 30%. Ground truth comes from cell-type eQTLs and cell fractions estimated in the full cohort using snRNA-seq where available and full-cohort deQTL otherwise. Mayo and MSBB serve as replication for whether hits survive.

The positive control is TMEM106B. Selecting on resilience should enrich excitatory-neuron differences, and the corrected model should recover the full-cohort TMEM106B cell-fraction effect size, where naive analysis should inflate it.

The null control is a semi-synthetic simulation. We plant cell-type eQTLs with known effects into real genotypes and snRNA-seq-derived profiles. We give the phenotype a composition dependence of varying strength and set genotype–composition effects to zero. This measures type I error as a function of how strongly the trait depends on cells. We expect inflation to grow with that dependence.

The decisive check is whether the tail-sampled corrected model matches full-cohort cell-type eQTL truth at a fraction of the samples, analogous to MREPS's 40%, while EPIC-unmix and deQTL on the same tails show inflated or sign-flipped effects for cell types tied to the phenotype. If neither the bias nor the correction shows up, extreme sampling is safe to combine with existing unmixing tools, and we will publish that conclusion with the same benchmark.