Genotype-Driven Error in Cell-Type Deconvolution

Much of human genetics runs on bulk tissue: each sample is ground up and its RNA read as one average over millions of cells. To see past that average, analysts routinely apply deconvolution, which estimates what fraction of each cell type a sample contains by matching its bulk expression against a reference of typical expression profiles for each cell type. The estimated fractions are then treated as measurements. They are tested for genetic association, used as covariates, and compared across ancestry groups. This note argues that the reference itself hides a genetic assumption that nobody checks. The genes deconvolution uses as cell-type markers often carry common variants that change their expression, and the alleles of those variants differ in frequency across populations. A donor who carries a high-expression allele at a T-cell marker will look as though they have more T cells. The resulting error in the estimated fraction is therefore partly a function of the donor's own genotype, and because allele frequencies track ancestry, it is also a function of ancestry.

If this is right, it has two consequences that current tools cannot tell apart from biology. Genome-wide association scans on estimated cell fractions will report hits at the marker genes' own regulatory variants. Comparisons of tissue composition across ancestries will partly reflect differences in marker-gene alleles rather than differences in cells. We propose to measure this error directly in a cohort where true fractions and genotypes are both known. We then correct it by making the reference profiles depend on each donor's genotype at the marker genes, and package the diagnostic as a single call that automated analysis agents can run by default.

Core Claim

Write the bulk expression of gene $g$ in donor $i$ as $y_{ig} = \sum_k \pi_{ik}\, s_{kg}(x_{ig})$. Here $\pi_{ik}$ is the true fraction of cell type $k$, and $s_{kg}(x_{ig})$ is that cell type's expression of $g$ given the donor's local (cis) genotype $x_{ig}$ near the gene. Standard deconvolution replaces $s_{kg}(x_{ig})$ with a population-average profile $\bar s_{kg}$. To first order, the estimation error then becomes a weighted genetic score:

$$\hat\pi_{ik} - \pi_{ik} \approx \sum_{g \in M} w_{kg}\, \beta_{kg}\,(x_{ig} - \bar x_g),$$

In this expression, $M$ is the set of marker genes, $w_{kg}$ is the weight the deconvolution solver effectively places on gene $g$ for cell type $k$, $\beta_{kg}$ is the cell-type-specific effect of the cis variant on expression, and $\bar x_g$ is the reference donors' mean genotype. The linear form comes from differentiating the least-squares or likelihood solution with respect to the profile matrix. Our claim is empirical and falsifiable: this genetic score explains a non-trivial share of deconvolution error, and it produces spurious genetic and ancestry signal downstream.

The correction is to use genotype-conditioned profiles, $s_{kg}(x_{ig}) = \bar s_{kg}\exp\{\hat\beta_{kg}(x_{ig}-\bar x_g)\}$, with $\hat\beta_{kg}$ taken from cell-type-specific expression-QTL estimates. An expression QTL (eQTL) is a variant associated with a gene's expression level. This requires only the genotypes that genetic cohorts already have.

Relation to Prior Work

Several papers motivate the gap. The SPLITR study (Single-cell deconvolution of 3,000 post-mortem brain samples) deconvolved 3,387 brain samples and corrected for age, sex, disease, platform, and neuron-versus-glia mRNA content. It then ran a "cell-fraction GWAS" that found TMEM106B as a driver of excitatory neuron loss, and it mapped 4,757 cell-type-specific eQTLs. That study shows genotype can move true composition. It never asks whether genotype also moves estimated composition through the reference, and separating those two effects is exactly our question.

Two other lines of work build more capable deconvolvers but share the same assumption. POPPY (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level) aligns bulk and single-cell embeddings. Count Bridges (Count Bridges enable Modeling and Deconvolving Transcriptomic Data) models integer counts with birth-death dynamics and reports beating CIBERSORTx and MuSiC. Both treat reference profiles as fixed across donors. Our correction is orthogonal to them and could be added on top of either.

Other papers supply the method. MTWAS (Partitioning and aggregating cross-tissue and tissue-specific genetic effects) separates cross-tissue from tissue-specific eQTL effects on OneK1K and DICE, and we reuse that kind of estimate for $\hat\beta_{kg}$. The population-stratification literature constrains the design. Apparent latent structure within the UK Biobank sample shows that structure survives adjustment for dozens of genotype principal components, and Leveraging the local genetic structure for trans-ancestry association mapping (LOG-TRAM) shows that genetic architecture differs across genomic windows. Both warn that a global ancestry covariate will not absorb an error concentrated at specific loci.

On the accessibility side, GenoMAS automates covariate adjustment for gene-trait tasks, and CellAgent automates single-cell workflows. Neither checks whether a derived covariate carries genetic signal. As automation lowers the skill needed to run these pipelines, fewer users will think to ask that question, so the check needs to be built in.

Research Program

The first stage is measurement in a setting with ground truth. OneK1K profiles peripheral blood from 982 genotyped donors across 14 immune cell types. For each donor we sum the donor's own cells into a pseudobulk sample, whose true fractions are known exactly. We build reference profiles only from other donors, using cross-fitting across donor folds. We estimate cell-type-specific eQTL effects for marker genes on the training folds only, compute each held-out donor's marker genetic score, and measure how much of the deconvolution error it explains. We use CIBERSORTx-style regression, MuSiC, and Count Bridges as the deconvolvers. A genotype-permutation control, which shuffles genotypes across donors, should drive the explained error variance to zero.

The second stage measures downstream harm. On the OneK1K pseudobulks we run cell-fraction GWAS using estimated fractions and using true fractions. Any locus significant with estimated fractions but not with true fractions, and colocalizing with a marker gene's cis-eQTL, is an artifact. We also regress estimated-minus-true fractions on genotype principal components to quantify artifactual ancestry differences in composition. We then move to real bulk brain tissue from ROSMAP, Mayo, and the GTEx brain regions. There we ask whether cell-fraction GWAS hits in SPLITR-style analyses sit at marker-gene eQTLs, and whether they disappear under genotype-conditioned deconvolution. TMEM106B serves as a positive control that should survive.

The third stage is delivery. The corrected deconvolution and a one-number diagnostic, the share of fraction variance explained by marker genetic scores, are exposed as a single function that takes a bulk count matrix, genotypes, and a reference. We add it to the Statistician agent's toolset in the GenoMAS codebase so that any agent-run cell-composition analysis reports the diagnostic automatically.

Baselines

The naive reference point is standard deconvolution with a fixed, population-average reference and no genetic adjustment. On pseudobulks this yields both the error decomposition and the spurious-hit count that everything else must reduce.

The strongest existing deconvolvers are CIBERSORTx, MuSiC, and Count Bridges. The Count Bridges paper reports outperforming the first two, but no numeric deconvolution accuracy on OneK1K pseudobulks was found on Lacuna for any of them. We will therefore establish these numbers ourselves. All three will be run on identical pseudobulks, identical cross-fitted references, and identical marker sets, each given the same hyperparameter search budget on a validation fold. Beating them means lower fraction error, measured as mean absolute error and per-cell-type correlation with true fractions. It also means fewer estimated-only cell-fraction GWAS hits at matched power for true loci.

The strongest stratification baseline is adding the top genotype principal components as covariates in the cell-fraction GWAS. We also run a linear mixed model, which accounts for relatedness through a genetic relationship matrix, via MaSk-LMM (Matrix sketching framework for linear mixed models). That paper reports up to 91% overlap with Regenie's significant associations on UK Biobank, and we use it because it controls genome-wide structure at scale. Beating these baselines means removing artifactual hits that they leave in place.

Two ablations isolate the key ingredient. The first drops every marker gene that has a significant cis-eQTL. The second uses genotype-conditioned profiles with permuted genotypes. The first tests whether conditioning is better than exclusion; the second tests whether the gain comes from the donor's actual genotypes rather than from added flexibility.

For the delivery stage, the reference is GenoMAS as published, which reports 60.48% gene-identification F1 on GenoTEX. We compare it with and without the diagnostic on semi-synthetic composition tasks with planted artifacts.

Why not just…?

Why not just adjust for genotype principal components? Principal components capture broad ancestry axes. The error we describe is concentrated at a few dozen marker loci, and at those loci the artifact is a genuine cis association between the donor's genotype and the estimated fraction. Adjusting for global ancestry does not remove a local genotype-to-estimate path, which is the same reason LOG-TRAM argues for windowed rather than global treatment. Stage two tests this directly by counting artifactual hits that survive principal-component and mixed-model adjustment.

Why not just drop marker genes that have eQTLs? This is a real threat. If exclusion matches genotype-conditioned profiles on error and on artifact removal, the simpler approach wins, and we will report that. Our expectation is that exclusion costs accuracy for rare cell types, because strong markers for small populations are few and many carry eQTLs. The exclusion ablation, broken down by cell-type abundance, settles this question. If the accuracy loss is negligible, we will recommend exclusion and keep only the diagnostic.

Why not just build ancestry-matched references? Matched single-cell references exist for few tissues and ancestries. They also do not address within-ancestry variation at marker loci, since the error term depends on each donor's genotype, not their population label. The genotype-conditioned approach needs only genotypes, which genetic cohorts already hold. Stage one's within-cohort cross-fitting shows whether error persists even when reference and target donors come from the same population.

Why not let an agent choose covariates? GenoMAS already automates covariate selection. However, nothing in its toolset tests whether a covariate is itself genetically driven. An agent can only choose among checks that its toolset makes available.

Proposed Experiments and Decisive Checks

The idea stands or falls on stage one. If, across deconvolvers, the cross-fitted marker genetic score explains essentially none of the pseudobulk fraction error, and the permutation control matches the real score, the claim is wrong and the project stops. The second decisive check is whether estimated-fraction GWAS on OneK1K pseudobulks yields hits absent from true-fraction GWAS that colocalize with marker cis-eQTLs, and whether genotype conditioning removes them while preserving hits found with true fractions. The third is real-data transfer: in the brain cohorts, we test whether any published-style cell-fraction associations lose support under conditioning while TMEM106B retains it.

For the agent stage, success means the augmented GenoMAS flags planted artifacts that the stock version reports as findings, with unchanged performance on tasks without artifacts. Every comparison uses the same donors, folds, and tuning budget, and we report results per cell type so that averages do not hide failures on rare populations.