When a bulk tissue sample shows higher expression of a gene in patients than in controls, there are two very different explanations. The gene may truly be turned up inside the cells that express it. Alternatively, the tissue may simply contain more of a cell type that always expresses that gene. Bulk RNA sequencing averages over thousands of cells and cannot tell these apart. Cell-type composition therefore acts as a hidden common cause: a variable nobody measured that drives both the phenotype contrast and the gene's bulk level. Current tools address this problem in one of two ways, and both miss something. Deconvolution methods produce a single best estimate of cell fractions and treat it as truth. Adversarial sensitivity analysis asks how strong an arbitrary hidden variable would need to be to erase each finding, but says nothing about what that variable is.
We propose to constrain the adversary to be biologically specific. The hidden confounder may only act by shifting cell-type proportions, within a set of compositions that stays consistent with the observed bulk profile, the reference signatures, and their uncertainty. The adversary may also add one previously unseen, nonnegative cell state. Every bulk association then receives a "composition robustness value": the smallest plausible shift in composition that nullifies it. This number can be checked directly. In single-nucleus cohorts where the true per-donor composition is known, genes that are fragile to composition shifts should be exactly those whose association disappears inside individual cell states. That check is the central empirical test of the proposal.
Core Claim
The closest existing work is sensGAN. It trains a generator to find worst-case latent confounders under two limits: $\kappa$, the variance the confounder explains in phenotype, and $\eta$, the variance it explains in expression. Its authors note two limitations. It runs on donor-level pseudobulk, which masks cell heterogeneity, and it cannot say whether a worst-case confounder is biology (for example, an unmeasured stromal subtype) or technical artifact. Our claim is that restricting the adversary to composition shifts turns an abstract robustness score into one that can be interpreted and falsified.
The restriction makes the analysis more interpretable because fragility now means "explainable by a plausible change in which cells are present." It also makes it more useful: at matched confounder strength, composition-fragility should predict within-cell-state failure better than unstructured fragility does, because the adversary spends its budget only on the confounding path that bulk averaging actually creates.
We are explicit about the target quantity. It is the change in expression within a cell state. For that quantity, composition is a nuisance path, whether it confounds disease or is caused by it. Neuron loss in Alzheimer's disease, for example, is a downstream consequence of disease, yet it still contaminates any claim that a gene is "upregulated in cells." This framing sidesteps the mediator-versus-confounder objection that would otherwise undermine adjustment.
Model
For donor $i$ and gene $g$, bulk expression is modeled as
$$x_{ig} = \lambda_{ig}\Big(\sum_{k=1}^{K} \pi_{ik}\, s_{kg} + \pi_{i0}\, u_g\Big) + \varepsilon_{ig}.$$
The terms are as follows. $\pi_i$ is donor $i$'s composition vector on the simplex. $s_{kg}$ is the reference signature of known cell type $k$. $u_g \ge 0$ is an adversarially chosen profile for an unseen cell state, with fraction $\pi_{i0}$. $\lambda_{ig}$ is a gene-specific platform scaling factor, handled as in MEAD. $\varepsilon_{ig}$ is count noise.
The adversary chooses compositions inside a credible set $\Pi_i(\epsilon)$, which contains every composition within divergence $\epsilon$ of a deconvolution posterior. It also chooses $u_g$ under a norm and sparsity budget. Its goal is to maximize the number of genes whose phenotype association, after adjusting for the chosen composition, becomes insignificant. For gene $g$, the composition robustness value $\epsilon^*_g$ is the smallest radius at which the adversary succeeds.
Here is how this differs from sensGAN. That method's adversary lives in an unconstrained space of latent variables. Ours lives in the space of mixtures consistent with the bulk data, so every worst case it returns is a concrete composition that can be inspected, for example "8% more activated microglia and a novel state enriched for genes A and B."
Research Program
The work has three parts.
The first part builds the credible composition set. We take posterior draws from a deconvolution model that reports uncertainty. MEAD supplies valid confidence intervals and corrects for platform scaling. TAPE supplies sample-adaptive signatures through its fine-tuned linear decoder. We use the union of their uncertainty regions, so the adversary is not given a composition set tighter than what the deconvolution methods support. A count-respecting variant uses the Poisson birth-death deconvolution of Count Bridges. This matters because the file flags a real tension: linear continuous models distort integer counts, and the fragility of lowly expressed genes may depend on how the counts are modeled.
The second part trains the constrained adversary. It is optimized by gradient ascent over simplex-parameterized compositions and a nonnegative unseen-state profile. We relax the count of significant genes to a smooth surrogate, as sensGAN does. The unseen-state term directly addresses the atlas-dependence limitation noted for POPPY: cell states missing from the reference can still be explored as candidate confounders.
The third part is the ground-truth validation, which carries most of the weight. We need cohorts where we can build bulk data ourselves and also know the answer. We aggregate single-nucleus or single-cell data per donor into pseudobulk and hide the cell labels. We compute each method's fragility scores. We then compare those scores against the within-state result, obtained by per-state differential expression that conditions on the true composition. A gene whose pseudobulk association vanishes within every state is labeled "composition-driven." A gene whose association persists within at least one state is labeled "intrinsic."
Baselines
The naive reference point is ordinary differential expression on bulk or pseudobulk with observed covariates only. It flags nothing as fragile. It gives the base rate of composition-driven findings that any method must beat.
The second baseline is surrogate variable analysis and principal-component adjustment. These give a single corrected estimate per gene. We convert them to a fragility score by measuring how much each gene's test statistic changes after adjustment. They receive the same covariates and a number of surrogate factors chosen by the standard permutation procedure.
The third baseline is plug-in adjustment for estimated cell fractions. We regress each gene on phenotype plus point-estimate fractions from TAPE, and separately from MEAD. Fragility is again the drop in the test statistic. This is the strongest competitor on our specific question, because it targets the same confounding path with far less machinery. It uses the identical references and deconvolution runs that build our credible sets, so any advantage of our method comes from bounding over uncertainty and unseen states, not from better deconvolution.
The fourth baseline is sensGAN itself. Each gene gets the minimal $(\kappa, \eta)$ that nullifies it. We train it with the same optimizer budget and the same restarts as our adversary. Its reported confounder recovery correlation on simulations is approximately 0.75 (sensGAN). On SEA-AD microglia it reports qualitative results: LGALS3 is robust and AHNAK is highly sensitive. We found no reported result on Lacuna for predicting within-cell-state failure of pseudobulk associations with any of these methods, so all comparisons on our main metric will be new. The one external anchor is qualitative: we will check whether our method reproduces sensGAN's LGALS3 robust / AHNAK fragile split and whether it attributes AHNAK's fragility to composition.
Two ablations isolate the key ingredients. The first fixes composition at the point estimate and removes the unseen state. The second keeps the credible set but drops the unseen state. A third control is an unstructured adversary matched to our adversary's explained variance in expression; it tests whether structure, rather than strength, drives any gain.
Beating the baselines means the following. Our method must achieve higher area under the ROC curve and precision-recall curve for classifying composition-driven versus intrinsic genes than plug-in adjustment and sensGAN. The gain must be consistent across held-out donors. It must be largest for genes marking rare or closely related states, where TAPE reports that point deconvolution struggles.
Why not just…?
Why not just adjust for deconvolved fractions as covariates? This is the serious threat. If point estimates are accurate, plug-in adjustment captures most of the composition path. MEAD argues that naive plug-in ignores estimation error and inflates false positives, and TAPE and POPPY both note degradation for rare states and states absent from the reference. Those are arguments, not measurements of this task. Our first ablation (fixed composition, no unseen state) is plug-in adjustment in sensitivity form, so the experiment settles the question directly. If the full method does not beat that ablation on rare and overlapping states in SEA-AD microglia, we will conclude that point adjustment is sufficient and report it.
Why not just run sensGAN with a larger confounder budget? Raising $\kappa$ and $\eta$ flags more genes but does not say why they are fragile. Fragility then partly reflects how weak a gene's signal is rather than how exposed it is to composition. The variance-matched unstructured adversary tests this. If it predicts within-state failure as well as our constrained adversary, the structural constraint adds nothing, and our method would reduce to a more interpretable way of reporting sensGAN's results.
Why not just use surrogate variable analysis? It returns one correction, not a sensitivity curve. Its factors are unlabeled mixtures of composition, batch, and demographics, so a large statistic change cannot be attributed to composition. It is included as a baseline mainly to show this confound.
Why not just do single-cell differential expression? Where single-cell data exist, one should, and we use it as ground truth. The method matters for the far larger bulk cohorts, such as TAPE's 532-patient Alzheimer's cohort and the 330-patient melanoma immunotherapy cohorts used by POPPY, where no single-cell data are available.
Why not use latent causal discovery such as GIN or CICA? GIN and CICA assume linear non-Gaussian models and search for unknown latent structure. Here the latent variable's structure is largely known: it is a simplex-valued composition acting through reference signatures. Discarding that knowledge to rediscover it would be wasteful. CICA's reported F1 of 0.522 is on a psychometric survey, not count data, which offers no evidence it transfers.
Proposed Experiments
The primary testbed is the SEA-AD microglia single-nucleus cohort. Microglial states play the role of hidden cell types, and the phenotypes are Alzheimer's disease and co-occurring Lewy body pathology. This is the same cohort sensGAN analyzed, which allows direct comparison.
The second testbed is the Onek1k blood-cell resource, used by Count Bridges. Its per-donor pseudobulk has well-annotated immune types, and we will use donor-level covariates available in that resource as phenotypes. A controlled setting uses 3CA tumors and Kassandra reconstructed mixtures. There we inject known composition shifts correlated with a synthetic phenotype, including withholding a cell type from the reference to test the unseen-state term. The primary metric is detection of composition-driven genes. Secondary metrics are calibration of $\epsilon^*_g$ against the true composition shift and, when an unseen state was withheld, recovery of its profile.
A robustness check replaces MEAD/TAPE credible sets with Count Bridges posteriors. It asks whether conclusions about low-count genes change under a count-respecting mixture model. The application study then runs the method on TAPE's Alzheimer's bulk cohort and the SLE cohort. In SLE, sensGAN recovered withheld sex and batch covariates, which lets us check whether our composition-only adversary correctly fails to "explain" associations that are truly driven by batch.
The decisive check is simple. On held-out SEA-AD and Onek1k donors, composition robustness values must separate composition-driven from intrinsic genes better than plug-in adjustment. If they do not, the proposal's central claim is false.
References
- High-Dimensional Sensitivity Analysis for Genomic Studies: An Adversarial Framework for Learning Worst-Case Latent Confounders
- Statistical inference for cell type deconvolution
- Deep autoencoder for interpretable tissue-adaptive deconvolution and cell-type-specific gene analysis
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Generalized Independent Noise Condition for Estimating Latent Variable Causal Graphs
- Conditional Independent Component Analysis for Estimating Causal Structure with Latent Variables