Most clinical gene-expression data is "bulk". A tissue sample is ground up and its RNA averaged over hundreds of thousands of cells. Computational methods called deconvolution try to undo that averaging, recovering which cell types were present and what each one was expressing. These methods are usually graded on whether they recover cell-type proportions. Scientists then use them for something else: they claim that a particular cell type responds to a drug, a mutation or a disease, and that this response drives the phenotype. Nobody has checked whether that second, regulatory kind of claim survives averaging and unmixing, because doing so requires knowing the true cell-type-specific response to an intervention.
That ground truth now exists. Large CRISPR screens read out by single-cell RNA sequencing (scRNA-seq, which measures the transcripts of each cell individually) knock down the same genes in several human cell lines. Each knockdown is a physical intervention whose effect in each cell type is directly measured. We propose to build bulk samples from these screens with controlled composition, controlled cross-context divergence and controlled composition shifts, and then ask a simple question. When a deconvolution or bulk-analysis method reports that knocking down gene $g$ shifts gene $j$ in cell type $k$, is that true? The central bet is that proportion accuracy does not predict effect accuracy. We also bet that failure follows a measurable law: it grows with how differently a regulator acts across the cellular contexts in the mixture.
Core Claim
Our claim is that heterogeneity masking of regulatory effects has two separable parts. One is cancellation, where a regulator pushes a target up in one cell type and down in another. The other is composition confounding, where the cell mix differs between conditions. For the mean bulk shift of gene $j$ under perturbation $g$, the identity is
$$\Delta^{\text{bulk}}_{gj} = \sum_k \pi_k\,\delta_{gjk} + \sum_k (\pi'_k - \pi_k)\,\mu'_{jk},$$
where $\pi_k$ and $\pi'_k$ are the fractions of cell type $k$ in the control and perturbed samples, $\delta_{gjk}$ is the true effect in cell type $k$, and $\mu'_{jk}$ is the mean expression of $j$ in perturbed cells of type $k$. The identity is trivial for means. Its consequence is not trivial. Any method that recovers $\delta_{gjk}$ from bulk must break this degeneracy using priors, cohort-level variation in $\pi$, or a reference atlas. How well it does so should depend on the cross-context dispersion of $\delta_{g\cdot k}$, which we call the divergence of the regulator.
We treat genomic complexity scaling as the lens that explains why this matters. In simple organisms, regulatory rules are local and transfer across contexts. In mammals they are distributed and context-dependent. Context dependence is exactly what makes $\delta_{gjk}$ diverge across cell types, and therefore exactly what averaging destroys. Our benchmark varies the number of contexts and their divergence, so we can measure how regulatory-claim accuracy degrades as heterogeneity grows. We expect this to be a much sharper diagnostic than proportion error.
Relation to Prior Work
The motivating evidence comes from three places. Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery showed that deep perturbation predictors look good on generic metrics but distort the gene-gene structure needed for causal inference, with effect sizes inflated about 23-fold. None of them beat a sparse ElasticNet. We ask the analogous question for the other way the field produces "virtual" single-cell data: unmixing bulk.
That paper also found that its own ground truth was weak. Regulatory networks inferred from real Perturb-seq data with the causal discovery algorithm GIES were "not robustly above chance" against the curated database TRRUST. This constrains our design. We do not use inferred networks as truth. Our truth is the directly measured, intervention-specific per-cell-type effect $\delta_{gjk}$, kept only where the E-test from scPerturb: Harmonized Single-Cell Perturbation Data calls it significant. We also respect that paper's guideline of roughly 300 cells per perturbation and 1,000 UMI counts per cell.
The methods we audit come from the deconvolution line. Count Bridges enable Modeling and Deconvolving Transcriptomic Data treats single-cell counts as latent variables that must sum to the observed bulk, and reports beating CIBERSORTx and MuSiC on proportions. It does not test whether perturbation effects are preserved. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level (POPPY) aligns bulk and single-cell encoders with ontology-aware contrastive learning. It attributes clinical outcomes to cell subsets through cross-attention, but it validates only integration and composition, not whether the attributed subsets carry the effect. We do not propose a better deconvolver. We propose the missing test of what deconvolvers are used for.
Two further papers show the confound in the wild. Identifying miRNA-mRNA regulatory relationships in breast cancer with invariant causal prediction treats breast-cancer subtypes as causal "environments" even though the subtypes differ in immune and stromal infiltration, so composition is entangled with invariance. Estimating drivers of cell state transitions using gene regulatory network models (MONSTER) found that subtracting edge weights between healthy and diseased bulk networks gave $R^2 < 0.01$ across cohorts, but it never separated composition shifts from rewiring. Our second masking term is precisely the quantity both papers leave uncontrolled.
Causal identification of single-cell experimental perturbation effects with CINEMA-OT separates confounders from treatment effects using optimal transport, but only when single cells are observed. We use it as the single-cell upper reference, not as a bulk competitor.
Research Program
The first component is a mixture generator built on Replogle CRISPRi screens in K562, RPE1 and HepG2. These are harmonized in scPerturb, and the HepG2 screen is the one used in Elucidating the Design Space of Generative Models for Single-Cell Perturbation Prediction. The generator works as follows:
- We restrict to perturbations targeting genes shared across lines. The exact overlap must be verified before committing, since the screens differ in scope.
- For each perturbation we sample perturbed and control cells from each line at specified proportions and sum their raw counts. This gives integer pseudo-bulk profiles, with replicates drawn from disjoint cell subsets.
- We add within-line contexts by splitting each line into cell-cycle-phase clusters. This raises the number of contexts $K$ from 3 to roughly 9 without new data.
The generator has three settings that can be tuned independently. The first is divergence, set by choosing perturbations whose measured $\delta$ vectors agree or disagree across contexts. The second is responder rarity, where only a fraction $f$ of the mixture comes from a responding context, mimicking a minority population. The third is composition shift, where $\pi' \neq \pi$ between control and perturbed bulk. A cohort mode generates $N$ bulk samples per condition with $\pi$ drawn from a Dirichlet distribution, which makes regression-based methods identifiable.
The second component is the evaluation protocol. Every method is given bulk profiles plus a single-cell reference from unperturbed cells, which is what a clinical analyst has. Each method must output per-context effect estimates $\hat\delta_{gjk}$. We score four things:
- Sign agreement and Spearman correlation with the measured $\delta_{gjk}$ over E-test-significant pairs.
- Recall of context-specific responders, meaning genes significant in one context but not another.
- False context attribution, meaning effects assigned to a context where none was measured.
- Proportion error, reported separately so its relationship to effect error can be measured directly.
For methods that return single-cell-level samples (Count Bridges), we also compute the E-distance between deconvolved and true perturbed distributions per context.
The third component is the analysis that turns scores into a claim. We regress per-perturbation effect error on divergence, $K$, $f$ and composition shift magnitude. We expect divergence and composition shift to explain most of the variance, and proportion error to explain little.
Baselines
The naive reference is the bulk shift itself. It assigns $\Delta^{\text{bulk}}_{gj}$ to every context, which is what an analyst does when interpreting bulk differential expression. Every method must beat it on context-specific responder recall. Otherwise unmixing adds nothing for regulatory claims.
The strongest simple competitor is cell-type-specific regression. Across the $N$ cohort samples, we regress bulk expression on estimated proportions with proportion-by-condition interaction terms, using an ElasticNet penalty. This is the most direct inversion of the identity above. The sparse-linear result in Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery is a reason to expect it to be hard to beat. The regularization strength is tuned by cross-validation on held-out perturbations, with the same tuning budget as all other methods.
The strongest existing deconvolvers are Count Bridges, run with its EM projection-guided sampler, and POPPY's cross-attention attribution. Both are trained on the same unperturbed reference and given the same compute. CIBERSORTx-style signature deconvolution, followed by per-type differential testing, serves as the classical representative.
The upper reference is pseudobulk per-context differential expression and CINEMA-OT run on the actual single cells. This shows how much signal is lost to aggregation at all.
The key ablation removes the composition-shift term. Methods are run with $\pi' = \pi$ forced in the generator and then with realistic shifts. This isolates how much failure comes from confounding rather than cancellation.
No reported result was found on Lacuna for per-context perturbation-effect recovery by any deconvolution method. Count Bridges' advantage over CIBERSORTx and MuSiC is reported only qualitatively, for proportions. POPPY's gains (up to 73% silhouette improvement, up to 714% higher composition correlation) are integration metrics, not effect metrics. "Beating" the regression baseline will therefore mean a statistically significant gain in Spearman correlation on high-divergence perturbations, held across at least five generator seeds. Seed robustness matters because PerturbCausal showed a single-seed STATE advantage that failed to replicate.
Why not just…?
Why not just report deconvolution proportion accuracy, as the field already does? Because the downstream claims are about effects, not proportions. The identity shows that perfect proportions still leave $\delta_{gjk}$ underdetermined when divergence is high. This is a falsifiable bet, not an assumption. If proportion error turns out to predict effect error well across the generator's settings, the benchmark's main message collapses. We would report that, since it would license the current practice.
Why not just regress bulk expression on proportions with interaction terms? This is a serious threat, and it may well win, so it is our strongest baseline rather than a straw man. Its known weakness is identifiability. It needs many samples with varied composition, and in cohorts with $N$ small or $\pi$ nearly constant its estimates should degrade sharply. The cohort-size sweep directly measures the crossover point. If regression dominates at all realistic $N$, the practical conclusion is that the simpler method suffices, and we would say so.
Why not just predict per-cell-type responses with a virtual cell model or foundation model? PerturbCausal already documents roughly 23-fold effect inflation and responses that were nearly identical across perturbations. That is precisely the failure that would erase cross-context divergence. We include GEARS-style prediction only as a secondary comparison, since the question here is what bulk data can support.
Why not just infer causal networks from deconvolved data and compare them to Perturb-seq networks? The GIES networks on real Perturb-seq data are themselves near chance against TRRUST, so agreement or disagreement with them is uninterpretable. Measured interventional effects avoid that problem entirely.
Decisive Checks
Three outcomes would settle the claim:
- Effect error should rise with divergence and composition shift, after controlling for proportion error, in the regression analysis over the generator.
- At least one method should beat the naive bulk reference on context-specific responder recall at $f \le 0.1$. If none does, rare-responder claims from bulk are unsupported by any current method, which is itself a publishable negative result.
- The pseudo-bulk results should hold in true bulk. Simulated pseudo-bulk may differ from real bulk in library preparation and ambient RNA, so we propose a small physical experiment as the final validation-gap test. For a handful of high- and low-divergence perturbations, CRISPRi-perturbed K562 and RPE1 cells would be mixed at known ratios before RNA extraction and bulk-sequenced. The ranking of methods and the divergence-error relation would then be checked against the pseudo-bulk predictions.
If the ranking reverses on real mixtures, the generator is not trustworthy and we would revise it before drawing clinical conclusions.
References
- Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery
- scPerturb: Harmonized Single-Cell Perturbation Data
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Identifying miRNA-mRNA regulatory relationships in breast cancer with invariant causal prediction
- Estimating drivers of cell state transitions using gene regulatory network models
- Causal identification of single-cell experimental perturbation effects with CINEMA-OT
- Elucidating the Design Space of Generative Models for Single-Cell Perturbation Prediction