When a bulk RNA sequencing experiment on diseased tissue reports that a pathway is "down," two different things may have happened. The cells may have switched the pathway off, or there may simply be fewer of the cells that run it. Standard pathway enrichment cannot tell these apart, and most current fixes first estimate how many cells of each type are in every sample, then try to subtract that out. This proposal takes a different route, borrowed from fluorescence microscopy. Ratiometric sensors divide an analyte-sensitive signal by a reference signal that suffers the same nuisance factors (light intensity, cell thickness, probe amount) but does not respond to the analyte. The nuisance cancels in the quotient. We apply the same trick to genes. For each gene we build a reference "channel" from other genes whose expression across cell types has the same shape, so that any change in tissue composition scales both identically. The log-ratio of gene to reference is then flat under pure composition change and moves only when expression inside cells changes. Pathway enrichment is run on these ratios.
The method needs a single-cell reference only to choose anchors. It never estimates cell proportions, so the downstream test inherits no proportion-estimation error. That error is exactly what Statistical inference for cell type deconvolution (MEAD) showed inflates false positives when estimated fractions are treated as known. If the idea works, a lab with an existing bulk cohort and a public single-cell atlas could report which pathway calls survive once composition is cancelled, with calibrated false-positive control.
Core Claim
A bulk expression value is a mixture. Write $Y_{g,s} = s_g \sum_k c_{k,s} X_{g,k}(1+\delta_{g,k,s})$, where $Y_{g,s}$ is gene $g$'s bulk level in sample $s$, $s_g$ is a gene-specific platform scaling constant, $c_{k,s}$ is the RNA-weighted abundance of cell type $k$ in sample $s$, $X_{g,k}$ is the gene's baseline per-cell expression in type $k$, and $\delta_{g,k,s}$ is the within-cell change we actually care about. Suppose we can find nonnegative anchor weights $w_{g,a}$ over a set of stable genes $a$ such that $\sum_a w_{g,a} X_{a,k} \approx \lambda_g X_{g,k}$ for every cell type $k$, with $\lambda_g$ a constant. When the anchors do not change ($\delta_{a,k,s}=0$), the ratio becomes
$$R_{g,s} = \frac{Y_{g,s}}{\sum_a w_{g,a} Y_{a,s}} \;=\; \frac{s_g}{\lambda_g \bar s_g}\cdot \frac{\sum_k c_{k,s} X_{g,k}(1+\delta_{g,k,s})}{\sum_k c_{k,s} X_{g,k}},$$
where $\bar s_g$ is the weighted platform scaling of the anchors. The first factor is constant across samples. The second equals one when $\delta_{g,\cdot,s}=0$, whatever the composition $c_{\cdot,s}$ is. In a case-versus-control contrast of $\log R_{g,s}$, both the composition term and the gene-specific platform scaling therefore cancel. What remains is an expression-weighted average of within-cell change.
This is the ratiometric condition stated in transcriptomic terms: the two channels share the confounder and differ in their dependence on the target. The claim is that per-gene tests on these ratios, followed by standard gene-set enrichment, hold nominal false-positive rates under pure composition shifts. Bulk enrichment does not. Composition-adjusted regression on estimated proportions is the alternative we must beat on calibration or power, and we treat it as the main threat. The ratio construction is linear, which matters because the mixture itself is linear in counts. Covariate adjustment of the log-scale models used by DESeq2 and limma is not.
Two limits should be stated up front. The ratio cannot recover a gene whose opposite changes in two cell types cancel at fixed composition. Bulk signal is truly flat there, and the ratio is flat too. Gap 1 of the research file is therefore not addressed. The ratio also says that a gene moved within cells, not which cell type moved it. Attribution is a secondary step, described below.
Research Program
Anchor construction. For each gene we solve a nonnegative least-squares problem in log-profile space over cell types, using a single-cell reference. Candidate anchors are restricted in three ways. They must be stable across conditions in the reference (no condition effect in any cell type). They must lie within a band of mean expression around the target. They must not belong to the pathway database's terms for that target. The expression-matching restriction is the transcriptomic analogue of pairing fluorophores with similar photostability. Its job is to blunt the expression-dependent capture and dropout biases that Bisque (Accurate estimation of cell composition in bulk expression through robust integration of single-cell information) and SCDC (SCDC) documented. Those biases break MEAD's assumption that scaling ratios are independent of expression level. Matched anchors should share most such bias with the target, so it largely divides out.
Each gene receives a profile-mismatch score, the residual of the anchor fit. Genes with poor fits, typically highly cell-type-unique genes with no shape-matched partners, are flagged rather than silently tested. The ratiometric literature warns that division cancels confounders only under strict proportionality, and that additive background distorts quotients. The analogue here is low counts and ambient RNA. We handle it with a count-scale pseudocount and by restricting anchors to genes above a detection floor.
Testing and enrichment. Log-ratios enter a standard linear model with condition and technical covariates. The resulting per-gene statistics feed both a threshold-based overrepresentation test against an assay-expressed background and a threshold-free rank test, so results do not hinge on arbitrary cutoffs. Anchor sharing induces correlation between genes' statistics. We estimate this correlation from the overlap of anchor weights and correct the gene-set test with a variance inflation factor, so pathway calls are not inflated.
Attribution. For pathways that survive, we decompose the shift into cell-type contributions. We regress ratio changes on each gene's cell-type expression weights $X_{g,k}/\sum_k X_{g,k}$ across the pathway's genes. This is a pathway-level, deconvolution-free counterpart to the cell-weighted fold-change in scMappR. It is reported as secondary output.
Baselines
The naive reference point is ordinary bulk differential expression with library-size normalization, followed by overrepresentation and rank-based gene-set tests. This is the pipeline the field uses by default. Under composition-only shifts we expect it to call many pathways, and that rate defines the problem.
The strongest competitor is composition-adjusted differential expression. Estimated cell-type proportions enter the per-gene model as covariates, and the same enrichment tests follow. We run it with proportions from three estimators: Bisque, SCDC, and MEAD. Bisque and SCDC are the most accurate deconvolution tools in the file. Bisque reports Pearson $R = 0.923 \pm 0.064$ on the 106-sample subcutaneous adipose benchmark, versus $R = 0.687 \pm 0.450$ for CIBERSORTx. SCDC reports $R = 0.991$ on a three-cell-line mixture at a known 6:3:1 ratio. MEAD is included because it supplies confidence intervals that can be propagated, which makes it the strongest uncertainty-aware version of the baseline. We found no reported result on Lacuna for pathway-level false-positive rate or pathway recovery for any composition-adjusted pipeline, so the numbers in this proposal will be measured, not quoted.
Fairness is ensured by giving every method the same single-cell reference, the same bulk counts, the same gene-set databases, the same background universe, and the same multiple-testing procedure. Tuning is limited to each package's documented defaults plus one shared sweep over the marker or anchor gene count. Beating this baseline means one of two things. The ratio method must keep the pathway false-positive rate under composition-only shifts at nominal level where adjusted regression exceeds it, or it must match calibration while giving higher area under the precision-recall curve for planted within-cell pathway changes.
The second published competitor is scMappR, with cwFold-changes followed by enrichment through g:Profiler. It reports an average Spearman $\rho$ increase of 0.0471 in cell-type specificity over bulk DEGs on the Monaco PBMC benchmark. It is run on the same inputs. It is strongest on attribution, so it is compared mainly there.
Two ablations isolate the key ingredient. The first divides by a single global housekeeping set, a ratio without profile matching. It tests whether matching the cell-type shape is what removes composition. The second drops the expression-band restriction on anchors. It tests whether abundance matching is what absorbs platform bias.
Why not just…?
Why not just add estimated proportions as covariates in DESeq2 or limma? This is the real threat, and the proposal is built around beating it rather than dismissing it. There are two reasons it may fall short. Proportion estimates carry error that MEAD shows inflates downstream false positives when ignored. The adjustment is also additive on the log scale, while composition acts additively on the count scale. The second mismatch matters most for genes expressed in several cell types. If MEAD-based adjustment matches the ratio method on both calibration and power across the simulations below, the honest conclusion is that the ratio method is a deconvolution-free convenience, not an improvement, and we will report it that way.
Why not just use scMappR's cell-weighted fold-changes? scMappR reweights bulk fold-changes by cell-type specificity and point-estimate proportions. It assigns DEGs to cell types but does not remove composition-driven fold-changes from the gene list it starts with. Under a pure composition shift, its inputs are already contaminated. The composition-only simulation measures this directly.
Why not just normalize to housekeeping genes, the obvious "ratiometric" move? A global reference shares the instrument-like confounders, such as library size, but not the cell-type composition of a specific target gene. It therefore cancels the wrong factor. This is the first ablation, and we expect it to behave like the naive bulk baseline for cell-type-restricted genes.
Why not just run single-cell sequencing? For new cohorts that can afford it, single-cell sequencing is better. The target is the large existing bulk cohorts, such as the 636-sample DLPFC or 77-donor islet sets. Dissociation also loses fragile cells, which SCDC documented for adipocytes. Bulk measurements avoid that loss.
Proposed Experiments
The decisive experiment is a controlled pseudobulk study. We use single-cell data with condition labels, aggregate cells into pseudobulk samples, and independently plant two kinds of change. Composition-only shifts change cell-type fractions with no within-cell change. Activation-only shifts multiply a chosen pathway's genes within one cell type at fixed composition. Ground truth is known exactly in both.
For references, we use the pancreatic islet atlases of Xin et al. and Segerstolpe et al. as distributed with SCDC and MEAD, and the Monaco PBMC data with 29 sorted immune types. Anchors are always fit on one dataset and tested on pseudobulk built from another. This forces cross-protocol shape transfer. The mouse mammary gland reference, with both 10X Chromium and Smart-seq2 profiles, provides the hardest protocol mismatch.
The metrics are pathway false-positive rate at nominal 5% under composition-only shifts, precision-recall for planted pathways under activation-only shifts, and both metrics under mixed shifts. We also report the correlation between each gene's anchor-fit residual and its observed miscalibration. This tells us whether the mismatch flag can be trusted. The result that would sink the idea is a ratio method that remains miscalibrated under composition-only shifts once anchors are fit across protocols. That would mean reference shapes do not transfer, and the core assumption fails.
On real bulk data, the islet cohort supplies a known composition effect. SCDC recovered the negative association between HbA1c and beta-cell proportion at $p = 0.0019$. Naive bulk enrichment should therefore report beta-cell secretory pathways as falling with HbA1c. We ask which of those calls survive the ratio test, and whether survivors agree in direction with within-beta-cell differences between diabetic and non-diabetic donors measured directly in the single-cell atlases. That agreement serves as held-out ground truth.
The adipose (106 samples) and DLPFC (636 samples) cohorts used by Bisque test scale and runtime against the composition-adjusted baselines. The mouse kidney regeneration series from scMappR tests attribution to cell types below 5% abundance. There, the question is whether the pathway-level decomposition reaches scMappR's attribution without its dependence on an upstream DEG list.
References
- Statistical inference for cell type deconvolution
- Accurate estimation of cell composition in bulk expression through robust integration of single-cell information
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- Single-cell mapper (scMappR): using scRNA-seq to infer cell-type specificities of differentially expressed genes