Gene co-expression networks built from bulk tissue are among the most widely used tools in genomics. Researchers measure RNA from thousands of tumor or brain samples, link genes whose expression rises and falls together across samples, and read the resulting clusters ("modules") as coordinated biological programs. Bulk tissue, however, is a mixture of cell types. Two genes can look co-regulated simply because both are made by immune cells, and some biopsies contain more immune cells than others. A real program running inside a rare cell population can also vanish entirely. Nobody currently reports how often either failure happens, or how rare or how weak a cell-type program can be before a bulk network stops detecting it.
This proposal borrows a habit from searches for physics beyond the Standard Model. When a rare particle decay is sought against an overwhelming background, the experiment does not just report what it found. It first predicts the background, then measures its own detection efficiency, and when nothing is seen it states the smallest signal it could have seen. Belle's search for $B^0 \to K^{*0}\tau^+\tau^-$, for example, reports an upper limit of $3.1\times10^{-3}$ at 90% confidence rather than a bare non-detection (Search for the decay B^0 -> K^{*0} tau^+ tau^- at the Belle experiment). We propose the same discipline for weighted gene co-expression network analysis (WGCNA). The plan has three parts: a composition-only "background" sample that has realistic cell mixtures but no within-cell co-regulation, a signal-injection protocol that plants known programs into chosen cell types at controlled abundance, and a predicted law for how fast signals are diluted. Together these yield a fake-module rate and a sensitivity limit for any bulk network pipeline.
Core Claim
Masking of cell-type programs in bulk co-expression is quantitative and predictable, and it is severe. Write the bulk expression of gene $g$ in sample $s$ as $y_{gs} = \sum_k p_{ks} x_{gks}$, where $p_{ks}$ is the fraction of cell type $k$ in sample $s$ and $x_{gks}$ is that cell type's mean expression of $g$ there. Two assumptions simplify the algebra: proportions vary independently of within-type expression, and cell types vary independently of each other. Under them, the covariance across samples splits into two parts:
$$\mathrm{Cov}(y_g, y_h) = \sum_{k,l} \mu_{gk}\mu_{hl}\,\mathrm{Cov}(p_k, p_l) \;+\; \sum_k \mathbb{E}[p_k^2]\,\mathrm{Cov}(x_{gk}, x_{hk}).$$
Here $\mu_{gk}$ is the average expression of gene $g$ in cell type $k$. The first term is pure composition: it produces modules even when no gene is co-regulated inside any cell. The second term is the biology people want, and it enters weighted by $\mathbb{E}[p_k^2]$. A program in a cell type making up 5% of the tissue therefore reaches the bulk covariance at roughly a four-hundredth of its within-cell strength. WGCNA then raises correlations to a power $\beta$ (typically 6 or more) to enforce a scale-free network shape. This step pushes weak correlations toward zero, so it compounds the dilution instead of correcting it.
The claim to be tested is threefold. Standard bulk WGCNA produces many modules on composition-only data. Recovery of injected programs follows the predicted dilution in mean fraction squared, sharpened by $\beta$. Subtracting a modeled composition term moves the sensitivity limit by a measurable amount and lowers the fake rate. Where proportion estimates are poor, as in the high-purity regime where deconvolution fails, that subtraction will itself leak background.
Relation to Prior Work
Two papers motivate the gap by stating it without measuring it. The Efnb2 knockout study builds bulk brain co-expression networks and concedes that it cannot tell whether a shift belongs to parvalbumin-positive neurons or to compensating glia (RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons). The TCGA breast cancer centromere study runs WGCNA on 1,118 bulk samples and only afterwards checks single-cell data to see which cells express its two genes (Advancing Breast Cancer Biomarkers: A Centromere-Related Gene Signature Integrated with Single-Cell Analysis for Prognostic Prediction). Neither study asks how many of its modules a composition-only dataset would also produce. That question is our primary measurement.
Three lines of work attack bulk averaging, and none of them builds or audits co-expression networks. SubNetDL splits tumors into subclones from DNA allele frequencies and propagates mutations over a fixed protein interaction network (A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer). It never infers RNA co-expression. POPPY maps bulk profiles into single-cell embedding spaces (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level). It produces no gene-gene network. Count Bridges deconvolves bulk counts into cells while respecting their integer nature (Count Bridges enable Modeling and Deconvolving Transcriptomic Data). We treat it as a possible plug-in estimator rather than a competitor.
The deconvolution benchmark on simulated breast tumor mixtures constrains our design (Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures). It shows that nearly all nine methods mislabel normal epithelium as cancer at high purity, so any method that subtracts estimated composition must be stress-tested in exactly that regime. The physics papers supply the method rather than the data. Recoil reconstruction against a fully reconstructed partner meson (Measurement of the B → X_s γ branching fraction and photon energy spectrum using the recoil method) is the template for subtracting a known, modeled component before looking for a faint residual. Reported upper limits are the template for stating sensitivity.
Research Program
The core artifact is a calibration benchmark built from the Wu et al. breast cancer atlas: 100,064 cells from 26 patients spanning ER+, HER2+ and triple-negative tumors, the same atlas used to build 38,000 simulated mixtures in the deconvolution benchmark above.
The first component is the background sample. We construct pseudo-bulk mixtures with realistic, patient-derived composition. Before summing, we shuffle each cell's expression across cells of the same type and same patient-subtype pool, gene by gene. This destroys within-type co-regulation but keeps cell-type means and the composition variation exactly. Any module found here is a fake by construction.
The second component is signal injection. We choose a gene set of size $m$ and a target cell type $k$. We impose within-type correlation $\rho$ on those genes, directly on integer counts, by thinning or boosting reads through a shared per-cell latent factor. We then aggregate the cells while sweeping the mean fraction of type $k$ from about 1% to 40% and setting purity independently.
The third component is the dilution law. From the covariance decomposition we predict the bulk correlation of injected genes, and from that a recovery curve, for each $\beta$. Comparing the predicted and observed curves tests whether masking is really governed by mean fraction squared or whether gene-level noise dominates.
The fourth component is a composition-subtracted estimator. We estimate proportions with BayesPrism and cell-type means from the atlas, compute the first term of the decomposition, subtract it from the sample covariance, and run soft-thresholding, topological overlap and clustering on the residual. A variant uses Count Bridges deconvolution in place of BayesPrism.
Every pipeline is scored the same way. The fake rate is the number of modules, and the fraction of genes in them, found on the background sample. Sensitivity is the smallest product of fraction and correlation at which an injected module is recovered with Jaccard overlap of at least 0.5 in at least half of the replicates. This is the analog of a 90% confidence limit, stated per cell type.
Baselines
The naive reference point is standard WGCNA on normalized bulk pseudo-bulk with the usual scale-free choice of $\beta$. This is exactly the pipeline used in the TCGA centromere study. We have found no reported result on Lacuna for module fake rate or recovery of injected programs for this pipeline, so all values will be measured here. For downstream context only, that study reports survival AUCs of 0.648, 0.609 and 0.624 at 1, 3 and 5 years on TCGA-BRCA.
The strongest simple competitor is proportion regression. Estimated cell fractions (BayesPrism) are regressed out of each gene before running WGCNA. It is the first thing a skeptic would try, and it removes the linear composition term to first order. Its weakness should appear only when proportion estimates are biased, and the TME benchmark tells us that happens at high purity.
The strongest existing method is cell-type-resolved WGCNA. We use the per-cell-type expression that BayesPrism infers for each sample and build a separate network for each cell type. BayesPrism is chosen because the TME benchmark reports that it and DWLS had the fewest combined false positives and negatives; the benchmark gives no numeric error on Lacuna.
All methods use the same mixtures and the same reference atlas. Each gets an identical $\beta$ grid and the same module-size and merge-threshold tuning budget, selected on a held-out set of patients. We report results stratified by tumor purity and subtype.
The upper bound is WGCNA on true per-type pseudo-bulk summed directly from the single cells. Two ablations isolate the contribution of subtraction. The first runs the subtraction estimator with true proportions in place of estimated ones, which separates the value of modeling composition from the damage done by deconvolution errors. The second sets $\beta=1$, which isolates how much soft-thresholding itself worsens masking.
A method beats proportion regression only under specific conditions. It must lower the fake rate on the background sample without raising the sensitivity limit for any cell type, and that improvement must hold in the purity range above 70%. A method beats cell-type-resolved WGCNA if it recovers injected programs in rare types (fractions below 10%) at lower injected strength.
Why not just…?
Why not just regress out estimated cell proportions? This is a real threat. If proportions were estimated without error, regression would remove the composition term almost exactly, and our estimator would add little. The deconvolution benchmark shows the estimates are not error-free at high purity, where normal epithelium is absorbed into the cancer fraction. The leftover composition then passes through regression untouched. The true-proportion ablation settles the matter. If regression with true proportions matches our estimator, and regression with estimated proportions does too, we report that regression is sufficient. In that case the contribution is the calibration benchmark and the measured limits, not a new estimator. We regard that as an acceptable and publishable outcome.
Why not just deconvolve into cell-type-specific expression and run WGCNA per type? Per-type expression inferred from bulk is largely the reference profile scaled by the sample's estimated fraction. For rare types it may carry little information about between-sample covariation beyond composition itself, and it may impose correlations through the prior. The injection experiment measures this directly. If per-type networks recover injected rare-type programs better than the subtraction estimator, we will recommend them. The benchmark is designed to tell, not to presuppose.
Why not just build networks from single-cell data? For discovery in a fresh tissue, one should. The POPPY authors and the TME benchmark both note, however, that clinical archives and cohorts like TCGA are bulk, often from fixed tissue, and cannot be redone at single-cell resolution. The question here is how far to trust the thousands of bulk modules that already drive biomarker papers. Answering it requires calibrating bulk pipelines themselves, and only single-cell-derived mixtures with known ground truth can do that.
Proposed Experiments and Decisive Checks
The first decisive check is the background test. If standard bulk WGCNA finds few modules on composition-only mixtures, the motivating concern is overstated, and the proposal reduces to a negative result worth reporting. We expect the opposite, with immune and stromal modules appearing at rates tracking composition variance.
The second check is the dilution law. On a log-log plot, recovery thresholds against mean fraction should show a slope near minus two in fraction, steepened by $\beta$. A flat curve would mean gene-level noise rather than averaging limits detection, which would redirect effort toward counting models such as Count Bridges.
The third check is the purity sweep. It should reproduce the deconvolution benchmark's failure as rising fake rates for estimated-proportion methods above 70% purity, with the true-proportion ablation showing how much of that rise deconvolution error explains.
Finally, transfer to real data. We apply the calibrated pipelines to TCGA-BRCA (1,118 samples) and report, for each published-style bulk module, its composition-explained share of covariance. We check these attributions against the Bassez and Pal single-cell cohorts, which were not used for calibration. For each cell type we also state the smallest program that bulk TCGA could have detected. This turns "we found no module for this cell type" into a stated limit.
References
- Search for the decay <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mrow><mml:mo>*</mml:mo><mml:mn>0</mml:mn></mml:mrow></mml:msup><mml:msup><mml:mi>τ</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>τ</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math> at the Belle experiment
- RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons
- Advancing Breast Cancer Biomarkers: A Centromere-Related Gene Signature Integrated with Single-Cell Analysis for Prognostic Prediction
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Measurement of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi><mml:mo>→</mml:mo><mml:msub><mml:mi>X</mml:mi><mml:mi>s</mml:mi></mml:msub><mml:mi>γ</mml:mi></mml:math>branching fraction and photon energy spectrum using the recoil method