Researchers studying medicinal plants, parasites and other non-model organisms often grind up a whole organ, sequence its RNA, and group genes into "modules" that rise and fall together across samples. They then read these modules as regulatory programs. The standard tool is weighted gene co-expression network analysis (WGCNA), which links genes by how strongly their expression correlates across samples and then clusters them. There is a problem with this reading. If one sample contains more xylem and another more storage tissue, every xylem-specific gene goes up together, even though nothing inside any cell was co-regulated. The resulting module describes a change in which cells were in the tube, not a regulatory program. In human biology this is fixed by estimating cell-type proportions against a single-cell reference atlas. Most organisms have no such atlas.
We propose a method that removes the cell-mixture component from bulk co-expression networks using only the bulk data. We also propose a benchmark that measures whether the removal worked, built from paired simulated tissues whose true composition-free network is known. The method uses a structural fact: on the linear (non-log) scale, a bulk sample is a nonnegative weighted sum of cell-type profiles, and the weights sum to one. Correlation caused by changing proportions therefore has a constrained low-rank form that genuine within-cell regulation does not have. We fit that form, carry its uncertainty forward, and give every module a score for how much of its behavior is explained by composition.
Core Claim
The claim is that a small-sample, reference-free mixture fit with simplex constraints separates composition-driven co-expression from within-cell co-expression better than the generic fix of regressing out top principal components. The advantage should be largest when a strong genuine regulatory program has high variance, because principal-component removal then deletes it along with the mixture signal.
Write $y_{gs}$ for the depth-normalized linear expression of gene $g$ in sample $s$. We model
$$y_{gs} = \sum_{k=1}^{K} w_{gk}\, p_{ks} + e_{gs}, \qquad w_{gk} \ge 0,\; p_{ks} \ge 0,\; \sum_k p_{ks} = 1,$$
where $K$ is the number of latent cell compartments, $w_{gk}$ is gene $g$'s mean expression in compartment $k$, $p_{ks}$ is that compartment's share of sample $s$, and $e_{gs}$ is everything else, including within-cell regulation. Under this model the composition part of the gene–gene covariance is $W\,\mathrm{Cov}(p)\,W^\top$. Here $W$ collects the $w_{gk}$ and $\mathrm{Cov}(p)$ is the covariance of the proportions across samples. Its loadings are nonnegative mean-expression profiles, while a principal component's loadings are unconstrained. We build the WGCNA network from correlations of $e$, conditional on estimated $p$, rather than from raw $y$.
The method will not recover single-cell networks. The target is the network the same bulk experiment would have produced if composition had stayed fixed across samples.
Relation to Prior Work
Three lines of work motivate the gap.
Single-Cell Co-expression Analysis Reveals Distinct Functional Modules, Co-regulation Mechanisms and Clinical Outcomes compared glioblastoma bulk and single-cell networks. Fewer than 10% of top co-expressed pairs were shared, and the bulk pairs were dominated by broad pathways and neighboring genes. That paper diagnoses the distortion but relies on human references to work around it.
Plant studies show the same confound in practice. The Peucedanum praeruptorum bolting study ran WGCNA on 24 de novo-assembled whole-root samples during a transition with heavy xylem lignification. The Alpinia officinarum genome paper found its AobHLH94 hub from homogenized organs.
The Efnb2 knockout study shows the problem is not specific to plants. Its authors explicitly could not say whether a bulk cortical module came from the targeted interneurons or from their neighbors.
The Artemisia trichome WGCNA study sidestepped the issue by isolating one cell type physically. That only works for surface structures.
Several methods supply partial solutions, but each needs something non-model systems lack.
- POPPY and Count Bridges deconvolve bulk RNA well but are trained against single-cell corpora and, for POPPY, a cell ontology.
- DeepCSCN builds cell-type networks but needs single-cell input.
- MVTLASSO removes study-specific noise but needs many independent datasets and treats each sample as homogeneous.
- SNAIL fixes spurious correlations created by quantile normalization of zero counts. That is a different artifact, and we adopt it as the normalization step rather than compete with it.
The closest method is DeClust. It is reference-free, fits a three-compartment cancer/immune/stroma linear model to estimate cancer-intrinsic expression, and needs roughly 200 samples. Our work differs from it in three ways:
- We target co-expression networks rather than subtype means.
- We work in the 12–50 sample regime typical of non-model studies.
- We choose the number of compartments from the data rather than fixing it at three.
Two papers constrain the design. Statistical inference for cell type deconvolution shows that treating estimated fractions as fixed inflates false positives downstream, so uncertainty must be propagated. The breast-cancer deconvolution benchmark shows that cell types sharing a developmental lineage get confused, so we expect related compartments to merge and we measure the cost.
Research Program
Fitting the model. The fit is a nonnegative factorization with simplex-constrained proportions on SNAIL-normalized linear counts, under a negative-binomial likelihood. We choose $K$ by held-out reconstruction across samples. Because small $n$ makes the factorization unstable, we refit on bootstrap resamples of samples and random restarts.
Building the network. For each fit we compute residual correlations, raise them to the WGCNA soft-threshold power, and compute topological overlap. We keep the consensus overlap across fits, so an edge must survive the uncertainty in $p$.
Composition share. For each module we report the fraction of its eigengene variance (the first principal component of the module's genes) explained by $\hat p$, averaged over fits. A practitioner can then label modules as mostly-composition, mixed, or mostly-regulatory.
Evaluation protocol. The protocol is the second contribution. From a real single-cell atlas we build paired pseudo-bulks from the same donors and cells:
- Fixed-composition arm: proportions are held constant and only within-type expression varies. Its WGCNA network is the ground truth.
- Varying-composition arm: proportions are drawn over a realistic range, as in the 5–95% purity sweep of the breast-cancer benchmark.
Every method sees only the varying arm and must recover the fixed arm's network. We also add strong planted within-type programs to test whether corrections destroy real signal.
Baselines
Uncorrected bulk WGCNA. This is the naive reference point: standard WGCNA on SNAIL-normalized varying-composition data. Every published non-model study above does this. Any correction must beat it on recovering the fixed-composition network.
Top-principal-component removal. Regress out the top $K$ principal components before WGCNA, with $K$ matched to the mixture model's choice and the same normalization and soft-threshold selection. This is the strongest generic competitor because it is what careful practitioners already do. It also serves as the ablation that isolates our key ingredient, the nonnegative simplex structure, since the two methods then differ only in constraints. We count as beating it when two things hold: significantly higher precision–recall area for top fixed-arm edges at the same $K$, and better retention of planted programs.
DeClust-style residuals. Run DeClust's compartment subtraction and compute WGCNA on the resulting intrinsic expression. This is the closest existing reference-free method. We run it at both its intended sample size and our small-$n$ settings.
Oracle-proportion and reference-based corrections. Regress out the true simulated proportions (oracle), or proportions from a reference-based deconvolution trained on an independent atlas cohort. These are upper bounds, not competitors. Closing most of the gap between uncorrected and oracle-proportion results is the success criterion.
Single-scale ablation. Our method with uncertainty propagation removed, meaning a single fit. This tests the false-positive warning of Statistical inference for cell type deconvolution.
All methods receive identical normalized input, the same gene filter, the same soft-threshold rule, and an equal budget for choosing $K$ or the number of components.
Reported numbers. No reported result was found on Lacuna for any method's accuracy at recovering a composition-free co-expression network, so we cannot quote a number for the strongest baseline on this exact task. The nearest published figures address other tasks:
- On normalization, SNAIL reduced cross-tissue false-positive associations from over 3,400 to 231 on GTEx V8, and correlation RMSE from 0.389 to 0.015 on ERCC spike-ins (SNAIL).
- On the naive state of bulk networks, fewer than 10% of top pairs overlapped with single-cell networks in glioblastoma (Wang et al.).
Why not just…?
Why not just regress out the top principal components? This is the real threat, because in many datasets composition is the dominant variance and principal components capture it well. Principal components are not constrained to be mixtures. When a genuine program, such as a stress response, carries variance comparable to composition, a component can absorb it and the correction erases it. The planted-program experiment settles this directly. If principal-component removal matches our method there, we will report that the simplex structure adds nothing and reduce the contribution to the benchmark and the composition-share diagnostic.
Why not just apply SNAIL? SNAIL corrects ties among zero counts during normalization (SNAIL). Samples that differ in true cell proportions remain correlated after SNAIL, because the confound is biological, not a normalization artifact. The uncorrected baseline is already SNAIL-normalized, so this comparison is built into the design.
Why not just use DeClust or borrow marker genes from a related species? DeClust fixes three compartments and needs about 200 samples, while non-model studies have around 24. We test whether it degrades at small $n$ rather than assuming it does.
Borrowed markers are the more serious alternative. A reviewer could argue that orthologs of known markers are enough to estimate proportions. We simulate this case by deriving markers from an independent atlas cohort, which mimics imperfect transfer. If borrowed markers do as well, the practical recommendation becomes "use them when available," and our method applies only where no related atlas exists.
Why not use graphical lasso or partial correlation instead of WGCNA? Partial correlation removes indirect links between genes. It does not remove a shared hidden driver like composition unless that driver is modeled. It is also unstable when genes far outnumber samples, as with 24 samples. We include it as a secondary comparison on the simulated arm.
Proposed Experiments
Primary benchmark. We use the breast-cancer atlas framework (100,064 cells, 26 patients) for paired pseudo-bulks at sample sizes 12, 24, 50 and 200. We score:
- precision–recall area for the fixed-arm top-edge set;
- adjusted Rand agreement with fixed-arm modules;
- retention of planted programs.
Transfer to a second tissue. We repeat the benchmark with GTEx single-cell data (DeepCSCN's 27-cell-type atlas) to check that results transfer across tissues. We expect performance to drop where lineage-related types merge, and we report how much.
Real-data check. We correct the 120-sample TCGA glioblastoma bulk network and ask whether its overlap with the single-cell network, and its enrichment for physically interacting protein pairs, rises above the published under-10% overlap (Wang et al.). Failure here would mean that removing composition alone does not bring bulk networks closer to within-cell biology.
Non-model application. We reanalyze the Peucedanum data (24 samples, SRA PRJNA714368, unigenes collapsed and expression-filtered) (Song et al.) and make a falsifiable prediction before seeing the results. The lignin and cell-death module should have a high composition share, consistent with xylem expansion. The coumarin-biosynthesis module should keep most of its structure if it reflects regulation. If both collapse, or both persist unchanged, the diagnostic is uninformative in this setting.
We exclude the Alpinia six-tissue design from network inference because six samples cannot support it.
References
- Single-Cell Co-expression Analysis Reveals Distinct Functional Modules, Co-regulation Mechanisms and Clinical Outcomes
- De novo Transcriptome Sequencing Coupled With Co-expression Analysis Reveal the Transcriptional Regulation of Key Genes Involved in the Formation of Active Ingredients in Peucedanum praeruptorum Dunn Under Bolting Period
- A chromosome-level genome assembly of Alpinia officinarum Hance sheds new light on its evolution and flavonoid biosynthesis
- RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons
- Trichome-Specific Analysis and Weighted Gene Co-Expression Correlation Network Analysis (WGCNA) Reveal Potential Regulation Mechanism of Artemisinin Biosynthesis in Artemisia annua
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Deep Learning Driven Cell-Type-Specific Embedding for Inference of Single-Cell Co-expression Networks
- Robust Multi-view Co-expression Network Inference
- Adjustment of spurious correlations in co-expression measurements from RNA-Sequencing data
- A reference profile-free deconvolution method to infer cancer cell-intrinsic subtypes and tumor-type-specific stromal profiles
- Statistical inference for cell type deconvolution
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures