When biologists want to know which genes work together in a disease, they often use weighted gene co-expression network analysis (WGCNA). The method finds groups of genes, called modules, whose expression levels rise and fall together across patient samples. Most of the available patient data, however, comes from bulk tissue: each measurement averages thousands of cells of many types. In bulk tissue a module can appear simply because one cell type is more abundant in some samples. If macrophages infiltrate a tumor, every macrophage gene rises together, whether or not any gene is regulating any other. The module then reflects the tissue's cell mix, not a regulatory program inside a cell. Deconvolution methods, which estimate the cell mix of a bulk sample, recover cell fractions or per-cell-type expression. They stop before asking which genes co-vary inside which cell type. Separating these two sources of co-variation from bulk data is underdetermined without extra information.
We propose that the extra information comes from chromatin accessibility, meaning which stretches of DNA are unwrapped and open to regulatory proteins in each cell type. Two genes can only be co-regulated inside cell type $k$ if their regulatory DNA is open in $k$ and carries binding sites for a regulator that $k$ actually expresses. Single-cell multiome references record this "open and partnered" state for each cell type. We use it as a gate that limits which cell types each gene pair's bulk covariance may be assigned to. The result is a set of cell-type-specific co-expression networks, with WGCNA-style modules, estimated from ordinary bulk RNA cohorts. It also yields a per-module estimate of how much of the module is cell mix and how much is within-cell regulation.
Core Claim
Write $y_s \in \mathbb{R}^G$ for the bulk expression of $G$ genes in sample $s$, and $p_{sk}$ for the fraction of cell type $k$ in that sample. Let $x_{sk}$ be cell type $k$'s mean expression in sample $s$, with population mean $\mu_k$ and covariance $\Sigma_k$ across samples. Assume the $x_{sk}$ are independent across cell types. Then the residual after removing the fraction-predicted mean has conditional covariance
$$\operatorname{Cov}\big(y_s - \textstyle\sum_k p_{sk}\mu_k \,\big|\, p_s\big) = \sum_k p_{sk}^2\, \Sigma_k .$$
A plain bulk correlation mixes this term with the composition term $\sum_{k,l}\operatorname{Cov}(p_k,p_l)\,\mu_k\mu_l^\top$. That composition term is what produces abundance-driven modules.
The within-cell covariances $\Sigma_k$ can be estimated by regressing residual outer products on the squared fractions. The problem is that each gene pair then has $K$ unknowns, and the squared fractions are usually strongly collinear. Our gate $\Gamma_k(i,j)\in\{0,1\}$ forces $\Sigma_k[i,j]=0$ unless three conditions hold in cell type $k$ of a single-cell multiome reference: both genes are detected, their linked regulatory elements are accessible, and those elements carry a motif for a transcription factor expressed in $k$. Once the gate is applied, most pairs have only one or two admissible cell types, so the regression becomes well posed.
On each gated $\hat\Sigma_k$ we run the usual WGCNA steps: soft-threshold power, topological overlap, hierarchical clustering. For every original bulk module we also report the share of its eigengene variance explained by the composition term versus each $\hat\Sigma_k$.
The claim is empirical and falsifiable. The partner-TF gate should recover planted and natural within-cell-type co-expression from bulk mixtures better than three alternatives: an ungated moment estimator, an expression-only gate, and per-cell-type expression profiles from deconvolution fed into WGCNA.
We were unable to check whether ungated moment estimators of cell-type-specific covariance already exist outside the indexed material. Reviewers should assume they might. The contribution we defend is the chromatin gate and the evaluation that isolates it, not the moment identity.
Where the Gate Comes From, and What Could Break It
The biology behind the gate comes from the human retina single-cell multiome study. That study profiled RNA and accessibility jointly in 20 donors and found 87–92% of expression QTLs to be cell-type specific. Accessibility at the linked regulatory elements, by contrast, was broadly shared across lineages. The authors' explanation is a hierarchy: broad pioneer factors open the chromatin, and cell-type-specific secondary factors bind those opened sites to drive transcription.
This finding both motivates and constrains our design. An accessibility-only gate would be weak, because open chromatin is shared too widely to say which cell type owns a co-expression edge. That is why the gate requires the secondary-factor motif together with the expression of that factor in $k$. The retina study tested gene–peak pairs one at a time and built no networks. We carry its hierarchy into a pairwise, network-level constraint.
The BMI-discordant twin preadipocyte study shows that co-accessibility across large active compartments tracks expression changes, with 21 of 52 differentially expressed genes correlated with local peak accessibility. It used purified cells and therefore avoided cell-mix masking by design. Work on dynamic nucleosome displacement by p53 (Nili et al.) warns that a static accessibility snapshot can miss transiently opened sites. This is the gate's main failure mode: genes that are truly co-regulated in a cell type but gated out. We measure this cost directly rather than assume it away.
On the cell-mix side, deconvolution methods supply the fractions $p_{sk}$ but not networks. TAPE uses an autoencoder whose decoder is a signature matrix and outputs per-sample, per-cell-type expression. Count Bridges samples single-cell count profiles constrained to sum to the bulk. POPPY aligns bulk and single-cell embeddings by cell ontology. We reuse their fraction estimates as inputs.
Single-cell co-expression methods show what the target networks look like but require single-cell input. DeepCSCN builds cell-type-specific networks from learned gene embeddings. NEBULA fits mixed models and found that APOE co-expresses with CLU in astrocytes but with TREM2 and C1q in microglia. That split is exactly the kind of structure bulk WGCNA blends into one module. NEBULA also warns that subject-level inference inflates false positives below about 30 subjects, which constrains how we build ground truth from 20 retina donors.
Research Program
The first step is to build the gate. From a multiome reference, we link each gene to its regulatory elements using peak–gene correlation within cell types. We call motifs, then set $\Gamma_k(i,j)=1$ when both genes have an accessible linked element in $k$ and those elements share a motif for a factor expressed in $k$. A relaxed variant requires only accessibility. A control gate uses random masks with the same per-type density.
The second step is the estimator. We obtain fractions from TAPE, and alternatively from true fractions in simulation to separate estimation error from fraction error. We form residuals $r_s = y_s - \hat M p_s$. For each gene pair we regress $r_{si}r_{sj}$ on $\{p_{sk}^2 : \Gamma_k(i,j)=1\}$ with a small ridge penalty. We group pairs by gate pattern so that each distinct design matrix is factored only once. This keeps a 5,000-gene, 10-type problem to minutes on a CPU.
The third step converts each $\hat\Sigma_k$ to a correlation matrix, then applies the WGCNA soft threshold, topological overlap, and modules. We also compute the composition-versus-within-cell variance split for every standard bulk module.
The fourth analysis asks which gate is best. It compares the three gate variants, the ungated estimator, and gate misspecification, as a function of sample size $n$ and of how collinear the squared fractions are.
Baselines
The naive reference point is standard WGCNA on bulk expression (Shen et al.). It is what almost every bulk co-expression study in this area runs, for example Lian et al.. We run it with the usual power selection, on the same samples and gene set as every other method. Beating it means two things. First, fewer modules whose eigengene correlates with an estimated cell fraction above 0.7. Second, higher recovery of ground-truth within-cell-type edges.
The strongest simple competitor is fraction-residualized WGCNA. Here each gene is regressed on the estimated fractions and WGCNA is run on the residuals. This removes the composition term but leaves a single network with no attribution to cell types. It is fair to compare it only on the de-confounding metric. It cannot score on per-type edge recovery at all, so on that axis we compare against it after it assigns each module to the cell type with the highest marker overlap.
The strongest existing method family is deconvolution-then-network. TAPE produces per-sample, per-cell-type expression matrices, and we run WGCNA on each one. TAPE reports lower mean absolute error than CIBERSORTx, Scaden and MuSiC, but the file gives no numeric value for its error. For fraction accuracy the only reported figure is from DECONbench, where RNA deconvolution by r_WNM reached a mean absolute error of 0.024 on in silico pancreatic tumor mixtures. We found no reported result on Lacuna for recovering cell-type-specific co-expression networks from bulk data. The comparison is therefore new, and we will not claim a margin over a published number.
TAPE and our estimator share the same fractions, reference, gene set and tuning budget: a grid over the ridge penalty and soft-threshold power chosen on held-out donors. Beating TAPE-then-WGCNA means a higher area under the precision–recall curve for per-type edges, and higher adjusted Rand index for per-type modules, averaged over held-out-donor folds.
Two ablations isolate the key ingredient. The ungated moment estimator tests whether the gate matters at all. The expression-only gate admits type $k$ whenever both genes are detected there, and tests whether chromatin adds anything beyond expression. A single-cell upper bound, DeepCSCN or plain per-type pseudobulk correlation on the true single-cell data, shows how much of the achievable signal remains.
Why not just…?
Why not just regress out cell fractions before WGCNA? This handles the composition term, and on the de-confounding metric alone it may perform close to our method. We say so openly. It cannot tell whether a residual module lives in astrocytes or microglia. That is the question raised by the Efnb2 parvalbumin-neuron study, which could not tell cell-autonomous rewiring from glial compensation, and by the NEBULA APOE split. If residualized WGCNA with post hoc marker assignment matches our per-type edge recovery, the attribution step adds nothing and we will report that.
Why not just deconvolve with TAPE or Count Bridges and run WGCNA on each cell type's profiles? This is the most natural pipeline and a real threat. Our concern is that per-sample cell-type profiles produced by a shared decoder may inherit bulk correlations, but the file contains no evidence either way. The main semi-synthetic experiment settles it. If TAPE-then-WGCNA recovers planted within-type modules as well as the gated estimator, the proposal reduces to recommending that pipeline.
Why not gate on expression alone, since the reference already says which genes each cell type expresses? The retina study's finding that accessibility is broadly shared makes this a serious alternative. A chromatin gate built on accessibility alone may be no more selective than expression. The partner-TF requirement is our bet that the hierarchy adds information. If the expression-only gate matches it, chromatin is not the useful ingredient, and we will narrow the claim to expression-masked decomposition.
Why not just profile samples with single-cell multiome directly? For new studies with budget, that is better. The retina study reached only 20 donors, and banked clinical cohorts are bulk. We treat direct single-cell networks as the upper bound, not as a competitor.
Proposed Experiments and Decisive Checks
The main benchmark is semi-synthetic and built from the retina multiome. We hold out donors for the reference in each fold: the gate is built on 10 donors, and bulk mixtures are built from the other 10. For each mixture we pick a held-out donor, draw fractions from a Dirichlet distribution matched to observed retinal composition (as DECONbench does for tumors), and sum sampled nuclei. We vary $n$ from 50 to 500 mixtures.
Ground truth comes in two forms. The first is planted modules: gene sets scaled by a sample-specific factor inside one cell type. Half are chosen to satisfy the gate and half to violate it, so we measure both the gain from the gate and its cost when wrong. The second is natural donor-level co-variation, measured from per-type pseudobulk in the held-out donors. Because only 10 donors carry the natural signal, and NEBULA flags inflated error below 30 subjects, we report natural-edge results with donor-level bootstrap intervals and weight the planted benchmark more heavily. We repeat the construction on the joint ATAC/RNA datasets in DANCE to check that results are not specific to retina.
Metrics are three. The first is per-type edge precision–recall against ground truth. The second is module adjusted Rand index. The third is the composition artifact rate: the fraction of reported modules whose eigengene correlates with a fraction above 0.7. We measure all of them under both true and TAPE-estimated fractions.
The decisive check is a two-way comparison on gate-violating planted modules, which isolates the chromatin contribution.
On the naturally varying edges, the partner-TF gate must beat the expression-only gate, and it must do so by a margin that exceeds the bootstrap interval. It should also lose less than the ungated estimator gains on gate-consistent modules. If it fails the first condition, chromatin adds no information. If it fails the second, the gate costs more than it helps.
A real-tissue check follows if a brain multiome reference can be obtained, since none is in the file. We would apply the method to a bulk Alzheimer's cohort such as the 532-patient set deconvolved by TAPE, and ask whether APOE–CLU is attributed to astrocytes and APOE–TREM2/C1q to microglia, as NEBULA found at single-cell resolution. Only the expression-only gate can run on this cohort without that external reference.
References
- Single-cell multiomics of the human retina reveals hierarchical transcription factor collaboration in mediating cell type-specific effects of genetic variants on gene regulation
- Increased body mass index is linked to systemic inflammation through altered chromatin co-accessibility in human preadipocytes
- p53 binds preferentially to genomic regions with high DNA-encoded nucleosome occupancy
- 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
- Deep Learning Driven Cell-Type-Specific Embedding for Inference of Single-Cell Co-expression Networks
- NEBULA is a fast negative binomial mixed model for differential or co-expression analysis of large-scale multi-subject single-cell data
- Session Introduction
- The co-expression networks of differentially expressed RBPs with TFs and LncRNAs related to clinical TNM stages of cancers
- DECONbench: a benchmarking platform dedicated to deconvolution methods for tumor heterogeneity quantification
- RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons
- DANCE: a deep learning library and benchmark platform for single-cell analysis