Do Co-Expression Hubs Survive Inside the Cell? A Matched RNA–Protein Tissue Test

A common workflow in cancer biology runs as follows. Researchers measure the RNA of every gene in many ground-up tumor samples. They group genes whose levels rise and fall together into modules, pick the most connected gene in each module (the "hub"), and then "validate" the hub by showing its protein changes in the same direction on a western blot of ground-up tissue. We argue that this validation is often circular. Most of the patient-to-patient variation in a tumor lump comes from how much of it is cancer cells, immune cells, or connective tissue. Two genes made mainly by immune cells will look co-expressed in the RNA simply because some tumors contain more immune cells. Their proteins will agree for the same reason on a blot of the same ground-up tissue. Neither measurement shows that the genes are co-regulated inside any cell.

We propose to measure, rather than assume, how much of a transcriptome-wide co-expression network survives when protein is quantified inside defined tissue compartments of the same tumors. The design uses matched RNA sequencing, western blots, and whole-slide multiplexed immunostaining on one lung adenocarcinoma cohort. Western blots serve as a gate on antibody quality, not as the validation. The immunostaining provides within-compartment protein measurements that cannot be driven by tissue composition. The output is a benchmark that settles which module-building choices produce modules that are real at the protein level: standard networks, networks corrected for cell composition, nonlinear knowledge-guided clustering, or noise-robust inference. It also gives a direct estimate of how often the standard "hub plus blot" validation certifies a composition artifact.

Background and the gap

Weighted gene co-expression network analysis (WGCNA) turns gene–gene correlations into a network. It raises absolute correlations to a power so weak links fade smoothly, clusters genes by shared neighbors, and summarizes each module by its first principal component, the "eigengene," which can then be correlated with clinical traits (Li Shen et al.). The method is used widely to link modules to stage (Lian et al.), survival (Weng et al.), and diagnostic hubs. The closest existing work is the CENPL/LSM4 breast cancer study. It combined eight microarray cohorts, WGCNA, and interaction networks to nominate four hubs with diagnostic AUCs up to 0.948 for LSM4. It then confirmed them only by RNA in situ detection in cultured cell lines. Cell lines contain one cell type, so that check cannot distinguish co-regulation from composition, and it never measured protein. Our proposal differs in its target: we do not nominate new hubs. We measure whether a module's co-variation holds for protein within tumor, stromal, and immune compartments of patient tissue.

Two methodological papers motivate the benchmark and supply candidate methods. KISL replaces Pearson correlation with distance correlation, adds Gene Ontology constraints, and uses semi-supervised k-means. It reports beating WGCNA on silhouette, Calinski–Harabasz, and Davies–Bouldin indices across eight TCGA cancers, and WGCNA gave negative silhouettes in lung adenocarcinoma. Robust multi-view co-expression inference models heavy-tailed noise and batch effects with multivariate t-distributions. Both evaluate only internal clustering geometry or synthetic recovery. A cleaner cluster is not the same as a biologically real one, and composition variation produces very clean clusters. We supply the missing external test.

On the protein side, Kurowski et al. showed that scoring ALDH1A1 separately in tumor and stromal cells gives distinct prognostic hazard ratios (1.95 and 2.23 for low expression). This is direct evidence that compartment matters and that a lysate average would mix the two signals. Blom et al. supply our imaging method. They built an 8-plex fluorescence-plus-chromogenic assay on paraffin sections with whole-slide single-cell segmentation, quantifying 128,894 immune cells and finding a within-cancer-cell AR–AMACR correlation of r = −0.26. That work used a fixed marker panel and no transcriptome. We point the same machinery at protein panels chosen by rule from RNA modules.

Western blotting constrains the design rather than defining it. It resolves proteins by size, so it can reveal off-target bands, cleaved forms, or immature isoforms that staining cannot distinguish (Zhan et al.). It also loses all spatial information. We use it where it is strong: confirming each antibody gives a single band at the expected size and a linear signal across a dilution series. We also use blot correlations on tumor lysates as an explicit measure of the composition-confounded signal. The broader point that RNA-derived candidates are hypotheses needing physical tests (functional validation requirements) motivates the whole study. The mRNA–protein gap documented by Yang et al. and Liao et al. is the second reason RNA co-expression may fail at the protein level, and we separate it from the composition effect.

Core Claim

Our claim is that, in bulk tumor RNA-seq, a substantial fraction of co-expression modules are driven by variation in cell-type proportions. For these modules, conventional validation on tissue lysates will succeed while within-compartment protein co-variation will be absent. Modules not tied to composition, and modules produced by methods that remove composition before clustering, will show within-compartment protein co-variation well above matched random gene pairs.

The central quantity is the within-compartment protein preservation of a module edge. For genes $i$ and $j$ in a module and compartment $c$ (tumor cells, stroma, or immune cells), let $\rho^{\text{prot}}_{ij,c}$ be the correlation, across patients, of the mean per-cell staining intensity of the two proteins within compartment $c$. A module's preservation score is the average of $\rho^{\text{prot}}_{ij,c}$ over its measured edges, in the compartment where the genes' RNA is predominantly expressed. That compartment is assigned beforehand from public single-cell expression annotations. We compute a second, stricter score within each patient, across individual cells of one compartment. It tests whether the two proteins co-vary cell by cell. This is the closest a tissue measurement comes to intracellular co-regulation.

Research Program

Module discovery uses TCGA lung adenocarcinoma RNA-seq. We chose this cancer because KISL reports WGCNA performs worst there, so method differences should be largest. We build modules four ways. The first is standard WGCNA. The second is WGCNA after regressing RNA-estimated tumor purity and immune and stromal fractions out of every gene. The third is KISL. The fourth is the robust multi-view method applied across TCGA and our cohort as two views. For each WGCNA module we record, before any protein work, how strongly its eigengene correlates with the estimated composition fractions. We label modules "composition-linked" or "composition-independent" by a preregistered threshold.

The wet-lab cohort is about 120 resected lung adenocarcinomas with matched fresh-frozen tissue (for RNA-seq and lysates) and paraffin blocks (for tissue microarrays and whole sections). The size comes from a power calculation. Detecting an across-patient correlation of 0.25 at 80% power and two-sided α = 0.05 needs roughly $n \approx \left((1.96+0.84)/\operatorname{atanh}(0.25)\right)^2 + 3 \approx 123$, from the Fisher z-approximation. We sequence all tumors and check that the discovered modules are preserved at the RNA level in this cohort, so that any protein-level failure is not a failure of replication.

Target selection follows a preregistered rule so antibodies cannot be cherry-picked. We take eight modules, four composition-linked and four composition-independent, spanning the methods. From each we take the hub and two members ranked by module membership, and we require a commercially available antibody. Every antibody must pass a western blot gate before use: a single dominant band at the expected molecular weight in a positive cell line and in pooled tumor lysate, reduced signal in a knockdown or negative line where available, and linear densitometry over a fourfold dilution series. Antibodies that fail are replaced by the next-ranked member, and all replacements are logged.

Imaging uses three serial-section multiplex panels in the style of Blom et al. Each panel carries the same compartment markers (pan-cytokeratin for tumor epithelium, CD45 for immune cells, a mesenchymal marker for stroma) plus up to five module proteins. Whole slides are segmented into single cells assigned to compartments, and per-cell nuclear and cytoplasmic intensities are extracted. Whole-slide analysis, rather than hotspot sampling, avoids the sampling variance that plagues regional scoring. Tumor lysates from the same patients are blotted for the same proteins, giving a lysate-level protein correlation for every edge.

Baselines

The naive reference point is a null distribution of protein pairs matched to the module edges on RNA expression level and compartment of origin but drawn from different modules. Every module method must beat it. The trivial prediction that every RNA edge carries over to protein corresponds to a preservation score equal to the RNA correlation, and we report the gap from it. No reported result was found on Lacuna for protein-level preservation of any co-expression method, so the comparison has no external number. The only reported comparison among the methods is KISL's qualitative superiority over WGCNA on clustering indices, with no values given in the file.

The strongest existing competitor is the standard pipeline itself: WGCNA hubs confirmed by lysate western blot, as in the family of studies represented by the CENPL/LSM4 study. It is the strongest because it is what the field accepts as validation, and it will look successful. We run it on identical patients and antibodies. The comparison is not whether it "passes." It is whether its pass rate differs between composition-linked and composition-independent modules, and whether passing predicts within-compartment preservation.

Composition-adjusted WGCNA is the serious methodological competitor, because it attacks the confound directly with no new data. KISL and the robust multi-view method are run with their published settings on the same TCGA matrix. Module size is equalized across methods by subsampling members, and each method receives the same antibody budget per module. A method beats standard WGCNA if its modules' mean within-compartment preservation is higher, with a patient-level bootstrap interval excluding zero.

Two ablations isolate our key ingredient, compartment resolution. The first computes the same imaging correlations from whole-slide mean intensity, ignoring compartments. The second uses lysate blot correlations. If compartment resolution matters, both should track RNA correlation for composition-linked modules while the within-compartment score does not.

Why not just…?

Why not just regress out estimated cell fractions and trust the resulting modules? This is the real threat, and we do not argue around it. We include it as one of the four module methods under test. RNA-based composition estimates are themselves inferred from marker genes, and they cannot correct the mRNA–protein gap. Whether adjustment is sufficient is an empirical question that only an external protein measurement can answer. If adjusted modules reach the same within-compartment preservation as composition-independent modules, that is a useful result: the field could then drop the expensive imaging step and adopt adjustment by default.

Why not just use single-cell or spatial RNA data to test co-regulation? These would address composition but not the second failure mode, that co-regulated transcripts need not yield co-varying proteins because of translation and degradation (Yang et al.). The within-patient, across-cell protein score measures both at once. The paired RNA and compartment protein data also let us estimate how much of any preservation loss comes from each source.

Why not just mine the Human Protein Atlas instead of staining a new cohort? Its images show one antibody per sample on unmatched tissues, so they cannot give protein co-variation across patients who also have RNA-seq. Automated reading of its subcellular patterns also still trails experts (macro F1 about 0.59 versus about 0.71, HPA competition analysis). Matched multiplexed sections are the only way to put two proteins in the same cells of the same patient.

Why not just skip the network and blot the top differentially expressed genes? Single-gene rankings miss connector genes with weak differential expression. scPPIN found EGFR ranked 1800th by p-value despite acting as a hub. More to the point, differential expression tests cannot pose our question, which concerns co-variation between genes.

Decisive Checks

The primary test is a preregistered interaction. We predict that composition-linked modules show lysate-blot correlations comparable to their RNA correlations but within-compartment protein preservation indistinguishable from the matched null. We predict that composition-independent modules exceed the null in both. If composition-linked modules also preserve within compartments, our central claim is wrong: bulk hubs are usually real, and the standard pipeline is vindicated. We will report that outcome as prominently as the alternative.

The secondary test ranks the four module methods by within-compartment preservation and asks whether this ranking agrees with their ranking by silhouette index. Disagreement would show that the clustering indices used by KISL do not measure biological validity.

A third check applies the stricter across-cell score within patients to the modules that pass. It separates co-variation between patients from co-regulation inside individual cells.

We will release the matched RNA, blot, and single-cell imaging tables as a benchmark, so future network methods can be scored against protein ground truth rather than cluster geometry alone.