Biologists often pick "hub genes", the most connected nodes in a network of genes, as their best guesses for disease drivers and drug targets. Those networks are usually built from messenger RNA (mRNA) measured in ground-up tissue. That practice rests on two quiet assumptions. The first is that mRNA levels stand in for protein levels, yet by one widely cited estimate the two agree only 40–60% of the time. The second is that correlations measured across mixed tissue reflect regulation inside cells, rather than differences in which cell types happen to be present. This proposal tests both assumptions directly. It uses public single-cell datasets that measure RNA and surface protein in the same cell, where a protein-level hub ranking can be computed and treated as ground truth.
The core claim is that hub rankings from RNA mix three separate effects: cell-type composition, co-regulation within a cell type, and genuine RNA-to-protein disagreement. Composition is the one people mostly forget to remove, and it can be removed cheaply. We will measure how much each effect contributes to the gap between RNA hubs and protein hubs. We will then build a hub score that pools networks computed within each cell type and reports, for every gene, a calibrated probability that its hub status holds at the protein level. If the result holds, the practical message is short: condition on cell type first, then trust only the RNA hubs whose RNA-to-protein coupling you can vouch for.
Core Claim and Mechanism
The mechanism behind composition-driven hubs is the law of total covariance. Suppose a sample mixes cell types $k$ in proportions $\pi_k$. Each type has mean expression vector $\mu_k$ and within-type covariance $\Sigma_k$, and the mixture mean is $\mu = \sum_k \pi_k \mu_k$. The covariance a mixed-tissue network sees is then
$$ \Sigma_{\text{mix}} = \sum_k \pi_k \Sigma_k + \sum_k \pi_k (\mu_k - \mu)(\mu_k - \mu)^\top . $$
The second term is non-zero whenever cell types differ in mean expression. It creates dense correlation blocks among each type's marker genes, and centrality measures then rank those markers as "hubs". Across bulk samples with varying $\pi$, this effect gets stronger, because proportion variance drives the correlations. The first term is the within-type co-regulation that hub analysis claims to find. Transcript-protein disagreement then acts on top of that first term.
We predict that most of the disagreement between mixed-tissue RNA hubs and within-type protein hubs disappears once the between-type term is removed. We also predict that the remainder concentrates in genes with weak within-type RNA-protein coupling, which can be estimated from paired data.
Position Relative to Prior Work
Three lines of work motivate the proposal by exposing a piece of the gap. First, A high-stringency blueprint of the human proteome establishes the 40–60% mRNA-protein agreement figure. Second, the ESCC mass cytometry study shows that transcript signatures miss functional immune states after therapy. Third, SubNetDL finds that once subclonal structure is modeled, the genes that predict therapy response are inversely related to classical centrality. None of these asks whether a given hub ranking transfers from RNA to protein.
The hub-finding methods we compare against are scGIR and the pipeline in the morusin aortic dissection study. scGIR ranks genes by PageRank, a random-walk importance score, on single-cell RNA correlation networks. The morusin study ranks genes on a static interaction database and then checks where they are expressed in single-cell RNA. Both stay entirely in RNA; we score them against protein.
The prize-collecting Steiner approaches in the EGFRvIII signaling map and patient-specific glioblastoma pathways find hidden protein hubs such as MEK1, but only in bulk tissue or cell lines. They show that protein-level hubs matter; our composition conditioning addresses the blind spot they leave, which is that bulk tissue mixes cell types.
The methods we reuse as components or baselines are totalVI and scChord. totalVI is a variational autoencoder that jointly models paired RNA and protein counts and separates ambient antibody background from true signal. scChord translates RNA to protein by flow matching; it captures the shape of single-protein distributions, but reportedly struggles to preserve correlations between proteins (the Correlation Matrix Distance metric) once panels reach hundreds of proteins. Hub rankings are built from exactly those correlations, so this is a warning about imputation, not a side note. For composition estimates on bulk data we use MEAD, a statistical deconvolution method that gives valid confidence intervals on cell proportions.
The idea is also constrained by the classical view in Network Centrality for Identifying Regulatory Hub Genes, which holds that high-centrality nodes are the regulators. We do not assume that view is right or wrong. We measure which centrality survives the move to protein.
Research Program
The ground truth comes from paired CITE-seq data, which measures RNA and a panel of antibody-tagged surface proteins in the same cell. We use GSE164378, which has more than 161,000 human blood immune cells, 228 proteins and multiple donors. We use SLN-all, which has about 33,000 mouse spleen and lymph node cells measured with two panels of 111 and 208 proteins. GSE100866 (13 proteins) serves only as a sanity check.
For each annotated cell type we build networks restricted to genes whose protein product is on the panel. Protein values are normalized after background correction using isotype controls and totalVI's background component. We deliberately do not use totalVI's latent-space denoising for the ground truth, because a shared latent space would itself manufacture correlations.
Edges are defined in two ways, and every result is reported for both. The first is correlation-thresholded co-expression. The second is STRING interaction edges weighted by within-type co-expression. The hub score is PageRank with degree and betweenness as secondary scores. "Protein hub rank within cell type $k$" is the target. RNA hub ranks are computed from the same cells, from pseudo-bulk mixtures, and from mixtures of donors with controlled compositions.
The first deliverable is a variance decomposition of hub-rank disagreement. We compare three RNA rankings against protein ranks within each cell type. The first is from mixed tissue. The second removes the between-type term by pooling within-type covariances weighted by $\pi_k$. The third is within-type RNA. This attributes the loss in rank agreement to composition versus RNA-protein disagreement.
The second deliverable is the proposed method: a composition-conditioned, calibrated hub score. It pools within-type networks and attaches to each gene a transfer probability, meaning the chance that its RNA hub status holds in protein. That probability comes from a small model trained on paired data. Its features are the within-type RNA-protein correlation, detection rate, cell-type specificity, and variance in the RNA data. The calibration model is trained on one subset of proteins and evaluated on held-out proteins. The SLN-all dual panel allows a split across panels, and GSE164378 allows donor-held-out splits.
The third deliverable is an external bulk check. We apply MEAD-estimated proportions to the Xiangya ovarian cancer bulk proteomes from the integrated proteomics and scRNA-seq study (154 tumors and 29 normal tissues, with more than 40,000 reference cells from 8 tumors). The question is whether composition conditioning changes which proteins rank as hubs there, and whether the conditioned hubs align better with the subtype-specific targets that study reported, CD40/CD40L and TYMP. This depends on data access; it is a check, not the core result.
One scope limit applies throughout. Panel proteins are mostly surface markers, not intracellular signaling hubs. Our claims therefore concern hub transfer among measured proteins. Whether the calibration model extrapolates to intracellular proteins is framed as an open question, not assumed.
Baselines
The naive reference points are two rankings that use no cell-type information. The first is degree centrality on the static STRING graph restricted to panel genes, as in the morusin pipeline. It uses no expression data and so measures pure literature bias. The second is PageRank on a pseudo-bulk RNA co-expression network built across donors, the standard WGCNA-style practice (WGCNA builds gene co-expression networks). Both are scored with the same hub definition and the same edge thresholds, tuned on a validation donor.
The strongest existing RNA-only method is scGIR-style single-cell PageRank, computed within each annotated cell type. Its authors report an ARI of 0.68 on Chu-time clustering (ARI, the adjusted Rand index, scores cluster agreement with known labels), but that is a clustering result. We found no reported result on Lacuna for agreement between RNA-derived and protein-derived hub rankings, for scGIR or any other method. This proposal therefore defines that number.
The strongest imputation competitor computes protein hubs from imputed protein. We use totalVI and, if code is available, scChord, both trained on the same training cells and hyperparameter budget. scChord's reported weakness on protein-protein correlations makes this a sharp test rather than a strawman.
Two ablations isolate the key ingredients. The first removes composition conditioning but keeps calibration. The second removes calibration and keeps only composition conditioning.
A further baseline is regressing out MEAD-estimated proportions from pseudo-bulk expression before building the network. It is the cheapest fix and is the main threat to the method.
Beating a baseline means a higher Spearman correlation with protein hub ranks within cell type, a higher overlap with the protein top-10% hubs, and better calibration of the transfer probability on held-out proteins. Each comparison uses donor-level bootstrap intervals that exclude zero difference.
Why not just…?
Why not just run scGIR within each cell type? That already removes the between-type term, so it may capture most of the gain. If it does, the decomposition still has value: it would show that composition, not RNA-protein disagreement, was the main problem, which is a finding. The second ablation compares within-type RNA PageRank against the calibrated score, specifically on genes with low RNA-protein coupling. If calibration adds nothing there, we will drop it and report composition conditioning alone.
Why not just impute protein with totalVI and rank hubs on the imputed values? Imputation models reconstruct protein through a shared latent space. That can inflate correlations and shrink them toward cell-type means, recreating the between-type term under another name. scChord's own report of weaker correlation preservation at 228 proteins suggests hub ranks will suffer. The experiment settles this directly: if imputed-protein hubs match measured-protein hubs as well as our score does, imputation is the simpler answer and we will say so.
Why not just regress out estimated cell proportions? For bulk data this is a real threat. Regressing out proportions removes proportion-driven covariance across samples. It does not remove the effect of cell types differing in their within-type covariances $\Sigma_k$, and it does not address RNA-protein disagreement. We test it head to head on pseudo-bulk mixtures, where the true proportions are known. If it matches the pooled within-type score, the recommended practice becomes proportion regression followed by calibration.
Why not just use the protein panel and skip RNA? Where a panel exists, one should. But panels cover roughly 10 to 200 proteins, while RNA covers the genome. The practical question is which of the unpanelled RNA hubs to trust, and only the held-out-protein calibration answers it.
Proposed Experiments and Decisive Checks
The first decisive check is the decomposition itself. On GSE164378 and SLN-all, the prediction is that the drop in rank agreement between within-type protein hubs and mixed-tissue RNA hubs mostly disappears after composition conditioning. If conditioning recovers less than half of the gap, composition is not the main driver and the paper's emphasis shifts to RNA-protein disagreement.
The second check is transfer prediction on held-out proteins across the SLN-all panel split and across donors. The calibration model must beat a single-feature rule that thresholds per-gene within-type RNA-protein correlation, measured by area under the ROC curve (AUROC) and expected calibration error. If it does not, we report the simple rule as the method.
The third check reruns everything with both edge definitions and three centrality scores, and keeps only conclusions that hold under all of them. Hub findings are notoriously sensitive to these choices, so this guards against reporting a result that belongs to one choice.
The fourth check is external. On Xiangya bulk proteomes, we ask whether conditioned hubs rank CD40, TYMP and related subtype markers differently from unconditioned hubs. This is interpreted cautiously, as consistency evidence rather than proof.
References
- A high-stringency blueprint of the human proteome
- High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- scGIR: deciphering cellular heterogeneity via gene ranking in single-cell weighted gene correlation networks
- Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection
- Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling
- Network Modeling Identifies Patient-specific Pathways in Glioblastoma
- Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- Statistical inference for cell type deconvolution
- Lethality and centrality in protein networks
- Integrated proteomics and scRNA-seq analyses of ovarian cancer reveal molecular subtype-associated cell landscapes and immunotherapy targets