Do Transcriptomic Hubs Survive Translation? Measuring Higher-Order Protein Structure in Paired Single-Cell Data

Biologists usually decide which genes "run" a disease or cell state by building a network from messenger RNA (mRNA) measurements and ranking genes by how connected they are. The top-ranked genes are called hub genes. Proteins, not transcripts, do the regulating, and mRNA tracks protein abundance only loosely: across human tissues the correlation holds only 40% to 60% of the time (A high-stringency blueprint of the human proteome). Paired single-cell assays such as CITE-seq measure both RNA and a panel of surface proteins in the same cell, so this question can now be answered directly. We propose to do that, borrowing a tool from neuroscience. Instead of ranking genes by raw correlations, we separate each marker's baseline abundance from its pure pairwise and three-way interactions with other markers, as Nakahara and Amari did for spike trains. We then compare the interaction structure, and the hubs it implies, between RNA and protein measured in the same cells.

The same measurement becomes an evaluation for a fast-growing class of models that predict a cell's proteins from its RNA. These models are scored on per-protein error and correlation. The strongest of them, scChord, reports that its fidelity to the protein–protein correlation matrix degrades as panels grow to hundreds of proteins. We ask a sharper question: when such a model generates proteins, does it reproduce the protein-level interaction network and its hubs, or does it hand back the transcriptomic network in protein coordinates? If it is the latter, then "predicted proteomes" cannot substitute for measured ones in hub discovery. We can say exactly which interactions are lost.

Core Claim

We claim two things. First, in paired CITE-seq data, hubs defined from pure (marginal-corrected, cell-state-conditioned) protein interactions differ measurably from hubs defined the same way on the matched transcripts. The gap is larger than the split-half noise ceiling of the protein measurement itself. Second, current RNA-to-protein translators recover per-protein values well but recover protein hubs and three-way protein interactions poorly. Their predicted networks sit closer to the RNA network than to the measured protein network. Three-way errors carry information beyond the correlation-matrix error scChord already reports: they predict failures on biologically standard tasks such as recovering the frequency of cell populations defined by joint three-marker gates.

The novelty is not in building another translator or another hypergraph network. It is a specific decomposition, a specific comparison (RNA vs. protein vs. predicted protein, in the same cells), and a test of whether higher-order structure matters beyond pairwise structure. If it does not, the result is still informative, and we say so.

Why this specific combination

The neuroscience direction supplies the method. Information-Geometric Decomposition in Spike Analysis writes a distribution over binary patterns in log-linear form. Its "natural" parameters $\theta_i, \theta_{ij}, \theta_{ijk}$ measure intrinsic interactions of increasing order. These parameters are orthogonal to the firing rates, so a pure triplet term $\theta_{ijk}$ can change without any change in pairwise correlations. This solves the isolation problem: telling a real three-way dependency apart from overlapping pairwise ones. That problem is exactly what afflicts hub scoring on co-expression graphs.

Surface proteins suit a binary treatment well. Immunologists already treat them as on/off through gating. totalVI, which fits a joint model of paired RNA and protein, gives each cell's protein count a probability of being real signal rather than antibody background, via a two-component mixture. That probability is a principled binarization rather than an arbitrary threshold. We reuse totalVI for this purpose only.

Two neuroscience results constrain the design. The primate retina study found pairwise terms explained roughly 99% of structure, a warning that higher-order terms may be negligible in small, simple populations. Bethge et al. found pairwise models fail on richer, higher-dimensional inputs, and Cayco-Gajic et al. showed that triplet structure among similarly tuned cells changes what a downstream readout can separate. We treat the retina result as a live null hypothesis, not a strawman.

The warning about shared latent input also carries over. Unobserved common drive can masquerade as a multi-unit interaction. We therefore condition every model on cell state, following the contextual-independence idea of probabilistic circuits for neural populations, in which units are independent in one context and coupled in another. In practice this means fitting within annotated cell types and including latent cell-state covariates.

The discordance direction supplies the objects of study. The three translators we evaluate are scChord, scTranslator and totalVI.

  • scChord smooths protein space with a zero-inflated autoencoder, then transports RNA embeddings into it with flow matching. It reports degradation on correlation matrix distance (CMD) at large panels and suggests, without implementing, "pathway-informed graph priors."
  • scTranslator is a Transformer pretrained on bulk and single-cell RNA–protein pairs. It extracts regulatory networks from attention weights but applies no centrality analysis and no higher-order decomposition.
  • totalVI imputes held-out proteins from a shared 20-dimensional latent space.

None of them evaluates network structure beyond pairwise correlation.

The hub direction supplies the question and the scoring. ESPACE improves partial-correlation network recovery by relaxing the penalty on edges of pre-specified hubs. It is strictly pairwise, transcript-only, and defines hubs by a degree threshold (degree > 7 and above the 0.95 quantile). HyperG-VAE models genes as nodes and cells as hyperedges. It beats pairwise inference in 91% of evaluated conditions on scRNA-seq benchmarks, but never sees protein data. Huang et al. link bulk proteomic and transcriptional changes through curated interaction databases. Our study differs from all three: hubs are defined on measured single-cell protein, compared with hubs on matched RNA from the same cells, and scored with a decomposition that separates pure interactions from marginal abundance.

Research Program

The first component is the estimator.

  • Unit. For each annotated cell type, take the markers that are expressed in at least a minimum fraction of cells. This is typically a few dozen out of a 228-protein panel.
  • Binarization. Binarize each protein with the totalVI foreground probability. Binarize the matched gene with a depth-aware detection model.
  • Model. Fit, separately for RNA and for protein, a log-linear model with singleton, pairwise and selected triplet terms. Covariates are library size and a low-dimensional cell-state embedding, so that shared drive is not counted as interaction. The model is

$$\log p(x \mid c) = \sum_i \theta_i(c)\, x_i + \sum_{i<j} \theta_{ij}\, x_i x_j + \sum_{i<j<k} \theta_{ijk}\, x_i x_j x_k - \psi(c).$$

Here $x \in \{0,1\}^n$ is a cell's binary marker vector, $c$ holds the covariates, and $\psi(c)$ is the normalizer. Only the singleton terms depend on covariates, so that the interaction terms are the covariate-adjusted pure interactions in the Nakahara–Amari sense.

  • Fitting. Exact fitting is impossible at this size. We use penalized neighborhood logistic regression (each marker regressed on the others plus pairwise products), which is a pseudo-likelihood approximation. Candidate triplets are pre-screened by support, and false discovery rate is controlled across triplets. On subsets of five to eight markers we fit the exact log-linear model, to verify that the pseudo-likelihood recovers the orthogonal decomposition.

The second component defines hub scores on the resulting structure.

  • A pairwise hub score is weighted degree, plus betweenness following the bridge-gene logic in the file, on the graph of significant $\theta_{ij}$.
  • A higher-order hub score counts participation in significant $\theta_{ijk}$ that remain after conditioning.
  • We report ranks and top-$k$ sets rather than fixed thresholds. This avoids the threshold dependence of ESPACE-style hub definitions.

The third component is evaluation of translators. We draw predicted protein profiles for held-out cells from each translator, pass them through the identical estimator, and compare their interaction parameters and hub rankings with those of measured protein. The comparison is against both the measured protein network and the measured RNA network.

The fourth component is conditional. If translators do lose protein hubs and triplets, we test the prior scChord proposes but leaves unimplemented: a penalty during scChord training on the discrepancy between the generated and training-set protein $\theta_{ij}$ and hub scores. This is a structural prior estimated from measured protein, not from curated pathways. We run it only if the diagnostic shows a gap worth closing.

Baselines

The naive reference point is the standard practice we are questioning. We build a hub list from the RNA of the same cells with a plain co-expression graph (WGCNA-style correlation with degree ranking) and take it as a stand-in for protein hubs. Every protein-level comparison is also reported against a noise ceiling: hub agreement between two random halves of the measured protein cells. "Beating" the naive approach means showing that RNA-derived hubs fall significantly below this ceiling. Without a ceiling, low RNA–protein agreement could be just measurement noise.

The strongest pairwise network baseline is ESPACE, alongside graphical lasso on continuous log-normalized protein and RNA counts. Both are run on exactly the cells, cell-type splits and covariates we use, with their own information-criterion tuning at a matched budget. This isolates the key ingredient: whether pure, marginal-corrected interactions change the hub ranking relative to partial correlations on continuous data. The ablation ladder runs from raw correlation, to partial correlation, to pairwise-only pure $\theta_{ij}$, to pairwise plus triplet $\theta$. Each step is scored by agreement with the protein split-half ceiling and by held-out log-likelihood of binary protein states.

The strongest translator baseline is scChord. The file states it achieved "the lowest Root Mean Square Error (RMSE) and highest Pearson Correlation Coefficient (PCC)" on GSE100866 and GSE164378, outperforming totalVI, sciPENN and scTranslator. No numeric RMSE, PCC or CMD value appears on Lacuna, so we will reproduce these metrics ourselves before measuring anything new. For scTranslator, the file reports a 10.7% cosine-similarity increase over sciPENN and cTP-net on gold-standard data, and similarity above 0.83 under unaligned gene sets (scTranslator). For totalVI's held-out protein imputation, no reported result was found on Lacuna. For higher-order scorers, HyperG-VAE is run on RNA as the strongest transcript-only multi-gene method. Its hub output is compared with measured protein hubs under the same top-$k$ scoring. No reported result was found on Lacuna for any method's protein-level hub recovery.

All translators are trained on the same GSE164378 donor split, with published hyperparameters and an equal tuning budget. They are scored on identical held-out donors. A translator "preserves structure" if its predicted-protein hub agreement with measured protein reaches the split-half ceiling within its confidence interval.

Why not just…?

Why not just use scChord's correlation matrix distance instead of triplet terms? CMD is the obvious and cheaper diagnostic, and scChord already reports that it worsens at large panels. It is a real threat to the higher-order part of this proposal. We settle it empirically. For each translator and cell type we compute CMD, pairwise-$\theta$ error and triplet-$\theta$ error. We then ask whether triplet error explains variance in a downstream failure that CMD does not: errors in the frequency of populations defined by three joint gates, such as naive T cells that are simultaneously CD4-positive, CD45RA-positive and CCR7-positive. If CMD alone predicts these errors equally well, we will report that triplet terms add nothing here and keep the pairwise-pure analysis as the contribution. The retina precedent makes this outcome plausible.

Why not just build a correlation network on the measured protein and rank hubs directly? This is the most natural baseline, and it is included. Raw protein correlations confound interaction with marginal abundance and with shared cell-state drive. In CITE-seq they also confound interaction with total antibody capture per cell, which totalVI notes it does not model as a nuisance. The Nakahara–Amari orthogonality is precisely what separates a change in how often a marker is on from a change in how it couples. The ablation ladder shows whether that separation moves the hub ranking. If raw-correlation hubs match pure-interaction hubs at the split-half ceiling, the decomposition is unnecessary for hub ranking and we will say so.

Why not just regress each protein on its own transcript and call the residual "discordance"? Per-gene residuals describe discordance one marker at a time. Hubs are a property of how markers couple, and two markers can have large residuals that are mutually independent. We include residual-based ranking as a baseline and test whether its top genes coincide with genes whose protein-level hub score differs most from their RNA-level hub score. Coincidence would mean the network analysis reduces to single-gene discordance. The ESPACE-style and bridge-gene logic suggests it will not, but the experiment decides.

Why not keep counts continuous instead of binarizing? Binarization discards information, and a continuous Gaussian graphical model is the fair alternative. We run graphical lasso on continuous counts as a matched baseline. We also vary the binarization rule (totalVI foreground probability versus fixed quantile thresholds) to show hub rankings are not artifacts of the cut. The argument for binary states is specific to surface proteins: bimodal markers such as CD4 are exactly what scChord shows regression-style models flatten, and gating biology is inherently binary.

Proposed Experiments

Datasets. The main dataset is the GSE164378 PBMC atlas: over 161,000 cells, 228 surface proteins, multiple donors for held-out splits. GSE100866 (about 8,600 cord blood cells, 13 proteins) serves as a small-panel control, where pairwise sufficiency is expected and triplet terms should be negligible. The murine SLN-all dataset (about 33,000 cells, two panels of 111 and 208 proteins) tests panel robustness: protein hub rankings on the shared markers should agree across panels if they reflect biology rather than panel composition. Simulated binary data from known log-linear models with planted pairwise and triplet terms, with CITE-seq-like background rates, calibrates the estimator's recovery and false positive rate before any real claim is made.

First decisive check. Is RNA–protein hub disagreement larger than the protein split-half noise ceiling? We measure Spearman rank correlation and top-$k$ overlap of hub scores per cell type, with donor-level bootstrap intervals. A null result here would substantially weaken the discordance motivation for hub studies, and it would be reported as such.

Second decisive check. Are translator-predicted protein networks closer to the RNA network or to the measured protein network? We measure the same agreement statistics for each translator's output against both references. The hypothesis is falsified if predicted-protein hubs match measured protein as well as measured protein matches itself.

Third decisive check. Does triplet structure matter beyond CMD, as measured by errors in gated three-marker population frequencies? The analysis is described under the first "Why not just" question.

Conditional experiment. If the second check shows a gap, we retrain scChord with the protein-interaction penalty. Success requires closing some of the hub and triplet gap without worsening per-protein RMSE and PCC relative to our reproduction of the unmodified model.

Controls. Every positive finding is checked against confounds. Triplet terms must survive conditioning on cell state and library size, since shared drive is the main alternative explanation the neuroscience literature warns about. They must also appear in both donor halves.