Composition-Aware Bipartite Modules of Transcript–Protein Discordance

When a cell's protein level departs from what its messenger RNA predicts, something downstream of transcription is at work: translation control, protein degradation, or modification. Mapping which groups of transcripts jointly shape which groups of such departures would give a structured picture of this regulation. The natural tool is bipartite community detection, which splits a two-sided graph (here, transcripts on one side and proteins on the other) into blocks where each group of transcripts is unusually tied to one group of proteins. The trouble is that the data come from mixtures of cell types. Across a mix of T cells, B cells and monocytes, almost every transcript correlates with almost every protein simply because both track cell identity. Any bipartite module found on pooled cells will therefore mostly rediscover cell-type markers, and not regulation.

We propose to measure this confound directly, then remove it with a bipartite modularity whose random-graph baseline is conditioned on cell state. Only associations that hold within a cell state then count as modular structure. We will test three things on paired RNA–protein single-cell data (CITE-seq, which reads each cell's transcriptome together with a panel of antibody-tagged surface proteins). First, how much of naive bipartite structure is composition. Second, whether the surviving modules recover planted regulation in semi-synthetic data and reproduce across cohorts. Third, whether they help generative RNA-to-protein models capture how proteins covary with one another.

Core Claim

The claim has two parts, and the first stands even if the second fails. The first part is empirical. On mixed immune populations, most of the RNA–protein bipartite communities found by standard bipartite modularity are explained by cell-type proportions. So bulk and pooled bipartite analyses in the style of CONDOR are reading out composition to a degree nobody has quantified. Bipartite Community Structure of eQTLs found 52 communities in bulk lung tissue, with disease variants acting as local leaders. It did this on homogenized tissue, and the heterogeneity literature, for example Genome 2011, warns that bulk averaging conflates proportions with regulatory switches. Nobody has put a number on that conflation for bipartite modules. Here we can, because CITE-seq gives ground-truth cells from which we can build pseudobulk mixtures with controlled composition.

The second part is methodological. We score edges and the random-graph baseline conditionally on an RNA-only estimate of cell state. The resulting modules are smaller and reproducible, and they carry information about protein covariance that unstructured translators miss. scChord maps RNA to protein with a flow-matching generator. Its authors report that correlation matrix distance (CMD, a measure of how well predicted protein–protein correlations match the real ones) degrades as the panel grows into hundreds of proteins, and they call for "pathway-informed graph priors." Our modules are one concrete candidate for such a prior.

Research Program

Take a CITE-seq dataset with cells $n$, transcripts $i$ and proteins $j$. We first build an RNA-only representation of each cell's state, $s_n$. This can be a nearest-neighbor graph on RNA embeddings, or an RNA-only variational autoencoder.

The RNA-only restriction is the central design choice. totalVI learns a shared 20-dimensional latent space from RNA and protein jointly. Conditioning on it would absorb exactly the protein-specific variation we want to study. scTEL shows why this matters: CD8+ naive versus effector-memory subsets are separated by CD45RA and CD38 protein but are invisible in RNA alone. With an RNA-only state, such protein-defined substructure counts as discordance, which is the honest label for it.

For each protein we then compute a discordance residual. This is the observed protein count minus a prediction from its own transcript and $s_n$, computed with a count model and cross-fitted so no cell is scored by a model that saw it. The bipartite edge weight $W_{ij}$ is the within-state association between transcript $i$ and the residual of protein $j$. We estimate it by pooling associations computed inside local neighborhoods of $s$ rather than across the whole population. The random-graph baseline is built by permuting transcript values only among cells of similar state. That permutation preserves each transcript's dependence on composition and destroys only its within-state coupling. The objective is Barber-style bipartite modularity with this conditional baseline:

$$Q_{\text{cond}} = \frac{1}{W}\sum_{i,j}\big(W_{ij} - \mathbb{E}[W_{ij}\mid s]\big)\,\delta(g_i, h_j).$$

Here $W$ is the total edge weight, $\mathbb{E}[W_{ij}\mid s]$ is the expected weight under the within-state permutation, and $g_i$, $h_j$ are the community labels of transcript $i$ and protein $j$.

Maximizing this is NP-hard, as shown in Maximizing Barber's bipartite modularity is also hard. We therefore use the iterative heuristic CONDOR uses and assess stability across restarts. Following CONDOR, we also compute each transcript's core score within its module, to ask whether the local-leader pattern found for disease variants reappears for post-transcriptional regulators.

Discordance is not a new concept here: A high-stringency blueprint of the human proteome motivates the whole effort by reporting that mRNA explains only part of protein variance. What we add is a way to give that unexplained variance structure while guarding against composition.

The work has a firm scope limit. CITE-seq panels cover roughly 10 to 300 surface proteins. Our modules therefore describe discordance in the surface proteome only, and we will not claim intracellular or modification-level regulation.

Baselines

The do-nothing reference is standard bipartite modularity (CONDOR's formulation) on pooled cells. It uses raw RNA–protein correlations as edges and the usual degree-based random-graph baseline. This is the method the field would apply by default, and it is what the first half of our claim says is mostly composition. It runs on the same cells, correlation estimator and heuristic optimizer as our method, with the same number of restarts. CONDOR reports 52 communities and a 20.3-fold median core-score enrichment for COPD GWAS SNPs on LGRC bulk eQTL data. That is not an RNA–protein result, and we found no reported result on Lacuna for bipartite RNA–protein community detection on CITE-seq.

The strongest competitor is a two-step procedure: regress each protein and transcript on discrete cell-type cluster labels, then run standard bipartite modularity on the residuals. It is cheap and directly targets composition, and a reviewer would reasonably expect it to capture most of the benefit. We give it the same clustering resolution search as our neighborhood size search, and the same optimizer budget. It shares our core step, and differs only in handling state as discrete labels rather than continuous neighborhoods, which can leave within-cluster gradients such as the CD45RA continuum intact.

Two further baselines are unipartite. WGCNA-style or Louvain clustering on a concatenated gene–protein correlation graph represents projection-based methods like the one PhenoMol uses on bulk multi-omics. Clustering of totalVI latent dimensions or attention maps represents dense-latent integration.

Three ablations isolate the key ingredients. The first keeps the conditional edges but uses the unconditional baseline. The second conditions on the joint RNA-plus-protein totalVI latent instead of an RNA-only state, and we predict it will erase true discordance modules. The third drops cross-fitting.

Beating the strongest baseline means two things. On the semi-synthetic benchmark, the adjusted Rand index (a 0-to-1 agreement score between recovered and true groupings) against planted modules must be higher. Across cohorts, module reproducibility must be higher. Both must hold at matched module counts, with the gain exceeding the spread across random restarts. For the downstream use, the strongest existing translator is scChord, which reports the lowest RMSE and highest Pearson correlation against totalVI, sciPENN and scTranslator on GSE164378 and GSE100866. The file gives no numeric values, so we will rerun it under our splits. We found no reported CMD value on Lacuna for scChord on GSE164378. Beating it means lower held-out CMD with no loss in per-protein Pearson correlation.

Why not just…?

Why not just regress out cell-type labels and run CONDOR? This is the real threat. If it matches our method on the semi-synthetic and reproducibility tests, the methodological half of the proposal collapses into "residualize, then cluster." The empirical finding about composition would still stand. We expect discrete labels to fail where the confound is a continuum, and scChord's heavy-tailed CD45RA distribution suggests such continua exist in these data. The semi-synthetic benchmark tests this directly: it plants confounds that vary continuously within clusters as well as confounds that are purely between clusters. If the label-based method wins on the purely discrete setting and loses only on the continuous one, we will report exactly that.

Why not just fit sparse regressions from transcripts to protein residuals and skip community detection? Sparse regression gives per-protein predictors but no shared blocks. The question is whether the blocks carry information. The CMD experiment settles it. If adding a module-block covariance prior to scChord does not lower CMD relative to an equal-parameter prior built from sparse-regression coefficients, then modules add nothing beyond sparsity, and we will say so.

Why not just project to a gene–protein graph and use WGCNA or Louvain? Dense Graphlet Statistics of Protein Interaction and Random Networks argues that triangle-based models cannot represent parallel bipartite modules. Projection also blurs which side of the graph each node sits on, which is what the core-score analysis needs. We still run projection as a baseline, because the argument is theoretical and its practical cost on CITE-seq is unmeasured.

Why not just read structure off totalVI's latent space or a translator's attention weights? Both are dense summaries that were never trained to separate composition from regulation. totalVI's latent space is fit jointly to protein, so any structure extracted from it is contaminated by the protein-defined states we want to keep. The joint-latent ablation measures this contamination directly.

Proposed Experiments

The first experiment quantifies masking. From GSE164378 (about 161,000 PBMCs, 228 proteins), we construct pseudobulk samples by resampling annotated cells with controlled, varying compositions, and run the do-nothing baseline on them. We then regress each community's aggregate signal on the known proportions. The key output is the fraction of modular structure explained by composition. This is the pooled-data analogue of the CONDOR setting, now with ground truth.

The second experiment is a semi-synthetic benchmark on the real cell backgrounds. We keep the observed RNA and plant protein-residual modules driven by chosen transcript groups, crossed with composition confounds that are either discrete or continuous. We score recovery of the planted modules with adjusted Rand index against all baselines and ablations.

The third experiment tests reproducibility. We learn modules separately on GSE164378 and on the H1N1 and Monocyte CITE-seq sets, restricted to shared proteins, and measure membership overlap with Jaccard similarity. A composition-free method should reproduce better than the pooled baseline across cohorts with different cell mixtures. The Monocyte set is nearly single-lineage, so it serves as an internal control where conditional and naive modules should agree.

The fourth experiment tests downstream value. We add module-block structure as a covariance prior to scChord's generator and compare held-out CMD and per-protein Pearson correlation against unmodified scChord, totalVI and a sparse-regression prior, all under the same splits and tuning budget. We run this on GSE164378 and on the 208-protein panel of the murine SLN-all data.

The decisive check combines the first two experiments. If composition explains little of the naive modules, or if label residualization equals our method on continuous confounds, the core claim fails in a clearly identifiable way.