Blood is the cheapest window onto disease, so a large literature reads pathway activity out of blood RNA. Examples include "neuroinflammation pathways are up in Alzheimer's blood" and "lithium rewires monocyte signaling." These claims are then treated as statements about cellular function. They are made with enrichment tests that count messenger RNA as if every transcript became working protein, yet mRNA and protein abundance agree only roughly half the time across the genome (A high-stringency blueprint of the human proteome). Enrichment analysis here means asking whether a curated gene group, such as a Gene Ontology term, is overrepresented among the changed genes.
Nobody has checked, gene set by gene set and cell type by cell type, how often these RNA-level pathway conclusions would hold if one could measure protein instead. One kind of dataset makes that check possible for blood specifically. CITE-seq measures RNA and a panel of surface proteins in the same single cells. Using it, we propose to measure how often RNA-based enrichment calls in blood immune cells agree with protein-based calls on the same cells. We then test whether a cheap, per-cell-type estimate of how well each gene's RNA tracks its protein can predict which pathway calls will fail. Finally, we apply that estimate to reweight enrichment in large bulk-blood surrogate cohorts where protein was never measured.
Core Claim
The claim is that transcript–protein disagreement is structured by cell type in blood. Because of that structure, RNA-derived pathway calls in surrogate-blood studies are unreliable in a predictable way. A per-gene, per-cell-type "concordance weight" learned from paired single-cell RNA and protein should predict which enrichment calls a protein measurement would overturn. It should do so better than a single genome-wide concordance filter, and about as well as full protein imputation, at a fraction of the complexity.
What is new is the target of evaluation. Prior RNA-to-protein models such as totalVI and scChord are scored on per-protein error (RMSE, Pearson correlation). Surrogate-tissue studies score only diagnostic accuracy. None of them asks whether the pathway-level biological conclusion changes. The scChord authors themselves leave the downstream effect of predicted protein on pathway analysis open.
Research Program
Each prior line of work supplies a piece, and each stops short of this question.
- Joint single-cell models (method supply). totalVI models CITE-seq RNA and protein jointly and separates true antibody signal from background, a known problem for markers like CD28. scChord improves protein prediction with flow matching, a generative method that learns a smooth transport from RNA embeddings to protein space. It reports beating totalVI, sciPENN and scTranslator on the GSE164378 and GSE100866 benchmarks. Both stop at predicting proteins.
- Single-cell enrichment scoring (method supply). DeepGSEA and classical cell-level scorers such as AUCell and GSVA score gene sets in single cells, but only from RNA.
- Surrogate-blood studies (motivation). These draw pathway conclusions from blood RNA with no protein layer:
- bulk microarray enrichment in neurodegeneration (Shared Molecular Signatures Across Neurodegenerative Diseases and Herpes Virus Infections);
- cell-type deconvolution of 1,730 whole-blood samples (Cell-type deconvolution of bulk-blood RNA-seq reveals biological insights into neuropsychiatric disorders);
- plasma cell-free RNA panels (Integrative single-cell and cell-free plasma RNA transcriptomics identifies biomarkers for early non-invasive AD screening).
The program has three stages.
Stage one: a protein-grounded enrichment benchmark. We use the GSE164378 PBMC atlas (more than 161,000 cells, 228 surface proteins) and the GSE100866 cord-blood set. Cells are typed with CellTypist, which reports a global F1 of about 0.9. We denoise protein with totalVI's background mixture.
For each contrast, we compute two enrichment calls on the same cells for gene sets restricted to genes whose protein is on the panel. The contrasts are cell type versus cell type, and condition versus condition within a cell type where the metadata permit. The "RNA call" uses the RNA of those genes. The "protein call" uses the denoised protein of the same genes, with an identical test.
Disagreement between the two calls is the ground-truth failure signal. The restriction to panel genes is deliberate. It is the only setting where the truth is observed rather than modeled, and we accept that it covers mostly surface-receptor and immune-activation biology.
Stage two: concordance weights, and whether they transfer to unmeasured genes. For gene $g$ in cell type $c$, the weight $\rho_{gc}$ is the within-cell-type correlation between denoised RNA and denoised protein across cells.
Enrichment is then recomputed with each gene's contribution scaled by $\rho_{gc}$ (clipped at zero). This choice keeps the method a drop-in change to any rank- or score-based test. We then ask the harder question: can $\rho_{gc}$ be predicted for genes without a protein measurement? The predictor uses only RNA-side features:
- mean expression and cell-to-cell variance;
- dropout rate;
- cell-type specificity;
- the RNA's agreement with neighboring genes in the same set.
It is evaluated by leaving whole genes out. If held-out $\rho_{gc}$ is not predictable, the reweighting cannot be exported beyond the panel. We will report that as a negative result rather than proceed.
Stage three: re-examining published-style surrogate conclusions. We apply the transferred weights to cell-type expression deconvolved from the dbGaP whole-blood psychiatric cohort (phs002856.v1). There we rerun the lithium-dependent cell-type analysis that produced 82 genes, versus one gene in bulk. We ask which enriched pathways survive reweighting.
The plasma cell-free RNA setting carries an explicit constraint. Our weights are learned in blood immune cells, so they can only speak to the blood-cell-derived part of plasma RNA. They say nothing about brain-derived leakage, and we will not apply them to the brain-specific genes in the 34-gene Alzheimer's panel.
Baselines
Naive reference. The do-nothing reference is the RNA call itself: standard enrichment on RNA with no correction. We score it against protein calls on the benchmark. It uses the same gene sets, cells, test, and false discovery rate threshold as every other method. Its agreement rate with protein calls is the number every method must beat, and we know of no reported value for it, which is the point of stage one.
Global concordance filter. The next reference drops or down-weights genes whose mRNA–protein correlation is low genome-wide, with no cell-type resolution. This is the natural reading of the blueprint paper's 40–60% figure. We implement it by pooling all cells when estimating $\rho_g$, so the only difference from our method is cell-type specificity. That makes it the key ablation for the proposal's central ingredient.
Strongest existing method: impute, then enrich. Here totalVI and scChord predict protein for every cell, and enrichment is run on predicted protein. We use scChord's published architecture and a matched hyperparameter search budget for both models. The training and evaluation donor splits are shared with our weight estimation, so no method sees test donors.
For scChord, the file reports that it achieved the lowest RMSE and highest Pearson correlation on GSE164378 and GSE100866 against totalVI, sciPENN and scTranslator. It gives no numerical values (no reported result found on Lacuna for scChord's aggregate RMSE/PCC). Its authors also note degraded correlation-structure preservation as the protein count grows into the hundreds, which is exactly the regime where pathway-level structure matters.
RNA-only deep enrichment. DeepGSEA is included as an RNA-only method with the strongest reported sensitivity. It reports significantly higher sensitivity than ssGSEA, GSVA, AUCell, Vision, SCPA and scGSEA in simulations (no numerical value on Lacuna). Its role is to test whether more sensitive RNA enrichment helps or hurts protein agreement, since greater sensitivity to untranslated signal could increase false pathway calls.
What counts as beating a baseline. Beating the naive and global baselines means a significant gain in the area under the precision–recall curve for recovering protein-level calls, paired across contrasts. We also require a reduction in calls that are RNA-positive but protein-negative. Matching impute-then-enrich within its confidence interval counts as success for the weight approach, because it is far simpler and transfers to bulk deconvolved data, where imputation models trained on single cells are not designed to run.
Why not just…?
Why not just measure circulating proteins directly? Plasma proteomics such as the mass-spectrometry panel of 133 proteins behind the Generation Scotland EpiScores (Methylome-wide association studies and epigenetic biomarker development for 133 mass spectrometry-assessed circulating proteins in 14,671 Generation Scotland participants) measures shed and secreted proteins. It does not measure cell-type-resolved protein inside immune cells, and most pathway claims in blood RNA studies concern the latter. It also cannot retroactively audit the many existing RNA-only cohorts, such as the 1,730-sample psychiatric set. The alternative is complementary rather than a competitor, and we do not claim to replace it.
Why not just impute proteins with totalVI or scChord and run ordinary enrichment? This is the real threat. Imputation may capture discordance as well as or better than a single correlation weight, because it models multimodal protein distributions such as bimodal CD4. Stage one settles it head-to-head on held-out donors. If imputation clearly wins on the panel, we still face its limit: it predicts only the proteins it was trained on and requires single-cell input. The weight predictor in stage two is the only route to unmeasured genes and bulk data. If that predictor also fails, the honest conclusion is that discordance correction cannot currently leave the antibody panel. We would publish the benchmark and the failure rates as the contribution.
Why not just use a global mRNA–protein correlation list? It is simpler and needs no single-cell data. We test it directly as an ablation. The case for cell-type weights rests on the observation from PINNACLE that protein roles are strongly cell-type dependent. That observation concerns interaction context, however, not translation. If the global filter matches cell-type weights on the benchmark, we adopt it and drop the cell-type machinery.
Proposed Experiments and Decisive Checks
Check one: is there a problem? On GSE164378 and GSE100866, we report the fraction of RNA enrichment calls overturned by protein calls, stratified by cell type and gene-set category. We attach donor-bootstrap intervals and a permutation null where cell labels are shuffled within contrasts. If agreement is near perfect, the premise fails and the study ends there with a useful negative.
Check two: does cell-type-resolved concordance predict failures? We measure precision–recall area for predicting overturned calls, with cell-type $\rho_{gc}$ against global $\rho_g$ against imputation. All three use the same splits.
Check three: is concordance predictable for genes never measured as protein? We test leave-genes-out prediction of $\rho_{gc}$ from RNA-only features, reported as correlation against held-out truth. A pre-set threshold decides whether stage three runs.
Check four: what changes in real surrogate cohorts? We apply the method to deconvolved cell-type expression in the psychiatric cohort. The deconvolution in that paper reports agreement of $R^2$ from 0.61 to 0.84 against the DICE single-cell reference, which bounds how much of any change we can attribute to biology rather than deconvolution error. We report which lithium-associated pathways lose significance under reweighting.
A falsifiable prediction is attached to this cohort. Pathways dominated by panel-measurable surface receptors should show the smallest shifts, if the weights are calibrated.
References
- A high-stringency blueprint of the human proteome
- Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- DeepGSEA: explainable deep gene set enrichment analysis for single-cell transcriptomic data
- Shared Molecular Signatures Across Neurodegenerative Diseases and Herpes Virus Infections Highlights Potential Mechanisms for Maladaptive Innate Immune Responses
- Cell-type deconvolution of bulk-blood RNA-seq reveals biological insights into neuropsychiatric disorders
- Integrative single-cell and cell-free plasma RNA transcriptomics identifies biomarkers for early non-invasive AD screening
- Cross-tissue immune cell analysis reveals tissue-specific features in humans
- Methylome-wide association studies and epigenetic biomarker development for 133 mass spectrometry-assessed circulating proteins in 14,671 Generation Scotland participants
- Contextual AI models for single-cell protein biology