Most clinical tumor cohorts have bulk RNA sequencing and survival follow-up but no protein measurements. Researchers therefore run the same chain over and over: find genes whose RNA level tracks survival, test which pathways those genes are enriched in, and report a mechanism. Two separate distortions undermine that chain. First, RNA is a weak proxy for protein. Across matched cancer samples, transcript and protein levels correlate only around $r \approx 0.3$ across samples (Proteomizer). Second, a bulk sample averages many cell types, so a "prognostic pathway" may only mean that some tumors contain more immune cells or stroma. Each distortion has been studied, but only on its own. Proteomizer predicts bulk protein from bulk RNA and ignores composition. Deconvolution tools such as TAPE, POPPY and scMoE recover composition and stay entirely in RNA space.
We propose to model the two distortions jointly, through a specific mechanism. Deconvolution means estimating the cell-type mixture inside a bulk sample. We treat a bulk protein measurement as a mixture of cell-type-specific RNA-to-protein translations, weighted by the deconvolved cell fractions. We fit that mixture using only bulk protein as supervision. The fitted model then serves as an instrument for one question: is a bulk-RNA prognostic pathway signal (a) a composition artifact, (b) an RNA signal that does not carry through to protein, or (c) a cell-intrinsic protein-level effect? The cohorts that make this testable already exist. The Cancer Genome Atlas (TCGA) has samples with RNA, reverse-phase protein array (RPPA) measurements and survival. These let us check every RNA-derived call against measured protein.
Core Claim
Our hypothesis is that much of the gap between RNA and protein in bulk tumors is driven by composition, because translation and protein turnover differ by cell type. A changed cell mixture therefore changes the apparent bulk RNA-to-protein relation even when no regulation inside any cell has changed. If this is right, two consequences follow.
The first is about generalization. A translator that factorizes protein prediction through cell fractions should transfer better to unseen tissue types than a pooled model. Cross-tissue failure is exactly the weakness Proteomizer's authors report.
The second is about prognostic enrichment. Pathway enrichment computed on composition-separated, translated protein estimates should reproduce the prognostic pathway calls made from measured protein. Enrichment on bulk RNA should do so less often.
The contribution is therefore not another RNA-to-protein regressor. It is a structured model whose components carry a stated meaning, and an evaluation protocol that scores prognostic and functional claims against measured protein.
Research Program
The model has the form
$$\hat{p}_{sg} = \sum_{k=1}^{K} \pi_{sk}\, h_{kg}\big(x_{skg}, m_s\big),$$
where the symbols are defined as follows. $\hat{p}_{sg}$ is the predicted bulk protein level of gene $g$ in sample $s$. $\pi_{sk}$ is the deconvolved fraction of cell type $k$. $x_{skg}$ is the cell-type-specific RNA estimate produced by a TAPE-style decoder. $m_s$ is the sample's microRNA profile, which Proteomizer found necessary for capturing post-transcriptional silencing. $h_{kg}$ is a small cell-type-conditioned translation network that shares parameters across genes.
One caveat is that RPPA and mass spectrometry report relative, normalized intensities rather than absolute counts. The mixture is therefore fit in a per-platform transformed space with a learned per-gene affine link. Additivity is treated as an approximation, and we check it on pseudo-bulks.
Training happens in three steps. First, fractions and cell-type RNA are initialized from single-cell atlases, including the 3CA atlas used by POPPY. Second, the translators $h_k$ are pretrained on paired single-cell RNA and surface-protein (CITE-seq) data, where cell-type truth exists. We use GSE164378, with more than 161,000 cells and 228 proteins, and GSE100866. Third, the whole model is fine-tuned on the 8,613 matched TCGA/CPTAC RNA–microRNA–protein samples, with bulk protein as the only target.
For survival, the cell-type-resolved protein estimates $\pi_{sk} h_{kg}$ go into Cox models, which relate covariates to the risk of death over time. We run these at the gene level and at the pathway level, with the Gene Ontology (GO) and KEGG pathway databases as gene sets. Each pathway's hazard association is then attributed in a fixed order: to composition (the variance explained by $\pi$), then to translation (the change from RNA to predicted protein), then to the cell-intrinsic residual.
We keep DeepSigSurvNet's caution in view (DeepSigSurvNet). Pathways that are active in every tumor of a cohort carry no prognostic signal, so enrichment is scored on between-patient variance rather than on activity.
Baselines
The naive reference point is the standard practice itself. We fit a Cox model on bulk RNA for each gene and run overrepresentation analysis (a hypergeometric test of whether a pathway's genes appear among the prognostic genes more often than chance) on the significant genes. For protein prediction, the matching naive reference is RNA used directly as a proxy, which Proteomizer reports at $r = 0.27$ on its TCGA/CPTAC data.
The strongest existing translator is Proteomizer. It is a multilayer perceptron on RNA plus microRNA and reaches Pearson $r = 0.68$ on matched TCGA/CPTAC samples. It is the right competitor because it uses the same inputs and the same cohort. We retrain it on our exact splits, with the same tuning budget and compute. Beating it means significantly higher held-out-tissue correlation; matching it within tissue is acceptable. A second strong competitor is composition-adjusted Cox: TAPE fractions added as covariates next to bulk RNA. This is the obvious fix for the composition problem.
For survival discrimination we report two numbers as context, not head-to-head targets, because the inputs differ. DeepSigSurvNet reports a test concordance index (C-index; the probability that the model ranks a random pair of patients correctly by survival time) of 0.7438 in TCGA LUAD, against 0.5593 for a random forest. DeepSurv reports 0.643 on METABRIC, against 0.630 for linear Cox. We rerun DeepSigSurvNet on our cohort splits with RNA inputs and with predicted-protein inputs. Our main claim is about whether calls agree with measured protein, not about C-index, so matching it suffices.
The key ablation removes the mixture structure. It feeds the same fractions to a pooled translator as extra input features, which isolates whether the factorization itself matters.
Why not just…?
Why not just add cell fractions as Cox covariates? This removes the composition confound but not the RNA-to-protein gap. It also cannot detect a pathway whose protein, but not RNA, varies within a cell type. Including it as a baseline settles the question. If its pathway calls agree with measured-protein calls as well as ours do, the translation component adds nothing, and we will report that.
Why not just run Proteomizer with fractions appended as inputs? This is the real threat, because a flexible network could in principle learn composition-dependent translation. The ablation tests it directly. Our prediction is that the gap appears mainly in leave-one-tissue-out evaluation, where the pooled model must extrapolate to unseen mixtures. If no gap appears there, the structural claim is wrong. The remaining value would then be interpretability, and we would say so.
Why not just use measured proteomics? Measured proteomics is the gold standard, and we use it as ground truth. Most survival cohorts, METABRIC among them, have no proteomics. scpDeconv also showed that deconvolving proteomes needs single-cell proteomic references, which have limited coverage. That said, it showed real prognostic value, with $p = 0.031$ for vertical-growth melanoma fractions.
Proposed Experiments
The first check is the structural hypothesis on pseudo-bulks. We mix GSE164378 cells into bulks with known fractions and cell-type-specific surface proteins. The question is whether the bulk-trained model recovers per-cell-type protein, scored by correlation per cell type and by correlation-matrix distance. scChord (scChord) reports degraded correlation-matrix distance at hundreds of proteins; no reported numerical value was found on Lacuna for that metric.
The second check is leave-one-tissue-out translation on TCGA/CPTAC, comparing our model against Proteomizer, the fractions-as-input ablation and the RNA proxy.
The third and decisive experiment is the prognostic audit. In each TCGA cohort with RPPA data, we compute GO and KEGG prognostic enrichment from five sources: measured protein, bulk RNA, composition-adjusted RNA, Proteomizer predictions, and our model. The primary metric is agreement with the measured-protein calls, specifically precision and recall of significant pathways and concordance of hazard direction.
Two controls anchor the audit. As a positive control, immune and stromal gene sets with known infiltration-driven prognosis should be attributed to composition. As a negative control, pathways active in every tumor of a cohort should stay non-prognostic.
Finally, we apply the model to METABRIC and report which RNA-derived prognostic pathways it reclassifies, with the reclassifications pre-registered before outcomes are examined.
The program fails if our model's agreement with measured-protein calls does not exceed both composition-adjusted RNA and Proteomizer. It also fails if the held-out-tissue advantage over the fractions-as-input ablation does not appear.
References
- Proteomizer: Leveraging the Transcriptome-Proteome Mismatch to Infer Novel Gene Regulatory Relations
- Deep autoencoder for interpretable tissue-adaptive deconvolution and cell-type-specific gene analysis
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- scMoE: single-cell mixture of experts for learning hierarchical, cell-type-specific, and interpretable representations from heterogeneous scRNA-seq data
- Investigating the relevance of major signaling pathways in cancer survival using a biologically meaningful deep learning model
- Deep Survival: A Deep Cox Proportional Hazards Network
- Deep domain adversarial neural network for the deconvolution of cell type mixtures in tissue proteome profiling
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation