When a cell decides how much of a protein to make, the mRNA transcript is only the first draft. Systematic surveys show mRNA abundance tracks protein abundance only 40% to 60% of the time, so up to half of the variance in a cell's actual machinery is invisible to transcriptomics alone (A high-stringency blueprint of the human proteome). This gap matters most exactly where preclinical modeling matters most: in patient-derived organoids — miniature three-dimensional tumors grown from a patient's own cells — where diffusion limits create radial oxygen and nutrient gradients, and in the mouse-to-human translation step where 85–90% of oncology drug candidates die in the clinic. A growing family of machine-learning models (scChord, scTEL, totalVI, PIKE-R2P) now predicts protein abundance from single-cell RNA data, but every one of them was trained and benchmarked on dissociated blood cells on flat, uniform 2D cultures. Nobody has asked the two questions that decide whether these models are useful in translational oncology: does the RNA-to-protein mapping itself change across a tumor's oxygen gradient, and is that mapping conserved between mouse and human?
I propose to build the first spatially resolved, cross-species benchmark of the RNA-to-protein transfer function in pancreatic ductal adenocarcinoma (PDAC) organoids, and to use it to train a gradient-conditional translation model — a predictor that takes single-cell RNA plus a microenvironmental coordinate (local oxygen tension or radial depth) and outputs protein abundance. The design deliberately reuses two validated platforms: the TRACER tissue-rolling system, which imposes controlled radial oxygen gradients on patient-derived PDAC organoids and showed that hypoxia depth linearly drives classical-to-basal-like subtype switching ($R^2 = 0.78$) (Analysis of an engineered organoid model of pancreatic cancer identifies hypoxia as a contributing factor in determining transcriptional subtypes), and matched murine PDAC organoids from the branched-organoid system, which already supports side-by-side mouse and human culture (Heterogeneity-driven phenotypic plasticity and treatment response in branched-organoid models of pancreatic ductal adenocarcinoma).
Core Claim
The claim to be tested is concrete and falsifiable: transcript-protein discordance is not a uniform property of a cell type but a function of the local microenvironment, and the shape of that function differs between species. If true, an RNA-to-protein model trained on normoxic shell cells will systematically mispredict protein states in the hypoxic core, and a model trained on mouse organoids will mispredict human organoids — not randomly, but in a structured way that includes pathway-level sign inversions, as foreshadowed by the sbv IMPROVER species translation challenge, where the most predictive rat-to-human pathway was negatively correlated between species (Inter-species inference of gene set enrichment in lung epithelial cells from proteomic and large transcriptomic datasets). If false — if the transfer function is robust to oxygen and conserved across species — that is also a publishable, decision-relevant result, because it would license the off-the-shelf use of existing translation models in organoid drug screening.
Why This Is Not Already Done
The two closest prior works each stop one step short. The TRACER study connected organoid gradients to transcript-protein discordance only in passing: its CyTOF protein measurements served as targeted validation for four pre-selected markers (GATA6, AGR2, Vimentin, Cytokeratin 5), not as a systematic, transcriptome-wide quantification of discordance, and it made no cross-species comparison and no use of computational translation models. The sbv IMPROVER study did quantify cross-species multi-omic prediction, but entirely in 2D bulk bronchial epithelial cultures at acute time points, with no single-cell resolution, no organoids, no spatial gradients, and no deep-learning translation architecture. Separately, SATURN aligns cross-species single-cell atlases using protein-sequence embeddings (Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN), but it aligns RNA landscapes — it never touches measured protein abundance, let alone the RNA-to-protein map. No paper in the retrieved literature combines patient-derived organoids, spatially resolved paired RNA-protein measurement, and cross-species validation of the translation function; that intersection is empty.
Research Program
Three components, in order of dependency. First, the benchmark: patient-derived PDAC organoids from a planned cohort of roughly 20 patients are grown in the TRACER configuration, unrolled, and sectioned into radial layers after 24 hours of gradient culture, following the published protocol. Each layer is split for paired single-cell measurement — scRNA-seq plus CITE-seq on an expanded antibody panel targeting the classical/basal markers, EMT machinery, and stress-response proteins identified in the original TRACER paper, supplemented by mass cytometry for intracellular phospho-markers (p-AKT, p-mTOR, GSK3β) following the dual-assay logic of paired transcript/protein quantification established in expression-analysis work such as the PHLDA2 study (PHLDA2 regulates EMT and autophagy in colorectal cancer via the PI3K/AKT signaling pathway). Murine PDAC organoids from matched genetic backgrounds undergo the identical protocol, giving a paired human–mouse, normoxia–hypoxia, shell–core factorial design. Second, the measurement: for every gene-protein pair in every layer, we compute a per-cell discordance score (the residual of protein abundance after regressing out transcript abundance), yielding a discordance map indexed by oxygen depth. The key readout is whether discordance varies systematically with layer depth — for instance, whether hypoxic-core cells show translational suppression of epithelial proteins that their transcripts do not announce, a mechanism directly motivated by ribosome-level regulation cases such as METTL5-driven CCND3 translation without transcript change (METTL5‑mediated m6A modification of 18S rRNA drives oral squamous cell carcinoma progression by enhancing CCND3 translation). Third, the model: a conditional extension of existing translation architectures, where the input is the single-cell RNA vector concatenated with the microenvironmental coordinate and species indicator. Architecturally this is a modest delta on scChord's two-stage design — a variational autoencoder with a zero-inflated negative binomial decoder smoothing the protein manifold, then an optimal-transport conditional flow-matching map from RNA to protein (scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation) — but conditioned on depth and species rather than treating the map as a fixed steady-state transformation. For cross-species evaluation, gene-level features are aggregated into pathway-level gene-set enrichment scores, because gene-level ortholog transfer is known to fail between species while pathway-level scores with eight principal components carried predictive signal in sbv IMPROVER (Inter-species inference of gene set enrichment in lung epithelial cells from proteomic and large transcriptomic datasets); we explicitly test for negatively correlated pathways rather than assuming positive conservation.
Baselines
The naive reference point is the assumption embedded in most transcriptomic drug-screening pipelines: that mRNA abundance alone stands in for protein state. Concretely, this baseline predicts protein abundance by rank-normalized transcript abundance of the same gene. Its expected ceiling is set by the literature: the 40–60% correlation between mRNA and protein (A high-stringency blueprint of the human proteome), and the demonstrated clinical consequence that RNA-defined subtypes can invert prognosis relative to proteomic subtypes in ovarian cancer (Integrated proteomics and scRNA-seq analyses of ovarian cancer reveal molecular subtype-associated cell landscapes and immunotherapy targets). Beating this baseline means achieving a per-protein Pearson correlation with measured protein that is materially higher and, critically, more uniform across organoid layers than the transcript-only proxy achieves.
The strongest existing competitor is scChord, the current best-performing RNA-to-protein translation model; on the GSE100866 cord-blood CITE-seq benchmark it reports the lowest RMSE and highest Pearson correlation among totalVI, sciPENN, and scTranslator, though no reported result found on Lacuna for its exact decimal RMSE/PCC values. It is the strongest competitor because it is the only method that generatively models full protein distributions (capturing bimodal markers like CD4), and its published training cost — 100,000 cells in 27 minutes on 229 MB of GPU memory — makes a fair comparison cheap to run. We will run scChord in two modes, both on our paired organoid CITE-seq data with identical train/test splits (held-out patients, held-out layers), identical antibody panels, and a matched tuning budget: zero-shot (pretrained weights, no organoid data) and retrained on our normoxic-shell layers only, then evaluated on hypoxic-core layers. Beating scChord means higher per-protein Pearson correlation and lower RMSE on held-out hypoxic-core layers specifically, plus better Correlation Matrix Distance on protein-protein co-expression structure — a metric where scChord already degrades as panels grow, so improvement there is meaningful. For the cross-species axis, the strongest baseline is the sbv IMPROVER recipe itself: gene-set enrichment scores, PCA to eight components, and a Naive Bayes classifier predicting human pathway activation from mouse data, which reported high accuracy in the original challenge though no reported result found on Lacuna for its exact classification accuracy; we reimplement it on our organoid data. Beating it means the conditional model predicts human organoid protein-state pathways from mouse organoid RNA with higher accuracy while additionally localizing where along the gradient the mouse-to-human map fails, which the pathway-level classifier cannot do because it discards spatial information. Finally, the ablation that isolates the proposal's key ingredient is the full conditional model with the depth/oxygen coordinate removed (and, separately, with the species indicator removed): if performance on hypoxic-core layers collapses to scChord's level without the gradient input, the microenvironmental conditioning — not extra parameters — is what earned the gain.
Why not just…?
Why not just sequence RNA deeply and skip the proteomics? This is the cheapest alternative and the one most worth arguing with, because deep scRNA-seq genuinely covers 20,000+ genes at scale. But the 40–60% correlation ceiling is a property of the biology, not the sequencing depth (A high-stringency blueprint of the human proteome): translation efficiency, protein half-life, and phosphorylation sit between transcript and function. The failure is not hypothetical — in esophageal squamous cell carcinoma, transcriptomics missed the post-neoadjuvant immune-microenvironment shifts that single-cell proteomics detected (High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy), and in ovarian cancer the RNA-defined immune-rich subtype looked favorable while the proteomic data showed protein-level checkpoint suppression and poor survival (Integrated proteomics and scRNA-seq analyses of ovarian cancer...). Our proposal settles this empirically within its own data: if the transcript-only baseline matches the conditional model on held-out layers, the discordance hypothesis fails and we will report that.
Why not just run an off-the-shelf model like scChord or scTEL on the organoid RNA? If this worked, the proposal would collapse to an application paper, so we test it directly as a baseline rather than arguing around it. Three published observations predict it will fall short. First, tissue specificity: PIKE-R2P's authors note that models trained on blood CITE-seq data do not generalize to solid tissues because translational regulatory networks differ by tissue (PIKE-R2P: Protein-protein interaction network-based knowledge embedding with graph neural network for single-cell RNA to protein prediction). Second, none of these models accepts a spatial or microenvironmental input, so a gradient-dependent transfer function is structurally invisible to them. Third, scChord's own Correlation Matrix Distance degrades on large panels, and scTEL's predictions are locked to the antibody panels in its training reference (A joint analysis of single cell transcriptomics and proteomics using transformer). That said, this is a real threat: if zero-shot scChord holds up on hypoxic-core layers, the microenvironmental component of the proposal is unsupported, and the zero-shot baseline is designed precisely to reveal that outcome either way.
Why not just use human organoids and skip the mouse arm entirely? Human patient-derived organoids do carry patient-specific genetics, and dropping the mouse arm would cut cost and timeline substantially. But the mouse arm is not a detour — it is the object of study. The premise being tested is whether the translation function is conserved, and sbv IMPROVER showed it can be anticorrelated at the pathway level between rat and human (Inter-species inference of gene set enrichment...), while SATURN found single genes like Myoc drifting into different functional contexts across species (Towards Universal Cell Embeddings...). Since the field will keep running preclinical drug screens in mouse organoids (as the branched-organoid work already does, using mouse models for metastatic-competence validation that no in vitro system provides), quantifying where the mouse-to-human protein map breaks is exactly the knowledge needed to interpret those screens. If the map turns out to be conserved, the mouse arm delivers a clean negative result that licenses faster translation — either outcome is informative, which is what makes this the right experiment rather than a redundant one.
Proposed Experiments and Decisive Checks
The experimental sequence has three gates. Experiment one establishes the discordance map: across the TRACER layers of the human cohort, test whether per-gene discordance scores vary with radial depth, with the pre-registered expectation that markers of translation and protein stability programs (unfolded protein response, glycolysis, EMT — the pathways the TRACER study found enriched in hypoxic zones) show the strongest depth-dependence, benchmarked against that study's reported $R^2 = 0.78$ for depth versus basal-like cell fraction as the scale of gradient effect we should be able to detect. Experiment two is the prediction contest: naive transcript proxy, zero-shot and retrained scChord, and the depth-conditional model, all evaluated on held-out patients' hypoxic-core layers with per-protein Pearson correlation, RMSE, and Correlation Matrix Distance; the decisive check is whether the conditional model's advantage concentrates in the deep hypoxic layers and disappears in the normoxic shell, which is the signature of a genuinely gradient-dependent transfer function rather than generic model capacity. Experiment three is the species test: train on mouse organoid layers, predict human organoid protein-state pathways, and compare against the reimplemented sbv IMPROVER GSEA-PCA-Naive-Bayes baseline; the decisive check is the sign structure — if any high-mutual-information pathway comes out negatively correlated between mouse and human organoids, that replicates the inversion phenomenon in a patient-derived 3D system for the first time and warns directly against naive preclinical transfer. As a translational tiebreaker, drug-response prediction for gemcitabine (the agent the TRACER study linked to hypoxic-zone resistance) is run twice — once from RNA-inferred state, once from protein-inferred state — and scored against measured layer-specific viability; if protein-inferred state predicts response where RNA-inferred state does not, the clinical stakes of the whole program are made concrete. Each experiment is sized so that a clean negative result at any gate is publishable and decision-relevant, which is the property that makes this proposal worth doing rather than merely combinable.
References
- A high-stringency blueprint of the human proteome
- Analysis of an engineered organoid model of pancreatic cancer identifies hypoxia as a contributing factor in determining transcriptional subtypes
- Heterogeneity-driven phenotypic plasticity and treatment response in branched-organoid models of pancreatic ductal adenocarcinoma
- Inter-species inference of gene set enrichment in lung epithelial cells from proteomic and large transcriptomic datasets
- Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN
- PHLDA2 regulates EMT and autophagy in colorectal cancer via the PI3K/AKT signaling pathway
- METTL5‑mediated m6A modification of 18S rRNA drives oral squamous cell carcinoma progression by enhancing CCND3 translation
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- Integrated proteomics and scRNA-seq analyses of ovarian cancer reveal molecular subtype-associated cell landscapes and immunotherapy targets
- High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy
- PIKE-R2P: Protein-protein interaction network-based knowledge embedding with graph neural network for single-cell RNA to protein prediction
- A joint analysis of single cell transcriptomics and proteomics using transformer