When a gene is transcribed, its RNA message still has to be translated into protein, and that second step is loosely coupled at best: across the human proteome, mRNA levels agree with actual protein abundance only 40% to 60% of the time A high-stringency blueprint of the human proteome. This matters clinically because almost everything a drug does, it does to proteins. Yet the field's pipeline for predicting patient drug response runs the other way: hundreds of thousands of drug-sensitivity measurements exist for cell lines grown in dishes, and the best transfer-learning models carry that knowledge to patients using RNA expression or DNA mutations alone. At the same time, the patients who most need better predictions are those with "receptor-negative" tumors — triple-negative breast cancer, driver-negative lung adenocarcinoma — where no surface receptor or classic driver mutation gives a targeted hook, and the treatment decision collapses onto cytotoxic combinations chosen partly by guesswork.
This proposal connects the three threads with one mechanism. I propose to learn the RNA-to-protein mapping on cell lines, where paired transcript and protein measurements exist, and then use that mapping to impute each patient tumor's functional proteome from its RNA-seq — so that drug response knowledge transfers through protein space rather than RNA space. The bet is that the translation-regulation code itself (which RNAs get translated efficiently, which proteins persist) is more conserved between dish and body than the abundance distributions of either RNA or protein, which shift wholesale when a tumor grows in living tissue.
Core Claim
A drug-response model whose input is an imputed, network-regularized proteome will predict patient response better than the same model reading RNA directly, because roughly half of protein variance — including the protein-level consequences of driver mutations in TP53 and ARID1A — is invisible in transcripts. The claim has a falsifiable core: if a sufficiently expressive RNA-input model, given the same pretraining and tuning budget, matches the imputed-proteome model, the intermediate layer adds nothing and the idea dies. Two findings motivate the bet. First, Drug Response Modeling across Cancers: Proteomics vs. Transcriptomics showed with an identical architecture (BDRN) that measured proteins beat RNA for predicting drug potency in cell lines — a median Pearson correlation of 0.5467 versus 0.5311 for RNA and 0.5261 for DeepOmicNet — but that model never left the dish. Second, the models that do leave the dish — DeepSADR on 1,426 transcripts, GANDALF on a 324-gene mutation panel, and $CSG^2A$ on 978 landmark genes — are all protein-blind by construction. Nobody has yet coupled the protein advantage to the transfer problem; that coupling, with the discordance modeled explicitly rather than averaged away, is the new object of study.
Research Program
The build has three stages. Stage A trains the imputer $g_\phi$ on the ProCan-DepMapSanger cell lines, where RNA-seq and mass-spec proteomics are paired for the same line. The architecture is a generative RNA-to-protein model, not a point regression: single-cell work has shown why this matters, since deterministic regression collapses to the conditional mean and erases multimodal protein states like bimodal CD4 scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation, and a hierarchical decoder that separates biological signal from technical background handles proteomics' missing values honestly A Joint Model of RNA Expression and Surface Protein Abundance in Single Cells. The imputer is constrained by protein-interaction priors (STRING), because scChord's authors report that RNA-to-protein fidelity degrades as protein panels grow without explicit network structure, and because gene-gene attention tied to STRING priors was exactly what let $CSG^2A$ transfer perturbation rules across contexts. The training objective combines three terms, one per supervision source:
$$\mathcal{L} = \underbrace{\|g_\phi(x_i) - p_i\|^2_{\text{paired lines}}}_{\text{RNA–protein pairs}} \;+\; \lambda\, \underbrace{\mathcal{L}_{\text{PPI}}(g_\phi)}_{\text{network smoothness}} \;+\; \mu \underbrace{\ell\big(f_\theta(g_\phi(x_i), d),\, y_i\big)}_{\text{drug response}},$$
where $x_i$ is a cell line's transcriptome, $p_i$ its measured proteome, $d$ a drug, $y_i$ the measured sensitivity, and $f_\theta$ the response head. The third term is what makes the imputer task-aware: it is steered to get the proteins right that matter for drug response, not to reproduce the whole proteome evenly.
Stage B couples the imputer to perturbation biology. LINCS L1000 (~200,000 profiles of drug-induced transcriptional change) teaches how drugs rewire gene-gene relationships, a signal shown to beat raw expression snapshots for identifying drug mechanisms Cell-Level Virtual Screening. The single existing paired perturbation proteomics dataset — TMT profiles of triple-negative breast cancer cells after CDK4/6 inhibition, in which palbociclib strips Rb phosphorylation and raises surface Trop-2 and MHC Proteomic characterisation of triple negative breast cancer cells following CDK4/6 inhibition — serves as an anchor: the model must reproduce those protein-level consequences from transcript-level input. Stage C performs the transfer: apply $g_\phi$ to patient RNA-seq from TCGA, feed the imputed proteome into $f_\theta$ with an adaptive readout of the kind DeepSADR showed is essential (removing theirs dropped average clinical AUC from 0.856 to 0.662), and fine-tune on the 555-patient, 5-drug TCGA benchmark. The receptor-negative setting is the primary stratification, not an afterthought: triple-negative breast cancer cases and the cytotoxic drugs they actually receive (paclitaxel, gemcitabine, cisplatin) are where the prediction is most consequential and where protein-level states like the palbociclib-induced senescence program — invisible in RNA — plausibly govern outcome.
Baselines
The naive reference point is a drug-structure-only predictor, which on the drug-disease retrieval benchmark DDR-Bench reaches a Hits@1 of just 0.0357 with molecular fingerprints versus 0.1250 for cell-level network models Cell-Level Virtual Screening; in receptor-negative tumors there is no receptor pocket to dock against in the first place, so molecule-only screening is structurally unfit here. The strongest existing competitor is DeepSADR DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction, the current best cell-line-to-patient model: it reports an AUC of 0.957 on Sorafenib (versus 0.811 for the next best baseline) and an average clinical AUC of 0.856 across its five patient drugs. I will reimplement DeepSADR and retrain it on exactly the same DepMap pretraining corpus (966 cell lines, 20 drugs, roughly 225,781 response samples) and the same TCGA fine-tuning split, with the same tuning budget and compute as my model, so the only difference is the input representation. Beating it means a higher mean clinical AUC with non-overlapping bootstrap confidence intervals across the five drugs, plus a preregistered check that the gain concentrates in the receptor-negative strata rather than being averaged in. A second comparator is the GANDALF recipe of diffusion-based domain augmentation, since it reports gains of up to 10.96% AUROC over DruID and PANCDR on the same chemotherapies; no absolute AUROC for GANDALF itself is reported on Lacuna, so it must be run, not quoted. Finally, on the in vitro side, BDRN with measured proteins (median Pearson 0.5467 on $ln(IC_{50})$) is the ceiling anchor Drug Response Modeling across Cancers: Proteomics vs. Transcriptomics: my imputed-proteome model should land between the RNA model's 0.5311 and that ceiling on cell lines. Three ablations isolate the ingredients: the full architecture with raw RNA instead of the imputed proteome (isolates imputation), the imputer without PPI priors (isolates the network constraint), and a ridge-regression imputer in place of the generative one (isolates distributional modeling).
Why not just…?
Why not just build a bigger RNA model? This is the serious objection, because the protein advantage on Lacuna is real but modest — 0.5467 versus 0.5311 is not a chasm. The counter-evidence is that the mechanism of the advantage is driver-mutation protein effects and post-translational state, which no RNA architecture can invent, and that roughly half of protein variance is unexplained by transcripts A high-stringency blueprint of the human proteome. But the objection can only be settled empirically, which is why the head-to-head ablation (same architecture, same budget, RNA input versus imputed-proteome input) is the decisive experiment, and why I state the claim as falsifiable rather than presuming it.
Why not just wait for patient proteomics and skip imputation? Where paired patient proteomics with treatment outcomes exists, direct measurement should win, and I concede that. The problem is scale: clinical cohorts with high-dimensional baseline omics and standardized responses are tiny — 555 patients across 5 drugs in DeepSADR's TCGA use — while RNA-seq patient cohorts are orders of magnitude larger. Imputation is what makes every existing RNA cohort usable as a proteomic input, and it will be calibrated against genuinely paired patient data (the ovarian cancer cohort combining tumor proteomics with single-cell RNA Integrated proteomics and scRNA-seq analyses of ovarian cancer reveal molecular subtype-associated cell landscapes and immunotherapy targets) so we know exactly how much signal survives the crossing.
Why not just do standard domain adaptation on RNA, GANDALF-style? Because the failure mode is documented: forcing cell lines and patients into one shared latent space discards patient-specific biology, and DeepSADR's authors showed the shift between dish and body is a shift in drug mechanism, not a global distribution mismatch — swapping their adaptive readout for standard pooling cratered clinical AUC from 0.856 to 0.662. GANDALF remains a real threat on the five chemotherapies, so it is run as a live baseline rather than argued away.
Why not just regress RNA onto protein with a simple model? Deterministic regression predicts the conditional mean and over-smooths away the multimodal and heavy-tailed protein states that generative translators like scChord were built to preserve, and protein-level states (a bimodal surface marker, a phosphorylation switch) are precisely the things a response model needs to see rather than average. The ridge-imputer ablation will show whether, for this task, only the mean matters — if it matches the generative imputer, I will report that the fancy layer was unnecessary.
Proposed Experiments
Experiment 1 tests the imputer itself: per-protein Pearson correlation between imputed and measured proteomes on held-out cell lines, and on patient tumors with measured proteomics from the ovarian cohort; the reference point is the raw-RNA-as-proxy baseline, which by the HUPO evidence should track protein only 40–60% of the time A high-stringency blueprint of the human proteome. Experiment 2 tests task-aware imputation in vitro: drug-response prediction on ProCan-DepMapSanger with imputed versus measured versus raw-RNA inputs, all through the same response head, with BDRN's 0.5467 as the measured-protein ceiling. Experiment 3 is the decisive clinical test on the DeepSADR benchmark: does the imputed-proteome transfer beat the reimplemented RNA transfer on mean clinical AUC, with a preregistered receptor-negative stratification. Experiment 4 is the perturbation check: given only transcriptomes of palbociclib-treated cells, does the model predict the protein-level senescence program — Rb-phosphorylation loss, Trop-2 and MHC upregulation — documented by Proteomic characterisation of triple negative breast cancer cells following CDK4/6 inhibition? Success there would mean the model captures the discordant layer, not just correlated abundance, and would justify extending the same machinery to predict which receptor-negative patients are primed for CDK4/6-sensitizing combinations. If Experiments 2 and 3 both fail against the RNA ablation, the honest conclusion is that current paired data cannot pin down the translation code well enough to pay for itself, and that the community should invest in paired patient proteo-transcriptomics rather than in imputation — a conclusion worth publishing either way.
References
- A high-stringency blueprint of the human proteome
- Drug Response Modeling across Cancers: Proteomics vs. Transcriptomics
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- GANDALF: Generative AttentioN based Data Augmentation and predictive modeLing Framework for personalized cancer treatment
- Transfer learning of condition-specific perturbation in gene interactions improves drug response prediction
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- A Joint Model of RNA Expression and Surface Protein Abundance in Single Cells
- Cell-Level Virtual Screening
- Proteomic characterisation of triple negative breast cancer cells following CDK4/6 inhibition
- Integrated proteomics and scRNA-seq analyses of ovarian cancer reveal molecular subtype-associated cell landscapes and immunotherapy targets