When a patient with prostate cancer starts androgen-deprivation therapy and his PSA level falls, clinicians read the decline as tumor response. But the drugs that suppress PSA transcription do not necessarily suppress the androgen receptor protein that keeps the tumor alive: in patient-derived xenografts, enzalutamide lowered secreted PSA while AR protein stayed overexpressed, and surgical castration actually induced a 14.5-fold increase in full-length AR inside stem-like tumor cells Human intermediate prostate cancer stem cells contribute to the initiation and development of prostate adenocarcinoma. In other words, the biomarker being tracked and the biology driving resistance are deliberately decoupled by the therapy itself — a clinically weaponized case of the general fact that mRNA levels predict protein levels only 40–60% of the time A high-stringency blueprint of the human proteome.
We propose to estimate, from routine serial bulk transcriptomic measurements, the hidden protein states of the cell subpopulations inside a tumor, and to track how fast those hidden states change under therapy — a quantity we call proteome velocity. The reader can carry the idea in two sentences: deconvolve each bulk sample into cell-state fractions and translate each state's transcriptome into a context-conditioned protein activity profile; then fit a trajectory model across serial samples and use the velocity of inferred AR and bypass-signaling protein activity within the tumor fraction as an early-warning signal of castration resistance, weeks to months before PSA kinetics fail.
Core Claim
The claim has three parts, and each corresponds to a gap that no single paper in the corpus closes. First, current deconvolution methods — Count Bridges, which exactly decomposes bulk RNA counts into single-cell count profiles via Poisson birth-death bridges, and POPPY, which maps bulk transcriptomes into single-cell foundation-model embeddings — stop at the RNA layer; they assume that recovering mRNA fractions recovers biological truth. But Adhikari's 40–60% discordance bound means a method can achieve near-zero deconvolution error in RNA space and still be wrong about the functional protein activity driving the phenotype A high-stringency blueprint of the human proteome. Second, RNA-to-protein translators — scChord, scTranslator, and totalVI — require an isolated single cell as input and are trained on untreated atlases, so they cannot see a therapy-induced 14.5-fold protein shift that has no counterpart in the training distribution. Third, longitudinal trajectory modeling exists only for imaging: IMMFM uses piecewise-quadratic flow-matching paths to forecast 3D MRI series, and MK-UNet tracks CT volume kinetics, while every liquid-biopsy deconvolution tool, including cfDecon, treats each blood draw as an independent snapshot.
Our proposal is the composition: a single estimator whose state is a pair $s_t = (f_t, p_t)$, where $f_t$ is a vector of cell-state fractions on the probability simplex and $p_t$ is the per-state protein activity profile, inferred jointly from bulk RNA at time $t$ and conditioned on the treatment context. The early-warning signal is the projected velocity of the tumor fraction's AR and bypass-pathway protein mass:
$$v^{AR}_t \;=\; \frac{d}{dt}\sum_k f_{t,k}\, p^{\,AR}_{t,k},$$
where $k$ indexes deconvolved cell states. The core prediction is that $v^{AR}$ rises under androgen deprivation while PSA falls, and that this divergence precedes clinical progression — which currently carries a median survival of only 16–18 months once it is diagnosed Comprehensive analysis of the progression mechanisms of CRPC and its inhibitor discovery based on machine learning algorithms.
Research Program
Component one is discordance-aware deconvolution-to-protein translation. We train a model on paired single-cell CITE-seq references (aggregated into pseudo-bulk mixtures with known composition) that maps a bulk RNA profile plus a treatment-context token directly to per-state protein abundances. Two design choices are dictated by documented failures: the translation function is state-specific and context-conditioned rather than global, because discordance is not a constant offset — METTL5-driven m6A modification selectively accelerates CCND3 translation without mRNA change METTL5‑mediated m6A modification of 18S rRNA drives oral squamous cell carcinoma progression by enhancing CCND3 translation, and PRMT5 methylates FUBP1 to accelerate progression through an entirely AR-independent post-translational axis PRMT5-mediated FUBP1 methylation accelerates prostate cancer progression; and protein-protein interaction graph priors regularize the output, because scChord's Correlation Matrix Distance degrades when panels scale beyond hundreds of proteins, indicating that co-expression structure is not learned implicitly scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation.
Component two is a trajectory model for the composite state $s_t$, built by porting IMMFM's piecewise-quadratic conditional probability paths and learned predictive-error diffusion coefficients from 3D MRI voxels to the composition simplex and protein activity space. The port is nontrivial and is the methodological contribution: IMMFM's smoothness machinery was designed for continuous image intensities Longitudinal Flow Matching for Trajectory Modeling, whereas our state mixes constrained fractions with heavy-tailed protein abundances, and clinical sampling is irregular — the exact open problem IMMFM's authors leave and the reason per-timepoint inference is not enough.
Component three is the CRPC application layer, where the biology of plasticity replaces genomics as the source of resistance signal. SubNetDL predicts therapy response from static DNA mutation clusters propagated over interaction networks; Mu et al. proved that CriPCSCs survive enzalutamide without any genetic modification Human intermediate prostate cancer stem cells..., so a mutation-driven subclone model is structurally blind to the dominant resistance mode in this disease. Our signal is phenotypic and longitudinal: the dedifferentiation of luminal cells into AR-overexpressing intermediate stem cells should appear as rising $v^{AR}$ in the deconvolved tumor fraction before PSA relapse.
Proposed Experiments
Study 1 benchmarks the translation layer on CITE-seq pseudo-bulk. We simulate bulk mixtures with known composition and held-out proteins from the PBMC atlas GSE164378 (over 161,000 cells, 228 surface proteins) and the CBMC dataset GSE100866 (roughly 8,600 cells, 13 proteins), reporting RMSE, Pearson correlation, and Correlation Matrix Distance, with a deliberate scale-up arm beyond 200 proteins to test whether graph priors fix the documented CMD degradation. On these benchmarks the strongest prior translator is scChord, which reports "lowest RMSE" and "highest Pearson correlation" against totalVI, sciPENN, and scTranslator on exactly these datasets scChord; no exact numerical RMSE value is printed in the retrieved summaries, so we will re-run all translators ourselves under matched budgets rather than rely on reported figures.
Study 2 benchmarks the trajectory layer on simulated serial mixtures with known drift, irregular sampling intervals, and therapy-switch events, comparing three forecasters: independent per-timepoint inference (the implicit default in cfDecon-style pipelines), pairwise linear flow matching, and our quadratic-path simplex model, measuring forecasting error at 1, 3, and 6-sample horizons. IMMFM's imaging evidence — full-trajectory evaluation lifted disease-classifier accuracy from 71.7% to 80.8% over static single-timepoint evaluation Longitudinal Flow Matching for Trajectory Modeling — motivates the comparison but provides no number in molecular space.
Study 3 is the biological calibration on the PrDX castration series of Mu et al. Human intermediate prostate cancer stem cells...: the inferred AR protein trajectory must reproduce the 14.5-fold castration-induced upregulation and the PSA–AR dissociation under enzalutamide. This is the decisive test of context-conditioning, because a translator trained on untreated CITE-seq atlases has never seen this shift. Note the constraint: primary CriPCSCs senesce after about four weeks in vitro, so longitudinal in vitro series are too short and the xenograft time series is the right substrate.
Study 4 is the clinical evaluation. No longitudinal bulk-RNA cohort for CRPC under androgen deprivation was found on Lacuna — this is a gap we must fill by assembling serial biopsy or liquid-biopsy RNA-seq with matched PSA and outcome records; until then, the static microarray cohorts GSE32269, GSE74367, and GSE66187 serve as a cross-sectional checkpoint, where our deconvolved proteomic features must at least match the static transcript biomarkers CCNA2 (AUC 0.995) and CKS2 (AUC 1.000) reported on training cohorts Comprehensive analysis of CRPC.... The primary endpoint of the assembled cohort is lead time: how many weeks before PSA-defined progression does $v^{AR}$ cross a pre-registered threshold, evaluated by time-dependent AUC and concordance index.
Baselines
The naive reference point is standard PSA kinetics on the same serial samples: fit the established clinical velocity measures to PSA alone and evaluate identical endpoints (lead time, time-dependent AUC). No reported performance number for PSA-kinetics-based progression prediction in CRPC was found on Lacuna, so we treat it as the "do nothing" anchor to be measured, not cited. Beating it means detecting progression strictly earlier at matched specificity.
The second baseline is pseudobulk averaging: average the deconvolved or measured transcriptomes per patient per timepoint and train the same downstream survival/progression model. This is a genuine threat because it is cheap, but the evidence against it is direct — on clinical prediction from single-cell distributions, the set-aware encoder scSet reached a weighted F1 of 0.95 while pseudobulk averaging reached 0.77 in the SLE cohort, because equal-and-opposite subpopulation shifts cancel in the average A Diffusion-Based Autoencoder for Learning Patient-Level Representations from Single-Cell Data. It will be run on identical data, sampling, and tuning budget; beating it means resolving shifts that the average conceals.
The strongest existing competitor is the cascade: run Count Bridges to deconvolve each bulk sample, then run scChord on each deconvolved single-cell profile, per timepoint and independently Count Bridges, scChord. This is the best approximation of our goal buildable from existing parts, and Count Bridges already outperformed CIBERSORTx and MuSiC on proportion recovery (no exact numerical error was reported). It gets the same data, compute, and tuning budget, with scChord's own reported configuration (its ZINB VAE plus optimal-transport conditional flow matching, 229 MB GPU memory footprint). We beat it only if context-conditioning and temporal coupling jointly improve protein RMSE and progression lead time beyond the cascade's, with the PrDX AR-induction study as the sharpest discriminator.
The key ablation removes exactly one ingredient at a time: a context-free translation (same architecture, no treatment token) isolates the discordance-conditioning claim; independent per-timepoint inference with our full translation isolates the trajectory-modeling claim; and an RNA-only variant that tracks composition velocity without any protein layer isolates the claim that modeling discordance matters at all.
Why not just…?
Why not just track PSA kinetics? Because the therapy engineers the biomarker to lie. Enzalutamide suppresses PSA transcription while AR protein persists in the very stem-like clones that drive relapse Human intermediate prostate cancer stem cells..., so a declining PSA trajectory under ARSI is partly a pharmacological artifact. The longitudinal-medicine literature is unanimous that trajectories beat snapshots — serial albumin revealed an S-shaped mortality relationship that admission baseline missed Association between albumin changes and prognosis in older patients with acute myocardial infarction, and hsCRP trajectories beat baseline values for metabolic-syndrome incidence Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome — but trajectory modeling of the wrong marker inherits the marker's blindness. Our design keeps PSA kinetics as the anchor baseline precisely so the lead-time comparison quantifies that blindness.
Why not just deconvolve each sample with existing tools and watch the fractions move? This is the RNA-only ablation, and we will run it rather than argue it away. The evidence for why it falls short is Adhikari's 40–60% mRNA–protein discordance A high-stringency blueprint of the human proteome plus the specific CRPC finding that the resistance-relevant molecule (AR protein) rises while its transcript story diverges under therapy; a composition trajectory in RNA space can be flat while the proteome of the same fraction is rearming. If the RNA-only ablation matches our full model on progression lead time, the discordance component is empirically unnecessary for this task and we would report that — but Study 3, where AR protein moves 14.5-fold under castration, is designed to expose exactly the regime where it cannot.
Why not just cascade Count Bridges and scChord? This is the real threat, and we concede it could carry much of the weight; it is our strongest baseline for that reason. It falls short in two documented ways. scChord's translators are trained on untreated CITE-seq distributions, so a castration-induced 14.5-fold AR induction — a shift absent from training data — sits outside what a static translator can reproduce, which is the specific hypothesis of Study 3. And the cascade treats each timepoint independently, discarding momentum; cfDecon shows the cost of that habit in the liquid-biopsy setting, where every draw is analyzed in isolation cfDecon: Accurate and Interpretable methylation-based cell type deconvolution for cell-free DNA. If the cascade matches our model on every endpoint, the field has learned that the cheap path suffices and our contribution reduces to the trajectory layer — an outcome the ablations are structured to detect honestly.
Decisive Checks
Three results would make or break the proposal. First, on the PrDX series, the context-conditioned model must reproduce the 14.5-fold AR induction and PSA–AR dissociation while the untreated-trained cascade fails — if the cascade also reproduces it, context-conditioning is dead. Second, on the assembled longitudinal cohort, $v^{AR}$ must cross its pre-registered threshold a clinically meaningful interval before PSA-defined progression at fixed specificity; a lead time near zero reduces the method to an expensive restatement of PSA kinetics. Third, the RNA-only ablation must underperform the full model on that same endpoint, establishing that the discordance layer — not just the deconvolution or the trajectory machinery — is what buys the early warning. Each check has a named failure mode, and we commit to reporting whichever one occurs.
References
- Human intermediate prostate cancer stem cells contribute to the initiation and development of prostate adenocarcinoma
- A high-stringency blueprint of the human proteome
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- A pre-trained large generative model for translating single-cell transcriptome to proteome
- Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells
- Longitudinal Flow Matching for Trajectory Modeling
- AI-based non-invasive profiling of the tumor immune microenvironment using longitudinal CT radiomics predicts immunotherapy response in lung cancer
- cfDecon: Accurate and Interpretable Methylation-Based Cell Type Deconvolution for Cell-Free DNA
- Comprehensive analysis of the progression mechanisms of CRPC and its inhibitor discovery based on machine learning algorithms
- METTL5‑mediated m6A modification of 18S rRNA drives oral squamous cell carcinoma progression by enhancing CCND3 translation
- PRMT5-mediated FUBP1 methylation accelerates prostate cancer progression
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- A Diffusion-Based Autoencoder for Learning Patient-Level Representations from Single-Cell Data
- Association between albumin changes and prognosis in older patients with acute myocardial infarction
- Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome:A Retrospective Cohort Study