Stress-Gated Protein Kinetics: Learning the Time-Lagged RNA-to-Protein Map Under Oxidative Stress

When a cell is hit by reactive oxygen species — chemically aggressive molecules like hydrogen peroxide that accumulate during UV exposure, metabolic stress, or vascular injury — its RNA story and its protein story come apart. The clearest demonstration is three decades old and still unresolved methodologically: in human skin, UV light raises c-Jun mRNA six-fold, and topical retinoic acid leaves that mRNA surge completely intact while cutting c-Jun protein by roughly 70% (Fisher et al.). Any method that reads only RNA — which includes essentially every single-cell dynamics model in use today — would conclude the drug failed. A second example drives the point deeper: under oxidative stress, the protective protein TERT is destroyed by the ubiquitin ligase UBR5 precisely when the cell needs it most, with no change in its transcript (Zhao et al.). Discordance between transcript and protein is not a niche problem; mRNA levels explain only 40–60% of protein variance in human cells (Adhikari et al., A high-stringency blueprint of the human proteome).

The methodological situation is a clean, four-way stalemate. RNA-to-protein predictors (scChord, scTEL, PIKE-R2P) learn a fixed, instantaneous map from RNA to protein and have never been tested under oxidative stress or across time. Single-cell kinetic models (FLeCS, cell2fate) model dynamics entirely in RNA space, with no protein layer at all. Deconvolution tools (Count Bridges, POPPY) unmask hidden cell types but only at the mRNA level. And the oxidative-stress biology itself rests on bulk endpoints that average over distinct cell lineages. This proposal builds the missing piece at the intersection: a kinetic model in which the RNA-to-protein map is itself a dynamical object, with a translation delay and a degradation rate that oxidative stress actively widens.

Core Claim

The claim is that under oxidative stress, the RNA-to-protein relationship is not a function $p(P_t \mid R_t)$ to be fit once, but a time-lagged, stress-gated dynamical system: protein lags RNA by a translation delay, and the rate at which specific proteins are destroyed rises with the cell's stress level — which is exactly what the UBR5–TERT and retinoic-acid–c-Jun results demand and what every current model forbids. If we learn this system jointly with per-cell-type structure, then (i) RNA-only time courses predict protein trajectories hours ahead, including post-transcriptional drug effects invisible to transcriptomics; (ii) bulk organoid measurements can be resolved into cell-type-specific protein dynamics, not just mRNA proportions; and (iii) we can predict, in silico, which interventions rescue protein-level outcomes before any wet-lab screen.

Research Program

The model — call it the kinetic discordance translator — extends the learnable-ODE approach of FLeCS, which simulates gene dynamics directly in expression space with regulatory-network priors and constant-memory adjoint gradients, by adding a protein layer whose kinetics are deliberately asymmetric. For each cell type $k$, with mRNA vector $x_k(t)$ and protein vector $p_k(t)$:

$$\frac{dx_{k,i}}{dt} = f_{k,i}(x_k) - \gamma_i\, x_{k,i}, \qquad \frac{dp_{k,i}}{dt} = a_i\big(x_k(t - \tau)\big) - d_i \big(1 + m_i\, s(t)\big)\, p_{k,i}$$

Here $f$ is a learnable production function constrained by regulatory-network priors as in FLeCS; $\tau$ is a translation lag; $a_i$ maps delayed mRNA to protein synthesis; $s(t)$ is a stress-activity scalar inferred from ROS-responsive transcripts (the Nrf2 target genes such as HMOX1 and NQO1 serve as a built-in stress gauge); and $m_i$ is the key new ingredient — a per-protein susceptibility of proteasomal degradation to stress. A protein like TERT, targeted by a stress-activated E3 ligase (a cellular machine that tags proteins for destruction), gets large $m_i$; a stable structural protein gets $m_i \approx 0$. This single parameter is what lets protein fall while mRNA holds, the signature both founding observations share. Bulk or organoid readouts are handled by treating single-cell states as latent variables in an expectation-maximization loop with count-conserving projection, reusing the discrete machinery of Count Bridges, so that integer count geometry is respected rather than blurred.

Three data streams feed the program. First, the public CITE-seq benchmarks (a technology measuring RNA and surface proteins in the same single cell): cord blood mononuclear cells from GSE100866 (~8,600 cells, 13 proteins) and the PBMC atlas of GSE164378 (161,764 cells, up to 228 proteins), used to verify that the model's instantaneous head matches static state of the art before any temporal claims are made. Second, existing temporal scRNA-seq — the 72-hour A549 drug-response course and K562 CRISPRi data used by FLeCS — to validate the mRNA dynamics layer in isolation. Third, and decisively, a new experiment we would run: a time course of human endothelial cells and hiPSC-derived vascular organoids under hydrogen peroxide and oxidized LDL, sampled at Fisher-style time points (15 minutes, 1, 4, 8, 24 hours), with paired single-cell RNA, a surface-protein panel, and targeted intracellular quantification of c-Jun, HO-1, and TERT — the low-abundance proteins that motivated the problem (Zhao et al.). Intervention arms (proteasome inhibition and UBR5 knockdown) are included not as endpoints but as identifiability levers: they decouple degradation from transcription, which is what pins down $m_i$ and $\tau$ separately. Where possible we reuse the released FLeCS codebase as the mRNA-layer scaffold.

Proposed Experiments

The first experiment is a static sanity check on CBMC and PBMC: the model must recover the multimodal protein distributions scChord is known for (bimodal CD4, heavy-tailed CD45RA) at RMSE, Pearson correlation, and correlation-matrix-distance parity with scChord and scTEL. The file records that scChord achieved the lowest RMSE and highest Pearson correlation against totalVI, sciPENN, and scTranslator on these benchmarks, and that scTEL reached 86.6% cell-type classification accuracy versus 77.0% for sciPENN and 76.1% for Seurat, but no numeric RMSE or PCC values are quoted for either system; we will establish those numbers ourselves on identical splits, and where we cannot compare to a published figure we will say so — no reported result found on Lacuna for numeric scChord or scTEL protein-prediction errors on CBMC/PBMC.

The second and central experiment is temporal generalization: train all methods on time points up to 4 hours of oxidative stress, predict protein states at 8 and 24 hours from RNA alone. The hypothesis, grounded in Fisher et al., is that static translators will systematically mispredict proteins whose degradation is stress-gated, while the kinetic model holds. We quantify this as the gap in per-protein RMSE between held-out late time points and in-distribution early ones, and separately by whether each method reproduces the two canonical signatures: ~70% protein suppression with intact six-fold mRNA under the translation-blocking arm, and TERT loss under peroxide that UBR5 knockdown rescues. The third experiment tests heterogeneity: in mixed organoids (endothelial tubes plus pericytes, as in Zhao et al.), we compare the mixture model against a pooled model and against two-stage pipelines that deconvolve first and translate second, scoring per-cell-type protein trajectory accuracy. The fourth is the payoff demonstration: given only an RNA trajectory and a proposed perturbation (UBR5 inhibition), predict the protein-level outcome before it is measured, then verify against the intervention arm.

Baselines

The naive reference point is using RNA itself as the protein proxy, plus per-protein mean imputation and ridge regression. This is not a strawman: it is what most transcriptomic studies implicitly do, and PIKE-R2P showed that even a random forest on raw counts loses to network-constrained prediction and blurs functionally distinct T-cell subsets, while scChord showed that conditional-mean regression collapses genuinely bimodal protein distributions. The strongest existing competitor is scChord, the current benchmark leader on RMSE and Pearson correlation over totalVI, sciPENN, and scTranslator; because the file quotes no numeric value for those wins, our comparison will re-run scChord on our exact splits with matched compute and tuning budget (scChord's reported 27-minute, 229 MB footprint on 100,000 cells means compute parity is cheap to honor). scTEL joins as the strongest multi-dataset system, with its published 86.6% classification accuracy as the number to beat on any cell-type task. Both are run two ways: trained once on pre-stress data and applied across time (the honest static deployment), and retrained per time point (the generous static deployment, which uses information our model never sees). The critical ablation is our own model with the translation lag fixed to zero and stress-susceptibility $m_i$ fixed to zero, which reduces it to a per-cell-type static translator and isolates exactly the ingredient the proposal stands on. Beating the baselines means, concretely: lower protein RMSE at held-out late time points under stress, recovery of the two canonical mRNA-protein divergence signatures where static methods fail, and no loss on the static CBMC/PBMC benchmarks.

Why not just…?

Why not just run scChord or scTEL separately at each time point? This is the most serious alternative, and the per-time-point retrained variant uses strictly more measurement information than our model. It falls short on the deployment we actually care about: predicting protein at time points where only RNA was measured, which is the realistic case in time courses (protein panels are expensive and, for intracellular targets like TERT, often unavailable). The Fisher result shows the conditional map itself is rewritten by intervention, so a map fitted at hour 4 is wrong at hour 8 in a direction no retraining at hour 4 can fix. Experiment two settles this directly, and honestly: if per-time-point static models match our held-out predictions, the temporal machinery is unjustified and we would report that.

Why not just deconvolve with Count Bridges or POPPY and then translate? A two-stage pipeline is a real threat and we run it as a baseline. But both stages lose information at the exact seam this proposal targets: Count Bridges deconvolves only integer mRNA counts, and POPPY attributes bulk profiles only against cell states already present in its reference atlas — and stressed, transitional endothelial states are precisely what atlases lack. Composition also drifts under stress, so a translator fitted on the deconvolved early mixture is applied to a different late mixture. Our joint model estimates cell-type proportions and protein kinetics together; experiment three, comparing per-type protein accuracy, is where the difference should appear.

Why not just use RNA velocity (cell2fate) and map proteins on afterwards? cell2fate improved on ten velocity methods by decomposing splicing kinetics into modules, but it works purely in RNA space and assumes degradation rates constant across lineages — an assumption the UBR5 result violates at the protein layer by construction, since stress turns degradation on for specific substrates. A velocity-plus-static-head hybrid is included as a baseline; it should fail specifically on stress-gated proteins, and that failure would be evidence for the model's central term rather than an embarrassment to hide.

Why not just measure the proteins directly at every time point? For the surface panel, one should — and we do. But intracellular, low-abundance targets are the ones that matter here: Zhao et al. note TERT sits at extremely low baseline levels requiring specialized detection, and proteome-scale single-cell measurement over dense time courses remains impractical. The model's value is filling the protein layer from RNA-only time courses exactly where measurement cannot go, validated at the sparse points where it can.

Decisive Checks

The proposal stands or falls on three falsifiable outcomes. First, if static translators trained per time point match the kinetic model's held-out protein predictions under stress, the delay-and-degradation machinery is unnecessary. Second, if ablating $m_i$ (stress-gated degradation) leaves the TERT and c-Jun signatures reproducible, the key ingredient is doing nothing and the model is an expensive regression. Third, if per-cell-type protein trajectories in mixed organoids are no better than the pooled model's, the heterogeneity component is decorative. Each check is built into the experimental plan above with a pre-specified metric, and any of them failing would kill the specific claim it tests — which is the standard a proposal at this intersection should be willing to meet.