Protein Remembers RNA: Lag Kinetics for Single-Cell Transcript–Protein Discordance During Metabolic Switching

When a cell measured by single-cell sequencing shows high messenger RNA for a gene but little of the matching protein, current models treat the gap as noise or as a fixed, mysterious translation factor. A large part of that gap may be simpler: time. A protein's level today reflects how much of its RNA was present over the past hours and how fast the protein is broken down. RNA and protein disagree most while cells are changing state. Metabolic reprogramming is a clear case. Under stress or activation, cells shift from mitochondrial respiration toward glycolysis, the fast, oxygen-independent breakdown of glucose, and the transporters and enzymes involved are induced on a timescale comparable to their protein lifetimes. We propose to predict each cell's protein levels from the reconstructed history of its RNA, not only its current RNA.

The mechanism is borrowed from spiking neural networks. There, the voltage of a leaky neuron is an exponentially decaying sum of past inputs, and recent work computes that sum in parallel across all time steps instead of one step at a time. The first-order law for protein abundance (made in proportion to RNA, removed at a constant rate) has exactly this form. The same parallel machinery therefore fits per-protein decay rates across long cell trajectories, for hundreds of proteins at once. The output is a predictor whose errors should shrink specifically in transitioning cells, along with interpretable per-protein lag constants.

Core Claim

The discordance problem is motivated by atlas-scale evidence. A high-stringency blueprint of the human proteome maps proteins across tissues and finds that mRNA tracks protein abundance only 40–60% of the time. Existing single-cell translators all condition on a single snapshot:

  • totalVI is a variational autoencoder for paired RNA and surface-protein counts. It models protein background and dropout.
  • PIKE-R2P is a graph network over protein–protein interaction priors.
  • scChord is a flow-matching model. It transports a smoothed RNA latent space onto protein distributions and recovers bimodal markers such as CD4.

None of them can represent the fact that two cells with identical current RNA can differ in protein because one has just induced the gene and the other has held it high for hours. scChord's "time" is an artificial integration coordinate, not biological time.

Our claim is that, for proteins whose lifetime is comparable to the duration of a state transition, a model given the cell's RNA history will beat any model given only current RNA. It will also beat a model given current RNA plus a time stamp. The gain should concentrate in cells mid-transition, which a direct test can check.

The kinetic core, per protein $g$, is

$$\frac{dp_g}{dt} = \kappa_g\, f_g(\mathbf{m}(t)) - \delta_g\, p_g(t).$$

In this equation:

  • $p_g$ is the latent protein abundance.
  • $\mathbf{m}(t)$ is the RNA state along the cell's reconstructed ancestry.
  • $f_g$ is a small learned network giving translational drive (it can absorb regulation such as the selective, RNA-independent translation boost shown in the METTL5–CCND3 study).
  • $\kappa_g$ is a translation rate.
  • $\delta_g$ is the decay rate.

Holding $\mathbf{m}$ constant over each trajectory interval of length $\Delta_k$ makes the update exactly affine:

$$p_{g,k} = a_{g,k}\, p_{g,k-1} + b_{g,k}, \quad a_{g,k} = e^{-\delta_g \Delta_k}, \quad b_{g,k} = \tfrac{\kappa_g}{\delta_g}\left(1-e^{-\delta_g\Delta_k}\right) f_g(\mathbf{m}_k).$$

Compositions of affine maps are associative, so all $K$ steps can be evaluated with a parallel scan in $O(\log K)$ depth. This is the construction Bullet Trains uses for sub-threshold neuron dynamics, where it reported up to a 44x speedup on Spiking Heidelberg Digits. We need only the linear, reset-free part of that method, which sidesteps its main limitation: speculative chunking for spike resets and restriction to feedforward topologies.

The scan is not the scientific claim; it is what makes per-protein exact solutions cheap enough to fit across 100,000-cell trajectories with hundreds of proteins. The claim is the history dependence.

Research Program

Reconstructing each cell's RNA history. Sequencing destroys the cell, so each cell's past must be reconstructed. We use two routes:

  • Where only one time point exists, we order cells along an inferred trajectory (pseudotime) and define a cell's history as the averaged RNA of cells upstream of it.
  • Where multiple real time points exist, we use optimal-transport couplings between time points. Alternatively, we roll RNA backward with a gene-network dynamics model such as FLeCS. FLeCS fits coupled ODEs over regulatory graphs (code) and simulated 72-hour A549 drug-response trajectories, but contains no protein equations. Our layer adds exactly those equations downstream of it.

Likelihood. The scan outputs $p_g$ at the measured cell. This value enters the totalVI-style protein count likelihood with a background component, so technical noise is handled as in established tools rather than absorbed into the lag constants.

Data regimes. We work in three regimes of increasing biological realism.

The first is simulation. We generate RNA trajectories with known $\delta_g$ and test whether lag constants are recoverable from destructive snapshots. We also measure how history-reconstruction error biases them.

The second is the existing static CITE-seq atlases in the file: GSE164378 PBMC with 228 surface proteins, GSE100866 CBMC, and PIKE-R2P's PBMC. We restrict attention to trajectories that involve metabolic reprogramming. The main example is T-cell activation, which is accompanied by a shift toward glycolysis. Within it we focus on panel proteins tied to nutrient uptake or activation, where present. Here time is pseudotime, so $\delta_g$ is identifiable only in pseudotime units. We state this limitation plainly.

The third is a targeted new dataset. It would be a real-time course of paired RNA and protein in cells under hypoxia, with intracellular staining for glycolytic gatekeepers: hexokinase 2, GLUT1, PDHK1, and phosphorylated pyruvate dehydrogenase. These enzymes are reported transcriptionally in hypoxia studies such as He et al. and the HIF-1α–JFK work of Yang et al., and their control is described biochemically by Larrieu et al.. The dataset is where the metabolic hypothesis from Bao et al. can be tested in real hours. Bao et al. review evidence that cells oscillate between glycolysis and respiration rather than switching once.

We keep phosphorylation-controlled activity, the PDHK gate on pyruvate dehydrogenase, out of the core model. It is a separate, switch-like layer and would confound the abundance claim.

Baselines

Naive reference. The naive reference is per-protein ridge regression from current RNA. It anchors how much any model gains. The single-cell benchmark of Luecken et al. found that simple linear models often beat deep models on noisy tasks, so this baseline is not a formality.

Strongest snapshot method. The strongest snapshot method is scChord. It reportedly outperforms totalVI, sciPENN and scTranslator on RMSE and PCC on both CBMC and the 228-protein PBMC atlas. The file reports no numeric RMSE or PCC for it: no reported result found on Lacuna for scChord's RMSE/PCC values on GSE100866 or GSE164378. The only figures reported are training on 100,000 cells in about 27 minutes with a 229 MB GPU footprint, which we use as the compute envelope. totalVI and PIKE-R2P are run as additional snapshot baselines. PIKE-R2P likewise reports only that it beats Random Forest and cTP-net, with no numbers on Lacuna.

Fair comparison. All methods use the same cell splits, the same denoised RNA input, matched hyperparameter search budgets (equal trial counts), and a GPU budget no larger than scChord's.

Ablations isolating history. Three ablations isolate the key ingredient:

  • a static model given current RNA plus pseudotime or real time as an extra input;
  • an MLP given a fixed window of lagged upstream RNA states, which tests whether history helps even without the kinetic form;
  • our model with $\delta_g$ forced very large, which collapses it to an instantaneous mapping.

What counts as beating them. Beating the baselines means a statistically significant improvement in per-protein PCC and RMSE over scChord, measured on held-out cells in the top quartile of trajectory speed. It also requires no loss in steady-state cells, and the gain must disappear in the large-$\delta_g$ ablation.

Why not just…?

Why not just give scChord or an MLP the pseudotime as an extra input? This is the most direct threat. Time alone tells the model where a cell is, not how fast it got there. Two cells at the same pseudotime that arrived along branches with different induction histories will receive identical predictions. The time-as-feature ablation measures this gap directly. If it matches our model, history adds nothing beyond position, and we would report that the discordance is state-dependent rather than lag-dependent.

Why not just feed a window of lagged RNA into a flexible network? This could match accuracy, and if it does, the honest conclusion is that history matters but the kinetic form does not. We expect the structured model to win in two places. The first is forecasting held-out later time points, where an unconstrained window has no reason to extrapolate. The second is interpretability: one lag constant per protein, checkable against known protein lifetimes. The lagged-window ablation is run head to head on both.

Why not just extend FLeCS with protein ODEs and its adjoint solver? That would work in principle. Because the protein layer is linear given RNA, our exact scan gives the same answer without numerical integration error and with logarithmic rather than linear depth. The advantage is computational, and we will report wall-clock time against an adjoint-ODE implementation instead of claiming more. scChord's own authors flag multi-step ODE latency as their main limitation, which makes this cost relevant.

Proposed Experiments

Model-free precheck. Before any modeling, we compute the cross-correlation between each protein and its RNA along pseudotime in the PBMC atlas. If no protein shows a peak at a positive lag, with protein trailing RNA, the premise fails on existing data and we go straight to the hypoxia time course.

Simulation. We test whether recovered $\delta_g$ falls within a set tolerance of the true value as history reconstruction degrades.

Main benchmark. On real atlases we report RMSE, PCC and the Correlation Matrix Distance (CMD), the metric on which scChord degrades beyond 200 proteins. Results are stratified by trajectory speed. The decisive signature is that our advantage over scChord grows with trajectory speed and with estimated lag.

Hypoxia time course. On the new data, we train on early and late time points, forecast intermediate ones, and compare against every baseline. We then test whether glycolytic enzymes carry longer lags than the panel median. A positive result would mean that snapshot RNA systematically misreports the metabolic state of cells caught mid-switch.