Is It Lag or Regulation? Separating Time Delay from True RNA–Protein Discordance

When a cell's messenger RNA (the transcript) and its protein disagree, biologists usually read the gap as regulation: translation control, protein degradation, or modification. In a snapshot measurement, though, much of that gap may simply be time. A protein made from mRNA hours ago can still be present after the transcript has decayed, and a new transcript burst may not have produced protein yet. In cells that are changing state, such as maturing T cells, a large mismatch can arise with no regulation at all. We propose to split each cell's per-protein mismatch into two parts. The first is the part explained by a simple delay along the cell's inferred developmental path. The second is the residual that the delay cannot explain. Only the residual is then passed to pathway enrichment, the standard test for whether a set of genes sharing a biological function is over-represented. We make this split separately within each cell type, so that mixing cell populations cannot create or hide a signal.

The motivation is that "discordance-driven pathways" are becoming an analysis target. The research file notes this explicitly: nobody has run enrichment on the discordance itself. If lag and cell-type mixing contaminate that signal, the first wave of such analyses will name pathways as "post-transcriptionally regulated" when they are only slow, or only enriched in one population. Transcript levels are known to explain only part of protein variance, with correlations of roughly 40–60% reported in A high-stringency blueprint of the human proteome. How much of that unexplained share is time, as opposed to regulation, has not been measured at single-cell resolution.

Core Claim

In single-cell data that pairs RNA with surface-protein counts, and where cells lie along a continuous trajectory, a protein-specific delay explains a measurable fraction of RNA–protein mismatch. Naive enrichment on raw mismatch will preferentially flag pathways whose members are long-lived or sit on dynamic lineages. After removing the delay-explained part within each cell type, the remaining discordance gives an enrichment result that is more stable and less tied to trajectory position.

The deliverable is not a new deep generative model. It is a decomposition and a testing procedure, plus a benchmark showing when the decomposition changes biological conclusions.

How This Differs from Existing Work

Three RNA-to-protein models supply our predictors, and all three treat the mapping as instantaneous. totalVI jointly models RNA and surface-protein counts with a background-noise mixture in a shared latent space. scChord uses a noise-absorbing autoencoder followed by optimal-transport flow matching, a generator that learns to move samples from an RNA embedding to the protein distribution, and it preserves bimodal and heavy-tailed markers such as CD4 and CD45RA. scMHVA fuses the two modalities with attention. Each predicts a cell's protein from that same cell's current RNA, so any delay is absorbed as irreducible error or as spurious "discordance." We reuse them as the instantaneous component and add the delay term on top.

DeepVelo supplies the trajectory side. It learns continuous RNA velocity fields, meaning the direction each cell's transcriptome is moving, using neural differential equations. It also reports that modules extracted along trajectories show GO enrichment two orders of magnitude above static clusters. That result motivates us: trajectory position matters for enrichment. But DeepVelo never sees protein.

Two lines of enrichment work supply the testing machinery, and both ignore time. ActivePathways fuses per-gene P-values across omic layers with Brown's correction and ranked hypergeometric tests. It works on bulk cohort statistics and has no notion of cell type or time. pmVAE and DeepGSEA test pathways within single-cell RNA data, and both are RNA-only and static. We borrow the ranked, covariance-aware testing of ActivePathways and apply it to the per-cell-type residual discordance vector.

Research Program

For protein $j$ in cell $i$, let $m_{ij}$ be the denoised RNA level of its coding gene and $p_{ij}$ the background-corrected protein level from totalVI's foreground component. Let $s_i$ be the cell's position along a trajectory, obtained from pseudotime or from integrating a velocity field. Suppose protein follows first-order kinetics along the trajectory, with synthesis proportional to mRNA and loss at a protein-specific rate. Then the delay-aware prediction is an exponentially weighted average of mRNA over the cell's own past:

$$\hat p_{ij} = a_j + b_j \int_0^{s_i} m_j(s')\, \lambda_j e^{-\lambda_j (s_i - s')}\, ds',$$

Here $m_j(s')$ is the smoothed mRNA level of gene $j$ at trajectory position $s'$, estimated from neighboring cells within the same lineage. The parameter $\lambda_j$ is a per-protein rate in pseudotime units; a small $\lambda_j$ means a long memory. The coefficients $a_j$ and $b_j$ are a per-protein offset and scale. The exponential kernel is simply the solution of the first-order kinetic equation. As $\lambda_j \to \infty$, the model recovers the instantaneous predictor.

The raw discordance is $p_{ij}$ minus the instantaneous prediction. The lag-explained part is the improvement gained by letting $\lambda_j$ be finite. The residual discordance is $p_{ij} - \hat p_{ij}$.

All fitting is done within annotated cell types and lineages. Discordance is then summarized per protein and per cell type, and ranked tests are run against GO and Reactome sets that are restricted to the measured protein panel. This restriction matters because the background must be the panel, not the genome.

Pseudotime has no physical units, so we interpret $\lambda_j$ only relative to other proteins on the same lineage. We do not read it as a half-life in hours. Surface panels cover 13 to 228 proteins, so enrichment power is limited, and that limitation shapes the design. We use the 228-protein PBMC atlas (GSE164378) for real-data enrichment. The smaller panels serve for prediction checks only.

Baselines

The naive reference is instantaneous discordance with no cell-type stratification. We regress protein on its own gene's RNA across all cells, take residuals, and run the same ranked enrichment. This is what an analyst would do today, and it is the point our method must move away from.

The strongest existing predictor is scChord. The file states it achieved the lowest RMSE and highest Pearson correlation against totalVI, sciPENN, and scTranslator. However, no reported numeric RMSE or PCC value was found on Lacuna for scChord on GSE100866 or GSE164378, so we will rerun it ourselves. We will retrain scChord and totalVI on identical train/test cell splits of GSE100866, GSE164378, and the mouse thymocyte CITE-seq data, with equal hyperparameter search budgets. Their instantaneous predictions become the $\lambda_j \to \infty$ component. Beating them means the delay term lowers held-out protein RMSE specifically in cells on dynamic lineages, with no degradation in static populations such as terminally differentiated cells. Winning only on static cells would indicate overfitting rather than lag.

The key ablation is a pseudotime-shift correlation. For each protein, we slide its RNA profile forward along pseudotime and keep the shift that maximizes correlation, with no kinetic model. If this ablation matches the kernel on both prediction and enrichment, the kinetic form adds nothing and we will say so. A second ablation removes within-cell-type fitting, which isolates the contribution of heterogeneity masking.

Why not just…?

Why not just run enrichment on residuals from a strong instantaneous predictor like scChord? This is exactly the naive pipeline we suspect is contaminated. A better instantaneous model reduces noise but cannot represent the past. A long-lived protein on a maturing lineage will stay "discordant" however well the current RNA is modeled. The simulation below tests this directly by checking whether lag-only proteins are falsely flagged under scChord residuals.

Why not just shift RNA along pseudotime and correlate? This is a genuine threat, because it captures most of the delay idea with one parameter. We include it as an ablation rather than arguing against it. The exponential kernel differs only in modeling accumulation, an average over the past, rather than a single past point. That difference should matter for transient transcript pulses. If the data show no difference, the simpler shift becomes our recommended tool, and the paper's contribution is the decomposition and enrichment finding rather than the kernel.

Why not just use RNA velocity and assume protein tracks spliced mRNA? This assumes the delay is zero and the scale is constant. The 40–60% correspondence reported in A high-stringency blueprint of the human proteome leaves too much variance for a constant scale to absorb. Our $\lambda_j \to \infty$ baseline embodies this assumption, so the prediction comparison tests it directly.

Why not just pseudobulk by cell type and run standard enrichment? Pseudobulking addresses mixing but averages cells across trajectory positions. By DeepVelo's result, this dilutes trajectory-linked signal. It is effectively our method with $\lambda_j$ fixed and no ordering within a cell type, and the ablations cover it.

Proposed Experiments

The first decisive check is a semi-synthetic benchmark with known answers. We take real RNA trajectories, from the pancreatic endocrinogenesis and dentate gyrus datasets used by DeepVelo, and generate protein counts. Each synthetic protein receives a known $\lambda_j$. True post-transcriptional regulation is injected only into proteins belonging to a chosen set of pathways, and counts are drawn with totalVI-style background noise. We score every pipeline on precision and recall for recovering the injected pathways. Critically, we also measure its false-positive rate on pathways enriched for small $\lambda_j$ but carrying no injected regulation. The hypothesis fails if naive discordance enrichment does not show elevated false positives on long-memory pathways.

The second check uses real data. On GSE164378 and the thymocyte data, we compare held-out protein prediction for dynamic versus static populations, using the stratified error comparison described in the Baselines section.

The third check tests enrichment stability. We split donors in the PBMC atlas and compare pathway rankings between the halves. Residual discordance should give higher between-half rank agreement and lower correlation with mean trajectory velocity than raw discordance does. Finding a protein-level lag explanation for markers such as CD45RA is exploratory and not a claim made in advance.

The fourth check tests masking. We compare pooled versus within-cell-type discordance enrichment, and we report pathways whose significance appears or reverses only under pooling. These cases are the concrete demonstration that mixing cell populations can manufacture discordance signals.