When a cell responds to a drug or differentiates, the messenger RNA (mRNA) for a gene and the protein made from it often move apart. Protein lags behind mRNA, translation can shut down, and degradation can speed up. Mass spectrometry proteomics, which measures thousands of proteins by weighing their fragments, can see this divergence. The standard way to read a time course of such data still compares each time point to baseline, collects the changed genes into a list, and asks which biological pathways are over-represented in that list. This approach throws away the shape of the trajectory. It also says nothing about whether the protein went where its mRNA predicted.
We propose a pathway test that asks a different question. Instead of asking "did this pathway's proteins change over time?", it asks "did this pathway's proteins depart from what their own mRNA trajectories, passed through simple synthesis-and-decay kinetics, would predict?" The test treats proteins that the mass spectrometer fails to detect as values below a detection limit, not as random blanks. It uses ELISA, an antibody-based assay that measures protein concentration in absolute units, as an independent check on both the detection-limit model and the flagged trajectories. The output is a ranked list of pathways under post-transcriptional control over time, together with the time window in which the control acts.
Core Claim
Existing tools each cover part of this question.
- Time-course gene set analysis. TcGSA fits mixed-effects spline models (models with gene-level random variation) to repeated transcript measurements and finds dynamic gene sets that snapshot testing misses. It was built for microarrays and RNA sequencing. It has no notion of a protein lagging its mRNA, and no handling of mass spectrometry missingness.
- Multi-omic pathway integration. ActivePathways merges gene-level P-values across omics layers with Brown's correlated extension of Fisher's combined test. Its design is cross-sectional. It rewards agreement between layers, whereas we look for disagreement.
- Kinetic proteomics. ApoE kinetic proteomics and the bortezomib time-resolved study measure turnover directly with heavy isotope labels. Their pathway interpretation is manual or absent.
Our claim is that a pathway-level likelihood ratio test can recover the pathways whose synthesis or degradation genuinely changes over time, using only paired mRNA and protein abundance time courses. The test needs two parts: a global-translation-adjusted kinetic null, and a censored likelihood for missing values. We validate it against isotope-labelled turnover measured on the same samples but withheld from the method. If that recovery holds, cheap abundance-only designs can be mined for post-transcriptional regulation.
Research Program
For gene $g$ with mRNA trajectory $R_g(t)$, the null model for its protein $P_g(t)$ is
$$\frac{dP_g}{dt} = k^s_g\, G(t)\, R_g(t) - k^d_g\, P_g(t),$$
where:
- $k^s_g$ and $k^d_g$ are gene-specific synthesis and degradation constants that do not change over time;
- $G(t)$ is a single global translation-capacity curve shared by all genes.
The global term is borrowed from the bortezomib study. That study found a roughly 40% global drop in translational efficiency, which ribosome profiling missed. Without $G(t)$, such a drop would make every pathway look discordant.
The alternative model lets the genes in a pathway $S$ share a time-varying spline deviation in synthesis or in degradation. Genes within the pathway get random effects, following TcGSA's way of handling heterogeneity inside a gene set. The mRNA trajectory is smoothed with splines before being fed through the equation, so the kinetic lag is built into the null.
Missing protein values enter the likelihood as left-censored observations. Each is assumed to lie below a run-specific detection limit, estimated from the intensity distribution. This matters because shotgun mass spectrometry drops low-abundance proteins preferentially, so missing values are not random. Pathway significance comes from a likelihood ratio test. Null calibration uses permutation within pathways matched for size and abundance. Each significant pathway is reported with three things: whether the deviation is in synthesis or degradation, the time window where it peaks, and an evidence breakdown in the spirit of ActivePathways.
ELISA plays two roles.
- Checking the detection-limit model. For a panel of about a dozen proteins spanning the mass spectrometry abundance range, ELISA on the same lysates measures absolute concentration. The panel deliberately includes proteins that drop out at some time points. If the censoring model is right, ELISA values at mass-spectrometry-missing points should fall below the inferred detection limit, in the predicted rank order.
- Checking flagged trajectories. Wherever an antibody pair exists, the ELISA trajectory of a hub protein from a flagged pathway should match the model's fitted protein trajectory. It should not match the mRNA-predicted one.
Baselines
Naive reference. Per-time-point differential testing of proteins against time zero, followed by Fisher's exact over-representation analysis. The background is restricted to detected proteins, and results are combined across time points by taking the minimum adjusted P-value. This is what most proteomics time-course papers do, and it sets the floor.
Strongest existing dynamic method. TcGSA applied directly to protein log-abundances, with missing values filled by the minimum observed intensity. It is the closest competitor because it already models pathway trajectories with random effects. On DALIA-1 (an HIV vaccine-trial microarray time course) the TcGSA paper reports 69 significant dynamic gene sets, against 0 for univariate testing and 0 for time-series gene set enrichment analysis.
Strongest integrative method. ActivePathways, fed per-gene P-values from separate mRNA and protein time-trend tests. Its METABRIC analysis reports 192 prognostic pathways, 33 of them visible only through integration.
Ratio baseline. TcGSA run on the log protein-to-mRNA ratio. It is the cheapest direct test of discordance.
Ablations. Three ablations isolate the proposal's ingredients:
- removing the kinetic lag, which reduces the model to the ratio model;
- removing $G(t)$;
- replacing censoring with minimum-value imputation.
Fair comparison. All methods see identical data, gene set libraries (GO Biological Process and KEGG), and minimum set size. Each gets the same spline-degree search budget, and all use the same permutation-based false discovery rate (FDR) calibration.
What counts as beating them. We have found no result on Lacuna for any pathway method scored against turnover ground truth in a proteomics time course, so there is no published number to beat. Success means a higher area under the precision-recall curve than every baseline for recovering turnover-defined positive pathways, with FDR held at nominal level under permutation.
Why not just…?
Why not just run TcGSA on protein abundance? TcGSA detects pathways whose proteins change. Most such pathways change because their mRNA changed, so its top hits will be dominated by transcriptional programs that are already visible in RNA. Our experiment scores both methods against withheld turnover ground truth. We expect TcGSA to recall well but have poor precision for post-transcriptional pathways. If it matches us, the kinetic null adds nothing, and we will say so.
Why not just use ActivePathways on mRNA and protein P-values? Brown's method combines evidence that both layers moved. A pathway whose protein falls while its mRNA stays flat gets weaker combined support than one where both move together, which is the opposite of what we are after. The comparison will show this directly. The 33 integration-only pathways in METABRIC show that integration finds convergent signal, not discordance.
Why not just test the protein/mRNA ratio? This is the real threat. If lags are short relative to sampling intervals, the ratio is nearly equivalent to our model. Under longer lags, however, the ratio swings transiently even with constant kinetics, which produces false positives. The ratio ablation on the bortezomib data, with its 48-hour span and known 40% global shift, settles it. If the ratio performs as well, we will drop the kinetic layer and reframe the contribution around the censoring and global-capacity terms.
Why not just measure turnover with isotope labelling? When pulsed labelling is affordable, it is the better measurement, and we use it here as ground truth. The bortezomib study, though, covered only 272 targeted peptides. The point of our method is to extract comparable conclusions from the far more common abundance-only designs.
Proposed Experiments
Benchmark on existing data. The first benchmark is the bortezomib MM1.S time course. The method receives mRNA-seq and protein abundance only. The pulsed SILAC (isotope-label) synthesis and degradation estimates are withheld, and pathways with significant time-varying deviations in those estimates define the positives. That study reports a baseline correlation of R = 0.80–0.84 between ribosome density and synthesis, which tells us baseline synthesis is well determined there.
Because 272 peptides support few pathways, we add a semi-synthetic benchmark. Kinetic deviations of known size and timing are injected into real mRNA trajectories. Missingness is imposed using the empirical detection patterns of a deep TMT dataset (TMT is isobaric tagging, which pools many samples in one mass spectrometry run). This benchmark lets us map power against effect size and lag.
Transcript-side sanity check. DALIA-1 checks that the mixed-effects machinery reproduces TcGSA's behavior when the protein term is switched off.
New data collection. We will run a deep TMT time course, with matched RNA-seq and parallel pulsed SILAC, in MM1.S under bortezomib at roughly eight time points over 48 hours. A twelve-protein ELISA panel will be run on the same lysates. Pulsed SILAC here provides proteome-wide ground truth.
Decisive checks. The method succeeds only if all four hold:
- it beats the ratio ablation on precision-recall against turnover ground truth;
- its FDR stays at nominal level under permutation;
- ELISA values at mass-spectrometry-missing points fall below the inferred detection limit more often than chance, while minimum-value imputation misranks them;
- ELISA trajectories of hubs from flagged pathways track the fitted protein curve more closely than the mRNA-predicted curve.
References
- Time-Course Gene Set Analysis for Longitudinal Gene Expression Data
- Integrative pathway enrichment analysis of multivariate omics data
- Quantitative and Kinetic Proteomics Reveal ApoE Isoform-dependent Proteostasis Adaptations in Mouse Brain
- Time-resolved proteomics vs. ribosome profiling reveals translation dynamics under stress