Which Pathways Are Actually Executed? Fidelity-Weighted Enrichment for Cross-Platform Immune Profiling

When a vaccine or an infection reprograms an immune cell, the cell's RNA and its proteins tell different stories. Transcript abundance predicts actual protein abundance only about half the time ($r \approx 0.40$–$0.60$), because translation rates, RNA decay, and protein turnover decouple the two layers (A high-stringency blueprint of the human proteome; Proteomizer: Leveraging the Transcriptome-Proteome Mismatch to Infer Novel Gene Regulatory Relations). Yet the standard way immunologists interpret such experiments — functional enrichment analysis, which asks whether whole biological programs (pathways) are coordinately switched on rather than eyeballing thousands of individual genes — is fed almost exclusively RNA. The result is that enrichment reports transcriptional intent, not functional execution: it can declare a T-cell activation program "enriched" even when the proteins that carry out that program were never made, because phosphorylation switches and translational control are invisible to RNA sequencing (Proteomics, pathway array and signaling network-based medicine in cancer).

This matters acutely in cross-platform immune benchmarking, where vaccine decisions increasingly rest on immunological measurements rather than new placebo-controlled trials. The clinical literature has converged on a striking principle: when antibody measurements from different platforms are anchored to a shared reference cohort, the relationship between titer and protection converges into one smooth, platform-agnostic gradient — but fixed physical conversion factors fail, with commercial assays showing divergent decay trajectories even after WHO unit conversion and up to 50-fold raw discrepancies (Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection; Studying temporal titre evolution of commercial SARS-CoV-2 assays reveals significant shortcomings of using BAU standardization for comparison). No analogous discipline exists at the cellular level: single-cell multi-omic models that quantify RNA–protein discordance (scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation; A joint analysis of single cell transcriptomics and proteomics using transformer; Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells) are benchmarked on healthy or observational cohorts and never tested for whether their discordance signatures survive a change of assay panel, sequencing platform, or laboratory — the exact failure mode that broke BAU standardization in serology.

Core Claim

I propose an enrichment method that separates pathway intent from pathway execution by weighting each gene's transcriptomic evidence by its measured transcript-to-protein fidelity, and a benchmark protocol that tests whether the resulting pathway signatures are stable across assay platforms the way anchored serological titers are. The method has three parts. First, from paired RNA–protein measurements (bulk TCGA/CPTAC samples or single-cell CITE-seq, where antibodies tagged with oligos let RNA and surface protein be read from the same cell), estimate a per-gene fidelity coefficient:

$$\rho_g = \mathrm{cor}\big(\log x^{R}_{g,\cdot},\; \log x^{P}_{g,\cdot}\big), \qquad p_g = \mathcal{B}\big(p^{R}_g,\; p^{P}_g;\; \rho_g\big)$$

where $\rho_g$ is the across-sample correlation between gene $g$'s log RNA and log protein abundance, $p^{R}_g$ and $p^{P}_g$ are its differential-expression $p$-values in each layer, and $\mathcal{B}$ is Brown's extension of Fisher's combination test — the machinery Integrative pathway enrichment analysis of multivariate omics data introduced for merging omic layers — modified so that a gene's RNA evidence is down-weighted when its fidelity is low, forcing the protein layer to speak for genes whose transcripts are poor proxies. Second, the fused $p_g$ values feed a regularized regression in the style of Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression (pareg), whose network fusion penalty on overlapping pathways prevents the redundant, correlated term lists that plain overrepresentation analysis produces. Third, the model reports a continuous execution gradient per pathway rather than a binary significant/not-significant call — mirroring the clinical finding that protection is a smooth function of titer, not a threshold (Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection) — so a pathway can be flagged as "high intent, low execution," the signature of translational arrest.

The cross-platform dimension is the evaluation regime and the stress test, not a new model component. Exactly as Studying temporal titre evolution of commercial SARS-CoV-2 assays... showed that fixed BAU factors fail over time while Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection showed that anchoring to a reference arm's geometric mean succeeds, I will test whether pathway-level discordance signatures computed from different antibody panels and sequencing platforms can be brought into agreement by anchoring to a shared reference cell population — and whether the fixed-factor alternative breaks.

Research Program

What gets built is modest and mostly reuses validated components. The fidelity estimator $\rho_g$ is a straightforward per-gene correlation on paired data, but it must be computed in a noise-aware way for single cells: raw CITE-seq counts carry RNA dropouts and nonspecific antibody background, so $\rho_g$ will be estimated on denoised quantities, using the probabilistic decoders of scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation (a zero-inflated negative-binomial model that absorbs dropouts into explicit noise parameters) or Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells (which separates true protein signal from antibody background with a two-component mixture). The fidelity-weighted Brown combination is a small modification of an established statistic, not a new architecture. Where protein measurements are missing entirely — different laboratories run different antibody panels — I will impute them with the masked-panel training trick of A joint analysis of single cell transcriptomics and proteomics using transformer (scTEL), which learns from the union of panels via an indicator matrix, and crucially, imputation confidence will itself be gated by $\rho_g$: proteins whose transcripts track them faithfully can be imputed; proteins with low fidelity cannot, and those genes fall back to protein-only or are flagged as unmeasured. This is the direct answer to the generalization collapse that Proteomizer: Leveraging the Transcriptome-Proteome Mismatch to Infer Novel Gene Regulatory Relations reports when its translation model meets unseen tissues or protocols.

The scientific payoff is a map of which immune pathways are governed transcriptionally versus post-transcriptionally, and whether that governance is stable across platforms. If, for example, interferon-response genes turn out to be high-fidelity while cytotoxic-granzyme genes are low-fidelity in circulating immune cells, then RNA-only enrichment studies of vaccine response are systematically biased in a predictable, correctable direction — an immunological analogue of the 62-fold precision gain Proteomizer: Leveraging the Transcriptome-Proteome Mismatch to Infer Novel Gene Regulatory Relations demonstrated for differential expression once the proteome was modeled.

Proposed Experiments

The first experiment is a controlled synthetic benchmark built by extending the simulation design of Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression: dysregulated pathways are planted at varying noise levels and pathway similarities, and I additionally simulate transcript–protein decoupling by letting a chosen fraction of member genes' protein abundances drift independently of their transcripts. The metric is area under the precision–recall curve for recovering the planted pathways. On its own benchmark, pareg achieved the highest mean precision–recall AUC against Fisher's exact test, topGO, MGSA, and blitzGSEA across all noise levels, though no reported result found on Lacuna for the actual AUC values. The decisive question here is different: how does each method degrade as the decoupling fraction rises? Plain RNA-fed pareg and Fisher overrepresentation should call planted "intent-only" pathways as hits; the fidelity-weighted model should not.

The second experiment uses the matched bulk cohort of Proteomizer: Leveraging the Transcriptome-Proteome Mismatch to Infer Novel Gene Regulatory Relations — 8,613 TCGA and CPTAC samples across more than 30 tissue types with paired transcriptomic, miRNomic, and proteomic measurements — where the paper reports a baseline RNA–protein correlation of $r \approx 0.5$ across genes and $r = 0.3$ across samples, improved to $r = 0.68$ by its translation model. Here I ask whether fidelity-weighted enrichment, run on held-out tissues, recovers known dysregulated cancer pathways that RNA-only enrichment misses, and whether per-pathway fidelity is itself conserved across tissues (if it is, fidelity transfer becomes a cheap correction for transcriptome-only cohorts; if it is not, that is a negative result worth publishing, since it would mean RNA-only enrichment bias is irreparable without local paired data).

The third experiment moves to immune challenge in single cells. The H1N1 CITE-seq cohort (53,201 peripheral blood mononuclear cells from influenza-infected patients) used by A joint analysis of single cell transcriptomics and proteomics using transformer provides the setting: does executed-enrichment identify T-cell activation programs that intent-only enrichment ranks differently, and do the predicted protein markers (CD45RA, CD38) that paper used to resolve CD8 subpopulations align with the execution calls? The cross-platform stress test uses the GSE164378 PBMC atlas (161,764 cells, 228 surface proteins) and the multi-platform PBMC benchmark of scTEL (161,764 cells across Illumina NovaSeq and HiSeq), splitting by platform and panel and measuring, with the correlation-matrix-distance metric that scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation showed degrades as panels scale into the hundreds, whether anchored harmonization of pathway signatures across platform splits converges better than fixed per-protein conversion factors — a direct cellular analogue of the BAU-versus-anchoring contrast in serology.

Baselines

The naive reference point is standard overrepresentation analysis: Fisher's exact test on an RNA differentially-expressed gene list with conventional cutoffs, the method still used in most vaccine immunology papers and described in Interpreting Gene Expression via Functional Enrichment Analysis. It is the floor to beat, and the synthetic benchmark already shows it loses to dependency-aware methods under noise (Coherent pathway enrichment estimation...); beating it is necessary but proves little.

The strongest existing competitor on the paired-omics question is ActivePathways (Integrative pathway enrichment analysis of multivariate omics data), because it already merges RNA and protein $p$-values with Brown's method and rationalizes which layer drove each enrichment — in the PCAWG pan-cancer cohort it recovered 100% of coding and 85% of non-coding consensus driver genes, and in METABRIC it found 192 prognostic pathways of which 33 were visible only through integration. It will be run on exactly my data with the same GO and Reactome sets, same compute, and the same multiple-testing correction, in two modes: RNA-plus-measured-protein, and RNA-plus-imputed-protein. On the synthetic discordance benchmark it will additionally be given the true planted decoupling fraction as an oracle-free upper-reference run. Beating it means two things together: higher precision–recall AUC as planted decoupling increases, and fewer semantically redundant reported terms (measured by overlap among called pathways, the pathology pareg was built to cure). I will also run pareg itself, fed protein-layer statistics and RNA-layer statistics separately, with its authors' regularization settings; no reported result found on Lacuna for pareg's numeric AUC on its own benchmark, so the comparison is head-to-head on shared data.

The key ablation is the fidelity-weighted Brown combination with all $\rho_g$ fixed to 1, which reduces the fusion step to exactly ActivePathways' unweighted combination while keeping everything else — the pareg-style regression, the continuous gradient, the anchoring protocol — intact. If this ablation matches the full method, the proposal's central ingredient is worthless and I will say so; the whole case for the method rests on a gap between these two runs that grows with the planted decoupling fraction. A secondary ablation replaces the fidelity-gated imputation with impute-always, to isolate whether gating by $\rho_g$ is what rescues cross-panel robustness.

Why not just…?

Why not just run Fisher's exact test on RNA DEGs, as most immunology papers do? It is interpretable and instant, but the documented failure modes are precisely what this proposal targets: it requires arbitrary significance cutoffs that discard coordinated sub-threshold shifts, evaluates overlapping pathways in isolation producing redundant term lists, and — decisively — reads only the transcript layer, which misrepresents functional state for roughly half the proteome (Coherent pathway enrichment estimation...; Proteomizer: Leveraging the Transcriptome-Proteome Mismatch to Infer Novel Gene Regulatory Relations). It remains the reference floor rather than a serious competitor.

Why not just use ActivePathways on paired RNA and protein data — doesn't it already solve multi-omic enrichment? This is the real threat, and I treat it as the primary baseline rather than arguing around it. ActivePathways combines layers symmetrically with fixed evidence weights, was built for static bulk cohorts, and has no pathway-dependency penalty, so it should over-call intent-only pathways when decoupling is planted and report redundant terms. But the honest position is that this is an empirical bet: if the fidelity-weighted ablation does not separate from unweighted ActivePathways on the synthetic and bulk experiments, the fidelity ingredient adds nothing and the proposal collapses to an application paper. The experiments are designed so that this outcome would be visible and reportable.

Why not just predict the proteins first (with Proteomizer or scTEL) and then run any enrichment tool on the predicted proteome? This two-stage pipeline is a legitimate simplification, and I will run it as a baseline: proteomize, then enrich. It falls short for two documented reasons. Proteomizer itself reports failure to generalize to unseen tissues and protocols, and a 2018-style transformer such as scTEL achieves 86.6% cell-type classification accuracy against sciPENN's 77.0% and Seurat's 76.1% — strong, but its imputations are unconditional, so it will confidently hallucinate proteins for exactly the low-fidelity genes where imputation is invalid. The fidelity gate is the specific mechanism that distinguishes my pipeline, and the impute-always ablation isolates it.

Why not just harmonize antibody panels with fixed conversion factors, like the WHO's binding antibody units? Because the serology literature already ran this experiment and it failed: commercial assays converted to BAU/mL still show divergent longitudinal trajectories (Studying temporal titre evolution of commercial SARS-CoV-2 assays...), while anchoring to a reference cohort's mean titer is what made protection curves converge across seven vaccine platforms, with a candidate at roughly 0.44-fold of the Moderna reference geometric mean titer predicted to exceed 80% efficacy against ancestral strains (Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection). I apply the same contrast at the pathway level, and the platform-split experiment will measure both approaches directly.

Decisive Checks

Three outcomes decide the proposal. On the discordance-augmented synthetic benchmark, fidelity weighting must beat the unweighted ablation and ActivePathways by a margin that grows with the planted decoupling fraction; without that gap, the idea is dead and the honest conclusion is that symmetric p-value combination already suffices. On the platform-split single-cell data, anchored harmonization must reduce inter-platform pathway-signature divergence relative to fixed conversion factors by a clear margin on the correlation-matrix-distance metric; if fixed factors work as well, the serology analogy does not transfer to cellular pathway signatures, which would itself be a useful negative finding. And on the H1N1 cohort, at least one immune program must be called differently by execution than by intent, with the discrepancy traceable to specific low-fidelity genes and their measured proteins — the concrete demonstration that transcriptional intent and functional execution can be told apart in a real immune response, which is the whole point of the exercise.