Enrichment on Discordance: Finding Fc-Effector Pathways That RNA Alone Misses

When an antibody's Fc region — the constant "tail" of the molecule that recruits immune cells rather than binding the target — triggers an effector response, much of the action happens without any new gene transcription. Trogocytosis, for instance, physically strips receptors like TIGIT off the cell surface, and antibody-dependent killing by NK cells is driven by protein-level receptor engagement and redistribution. Yet the standard way biologists interpret such experiments — take differentially expressed genes, run functional enrichment analysis against pathway databases like Gene Ontology or KEGG — sees only RNA. Meanwhile, we know from proteome-wide surveys that mRNA levels match protein levels only 40% to 60% of the time A high-stringency blueprint of the human proteome, and single-cell studies of tumors keep finding immune states that transcriptomics cannot see High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy. The enrichment literature and the transcript-protein discordance literature have simply never been connected.

This proposal builds that connection, with Fc-mediated effector biology as the application regime where it matters most. The core deliverable is a discordance-aware enrichment method: instead of asking "which pathways are overrepresented among genes whose RNA changed?", it asks "which pathways are overrepresented among genes whose RNA and protein disagree?" — and it models pathway-to-pathway overlap so the answer is a coherent, non-redundant list rather than a sprawling one. We will evaluate it where the blind spot is sharpest: Fc-competent versus Fc-silent antibody treatment of human immune cells, where the Ociperlimab preclinical work already showed that efficacy depends entirely on the Fc domain while the Fc-silent variant shows minimal efficacy An Fc-Competent Anti-Human TIGIT Blocking Antibody Ociperlimab (BGB-A1217) Elicits Strong Immune Responses and Potent Anti-Tumor Efficacy in Pre-Clinical Models.

Core Claim

Standard enrichment analysis is structurally blind to functional states governed by protein turnover, receptor stripping, and post-translational activation, because its input is a univariate RNA statistic. We claim that a per-gene discordance statistic — the gap between the RNA-level and protein-level differential signal, measured in the same single cells by CITE-seq (an assay that reads out transcriptomes and hundreds of surface-protein levels simultaneously) — is itself a valid enrichment input, and that enriching on it recovers Fc-effector pathways (Fcγ receptor signaling, trogocytosis, antibody-dependent cytotoxicity modules) that RNA-only enrichment provably misses. The closest existing method, pareg, already solves the statistical half of this — it models inter-pathway dependencies with a network fusion penalty and beats Fisher's exact test, topGO, MGSA, and blitzGSEA on precision-recall in noisy settings Coherent pathway enrichment estimation by modeling inter-pathway dependencies using regularized regression — but it has only ever been fed RNA p-values. We extend its input, not its machinery, and that single change is what opens the Fc-effector blind spot.

Research Program

Three components, in increasing order of biological commitment.

First, the discordance statistic. For each gene $g$ with a paired protein measurement, we compute $z^{RNA}_g$, a standardized differential-expression statistic contrasting two cell states (e.g., effector-cell populations before versus after antibody engagement), and $z^{P}_g$, the analogous statistic for the protein. The discordance is the signed difference $d_g = z^{RNA}_g - z^{P}_g$, converted to a per-gene p-value by permutation within matched cell types and donors, so that genes whose protein moves without their RNA — the signature of trogocytosis, shedding, or translation-specific regulation like the METTL5–CCND3 case, where an RNA modification boosts translation of a protein without touching its transcript METTL5‑mediated m6A modification of 18S rRNA drives oral squamous cell carcinoma progression by enhancing CCND3 translation — land at the extremes of the discordance distribution.

Second, the enrichment engine. We feed these discordance p-values into a pareg-style generalized linear model: gene-level statistics are regressed on a binary gene-pathway membership matrix with a LASSO penalty for sparse pathway selection plus the network fusion penalty that forces functionally similar pathways (by Jaccard or semantic similarity) toward similar scores. We deliberately start with this linear engine rather than a graph neural network, because the file's own evidence cuts both ways: pathway-informed graph attention networks achieved an 81% reduction in MSE versus an MLP baseline on the TP53–MDM2–MDM4 time series Biological Pathway Informed Models with Graph Attention Networks (GATs), but that demonstration was a three-gene circuit at the RNA layer only; we will swap in a graph-attention variant only if a pre-registered diagnostic shows the linear model's residuals carry pathway structure it cannot absorb.

Third, the Fc-effector benchmark, which is the genuinely new data collection. Following the Ociperlimab design directly — wild-type IgG1 versus Fc-silent anti-TIGIT, where the intact Fc drives Treg depletion by ADCC and Fc-dependent trogocytosis of TIGIT An Fc-Competent Anti-Human TIGIT Blocking Antibody Ociperlimab (BGB-A1217) Elicits Strong Immune Responses and Potent Anti-Tumor Efficacy in Pre-Clinical Models — we will run paired CITE-seq on human PBMC co-cultures at early time points (0, 2, 6, 24 hours), chosen because trogocytosis and receptor internalization act on protein timescales that precede transcriptional response. This gives the method a ground truth with a built-in negative control arm: whatever enrichment finds in the Fc-competent arm but not the Fc-silent arm, at time points where RNA differential expression is near zero, is by construction a post-transcriptional Fc-effector program.

Baselines

The naive reference point is RNA-only differential expression followed by Fisher's exact test overrepresentation analysis, exactly as used across the enrichment literature, for instance in the microglia single-cell study that mapped NF-κB and chemokine changes in traumatic brain injury from KEGG and GO terms Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases. This baseline matters because it is what nearly every lab currently does, and because it should fail in a specific, informative way on the Fc-competent arm: if our early time points indeed precede transcription, it should return nothing, which is precisely the blind spot we claim to fill. The strongest existing competitor is pareg run on RNA p-values alone: it reports the highest mean Area Under the Precision-Recall Curve compared with Fisher's exact test, topGO, MGSA, and blitzGSEA across noise levels on its synthetic benchmark, though no scalar AUC value is reported on Lacuna, so our comparison against it will rest on effect direction plus our own re-run under identical inputs. It is the strongest competitor because it shares our statistical engine, meaning any advantage we show is attributable to the discordance input and not to better handling of pathway overlap. It will be given the same CITE-seq RNA layer, the same pathway database, and the same compute and tuning budget (cross-validated penalty selection, as in the original, which the authors note can take up to roughly 30 minutes in parallelized mode). Beating it means recovering the FcγR-signaling and trogocytosis pathways in the Fc-competent arm with higher precision at matched recall, and — the decisive test — flagging pathways that its RNA-only run scores near zero. Two ablations isolate our key ingredient: the same network-fusion GLM fed protein-only p-values (testing whether discordance beats either modality alone, since protein-only enrichment also exists in principle but is limited to the ~200 measured markers), and the same discordance input fed to plain LASSO without the fusion penalty (testing whether coherent pathway scoring matters once the input is discordance rather than expression).

Why not just…?

Why not just run RNA-seq and standard GSEA, since it's cheaper and mature? Because the failure is documented, not hypothetical: mRNA predicts protein only 40% to 60% of the time proteome-wide A high-stringency blueprint of the human proteome, and in esophageal tumors, transcriptomic profiling failed to capture the functional immune state, missing the CD16+ NK-cell activation that protein-level CyTOF revealed as the driver of checkpoint-blockade response High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy. Our first experiment makes this failure quantitative on our own benchmark: we will report exactly what RNA-only GSEA recovers in the Fc-competent arm at each time point, and we expect near-total silence at 2 and 6 hours. If RNA-only enrichment in fact recovers the Fc-effector programs, our premise is wrong and we will say so.

Why not just measure the proteins directly with flow cytometry or CyTOF and skip the modeling? This is a real threat, and part of our answer is to concede it a role: the 42-antibody CyTOF panel of the esophageal study is our external validation target, not a competitor — if our discordance enrichment on independent paired data flags FcγR and NK-cytotoxicity terms consistent with that study's protein-level findings, that is convergent evidence. But cytometry cannot be the whole answer: panels are capped at roughly 40–200 pre-selected surface epitopes and cannot see the thousands of intracellular signaling molecules involved, and it measures no transcripts, so per-gene discordance — the quantity that separates translational regulation from receptor stripping — is undefined without paired modalities. The proposal's value is precisely in ranking pathways genome-wide, which no panel can do.

Why not just impute the missing proteins with an existing translation model like scTranslator or scChord, and enrich on the imputed proteome? We will run this as a third competitor, honestly, but we expect a structural problem: a model that predicts protein from RNA makes the "protein" channel a deterministic function of RNA, so the discordance it can express reflects model error and training bias rather than biology. The file documents this limitation from inside the field — scChord's own authors report that as panels scale to hundreds of proteins, performance degrades on Correlation Matrix Distance, and they propose pathway-informed priors as future work scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation, and translation models depend on CITE-seq reference panels biased toward pre-selected antigens A pre-trained large generative model for translating single-cell transcriptome to proteome. Our design settles it empirically: on held-out measured proteins, we compare discordance enrichment driven by measured versus imputed protein channels, and if imputation-based discordance performs as well, the cheaper route wins and we will report that.

Why not just feed protein p-values into pareg and call it done? That is exactly our protein-only ablation, and the reason it is not the proposal is that neither modality alone distinguishes the two mechanisms we care about: a pathway can be transcriptionally activated (RNA and protein both move, no discordance) or post-translationally remodeled (protein moves, RNA silent, maximal discordance). Only the joint statistic separates them, and the Fc-effector setting is where the second regime dominates.

Proposed Experiments and Decisive Checks

The method's statistical soundness will be established first on an extension of pareg's own synthetic benchmark: we simulate pathway activation in two modes — transcriptional and post-transcriptional-only — with controlled noise and pathway overlap, and measure precision-recall for recovering the post-transcriptional ground-truth pathways. Calibration and robustness will be established on the GSE164378 PBMC atlas (161,764 cells, 228 paired proteins) scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation and the ~33,000-cell SLN-All murine atlas with its 111- and 208-protein panels Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells, where NK-cell and monocyte compartments let us verify that baseline discordance is low for most pathways, so the method is not simply flagging noise.

The decisive experiment is the in-silico trogocytosis perturbation followed by the wet-lab arm. In the perturbation, we artificially strip the protein channel for CD16 and co-receptors on atlas NK cells while leaving RNA untouched, mimicking what the Ociperlimab study showed happens physically in vivo; RNA-only enrichment should score FcγR signaling near zero, while discordance enrichment must rank it at the top for the claim to stand. The wet-lab arm then tests the same prediction with real biology: Fc-competent versus Fc-silent anti-TIGIT co-cultures, paired CITE-seq across four early time points, with pre-registered success criteria — discordance enrichment flags FcγR-mediated trogocytosis and ADCC modules in the Fc-competent arm only, at time points where RNA differential expression is statistically indistinguishable from zero. As external validation, we check whether the enriched terms on an independent esophageal tumor cohort are consistent with the CD16+ NK and chemokine-macrophage findings that the CyTOF study established at the protein level (that study's own 7-chemokine survival model separated TCGA patients with p = 0.000073 High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy). A secondary payoff, if the method works, is answering an open question the engager literature poses explicitly: whether CYT-303's reversal of NK exhaustion is transcriptional or post-translational Derivation and Preclinical Characterization of CYT-303, a Novel NKp46-NK Cell Engager Targeting GPC3 — a question our discordance scores answer by construction, since the two mechanisms fall on opposite ends of the same axis.