Assembly-Gated Hubs in Developmental Program Reactivation

When an adult cell turns cancerous or chronically inflamed, it rarely invents new behavior. Instead it re-runs embryonic programs that were epigenetically silenced after development finished: medulloblastomas land squarely in the mouse postnatal day 5–7 window of granule-cell migration Conserved mechanisms across development and tumorigenesis revealed by a mouse development perspective of human cancers, adult atopic skin re-expands fetal macrophage and endothelial states Developmental cell programs are co-opted in inflammatory skin disease, and aggressive gliomas reboot a germ-cell transcriptional identity Activation of embryonic/germ cell-like axis links poor outcomes of gliomas. All of this is read almost exclusively from mRNA. Yet mRNA levels predict actual protein abundance only 40–60% of the time A high-stringency blueprint of the human proteome, and the functional output of a protein depends on whether it physically assembles with its partners — something no transcript measurement can show.

This proposal asks a single question: which physical protein complexes actually execute developmental reactivation? The answer requires connecting four things that have never been connected. Developmental timelines tell us when the co-opted programs normally run; transcript-protein discordance tells us that the transcriptome is an unreliable proxy for the machinery; network centrality tells us which genes matter topologically; and co-immunoprecipitation (Co-IP — pulling one protein out of a cell lysate with an antibody and seeing which partners come with it) tells us which predicted interactions are physically real.

Core Claim

The central claim is that developmental reactivation is carried by a specific, identifiable subset of interactome hubs — "assembly-gated hubs" — that satisfy three conditions at once: they are topologically central, they are present as protein at the reactivated developmental stage (not merely transcribed), and they physically assemble into complexes only in the reactivated state. Standard hub analysis on static, unweighted interaction graphs cannot find these, because a gene can be a perfect hub on paper while its protein is absent, degraded, or stoichiometrically mismatched with its partners in the actual cell. A secondary, mechanistic claim follows directly: if reactivated states assemble complexes from latent, pre-existing protein pools rather than from newly induced transcription, then complex formation will precede transcript induction — a prediction Co-IP time courses can test and that current reactivation models, which assume transcriptional re-expression, cannot even pose.

Research Program

The program has three coupled components.

First, we quantify transcript-protein discordance at cell-state resolution. The skin atlas of Reynolds et al. Developmental cell programs are co-opted in inflammatory skin disease is the ideal substrate: 528,253 cells across fetal, healthy adult, and inflamed skin, with single-cell RNA sequencing paired with mass cytometry protein measurements, and already-resolved reactivated states (the fetal-like Mac2 macrophages and VE3 endothelial cells). For each gene with a matched protein marker, we compute the rank divergence between transcript and protein abundance per cell state, then impute protein abundance for the rest of the genome with an RNA-to-protein model in the family of totalVI Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells and scChord scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation. Notably, scChord's own evaluation shows degradation on the Correlation Matrix Distance metric when the protein panel scales to 228 proteins — evidence that multi-protein correlation structure cannot be learned from RNA alone, which is precisely why we inject an interaction-graph prior rather than scaling the generative model.

Second, we build a developmental-stage-resolved, protein-gated interactome. Nodes carry estimated protein abundance per state; an edge between proteins $u$ and $v$ at developmental stage $s$ is active only if both partners are present as protein and stoichiometrically balanced:

$$w_{uv}(s) = \mathbb{1}\!\left[\min(p_u(s), p_v(s)) > \tau\right] \cdot \frac{\min(p_u(s), p_v(s))}{\max(p_u(s), p_v(s))},$$

where $p_u(s)$ is the imputed or measured protein abundance of $u$ at stage $s$, $\tau$ is a presence threshold, and the ratio term penalizes partnerships whose members exist at wildly unequal concentrations. Degree and betweenness centrality are then recomputed on the gated graph at each stage, using the mouse cerebellar P1–P60 timeline from Kho et al. Conserved mechanisms across development and tumorigenesis... as the temporal scaffold. The headline analysis compares hub identity between the static aggregate graph and the P5-gated graph: we predict that a substantial fraction of static hubs (especially well-studied oncoproteins whose degree reflects publication density) drop out, while previously invisible hubs — proteins abundant only in the migratory window — emerge. Existing work motivates this suspicion: high-betweenness bridge genes in HIV progression were found precisely because they were not differentially expressed, showing that transcript screens miss structural players; and the ESPACE work showed that uniform statistical penalties systematically miss true super-connectors Enhanced construction of gene regulatory networks using hub gene information, a failure mode our gating is designed to repair at the protein level.

Third, we validate physically. Top-ranked assembly-gated hub baits are subjected to Co-IP followed by unbiased mass spectrometry (Co-IP-MS), with chemical cross-linking to stabilize the transient kinase-substrate and ligand-receptor interactions that the Co-IP literature itself flags as lost during washing. We compare complex composition between the reactivated state and the differentiated control state, in glioma lines (U251, LN229, A172) carrying the germ-cell axis described by Ma et al. Activation of embryonic/germ cell-like axis links poor outcomes of gliomas and in an endothelial-macrophage co-culture modeling the ACKR1–CXCL8 circuit that Reynolds et al. predicted computationally but never verified physically. The IQGAP3–GLI1 interaction, already proven by targeted Co-IP IQGAP3 activates Hedgehog signaling to confer stemness and metastasis via up-regulating GLI1 in lung cancer, serves as a positive control and calibration point. Because cell lysis can create artificial post-lysis associations, top hits are cross-checked in situ with proximity ligation assays, which fluoresce only when two targets sit within about 40 nm in intact tissue.

The mechanistic payoff experiment is a time course: after inducing the reactivated state, we track complex assembly (Co-IP) and transcript induction (RNA-seq) in parallel. If physical complexes form within minutes and resist transcriptional blockade, reactivation is partly an assembly-of-latent-pool phenomenon, consistent with the "shadow proteins" that persist without transcript support Persistently Increased Expression of PKMzeta and Unbiased Gene Expression Profiles Identify Hippocampal Molecular Traces of a Long‐Term Active Place Avoidance Memory and "Shadow" Proteins. If assembly strictly tracks transcription, the field's transcription-centric model is vindicated — also a decisive answer.

Baselines

The naive reference point is conventional hub analysis: degree and betweenness centrality computed on a static, unweighted aggregate interaction graph, exactly as routinely practiced in the COPD, pituitary adenoma, and hydronephrosis studies cited in the literature. This baseline shares our interactome, our gene lists, and our centrality code; only the edge gating is removed. Beating it means the protein-gated graph's top hubs recover significantly more Co-IP-validated, state-differential complexes at matched list size, measured as precision@k against Co-IP-MS prey recovery.

The strongest existing competitor is the Prize-Collecting Steiner Tree/Forest family, which already integrates proteomic and transcriptomic signals on the interactome and has the best demonstrated track record of network predictions that survive wet-lab scrutiny: Huang et al. Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling recovered Steiner hubs (CREBBP, β-catenin, p300) that drug screening confirmed, and Tunçbağ et al. Network Modeling Identifies Patient-specific Pathways in Glioblastoma predicted MEK1 as a universal hub that Western blot confirmed as phosphorylated in 7 of 8 glioblastoma xenograft lines, with NUMB hub presence predicting low invasiveness at $p = 1.61 \times 10^{-14}$. We will run PCST/PCSF on identical inputs (same interactome, same RNA and protein measurements, same top-k output budget, prize and penalty parameters swept over a matched grid) and compare precision@k of physical-complex recovery. PCST is the fiercest rival because it, like us, fuses protein data with topology; the honest difference is that PCST uses protein measurements as node "prizes" on a static graph, without stage resolution, stoichiometric edge gating, or any physical-complex validation, so the comparison cleanly isolates whether gating and temporal resolution add value beyond better prize placement.

The key ablation replaces protein gating with mRNA weighting: identical pipeline, identical centrality metrics, but edges weighted by transcript abundance (an ESPACE-like construction Enhanced construction of gene regulatory networks using hub gene information). If mRNA-weighted hubs match protein-gated hubs in Co-IP precision, the discordance ingredient is decorative and the claim fails. For the imputation sub-component, scChord is the strongest baseline; it reports the lowest RMSE and highest Pearson correlation on the GSE164378 benchmark, though no exact numeric values for those metrics were given, so no reported result found on Lacuna for the precise RMSE/PCC figures — our comparison will therefore report absolute numbers on the same datasets (GSE164378 with its 161,000+ cells and 228 proteins, GSE100866 with ~8,600 cells and 13 proteins, and the SLN-all murine set of ~33,000 cells) under matched compute, using scChord's own published 100,000-cell, 27-minute training run as the runtime reference.

Why not just…?

Why not just run PCST/PCSF on paired RNA-plus-protein data? This is the most serious alternative, and we treat it as our primary baseline rather than arguing around it. It falls short on two counts that the experiments will test directly. PCST has no concept of developmental stage, so it cannot say when a hub is active; and it treats proteins as independent prize signals, so a stoichiometrically mismatched pair — both "prized" but unable to assemble — is indistinguishable from a functional complex. Our decisive check is the head-to-head: if PCST at matched budget recovers as many Co-IP-validated differential complexes as the gated graph, PCST wins and we say so.

Why not just do Co-IP-MS on a few expert-chosen developmental regulators (DAZL, GLI1, Snail) and skip the modeling? This is cheap and partially informative, but it cannot answer the prioritization question: Co-IP-MS is bait-limited, antibody-constrained, and vulnerable to lysis artifacts, so the scientific value lies in which baits to pull. The CtBP–Snail study illustrates the trap — its central physical interaction rests on a STRING prediction score of 0.923 and heterologous assays, with no endogenous Co-IP in invading tissue CtBP modulates Snail-mediated tumor invasion in Drosophila — and the glioma study's 12-gene signature is a flat, unweighted list with no protein or complex layer Activation of embryonic/germ cell-like axis links poor outcomes of gliomas. We will settle it with a matched-bait contest: an equal number of model-selected baits versus expert-chosen baits, scored on Co-IP-validated, state-differential complex recovery. If the experts match the model, the model earns nothing beyond convenience.

Why not just build a bigger RNA-to-protein imputation model? Because per-protein accuracy is not the bottleneck — correlation structure across many proteins is. scChord's own reported degradation on the Correlation Matrix Distance metric at 228 proteins scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation shows that scaling the generative model alone does not capture joint protein behavior, which is exactly the information a physical interaction graph supplies for free. Our graph-prior variant will be evaluated specifically on whether it holds CMD flat as the panel grows, on the same benchmarks where scChord's weakness was documented.

Proposed Experiments and Decisive Checks

Experiment one calibrates discordance-aware protein imputation on GSE164378, GSE100866, and SLN-all, reporting RMSE, Pearson correlation, and CMD for the graph-prior model against scChord and totalVI, with matched training budgets (scChord's published run: 100,000 cells in roughly 27 minutes on under 230 MB of GPU memory sets the compute anchor). Success means matching per-protein accuracy while keeping CMD stable as the panel scales, where the baseline degrades.

Experiment two builds the gated, stage-resolved cerebellar interactome and measures how much hub identity changes between the static graph and the P5-gated graph, the window where over 90% of human medulloblastomas land and where metastatic tumors show an odds ratio of 7.0 Conserved mechanisms across development and tumorigenesis.... A sanity check tests whether the known metastatic shortlist (ASPM, MET Integrative Genome Comparison of Primary and Metastatic Melanomas) and the 12-gene glioma signature are recovered as gated hubs, and whether the 443-day versus 3,411-day survival stratification of that signature Activation of embryonic/germ cell-like axis... sharpens when weighted by gated-hub score. For ESPACE's own benchmark performance the file reports only directional improvements in false discovery rate and Matthews correlation against RegulonDB ground truth — no reported result found on Lacuna for the exact values — so transcript-only comparisons will be run by us on identical data.

Experiment three is the wet-lab core: Co-IP-MS with cross-linking on the top assembly-gated baits in reactivated versus control glioma cells and in the endothelial-macrophage co-culture, with proximity ligation confirmation in situ, the IQGAP3–GLI1 pair as positive control, and the ACKR1–CXCL8 circuit of Reynolds et al. as a named, previously unverified target. Primary metric: precision@k of Co-IP-validated state-differential interactions, scored against the static-centrality, PCST, and mRNA-weighted ablation baselines at matched budgets. The time-course experiment — complex assembly versus transcript induction, with and without transcriptional blockade — delivers the mechanistic verdict on whether reactivation runs on newly made proteins or pre-built ones. Either answer changes how the field reads its transcriptomes.