Tipping-Point Hubs as a Prognostic Signal in Lung Adenocarcinoma

When a cancer cell moves from a tightly packed epithelial state to a mobile mesenchymal one, the change is not gradual. Single-cell studies of this epithelial–mesenchymal transition (EMT) find an unstable halfway state. At that point a small set of genes briefly becomes the most connected nodes in the regulatory network, the "hubs," and afterwards they fade back to the periphery. Clinical prognostic signatures are built differently. They come from bulk tumor RNA, where every cell in the biopsy is averaged together, and they pick genes by how strongly raw expression tracks survival. Two things are lost in that process. The genes that drive the transition are often not strongly differentially expressed, and the bulk average mixes the cancer cells' signal with stroma and immune cells.

This note proposes a clean test of whether the missing information matters clinically. We will derive a gene set purely from single-cell EMT time courses, choosing genes by how sharply their regulatory influence spikes at the tipping point. No survival data are used in the selection. We then freeze that gene set and score bulk lung adenocarcinoma (LUAD) patients after removing the part of each score explained by non-tumor cell content. If these "transient hubs" carry survival information that bulk-trained LASSO signatures, generic EMT scores, and static network hubs do not, that would show three things at once. Dynamic network position is a better criterion for choosing biomarkers than static centrality. Averaging across cell types is actively hiding it. And the gene set was never fitted to the outcome it predicts.

Core Claim

The claim is specific. Genes ranked by their peak outgoing regulatory influence inside the unstable EMT window predict LUAD overall survival better than two alternatives built from the same single-cell network: genes ranked by time-averaged influence, and a size-matched generic EMT score. The advantage should grow, rather than shrink, once the score is adjusted for tumor purity and stromal fraction. Genes are ranked by the Jacobian, the matrix of how much each gene's rate of change depends on every other gene.

The last prediction separates this idea from a restated EMT signature. Bulk EMT scores are widely suspected to read out fibroblast content, which the file flags as a general weakness of bulk signatures. That suspicion predicts a generic EMT score should weaken under purity adjustment. A cancer-cell-intrinsic transition signal should not weaken.

Relation to Prior Work

The single-cell side supplies the method. spliceJAC fits models of how unspliced and spliced mRNA are produced and degraded, then linearizes around each cell state to get local Jacobians. It shows that networks become denser and less modular in transition states, and that transition genes often lack strong differential expression. Tipping points in epithelial-mesenchymal lineages locates the saddle-node tipping point in OVCA420 EMT, where the largest Jacobian eigenvalue peaks around three days post-induction. Neither paper asks whether its transition genes stratify patients. That is the step we add.

On the clinical side, three studies motivate the gap but constrain the design. Zhao et al. use single-cell clusters only as a marker filter before static co-expression modules (WGCNA) and LASSO. Ding et al. take static hubs from the STRING protein-interaction database via the CytoHubba tool and then fit LASSO. Saha et al. build patient-specific regulatory networks from bulk GTEx and TCGA data and derive a prognostic aging signature (p = 0.015 versus p = 0.147 for chronological age). They cannot tell rewiring inside cells apart from shifts in cell composition.

BCTI detects critical stages in bulk TCGA and links them to worse survival. It works on bulk stage labels and outputs a whole-system score rather than a portable gene panel.

SubNetDL is the result a reviewer might see as contradicting us. It finds that feature importance for therapy response runs inversely to static centrality in STRING. We read this as consistent with our hypothesis: static centrality is the wrong measure, and centrality inside the transition window may be the right one. We test that interpretation directly rather than assume it.

Research Program

Stage 1: select genes from single-cell data only. We run spliceJAC-style kinetic inference on the A549 TGF-β1 EMT time course, which is lung adenocarcinoma-derived and so matches the clinical target. We use a sliding pseudotime window and restrict to the few hundred most variable genes, since these methods do not scale further.

For each gene $g$ and window $w$, we define outgoing influence as $s_g(w)=\sum_{h\neq g}|J_{hg}(w)|$, where $J(w)$ is the window's Jacobian. The tipping window $w^*$ is where the largest real eigenvalue of $J$ peaks. A gene's transience score is its peak-to-baseline ratio, $T_g = s_g(w^*)/\mathrm{median}_w\, s_g(w)$.

The top $k$ genes by $T_g$ form the transient-hub set, with $k$ fixed in advance at 10 and 20 to sit in the usual signature range. Each gene is signed by whether it rises toward the mesenchymal side. OVCA420 serves as an independent replication of the gene ranking. Genes that rank highly in both cell lines are the primary set.

Stage 2: score patients without fitting to survival. A patient's score is the signed mean of standardized expression over the frozen set. No survival-trained weights are used. In RNA-seq cohorts (TCGA-LUAD, n = 499), we compute the score two ways. The first uses bulk expression residualized on estimated cell-type fractions. The second uses malignant-compartment expression estimated with Count Bridges, a generative model that splits integer bulk counts into cell-level counts, with a LUAD single-cell reference from the Curated Cancer Cell Atlas. The microarray validation cohorts are not counts, so there we use only the fraction-residualized version. We treat this asymmetry as an honest limit, not something to hide.

Stage 3: test whether prognostic genes are dynamic hubs. Across all genes in the single-cell network, we correlate each gene's univariate Cox association in TCGA-LUAD with three quantities: $T_g$, time-averaged $s_g$, and STRING degree. This asks whether the SubNetDL-style inverse relation with static centrality holds while a positive relation with transience also holds.

Baselines

Do-nothing reference. A Cox model with stage, age, and sex, with and without estimated tumor purity and stromal and immune fractions. Every signature must add C-index over the purity-adjusted clinical model, not merely over stage.

Strongest existing method. The standard bulk pipeline: univariate Cox filtering, then LASSO-penalized Cox on TCGA-LUAD, validated on GSE31210, GSE50081, GSE11969, and GSE42127. We run it ourselves with nested cross-validation and the same candidate-size range. We also re-implement the published LUAD panels on identical splits.

The strongest reported LUAD numbers in the file are these. The 3-gene MMP12/NFE2/HOXC8 panel in early-stage TCGA-LUAD reports time-dependent AUC 0.916, 0.90, and 0.863 at 1, 3, and 5 years (HR 3.491). The 6-gene aging nomogram reports 0.754, 0.73, and 0.742 in TCGA and 0.923, 0.803, and 0.752 in GSE31210. These are training-cohort or subset figures and not directly comparable to each other, which is why we re-run everything on common splits.

Beating the LASSO pipeline means a higher pooled external-cohort C-index. Matching it would still be meaningful, because our gene set saw no survival data. We will call the result a win if the frozen set matches or exceeds LASSO externally and loses less performance from TCGA to GEO.

Ablations. Two ablations isolate the key ingredient. The time-averaged-hub set uses the same single-cell networks and the same $k$, but ranks by mean $s_g$. The static-hub set uses the same $k$ from STRING degree and CytoHubba ranking, following Ding et al. A third control is a size-matched set of EMT markers that are top differentially expressed along pseudotime. If time-averaged hubs match transient hubs, the dynamics add nothing.

Null. Random gene sets of equal size, matched on mean expression and on correlation with a proliferation score. This matters because arbitrary gene sets in cancer are often prognostic through proliferation.

Why not just…?

Why not just run LASSO on bulk differentially expressed genes? This is the main competitor, and it may well win on in-sample TCGA AUC because it is fitted to that outcome. The file's evidence against it is structural. spliceJAC reports that transition drivers are often not differentially expressed, so a differential-expression filter removes them. BCTI argues that collinearity near transitions makes regression pick arbitrary surrogate genes. Neither point shows worse survival prediction, so the external-cohort comparison decides it. If LASSO holds its performance across the GEO cohorts as well as the frozen set does, our claim reduces to interpretability and we will say so.

Why not just add cell-type fractions as covariates to any signature? This is a real threat to the deconvolution half of the proposal. We include fraction-residualized scoring as a first-class arm and compare it against Count Bridges malignant-compartment scoring. If residualization does as well, we drop generative deconvolution and keep the simpler step. The purity-adjustment prediction still stands either way: transient hubs should survive adjustment and generic EMT scores should not.

Why not just use a generic EMT or hybrid-EMT marker score? Transient hubs might simply be EMT genes relabeled. The size-matched pseudotime-marker control and its behavior under purity adjustment test this. We also report gene overlap. If the two sets overlap heavily and behave identically, the dynamic ranking is not buying anything.

Why not just take static hubs from STRING? SubNetDL's inverse-centrality finding already argues against static hubs for predicting therapy response, and database degree is biased toward well-studied genes. The static-hub ablation and the Stage 3 correlation analysis test this directly for survival.

Proposed Experiments and Decisive Checks

The primary endpoint is overall-survival C-index and time-dependent AUC at 1, 3, and 5 years. Each signature is added to the purity-adjusted clinical Cox model, and results are pooled across the four GEO LUAD cohorts with TCGA as the discovery set for the LASSO arms only. We also report hazard ratio per standard deviation of score and incremental C-index with bootstrap intervals.

Four checks decide the outcome. First, transient hubs must beat time-averaged hubs from the identical networks externally. This is the core test, and failing it falsifies the idea. Second, the transient-hub hazard ratio must stay stable or rise after purity adjustment while the generic EMT score falls. Third, the frozen set must beat 95% of matched random sets. Fourth, the OVCA420-derived ranking must reproduce the A549 ranking well enough that the primary set is not an artifact of one cell line.

A secondary check tests the spliceJAC claim at clinical scale. We count how many transient hubs fall below conventional tumor-versus-normal differential-expression cutoffs in TCGA-LUAD. This shows how many would have been discarded by the standard pipeline.

The main risks are that kinetic inference in two cell lines may not reflect the transitions happening in patient tumors, and that the gene pre-filter required for tractable Jacobian inference could exclude real drivers. We report sensitivity to the pre-filter size to address the second.