Whose Metabolism Predicts Survival? Separating Cell Mix from Cell State in Bulk Tumor Signatures

Most cancer survival studies still measure gene activity from bulk tissue. A biopsy is ground up, and the RNA from millions of cells (tumor cells, immune cells, connective-tissue cells) is read out as one average. When such a study reports that a "glycolysis score" (a summary of how strongly the genes for fast, oxygen-independent sugar burning are switched on) predicts poor survival, at least three different stories could produce the same number. The tumor cells themselves may have rewired their metabolism. The biopsy may simply contain more macrophages, which also run glycolysis. Or the score may be a disguised readout of how fast the tumor is dividing, which is already known to predict outcome. Each story points to a different therapy. A standard statistical "independence" check against stage and age cannot tell them apart.

We propose a method and an audit that split each bulk pathway signal into two parts: a part explained by which cells are present (composition) and a part explained by what the cancer cells are doing (state). We then ask which part carries survival information beyond stage, grade, age, tumor purity, and proliferation. The method's key step reconstructs per-patient, cell-type-specific expression from bulk counts. Pathway enrichment and survival modeling are then done on the malignant compartment. The headline test is whether malignant-cell-intrinsic metabolic reprogramming carries prognostic signal that composition and proliferation do not. We test this in hepatocellular carcinoma, in IDH-wild-type glioblastoma, and in ER-negative and HER2-positive breast cancer, where proliferation markers are known to fail.

Core Claim

Published metabolic prognostic signatures mix composition and state, and current independence testing cannot detect the mix. The motivating evidence comes from breast cancer. In Wirapati et al.'s meta-analysis, dozens of apparently distinct signatures across 2,833 profiles turned out to track one proliferation module. The same study showed that a whole "normal-like" subtype was an artifact of stromal contamination. The open-problems literature also states plainly that it is unresolved whether bulk prognostic scores reflect cancer-cell aggressiveness or shifts in stromal and macrophage infiltration.

Our claim is that once composition and proliferation are removed at the level of the cells, not just as covariates, a measurable malignant-intrinsic metabolic term either survives or does not. Its survival is a falsifiable, cancer-specific finding. It matters most where proliferation is uninformative.

The decomposition is simple to state. For patient $i$, let $\pi_{ik}$ be the fraction of cell type $k$ and $s_{pk}(i)$ the activity of pathway $p$ inside cell type $k$. To first order the bulk pathway score is $S_p(i) \approx \sum_k \pi_{ik} s_{pk}(i)$. Its deviation from the cohort mean, written with bars for cohort averages, splits as

$$\Delta S_p(i) \approx \underbrace{\sum_k \Delta\pi_{ik}\,\bar s_{pk}}_{\text{composition}} + \underbrace{\sum_k \bar\pi_k\,\Delta s_{pk}(i)}_{\text{state}} + \text{interaction}.$$

We enter the malignant state term $\Delta s_{p,\text{mal}}(i)$ into a Cox proportional hazards model. That model relates a patient's covariates to their instantaneous risk of death. Alongside the state term go stage, grade, age, all composition fractions, and a malignant-compartment proliferation score. A metabolic biomarker is called "independent" only if this term survives.

Relation to Prior Work

Two studies are the closest precedents, and both motivate the proposal while leaving its central question open.

The SF3A3 hepatocellular carcinoma study scored metabolic pathways in single cells. It found a late-trajectory malignant cluster enriched for inositol phosphate metabolism. It then collapsed the result to one splicing gene, SF3A3, and tested that gene's bulk expression for independence in TCGA and ICGC. That bulk test does not ask whether SF3A3's bulk level reflects malignant state or the share of malignant cells.

The glioblastoma gene-pair study built a 21-pair metabolic signature with external C-indexes between 0.648 and 0.801. It confirmed independence by multivariate Cox regression. Only afterwards did it use single-cell data to discover that a key gene, ABCA1, is expressed by macrophages. That is a direct example of a "metabolic" signature partly driven by a non-malignant compartment, found after the fact.

We reverse the order. Cell attribution happens before the survival test and is part of it.

The method layer comes from deconvolution work. Count Bridges reconstructs latent single-cell integer counts that sum exactly to the observed bulk counts. It reports outperforming CIBERSORTx and MuSiC on cell-proportion recovery. We reuse it as the engine that produces $s_{pk}(i)$. Its authors did not connect it to survival or pathway analysis. POPPY aligns bulk and single-cell profiles in a shared embedding and attributes responder status to cell states. It handles only binary response, not time-to-event outcomes adjusted for clinical covariates, so we use it as an alternative attribution engine. SubNetDL shows that resolving subclonal structure beats bulk tumor mutational burden for therapy response, with a median AUROC of 0.74 across 10 TCGA cancer-drug pairs. It works only on DNA mutations.

On the metabolism side, CaMeRe ranks metabolic routes by rate-limiting enzyme expression and by how coherently the enzymes along a route are expressed. It does so on bulk TCGA expression and cannot say which cells the enzymes come from. We borrow its scoring idea and apply it within the malignant compartment as a topology-aware alternative to "bag of genes" enrichment.

The PDHK glioblastoma study acts as a constraint. It found PDHK1 in hypoxic cores and PDHK2 at invasive margins. A malignant-intrinsic term therefore still averages over regions within the tumor, and we treat it as a lower bound on resolvable state, not a full picture.

Research Program

We first assemble single-cell references with malignant cells labeled by inferred copy-number changes, following the SF3A3 study's approach. For glioblastoma we use the GSE131928 single-cell cohort used in the gene-pair study. For liver and breast we use the public atlases those studies drew on.

Bulk RNA-seq counts from TCGA, ICGC, and GSE7696 are then deconvolved with Count Bridges. This gives, per patient, cell-type fractions and integer-count profiles for each compartment. On the malignant profiles we compute pathway scores for KEGG and Gene Ontology metabolic sets: glycolysis, oxidative phosphorylation, the hexosamine pathway, lipid transport, inositol phosphate metabolism, and hypoxia response. Each is computed two ways: as a rank-based set score, and as a CaMeRe-style bottleneck-and-coherence score. We compute a malignant proliferation score from the AURKA-centered module defined by Wirapati et al.

Survival enrichment replaces the usual comparison of genes that differ between high-risk and low-risk patients. For each compartment we rank genes by their composition-adjusted Cox association. We then test pathway enrichment along that ranking, which yields a per-compartment pathway hazard map. This directly addresses the enrichment field's stated limitation that bulk enrichment cannot say which cell type ran the pathway.

Baselines

The naive reference point is the bulk pathway score itself: the same metabolic gene set scored on bulk expression and entered into a Cox model with stage, grade, and age. This is what most published metabolic signatures do. It sets the bar for discrimination and is the object whose independence claims we audit.

The strongest existing method for robust metabolic prognosis is the metabolism-related gene-pair LASSO Cox signature of the glioblastoma study. Its reported validation C-index ranges from 0.648 to 0.801 across internal and external cohorts, including GSE7696. We re-implement it on the identical training and validation splits and give it the same hyperparameter search budget as our model. Beating it means matching or exceeding its external C-index while adding a malignant-intrinsic term that survives adjustment for composition and proliferation; the gene-pair signature, by construction, cannot make that claim. For SF3A3 in liver cancer, no reported C-index was found on Lacuna. We rerun the single-gene multivariate Cox analysis on TCGA and ICGC as a second strong reference.

The most dangerous competitor is covariate adjustment. Here the bulk pathway score is adjusted for ESTIMATE purity, CIBERSORT or xCell fractions, and bulk proliferation, all as linear Cox covariates. These fractions are produced by the same tools the gene-pair study used. We give this baseline the same reference atlas wherever its tools accept one. We compare it to our model on the semi-synthetic benchmark below, where the true source of the survival signal is known.

Three ablations isolate the key ingredient, which is cell-resolved state estimation. The first swaps Count Bridges for CIBERSORTx high-resolution expression. The second swaps it for POPPY attribution. The third removes the composition step entirely and restricts pathways to genes that are malignant-specific in the atlas.

Why not just…?

Why not just add purity and immune fractions as Cox covariates? This is a real threat. It is cheap, and it may suffice whenever metabolic genes are expressed mainly by one cell type. It fails in principle when the non-malignant cells also vary in metabolic state, for example macrophages that are more or less glycolytic across patients. Adjusting for how many macrophages are present then leaves their state variation inside the bulk score. The semi-synthetic benchmark settles this directly. If covariate adjustment matches our decomposition at attributing a planted malignant signal versus a planted macrophage-state signal, we will report that. The contribution would then shrink to an audit that uses the cheaper tool.

Why not just analyze single-cell data directly, as the SF3A3 study did? Single-cell cohorts large enough to carry survival follow-up do not exist for these cancers. The SF3A3 work itself had to return to bulk TCGA and ICGC data for its Cox models, and there it lost cell attribution. Our pipeline keeps single-cell resolution through the step where survival is tested.

Why not just restrict signatures to malignant-specific marker genes? The rate-limiting enzymes of glycolysis and hypoxia response are shared across compartments; the file highlights HK2 as a hypoxia-induced gene in placental tissue. Restricting to malignant-only genes would discard most of each pathway. The marker-restriction ablation measures exactly how much prognostic signal that discards.

Why not just use the gene-pair approach for batch robustness? Gene pairs solve cross-platform scaling, not attribution. The ABCA1 macrophage finding in the glioblastoma study shows that gene-pair signatures can be partly non-malignant without the method noticing.

Proposed Experiments

The first decisive check is a semi-synthetic benchmark built from single-cell references. We form pseudo-bulk samples with realistic variation in composition. In some samples we plant a survival effect, simulated through a Cox model with known hazard ratios, in one of three places: malignant glycolysis state, macrophage fraction, or macrophage glycolysis state. Every method must say whether a malignant-intrinsic effect exists. The metrics are the attribution false-positive rate when the effect is planted only in macrophages, and power when it is planted in malignant cells. We sweep reference-atlas completeness by deleting cell states from the reference. This probes the known dependence of deconvolution methods on complete atlases.

The second check is the real-data audit. In TCGA liver cancer (validated in ICGC) and in TCGA glioblastoma (validated in GSE7696), we estimate what share of each bulk metabolic score's hazard comes from composition, malignant state, and their interaction. We report Harrell's C-index, which measures how well a score orders patients by survival time, together with hazard ratios on validation cohorts. A malignant metabolic term is called independent only if it remains significant on external data after stage, grade, age, all fractions, and malignant proliferation are included.

The third check is the application where the stakes are highest: TCGA ER-negative and HER2-positive breast tumors, where Wirapati et al. found that proliferation fails. In ER-positive/HER2-negative disease, proliferation stratified untreated patients with a reported hazard ratio of 2.13. We ask whether the malignant-intrinsic hypoxia or glycolysis term stratifies the high-proliferation subtypes where proliferation does not.

A negative result is informative. If the state term survives nowhere while composition explains most bulk metabolic hazard, that finding directly reinterprets the published metabolic biomarker literature.