Thousands of transcriptomic studies conclude that cells have undergone metabolic reprogramming because a gene-set test (a statistical check of whether a curated group of genes, such as "glycolysis," is over-represented among changed genes) comes back significant. That conclusion assumes that more transcripts for pathway enzymes means more chemical flow through the pathway. Nobody has measured how often this holds. There is a specific reason to expect it to fail. When a drug or stress chokes a pathway, cells often react by making more of that pathway's enzymes. The transcript signal can then point the opposite way from the flux. We propose to build a matched dataset of cell perturbations. Each one would be profiled by RNA sequencing and by live-cell measurements of oxygen use and acid output. With it we would measure how often standard enrichment calls get the direction of flux change right, when they fail, and whether a direction-aware pathway score repairs the failures.
The core claim is narrow and testable. Enrichment scores track flux sign reliably only when the perturbation acts upstream of transcription, such as hypoxia signalling or gene knockdown. They fail systematically, sometimes with the wrong sign, when the perturbation acts on enzymes directly and transcription responds by compensating. A score that knows which genes control flux and in which direction should recover part of that loss. The same matched data also lets us quantify a second, practical failure. The most common viability assays read out dehydrogenase activity, so they can report a metabolic shift as a change in cell number.
What exists and what is missing
Tools that infer metabolic rewiring from expression have never been checked against physical flux. CaMeRe (CaMeRe) weights the HumanCyc reaction network with TCGA expression and ranks routes by how highly their bottleneck enzymes are expressed. It recovers 7 of 8 known cancer pathways in its top three results, but its benchmark is literature recall, not measured flux. SNODEP (Predicting time-varying flux and balance in metabolic systems using structured neural-ODE processes) predicts time-varying flux from single-cell time series. Its "ground truth" is itself a computational flux estimate derived from expression, so the evaluation is circular with respect to our question.
The enrichment field has moved toward better gene sets, but it validates them only against knowledge bases. GSPA (Gene set proximity analysis) expands sets with protein-interaction neighbours and beats GSEA on AUPRC in 75.8% of GEO2KEGG datasets. gtCLIP (Contrastive Learning for Gene Set Enrichment Analysis Post-Processing) clusters redundant enriched sets using text embeddings. Neither tool encodes whether a gene speeds or slows its pathway.
Precily (Gene expression based inference of cancer drug sensitivity) shows that pathway-level scores computed with gene set variation analysis (GSVA) predict drug IC50 well. Its endpoint, though, is static viability rather than bioenergetics.
On the measurement side, three studies show what live and spatial readouts can reveal. Extracellular flux analysis (Extracellular Flux Analysis of Cellular Bioenergetics) measures oxygen consumption rate (OCR, a readout of mitochondrial respiration) and extracellular acidification rate (ECAR, a readout of glycolytic lactate export) in living cells. It can split these into basal, ATP-linked and spare capacity with sequential drug injections. FLIM imaging of NAD(P)H (Quantitative Spatial Analysis of Metabolic Heterogeneity) shows that metabolic state varies spatially in 3D cultures. The AML pharmacoscopy study (Single-cell landscape of innate and acquired drug resistance in AML) links an OXPHOS shift to venetoclax resistance. None of these three tests enrichment calls against flux.
The biology that motivates the "wrong sign" hypothesis comes from two sources in the file. PDHK enzymes inhibit pyruvate entry into mitochondria (Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development). A "bag of genes" score therefore counts their upregulation toward oxidative-metabolism terms when it actually lowers oxidative flux. And glycolysis and OXPHOS coexist in hybrid states rather than one replacing the other (Absolute dynamic and relative static).
Research Program
We would generate a perturbation atlas. It would cover eight cell lines chosen to match contexts in the file: glioblastoma, clear cell renal carcinoma, pancreatic ductal adenocarcinoma, hepatocellular carcinoma, breast, and colorectal lines, plus two non-transformed lines.
Each line receives about twenty perturbations, falling into three classes. Upstream-of-transcription perturbations are hypoxia at 1% oxygen, HIF stabilisation, and siRNA against HK2, PDK1, ACOT8 (ACOT8) and ABCG1 (ABCG1). Direct enzyme or transporter inhibitors are 2-deoxyglucose, dichloroacetate, metformin, low-dose oligomycin, and glutamine withdrawal. Non-metabolic negative controls are cytoskeletal and cell-cycle perturbations that should not move flux.
Each condition is sampled at 6, 24 and 72 hours, because transcript and flux responses can separate in time. Every well gets matched readouts:
- bulk RNA-seq;
- an extracellular flux mitochondrial stress test and glycolytic rate assay, normalised to imaged nuclear counts;
- EdU incorporation, which labels cells actively copying DNA;
- direct nuclear counting;
- CCK-8, the common dehydrogenase-based viability dye;
- total ATP.
The flux readouts give three ground-truth targets per condition: change in basal ECAR, change in basal OCR, and change in spare respiratory capacity. Each is labelled up, down or unchanged with replicate-based confidence intervals.
The proposed method is a signed, control-weighted pathway score. We restructure HumanCyc (CaMeRe) glycolysis, TCA, OXPHOS, glutaminolysis and fatty-acid oxidation modules in four ways:
- Each gene carries a sign, so it counts toward the pathway's flux if it is an activator and against it if it is an inhibitor; PDK1–4 count against PDH flux.
- Each gene carries a weight reflecting whether it sits at a committed or rate-limiting step, using CaMeRe's bottleneck logic.
- Transcript changes are converted into a predicted flux direction per module.
- A small regression maps module scores to the three flux targets. It is trained only on the upstream-of-transcription class and then tested on the held-out inhibitor class.
Training on one class and testing on the other directly tests whether the score transfers across mechanism. We also evaluate leave-one-cell-line-out.
Baselines
The do-nothing reference predicts "no flux change" for every condition. A second naive reference predicts the majority sign for the perturbation's class. Any enrichment method must beat both on balanced sign accuracy, or it adds nothing beyond knowing what drug was added.
The practice baseline is what papers actually do. We run ORA (a Fisher's exact test for over-representation of changed genes) on KEGG glycolysis and OXPHOS terms, and we read the direction of the changed genes as the flux direction. The rank-based variant is GSEA, scored the same way. GSVA scores of the kind Precily uses (Precily) are fed into the same regression head as our score, so both get an identical learned calibration step. Precily reports ρ = 0.88 and R² = 0.77 for GDSC IC50 prediction. That task is different, and we found no reported result on Lacuna for predicting OCR or ECAR change from transcriptomes, so our study would set that number.
The strongest existing enrichment method is GSPA's network-expanded sets (GSPA), using the published radius setting and the same regression head.
The most dangerous baseline is a hand-picked marker panel: HK2, LDHA, SLC2A1, PDK1 and a few OXPHOS subunit genes, signed by textbook knowledge.
All methods see identical normalised counts and identical cross-validation splits. Any tunable parameter, such as the GSPA radius or regression regularisation, gets the same grid budget tuned on training folds only.
Two ablations isolate our ingredients. Removing gene signs tests direction-awareness, and flattening the weights tests bottleneck weighting. A label-shuffled-sign control checks that any gain is not just from having fewer, cleaner genes.
Beating the field means three things. On the held-out inhibitor class, our score's balanced sign accuracy must exceed both ORA/GSEA and the marker panel, with confidence intervals that do not overlap across cell-line bootstraps. It must also improve Spearman correlation with flux magnitude. And the signed-weight ablation must account for most of the gain.
Why not just…?
Why not just run a flux assay whenever metabolism matters? For new experiments, that is the right answer, and our results should make that recommendation quantitative. The audit still matters for two reasons. First, most existing transcriptomic claims of reprogramming, including the many enrichment-only studies in the file, have no flux data and cannot be rerun. Second, a measured error rate per perturbation class tells readers which of those claims to trust. The study is also the only way to check whether "just run flux" is needed always or only for inhibitor-type perturbations.
Why not just use a handful of marker genes? This is a real threat. For glycolysis, HK2 and LDHA may carry most of the signal; chronic hypoxia reliably raises HK2 (Elevated expression of glycolytic genes in early-onset preeclampsia). The marker panel is therefore run as a full baseline under identical calibration. If it matches our signed score on the held-out inhibitor class, we drop the topology claim and report that a signed marker panel is sufficient. That outcome is still useful, because the panel is signed and unsigned enrichment is not.
Why not just use computational flux estimators instead of measuring flux? Estimators like those SNODEP trains against (SNODEP) are themselves computed from expression. If compensatory transcription flips the sign, they inherit the same error. We include such an expression-derived flux estimate as an additional baseline, and we expect it to fail in the same conditions where ORA fails.
Why not just pair CCK-8 with a cell count instead of predicting discordance? We agree that this is the fix. Our contribution is to measure how often the two disagree under metabolic perturbation. CCK-8 is used as a viability readout without such pairing in, for example, the miR-10b-3p hepatocellular carcinoma study. We also test whether a redox or dehydrogenase enrichment score flags the disagreeing conditions in existing datasets that lack counts.
Proposed Experiments
The first decisive check is the sign-flip test. For 2-deoxyglucose, oligomycin and metformin at 24 and 72 hours, does measured ECAR or OCR move opposite to the enrichment-implied direction in a meaningful fraction of cell lines? The comparison is against hypoxia and siRNA conditions, where we expect agreement. If ORA sign accuracy is high and indistinguishable across the two classes, the central hypothesis is false. We would then report the reassuring result that enrichment calls are trustworthy.
The second check is held-out transfer. The signed score is trained on upstream perturbations and evaluated on inhibitors against every baseline above. Accuracy is reported per flux target and per time point, so we can see whether agreement is best at a particular lag.
The third check is the viability discordance study. For every condition we compute the ratio of CCK-8 signal to nuclear count and to EdU fraction. We then test whether conditions with large discordance are predicted by the measured OCR change and by enrichment of NAD(P)H-producing dehydrogenase sets. This separates metabolic artifact from true growth change.
A final, smaller extension repeats a subset of conditions in 3D spheroids with NAD(P)H lifetime imaging, following the FLIM heterogeneity study. It asks whether bulk enrichment accuracy degrades when metabolic states are spatially mixed. This extension is reported separately and does not bear on the core claim.
References
- CaMeRe: A Novel Tool for Inference of Cancer Metabolic Reprogramming
- Predicting Time-Varying Flux and Balance in Metabolic Systems using Structured Neural ODE Processes
- Gene set proximity analysis: expanding gene set enrichment analysis through learned geometric embeddings
- Contrastive Learning for Gene Set Enrichment Analysis Post-Processing
- Gene expression based inference of cancer drug sensitivity
- Eugenol inhibits oxidative phosphorylation and fatty acid oxidation via downregulation of c-Myc/PGC-1β/ERRα signaling pathway in MCF10A-ras cells
- Quantitative Spatial Analysis of Metabolic Heterogeneity Across in vivo and in vitro Tumor Models
- Single-cell landscape of innate and acquired drug resistance in acute myeloid leukemia
- Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development
- Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development
- Acyl-CoA thioesterase 8 induces gemcitabine resistance via regulation of lipid metabolism and antiferroptotic activity in pancreatic ductal adenocarcinoma
- ABCG1 promotes the proliferation and migration of clear cell renal cell carcinoma and reduces its apoptosis
- Elevated expression of glycolytic genes as a prominent feature of early-onset preeclampsia: insights from integrative transcriptomic analysis
- Up‐regulation of miR‐10b‐3p promotes the progression of hepatocellular carcinoma cells via targeting<scp>CMTM</scp>5