A Substrate Stress Test for Deconfounding Metabolic Reprogramming

When biologists measure the metabolism of living cells, they usually do it with extracellular flux analysis — a live assay in which optical sensors hover a few dozen micrometers above a cell monolayer, forming a temporary sealed microchamber, and track how fast the cells consume oxygen (the oxygen consumption rate, OCR) and acidify their surroundings (the extracellular acidification rate, ECAR). The catch is that the cells sit in an artificial bath whose nutrient composition was chosen for instrument convenience — unphysiologically high glucose, no fatty acids — and when the cells come from human donors, almost nothing is recorded about what that donor ate or how long they fasted before the blood draw. Clinical epidemiology has documented for years what this kind of omission costs at the organism level: unrecorded diet can completely mask a genetic defect in fatty-acid processing, so that carriers of the slow FADS2 rs174575 allele look normal whenever dietary intake happens to be high in preformed long-chain polyunsaturated fats Effects of the rs174575 single nucleotide polymorphism in FADS2.... Our claim is that the same confounding operates inside the flux assay itself: the medium is an unmeasured treatment, the donor's nutritional history is an unmeasured prior state, and apparent metabolic reprogramming — the active rewiring of fuel use that stressed, cancerous, or activated immune cells perform — can be created or erased purely by substrate supply.

We propose to build a substrate stress test as the mirror image of the mitochondrial stress test. The mitochondrial stress test probes respiratory machinery with drugs (oligomycin, FCCP, rotenone/antimycin A) to reveal ATP-linked respiration, maximal capacity, and spare capacity; the substrate stress test probes fuel-handling machinery with nutrients. Candidate fuels — glucose, pyruvate, lactate, glutamine, a palmitate–albumin complex, ketone bodies, and succinate — are injected sequentially through the same drug ports, and the immediate OCR/ECAR response to each defines a nutrient response fingerprint. Coupled to a probabilistic flux model that treats exogenous nutrient supply as a latent variable, this fingerprint lets us separate cell-intrinsic reprogramming from substrate availability, and to detect fuels that leave no trace in expression data.

Core Claim

Apparent reprogramming phenotypes measured by standard extracellular flux analysis are not portable across substrate conditions or donor nutritional states. Specifically: (i) a cell's glycolytic-versus-oxidative classification can flip when the assay medium is moved toward physiological composition; (ii) the donor's fasted-versus-fed state leaves a measurable signature in the nutrient response fingerprint of their isolated immune cells; and (iii) combining the fingerprint with a substrate-conditioned Bayesian flux model predicts unobserved intracellular fluxes better than basal flux data alone or static expression-based inference. The motivation is concrete. The MCPyV small T antigen study found that viral transformation doubles glycolytic acidification and makes the lactate exporter MCT1 an indispensable "exhaust pipe," yet evaluated this in a single nutrient-replete medium, never asking whether lactate-rich or acidified conditions — the actual tumor microenvironment — change that dependence Merkel Cell Polyomavirus Small T Antigen Promotes Pro-Glycolytic Metabolic Perturbations.... The adenosinergic NK cell study showed that a tumor metabolite suppresses both OCR and ECAR while interferon transcripts paradoxically rise — expression and flux disagree — again in uniform, high-nutrient media with unmeasured donor backgrounds Adenosinergic Signaling Alters Natural Killer Cell Functional Responses. And the 2P-FLIM macrophage study documented substantial donor-to-donor variability in metabolic phenotypes while recording nothing about the donors' nutritional state Non-invasive classification of macrophage polarisation by 2P-FLIM and machine learning. These are not isolated sloppinesses; they are the field's standard operating procedure.

Research Program

The program has three components: an assay, a model, and a donor cohort.

The assay extends the standard Seahorse-style workflow. After a standard mitochondrial stress test on each well, we run the substrate roster: a small bolus of one candidate fuel, a short equilibration, measurement of the OCR/ECAR transient, then washout by the cartridge lift before the next injection. For each injected substrate $s$ we record the response coefficient

$$r_s = \frac{\mathrm{OCR}(s^{+}) - \mathrm{OCR}_{\text{base}}}{\mathrm{OCR}_{\text{base}}},$$

where $\mathrm{OCR}(s^{+})$ is the oxygen consumption rate in the minutes after injecting $s$. The vector of $r_s$ values across the roster, plus the ECAR counterparts, is the nutrient response fingerprint. Substrates like succinate are included deliberately: Brandes et al. showed that expression-constrained flux models cannot detect carbon sources that enter central metabolism without dedicated entry enzymes — succinate failed completely in their bacterial inference — whereas a respiratory surge upon succinate injection is directly visible to flux sensors Correction: Inferring Carbon Sources from Gene Expression Profiles Using Metabolic Flux Models. We invert their logic: instead of inferring the nutrient environment from expression, we infer it from kinetic responses.

The model extends Bayesian Metabolic Flux Analysis, which reformulates flux estimation as probabilistic inference — placing Gaussian priors on fluxes, relaxing the rigid steady state, and sampling a truncated multivariate normal posterior over all intracellular fluxes, achieving an F1 score of 41% in predicting unobserved intracellular fluxes from exchange data versus 9% for standard Flux Variability Analysis Bayesian Metabolic Flux Analysis reveals intracellular flux couplings. Our extension adds a latent exogenous supply vector $u$ (uptake capacities for each roster substrate) and conditions the likelihood on the injection experiment:

$$p(v, u \mid y) \propto p\!\left(y \mid v, u\right)\, p(v)\, p(u),$$

where $v$ is the intracellular flux vector on a human genome-scale network, $y$ collects the measured OCR/ECAR transients and basal exchange rates, and $p(u)$ encodes physiologically plausible plasma ranges. The posterior over $u$ is the model's estimate of the cell's effective nutritional environment; covariance between $u$ and $v$ — the property Heinonen et al. established — is what lets exchange data constrain internal pathway usage. We deliberately do not use rank-based metabolite harmonization as the modeling backbone: MIRTH imputes ~89% of masked metabolites with positive rank concordance across nine pan-cancer datasets, but rank transformation permanently discards absolute concentrations, making its outputs unusable in stoichiometric balance equations MIRTH: Metabolite Imputation via Rank-Transformation and Harmonization. Stoichiometry is exactly what we need.

The donor cohort addresses nutritional confounding at its source. We recruit healthy donors in a within-person crossover: each donor is sampled once after an overnight fast and once after a standardized meal, with plasma glucose, lactate, non-esterified fatty acids, ketones, and the arachidonic-acid-to-DGLA ratio (the FADS2-relevant lipid pair) measured at each draw, plus rs174575 genotyping. From each draw we differentiate monocytes into macrophages (the system of the 2P-FLIM benchmark) and isolate NK cells, then run the substrate stress test. The crossover design holds genetics fixed, so any flux difference between the fasted and fed visits is nutritional, not genetic — the in vitro counterpart of the gene–diet masking that Motong Liu et al. could only observe across cohorts. This design is forced on us by the epidemiological evidence: simple covariate adjustment fails when confounding is non-linear, as in the U-shaped mortality curve for serum uric acid in hemodialysis, where low uric acid marks protein-energy wasting rather than protection paper on uric acid and mortality in hemodialysis, and self-reported intake is systematically biased relative to doubly labelled water measurement doubly labelled water analysis.

A secondary system tests reprogramming directly: 3D human adipocyte spheroids in Matrigel, the model Wagner et al. used to show that a lipid mixture blunts maximal respiration and that rosiglitazone rescues it — but with only two donors and one lipid condition In vitro model of human subcutaneous adipocyte spheroids.... We replicate their lipid-overload challenge across a graded palmitate series crossed with varying background glucose and lactate, asking whether the "lipid overload phenotype" and its rescue are stable or substrate-dependent. No reported numerical flux rates were found on Lacuna for that study, so we re-measure its conditions under our protocol to obtain comparable numbers.

Baselines

The naive reference point is current standard practice: the standard single-medium extracellular flux assay, with cells washed into unbuffered assay medium at fixed high glucose and no donor nutritional information recorded — the protocol used by the Wagner adipocyte study and the adenosinergic NK study cited above. We run it on the very same donors, plates, and instrument sessions, so any difference is attributable to the protocol, not the sample. Beating it means showing that its readouts do not transport: concretely, that a polarization classifier trained on standard-medium flux data loses substantial accuracy when applied to identical cells in physiological media variants, and that basal OCR/ECAR alone cannot distinguish fasted from fed donors while the fingerprint can.

The strongest existing methodological competitor is Bayesian Metabolic Flux Analysis itself Bayesian Metabolic Flux Analysis reveals intracellular flux couplings, which reported an F1 score of 41% when predicting unobserved intracellular fluxes from exchange measurements in Clostridium acetobutylicum, against 9% for Flux Variability Analysis. That 41% is the number to beat in spirit, but it was obtained on bacterial $^{13}$C data, so for a fair comparison we re-implement BMFA with its published sampler settings and give it only basal exchange fluxes from our human cells, with the same compute and tuning budget as our model. Success is a materially higher held-out F1 for our substrate-conditioned version — we pre-register a target of at least 0.55 — with held-out fluxes validated against in-house $^{13}$C labeling or NAD(P)H fluorescence-lifetime measurements as ground truth, the same pairing of optical and flux modalities used in the 2P-FLIM macrophage work.

The strongest competitor for the reprogramming-inference question is expression-based pathway ranking: CaMeRe, which searches a HumanCyc-scale network weighted by TCGA transcriptomes and recovered 7 of 8 canonical cancer metabolic pathways (87.5%) within its top three ranks CaMeRe: A Novel Tool for Inference of Cancer Metabolic Reprogramming. We run CaMeRe on matched RNA-seq from our macrophage, NK, and adipocyte cultures under the same conditions we assay by flux, with equal tuning effort. Beating it means that where expression-ranked pathway activity and measured flux response disagree, targeted perturbation experiments — for example MCT1 blockade in the manner of the MCPyV study — side with the flux-based call more often than the expression-based call.

Finally, an ablation isolates each ingredient on the same donor material: the full model fed only basal exchange data (removing the substrate roster), and the raw fingerprint fed to a plain classifier with no stoichiometric model (removing the Bayesian coupling). The proposal's specific contribution is only validated if the full pipeline beats both ablations on held-out flux prediction and donor-state classification.

Why not just…?

Why not just run the standard assay in a few different media and compare? Honestly, the first phase of the experimental program is that comparison, because it is the cheapest way to establish that the confounding exists at all. But the simple factorial stops at detection: without a stoichiometric model you cannot tell which internal pathways shifted when the medium changed, only that the two bulk rates moved, and without substrate injections you cannot see fuels that enter central metabolism directly — the succinate blind spot that Brandes et al. demonstrated for expression models applies to basal bulk rates too, since succinate oxidation changes OCR only when succinate is actually supplied. If, in the ablation, the raw fingerprint without the flux model matches the full pipeline's held-out prediction accuracy, the modeling layer is unnecessary and we will say so; that ablation is precisely the decisive check.

Why not just infer reprogramming from RNA-seq, as CaMeRe does? Because expression and flux are demonstrably decoupled at exactly the points that matter here. The adenosinergic NK study found transcripts and protein for interferon-gamma going up while the cells' actual bioenergetics and tumor-killing capacity collapsed — an expression-only reading would have concluded hyper-activation. Enzyme abundance also does not determine flux: pyruvate dehydrogenase kinase phosphorylates and shuts off pyruvate entry into the TCA cycle allosterically and post-translationally, with no necessary change in transcript levels Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development, and a recent review emphasizes that tumors adopt hybrid, dynamically shifting glycolysis/oxidation phenotypes that static snapshots mislabel Absolute dynamic and relative static.... The CaMeRe comparison in our experiments settles this empirically rather than by argument.

Why not just record donors' diets and add them as covariates? Because the epidemiological record shows linear adjustment fails on this exact problem. The FADS2 meta-analysis found a standardized mean difference of −0.7 in arachidonic acid for minor-allele carriers that vanishes in populations with high dietary PUFA intake — a gene-by-diet interaction, not an additive shift FADS2 rs174575 meta-analysis; the hemodialysis uric-acid curve is U-shaped, not linear hemodialysis uric acid study; and self-reported intake is biased against objective doubly labelled water measurement doubly labelled water. The multicenter carbohydrate-loading trial in 1,953 women in labor illustrates the interventional version of the same problem: pre-enrollment energy reserves were completely unrecorded, so the null result on cesarean delivery (11.3% versus 10.9%, RR 1.04) cannot be separated from baseline masking carbohydrate-rich beverage RCT. Our crossover design with objective plasma measurements at the draw replaces the questionnaire altogether. A residual threat remains — donor nutritional state may imprint weakly after days of ex vivo culture, in which case the donor signal will be small; the crossover with matched plasma metabolomics is designed to detect even a modest imprint, and a null result there would itself be a publishable, useful negative finding for assay standardization.

Proposed Experiments and Decisive Checks

Experiment 1 establishes the confounding. M1/M2 macrophage polarization and NK-cell activation are assayed by the substrate stress test in a media matrix spanning physiological glucose (5 mM versus the standard 10–25 mM), lactate (0 versus 10 mM), and palmitate (0 versus 0.4 mM). Decisive outcome: the polarization classifier benchmarked at a reported ROC-AUC of 0.944 in the 2P-FLIM macrophage study (out-of-bag error 16.67%) is retrained on standard-medium data and evaluated across media variants; a drop in AUC of more than about 0.1 in any variant demonstrates non-portability of the phenotype, and the fingerprint features that recover the classification identify which fuels drive the flip.

Experiment 2 tests the donor-nutrition signal. In the fasted/fed crossover, we test whether the nutrient response fingerprint classifies visit state within donors above chance, and whether plasma metabolites at the draw predict the fingerprint. Decisive outcome: within-donor classification accuracy significantly above 50% across donors, with the arachidonic-acid/DGLA ratio as a mechanistically interpretable predictor given the FADS2 result of SMD −0.7 for arachidonic acid in minor-allele carriers.

Experiment 3 benchmarks the model. Held-out intracellular fluxes (masked exchange and internal reactions, validated against $^{13}$C labeling or NAD(P)H lifetimes) are predicted from basal-only data through re-implemented BMFA versus the full substrate-conditioned model. Decisive outcome: the full model exceeds the BMFA baseline's F1 on our data; the published 41% versus 9% comparison in C. acetobutylicum BMFA sets the scale, and the basal-only ablation of our own model isolates how much the substrate roster itself contributes.

Experiment 4 adjudicates expression versus flux. CaMeRe-style ranking on matched RNA-seq is compared with measured flux shifts, with MCT1-blockade and PDHK-inhibitor perturbations as adjudicating interventions in the adipocyte and macrophage systems. Decisive outcome: flux-based pathway calls are confirmed by perturbation more often than expression-based calls, quantified as a concordance rate difference with confidence intervals.

If Experiments 1 and 3 succeed, the deliverable is a standardized, open protocol and dataset that turns the assay medium and donor nutritional state from unmeasured confounders into measured, modelable variables — the same discipline clinical epidemiology learned the hard way, applied one layer down, at the live cell.