Inferring Tumor Carbon Routing by Fusing Respirometry with MR Spectroscopy

A cancer cell that "reprograms" its metabolism changes where carbon goes. Glucose can be burned in the mitochondria, turned into lactate and exported, or sent into pathways that build new cells. Neighboring tumor cells may also trade lactate, with one cell exporting it and another importing it as fuel. The two main tools for watching energy metabolism each see only part of this picture. Extracellular flux analysis measures how fast living cells in a dish consume oxygen and acidify their medium, and it uses timed drug injections to reveal how much respiratory reserve the cells hold. It works only on cells grown in plastic wells in artificial media. Magnetic resonance spectroscopy (MRS) reads the chemical makeup of living tissue without harming it, sorting molecules by the small frequency shifts their nuclei show in a magnet. It mostly reports how much of a molecule is present, not how fast it is made or used. In practice the two are run in separate labs and compared informally. The usual summary, "this tumor is glycolytic," is a label, not a measured flux.

We propose to treat both instruments as noisy, partial views of one hidden quantity: the vector of reaction rates through a mammalian core metabolic network. We build a single Bayesian model in which the in-dish stress test sets a prior on what the cells' mitochondria and glycolysis can do. The in-tissue spectra, with their full uncertainty kept rather than collapsed to point values, supply evidence about what the cells are doing. An explicit term absorbs the gap between dish and tissue. The claim to be tested is specific. Such a model can recover internal carbon routing that neither instrument measures directly, such as the fraction of tumor lactate that is oxidized rather than exported, with calibrated uncertainty. It should do better than direct regression from instrument readouts and better than flux models fed point estimates.

Core Claim and What Is New

The closest existing work comes from three single-pairing studies. Hyperpolarized 13C pyruvate MRS of rat HCC injected a magnetically "supercharged" carbon-13 pyruvate tracer into rats with liver tumors and fitted a two-pool exchange model. It reported a pyruvate-to-lactate conversion rate $k_{pl}$ of 0.041 ± 0.008 s⁻¹ in high-lactate tumors versus 0.028 ± 0.01 s⁻¹ in low-lactate tumors. That study observes one enzymatic step, treats the tumor as one well-mixed pool, and has no respiration readout. As a result, a low $k_{pl}$ cannot be told apart from active oxidation of pyruvate in the mitochondria or from lactate being cleared by neighboring cells.

Bayesian Metabolic Flux Analysis (BMFA) supplies the inference method. It places a truncated multivariate normal posterior over genome-scale reaction rates, relaxes the strict steady-state constraint $Sv = 0$ to a Gaussian penalty, and conditions on measured exchange rates. In leave-one-out prediction of unobserved bacterial fluxes it reached an average F1 of 41%, against 9% for flux variability analysis (FVA). BMFA has never been given mammalian respirometry or spectroscopy.

Physics-informed Sylvester normalizing flows supply the spectral side. They are flexible neural density models constrained by the MRS signal equation, and they return full posteriors over metabolite amounts. On simulated 7T brain spectra they reached a mean absolute error of 0.797 versus LCModel's 1.339, and they exposed strong correlations between overlapping peaks that point fits hide.

None of these studies joins the three pieces. This proposal does three things none of them do. It feeds spectral posteriors, not point values, into a flux posterior. It turns the stress-test numbers into mechanistic priors rather than descriptive phenotypes. It models the tumor as two metabolic compartments exchanging lactate, so that symbiosis becomes an inferable quantity rather than an unmodeled confound.

The Model

Let $v$ be the reaction rates of a core mammalian network covering glycolysis, lactate dehydrogenase, alanine transaminase, pyruvate dehydrogenase, the TCA cycle, oxidative phosphorylation, and lactate transport. Let $S$ be its stoichiometric matrix. In tissue, $v$ is split into two compartments, $v^{(1)}$ and $v^{(2)}$, coupled by lactate exchange. Their unknown mixing fraction is $\phi$.

The joint posterior we target is

$$p(v, \phi, \delta \mid E, M) \propto p(E \mid v_{\text{dish}})\, p(M \mid v, \phi)\, p(v \mid v_{\text{dish}}, \delta)\, p(\delta)\, \mathcal{N}(Sv; 0, \Sigma_v).$$

The symbols are defined as follows:

  • $E$ is the extracellular-flux data: basal, oligomycin, FCCP, and rotenone/antimycin oxygen consumption, plus acidification.
  • $M$ is the MRS evidence. It includes proton and phosphorus-31 spectra and, where available, hyperpolarized carbon-13 time courses.
  • $v_{\text{dish}}$ are the rates of the same cells in culture.
  • $\delta$ is a discrepancy that maps dish rates to tissue rates. It is shrunk toward zero and is shared across cell lines.
  • $\Sigma_v$ is BMFA's tolerance on departures from steady state.

Three modeling choices carry the idea.

Acidification is modeled mechanistically. The likelihood for acidification is written as lactate-linked proton efflux plus protons from respiratory CO₂, through a buffer-capacity term. Oxygen consumption therefore corrects acidification inside the model. Usually this correction, the known way acidification overstates glycolysis, is applied as a separate ad hoc conversion.

The stress test becomes constraints, not features. The FCCP response becomes a soft upper bound on the electron-transport rate. The oligomycin drop constrains the ATP-coupled fraction.

Spectra enter through their posteriors. The MRS likelihood $p(M \mid v, \phi)$ uses spectral posteriors from the normalizing flow. Correlated pairs therefore propagate their joint ambiguity, for example lactate against the overlapping 1.3 ppm lipid signal. For hyperpolarized data, the forward model is a two-compartment extension of the HCC study's exchange model with explicit recirculation.

Inference reuses BMFA's Gibbs sampler from the bamfa codebase on the expanded network. A Hamiltonian Monte Carlo sampler is the fallback if the mixing fraction $\phi$ couples poorly to the other variables.

Research Program

The work proceeds in three stages. Each stage tests the claim against ground truth before the next adds realism.

Stage one: a closed loop in culture. We use the high- and low-lactate HCC cell lines from the hyperpolarized study and several further lines chosen to span oxidative and glycolytic phenotypes. Each line receives an extracellular-flux stress test. In parallel, the same line is tracer-labeled with [U-¹³C]glucose and [3-¹³C]lactate, and conventional ¹³C metabolic flux analysis is run on the extracts. That analysis serves as ground truth for internal routing. We also run proton NMR on extracts and media as the "spectroscopic view." Because every instrument sees the same cells, there is no dish-to-tissue gap, and the stage tests only whether the fusion recovers internal fluxes that no instrument observes directly.

Stage two: identifiability on synthetic data. Before any animal work, we simulate two-compartment tumors with known mixing fraction and known lactate shuttle rate. We render them through the spectral and hyperpolarized forward models across the SNR range of the simulated 7T benchmark (4–68 dB). We then ask whether the shuttle rate is identifiable at all. The 9.4T dividing-cell study acts as a constraint here. Its spectra did not detect a threefold increase in neurogenesis but did detect glioblastoma, where about 22% of tissue was dividing cells. This warns that small compartments may be invisible to spectra. We will map the fraction below which the posterior on the minority compartment falls back to its prior, and we will report that threshold rather than claim beyond it.

Stage three: transfer to tissue. We implant the same lines subcutaneously in nude rats. We acquire hyperpolarized pyruvate time courses and proton/phosphorus-31 spectra. We then infuse ¹³C-lactate before resection, so that ex vivo isotope labeling of tumor extracts gives a ground-truth lactate-oxidation fraction. Ex vivo LDH and transaminase activity ratios, measured as in the HCC study, serve as a secondary check.

Baselines

Naive reference: population prior only. This is the flux posterior conditioned on the network and measured glucose uptake alone, with no line-specific information. It sets the floor that any informative model must beat.

FVA. Flux variability analysis is standard linear-programming bound-finding on the same network with the same exchange constraints. It is the field's default. The only reported comparison is BMFA's 41% versus FVA's 9% F1 on C. acetobutylicum. For mammalian cells conditioned on respirometry and spectra, no reported result was found on Lacuna, so both are rerun here.

BMFA with respirometry only. This is BMFA as published, conditioned only on PER-corrected exchange rates derived from the stress test. It is the strongest existing method and the most direct competitor. It shows how much the spectra add.

BMFA with LCModel point estimates. This is the same full model, but spectra enter as fixed point values with Cramér-Rao bounds treated as independent Gaussian errors. It is the key ablation: it isolates whether carrying spectral posteriors forward, the proposal's central ingredient, matters. For reference, the reported spectral-level gap is MAE 0.797 versus 1.339 on simulated 7T spectra. Whether that gap survives into flux estimates is unknown and is what this ablation measures.

Direct regression. This is a gradient-boosted regressor and a ridge regressor mapping the concatenated stress-test parameters and fitted spectral features straight to each ground-truth flux. It has no network model. It is the most plausible simple threat.

$k_{pl}$ alone. In tissue, the hyperpolarized conversion rate is used as a scalar index of glycolysis, as in the HCC study.

Fairness. All methods see identical inputs per line. All receive the same hyperparameter budget, tuned by nested leave-one-cell-line-out cross-validation. Sampling budgets for the Bayesian variants are matched by effective sample size.

What counts as beating them. On held-out cell lines, the full model must beat respirometry-only BMFA, the regression, and the point-estimate ablation on three criteria. First, it must have lower normalized error on the five routing fluxes not directly measured: pyruvate dehydrogenase, lactate oxidation, TCA turnover, the pentose phosphate split, and anaplerosis. Second, it must reach the leave-one-out F1 used in BMFA. Third, its nominal 90% intervals must cover truth between 85% and 95% of the time. "Beating" means doing so with paired bootstrap confidence intervals that exclude zero.

Why not just…?

Why not just regress the ¹³C fluxes on the instrument readouts? With a few dozen cell lines, regression may well match the model on fluxes strongly tied to a single readout, such as glycolytic rate from proton efflux. That is a real threat. The proposal's case rests on fluxes that no single readout fixes, and on extrapolation to unseen lines and to tissue, where a network-constrained model should degrade more gracefully. Stage one settles this directly. If regression equals the model on the routing fluxes under leave-one-line-out validation, we will report that the mechanistic model buys only calibration, and we will narrow the claim accordingly.

Why not just apply the standard proton-efflux correction to acidification and read glycolysis off the stress test? That gives a cleaner glycolytic rate in culture. It still says nothing about tissue, and it cannot split lactate production from lactate oxidation. The review on glycolysis and OXPHOS in cancer argues that lactate oxidation is exactly where plasticity and therapy-induced rebound occur. The respirometry-only BMFA baseline already includes this correction, so its gap to the full model measures what the spectra add.

Why not just use FDG PET, or hyperpolarized $k_{pl}$, as the in vivo glycolysis readout? FDG reports glucose uptake and trapping only. Work on high-FDG lung adenocarcinoma shows that high uptake can reflect hexosamine routing rather than lactate production. $k_{pl}$ reflects one exchange step and is confounded by mitochondrial pyruvate use. The stage-three comparison tests whether $k_{pl}$ alone ranks tumors by ex vivo lactate-oxidation fraction as well as the fused posterior does. If it does, the simpler readout wins.

Why not just run the stress test on fresh tumor slices instead of modeling the dish-to-tissue gap? Slice respirometry is a reasonable complement. It is still ex vivo and destructive, so it gives one time point per animal, and it does not remove the need to explain the in vivo spectra. Where feasible, we will run it as an additional check on the learned discrepancy $\delta$.

Decisive Checks

The first decisive check is the stage-one comparison on routing fluxes against regression and against the point-estimate ablation. If carrying spectral posteriors forward gives no gain over point estimates, the central methodological claim fails.

The second is the stage-two identifiability map. If the lactate shuttle between compartments is not recoverable at realistic SNR, we drop the two-compartment claim and keep the single-pool fusion.

The third concerns transfer. The posterior on the tissue lactate-oxidation fraction must rank tumors in the same order as the ex vivo ¹³C-lactate measurement, with Spearman correlation reliably above that of $k_{pl}$ alone. In addition, the high- versus low-lactate contrast reported in vivo must be reproduced from culture-plus-spectra inputs without using that tumor's own isotope data.