Is Inferred Enzyme Regulation a Shadow of Cell-to-Cell Variation?

One way to model a metabolic network when its enzyme rate constants are unknown is to assume that it settles into the steady state that dissipates energy fastest. This is the maximum entropy production principle: among the steady states a driven system could reach, it favors the one that produces entropy at the highest rate. Cannon and colleagues used this idea on the central metabolism of the fungus Neurospora crassa (Prediction of Metabolite Concentrations, Rate Constants and Post-Translational Regulation…). Their method first finds the flux pattern that maximizes entropy production without using any rate constants. It then compares the predicted metabolite concentrations with measured physiological ranges. Wherever the two disagree, it infers that an enzyme is being switched down by regulation. In that study, the inference recovered allosteric control at phosphofructokinase (PFK) and mitochondrial pyruvate dehydrogenase (PDHm).

That inference rests on two assumptions. The cells are treated as one homogeneous population, and the comparison rests on one organism under one condition. The cell-culture literature suggests the first assumption can fail in ways that matter. When individual cells are allowed to occupy different metabolic phenotypes, the population's steady-state behavior changes qualitatively (Maximum entropy and population heterogeneity in continuous cell cultures). This proposal asks whether a model that sees only population-averaged ("bulk") measurements would mistake that averaging for regulation. We also want to measure how often it does so, so that the N. crassa result stops being a single case study and gets an error rate.

Core Claim

Mass-action thermodynamics computes driving forces from the logarithms of concentrations, but bulk assays report the arithmetic mean of concentrations over cells. Suppose a fraction $w_k$ of cells sits in subpopulation $k$, with concentration $c_{j,k}$ of metabolite $j$. The gap between what bulk measurement reports and what an averaged thermodynamic state would imply is

$$ \Delta_j = \log\Big(\sum_k w_k c_{j,k}\Big) - \sum_k w_k \log c_{j,k} \;\ge\; 0, $$

by Jensen's inequality (the log of an average is never smaller than the average of the logs). For modest spread, $\Delta_j$ is approximately half the cell-to-cell variance of $\log c_j$, which follows from a second-order Taylor expansion of the logarithm.

The bias always points the same way. Bulk metabolites look more concentrated than any homogeneous steady state predicts. To a homogeneous model, that pattern is exactly the signature of a slowed enzyme just downstream. We therefore predict three things. First, a homogeneous entropy-production model run on bulk data will produce spurious regulation calls, and their number will scale with population variance. Second, the spurious calls will sit just downstream of the most variable metabolites. Third, a model that represents the population as a small mixture of subpopulations, each held to the entropy-production principle, can tell genuine regulation apart from averaging artifacts.

What is new here is not applying maximum entropy to heterogeneous populations, which Fernandes et al. already did for phenotype distributions. It is showing, and then correcting, a specific failure mode of the rate-constant-free kinetic inference that the homogeneous method relies on.

How the Three Directions Fit

The entropy-production principle is the mechanism being tested. Cellular heterogeneity masking is the suspected confound. The deconvolution literature shows that this confound produces systematic, structured misattribution rather than random noise. In bulk RNA deconvolution, for example, most methods mislabel normal epithelium as cancer epithelium as tumor purity rises (Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures). CSsingle shows that ignoring cell-to-cell differences in RNA content biases cell-type estimates in predictable directions (CSsingle). Neither paper concerns metabolism, so both serve only as motivation by analogy.

The single-case-study limitation is our evaluation lens. Clinical case reports such as the 94.0% tumor-volume reduction under 5-FU/leucovorin (A resected case of advanced esophageal carcinoma…) show that something can happen, but they have no denominator. Cannon's PFK/PDHm inference is the same kind of evidence: it proves the method can recover known regulation once. It says nothing about how often the method invents regulation that is not there. We supply that denominator.

We deliberately do not take up the tempting three-way idea of scoring individual patients' tumors as stable or unstable entropy-production attractors. Nothing in this file validates entropy-production state selection in tumors. Even in lattice models the principle is confirmed only at very small sizes ($L \le 10$) (MEP vs. Kolmogorov-Sinai Entropy in a Constrained ASEP Model). Building clinical claims on it would compound unverified assumptions.

Research Program

The work has three stages.

Stage 1: a ground-truth generator. We take two networks with known structure. The first is Cannon's N. crassa subnetwork covering glycolysis, the pentose phosphate pathway (PPP) and the TCA cycle. The second is a reduced core of the CHO-K1 genome-scale model used by Fernandes et al. For each network we build synthetic cell populations in which we control two things independently: the true set of regulated enzymes, and the degree of heterogeneity.

Heterogeneity enters in two ways. In the first, we sample discrete subpopulations, each at its own entropy-production-maximizing steady state under different boundary conditions. The boundaries vary in glucose-6-phosphate and NADPH demand, the latter mimicking nitrogen limitation. In the second, we draw continuous phenotype distributions with the Fernandes expectation-propagation sampler at finite heterogeneity parameter $\beta$, and then compute concentrations from them. Per-cell concentrations are averaged with known weights to produce bulk profiles. We then add measurement noise matched to the mass-spectrometry ranges in Cannon et al.

Stage 2: auditing the homogeneous method. We run Cannon's four-step pipeline unchanged on each bulk profile and record which enzymes it calls regulated. Because the true answer is known, this yields precision and recall for regulation calls as a function of variance. It also lets us test the localization prediction: are false calls concentrated downstream of the metabolites with the largest $\Delta_j$?

Stage 3: the mixture model. The proposed method replaces the single steady state with $K$ steady states, each obeying the entropy-production principle. These states share the same network and the same rate constants, and differ only in boundary fluxes. The model fits the subpopulation weights and the regulation set jointly to the bulk data, and chooses $K$ and the regulation set by held-out likelihood.

The known stiffness of the solvers when regulation constraints are removed is a real cost. We will report wall-clock time, use implicit solvers throughout, and warm-start each subpopulation from the homogeneous solution.

Baselines

The naive reference is homogeneous entropy-production optimization with no regulation step, which is Cannon's step 1 alone. It sets the floor for concentration error and calls nothing regulated.

The strongest existing competitor is the full Cannon pipeline. It is the only method in this file that infers post-translational regulation without rate constants, and it is the one whose reliability is in question. It runs on identical bulk profiles, solver tolerances and loss function. The mixture model gets no extra data, only extra latent structure. No reported result on Lacuna gives precision or recall for its regulation calls. The paper reports only that concentrations returned to the "~10⁻³ M range" and that PFK and PDHm regulation was recovered. Our audit therefore produces the first quantitative baseline numbers.

We also run a heterogeneity-only ablation: the mixture model with regulation disabled. It isolates how much of the discrepancy averaging alone explains.

A generic statistical alternative completes the set. It fits a Gaussian mixture in log-concentration space with the same $K$, then attributes regulation by flux-control analysis on each component. This separates the contribution of the thermodynamic constraint from that of simply having mixture components.

The proposal beats the baselines if it meets three conditions:

  • it raises precision of regulation calls over the Cannon pipeline at matched recall, once log-variance exceeds the noise floor;
  • it beats the statistical mixture on that same measure;
  • it does not lose recall on homogeneous populations, where it should collapse to $K=1$.

Why not just…?

Why not just widen the tolerance in Cannon's loss so noise-sized discrepancies are ignored? A wider tolerance treats the bias as symmetric noise, but $\Delta_j$ is always positive and concentrated at specific metabolites. Widening tolerance until the false calls disappear should also erase real regulation, because true PFK control produces deviations of the same sign. We test this directly by sweeping the tolerance and plotting the precision-recall curve against the mixture model's. If a single tolerance value matches the mixture model across heterogeneity levels, the mixture machinery is unnecessary, and we would say so.

Why not just use the Fernandes maximum-entropy flux distribution, which already models heterogeneity? That method describes the distribution of fluxes across cells under flux-balance constraints. It contains no concentrations or rate constants, so it cannot express or test post-translational regulation at all. It is a component of our generator, not a competitor for regulation inference. Its mean-field assumption also means it says nothing about local microenvironments, which we inherit as a limitation.

Why not just fit a statistical mixture without any thermodynamic principle? This is the most serious threat, and it is why the Gaussian-mixture baseline is included. With bulk data alone, the number of components and their positions are poorly identified. The entropy-production constraint makes them identifiable because every component must be a physically admissible steady state of the same network. If the unconstrained mixture matches our precision, the thermodynamic ingredient adds nothing and the paper becomes a pure audit of the homogeneous method. That is still a worthwhile result.

Proposed Experiments and Decisive Checks

On the synthetic suite, the primary metrics are three: precision and recall of regulation calls, error in recovered subpopulation weights, and log-concentration error on held-out metabolites. Each is plotted against controlled log-variance, with 50 random regulation sets per network and heterogeneity level. That sample is large enough to give each method an actual false-call rate, which is the denominator a single case lacks.

The first decisive check concerns the localization prediction. If spurious calls under the homogeneous pipeline do not concentrate downstream of high-$\Delta_j$ metabolites, the Jensen mechanism is wrong and the proposal fails at its foundation.

The second decisive check is a reanalysis of the real N. crassa data. PFK regulation is biologically well established, so it is a sanity check. If the mixture model explains it away as averaging, the model is over-flexible and fails. If PFK survives but PDHm does not, that is a specific, testable claim about mitochondrial heterogeneity worth pursuing experimentally.

The third check repeats the audit on the CHO-K1 core under the Fernandes heterogeneity levels. This tests whether the failure rate generalizes beyond one organism, which is the single-case limitation applied to the method itself.