Which Genes Actually Steer Cellular Energy Use? Testing Hub-Gene Rankings Against Live Respiration Under Matched Starch–Lipid Complexes

Biologists routinely name "hub genes", the highly connected nodes in a gene network that are presumed to control a process. The trouble is that a hub claim is almost never checked against a measured cellular function. This note proposes a test that does that check. The function we use is mitochondrial respiration and glycolysis in living gut cells. Both can be read in real time with extracellular flux analysis, a plate-based instrument that seals a tiny volume of medium above a cell layer and measures how fast oxygen disappears and acid appears. We perturb the cells with digests of two starch–lipid complexes that look nearly identical in crystal structure but break down at very different speeds. We profile gene expression and respiration side by side at many time points. Then we rank genes by how strongly a fitted dynamic model says they drive the respiration curve, and we confirm the top genes by knocking them down and measuring respiration again.

The claim is concrete. A gene's importance should be scored by how much perturbing it moves a measured energy-use trajectory over time, not by its connectivity in a static co-expression or protein-interaction graph. If that is right, a ranking anchored to the flux readout and fitted to time-ordered data should pick more genes whose knockdown changes respiration than either standard intramodular connectivity or simple gene-to-flux correlation. If it is wrong, the knockdown panel shows it plainly. Either way the paired expression, flux, and knockdown data become a benchmark with real ground truth. That kind of ground truth is currently missing: the geneRNIB living benchmark found that Pearson correlation and GRNBoost2 beat elaborate multi-omic pipelines and foundation models when network inference is scored causally.

Why this perturbation, and why this readout

The perturbation comes from a study of high-amylose starch complexed with stearic acid versus potassium stearate. That study found both complexes carry V-type crystallinity, meaning amylose helices wrapped around a lipid. Yet the free-acid complex is 16.1% to 29.2% resistant starch, while the salt complex forms a charged dispersion (ζ-potential −22.2 to −32.8 mV) and is only 0.1% to 0.2% resistant. The pair holds the ingredients and the crystal form nearly constant while sharply changing how fast glucose is released. That makes it a cleaner contrast than comparing a resistant starch with an ordinary one.

The physiological motivation comes from the human resistant-starch trial. There, 40 g/day of type 2 resistant starch produced a mean −2.8 kg weight change and raised gut ANGPTL4 through Bifidobacterium adolescentis. That trial tested unmodified type 2 starch only. It measured no cellular respiration and no time course of gene regulation, so it motivates our question but does not answer it. Zhao et al.'s pig preadipocyte time course independently named ANGPTL4 as a hub of co-expression modules across four days of differentiation. In their hands, siRNA knockdown above 90% increased lipid accumulation and raised LPL. They never measured respiration and used static, pooled network analysis.

Together these papers make ANGPTL4 a candidate that two unrelated systems both point to. They give no evidence about whether it moves bioenergetics. We therefore register ANGPTL4 in advance as a named prediction. The method is not built around it.

Extracellular flux analysis is the readout because the mitochondrial stress test isolates quantities that endpoint ATP assays cannot. The test injects oligomycin, then FCCP, then rotenone with antimycin A, which separates ATP-linked respiration, proton leak, and spare respiratory capacity. The method is described in Strickertsson et al. and Li et al., and it gives a functional target with several quantities that can be measured repeatedly on living cells.

Research Program

The experimental system is differentiated Caco-2 intestinal monolayers, with HepG2 hepatocytes as a replication line. Each complex goes through a standardized static digestion. Digesta are collected at the digestion sampling times, heat-inactivated, and ultrafiltered to remove enzymes and bile salts that would damage the cells.

There are five exposure arms. The first two are the stearic-acid complex and the potassium-stearate complex. The third is the starch alone. The fourth is matched stearate without starch. The fifth is a "kinetics replay" medium. It contains no complex, but its glucose and stearate concentrations are stepped to follow the measured release curve of the resistant complex. The replay arm asks whether the complex's structure matters to the cell beyond the nutrient-delivery curve it produces. That is the one scientific question the starch-chemistry literature cannot answer by itself.

For each arm, cells are sampled at 0, 0.5, 1, 2, 4, 8 and 24 hours after exposure begins, in triplicate. Bulk RNA-seq is run on parallel wells. Acute flux traces are recorded on the instrument by injecting digesta through its ports, and a full stress test is run at each sampled time. This gives about 105 transcriptomes, each paired to a vector of flux quantities.

The model extends FLeCS, which describes how expression changes over time as coupled differential equations restricted to a prior regulatory graph. FLeCS predicted 72-hour drug trajectories and matched known biology in 11 of 12 CRISPRi knockdowns. Our change is to add a readout map from gene states to flux:

$$\frac{dx}{dt} = f_\theta\big(x, u(t)\big), \qquad y(t) = g_\phi\big(x(t)\big).$$

Here $x$ is the expression of roughly 800 temporally variable genes. $u(t)$ is the exposure arm and its delivery curve. $f_\theta$ is the network-restricted rate model. $y$ is the vector of oxygen consumption rate, extracellular acidification rate, proton leak and spare capacity, and $g_\phi$ is a sparse linear map from expression to that vector.

The model is fitted jointly on the expression and flux trajectories, with conditions held out for validation. Gene $i$'s flux-anchored score is the time-integrated change in predicted flux when its production is silenced in silico:

$$s_i = \int_0^T \big\lVert y(t; x_i \text{ silenced}) - y(t) \big\rVert \, dt.$$

We also compute the difference in $s_i$ between the two complexes. This flags genes predicted to matter more under fast than under slow carbohydrate delivery.

With only seven time points, the model is underdetermined. We therefore keep the gene set small, use ESPACE-style sparse priors, and report rankings averaged over bootstrap refits rather than a single fit.

Baselines

The naive reference is the ranking a typical study would publish: genes ordered by fold change and adjusted p-value between each complex and the starch-alone arm at 24 hours. Random genes matched for expression level serve as the floor.

The strongest existing method for this exact use is WGCNA, the co-expression module method of Zhao et al.. It is run on the same 105 samples, takes the module whose eigengene correlates best with spare respiratory capacity, and ranks genes by intramodular connectivity. Its power and module-size settings are tuned over the same number of configurations as our model, and selection uses held-out-condition stability. It is the strongest competitor because it is what the field already uses to name hubs from time courses with a phenotype attached. No reported result was found on Lacuna for how often WGCNA hubs change a measured cellular function when knocked down; Zhao et al. validated one gene.

Because geneRNIB found that plain correlation is hard to beat, we treat per-gene Pearson correlation with each flux quantity as a co-strongest baseline. We also run a time-lagged version that scores each gene by its correlation with flux one sampling step later. That lagged version is the cheapest way to capture "comes before". No reported number was found on Lacuna for geneRNIB's metric values; the file records only that correlation and GRNBoost2 ranked highest.

Graph baselines are degree and betweenness in a protein-interaction network built from the union of differentially expressed genes, and ESPACE hub degree. We also include temporal betweenness on time-sliced co-expression graphs, using the prefix-foremost path definition, because Buß et al.'s analysis of temporal betweenness (Algorithmic Aspects of Temporal Betweenness) shows that definition is computable in polynomial time while foremost and fastest variants are #P-hard. All methods see the same samples, gene universe and prior graph.

Three ablations isolate the key ingredients. The first removes the flux readout, leaving plain FLeCS sensitivity on expression. The second trains only on the two complex arms without the replay and single-ingredient controls. The third replaces in-silico silencing with summed learned edge weights.

Rankings are frozen and deposited before any knockdown. "Beating" a baseline has two parts. First, our top 10 must contain significantly more genes whose knockdown shifts spare capacity or acidification rate by more than two standard deviations of non-targeting controls; this is precision at 10, compared by a permutation test across the shared panel. Second, the Spearman correlation between predicted score and measured knockdown effect across the whole panel must be higher.

Why not just…?

Why not just correlate every gene with oxygen consumption and knock down the top hits? This is the real threat, and geneRNIB gives good reason to expect it to be strong. Correlation cannot tell a gene that drives respiration from one that responds to it, because both move together. The dynamic model and the lagged-correlation baseline both use time order, and the replay arm separates effects of the delivery curve from effects of structure. If lagged correlation matches the model on the knockdown panel, we will report that simple correlation is enough. The study's contribution then becomes the ground-truth panel, not the model.

Why not just use WGCNA as Zhao et al. did? It is included as a primary baseline, run with the same tuning budget. It has two built-in limits. Pooling time points discards ordering. Intramodular connectivity measures how central a gene is to the expression pattern, not to the flux phenotype. The knockdown panel tests whether those limits cost accuracy in practice.

Why not just test commercial type 2 resistant starch, or slow digestion with acarbose? Either swap changes digestion kinetics together with structure or chemistry, so effects cannot be attributed to either. The stearic-acid versus potassium-stearate pair keeps V-type order and the stearate ligand while changing resistant starch about a hundredfold. The replay arm then separates kinetics from structure, which acarbose cannot do because it adds a drug acting on the cells' own enzymes.

Why not skip transcriptomics and just run the flux assays? That would answer whether the complexes change cellular energy use, which is worth knowing. It cannot nominate which genes to target, and target nomination is the question the hub literature claims to answer.

Proposed Experiments and Decisive Checks

The knockdown panel is the union of the top 10 genes from every method, plus 10 expression-matched random genes and non-targeting controls, or about 60 to 80 siRNAs. Every siRNA is tested under both complex exposures with a full stress test, and the main results are replicated in HepG2. Knockdown efficiency is confirmed by qPCR, and genes knocked down below 70% are excluded from all methods equally.

Three outcomes would settle the core claim. The first is whether flux-anchored scores beat both WGCNA connectivity and lagged correlation on precision at 10 and on rank correlation. The second is whether removing the flux readout erases that advantage. If not, the anchoring is not the active ingredient. The third is whether the genes predicted to depend on the complex actually show knockdown effects that differ between the two complexes.

A secondary check asks whether the replay arm reproduces the resistant complex's flux and expression trajectories. If it does, structure acts only through glucose delivery. If it does not, the cells respond to the complex itself, for example to its surface charge. The pre-registered ANGPTL4 prediction is scored the same way as every other gene. A null result there is informative, because it would show that being called a hub in two different systems does not by itself predict control of respiration.