Extracellular flux analysis (EFA) is a live-cell assay that measures, in real time, how fast cells consume oxygen and release acid into their surroundings — a direct readout of the two energy engines of mammalian cells, mitochondrial respiration and glycolysis. To run it on patient tissue rather than immortalized cell lines, cells must first survive an obstacle course: surgical resection, mincing, enzymatic digestion, centrifugation, attachment to the bottom of a measurement well, and a sequence of drug injections. Every published study reports flux values measured only on the survivors. Yet in exactly the diseases where EFA matters most, the frailest cells are the most informative ones: cigarette-smoke-damaged lung cells in COPD, energetically stressed placental cells in pre-eclampsia. If sick cells preferentially die during isolation, the measured difference between patients and controls is shrunken or distorted — the laboratory twin of what retrospective clinical studies suffer when only patients healthy enough for surgery, and compliant enough to return for follow-up, end up in the dataset.
The clinical literature has built serious machinery for this problem — collider bias, shadow variables, and learned corrections such as ShadowCatcher and Two-Stage Shadow Inclusion (2SSI) — but all of it strains under a handicap: in the clinic, the selection process itself is invisible. In the lab it is not. You can count cells at every stage of the funnel, from the suspension that goes into the plate to the cells that attach and remain measurable after the stress injections. This proposal's core move is to treat those counts as data about selection, rather than as waste, and to correct measured flux distributions back to the tissue population they came from.
Core Claim
Cell-level selection in primary-tissue EFA is substantial, quantifiable, and invertible. Concretely, I claim three things. First, group contrasts in oxygen consumption rate (OCR) and extracellular acidification rate (ECAR) reported from primary tissue are contrasts among a fitness-selected subpopulation, and in diseases that impair mitochondrial or glycolytic function this selection attenuates the true tissue-level effect. Second, because per-stage cell counts are directly observed, the selection mechanism can be modeled and the attenuation corrected under assumptions far weaker than those required when selection is unobserved — the counts play the role that hand-crafted or learned shadow variables must play in electronic health record (EHR) settings. Third, the correction is empirically checkable in a way clinical bias corrections never are: a paired reference measurement of the entire cell suspension in a stirred closed chamber, which requires no attachment step, exposes the unselected flux value for the same sample, so the bias can be observed per patient rather than merely argued about.
The estimand is the tissue-level contrast
$$\Delta = \mathbb{E}[y \mid D{=}1] - \mathbb{E}[y \mid D{=}0],$$
where $y$ is a flux quantity (basal OCR, spare respiratory capacity, ECAR), $D$ is disease status, and the expectation runs over all viable cells in the tissue. Standard practice estimates this contrast over $S{=}1$, the cells that survived isolation, adherence, and the stress protocol.
Research Program
The first component is a hierarchical selection model. Each cell carries a latent fitness $\theta$ that influences both its chance of surviving each stage and its flux values. Survival is a product of stage hazards, $\Pr(S{=}1 \mid \theta) = \prod_k p_k(\theta)$, and the observed stage counts arise by binomial thinning of the previous stage, which makes the funnel itself a likelihood object: the counts tell us about the distribution of $\theta$, and the flux measurements on survivors tell us about $y$ given $\theta$. Fitting both jointly recovers an estimate of the $\theta$-distribution in the unselected population and hence a reweighted or model-based estimate of $\Delta$. Identification is not free — with only aggregate counts one needs a parametric form for the stages and a coupling $\lambda$ between $\theta$ and $y$ — so the program includes a tipping-point sensitivity analysis: for each published finding, we report how strong the fitness–flux coupling would have to be before the reported group difference is fully explained by selection. The model extends one level up to patient-level selection (who supplies tissue at all) using consecutive-screening logs, which is the same funnel logic applied to the surgical-availability filter that, for example, the COPD cohort of Li et al. never adjusted for.
This is deliberately different from the two closest methods. Learning Shadow Variable Representation for Treatment Effect Estimation under Collider Bias learns a latent shadow representation from covariates because selection is unobserved, and it was validated only on synthetic and standard causal datasets (IHDP, ACIC 2016); we observe the selection process directly, so we need neither the refutable-independence search nor the Fredholm-equation machinery, though we adopt its odds-ratio framing of the selected-versus-unselected gap as a comparison point. Two-Stage Shadow Inclusion Estimation: An IV Approach for Causal Inference under Latent Confounding and Collider Bias constructs a shadow variable from instrumental-variable residuals; we need no instrument at the cell level because the counts are data. Both papers are unapplied to any biological measurement — this proposal is the first biological test of whether collider corrections matter where the biology plausibly induces them.
The motivating evidence comes from the EFA side. Role of mitochondrial fusion proteins MFN2 and OPA1 on lung cellular senescence in chronic obstructive pulmonary disease measured respiration in primary alveolar cells from 40 male subjects across three clinical strata, selected by tissue availability, with no accounting of which cells survived isolation. <scp>AMPK</scp> regulates homeostasis of invasion and viability in trophoblasts by redirecting glucose metabolism: Implications for pre‐eclampsia (AMPK regulates homeostasis of invasion and viability in trophoblasts by redirecting glucose metabolism: Implications for pre‐eclampsia) profiled just 15 pre-eclamptic and 15 control placentas, in a disease that itself stresses trophoblast energetics — precisely the setting where the cells most affected by the pathology are the ones lost before measurement. And the clinical-selection literature tells us what happens if we ignore this: Randomized Clinical Trials or Convenient Controls: TREWS or FALSE? showed that even after regression adjustment, a completely random alert produced an adjusted mortality odds ratio of 2.21 in one cohort and 1.6 in another — selection and process confounding survive standard adjustment.
Baselines
The "do nothing" reference point is standard practice: analyze the adherent surviving cells, as every cited EFA study of primary tissue implicitly does, including the COPD work of Li et al. and the pre-eclampsia trophoblast study. This baseline defines the bias we measure; it is not a strawman, it is what the field currently publishes. The strongest clinical-practice competitor is propensity score matching plus multivariable regression, as executed by Prognostic prediction and treatment options for gastric signet ring cell carcinoma: a SEER database analysis, which matched patients 1:1 on age and stage and built Cox nomograms reaching a C-index of 0.810 in the SEER training set and 0.859 in the external Soochow cohort of 257 patients. I will run PSM on the same simulated and pilot data, with the same covariates, the same folds, and an equal grid-search budget over matching variables, so the comparison isolates the correction rather than tuning effort. Beating it means leaving less residual bias against a known ground truth: in simulations where selection depends on unrecorded cell fitness, PSM should retain most of the attenuation while the funnel model removes it, and in the pilot, PSM-corrected estimates should still deviate from the paired suspension reference where the funnel-corrected estimates do not.
The strongest causal competitor is ShadowCatcher from Learning Shadow Variable Representation for Treatment Effect Estimation under Collider Bias, which reduced estimation error by roughly 30% to 50% relative to the best alternatives on IHDP and ACIC 2016. It is the right competitor because it attacks the same estimand class, and it is sensitive to the rejection threshold $\alpha$, so I will tune $\alpha$ on validation folds with the same budget given to my model. Where an instrument is available (the pilot randomizes isolation harshness, which acts as one), I will also run 2SSI, whose unobserved-data MSE was 5x to 10x lower than DeepIV and 2SRI on the Demand benchmarks. The key ablation deletes the funnel counts from my own model — selection then depends only on recorded covariates, collapsing the method to an inverse-probability-of-selection weighting variant. If the ablated model performs as well as the full one, the counts are not the ingredient and the idea fails; that is the point of running it.
Why not just…?
Why not just enroll patients prospectively? Prospective consecutive enrollment fixes patient-level sampling but not cell-level selection, and it is slow and small: the COPD study gathered 40 male subjects across three strata, and the radiomics study of A novel computed tomography radiomic nomogram for early evaluation of small airway dysfunction development assembled 272 patients over roughly four years, while registries hold thousands. Even a perfect cohort still loses frail cells at the dish. I adopt prospective elements where they are cheap (screening logs in the pilot) because they complement, not replace, the correction.
Why not just use PSM or multivariable regression? The TREWS critique is the decisive evidence: after backdoor adjustment, a coin-flip alert still showed an adjusted mortality odds ratio of 2.21 (Cohort A) and 1.6 (Cohort B), and the study arm mortality difference of 3.3% versus 6.4% ($p < 0.001$) was pure process confounding. Yu et al. themselves concede that PSM cannot see unmeasured confounders in a 30-year registry. In our setting the unmeasured variable is cell fitness, which no recorded covariate captures; the simulation and pilot run PSM head-to-head to settle this empirically rather than by argument.
Why not just use static assays that don't need live cells? This is the most serious alternative: ATP bioluminescence, Western blots, or transcriptomics on frozen tissue avoid the adherence funnel entirely. But they measure a different quantity. In Adenosinergic Signaling Alters Natural Killer Cell Functional Responses, RNA-seq missed the suppression of both respiration and glycolysis that real-time flux analysis caught. In the pre-eclampsia study, static ATP was elevated in diseased placentas while AMPK redirected glucose away from lactate production — a static readout would have pointed the wrong way. And in the COPD work, blots showed protein changes but only live-cell measurement showed that pharmacological agents actually restored respiration. The pilot includes the head-to-head: if static markers predict the unselected suspension reference as well as corrected flux does, I will report that the correction is unnecessary, and that result would itself be useful.
Proposed Experiments and Decisive Checks
The first experiment is a calibrated simulation. Effect sizes come from published magnitudes: eugenol reduced MCF10A-ras viability by roughly 64% and intracellular ATP by 57.7% while sparing untransformed MCF10A (Eugenol inhibits oxidative phosphorylation and fatty acid oxidation via downregulation of c-Myc/PGC-1β/ERRα signaling pathway in MCF10A-ras cells), and the glyphosate PBMC work gives a toxicity ladder from an IC50 above 10 mM down to 7.9 mM for PMIDA with ATP collapse to 2.1% of control (The Impact of Glyphosate, Its Metabolites and Impurities on Viability, ATP Level and Morphological changes in Human Peripheral Blood Mononuclear Cells). I simulate a disease that both lowers flux and raises isolation mortality, then compare naive analysis, PSM, ShadowCatcher, 2SSI, the ablated model, and the full funnel model on relative bias in $\Delta$, RMSE, and 95% interval coverage. Success is coverage near nominal with bias at least halved relative to PSM at matched tuning budget.
The second experiment is a prospective paired pilot in peripheral blood mononuclear cells from roughly 40 to 60 donors, the sample type already used in the glyphosate work. For every donor we log the full funnel (input count, adherent count, measured count), run the standard stress protocol, and in parallel measure whole-suspension respiration in a stirred closed chamber that requires no attachment — the unselected reference. The decisive checks are three. The naive adherent-only estimate should deviate from the suspension reference by an amount the funnel model predicts from the counts alone. A lab analog of the coin-flip experiment should pass: randomize wells to harsh versus gentle isolation, and show that naive analysis manufactures a spurious group difference while the corrected estimate stays null. And the fitness–flux sensitivity curve should show whether plausible selection strengths could flip the sign of a reported group contrast.
The third experiment applies the machinery retrospectively, where no counts exist. For the COPD cohort of Li et al. and the pre-eclampsia cohort of the AMPK trophoblast study, I run the tipping-point analysis: how strong would cell-level selection have to be for the reported group differences to be fully or partially artifacts? No reported result found on Lacuna for baseline OCR values in the COPD cohort, so the calibration uses the published direction of effect (reduced basal respiration and ATP-linked respiration under cigarette smoke extract, restored by Leflunomide and BGP15) plus the simulation parameterization. The output is a quantitative statement the field currently cannot make: whether published bioenergetic contrasts in primary tissue are robust to the survivor funnel, or small enough to be plausibly produced by it.
References
- Learning Shadow Variable Representation for Treatment Effect Estimation under Collider Bias
- Two-Stage Shadow Inclusion Estimation: An IV Approach for Causal Inference under Latent Confounding and Collider Bias
- Role of mitochondrial fusion proteins MFN2 and OPA1 on lung cellular senescence in chronic obstructive pulmonary disease
- <scp>AMPK</scp> regulates homeostasis of invasion and viability in trophoblasts by redirecting glucose metabolism: Implications for pre‐eclampsia
- Randomized Clinical Trials or Convenient Controls: TREWS or FALSE?
- Prognostic prediction and treatment options for gastric signet ring cell carcinoma: a SEER database analysis
- A novel computed tomography radiomic nomogram for early evaluation of small airway dysfunction development
- Adenosinergic Signaling Alters Natural Killer Cell Functional Responses
- Eugenol inhibits oxidative phosphorylation and fatty acid oxidation via downregulation of c-Myc/PGC-1β/ERRα signaling pathway in MCF10A-ras cells
- The Impact of Glyphosate, Its Metabolites and Impurities on Viability, ATP Level and Morphological changes in Human Peripheral Blood Mononuclear Cells