When biologists measure how living cells generate energy, they lower a sensor probe to form a tiny sealed chamber — under 10 microliters — over a sheet of cells growing at the bottom of a well, and watch how fast the cells consume oxygen and release acid. Those two rates track mitochondrial respiration and glycolysis, the twin engines of cell metabolism (Yan et al.; Bie et al.). The catch is that the rates are totals: a well with twice as many cells reads twice as high. Every reported number is therefore a division — rate divided by cell count — and that count usually comes from a person tracing cells in a microscope image. People disagree about where a cell ends, especially when cells are swollen, dying, or piled on top of each other. We propose to measure, for the first time, how this human disagreement propagates into the final metabolic numbers, and to replace the single divisive count with a probability distribution over counts carried all the way through to the flux estimate.
The regime where this matters most is toxicology under air pollution. Atmospheric chemistry has established that severe urban haze is not accumulated soot but a chemically transformed secondary aerosol — sulfate, nitrate, ammonium, and oxygenated organics formed from gases, swollen by humidity — making up up to 75% of aerosol mass in intense episodes (Z. B. Wang et al.; Xuehui He et al.). Such aerosol is acidic and reactive (Cynthia Wong et al.), and it is precisely the exposure that kills and deforms cells in a monolayer — exactly the conditions under which segmentation becomes ambiguous and cell-counting disagreement explodes. Meanwhile, the one study that has watched immune cells clearing inhaled particles in intact lungs showed that alveolar macrophages actively crawl and internalize 83–93% of deposited nanoparticles within 24 hours, and that this clearance halts when cellular metabolism is shut down (Multimodal imaging and deep learning...) — yet nobody has measured the metabolic flux that powers that clearance, let alone under chemically realistic haze.
Core Claim
Inter-observer segmentation variability is a first-order, currently unquantified error source in extracellular flux analysis, and it is not constant: it grows exactly in the exposure regimes toxicologists care about, because chemically transformed secondary aerosol produces the ambiguous, partially dead monolayers on which human raters disagree most. A probabilistic normalization that models individual raters — extending the latent rater distribution of Arne Schmidt et al. (Pionono)) — and propagates the resulting uncertainty over cell counts into the flux parameters converts this noise into calibrated error bars, and in doing so unmasks bioenergetic effects of secondary aerosol that crude normalization hides.
Research Program
Three components. First, an exposure platform: a bench-scale photochemical aging and aerosolization setup that converts a gas mixture (SO₂, NOₓ, VOCs) into secondary aerosol under controlled humidity, monitored with the molar conversion ratios used in field campaigns, the Sulfur and Nitrogen Oxidation Ratios (Mei Li), so that the delivered aerosol is genuinely secondary rather than resuspended dust. Primary soot and clean air serve as controls. The cell model is alveolar-macrophage-like, motivated by the finding that macrophage migration is the active engine of lung particle clearance (Multimodal imaging and deep learning...). Cells are exposed by aerosol deposition onto the monolayer rather than pipetted suspension, because liquid instillation produces patchy, centrally concentrated deposits while aerosol delivery deposits uniformly, with ventilator-assisted delivery measured as roughly 20 times more efficient than nose-only inhalation in the murine study (Multimodal imaging and deep learning...).
Second, a multi-rater microchamber image benchmark. Before and after exposure, each well is imaged and a mitochondrial stress test is run — sequential oligomycin, FCCP, and rotenone/antimycin injections that decompose raw oxygen consumption into ATP-linked respiration, proton leak, and spare respiratory capacity (Strickertsson et al.; Li et al.). Six trained raters independently segment viable-cell regions per well, with a repeated session per rater to capture intra-observer drift. This is the first dataset, to our knowledge, pairing multi-rater segmentation of live-cell microchamber images with matched kinetic flux readouts.
Third, the method. Where $\text{OCR}_k$ is the oxygen-consumption rate in well $k$ and $N_k$ the viable-cell count, every reported flux parameter today is effectively $\theta_k = \text{OCR}_k / \hat{N}_k$ with $\hat{N}_k$ a single count from one rater or one automated model. We replace this with
$$p(\theta_k) = \int \delta\!\left(\theta_k - \frac{\text{OCR}_k}{N}\right)\, p(N \mid r_1, \dots, r_R)\, dN,$$
where $p(N \mid r_1, \dots, r_R)$ is induced by a Pionono-style latent rater model fitted to the $R$ raters' masks: the latent mean captures each rater's systematic tendency to include or exclude ambiguous cells, and the variance captures their internal inconsistency (Arne Schmidt et al.). The output is a posterior over spare respiratory capacity and proton leak per viable cell, not a point estimate. Conceptually this imports the lesson of Bayesian Metabolic Flux Analysis, which quantified uncertainty in intracellular fluxes from isotope tracer data, but our uncertainty source is human disagreement over images — a channel that work never touches.
Baselines
The naive reference point is doing nothing: reporting raw, unnormalized oxygen-consumption and acidification rates. This is what the microchamber actually measures, and it anchors how much normalization is buying at all; no reported result found on Lacuna for the reproducibility of raw unnormalized rates across replicate wells.
The incumbent biochemical baseline is destructive endpoint normalization: after the flux run, lyse the cells and divide all rates by total protein from a bicinchoninic acid assay or total ATP from luciferase bioluminescence. This is the strongest competitor because it needs no images, no raters, and no model, and it is the field's default; no reported result found on Lacuna for BCA-normalized flux reproducibility, so we will establish that number ourselves on the same wells, same exposure conditions, and same number of replicate wells — an exactly matched comparison.
The strongest segmentation-based baseline is a deterministic consensus pipeline: a single automated segmenter of the nnU-Net family, whose 3D variant reached Dice scores of 0.89–0.92 on light-sheet microscopy of cleared murine lungs (Multimodal imaging and deep learning...), trained on a majority-vote or STAPLE consensus of our raters' masks, then used to produce one count per well. Consensus fusion is genuinely strong — hierarchical majority voting of the top ten models rivaled human expert overlap on the BRATS brain-tumor benchmark (BRATS benchmark) — which is exactly why it is the right baseline rather than a straw man. We will train it on the same images with the same compute and tuning budget as our model, with rater disagreement measured by unweighted Cohen's kappa, the metric on which the latent rater model of Arne Schmidt et al.) reached 0.758 on the six-pathologist Gleason 2019 dataset and 0.711 on a 25-rater breast-cancer dataset — the strongest published multi-rater segmentation numbers in our file.
Finally, the key ablation isolates uncertainty propagation itself: the same latent rater model, but collapsed to its posterior mean as a point count before division. If full propagation beats point-estimate normalization on interval calibration, the gain comes from carrying disagreement through the division, not merely from better masks.
Beating the baselines means three concrete things: higher between-replicate reproducibility (intraclass correlation) of normalized spare respiratory capacity in exposed wells than under BCA or consensus normalization; 90% posterior intervals that cover held-out replicate means at least 90% of the time while being narrower than intervals obtained by naively bootstrapping over raters; and recovery of a secondary-versus-primary aerosol metabolic effect that at least one baseline normalization fails to detect.
Why not just…?
Why not just lyse the cells and measure protein? This is the real threat, and we treat it as a head-to-head baseline rather than arguing around it. The known weaknesses are that endpoint lysis destroys the sample and any longitudinal readout (Extracellular Flux Analysis summary; The Impact of Glyphosate, Its Metabolites and Impurities on Viability, ATP Level and Morphological changes in Human Peripheral Blood Mononuclear Cells), and that total protein conflates viable and nonviable cells: an exposed well that has lost half its cells to death still carries most of its protein, so per-protein normalization systematically underestimates the per-viable-cell metabolic injury — the very quantity aerosol toxicology wants. But if BCA matches our intraclass correlation in exposed wells, the honest conclusion is that image-based normalization matters only for longitudinal or viability-weighted designs, and our experiments are constructed to settle exactly that.
Why not just use a consensus mask? Because the consensus answers a different question. Majority voting and STAPLE collapse raters into one smoothed boundary; Arne Schmidt et al.) showed this averages away genuine ambiguity and still loses to explicit rater modeling on Gleason 2019 (0.758 versus lower kappa for STAPLE-style baselines). More fundamentally, a consensus mask yields a point count and therefore no error bar on the flux parameter at all. Our coverage test makes this failure measurable: if consensus-normalized rates cannot produce calibrated intervals, no amount of segmentation accuracy fixes the missing uncertainty.
Why not just expose cells to filter-collected PM2.5? Because filter extraction discards the chemistry that defines the exposure. In the Yinchuan decadal campaign, secondary organic aerosol made up roughly 74% of organic mass (Wei Zhou et al.), and heterogeneous uptake of small dicarbonyl gases alone contributes about 21% of secondary organic aerosol in polluted subtropical environments (Zhenhao Ling et al.) — precisely the volatile and semivolatile fraction lost on filters. Filters also lose the humidity-driven hygroscopic swelling and aqueous oligomer formation that shape what actually reaches cells (Xuehui He et al.; Yonghong Wang et al.). The murine delivery study makes the physical point directly: liquid instillation produced patchy aggregates while aerosol deposition was uniform (Multimodal imaging and deep learning...). A filter-suspension arm is cheap, so we will include it as a secondary control and let the data show how much of the metabolic signature it misses.
Proposed Experiments and Decisive Checks
The first experiment tests the premise: inter-rater agreement, measured by Cohen's kappa, should fall as secondary-aerosol exposure dose rises, dropping from near-ceiling in clean confluent wells toward the 0.71–0.76 range reported for genuinely ambiguous histopathology (Arne Schmidt et al.) and toward the 74–85% Dice overlap that human experts achieve on ambiguous brain-tumor boundaries (BRATS benchmark). A dose-response curve of disagreement is itself a result: it quantifies the ground-truth fallacy — that a single annotator's mask is one opinion, not truth — in a domain where nobody has measured it. If agreement instead stays uniformly high even in dying wells, the core claim is weakened and we would report that.
The second experiment is a variance decomposition: across replicate exposed wells, what fraction of the variance in normalized spare respiratory capacity is attributable to rater identity versus biological replicate? We predict rater identity contributes a substantial share in exposed wells and a negligible share in clean wells, which would explain a chunk of the irreproducibility that plagues the field's known normalization problem (Anita Wagner et al.).
The third experiment is the biological payoff and the calibration check together. We predict that aged secondary aerosol suppresses macrophage spare respiratory capacity — the reserve that migration presumably draws on, given that metabolic shutdown halts macrophage crawling (Multimodal imaging and deep learning...) — relative to primary soot at matched delivered mass, and that this effect is visible only under observer-aware or viability-weighted normalization, because protein and consensus normalizations absorb it into the cell-death signal. Prior inhalation work shows environmental stress can rewire metabolism, but only for pure oxygen deprivation without any aerosol chemistry (Emilie Montellier et al.); showing a chemically specific, aerosol-composition-dependent bioenergetic signature would be new. The decisive readouts are the intraclass correlations, interval coverage rates, and effect recovery rates defined against the baselines above; the whole proposal stands or falls on whether calibrated uncertainty over the cell count, not just a better count, is what makes the difference.
References
- Eugenol inhibits oxidative phosphorylation and fatty acid oxidation via downregulation of c-Myc/PGC-1β/ERRα signaling pathway in MCF10A-ras cells
- β-Boswellic Acid Suppresses Breast Precancerous Lesions via GLUT1 Targeting-Mediated Glycolysis Inhibition and AMPK Pathway Activation
- New particle formation in the presence of a strong biomass burning episode at a downwind rural site in PRD, China
- The significant contribution of biomass burning to methanol-soluble nitrogenous organics and its evolution during the highly-humid haze event in urban Wuhan
- Highly Acidic Conditions Drastically Alter the Chemical Composition and Absorption Coefficient of α-Pinene Secondary Organic Aerosol
- Multimodal imaging and deep learning unveil pulmonary delivery profiles and acinar migration of tissue-resident macrophages in the lung
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- Size-segregated Characteristics and Formation Mechanisms of Water-soluble Inorganic Ions during Different Seasons in Heshan of Guangdong, China
- Enterococcus faecalis Infection Causes Inflammation, Intracellular Oxphos-Independent ROS Production, and DNA Damage in Human Gastric Cancer Cells
- Role of mitochondrial fusion proteins MFN2 and OPA1 on lung cellular senescence in chronic obstructive pulmonary disease
- Bayesian metabolic flux analysis reveals intracellular flux couplings
- The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)
- The Impact of Glyphosate, Its Metabolites and Impurities on Viability, ATP Level and Morphological changes in Human Peripheral Blood Mononuclear Cells
- Decadal transition of summertime PM 2.5 –O 3 coupling and secondary organic aerosol dominance in northwest China
- Formation and sink of glyoxal and methylglyoxal in a polluted subtropical environment: observation-based photochemical analysis and impact evaluation
- Increased inorganic aerosol fraction contributes to air pollution and haze in China
- In vitro model of human subcutaneous adipocyte spheroids for studying mitochondrial dysfunction and mitochondria activating compounds
- Chronic intermittent hypoxia reshapes circadian metabolic architecture in a model of sleep apnea