When water treatment destroys a pollutant but leaves most of its carbon behind as intermediate chemicals, safety claims rest on two weak instruments: a carbon meter (Total Organic Carbon, or TOC) and a jar of water fleas. The canonical example in the literature is tetrabromobisphenol A (TBBPA): ozonation removes the parent molecule from a 100 mg/L solution within 6 minutes at pH 8.0, acute toxicity to Daphnia magna falls from nearly 100% to 20% in that same window, and yet only 22% of organic carbon is mineralized after 2 hours, with 17 intermediates identified by mass spectrometry and bromate — a regulated carcinogen — formed along the way (Rapid Removal of Tetrabromobisphenol A by Ozonation in Water). Whether the remaining 78% of carbon does anything to tissue in a whole animal over weeks is simply unknown; the study itself flags chronic toxicity of unmineralized intermediates as unassessed. Meanwhile, the tools that could answer this — quantitative histology such as villus-height-to-crypt-depth morphometry and H&E necrosis scoring — exist in adjacent literatures (Effects of Acute Salinity Stress on the Histological and Bacterial Community Structure and Function in Intestine of Stichopus monotuberculatus; Tetrandrine attenuates lipopolysaccharide-induced fulminant hepatic failure in d-galactosamine-sensitized mice) but have never been pointed at transformation-product mixtures.
There is a second, methodological reason this has not been done: the moment tissue slides come from more than one laboratory, staining and protocol differences swamp the biological signal. On Camelyon17-WILDS, a standard model's accuracy drops from roughly 93% on in-distribution hospital slides to roughly 70% on an unseen hospital, and popular domain-invariance algorithms (IRM, CORAL) fail to beat plain pooled training in 8 of 10 real-world benchmark datasets (WILDS: A Benchmark of in-the-Wild Distribution Shifts). The one method that works well there — targeted augmentation that randomizes hospital-specific stain color while preserving cell geometry — lifted out-of-distribution accuracy from 75.3% to 90.5% on Camelyon17-WILDS (Out-of-Distribution Robustness via Targeted Augmentations) — but it was designed and validated only for binary tumor detection, with a single "domain" axis (which hospital stained the slide).
Core Claim
I propose to build and analyze the first benchmark in which the biological endpoint is tissue-level damage from an incompletely mineralized effluent, and the heterogeneity problem has two labeled axes: a spurious axis (which laboratory ran the exposure and stained the slide — this should be randomized away) and a predictive axis (which stage of the degradation cascade the organism was exposed to — this must be preserved, because it is exactly the signal we want to detect). The core methodological claim is that the targeted-augmentation recipe of Out-of-Distribution Robustness via Targeted Augmentations must be split along these two axes to work here: randomizing all domain-dependent variation, as single-axis stain jitter effectively does, risks discarding exposure-dependent remodeling — the very thing being measured. This is the concrete instantiation of the tension that paper itself identified, that domain-dependent features can carry valid predictive signal and total invariance destroys it.
Formally, we adopt their feature partition and refine it with two domain axes:
$$x \;=\; \bigl(x_{\text{obj}},\; x_{\text{noise}},\; x^{\text{lab}}_{d:\text{spu}},\; x^{\text{stage}}_{d:\text{pred}}\bigr)$$
where $x_{\text{obj}}$ is objective tissue morphology, $x_{\text{noise}}$ is domain-independent noise, $x^{\text{lab}}_{d:\text{spu}}$ encodes lab-specific staining and protocol artifacts, and $x^{\text{stage}}_{d:\text{pred}}$ encodes mineralization-stage-dependent remodeling. The augmentation policy randomizes the lab term (stain-color and thickness jitter) while conditioning the model on, and preserving, the stage term. The benchmark then asks a question no existing dataset can: does a model trained this way detect effluent-stage-dependent tissue damage on a laboratory it has never seen, and on a degradation stage it has never seen?
Research Program
The program has three coupled parts: a wet-lab data generation effort, a benchmark construction with explicit domain labels, and a two-axis robustness method evaluated against strong baselines.
For the chemistry, we replicate the TBBPA ozonation system of Rapid Removal of Tetrabromobisphenol A by Ozonation in Water, which gives us the best-characterized degradation gradient in the available literature: complete parent removal by 6 minutes, 22% TOC removal at 2 hours, a known 17-intermediate cascade, and a known hazardous byproduct (bromate). Effluent is sampled at four mineralization stages — parent-dominated (under 6 minutes), early-intermediate (roughly 15–30 minutes), intermediate-rich (2 hours), and a fully mineralized reference achieved by extended treatment — and each stage is chemically fingerprinted by LC-MS and TOC so the exposure matrix is known, not assumed. A second chemistry, lincomycin degradation by an optimized TiO2/HAP/Ge composite, where the literature already documents a stark gap between high target-molecule removal and low TOC mineralization (A Study on Characteristics of Lincomycin Degradation by Optimized TiO2/HAP/Ge Composite using Mixture Analysis), is run in parallel to test generality across pollutant classes. Both were chosen because they exhibit the field's recurring pathology that Kinetics and Mechanism Study of Oxidation of Ethylenediamine and Ethanolamine by Potassium Ferrate (VI) in Alkaline Media identifies generally: kinetic analyses characterize the first attack steps and leave the stalled downstream fragments structurally undefined.
For the biology, a sub-chronic exposure (a fish model, gut and liver, roughly three to four weeks) is run on the stage-sampled effluents, deliberately mirroring the design logic of the sea-cucumber salinity study, which showed that an abiotic stressor produces measurable villus shortening, crypt loss, and epithelial necrosis with 48 hours as a critical remodeling threshold (Effects of Acute Salinity Stress on the Histological and Bacterial Community Structure and Function in Intestine of Stichopus monotuberculatus), and the hepatic H&E workflow of the tetrandrine study, which couples structural scoring to biochemical markers like serum ALT/AST and caspase-3 (Tetrandrine attenuates lipopolysaccharide-induced fulminant hepatic failure in d-galactosamine-sensitized mice). The critical design decision is that three independent laboratories each run their own exposure and their own histology core, so that staining, section thickness, and protocol heterogeneity arise naturally rather than being simulated. Endpoints are continuous morphometric quantities (villus height, crypt depth, their ratio, necrotic-area fraction, collagen-area fraction via Masson's trichrome), not ordinal grades, because the histopathology literature itself identifies ordinal "mild/moderate/severe" scoring as a major source of reader subjectivity and irreproducibility.
For the machine learning, the benchmark — call it MinPath — consists of patch-level H&E images with two labels per image: effluent stage (four levels, the predictive axis) and source laboratory (three levels, the spurious axis). Two tasks are defined: lesion versus non-lesion patch classification (the Camelyon17-analog task, allowing direct comparison to documented numbers) and regression of the villus-to-crypt ratio from image patches. The method is the two-axis targeted augmentation defined above: a stain-and-protocol jitter operating in H&E color space plus geometric/thickness perturbations for the lab axis, with the stage label always preserved and conditioned on. Splits test three generalization regimes: leave-one-laboratory-out, leave-one-stage-out, and the combined shift.
Proposed Experiments
Experiment 1 is the biological headline test of Tension 1 in the literature: does intermediate-rich effluent (the 2-hour, 78%-carbon-remaining stage) produce measurable tissue remodeling relative to both the parent-dominated and fully mineralized stages, in organisms where the Daphnia acute assay reports the effluent as largely detoxified? The endpoint is the effect size and confidence interval on the villus-to-crypt ratio and necrotic fraction, plus their correlation with the LC-MS intermediate fingerprint. If the effect sizes are null across all stages, the biological premise fails and the benchmark becomes a negative result about acute-bioassay adequacy — still publishable, but the ML program loses its motivating signal; we state this honestly as a real risk.
Experiment 2 is the domain-generalization evaluation on MinPath. All methods share the same pretrained backbone, the same training data, and the same tuning budget (hyperparameters selected only on held-in domains, never on the test domain, following the validation protocol that WILDS: A Benchmark of in-the-Wild Distribution Shifts shows is essential). Metrics are out-of-distribution accuracy for lesion classification and Spearman correlation between predicted and measured villus-to-crypt ratio for morphometry, reported per held-out domain plus worst-domain. No prior work reports a morphometry-correlation number on heterogeneous toxicopathology slides; no reported result found on Lacuna for that quantity, which is itself part of the case for building the benchmark.
Experiment 3 is the ablation that isolates the proposal's key ingredient: single-axis stain jitter (randomizing everything domain-dependent, as in Out-of-Distribution Robustness via Targeted Augmentations) versus the two-axis policy that preserves the stage signal. The decisive comparison is generalization to an unseen degradation stage, where the single-axis version has no mechanism to protect exposure-dependent remodeling from being jittered away. A complementary reader study compares ordinal cross-lab pathologist scoring against the continuous morphometry on the same slides, testing whether the subjectivity concern documented in the histopathology literature materializes at measurable cost.
Baselines
The do-nothing reference point is current standard practice in the mineralization literature: judge effluent safety by TOC reduction plus an acute invertebrate bioassay, exactly as in Rapid Removal of Tetrabromobisphenol A by Ozonation in Water, where those instruments declared a 22%-mineralized, bromate-containing effluent largely detoxified. This baseline is run on every stage sample in Experiment 1. Beating it means demonstrating that a stage the acute assay passes (low immobilization, moderate TOC residual) nonetheless produces statistically clear tissue remodeling — or, if no such dissociation exists, showing that bioassay and morphometry track each other, which would be the first direct evidence that the acute assay is sufficient.
The strongest existing machine-learning competitor is targeted augmentation via Stain Color Jitter in H&E space, the method that holds the best documented out-of-distribution number on heterogeneous histopathology: 90.5% OOD accuracy on Camelyon17-WILDS, up from a 75.3% baseline, a 15.2-point gain (Out-of-Distribution Robustness via Targeted Augmentations). We run it verbatim on MinPath with the same backbone, data, and tuning budget as our two-axis variant. It is the strongest competitor because it is simple, domain-informed, and empirically dominant where it was validated; beating it requires both higher leave-one-lab-out accuracy than its single-axis policy achieves and, more importantly, materially better leave-one-stage-out morphometry correlation, a regime it was never designed for. Pooled Empirical Risk Minimization is included as the reference learner, with the expectation, grounded in the documented 93%–to–70% Camelyon17 drop under hospital shift (WILDS: A Benchmark of in-the-Wild Distribution Shifts), that it will degrade sharply on unseen labs. IRM and CORAL are included as documented negative controls, since the same benchmark showed such invariance methods fail to beat ERM in 8 of 10 datasets. Finally, the single-axis-jitter ablation from Experiment 3 serves as the baseline that isolates the two-axis decomposition as the proposal's specific contribution: any gain of the full method over this ablation on the stage-shift split is attributable to the new ingredient and nothing else.
Why not just…?
Why not just run the exposure study and have pathologists score the slides? If a single lab with a single reader could answer the biological question, that would be the cheaper and better design, and we treat this as a genuine threat. Two pieces of evidence argue against it: ordinal grading is documented as reader-subjective in the histopathology literature, and the study-heterogeneity literature shows that incompatible measurement definitions across sites routinely destroy comparability — as when sarcopenic obesity studies using BMI versus skeletal-muscle-index definitions skewed pooled mortality estimates. Our Experiment 3 reader study settles this empirically: if cross-lab ordinal agreement is high and effect sizes are unambiguous, the simple design suffices and we will say so.
Why not just rely on Daphnia and TOC? The TBBPA study itself is the counterevidence: acute toxicity fell to 20% while 78% of organic carbon remained and a carcinogenic byproduct formed, with chronic effects explicitly unassessed (Rapid Removal of Tetrabromobisphenol A by Ozonation in Water). The decisive check is the correlation between acute-bioassay response and morphometric endpoints across the four stages: if they are concordant everywhere, the bioassay is vindicated; if they dissociate at the intermediate-rich stage, the bioassay is demonstrably insufficient.
Why not just fine-tune a large pretrained pathology model on pooled data? Plain pooled training is the documented failure mode under exactly this kind of shift: roughly 93% in-distribution collapsing to roughly 70% on an unseen hospital's H&E slides on Camelyon17 (WILDS: A Benchmark of in-the-Wild Distribution Shifts), and the same paper showed the fashionable invariance algorithms do not reliably rescue it. We include ERM as the reference learner precisely so the magnitude of its MinPath drop is measured, not assumed.
Why not just apply Stain Color Jitter as-is? This is the most serious alternative, and we do not argue around it — we run it as the strongest baseline. The specific reason it should fall short is that its single-axis randomization has no way to distinguish lab-protocol color variation from exposure-dependent biological coloration change, the exact invariance-versus-preservation tension its own authors documented; randomizing all domain-dependent variation can erase the remodeling signal that defines our task. The leave-one-stage-out split is built to adjudicate this cleanly.
Why not just meta-analyze existing studies instead of generating data? There is nothing to pool: no study in the mineralization literature measures tissue histopathology of transformation-product exposure, which is precisely Gap 1. The same literature shows what happens when heterogeneous cohorts are pooled without standardization — quantitative meta-analysis collapses into narrative review. Benchmark creation is the prerequisite, not a substitute, for synthesis.
Decisive Checks
Three findings would make or break the proposal. First, a significant villus-to-crypt or necrosis effect at the intermediate-rich stage that the acute bioassay passes, which would resolve the field's open detoxification-versus-intermediate-hazard tension with tissue-level evidence. Second, the two-axis method beating single-axis stain jitter specifically on leave-one-stage-out morphometry while matching it on leave-one-lab-out accuracy, which would establish that the two-axis decomposition, not stain jitter alone, is what robust toxicopathology needs. Third, degraded ERM performance on unseen labs reproducing the documented Camelyon17 pattern, confirming that MinPath captures a real, transferable heterogeneity problem rather than a constructed one. A null biological result in Experiment 1 is the main failure mode, and it is reported as such if it occurs.
References
- Rapid Removal of Tetrabromobisphenol A by Ozonation in Water: Oxidation Products, Reaction Pathways and Toxicity Assessment
- Effects of Acute Salinity Stress on the Histological and Bacterial Community Structure and Function in Intestine of Stichopus monotuberculatus
- Tetrandrine attenuates lipopolysaccharide-induced fulminant hepatic failure in d-galactosamine-sensitized mice
- WILDS: A Benchmark of in-the-Wild Distribution Shifts
- Out-of-Distribution Robustness via Targeted Augmentations
- A Study on Characteristics of Lincomycin Degradation by Optimized TiO2/HAP/Ge Composite using Mixture Analysis
- Kinetics and Mechanism Study of Oxidation of Ethylenediamine and Ethanolamine by Potassium Ferrate (VI) in Alkaline Media