Lineage-Asymmetric Ultrastructural Injury: TEM Profiling of Primitive versus Definitive Macrophages under Nanoplastic–Pesticide Co-Exposure

When a developing tissue is hit by a mixture of environmental contaminants — say, nanoplastics carrying adsorbed pesticides — the damage is rarely the simple sum of the parts. Nanoplastics act as "Trojan horse" carriers that shuttle concentrated toxicants into cells, and mixture toxicity can flip between synergy and antagonism depending on concentration, in ways that standard additive-toxicity math cannot predict (Enhanced bioaccumulation and toxicity of Fenpropathrin by polystyrene nano(micro)plastics in the model insect, silkworm (Bombyx mori); Biphasic toxicity of heavy metals and microplastic leachates). Yet almost all of this work is done in immortalized monocultures or in whole animals whose individual cell populations are never separated. In particular, nobody knows whether the immune cells that reside in developing tissues — some descended from short-lived embryonic (yolk sac) progenitors, others from the adult bone-marrow line — take the same kind of subcellular damage from these mixtures, or even the same kind of cell death. The question matters because these two lineages have different metabolic programs and different roles in tissue maturation, and losing one but not the other during a sensitive developmental window could quietly reshape an organ for life.

The reason the question is open is largely a measurement problem. The relevant injuries — ferroptotic mitochondria with dissolved cristae, ruptured outer membranes, particle aggregates in endosomes — live at the nanometer scale, where transmission electron microscopy (TEM), which images with electron beams of sub-nanometer resolution, is the only ground truth. But TEM in toxicology papers today is descriptive: a handful of qualitative snapshots per condition (Molecular mechanism differences between nanoplastics and microplastics in colon toxicity: nanoplastics induce ferroptosis-mediated immunogenic cell death, while microplastics cause cell metabolic reprogramming). Meanwhile, the computer-vision community has built automated organelle segmentation and morphological feature extraction for volume EM that work superbly — but only on healthy reference tissue. This proposal moves those tools into toxicology and asks a single sharp biological question with them: do primitive, embryonic-lineage resident macrophages and definitive, bone-marrow-lineage macrophages show measurably different ultrastructural responses — including different cell-death modes — to synergistic nanoplastic–pesticide mixtures?

Core Claim

The claim is that sublethal, multi-contaminant injury in developing tissue is lineage-asymmetric at the organelle level, and that this asymmetry is detectable and quantifiable only when TEM is turned from an anecdotal illustration into a high-throughput, per-cell measurement platform. Concretely, we predict that 100 nm polystyrene nanoplastics carrying a pyrethroid pesticide will induce ferroptotic mitochondrial pathology (shrunken, electron-dense mitochondria with cristae loss, per Chen et al.) preferentially or at lower doses in primitive macrophages, while definitive-lineage macrophages under the identical exposure shift toward adaptive remodeling or metabolic reprogramming of the kind seen with excluded microplastics in Wang et al.. We further claim that this per-cell ultrastructural signature is a more sensitive detector of mixture synergy than organism-level lethality: a morphology-derived synergy index should register non-additive injury at doses where survival-based indices still look additive.

This differs from the closest existing work in three ways. Wang et al. used TEM to pin down size-dependent cell-death modes, but in a transformed epithelial line with a single contaminant and no immune cells; we test mixtures and lineage-resolved immune cells in an intact developing animal. The zebrafish multi-omics study of Cr(VI) plus chlorfenapyr (Multi-omics analysis reveals the synergistic and intergenerational toxicity of Cr(VI) and chlorfenapyr on zebrafish: disruption of gut-liver axis and mitochondrial impairment) measured mixture synergy with bulk transcriptomics and survival, but could not say which early progenitor populations harbored the initial damage; we resolve exactly that, cell by cell, in the same organism and with the same mixture as an anchor. And the AI-EM tool papers themselves — ATFormer, DeepOrganelle, MorphoFeatures, and the cross-tissue transfer-learning study (Transfer learning improves performance in volumetric electron microscopy organelle segmentation across tissues) — supply the method but have never been pointed at chemically injured tissue. We reuse their techniques; we do not modify their architectures.

Research Program

The experimental host is the zebrafish larva, chosen deliberately so the organism-level synergy benchmark carries over: the Cr(VI) + chlorfenapyr study reports Marking's Additive Index values of 1.09 in adults and 1.28 in embryos in this species, and zebrafish larvae are small enough that serial block-face volume EM (vEM) can cover an entire gut region including its resident immune cells. Zebrafish also offer the lineage separation we need: primitive macrophages arise from the early embryonic hematopoietic wave in a spatially and temporally distinct way from the definitive wave that later seeds the gut, and reporter lines plus macrophage-deficient mutants let us identify or remove each population. The lineage-resolved question is grounded in the human yolk sac atlas of Goh et al. (169,798 cells), which mapped the accelerated, embryonic route to tissue-resident macrophages and showed these primitive cells carry metabolic and transcriptional programs distinct from definitive ones — a plausible source of differential ferroptotic vulnerability, since ferroptosis is fundamentally a lipid-metabolism failure.

The measurement platform is the second thing we build. We will fine-tune a pretrained 3D ResUNet, following the transfer-learning recipe of the cross-tissue vEM study, to segment mitochondria, endoplasmic reticulum, and intracellular particles in zebrafish larval gut vEM volumes, and we will use ATFormer (Pan et al.) as the instance-segmentation engine for mitochondria, since its adaptive templates were designed exactly for the mix of tiny and sprawling mitochondria that damaged macrophages present. On top of the segmentations we compute per-cell ultrastructural features in the spirit of MorphoFeatures — organelle shape, texture, and contact-site densities — and ask whether these features alone classify exposure condition (vehicle, particle only, pesticide only, mixture) and lineage (primitive versus definitive), compared against manually designed morphological features. The full design is a three-factor grid: lineage (primitive, definitive, macrophage-ablated sibling control), exposure (four conditions at three concentrations spanning the biphasic switch reported by Du et al.), and tissue context (gut, where the Cr(VI)+chlorfenapyr phenotype is best characterized). The macrophage-ablated arm tests the second half of the claim: whether removing resident immune cells changes the epithelium's own ultrastructural injury under mixture exposure, which current organoid-based risk-assessment frameworks explicitly cannot ask because, as Bang et al. state, their organoids lack resident immune populations.

Proposed Experiments

Experiment 1 establishes the measurement platform on injured tissue. We acquire vEM volumes of larval gut at 5 days post-fertilization across the exposure grid, annotate a seed set of roughly 100 patches per organelle class (the amount the transfer-learning study shows is sufficient, reporting 92.37% IoU for mitochondria and 77.12% for ER after fine-tuning with 100 patches), and report our segmentation quality against the same benchmark numbers. Success is mitochondria IoU within a few points of the 92.37% reported on rat liver, demonstrating that the transfer-learning approach survives chemically injured, fixation-stressed tissue. Where fixation threatens artifacts, we adopt the buffer-optimization lesson of Anatomics MLT, which showed that standard 150 mM buffers collapse cells and hypotonic preparation is required for fragile specimens.

Experiment 2 is the decisive biological test. For each lineage and exposure condition we compute the per-cell ultrastructural phenotype: a ferroptosis morphology score built from mitochondrial shrinkage, membrane electron density, and cristae continuity (the hallmarks fixed by Chen et al. and used as ground truth by Wang et al. and by the microglial ferroptosis study of Wang et al.); ER–mitochondria contact density, the metric DeepOrganelle proved trackable at scale; and label-free intracellular particle counts. We then compute a morphology-based Additive Index, comparing the mixture condition against the two single-exposure conditions within each lineage, and test whether it diverges from the organism-level index of 1.28 reported for embryos by the Cr(VI)+chlorfenapyr study at matched doses. The key readout is a lineage-by-exposure interaction: a statistically significant difference in injury score between primitive and definitive macrophages under the mixture but not under vehicle, with the direction of the difference telling us whether primitive cells are the weak link. The macrophage-ablated siblings give the causal complement: if epithelial ferroptosis scores rise when resident macrophages are absent, the resident immune population is protective, not merely a victim.

Experiment 3 addresses the 2D-versus-3D methodological tension directly, because it is the main threat to scalability. DeepOrganelle validated 2D EM organelle-contact ratios as a proxy for 3D on healthy, relatively isotropic testis tissue, while cautioning that polarized tissues bias 2D estimates; MultiOrg independently showed that compressing 3D data into 2D projections loses spatial information. The larval gut is exactly the polarized case in question. We will therefore run paired 2D tile-scans and 3D vEM on a subset of conditions and measure the disagreement between 2D-derived and 3D-derived contact and injury metrics, publishing the bias as a function of polarization. If 2D is adequate, the platform scales to the full grid; if not, we will say so and quantify what 2D screening actually costs in this regime.

A small human-arm validation closes the loop back to the atlas. We will expose iPSC-derived primitive-like macrophages (the accelerated yolk sac route mapped by Goh et al.) and monocyte-derived macrophages to the same mixture in 3D culture, embedding them in macrophage-containing organoids of the kind captured in the "Macros" subset of MultiOrg, and run 2D TEM on the ferroptosis morphology score. Concordance with the zebrafish directionality would argue the asymmetry is a general property of primitive macrophage biology rather than a fish-specific artifact — a live concern, since Goh et al. document species divergence in embryonic hematopoiesis timelines.

Baselines

The naive reference point is expert qualitative TEM scoring, which is what the multi-contaminant literature actually does: Wang et al. and Muhammad et al. verify ferroptosis and tissue damage by eye on a small number of micrographs, and the zebrafish mixture study (Multi-omics analysis) describes villi detachment qualitatively. We will run this baseline properly — three blinded expert raters scoring the same images our pipeline processes, on the same cells — with the metric being inter-rater agreement and the ability to rank exposure conditions. Beating it means the automated per-cell metrics resolve differences between lineage and mixture conditions that experts either cannot rank consistently or rank with low agreement; if experts match the pipeline, the pipeline still wins on throughput, but the honest claim would then be scale, not insight, and we would report it that way.

The strongest existing methodological competitor is the transfer-learned 3D ResUNet pipeline of Transfer learning improves performance in volumetric electron microscopy organelle segmentation across tissues, which is the current best recipe for segmenting organelles in a scarce-data, novel tissue. It reports 92.37% IoU for mitochondria and 77.12% for ER on SBF-SEM rat liver after fine-tuning with 100 patches, and an improvement from 0.811 IoU with random initialization to 0.904 with pretraining at only 25% of the training data — numbers we adopt as the target our zebrafish-injury segmentation must approach. It will be run with the same fine-tuning patch budget, the same annotation effort, and the same single-GPU compute (the transfer-learning paper's regime, comparable to the 10–24 hour trainings reported by Anatomics MLT on one RTX A6000), so that any difference is attributable to tissue condition, not resources. Beating it is not the goal for segmentation per se; matching it on injured tissue is the goal, and falling more than a few IoU points below it would mean injury-associated appearance shift is a real problem, at which point the proposal's fallback is ATFormer, whose adaptive templates were built for appearance variation and which reports 78.2% mAP on MitoEM-R and 68.2% on MitoEM-H (Pan et al.).

The ablation that isolates our key analytical ingredient is feature-set ablation, not segmentation ablation. Following the design of MorphoFeatures, where learned features achieved 96% cell-classification accuracy against 94% for manually defined features, we will classify lineage and exposure condition twice: once from learned self-supervised features and once from a carefully hand-built feature set (mitochondrial aspect ratio, cristae continuity score, contact density, particle count). If the manual features match the learned ones, the finding is stronger and cheaper — the asymmetry is visible in interpretable quantities. The second ablation, transfer versus random initialization on our injured zebrafish data, isolates the contribution of pretrained EM representations and directly tests whether the 0.811-to-0.904 gain survives the domain shift to toxicologically injured, polarized larval tissue, for which no reported result exists on Lacuna.

Why not just…?

Why not just use confocal or super-resolution fluorescence microscopy? Fluorescence cannot see the structures that define the question. The ferroptotic signature that distinguishes this death mode from apoptosis — cristae dissolution, outer-membrane rupture, increased membrane electron density — is a sub-200 nm ultrastructural phenotype that the light diffraction limit cannot resolve, which is why Wang et al. and Wang et al. both reached for TEM as the deciding evidence. There is also a chemistry problem: attaching fluorophores to 100 nm polystyrene particles changes their surface properties and therefore their adsorption of the co-contaminant, corrupting the very mixture effect under study. Fluorescence has a real supporting role in our program — lineage reporters for sorting — but it cannot substitute for the organelle-level measurement, and Experiment 1's per-cell TEM metrics are precisely the readout fluorescence cannot provide.

Why not just run bulk biochemical and molecular assays? These are cheap and scalable, and we will run them alongside (GPX4 Western blot, lipid peroxidation, qPCR of the mitochondrial genes the zebrafish mixture study tracked). But they fail in two documented ways that matter here. First, they lose spatial localization: in a tissue where resident macrophages are a small minority among epithelial cells — the McGadey et al. Harderian gland work shows resident immune populations can be grossly sex- and hormone-skewed, so abundance varies wildly by condition — a bulk homogenate dilutes the lineage-specific signal we are hunting into noise. Second, bulk markers are ambiguous between death modes, which Wang et al. note as the reason TEM is the ground truth for ferroptosis. The proposal's decisive check against this alternative is built in: if bulk assays detect the lineage asymmetry as sensitively as per-cell ultrastructural scoring across the exposure grid, the simpler method wins and we will say so; our prediction, grounded in the dilution argument, is that they will not.

Why not just predict mixture toxicity with concentration-addition or independent-action models? The compiled evidence is unambiguous that these models fail in exactly our regime. Du et al. showed heavy-metal-plus-leachate mixtures switch from antagonistic complexation to synergistic oxidative stress as concentration rises, and Muhammad et al. showed nanoplastics physically vector a pesticide into tissues in a way no active-ingredient math model captures; Luo et al. even found dissolved organic matter amplifying rather than buffering toxicity. Our design leans into this rather than around it: the concentration series in Experiment 2 deliberately spans the biphasic switch, and the morphology-based Additive Index is explicitly a way to measure where and in which cells the additivity assumption breaks. If the Additive Index computed on ultrastructure tracks the bulk survival index perfectly at all doses, that would be evidence the mixture acts additively at the subcellular level, and the asymmetry claim would be weakened — the experiment settles it either way.

Why not just use existing 2D cell cultures, the cheapest option? This alternative is a genuine threat to the human arm, and we concede part of it: 2D cultures are the right first screen, and Bang et al. build their fast-track framework on exactly that logic. But the same authors identify the ceiling: their organoids lack resident immune cells, so the question of immune-mediated and lineage-specific injury — the entire point of this proposal — cannot be asked in those systems at all. Moreover, Wang et al. showed particle uptake is size- and mechanics-dependent, and monolayer culture artificially forces particle–cell contact, distorting uptake relative to intact 3D tissue. Our compromise is explicit: the in vivo zebrafish work carries the causal and lineage claims, and the 2D human-cell arm is framed as directional validation only, with macrophage-containing 3D organoid cultures (MultiOrg) rather than monolayers as the human testbed.

Decisive Checks

Three outcomes would decide the proposal. First, a significant lineage-by-exposure interaction on the ferroptosis morphology score, with the mixture condition separating primitive from definitive macrophages where single exposures do not — this is the core claim confirmed. Second, a morphology-based Additive Index that flags synergy at dose tiers where the survival-based benchmark stays near additivity, establishing per-cell ultrastructure as the more sensitive toxicity endpoint. Third, segmentation within striking distance of the 92.37% mitochondria IoU that the transfer-learning study achieved on healthy rat liver, proving the AI-EM toolkit survives injured tissue and can therefore be handed to the toxicology community as a working measurement instrument. A null result on the first check — symmetric injury across lineages — would itself be informative, ruling out primitive-lineage hypersensitivity as the mechanism behind early developmental mixture toxicity and pointing the field back to the epithelium, but we would report it as a constraint on the hypothesis, not as a success.

References

  1. Enhanced bioaccumulation and toxicity of Fenpropathrin by polystyrene nano(micro)plastics in the model insect, silkworm (Bombyx mori)
  2. Joint toxicity of cadmium (II) and microplastic leachates on wheat seed germination and seedling growth
  3. Molecular mechanism differences between nanoplastics and microplastics in colon toxicity: nanoplastics induce ferroptosis-mediated immunogenic cell death, while microplastics cause cell metabolic reprogramming
  4. TXNIP mediates ferroptosis in a bronchopulmonary dysplasia mouse model by regulating the SLC7A11/GPX4 pathway
  5. Multi-omics analysis reveals the synergistic and intergenerational toxicity of Cr(VI) and chlorfenapyr on zebrafish: disruption of gut-liver axis and mitochondrial impairment
  6. Adaptive Template Transformer for Mitochondria Segmentation in Electron Microscopy Images
  7. AI-driven Large-scale Electron Microscopy enables Whole-tissue Subcellular Digitization
  8. MorphoFeatures for unsupervised exploration of cell types, tissues, and organs in volume electron microscopy
  9. Transfer learning improves performance in volumetric electron microscopy organelle segmentation across tissues
  10. Yolk sac cell atlas reveals multiorgan functions during human early development
  11. Developing a feasible fast-track testing method for developmental neurotoxicity studies: alternative model for risk assessment of micro- and nanoplastics
  12. Anatomics MLT, an AI tool for large scale quantification of ultrastructural traits
  13. Minocycline alleviates microglia ferroptosis by inhibiting HO-1 during cerebral ischemia–reperfusion injury
  14. MultiOrg: A Multi-rater Organoid-detection Dataset
  15. Mast cells in the hamster Harderian gland: sex differences, hormonal control and relationship to porphyrin.
  16. Toxicological effects, bioaccumulation, and metabolic pathways of tricresyl phosphate in Scenedesmus obliquus