Ultrastructural Birth-Dating of Mast Cells in the Castrated Hamster Harderian Gland

Some tissues hold cell populations whose origin nobody has worked out. A clear case is the golden hamster's Harderian gland, a tear-film gland behind the eye. Females carry on average forty times more mast cells there than males. Mast cells are immune cells packed with histamine-filled granules. When a male is castrated, his mast cell numbers climb to female levels over about five months, and testosterone replacement prevents the climb (Mast cells in the hamster Harderian gland). That same study asked, and could not answer, whether the new cells come from progenitors recruited out of the blood or from quiet progenitors already in the gland that divide and mature on the spot. Its transmission electron microscopy (TEM, imaging thin tissue sections with an electron beam at nanometer resolution) found only mature, normal-looking granules. Static snapshots of finished cells cannot say where those cells came from.

We propose to turn this hormone-triggered expansion into a controlled experiment on cell origin. Castration starts a synchronized wave of new mast cells. We would pulse-label newly born cells with a DNA tag at a known time, then image the gland in three dimensions with volume electron microscopy. That gives nanometer-scale pictures of cells whose age is known. From these pictures we would learn an "ultrastructural pseudotime", a continuous ordering of cells from immature to mature built only from what their organelles look like, and check it against the known cell ages. With that ordering in hand, we then ask where the youngest mast-lineage cells sit: inside or against capillaries, which would favor recruitment, or dividing among the gland's connective tissue cells, which would favor local expansion. Part of the value is methodological. Nobody has yet tested whether electron microscopy morphology can recover developmental time rather than just a fixed list of cell types.

Core Claim

Our central claim is that the order in which organelles mature, including granule condensation, the growth and then shrinkage of the protein-making rough endoplasmic reticulum, nuclear shape, and contacts with vessels, carries enough information to rank cells by developmental age. We think this ranking will extend back past the point where granules become visible. If so, ultrastructure can place cells on a lineage path without transcriptomics. The Harderian gland can then answer the recruitment-versus-proliferation question, which neither cell counts nor sequencing has settled.

Two pieces of prior work show that electron microscopy morphology identifies cell types. Elabbady et al. classified cortical cells from the cell body and nearby processes at 91% accuracy (Perisomatic ultrastructure efficiently classifies cells in mouse cortex). MorphoFeatures learned self-supervised shape and texture embeddings across a whole Platynereis larva (MorphoFeatures). Both work on a single time point in adult or fixed-stage tissue and produce a static set of types. Neither has any ground truth about time.

On the trajectory side, diffusion pseudotime orders cells along branching lineages (Diffusion pseudotime robustly reconstructs lineage branching), but only in gene-expression space. Optimal-transport integration methods align unpaired modalities, again without electron microscopy. scConfluence works on 2D light-microscopy morphology (scConfluence), and MatchCLOT matches molecular modalities to each other (MatchCLOT).

The obvious combination would be to align electron microscopy embeddings to a single-cell RNA-sequencing (scRNA-seq) atlas and borrow its pseudotime. We deliberately avoid making that the core. It would inherit the atlas's time ordering without ever testing it, and no hamster Harderian atlas exists to borrow from. Our ground truth is physical birth-dating, which none of these papers has.

Research Program

The biological setup follows Payne et al. We would use intact males, intact females, castrated males sampled at 2, 4, 8, 12 and 20 weeks, and castrated males given testosterone. The DNA tag is EdU (5-ethynyl-2'-deoxyuridine), a thymidine analog that labels any cell that copies its DNA while the tag is present. It is given in two regimes.

The first regime is a short pulse, a few hours before sacrifice. This marks cells dividing inside the gland at that moment. The second regime is a pulse at week 2 after castration followed by chase periods of 1, 4 and 10 weeks. This birth-dates the cohort born in that window, whether the cells divided in bone marrow and were later recruited or divided in the gland. Because the tag marks the time of birth rather than the place, the chase design gives known-age cells under either hypothesis.

For imaging, we would first photograph fluorescent EdU and a mast-lineage marker in thick sections. The marker would be c-Kit (CD117) or, if a hamster-reactive antibody can be validated, the selectin ligand sialyl-Lewis x, which circulating mast progenitors need in order to roll along vessel walls (FUT6 deficiency compromises basophil function). The same tissue would then be imaged by serial block-face electron microscopy, which repeatedly images and shaves the face of a resin block to build a 3D volume. Registering the two gives a correlative light and electron dataset. We would collect interstitial and capsular volumes of several hundred micrometers per side, from three animals per condition.

The computation proceeds in three steps. First, we segment cells and nuclei in the 3D volumes, starting from an encoder pretrained on the CEM500K collection of heterogeneous electron microscopy images (CEM500K). Second, we compute per-cell embeddings in the MorphoFeatures style: shape, coarse texture and fine organelle texture. Alongside them we compute interpretable measurements: granule count, volume and electron density, rough endoplasmic reticulum volume fraction, nuclear infolding, and distance to the nearest capillary lumen. Third, we build a neighbor graph over marker-positive and marker-adjacent cells and run diffusion pseudotime from a root chosen without looking at EdU, namely the least granulated marker-positive cell. The EdU chase labels are held out entirely for evaluation.

Baselines

The naive reference point is to order cells by granule volume fraction alone. This is what a histologist would do, and it encodes the assumption that maturity simply means more granules. If the learned ordering cannot beat it on known-age cells, the deep representation adds nothing.

The strongest existing method is the MorphoFeatures pipeline: its self-supervised shape and texture encoder, retrained on our volumes with the same architecture and training schedule, fed into diffusion pseudotime. It is the best published general-purpose cell representation for volume electron microscopy. In its original setting it beat 140 handcrafted features at classifying seven broad cell types (96% versus 94% accuracy, MorphoFeatures).

The second strong competitor is an Elabbady-style handcrafted feature vector (cell volume, nuclear infolding, organelle fractions) fed into the same pseudotime procedure. That design classified cortical cells at 91% overall and 97.5% for non-neuronal cells (Perisomatic ultrastructure efficiently classifies cells in mouse cortex).

To keep the comparison fair, all methods get the same segmentations, the same cells, the same root rule, the same pseudotime parameters, and an equal hyperparameter search budget. Tuning is done on animals separate from the held-out EdU-chase animals.

The key ablation removes vessel-contact and position features from the embedding. This checks that the ordering reflects intrinsic organelle maturation rather than location, which matters because location is what we later use to infer origin. A second ablation removes the fine-texture channel to test whether organelle-scale detail, not cell shape, carries the time signal.

For reference, diffusion pseudotime reached a Kendall rank correlation of 0.78 with experimental time on mouse embryonic stem cell RNA data (Diffusion pseudotime). No reported result exists on Lacuna for pseudotime inferred from electron microscopy morphology on any dataset.

We would count the proposal as succeeding if the learned ordering reaches a Kendall correlation with EdU chase age that is significantly higher than both the granule-only ordering and the handcrafted-feature ordering, as judged by bootstrap over animals. We would treat roughly 0.6 as the bar for a useful clock.

Why not just…?

Why not just count EdU-positive mast cells after a short pulse? This is a real threat to half the question, and we include it as an arm of the design. A short pulse with immediate sampling directly measures division inside the gland. If many granulated cells are EdU-positive, local proliferation is real. It cannot rule out recruitment followed by local division, though, and it says nothing about how agranular precursors become recognizable mast cells. If the short-pulse arm alone shows that local division fully accounts for the observed growth rate, we will say so, and the origin question narrows to what the resident precursor is.

Why not just do scRNA-seq and pseudotime on the gland? Sequencing dissociates the tissue. It loses exactly the variable that separates the two hypotheses, which is whether the youngest cells sit in or against capillaries. It also measures transcripts rather than organelle state, and Payne et al. already showed that endocrine state can change cell numbers fortyfold while leaving granule structure unchanged. A small scRNA-seq arm aligned with scConfluence-style transport remains a useful cross-check. If RNA pseudotime ranks known-age cells better than the ultrastructural ordering, electron microscopy loses its claim to be a lineage clock, and the experiment would show that directly.

Why not just do parabiosis, or use mouse fate-mapping? Parabiosis in hamsters, surgically joining two animals so they share a blood supply and tracking donor-derived cells, would directly test recruitment. It is heavy, and it confounds hormonal state across the joined partners, so we keep it as a follow-up only if the spatial results come out ambiguous. Mouse genetic fate-mapping is cleaner in principle, but the laboratory mouse does not show this dimorphism at comparable magnitude, and the amplified hamster phenotype is what makes the wave measurable.

Proposed Experiments and Decisive Checks

The first decisive check is the clock test. Among EdU-positive, marker-positive cells in the chase animals, the ultrastructural pseudotime must rise with chase duration. We also expect the fraction of cells in early pseudotime to peak at 2 to 4 weeks after castration and stay near zero in testosterone-replaced and intact males.

The second check is the origin test, and it applies only if the clock test passes. We compare the distances from earliest-pseudotime cells, and from 1-week-chase EdU-positive cells, to capillary lumens against a null model that places the same cells randomly in the interstitium. We also score intravascular or endothelium-contacting agranular marker-positive profiles. The two hypotheses make different predictions:

  • Recruitment: young cells cluster at vessels, and short-pulse local division is low.
  • In situ expansion: young cells are dispersed through the connective tissue, and mitotic figures and short-pulse labeling are common among partly granulated cells.

A mixed result is quantifiable as the fraction of growth explained by each route.

The main risk is that agranular precursors cannot be told apart from lymphocytes or fibroblasts by ultrastructure alone. The correlative marker imaging is the guard against this. If the marker cannot be validated in hamster tissue, the trajectory will be restricted to granule-bearing stages and the claim scaled back accordingly.

A secondary transfer test, which is not decisive, would project bullfrog adrenal "summer cells", histamine-storing cells of unresolved lineage (Occurrence and release of histamine-containing granules in summer cells), into the learned embedding. It would ask whether they fall on the mast-cell maturation path or next to the steroid-producing lipid cells that surround them.