Three-dimensional modeling of sensory nerve architecture and eosinophil and mast cell interactions in eosinophilic gastrointestinal disease

Abstract

Increased sensory nerve density has been described in type 2 inflammatory conditions and is linked to eosinophil and mast cell infiltration and neuropathic pain. To examine these relationships in eosinophilic gastrointestinal diseases, we developed a protocol using Ce3D (clearing-enhanced 3D) tissue clearing to evaluate mucosal gastrointestinal (GI) nerve remodeling and interactions with eosinophils and mast cells in whole-mount GI biopsies. Human gastric and esophageal biopsies were processed for 3-dimensional imaging. Tissues were fixed, permeabilized, and immunolabeled for mucosal nerves (Protein Gene Product 9.5), sensory neurons (substance P), eosinophils (eosinophil peroxidase), mast cells (tryptase), and epithelial cells (pan-cytokeratin). Vacuum-microwave–assisted staining was employed to enhance antibody penetration and protocol efficiency. Samples were cleared with Ce3D and imaged using confocal microscopy. Machine learning–enhanced quantitative analysis and modeling of nerve morphology and cellular interactions, including nerve length, branching, and spatial relationships between eosinophils, mast cells, and sensory neurons were performed. We established a 3-dimensional imaging method for whole-mount GI biopsies to characterize nerve architecture and eosinophil and mast cell interactions in the human esophagus and stomach. This approach enables high resolution and volumetric analyses and can be modified to assess other spatial neuroimmune interactions in human GI mucosal biopsies.

This 3-dimensional whole-mount tissue imaging method enables visualization and quantification of the spatial relationships among sensory nerves, eosinophils, and mast cells in human gastric and esophageal tissues, providing insight into neuroimmune interactions in gastrointestinal disorders, including eosinophilic gastrointestinal disorders.

1. Introduction

The gastrointestinal (GI) tract is regulated by the enteric nervous system (ENS), a complex network of neurons and glia essential for motor, sensory, and secretory functions.^1,2 Recent advances highlight the critical role of neuroimmune interactions—particularly between enteric neurons, mast cells, and eosinophils—in maintaining gut homeostasis and contributing to the pathophysiology of disorders of gut-brain interaction, including irritable bowel syndrome and functional dyspepsia. Inflammation in the setting of inflammatory bowel disease or infection may also disrupt these interactions, contributing to altered neuronal function, visceral hypersensitivity and dysmotility.^3–10 Recent research has expanded our understanding of the relationship between sensory nerves and type 2 inflammatory conditions.¹¹ For example, increased airway sensory innervation is observed among individuals with persistent eosinophilic asthma and is mediated by eosinophils.^12,13 Similarly, cutaneous nerves are increased in atopic dermatitis and eosinophils promote sensory neuron branching in vitro.¹⁴ Eosinophilic gastrointestinal disorders (EGIDs), including eosinophilic esophagitis (EoE) and eosinophilic gastritis (EoG), are characterized by exuberant T helper 2 responses in the affected tissue^15–17 with eosinophil and mast cell infiltrates.^18–20 In EoE, mast cells contribute to esophageal barrier dysfunction and may modulate pain associated with EoE.^19,21,22 Even after histological remission, esophageal mast cells can remain transcriptionally activated, which may perpetuate clinical symptoms despite resolution of tissue eosinophilia.^23,24 Similarly, infiltrating mast cells in the gastric mucosa of patients with EoG exhibit evidence of activation and degranulation.²⁵ Furthermore, direct and indirect interactions between mast cells and eosinophils are well established, resulting in increased bidirectional cell activity and mediator release that likely amplifies local disease pathology.²⁶ Although increasing evidence suggests that eosinophils and mast cells play crucial roles in neurogenic inflammation and potentially contribute to ENS dysfunction, their specific interactions with the ENS have not been thoroughly examined in the context of EGIDs.

Tissue imaging techniques have provided valuable insights into disease pathogenesis by identifying and mapping cellular microenvironments. However, GI nerve structure is challenging to study using conventional histology, which typically relies on thin (<10 µm) formalin-fixed, paraffin-embedded sections. Two-dimensional (2D) sections are limited in their ability to capture the full complexity of nerve networks, leading to sampling bias, loss of spatial context, and inaccurate estimates of nerve density and volume.^27–29 Nonuniformly distributed cells in tissues can produce misrepresentations of true cell volume per area depending on the section sampled, as rare cellular populations may be insufficiently captured in a thin section.³⁰ 3D microscopy techniques overcome these limitations by enabling volumetric visualization and quantification of nerve architecture within intact tissue samples.³¹ Moreover, cleared confocal microscopy using whole-mount samples allows for detailed, high-resolution analysis of spatial neuroimmune cell interactions.^13,31–34 While 3D techniques like CUBIC, enGLOW, and those described in Graham et al.²⁷ offer in-depth tissue visualization, these protocols are time-consuming, taking 2 to 4 wk, from sample collection to data acquisition.^32,35,36 Microwave-assisted immunostaining can significantly enhance antibody penetration and reduce staining duration compared with conventional benchtop methods, allowing for improved image quality in less time.^37–39

Given the limitations of 2D methods to accurately investigate the morphological distribution and relationships between nerves and myeloid cells in GI tissues, we developed a 3D imaging protocol. This technique utilizes the Ce3D tissue clearing method and vacuum-microwave–assisted immunostaining to comprehensively determine stomach and esophageal nerve architecture and eosinophil and mast cell interactions. Our study presents an efficient and detailed approach for spatial modeling and quantitative analysis of nerve structures and cellular involvement in whole-mount, gastric, and esophageal endoscopic biopsies.

2. Methods

2.1. Patient information

Gastric antrum and proximal and middle esophageal research biopsies were collected during clinically indicated endoscopic procedures from participants enrolled on an National Institutes of Health institutional review board–approved research protocol (NCT00001406) to study eosinophilic disorders. Participant clinical and biopsy sample information is summarized in Table S1. All participants signed informed consent.

2.2. Tissue clearing and immunostaining

Human gastric antrum and esophageal endoscopic biopsies were fixed with BD Cytofix/Cytoperm (BD Biosciences), dehydrated with 30% sucrose, and frozen in optimal cutting temperature compound for storage at −80 °C. After thawing, samples were washed in phosphate-buffered saline and blocked with Human Fc block (BD Pharmingen) overnight. GI nerves were labeled with primary antibodies for the pan neuronal marker, Protein Gene Product 9.5 (PGP9.5) (Invitrogen) and sensory neurons with substance P (SP) (BD Pharmingen), followed by secondary antibodies Alexa Fluor 555 and 647, respectively (Invitrogen). Eosinophils, mast cells, and epithelial cells were labeled with conjugated antibodies targeting eosinophil peroxidase (EPX) Alexa Fluor 488 (Mayo Clinic), mast cell tryptase (MCT) Alexa Fluor 790 (BioLegend), and pan-cytokeratin Alexa Fluor 594 (BioLegend), respectively (Table S2). EPX and MCT antibodies were conjugated using Alexa Fluor Antibody Labeling Kits (Invitrogen). Staining was performed using a vacuum chamber in a tissue processing microwave (PELCO BioWave Pro). Samples underwent 1-h cycles at 20 mm Hg with 0 to 150 W intervals at 20 °C across 4 d of staining. On the first and third days of staining (after primary and secondary/conjugated staining respectively), samples underwent 1 microwave cycle. On the second and fourth days, samples underwent 4 cycles. Samples were gently rocked between each of the staining cycles. Following staining, tissues were postfixed with 1% paraformaldehyde, washed with phosphate-buffered saline, and optically cleared using clearing-enhanced 3D (Ce3D) tissue clearing solution (1.455 g/mL Histodenz [Sigma-Aldrich] in 40% N-methylacetamide [Sigma-Aldrich]) for 48 to 96 h. Samples were mounted in Ce3D to glass slides, covered with a cover slip, and sealed with nail polish.

2.3. Image acquisition and processing

Images were acquired using a Leica Stellaris confocal microscope equipped with a tunable white light laser, PL APO 20×/0.75 objective, HyD and APD detectors, and a motorized stage. Images were captured with a pixel size of 0.758 × 0.758 × 3 um and individual fields were mosaic-merged to generate a continuous representation of the tissue in the LAS X software. To minimize imaging artifacts, corrections were applied for motion-induced distortions, 3D alignment inconsistencies, and thermal drift across sequential z-sections. Additionally, crosstalk and color calibration adjustments were performed using Huygens Pro (version 24.04.0p3, 64-bit; Scientific Volume Imaging BV). Image deconvolution was conducted within the same software package to enhance signal resolution.

2.4. 3D mapping and machine learning implementation

Reconstructed images were analyzed using Imaris (version 10.2.0; Bitplane AG). A combination of colocalization analysis, ChannelArithmetics Xtension, machine learning (ML)–based classification, and masking techniques was employed to delineate distinct structures such as filaments and cells. ML was incorporated into the image analysis workflow to improve the accuracy and reproducibility of filament reconstruction and cellular segmentation within heterogeneous 3D biopsy datasets. After deconvolution and channel correction, voxel-level signal classification was performed using the Imaris ML segmentation framework, which applies a supervised classifier trained to recognize structural features characteristic of neuronal filaments and immune cell surfaces. The framework enabled more reliable reconstruction of nerve trajectories through regions of low-contrast, discontinuous staining, or high background fluorescence. ML also served to refine cell segmentation, ensuring consistent surface generation across samples with variable staining intensity. These integrated ML-guided approaches improved structural fidelity and supported downstream quantification of nerve length, branching, filament-to-cell distances, and surface-surface contact.

2.5. Filament and surface model generation

Neuronal SP expression was measured by identifying the colocalization of SP-positive voxels within PGP9.5-positive nerve axons. Filament objects were converted into spots using the Filaments to Spots Xtension in Imaris, allowing for assessment of nerve length, branch points, and distance measurements between filaments and cell surfaces. EPX and MCT signals were quantified in each GI compartment and segmented as surface objects, yielding the eosinophil and mast cell segmentations with their centroids serving as the spot location for these cells. For tissue samples with uneven signal distribution and reduced signal intensity, ML-based classification allowed for partial-volume segmentation of regions of interest. These were subsequently merged to reconstruct complete filament networks avoiding over- or underthresholding. Constructed Imaris filament and cell objects were used for tissue visualizations and their coordinates (X, Y, Z in µm) and edge connectivity were exported for spatial analysis. The topological degree of nerves was classified as terminal (degree = 1), continuation (degree = 2), or branch point (degree ≥ 3). The Surface module was employed to measure surface contacts, in which unsmoothed surface objects representing filaments and cells were created to capture regions of physical overlap using the Surface–Surface Overlap Xtension. For standardization, 3 individual stacks were randomly selected per patient from the entire stained tissue sample. The in-plane field of view and depth of the tissue z-stack was approximately 512 µm × 512 µm × 100 µm.

2.6. Spatial and morphologic organization of eosinophils and mast cells relative to nerve fibers

Nerve and cell models were generated for each tissue z-stack. Three nonoverlapping z-stacks per participant biopsy were analyzed independently, and per-patient values were derived as cell count–weighted means across stacks. Total nerve length and the number of branch points were quantified from the filament model. Eosinophil and mast density was measured as the number of cell surfaces per volume of the tissue stack (cell/µm³). SPACE (Spatial Patterning Analysis of Cellular Ensembles) was performed in R to assess nerve–cell interactions in 3D whole-mount tissue samples. Using annotated cell and nerve identities with their x-y-z coordinates, SPACE applies an information-theoretic framework to detect and rank nonrandom spatial relationships.⁴⁰ Cell-to-nerve distance was defined as the minimum 3D Euclidean distance from each cell centroid to the nearest point on any nerve segment. Computations were performed on an NVIDIA L40S GPU using PyTorch. All coordinates were centered by subtracting the volume minimum prior to distance calculation to maintain floating-point precision at the submicrometer scale. Cell-nerve contact was evaluated across a range of distance thresholds and contact rate was defined as the fraction of cells within each threshold distance of the nearest nerve fiber. For cells and nerves within contact distance thresholds, the microarchitecture such as terminal, continuation segment, and branch points were calculated and their proportion of each associated to cell type was assessed.

2.7. Epithelial surface reconstruction and cell compartmentalization

The epithelial layer was reconstructed in 3 dimensions to measure nerve and immune cells' proximity, density, and distribution. The pan-cytokeratin signal was segmented using 3D U-Net boundary prediction model (PlantSeg), followed by watershed segmentation and agglomerative clustering to produce labeled tissue regions. Cells with negative signal distance were classified as intraepithelial and cells with positive distances were classified as epithelial distances. Cell and nerve depth distribution used vertex coordinates to map objects across the pan-cytokeratin signal with 0 µm marking the luminal side of the epithelium and 500 µm marking the serosal. Whole-mount samples and stacks were manually reviewed to assess the orientation of the tissue.

2.8. Statistical analysis

Analysis was performed blinded to participant clinical status and diagnosis. For the proof-of-concept application of the spatial modeling workflow, analyses were intended to demonstrate the spatial and quantitative capabilities rather than compare across clinical subtypes. Given the heterogeneity of clinical presentation and a small sample size, statistics reported are observational and descriptive. Descriptive morphometric and density data are expressed as mean ± SEM of 3 nonoverlapping z-stacks per participant biopsy (n = 10) unless otherwise specified. Median cell-to-nerve distance was measured with a 2-tailed P value calculated from the null distribution. Effect size was quantified as the difference between the null mean and observed median, normalized by the null standard deviation (null deviation Z). The minimum reportable P value is 1 × 10⁻⁴ given 10,000 permutations. Per-stack P values were combined at the patient level using Fisher's method. All patient-level tests across cell types and filament types were corrected for multiple comparisons using the Benjamini-Hochberg false discovery rate procedure. Per-patient contact rates were computed as cell count–weighted means across stacks. Cell-to-nerve distances were compared against a null model of complete spatial randomness. For each stack, null populations were generated by sampling points within the convex hull of observed cell positions. Distances were calculated using the same point-to-segment algorithm applied to observed data.

3. Results

3.1. Tissue processing, immunostaining and clearing of whole-mount human gastric biopsies

Endoscopic biopsies of the gastric antrum were labeled with antibodies against PGP9.5, SP, pan-cytokeratin, EPX, and MCT; optically cleared using Ce3D; mounted; and imaged (Fig. 1A, B;  Fig. S1). The median tissue biopsy dimensions in the length (X), height (Y), and depth (Z) were 3,669 µm (interquartile range [IQR] 2,352–4,667 µm), 2,185 µm (IQR 1,992–2,471 µm), and 105 µm (IQR 101–517 µm), respectively (n = 10). A cross-sectional view of a whole-mount biopsy is shown in Fig. 1C and Fig. S2. Neurons, epithelial cells, eosinophils, and mast cells were visually represented throughout the sample, allowing for simultaneous visualization of neuron distribution and proximity to cells (Movie S1). Neuronal filaments were observed in the lamina propria frequently associating with eosinophils and mast cells. These filaments closely followed the antral foveolae and glands, demonstrating a spatial relationship between nerve fibers and the gastric epithelium (Fig. 1D;  Fig. S2).

Figure 1.

Summary workflow and cross-sectional imaging of immunostained whole-mount gastric antrum tissue. (A) Endoscopic biopsies are collected, then fixed and dehydrated for cryoprotection and can be stored for later staining. Samples then undergo vacuum-microwave–assisted immunostaining followed by optical cleaning with Ce3D. Slides are then imaged using a confocal microscope. (B) Whole-mount gastric antrum tissue before and after 72 h of Ce3D clearing. (C) Representative gastric antrum sample illustrating innervation and cellular interactions. Scale bar = 200 µm. The boxed area indicates the field of view for the inset (D). Nerves are labeled with pan-neuronal marker, PGP9.5 (green), sensory nerves (SP, yellow), epithelial cells (pan-cytokeratin, aqua), eosinophils (EPX, pink), and mast cells (tryptase, gray). (D) Slice view images of each marker: PGP9.5, SP (top left to right), EPX, and tryptase (bottom left to right). Scale bar = 50 µm.

3.2. Identifying and mapping sensory neurons in whole-mount gastric antrum samples

To examine the structure of the gastric mucosal ENS, we assessed the distribution of the pan-neuronal marker PGP9.5. PGP9.5-positive neurons demonstrated diffuse arborization throughout the gastric antral tissue (Fig. 2A), underscoring the extensive networks of the ENS. To verify the accuracy of PGP9.5 in identifying mucosal nerves, we compared it with beta-III-tubulin expression, demonstrating similar distribution of neuronal structures (Fig. S3). The sensory nerve subset was identified from SP expression among the PGP9.5-positive neurons (Fig. 2B) based on colocalization of the respective signals (Fig. 2B). 3D nerve filament models were generated for PGP9.5+ gastric neurons and the SP+ sensory nerve subset. The sensory nerve subset volume was quantified (Fig. S4A) and accounted for an average of 12% (range 0%–50%) of total gastric nerve volume across participants (n = 10) (Fig. S4B). The nerve models were generated into spot filaments to create continuous sensory nerves and allow for quantifying distance between filaments and other objects (Fig. 3A;  Movie S2). The generation of 3D nerves allowed for quantitative analysis of nerve density including axon length, and the number of branch and terminal points (Fig. S5). PGP9.5+ gastric neurons demonstrated a median length of 5,407 µm (IQR 712.8–17,043 µm) and 77 branch points, whereas the SP+ sensory nerve subset presented a length of 49.4 µm (IQR 32.7–243.0 µm) and 1.3 branch points. Across both nerve types, the quantities are variable between different disease conditions (n = 9) (Fig. 3B, C). Both PGP9.5+ gastric nerves and SP+ nerve subset was modeled relative to the epithelial layer (Fig. 3D, E;  Fig. S6). The PGP9.5+ gastric nerves were observed throughout the entirety of the tissue stacks (Fig. 3D) and density peaked within 20 µm of the epithelial surface (range −15 to +18 µm across patients) (n = 5) (Fig. 3E). The SP+ sensory nerve subset was largely observed in the superficial epithelium in closer proximity to the luminal surface (Fig. S6A, B).

Figure 2.

Identification and mapping sensory neurons in human gastric antrum of a patient with EoG. (A) Pan-neuronal PGP9.5 signal (green) in whole-mount gastric antrum tissue (oblique view). Scale bar = 300 µm. (B) Magnified area from panel A representing pan-neuron (PGP9.5, green), SP (yellow), merged image, and colocalized sensory nerves (blue). Scale bar = 30 µm.

Figure 3.

3D model generation, morphologic quantification, and spatial distribution of gastric antrum neurons. (A) Whole-mount tissue from a patient with EoG, oblique view demonstrating the SP-positive subset signal (top), the generated 3D model of the sensory nerve subset filament (middle), and overlay of spots (white) tracing the filament model (bottom). Scale bars = 300 µm (left) and 20 µm (right). (B) Total PGP9.5+ gastric mucosal nerve length and (C) branch points across participant samples. Data are represented as the mean ± SEM of 3 nonoverlapping z-stacks for each participant biopsy (n = 9). (D) PGP9.5+ gastric nerve depth distribution from the epithelial surface. Number of PGP9.5+ gastric nerves vertexes (green) and pan-cytokeratin (PanCK) signal intensity (orange) measured from the luminal (0 µm) to the serosal side (500 µm) of the tissue stack (n = 5). (E) PGP9.5+ gastric nerve density throughout the epithelial surface. Data are represented as the mean of 3 nonoverlapping z-stacks for each patient within (−50 to 0 µm) and outside (0 to 200 µm) the epithelium with dashed threshold at epithelial surface (0 µm) (n = 5). DGE, delayed gastric emptying; P, patient.

3.3. Spatial organization of eosinophils, mast cells, and nerve fibers in the gastric antrum

Eosinophil and mast cell tissue density and spatial distribution were mapped in whole-mount gastric tissue to evaluate nerve and cell interactions (Figs. S7 and S8 and Movie S3). Representative patient stacks showed both eosinophils and mast cells closely associated with the sensory nerve filaments throughout the tissue (Fig. 4A–D;  Fig. S9A–D and Movie S3). 3D surface models were generated using EPX and MCT signals to determine spatial proximity between eosinophils, mast cells, and nerves along with their distribution within the tissue. Cell voxel and filament vertex coordinates were used for spatial analysis using SPACE. Significant cellular associations were observed for all 3 variables at the 50 µm censused neighborhoods (|cisMI| = 0.35 bits; P < 0.0001) (Fig. 4A, B;  Fig. S9A, B). The representative whole-mount gastric sample was primarily dominated by PGP9.5+ gastric nerves (37%–100% of latent path) and secondarily by eosinophils (0%–9% and 21%–33%). Mast cells were noted at 10% to 20% and 53% to 59% of the latent path. These trends demonstrate that gastric nerves are the most prominent object, and both cell types largely reside in spatially distinct locations. It also demonstrates that mast cells density is increased when eosinophils and PGP9.5+ gastric nerves are localized within the same neighborhood (Fig. 4B). Both cell types and the SP+ sensory nerve subset showed significant associations across the neighborhoods (|cisMI| = 0.21 bits; P < 0.0001) (Fig. S9A, B). SP+ sensory nerve filament distribution followed a similar trend as the PGP9.5+ gastric nerve's latent path (35%–52% and 57%–100% latent path) (Fig. S9B). Similarly, mast cells were increased when eosinophils and SP+ sensory nerve subset filaments were localized together; however, this pattern was seen in a different region of the tissue (51%–56% latent path) than that of PGP9.5+ gastric nerves.

Figure 4.

Spatial and quantitative relationships of eosinophils and mast cells to gastric nerves and the epithelium in whole-mount gastric antrum tissue. (A) 3D representative rendering of PGP9.5+ gastric nerve (green), eosinophil (pink), and mast cell (white) models in the gastric antrum. Scale bar = 50 µm. (B) SPACE covariation plot for PGP9.5+ gastric nerves, eosinophils, and mast cells with mean, 95% confidence interval (CI), and continuous enrichment score versus randomized expectation plot. (C) Eosinophil (left) and mast cell (right) distance to radius of nerve by patient with complete spatial randomness (CSR) null (95% CI, gray, n = 9). (D) Eosinophil (left) and mast cell (right) contact rate at 10 µm from PGP9.5+ gastric nerve. Dashed line (gray) represents no association between observed and expected (CSR null) contact. (E) Eosinophil and mast cell depth distribution from the epithelial surface. Number of eosinophils and mast cells measured against the pan-cytokeratin (PanCK) signal intensity (orange) measured from the luminal side (0 µm) to the serosal side (500 µm) (n = 5). (F) Cell distance (eosinophil, left; mast cell, right) within and from the epithelial surface. Data are represented from inside (−50 to 0 µm) and outside (0 to 150 µm) the epithelium with dashed threshold at epithelial surface (0 µm) (n = 5). Data are represented from 3 nonoverlapping z-stacks for each patient biopsy (C–F). P, patient.

Eosinophils and mast cells were localized near PGP9.5+ gastric nerves and SP+ sensory nerve subset in which eosinophils' relative positioning to gastric nerves was nonrandom across all patients (n = 9, Fisher's combined P values, false discovery rate corrected, P < 0.05) (Fig. 4C, left). For mast cells, 6 of 7 patients with mast cell data demonstrated nonrandom association with PGP9.5+ gastric fibers (false discovery rate–corrected P < 0.05) (Fig. 4C, right). The contact rate for eosinophils measured at 10 µm from PGP9.5+ gastric nerves varied from 2.6% to 87.1% (n = 9) (Fig. 4D, left), and contact rates for mast cells with PGP9.5+ gastric nerves ranged from 4.9% to 52.8% (n = 7) (Fig. 4D, right). Cells in contact or close proximity (5 to 20 µm) to the nerves concentrated primarily at continuous filament segments compared with branch points and terminal points (Fig. S10). Cellular spatial analysis was also performed with the sensory nerve subset (Fig. S9), demonstrating localization with both eosinophils and mast cells. Similarly, nonrandom positioning of eosinophils to SP+ sensory nerve fibers was observed across all 9 patients (n = 9, Fisher's combined P values, false discovery rate corrected, P < 0.05) (Fig. S9C, left), and 6 out of the 7 patient's mast cell positionings were found to be nonrandom (n = 7, Fisher's combined P values, false discovery rate corrected, P < 0.05) (Fig. S9C, right). At the 10 µm threshold, eosinophils and mast cells in contact with SP+ sensory nerve filaments ranged from 1.0% to 33.9% (n = 9) (Fig. S9C, left) and 4.4% to 42.9% (n = 7) (Fig. S9C, right), respectively.

Cells were also measured based on their distribution across the epithelial surface (Fig. 4E;  Figs. S11 and S12). Eosinophil density was highest in the superficial epithelium, closer to the luminal surface in participants with EoG (P1 and P2), consistent with histopathologic observations of intraepithelial infiltration in EoG.^41,42 In contrast, those without EoG demonstrated the highest eosinophil density deeper in the tissue. Mast cell distribution was heterogenous among the 3 participants with mast cell data, with 2 (P2 and P3) demonstrating higher mast cell density closer to the luminal surface compared with P3 (Fig. 4E;  Fig. S12). Spatial separation between the two cell types was also observed, consistent with the SPACE analysis (n = 3) (Fig. 4B, E). Interpatient variability was observed in the proximity of eosinophils and mast cells to the epithelial surface (Fig. 4F), with 26.2% of eosinophils (IQR 22.9%–44.1%; n = 5) and 21.4% of mast cells (IQR 17.6%–38.7%; n = 3) residing within <20 µm of the epithelial surface.

3.4. Human esophageal innervation and eosinophil and mast cell interactions

The Ce3D clearing and staining method described for gastric biopsies can be adapted to other GI tissue types, including the esophagus. We applied this technique to esophageal tissue in context of EoE, following the same processing, staining, imaging, and modeling procedures. Neurons were observed below the epithelial surface within the lamina propria (Fig. 5A, left). Filaments and cell surfaces were modeled as previously described in whole-mount esophagus samples (Fig. 5A, right). Like the models in the gastric samples, nerve length and branching can be quantified in which the average (mean ± SEM) nerve length was 534.2 ± 288.6 µm and had 7 branch points (n = 1). From these, surface rendering models were generated to analyze cell-filament surface interactions in greater detail (Movie S4). This approach allowed for precise identification of nerve-eosinophil interactions, specifically targeting regions where cells and nerves come into contact. The surface rendering technique enabled visualization of colocalized areas, providing detailed insights into the interactions between eosinophils and neurons in whole-mount GI tissue samples (Fig. 5B–E;  Fig. S13).

Figure 5.

Mapped neuron and eosinophil surface interactions in EoE. (A) Orthogonal PGP9.5-positive neurons (green) localized in the lamina propria beneath pan-cytokeratin–positive epithelial layer (aqua, left) and view of pan neuronal filaments and eosinophil (pink) cell models (right). Scale bar = 50 µm. (B) Magnified inset demonstrating total neuron and eosinophil surface area. (C) Neuronal surface area with eosinophil surface–rendered model (pink) with specific eosinophil surface contact (yellow) with neuronal surface areas. (D) Neuronal surface–rendered model (green) with specific neuron surface contact (light blue) with eosinophil surface area. (E) Eosinophil-neuron surface interaction, in which violet marks the colocalized contact region between eosinophils and sensory nerves. Scale bar = 5 µm.

4. Discussion

EGID pathogenesis remains under investigation, and the role of mucosal sensory nerves may provide greater insight into these diseases. A key challenge is the disconnect between histologic remission (reduced tissue eosinophils) and symptom persistence in EoG and EoE.^43,44 Persistent symptoms may result from mucosal nerve remodeling from prior inflammation, causing lasting damage despite eosinophil reduction.^9,45–47 Accurate spatial assessment of neuronal and cellular interactions in the esophagus and stomach of patients with EGID may offer deeper insight into the neuroimmune mechanisms underlying EGIDs and associated symptoms.

A growing body of literature has highlighted bidirectional neuroimmune interactions between peripheral nerves, eosinophils, and mast cells, whereby type 2 inflammation drives neuronal activation and sensitization.¹¹ Mast cells localize near sensory nerves and respond to neuropeptides such as SP and CGRP (calcitonin gene-related peptide) through the mast cell–specific receptor MRGPRX2 (human) and its mouse homolog Mrgprb2, which trigger degranulation and chemokine production.^48,49 SP-driven MRGPRX2 activation creates a positive feedback loop whereby mast cell–derived chymase stimulates dorsal root ganglion cells to release additional SP, perpetuating neurogenic inflammation and vascular hyperpermeability.⁴⁹ Moreover, mast cell mediators including histamine, tryptase, chymase, and proinflammatory cytokines can sensitize and stimulate neighboring nerves, contributing to mechanical and thermal hyperalgesia.⁵⁰ Similarly, eosinophils localize to enteric nerves in inflammatory bowel disease through neurally expressed ICAM-1 and eotaxin-3, with eosinophil numbers correlating with disease activity.⁵¹ Eosinophil inflammatory mediators impair the intestinal barrier and stimulate and damage enteric nerve fibers, inducing visceral hypersensitivity and motility disturbances in the setting of functional GI disorders.⁵² In functional dyspepsia and eosinophilic colitis, increased numbers of eosinophils and mast cells near submucosal plexus neurons are associated with decreased neuronal responsiveness, fine nerve fiber sprouting, and perineural mast cell degranulation with IgE and tryptase colocalizing on enteric neurons.^53,54 This reciprocal neuroimmune communication sustains inflammation and likely contributes to the pathophysiology of visceral hypersensitivity, pain, and motility disturbances observed in gastrointestinal disorders.^11,52

Existing research has relied on traditional thin sectioning for 2D imaging of ENS anatomy and neuroimmune interactions; however, this approach is prone to sample collection errors and provides a limited view of ENS spatial distribution.^27,33,34,55 Whole-mount sampling for 3D visualization overcomes these limitations by capturing large-scale and thorough microanatomy and organization. Effective and diffuse immunolabeling can be difficult when using larger tissue samples due to antibody penetration challenges.^32,56 Antibody penetration can take several days to uniformly enter deeper tissue, requiring optimization for specific markers and fluorophores.^32,57 To improve antibody penetration, we implemented vacuum-microwave–assisted immunostaining.^37,38 The microwave's electromagnetic waves accelerate antigen-antibody binding kinetics, while the vacuum component enhances penetration into tissue, ultimately increasing rates of antibody diffusion.⁵⁸ Implementation of vacuum-microwave–assisted immunostaining shortened the staining duration and processing time from 28 d to 10 d from sample collection to image acquisition. Tissue opacity has also been a common challenge due to the varying optical properties of the tissue, resulting in poor-quality images.⁵⁹ Clearing methods have been implemented to enable high-resolution imaging of large tissue samples including whole murine models, allowing for the visualization of enteric neurons. Utilization of Ce3D provided rapid tissue clearing within 24 h, eliminating the need for an electrophoretic chamber,⁶⁰ and is an affordable, commercially available option for adoption.^28,32 These methods allowed for deep imaging of biopsy samples, allowing visualization of distinct neuron filaments and distribution patterns throughout the gastric mucosa.

Despite advances in whole-mount tissue clearing and fluorescence microscopy approaches such as iDISCO^61,62 and CLARITY,⁶³ and even methods like TESOS⁶⁴ and enGLOW³⁶ for mapping neural networks, whole-mount 3D imaging of the ENS can be challenging in the setting of low expression of neuropeptide signals.^65–67 To accurately reconstruct sensory ENS networks, we employed a computationally integrated imaging workflow combining deconvolution, ML-based signal classification, and voxel-to-structure modeling. ML-based classification played a central role in improving the accuracy of mucosal ENS reconstruction by generating probability-based structural maps that distinguished true neuronal signal from background, epithelial, and immune cell fluorescence, while enhancing detection of fine sensory branches and continuity of low-intensity fibers.⁶⁸ When combined with Huygens-based deconvolution and Autopath tracing, this ML-guided approach reduced threshold-dependent variability by maintaining filament tracing across weak or spatially fragmented regions for more reliable structural representations of neuronal networks compared with traditional model generation.^69,70 By integrating high-fidelity signal correction, structural reconstruction, and quantitative spatial analysis, our method overcomes the limitations of conventional imaging and enables reproducible, high-resolution modeling of mucosal neurons and eosinophil and mast cell interactions. Previous studies of neuroimmune proximity in tissue have relied on visual colocalization or simple distance measurements without formal comparison with a null expectation, carrying the limitation that immune cells and nerves can appear close in tissue simply because both concentrate in the same mucosal compartment.^34,51 Using a permutation approach by generating null cell populations within each volume and comparing observed distances to the resulting distribution can helps avoid distributional assumptions and is appropriate for the small per-patient sample sizes inherent to biopsy-based studies.

Samples for 3D gastric neuron imaging have primarily relied on animal models^71,72 or human samples obtained through surgical gastric resection (i.e. gastrectomy).⁷³ In contrast, our study demonstrates that endoscopic biopsy samples present several advantages for examining gastric mucosal neuron circuits, in which samples are more readily accessible and collection is minimally invasive.³²

This study has several limitations that should be acknowledged. This protocol was developed and optimized for gastric antrum and esophageal samples; thus, adoption of this technique for other tissue types may require additional modifications. Microscopy of large tissue samples may require 4 to 12 h of uninterrupted imaging acquisition and may not be feasible for all situations; thus, imaging stacks or partial samples should be considered to reduce imaging time and better support standardized data analysis. In terms of clinical samples, endoscopic biopsies often vary in size and penetration depth, which may impair visualization of certain structures. Routine endoscopic biopsies may not adequately sample the subepithelium,^67,74,75 preventing assessment of esophageal mucosal nerves confined to the lamina propria. The spatial statistics framework has its own constraints. The convex hull null model assumes that cells could occupy any position within the tissue volume with equal probability; it does not account for architectural features such as glands or vessels that may restrict cell migration. Both nerves and immune cells are distributed near the epithelial surface, so some fraction of the observed proximity may reflect shared depth preference rather than direct spatial association. Finally, the focus of this report was to describe the imaging technique, and samples used were based on clinical status and pathologic features necessary to confirm the technique's ability to visualize key markers of the disease. Therefore, statistics presented here are not meant to be representative of gastric and esophageal nerve volume.

In conclusion, we demonstrate a method for investigating neuroimmune relationships in gastric and esophageal biopsies. This technique enables tissue clearing with Ce3D and efficient immunolabeling utilizing vacuum-microwave–assisted staining, providing high-quality 3D imaging in whole-mount GI biopsy samples. We established a spatial modeling workflow best suited to analyze mucosal neuron architecture and eosinophil and mast cell distributions within the stomach and esophagus. Quantitative assessment can be applied to compare tissue innervation as well as the interactions between eosinophils, mast cells, and neurons in EGID. Additionally, this methodology can be extended and optimized for mucosal biopsies from other regions of the GI tract and to examine cells that may play a role in other diseases with suspected neuroimmune pathology.

Author contributions

A.M.L. (Investigation, Methodology, Writing—original draft [lead]), M.K. (Investigation, Methodology [equal], Writing—review & editing [supporting]), J.H. (Formal analysis [lead], Methodology, Writing—review & editing [supporting], Visualization [equal]), J.K. (Formal analysis, Writing—review & editing [supporting], Methodology, Software [lead], Visualization [equal]), M.S. (Visualization [supporting]), A.A.-M. (Methodology, Writing—review & editing [supporting]), M.B.K. (Resources, Writing—review & editing [supporting]), D.S. (Resources, Writing—review & editing [supporting]), A.T. (Resources [supporting]), S.K. (Resources, Supervision [supporting]), P.K. (Conceptualization, Writing—review & editing [supporting]), A.D.K. (Conceptualization, Investigation, Supervision, Writing—review & editing [supporting], Funding acquisition, Resources [lead]), and G.M.C. (Conceptualization, Methodology, Supervision [lead], Data curation, Formal analysis, Investigation [supporting], Writing—original draft, Writing—review & editing [equal])

Supplementary material

Supplementary material is available at Journal of Leukocyte Biology online.

Funding

This research was supported by the Intramural Research Program of the National Institutes of Health. The contributions of the National Institutes of Health author(s) were made as part of their official duties as National Institutes of Health federal employees, are in compliance with agency policy requirements, and are considered works of the U.S. government. However, the findings and conclusions presented in this paper are those of the author(s) and do not necessarily reflect the views of the National Institutes of Health or the U.S. Department of Health and Human Services.

Data availability

Data will be made available upon reasonable request.