Cellular and molecular profiling of collagenous gastritis implicates pathogenic CD4+ T cells.

M.D. performed EGDs and collected patient samples. J.D., A.E-I., C.E.K., and K.M.S. oversaw sample storage and recruited patients to the study. M.J.W., L.A., J.S., and L.Q. processed samples for flow cytometry and scRNA-seq. M.J.W., L.A., and J.S. performed scRNAseq library preparation and flow cytometry analysis. M.J.W. performed scRNA-seq analysis. E.C. provided independent confirmation of computational data analysis. Y.K. and M.G. performed histopathologic evaluation and analysis with the supervision of M.M-K. M.J.W., L.A., Y.K., E.C., M.T., M.M-K., S.K.D., and M.D. contributed to data interpretation. S.K.D. and M.D. supervised the experiments and analysis. M.J.W., M.M-K., S.K.D., and M.D. wrote the article with input from all authors.

To characterize collagenous gastritis at cellular and molecular resolution, we identified and recruited seven patients with collagenous gastritis ( Table 1 ). Patients underwent esophagogastroduodenoscopies in which biopsies from the stomach body and second portion of the duodenum were collected for downstream processing; for histopathology only, antral biopsies were also collected. Tissue biopsies were analyzed using three methods: histopathology, flow cytometry, and scRNA-seq ( Figure 1 A). Control stomach biopsies from patients undergoing upper endoscopies for non-gastric indications ( Table S1 ) were also collected and processed similarly. These control samples had no visible inflammation and normal stomach histopathology compared to their CG counterparts. Patients with CG were also profiled for the presence of Helicobacter pylori , and some patients were tested for anti-parietal cell, anti-enterocyte, or anti-intrinsic factor auto-antibodies. Results of these tests performed on patients with CG were negative ( Table S2 ). Table 1 CG patient characteristics Patient A B C D E F G Age 26 67 14 37 33 16 40 Sex F M F F M F F EGD findings Gastric erythema/hypertrophy; normal duodenum Gastric erythema/hypertrophy; normal duodenum Gastric erythema; normal duodenum Nodular gastric mucosa; microscopic colitis; normal duodenum Focal surface erosion; normal duodenum Focally active CG; normal duodenum Diffuse gastric glandular hypertrophy; normal duodenum Iron deficiency anemia (IDA) IDA on IV iron; not deficient – IDA; deficient IDA on oral iron; not deficient IDA on IV iron; not deficient IDA on oral iron; not deficient IDA; deficient Nausea/Vomiting – – + + + – + Weight loss – – + – – – + Fatigue – – + + + + + Diarrhea + – + + – – + Decreased Appetite + – + – – – – Abdominal pain + + + + + – + Abdominal bloating + + – – – – + Heartburn – + – – – + – Joint pain + – + – + – – PPI or NSAID at baseline? None Naproxen None None None None None Treatments related to CG Prednisone (post 1st biopsy), iron supplements/infusions, diet adjustments None SNRIs, Zofran, trazodone, and iron infusions Diet adjustments (dairy & gluten-free), budesonide for colitis fluvoxamine, iron infusions Vitamin D, diet adjustments (dairy free), iron supplements, mesalamine Iron, B12 supplements, diet adjustments, MiraLAX, Gavison, PeptoBismol Figure 1 Single cell analysis of patients with collagenous gastritis (A) Workflow for the collection of biopsies from the duodenum and stomach of control and patients with CG. Black and gray arrows indicate which tissues (stomach and duodenum, respectively) were processed by which methods. X indicates the approximate location of all biopsies. (B) Controls (H1-6) and CG biopsies were processed as fresh, unsorted biopsies for gene expression (GEX) libraries; frozen biopsies were CD45-sorted and processed for GEX and T cell receptor (TCR) libraries. CG patient “A” had a pre- and post-steroid biopsy collected for CD45-sorting (“A”; “Apost”), and “D” had a second biopsy (“D2”). (C) UMAP of all patients/samples, both unsorted and CD45-sorted. (D) Violin plot of canonical cluster-defining gene expression across clusters from (C). (E) Cluster frequency per sample. See also Figures S1 and S2 . CG patient characteristics Single cell analysis of patients with collagenous gastritis (A) Workflow for the collection of biopsies from the duodenum and stomach of control and patients with CG. Black and gray arrows indicate which tissues (stomach and duodenum, respectively) were processed by which methods. X indicates the approximate location of all biopsies. (B) Controls (H1-6) and CG biopsies were processed as fresh, unsorted biopsies for gene expression (GEX) libraries; frozen biopsies were CD45-sorted and processed for GEX and T cell receptor (TCR) libraries. CG patient “A” had a pre- and post-steroid biopsy collected for CD45-sorting (“A”; “Apost”), and “D” had a second biopsy (“D2”). (C) UMAP of all patients/samples, both unsorted and CD45-sorted. (D) Violin plot of canonical cluster-defining gene expression across clusters from (C). (E) Cluster frequency per sample. See also Figures S1 and S2 . For the scRNA-seq cohort, samples were either processed fresh immediately following biopsy by manual and enzymatic dissociation into single cell suspensions or were frozen and later processed similarly, but also sorted for CD45 + immune cells by flow assisted cell sorting. Gene expression (GEX) libraries were generated for fresh, unsorted tissue samples, while both GEX and T cell receptor (TCR) libraries were created for CD45-sorted samples ( Figure 1 B). One CG patient, patient “A”, also had a biopsy performed following a course of systemic glucocorticoids, while another patient, “D”, had a second biopsy performed several months after the first as part of clinical follow-up. Initially, the fresh, unsorted samples were combined into one dataset, while frozen, CD45-sorted samples were combined into a separate dataset ( Figures S1 A and S1B). The clusters in both populations were quite similar, apart from large increases in non-immune, gastric cells present in the unsorted samples. Some of these cells still appeared in smaller amounts in the CD45-sorted samples ( Figures S1 and S1B). Next, the unsorted and CD45-sorted samples were combined in one integrated dataset containing 76,738 quality cells from both control and patients with CG ( Figures 1 C, 1D, S2 A, and S2B). These clusters included immune cells such as T lymphocytes, IgA positive plasma B cells, mast cells, and macrophages, all of which were enriched within the CD45-sorted samples ( Figures 1 C–1E). Non-immune cells comprised gastric chief cells, mucus-secreting cells, parietal cells, and fibroblasts: most of the expected major cell populations of the gastric mucosa were successfully captured through sequencing ( Figures 1 C–1E). Comparing fresh versus frozen sample gene expression across CD45 + cell clusters revealed a correlation between each gene as a whole and within each sample ( Figures S1 C and S1D). Additionally, across samples, patients separated by disease status in a PCA plot when comparing either CD45 − or CD45 + clusters, suggesting there were overall differences in gene expression between control and CG ( Figure S1 E). Comparison of control and CG samples revealed certain populations were greatly reduced in patients with disease ( Figure 2 A). For example, the percentage of fibroblasts in CG patient samples was more than 2-fold decreased compared to controls ( Figure 2 B), though this could have been due to the highly fibrotic tissue making fibroblast recovery more difficult. The fibroblasts that were present in patients with CG expressed significantly higher levels of genes encoding collagens, including COL6A3 , which was one of the highest upregulated genes when comparing the two groups, expressed at almost 6-fold higher levels ( Figure 2 C). Though fibroblasts are known to be the typical producers of collagen, these data show that on a transcriptional level, collagen is more highly expressed in this population within patients with CG. Figure 2 Excessive collagen expression and disruption of the gastric mucosa in patients with CG (A) Cells captured from scRNA-seq split by the disease status of the patient. (B) Percent of fibroblasts captured by scRNA-seq in unsorted samples. Significance assessed by the Wilcoxon rank test. (C) Differentially expressed genes encoding collagens, displayed as violin plots of normalized gene expression. Adjusted p -value and log2 fold-change (LFC) displayed. (D) Percent of gastric mucosal cells captured by scRNA-seq in unsorted samples. Significance assessed by the Wilcoxon rank test. N.S. notsignificant. (E) Differentially expressed genes for anemia-related genes, displayed as violin plots of normalized gene expression. Adjusted p -value and log2 fold-change (LFC) displayed. LCN2 = lipocalin 2 (iron chelator), LTF = lactoferrin (iron transporter), TCN1 = transcobolamin-1 (B12 binder/transporter). (F) Differentially expressed MHC class I genes. (G) Representative H&E images of gastric gland mucosa. Black arrows indicate atrophic glands in CG. Yellow arrows indicate characteristic collagen deposits. Microscope objective magnification shown. (H) Quantification across control and CG biopsies for chief, mucus-producing, and parietal cells from H&E slides (F). Total cells per 5 high power fields (HPFs). (I) Quantifications of eosinophils from H&E slides. Data are mean ± SEM; gene expression and histological quantifications were compared with a Wilcoxon test followed by a Benjamini-Hochberg correction. ∗ p < 0.05. See also Figures S3–S5 . Excessive collagen expression and disruption of the gastric mucosa in patients with CG (A) Cells captured from scRNA-seq split by the disease status of the patient. (B) Percent of fibroblasts captured by scRNA-seq in unsorted samples. Significance assessed by the Wilcoxon rank test. (C) Differentially expressed genes encoding collagens, displayed as violin plots of normalized gene expression. Adjusted p -value and log2 fold-change (LFC) displayed. (D) Percent of gastric mucosal cells captured by scRNA-seq in unsorted samples. Significance assessed by the Wilcoxon rank test. N.S. notsignificant. (E) Differentially expressed genes for anemia-related genes, displayed as violin plots of normalized gene expression. Adjusted p -value and log2 fold-change (LFC) displayed. LCN2 = lipocalin 2 (iron chelator), LTF = lactoferrin (iron transporter), TCN1 = transcobolamin-1 (B12 binder/transporter). (F) Differentially expressed MHC class I genes. (G) Representative H&E images of gastric gland mucosa. Black arrows indicate atrophic glands in CG. Yellow arrows indicate characteristic collagen deposits. Microscope objective magnification shown. (H) Quantification across control and CG biopsies for chief, mucus-producing, and parietal cells from H&E slides (F). Total cells per 5 high power fields (HPFs). (I) Quantifications of eosinophils from H&E slides. Data are mean ± SEM; gene expression and histological quantifications were compared with a Wilcoxon test followed by a Benjamini-Hochberg correction. ∗ p < 0.05. See also Figures S3–S5 . In addition, in analyzing other cell types that were not highly captured in patients with CG, we discovered that gastric chief cells were relatively reduced, while mucus-producing foveolar cells and MUC6+ mucus-cells were near control levels ( Figures 2 A and 2D). Surprisingly, parietal cells were all but missing in patients with CG ( Figures 2 A and 2D), and ghrelin-expressing P/D1 cells were significantly reduced, potentially contributing to the decreased appetite frequently seen in these patients ( Figure S3 A). Tuft cells were increased in collagenous gastritis samples ( Figure S3 A). Glandular atrophy can occur in CG, 5 , 11 and our data suggest parietal cells and P/D1 are most affected, while other glandular cell types, such as mucus-producing cells or neuroendocrine enterochromaffin-like (ECL) cells, are relatively preserved in CG ( Figures 2 D and S3 A). Differential gene expression analysis on the four major glandular cell populations (chief, foveolar, mucus, parietal) revealed that some of the top upregulated genes in patients with CG included LCN2 , LTF , and TCN1 ( Figure 2 E). LCN2 encodes lipocalin-2, which sequesters siderophores and is involved in innate immunity to prevent iron acquisition by bacteria. However, it has also been shown to be involved in a transferrin-independent iron uptake pathway, and is upregulated during anemia. 16 LTF encodes lactoferrin, an iron-binding protein also involved in innate immunity as well as iron transport, and has been used to successfully treat patients with anemia. 17 TCN1 encodes transcobalamin-I or haptocorrin, a protein involved in the protection of vitamin B 12 from acidic degradation in the stomach. Vitamin B 12 is bound to haptocorrin until proteolytic cleavage in the duodenum, at which point B 12 associates with intrinsic factor, which is secreted by parietal cells. 18 Although these patients with CG did not exhibit B 12 deficiency, they had very few parietal cells, as found by scRNA-seq. In addition to these genes, we also saw the upregulation of all major histocompatibility class (MHC)-I genes, HLA-A, HLA-B, and HLA-C within these cell types lining the gastric glands ( Figure 2 F). To confirm the glandular atrophy and diminished parietal cell counts, we compared CG to control H&E-stained sections. A gastrin immunostain was conducted to confirm the tissue of origin as gastric body mucosa, in which parietal cells should normally exist, rather than pyloric gland mucosa ( Figure S4 ). Controls had typical gastric body architecture, while patients with CG had widespread glandular atrophy ( Figure 2 G, black arrows). Large subepithelial deposits of collagen, characteristic of the disease, were also seen ( Figure 2 G, yellow arrows). Quantification of chief, mucus, and parietal cells in tissue sections confirmed that while chief and mucus cells were seen in normal quantities, parietal cells were significantly reduced in CG tissue ( Figure 2 H). An increase in eosinophils has previously been seen in the gastric tissue of patients with CG with glandular atrophy. 7 Compared to the control H&E-stained sections, eosinophils were seen at significantly higher levels in the stroma, but not in the intraepithelial region ( Figures 2 I and S5 ). To accurately quantify the immune cell frequencies in CG, we profiled both the stomach and duodenum by flow cytometry of fresh samples ( Figure S6 ). Additional controls not analyzed by scRNA-seq were also included ( Table S3 ). As these patients with CG did not exhibit duodenal disease, the duodenum served as an additional reference for uninflamed tissue. As expected, CD4 + and CD8 + T cell infiltration in both control and patients with CG was equivalent in the duodenum ( Figures 3 A and 3B, top). In the stomach, however, CD8 + T cells were significantly reduced in patients with CG, whereas CD4 + T cells were found at much higher frequencies compared to controls ( Figures 3 A and 3B, bottom). Thus, the CD4:CD8 ratio was inverted in patients with CG ( Figure 3 C). As expected by the lack of duodenal inflammation, activated PD-1 + CD4 + and CD8 + T cells were virtually absent in the duodenums of both control and patients with CG ( Figures 3 D and 3E). We identified significantly more PD-1 + CD4 + and PD-1 + CD8 + T cells in the stomachs of patients with CG, suggesting ongoing antigen recognition by these T cells ( Figures 3 F and 3G). Figure 3 Patients with Collagenous gastritis have an influx of highly activated T cells (A) CD4 + and CD8 + T cells were quantified by flow cytometry of the duodenum (top row) or stomach (bottom row). (B) Representative flow plots from (A). (C) CD4:CD8 ratio from (A). (D and E) PD-1 expression on duodenal CD8 + (D) or CD4 + T cells (E). (F and G) PD-1 expression on gastric CD8 + (F) or CD4 + T cells (G). (H) Representative CD8 and CD4 stains from tissue sections. All images at 20x objective. Scale bars 500 μm. (I) Quantification of (H) for intraepithelial (top row) or intraepithelial and stromal lymphocytes (bottom) from slides. Total cells per five high power fields (HPFs). (J) Quantification of PD-L1 and PD-1 stained tissue sections. (K) CD4: CD8 ratio from T cells captured by scRNA-seq (see Figure 1 D) in CD45-sorted samples. Significance assessed by the Wilcoxon rank test. (L) Heatmap of differential gene expression comparing control patients to each CG patient sample for CD8 and CD4 scRNA-seq clusters from CD45+ sorted cells. Genes involved in T cell activation/exhaustion are shown. Dot indicates significantly different expression between a patient sample and all control controls. Data are mean ± SEM; flow percentages were compared with a Student’s t test (flow) or Wilcoxon test (histology) followed by a Benjamini-Hochberg correction. Gene expression data compared with a Benjamini-Hochberg-corrected Wilcoxon rank-sum test. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 See also Figures S6 and S7 . Patients with Collagenous gastritis have an influx of highly activated T cells (A) CD4 + and CD8 + T cells were quantified by flow cytometry of the duodenum (top row) or stomach (bottom row). (B) Representative flow plots from (A). (C) CD4:CD8 ratio from (A). (D and E) PD-1 expression on duodenal CD8 + (D) or CD4 + T cells (E). (F and G) PD-1 expression on gastric CD8 + (F) or CD4 + T cells (G). (H) Representative CD8 and CD4 stains from tissue sections. All images at 20x objective. Scale bars 500 μm. (I) Quantification of (H) for intraepithelial (top row) or intraepithelial and stromal lymphocytes (bottom) from slides. Total cells per five high power fields (HPFs). (J) Quantification of PD-L1 and PD-1 stained tissue sections. (K) CD4: CD8 ratio from T cells captured by scRNA-seq (see Figure 1 D) in CD45-sorted samples. Significance assessed by the Wilcoxon rank test. (L) Heatmap of differential gene expression comparing control patients to each CG patient sample for CD8 and CD4 scRNA-seq clusters from CD45+ sorted cells. Genes involved in T cell activation/exhaustion are shown. Dot indicates significantly different expression between a patient sample and all control controls. Data are mean ± SEM; flow percentages were compared with a Student’s t test (flow) or Wilcoxon test (histology) followed by a Benjamini-Hochberg correction. Gene expression data compared with a Benjamini-Hochberg-corrected Wilcoxon rank-sum test. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 See also Figures S6 and S7 . Immunohistochemistry staining of tissue sections for CD4 or CD8 ( Figure 3 H) revealed increased CD4 + and CD8 + T cells in patients with CG, including in the intraepithelial region, which also confirmed a skewed CD4:CD8 ratio as we had observed by flow cytometry ( Figure 3 I, top). This skewed CD4:CD8 ratio was confined to the intraepithelial region, although increased CD4 + and CD8 + T cells were observed throughout the tissue sections of patients with CG ( Figure 3 I, bottom). PD-L1 and PD-1 staining of tissue sections demonstrated a significant increase in PD-L1 + lymphocytes, macrophages, and epithelial cells from CG tissue ( Figure 3 J). PD-1 + lymphocytes were significantly increased within CG tissue sections ( Figure 3 J). Cell type specificity of PD-1/L1 staining was confirmed with dual staining ( Figure S7 ). From our scRNA-seq dataset, we also demonstrated a skewed CD4:CD8 ratio in favor of CD4 + cells in patients with CG ( Figure 3 K). Performing differential gene expression analysis across each CD45-sorted patient sample versus controls, we found that some of the highest upregulated genes in the CD4 + and CD8 + T cells were those associated with T cell activation and exhaustion, including PDCD1 (PD-1), CTLA4 , TIGIT , HAVCR2 (TIM-3), LAG3, TNFRSF4 (OX40), TNFRSF9 (4-1BB), TNFRSF18 (GITR), and ENTPD1 (CD39), 19 which were shared across patients with CG ( Figure 3 L; Table S4 ). We found significant GZMK and KLRG1 expression within the CD8 + T cells as well. These results suggest that highly activated or, less likely, dysfunctional T cells are abundant in the gastric tissue of patients with CG. In addition to the altered ratio of CD4:CD8 cells, we also saw significantly reduced γδ T cells, NK cells, and increased cycling T cells, IgA+ plasma cells, B cells, and macrophages by scRNA-seq in collagenous gastritis samples ( Figure S3 B). Although the quantities of these cell types differed, we saw fewer relevant gene changes compared to CD4 + and CD8 + T cells ( Table S4 ). Based on the flow cytometry and scRNA-seq data demonstrating an increased presence of activated/exhausted T cells, we wished to explore T cells in more depth from the scRNA-seq dataset. We therefore sub-clustered T cells to further delineate heterogeneity ( Figures 4 A and 4B). These sub-clusters included a tissue-resident (T RM ) CD8 + subset with ITGAE expression, an IFNG CD8 + subset also expressing TNF , a cytotoxic T lymphocyte (CTL) population expressing high levels of various granzymes including GZMK, GZMB, and GZMA , two CD4 + populations with high expression of either LTB (lymphotoxin-β) or CTLA4 , as well as a cycling population ( Figure 4 B). Patients with CG had a significant relative reduction in the T RM CD8 + and IFNG + CD8 + subsets ( Figures 4 C–4E). Concomitant with these decreased CD8 + populations, there was a slight increase in the granzyme+ CTL sub-cluster and significantly increased CTLA4 + CD4 + and cycling T cells ( Figures 4 C–4E). Figure 4 Resident memory CD8 T cells and expanded T cell clonotypes are reduced in patients with CG (A) T cells were sub-clustered from all patients/samples. T RM = Tissue resident memory; T EM = Effector memory; CTL = cytotoxic T lymphocyte. (B) Dot plot of cluster-defining genes for T cell sub-clusters. (C) Sub-cluster frequency per sample. (D) Distribution of T cells split by patient disease status. (E) Barplot of sub-clusters grouped by disease status of the patient. Mean ± SEM; adjusted p -value from BH-adjusted Wilcoxon rank test shown for p-adj <0.1. (F) UMAP from (A) of TCR expansion across all CD45-sorted samples; cells are colored by the TCR frequency. (G) Barplot of unique TCR clonotypes; expanded clonotypes are those with 2+ cells per unique TCR. Percentages above bars indicate the percentage of expanded clonotypes out of the total number of unique clonotypes. (H) Percentage of expanded T cells (2+ cells/clonotype) out of all T cells with TCRs sequenced. (I) Percentage of unique, expanded clonotypes out of all unique clonotypes; similar to (G), but grouped by disease status. Error bars are ±SEM; percentages were compared with a Wilcoxon rank test. See also Figure S8 . Resident memory CD8 T cells and expanded T cell clonotypes are reduced in patients with CG (A) T cells were sub-clustered from all patients/samples. T RM = Tissue resident memory; T EM = Effector memory; CTL = cytotoxic T lymphocyte. (B) Dot plot of cluster-defining genes for T cell sub-clusters. (C) Sub-cluster frequency per sample. (D) Distribution of T cells split by patient disease status. (E) Barplot of sub-clusters grouped by disease status of the patient. Mean ± SEM; adjusted p -value from BH-adjusted Wilcoxon rank test shown for p-adj <0.1. (F) UMAP from (A) of TCR expansion across all CD45-sorted samples; cells are colored by the TCR frequency. (G) Barplot of unique TCR clonotypes; expanded clonotypes are those with 2+ cells per unique TCR. Percentages above bars indicate the percentage of expanded clonotypes out of the total number of unique clonotypes. (H) Percentage of expanded T cells (2+ cells/clonotype) out of all T cells with TCRs sequenced. (I) Percentage of unique, expanded clonotypes out of all unique clonotypes; similar to (G), but grouped by disease status. Error bars are ±SEM; percentages were compared with a Wilcoxon rank test. See also Figure S8 . From the CD45-sorted samples, we also performed T cell receptor (TCR) clonotype tracking, which can be used to infer clonal relationships over time. 20 , 21 , 22 , 23 We identified 5,305 unique TCR clonotypes based on cells expressing distinct TCRα and TCRβ chains ( Figure 4 F). The GI tract is typically characterized by the expansion of a few resident memory T cell clones, while inflammatory diseases show the replacement of resident memory cells with a diverse set of both T cell subtypes and clonotypes, resulting in fewer individual clones being highly expanded. 23 Across patients with CG, clones had lower levels of expansion (<64+ cells/clonotype) compared to several highly expanded clones from controls ( Figures S8 A and S8B). Not only were some clones less expanded in patients with CG ( Figure S8 B), accounting for significantly less total T cell expansion ( Figure 4 H), but also patients with CG had significantly fewer unique clonotypes that were expanded compared to controls ( Figures 4 G, 4I, and S8 B). Because many more T cells were sequenced from controls, we also randomly downsampled across patients to simulate more equivalent numbers of total T cells captured in control vs. CG samples. With random downsampling, we still found a significant decrease in the proportion of expanded T cells as well as fewer unique, expanded clonotypes in patients with CG ( Figure S8 C). While controls predominantly had T RM and IFNG + CD8 + expansion, similar to their original levels of frequency, patients with CG in general had more granzyme+ CTL and CD4 + expansion ( Figure S8 D). We also examined whether expanded TCR clonotypes were shared between T cell subclusters. Similar to the levels of frequency, the predominant sharing of clonotypes occurred between T RM and IFNG+ CD8 + clusters, which was particularly enriched in control patients, even with random downsampling ( Figures S8 E and S8F). Based on these findings and that CD4 + T cells were present in greater proportions within CG patient biopsies by flow cytometry ( Figure 3 A) and from T cell sub-clustering ( Figure 4 E), we further sub-clustered CD4 + T cells from the total T cell sub-clustering ( Figure 5 A). We identified nine sub-clusters of CD4 + T cells, including a naive population with TCF7 and SELL (CD62L) expression, an effector memory (T EM ) population expressing CCL5 , Th17 cells with IL17A and IL23R , a cytotoxic T CD4 + subset expressing various granzymes, Th1 cells expressing EOMES and IFNG , and Th2 cells with IL13 and IL5 ( Figure 5 B). There were two different Treg populations expressing FOXP3 , with varying levels of their canonical genes; the “activated” Treg population had significantly elevated levels of expression of these genes ( Figure 5 B). The distribution of CD4 + subsets was significantly different between controls and those with CG ( Figures 5 C–5E). Like the total T cell sub-clustering, the memory subset of CD4 + T cells was significantly reduced in patients with CG compared to control samples ( Figure 5 E). Additionally, naive CD4 + and Th17 cells were significantly reduced in patients with CG compared to controls ( Figure 5 E). Conversely, cytotoxic CD4 + T cells, both Th1 and Th2, Tregs, and cycling CD4 + T cells were significantly increased in patients with CG ( Figure 5 E). The only population not different was the activated Treg subset ( Figure 5 E). Though Th1 predominated in the CG mucosa, both Th1 and Th2 were significantly increased, indicating that the CD4 + T cells are not uniformly polarized in one direction in this disease. Figure 5 Memory, Th17, and naive CD4 T cells are reduced, while cytotoxic CD4, Th1, Th2, and Treg populations are increased in patients with CG (A) CD4 T cells were sub-clustered from all patients/samples. (B) Dot plot of cluster-defining genes for CD4 T cell sub-clusters. (C) CD4 sub-cluster frequency per sample. (D) Distribution of T cells, colored by patient disease status. (E) Barplot of CD4 sub-clusters grouped by disease status of patient. Mean ± SEM; adjusted p -value from BH-adjusted Wilcoxon test shown for p-adj <0.1. (F) UMAP from (A) of CD4 clonal expansion across all CD45-sorted samples; cells are colored by the TCR clone frequency. (G) Barplot of unique CD4 TCR clonotypes; expanded clonotypes are those with 2+ cells per unique TCR. Percentages above bars indicate the percentage of expanded clonotypes out of the total number of unique clonotypes. “Patient H5” had fewer than 20 TCRs sequenced and was omitted from the plot. (H) Percentage of expanded T cells (2+ cells/clonotype) out of all T cells with TCRs sequenced (left); percentage of unique, expanded clonotypes out of all unique clonotypes, similar to (G), but grouped by disease status. Error bars are ±SEM; percentages were compared with a Wilcoxon rank test. N.S. not significant. See also Figure S9 . Memory, Th17, and naive CD4 T cells are reduced, while cytotoxic CD4, Th1, Th2, and Treg populations are increased in patients with CG (A) CD4 T cells were sub-clustered from all patients/samples. (B) Dot plot of cluster-defining genes for CD4 T cell sub-clusters. (C) CD4 sub-cluster frequency per sample. (D) Distribution of T cells, colored by patient disease status. (E) Barplot of CD4 sub-clusters grouped by disease status of patient. Mean ± SEM; adjusted p -value from BH-adjusted Wilcoxon test shown for p-adj <0.1. (F) UMAP from (A) of CD4 clonal expansion across all CD45-sorted samples; cells are colored by the TCR clone frequency. (G) Barplot of unique CD4 TCR clonotypes; expanded clonotypes are those with 2+ cells per unique TCR. Percentages above bars indicate the percentage of expanded clonotypes out of the total number of unique clonotypes. “Patient H5” had fewer than 20 TCRs sequenced and was omitted from the plot. (H) Percentage of expanded T cells (2+ cells/clonotype) out of all T cells with TCRs sequenced (left); percentage of unique, expanded clonotypes out of all unique clonotypes, similar to (G), but grouped by disease status. Error bars are ±SEM; percentages were compared with a Wilcoxon rank test. N.S. not significant. See also Figure S9 . From the total TCR clonotype analysis ( Figure 4 F), the CD4 + subset appeared to be mostly singletons (clonotype size = 1), though we wished to know which CD4 + sub-clusters were expanded. Therefore, we used the new CD4 + sub-clustering to map TCR expansion ( Figure 5 F). As seen with the total TCR analysis, most of the CD4 + cells were singletons ( Figure 5 F). However, some expansion was occurring in the Th1 and Th17/T EM sub-clusters ( Figure 5 F). Unlike the total TCR analysis, the CD4 + TCRs had more equivalent levels of expansion between CG and controls ( Figures 5 G, 5H, and S9 A–S9C). Th17/T EM CD4 + T cells accounted for most of the control sample expansion, while primarily Th1 cells were found to be expanded in patients with CG ( Figure S9 C). As with the total T cell compartment, we assessed the sharing of expanded TCR clonotypes between CD4 subclusters. In control patients, the predominant sharing of clonotypes across clusters occurred between T EM and Th17/T EM ( Figure S9 D). Patients with CG had a distinct profile of TCR clonotypes that overlapped between Th1 and cytotoxic CD4s ( Figure S9 D). Our findings indicate that Th1 (or perhaps Th2) cells could be inducing pathological damage followed by subsequent Treg recruitment. These data are consistent with flow cytometry showing increased PD-1 + CD4 + T cells in the stomachs of patients with CG ( Figure 3 ). As CG is an inflammatory disease, corticosteroids have been used to treat patients, with mostly limited effect. 12 Nevertheless, one CG patient, “A”, was given systemic glucocorticoids with the intention of limiting inflammation and resolving disease. Although glucocorticoids are commonly known to act on T cells, they also act on other cells, including myeloid cells. 24 When we examined changes in gene expression between all cells from CD45-sorted samples of patient “A” by Gene Set Enrichment analysis, we found major reductions in pathways involved in cell proliferation, metabolic activity, and activation following glucocorticoid administration ( Figure 6 A). These pathways reduced by steroid treatment included Myc targets, oxidative phosphorylation, interferon gamma response, and NF-κB signaling ( Figure 6 A). Figure 6 Effects of steroid treatment on pre-existing TCR clonotypes and possible future therapies for CG (A) Hallmark Gene Set Enrichment Analysis of all cells from CD45-sorted patient “A”, comparing pre- and post-steroid gene expressions. (B) TCR clones sequenced from patient “A” show the distribution of clonotypes that were shared between time points (pre- and post-steroids) and expanded (2+ cells); expanded but not shared; shared but not expanded; or neither shared nor expanded (not). Percentage of total clone pool for each timepoint shown. (C) Alluvial plot from patient “A” showing unique TCR clones that were expanded (2+ cells) and shared between the two time points (pre- and post-steroids). Flow between bars indicates which sub-cluster the TCR clones were shared between. (D) The top four expanded TCR clones shared between pre- and post-steroids for patient “A” and their T cell sub-clusters of origin. Y axis indicates the total number of T cells captured in each shared clonotype. TCR chains are listed below every four clones. (E) Heatmap of scaled gene expression for T cell activation genes across T cell sub-clusters. Dot indicates significantly different expression between the CG and control samples for a particular gene/cluster. (F) Same as (E), but across CD4 sub-clusters. (G) Expression of IL5 and IL13 in the Th2 subcluster. (H) Differential gene expression analysis was performed across each T cell sub-cluster. Normalized JAK3 expression shown for T cell subclusters in which p-adj ≤0.05. Adjusted p -value and log2 fold-change (LFC) displayed. (I) Same as (H) but for CD4 sub-clusters in which p-adj ≤0.05. (J) Normalized ITGA4 expression shown for T cell subclusters in which p-adj ≤0.05. Adjusted p -value and log2 fold-change (LFC) displayed. (K) Same as (J), but for CD4 sub-clusters in which p-adj ≤0.05. (L) Normalized MADCAM1 (ligand for α4 integrin) on endothelial cells. Gene expression significance assessed by the Wilcoxon rank-sum test with a Benjamini-Hochberg correction. See also Figures S10–S13 . Effects of steroid treatment on pre-existing TCR clonotypes and possible future therapies for CG (A) Hallmark Gene Set Enrichment Analysis of all cells from CD45-sorted patient “A”, comparing pre- and post-steroid gene expressions. (B) TCR clones sequenced from patient “A” show the distribution of clonotypes that were shared between time points (pre- and post-steroids) and expanded (2+ cells); expanded but not shared; shared but not expanded; or neither shared nor expanded (not). Percentage of total clone pool for each timepoint shown. (C) Alluvial plot from patient “A” showing unique TCR clones that were expanded (2+ cells) and shared between the two time points (pre- and post-steroids). Flow between bars indicates which sub-cluster the TCR clones were shared between. (D) The top four expanded TCR clones shared between pre- and post-steroids for patient “A” and their T cell sub-clusters of origin. Y axis indicates the total number of T cells captured in each shared clonotype. TCR chains are listed below every four clones. (E) Heatmap of scaled gene expression for T cell activation genes across T cell sub-clusters. Dot indicates significantly different expression between the CG and control samples for a particular gene/cluster. (F) Same as (E), but across CD4 sub-clusters. (G) Expression of IL5 and IL13 in the Th2 subcluster. (H) Differential gene expression analysis was performed across each T cell sub-cluster. Normalized JAK3 expression shown for T cell subclusters in which p-adj ≤0.05. Adjusted p -value and log2 fold-change (LFC) displayed. (I) Same as (H) but for CD4 sub-clusters in which p-adj ≤0.05. (J) Normalized ITGA4 expression shown for T cell subclusters in which p-adj ≤0.05. Adjusted p -value and log2 fold-change (LFC) displayed. (K) Same as (J), but for CD4 sub-clusters in which p-adj ≤0.05. (L) Normalized MADCAM1 (ligand for α4 integrin) on endothelial cells. Gene expression significance assessed by the Wilcoxon rank-sum test with a Benjamini-Hochberg correction. See also Figures S10–S13 . Despite these transcriptional effects, steroid treatment did not reduce symptoms in this patient or change the appearance of the lesions in the gastric lining as seen during upper endoscopy. We hypothesized that given the changes in T cells seen by flow cytometry and scRNA-seq, T cells likely had a causative, pathological role in disease. Despite capturing fewer T cells post-steroids (as measured as a percentage of the total sample) by scRNA-seq ( Figure 1 E), pre- and post-steroid T cell sub-clustering looked similar, with the biggest difference being a counterintuitive increase in cycling T cells post-steroids ( Figure 4 C). As we had captured two longitudinal biopsies from patient “A”, we explored whether any TCR clonotypes were either expanded (2+ cells) and/or shared between the time points. We discovered that 12.6% and 11.1% of clonotypes were shared between pre- and post-steroid biopsies, respectively ( Figure 6 B). Between 25 and 45% of clones that were shared were also expanded at the respective timepoints ( Figure 6 B). When we looked at how these shared and expanded clones changed over time, we observed some flux of clones moving from one sub-cluster to another, though the overall profile of expanded and shared clonotypes remained similar ( Figure 6 C). This analysis accounted for all expanded, shared clonotypes, yet some clonotypes were found at a much greater frequency than two cells per clone at each timepoint: the most expanded/shared clone had over 50 cells in each timepoint ( Figure 6 D). These results suggest that there were shared, expanded T cell clones that did not diminish or greatly change due to steroid treatment. Though T cells are one of the main targets of glucocorticoids, 24 the lack of pathological response we and others 12 have seen with their use prompted us to revisit changes occurring in the T cells of these patients compared to controls. Differential gene expression analysis across each T cell sub-cluster demonstrated that some of the highest upregulated genes in T cells across multiple different sub-clusters were genes involved in T cell activation/exhaustion. These genes included CTLA4, PDCD1, TIGIT, HAVCR2, LAG3 , TNFRSF4/9/18, ENTPD1, and TOX ( Figure 6 E; Table S5 ), very similar to our findings across each patient for the total CD4 + and CD8 + clusters ( Figure 3 L). High GZMK (granzyme K) expression in several subsets also suggested a pathological role for these T cells ( Figure 6 E; Table S5 ). CD4 + subsets also had a significant increase in many of these genes across the various sub-clusters ( Figure 6 F; Table S6 ). Despite the significantly higher activation/exhaustion gene expression in patients with CG, steroid treatment did little to reduce the expression of these genes across T cell sub-clusters in patient “A” ( Figure S10 ), again consistent with the absence of a clinical response to corticosteroid treatment. Based on the high levels of activation/exhaustion markers on these T cells, we looked for cytokine targets that might be elevated that could be blocked with approved biologics. We saw a 1.5– to 2.5-fold elevation in the expression of Th2 cytokines IL5 and IL13 in this sub-cluster, although there were too few cells to yield significance ( Figure 6 G; Table S6 ). Across several other cytokines with approved biologics, such as IL-4, IL-17, and IL-23A, we saw little signal of expression or significant increase in T and CD4 + sub-clusters ( Figures S11 A and S11B; Tables S5 and S6 ). Although we saw an increase in Th1 cells ( Figure 5 E), the expression of the typical cytokines they secrete, TNF and IFNG , was reduced in CG compared to controls ( Figures S11 A and S11B; Tables S5 and S6 ). We also saw no increase in the expression of the innate, pro-inflammatory cytokines, IL1B and IL6 , from macrophages and neutrophils ( Figure S11 C). Beyond cytokine targets, the signaling molecules downstream of cytokines, such as Janus kinases (JAKs), have approved selective kinase inhibitors. We saw slight, but significant increases in JAK3 expression across several T and CD4 + sub-clusters ( Figures 6 H and 6I). We also saw increases in JAK2 expression across non-immune cell types in the gastric glands ( Figure S11 D; Table S7 ). Though not directly targetable, the signal transducer and activator of transcription (STAT) proteins associated with JAK2 and JAK3 were elevated. STAT1 , STAT3 , STAT4 , and STAT5B expression was elevated across T cells, CD4 + T cells, and several non-immune cells ( Figures S11 D–S11F; Tables S5 , S6 , and S7 ). Despite finding an increase in CD4 + T cells by flow cytometry and demonstrating that multiple T cell subsets express significantly higher activation/exhaustion genes, TCR clonotype analysis suggested significantly lower levels of clonal T cell expansion in patients with CG compared to controls. With more T cells present in the CG mucosa (as seen by flow/histopathology), but fewer unique clones, our results pointed to an increased capacity for non-resident T cell clones to infiltrate the tissue. We discovered a significant increase in ITGA4 , which encodes integrin α4, in multiple T cell sub-clusters, especially in CD4 + T cells, with the highest fold change observed in cytotoxic CD4 cells ( Figures 6 J and 6K). Integrin α4 can form a heterodimer with integrin β1 ( Figures S12 A and S12B) and bind vascular cell adhesion molecule-1 (VCAM-1), typically present on endothelial cells. 25 While we did not see increased VCAM-1 expression on endothelial cells, expression was detectable on Fibroblasts ( Figure S12 C; Table S7 ). Integrin α4 also forms a heterodimer with integrin β7, and together this α4β7 heterodimer ( Figure S13 ) serves as a receptor for mucosal addressin cell adhesion molecule-1 (MAdCAM-1) present on endothelial cells. Mucosa-homing T cells, especially those homing to the gut, use α4β7 to enter the mucosal tissue. 26 , 27 These data, and significant increases in MADCAM1 expression on the surface of endothelial cells from patients with CG ( Figure 6 L), suggest a mechanism for the enhanced infiltration of T cells into the gastric tissue and subsequent activation and tissue damage. Biologics that target α4β7 are approved, and this targeted approach at blocking T cells from entering the gastric mucosa may be an effective and more selective method to treat patients with CG.

Further information and requests for resources and reagents should be directed to and will be fulfilled by Michael Dougan ( [email protected] ). This study did not generate any new materials. scRNA-seq data have been uploaded to NCBI GEO at GSE254513 ( https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE254513 ). All original code has been deposited at GitHub and is publicly available. The URL is listed in the key resources table . Any additional information needed to reanalyze the data reported in this article is available from the lead contact upon request.

Here we describe the first systems approach at examining the underlying etiology of collagenous gastritis using histopathology, flow cytometry, and single-cell RNA- and TCR-sequencing. We found increased frequencies of newly infiltrating ITGA4 + activated CD4 + T cells associated with pathology, and we propose integrin blockade to prevent influx of CD4 + T cells into the gastric tissue as an attractive option for patients with CG. Other therapeutic alternatives could include blocking Th2 cytokines or blocking JAK2/3 signaling. From the patients we biopsied, we found several disrupted non-immune populations, including a near absence of parietal cells. In autoimmune gastritis (AIG), another chronic inflammatory disease of the stomach, auto-antibodies targeting intrinsic factor or the H + /K + ATPase proton pump on parietal cells lead to the destruction of these cells and glandular atrophy. 28 , 29 Ultimately, patients with AIG develop pernicious anemia due to intrinsic factor deficiency, a necessary cofactor for proper vitamin B 12 absorption. 28 , 29 Patients with CG have not been found to have anti-intrinsic factor or anti-parietal cell antibodies, 30 but a significant proportion of patients with CG do have glandular atrophy, 5 , 11 , 30 a feature which we also observed. In patients with CG who were tested clinically, autoantibodies were not detected, including negative testing for anti-intrinsic factor or anti-parietal cell antibodies. However, some patients with CG have previously been found to have anti-smooth muscle or anti-nuclear antibodies, 31 which we did not have available for the patients in this study. The lack of autoantibodies against intrinsic factor and parietal cells suggests a different etiology from AIG. Patients with CG tend to develop iron deficiency anemia, though pernicious anemia is not known to co-occur. 10 , 13 , 15 , 30 , 32 Perhaps the incomplete destruction of parietal cells in patients with CG allows sufficient intrinsic factor to aid in B 12 absorption. Decreased absorption of iron itself could be due to tissue inflammation, which is a cause of iron-deficiency anemia in inflammatory bowel disease (IBD). 33 However, iron is typically absorbed in the duodenum and upper jejunum, so stomach inflammation may not be limiting iron absorption directly. Nevertheless, non-heme iron from dietary intake must be reduced from its ferric form (Fe 3+ ) to ferrous (Fe 2+ ) iron before it can be absorbed, and this is most readily achieved through hydrochloric acid in the stomach. 34 Therefore, iron-deficiency anemia in patients with CG could be due to the destruction of parietal cells leading to incomplete stomach acidification and lack of iron reduction. In patients with AIG who lack parietal cells, iron deficiency often precedes pernicious anemia. 28 Despite the reduction in parietal cells we found in patients with CG, the increase in iron transport and B 12 -related genes may indicate an adaptation due to their iron deficiency. However, whether LTF and LCN2 are aiding in iron transport or preventing uptake through sequestration is unclear; iron sequestration during chronic inflammation has been known to be a cause of anemia. 35 The cause of glandular atrophy in CG is not currently known. We hypothesize that the infiltration of pathogenic T cells may result in deposits of collagen from fibroblasts. As collagen deposition is a hallmark of CG, our finding that collagen genes were some of the highest upregulated genes in fibroblasts from patients with CG was somewhat expected. In many autoimmune conditions, such as Sjögren’s Syndrome, fibrotic tissue, including collagen deposits, can accumulate in response to damage and interfere with the normal function of glands, 36 which we hypothesize could also be occurring in the stomachs of patients with CG. In the case of Sjögren’s Syndrome, while the causative antigen is unknown, CD4 + T cell infiltration and proliferation are strongly implicated in causing tissue damage that leads to fibrosis. 36 , 37 We saw that patients with CG had a preferential influx of CD4 + T cells over CD8 + T cells, though both cell types were present at high levels and had high expression of activation/exhaustion genes. However, whether factors are directly released from T cells to induce fibroblast collagen deposition remains unclear. We saw an increase in granzyme+ T cells, including cytotoxic CD4 + T cells and significant increases in GZMK expression across T cell sub-clusters. Granzyme K has been shown to be increased in inflamed tissues, and granzymes in general can cause tissue damage, 38 which may ultimately lead to fibrosis. 39 Cancer, chronic infection, and autoimmune diseases can result in continual T cell activation, and ultimately T cell exhaustion. 40 In patients with Crohn’s Disease, markers of T cell exhaustion, such as CD39 and PD-1, were significantly enriched compared to controls, and these markers were expressed at even higher levels in patients with active disease compared to those in remission. However, CD39 + PD1 + CD8 + T cells in patients with Crohn’s disease were more dysfunctional in non-inflamed tissue compared to active lesions, suggesting different exhaustion programs. In patients in remission, the upregulation of TOX was seen in the CD39 + PD1 + T cells, which ameliorated the inflammatory signature and suggested this gene was critical for true exhaustion. 41 We saw that TOX was significantly upregulated in many T cell subsets in patients with CG, suggesting that at least some of these cells may have been exhausted at this stage rather than activated. However, we also saw the upregulation of KLRG1 in the main CD8 + cluster across several patients. T cells exposed to antigen exhibit a fate decision in which they can become effectors or progress to an exhausted state. During this binary fate decision, KLRG1 Lo PD-1 + cells often become exhausted, while KLRG1 Hi cells are terminal effectors. 42 Perhaps in the context of CG, CD8 + T cells are activated effectors, while CD4 + T cells, which had lower KLRG1 expression across several patients, have a more dysfunctional phenotype. Conversely, the lack of clonal T cell expansion uncovered by TCR sequencing in patients with CG suggests T cells could be exhausted, as proliferative capacity is greatly diminished in dysfunctional T cells, 43 and yet we also saw slightly higher levels of cycling T cell sub-clusters in patients with CG by scRNA-seq. The state of activation and exhaustion in the T cells in patients with CG may be dynamic, such as in patients with Crohn’s disease, 41 with some cells exhibiting markers of exhaustion but still capable of causing inflammation. Some clones, such as those that were found at both timepoints in patient “A”, could be capable of causing tissue damage. Of the few CD4 + T cell clones that were proliferating, up to 50% of them were Th1 cells. The Th1 clones that were expanded were also found in the cytotoxic CD4 subcluster, suggesting a possible pathogenic role for these cells. We have previously reported a significant increase in Th1 effector cells present during checkpoint inhibitor-induced colitis, as well as enhanced IFNG and GZMB expression in reactivated tissue-resident memory CD8 + T cells during this form of colitis. 23 However, the expression of the pro-inflammatory cytokines, such as TNF and IFNG , was significantly reduced in the Th1 cells from patients with CG. Nevertheless, as there were increased numbers of Th1 cells, the overall concentration of these cytokines could still be higher in patients with CG. The increased MHC-I expression observed on gastric cells suggests increased local concentrations of interferons and Th1 cytokines. The local concentration of these cytokines may be high enough to result in parietal cell death. This could be another possible mechanism to explain glandular atrophy in these patients. Future studies to characterize the cytokine protein expression within the tissue should be carried out to answer this question. The Th1 cells in collagenous gastritis may be less functional and play a different role than in checkpoint colitis. Similar to checkpoint colitis, we observe a marked reduction in T RM cells in patients with CG, with an increase in more cytotoxic/effector T cell subtypes. However, in checkpoint colitis, the T RM clonotypes are shared with newly expanded cytotoxic effector cells, suggesting the pathogenic T cells originated from tissue-resident populations. 23 In patients with CG, we found reduced clonal expansion, which suggests that rather than tissue-resident memory cells becoming effectors, newly infiltrating cells of diverse clonotypes were replacing the memory population. Interestingly, during checkpoint colitis, Tregs also expand, though these cells do not appear to be sufficient to prevent inflammation. 23 We also found a significant increase in Tregs present in patients with CG, though the extent of their functionality is unclear. The Treg population that was increased had lower levels of expression of canonical Treg genes, including FOXP3 and CTLA4 . In mice, Tregs with the lower expression of Foxp3 have greatly reduced suppressive capabilities. 44 Therefore, this Treg subset seen only in patients with CG may be less functional. Though patients with CG have an increase in Tregs, the cause of the underlying inflammation in CG is unknown. Across many patients with CG, Helicobacter pylori infection prevalence is rare or non-existent. 5 , 7 All patients with CG in our study were negative for the presence of H. pylori. Other environmental triggers, such as a chronic bacterial or viral infection, could lead to lasting inflammation and dysfunctional T cells discovered in patients with CG. However, these responses toward a trigger are often skewed, such as a Th1 response for viral infections and a Th2 response for extracellular bacteria. Indeed, the balance between Th1 and Th2 responses is thought to influence outcome in autoimmune diseases. 45 Although we saw higher levels of Th1 cells compared to Th2, both subsets were significantly increased compared to controls, suggesting no preferential skewing toward either Th1 or Th2 responses. Both elevated Th1 and Th2 cytokines have previously been seen in patients with collagenous gastritis. 46 Eosinophils have been previously detected in CG, 7 and their presence suggests an allergic, Th2 response, though the Th1 expansion we find suggests a more dynamic immune response is occurring. Notably, Th17 cells, which are involved in mucosal defense and can play a pathogenic role in IBD, 47 were reduced in patients with CG. This suggests Th17 cells and the microbes they protect against are not a direct cause of the disease. Pinpointing a potential environmental trigger may be difficult, and the rarity of the disease also makes studying genetic risk factors challenging. Though our study was the first systems approach to studying this disease, it was still limited by the relatively small number of patients. Within the small cohort of patients with CG, we found variability that may have been driven by a variety of factors, including the age of the patients, with young and old patients having slightly different cell composition, which may explain the different symptom presentation for this disease. 6 , 7 Nevertheless, in many of the patients with CG, almost 50% of the CD4 T cells were either Th1 or the phenotypically similar cytotoxic CD4 T cells. These data and the skewed CD4:CD8 ratio observed by flow cytometry and scRNA-seq strongly imply pathogenic CD4s as a causative agent in the disease. In patients with celiac disease, there is a high prevalence of the HLA-DQ2.5 (DQ2.5) haplotype. The DQ2.5 chain has been shown to bind and present gluten peptides to pathogenic CD4 T cells in celiac disease. 48 In patients with CG, an increased prevalence of the DQ2.5 haplotype has also been seen. 31 There is overlap between CG and celiac disease, with up to 30% of patients with CG having both diseases. 7 However, it is unclear if this overlap could be partially attributed to haplotype, though the pathogenicity of CD4 T cells in celiac disease and observed in this study for CG, suggests MHC-II might play an important role. DQ2.5 is also highly prevalent in type 1 diabetes, 49 further implicating this haplotype in the development of autoimmunity. The DQ2.5 chain may be responsible for presenting a driving antigenic peptide to CD4 T cells in CG, similar to Celiac Disease, but to date, no such analogous autoantigen or external antigen has been identified in CG. Without a distinct trigger for the disease, the treatment of CG must use other strategies to limit inflammation. Steroids have largely proved to be unsuccessful in managing CG, 5 and as we show, the transcriptional effect of steroids on T cells in the stomach is minimal. A more targeted approach may be more effective. One therapeutic strategy to consider for the treatment of CG is B cell depleting agents. Although we did not see strong transcriptional changes from B cells, there were significant increases in both B cells and IgA+ plasma cells in patients with CG. B cells and antibodies secreted from plasma cells have significant roles in a variety of autoimmune diseases, such as multiple sclerosis, rheumatoid arthritis, and systemic lupus erythematosus. 50 Because of the role that these cells play in various autoimmune conditions, B cell depleting antibodies have been extensively studied and approved for several of these diseases. However, anti-CD20 and anti-CD19 therapeutic antibodies do little to curtail levels of plasma cells, which are CD20 − /CD19 − and long-lived within the gut. 50 , 51 Beyond our finding of increased levels of these cells, it is unknown if B cells, plasma cells, and/or autoantibodies play a role in CG; more research is needed to understand the role these factors play and if patients would benefit from B cell depletion. Based on our data, T cells are an ideal target for therapeutic intervention. With the increase in Th2 cells, we saw a slight increase in the expression of Th2 cytokines IL5 and IL13. These cytokines have approved biologic therapeutics such as mepolizumab and dupilumab for targeting either the cytokines themselves or their respective receptor, such as IL-4Rα which also binds IL-13. 52 , 53 Diseases treated with these biologics include eosinophilic granulomatosis and eosinophilic esophagitis. 52 , 53 Eosinophils have been found to infiltrate the tissue of patients with CG, 7 but due to their fragility during the scRNA-seq processing, granulocytes are often lost, 54 and eosinophils were absent from our dataset. However, we did see increased levels of stromal eosinophils from histopathology. These findings, along with the Th2 signature, suggest these biologics could be useful in the treatment of CG. In addition to blocking Th2 cytokines, we saw increases in JAK3 expression across T cell sub-clusters. JAK3 signals downstream of cytokines that bind the common gamma chain (γ c ) receptors, such as IL-2, IL-4, IL-7, IL-15, and IL-21. 55 The JAK3 inhibitor, tofacitinib, is approved for several indications, including ulcerative colitis. 56 Based on our data, this drug could also potentially be repurposed for CG. However, JAK inhibitors are typically more immunosuppressive than targeting a specific cytokine, 57 and target multiple cell types. Therefore, we propose a more targeted approach to limit T cell infiltration through the blockade of α4β7 integrin. Blockade of the α4β7 has been studied for decades for the treatment of IBD to prevent the migration of mucosal-homing T cells into the gut. Both the expression of the α4β7 on T cells, as well as its ligand, MAdCAM-1, present on the endothelium, is significantly increased in patients with IBD. 58 These findings ultimately led to several clinical trials and the approval of the anti-α4β7 antibody, vedolizumab, for both Ulcerative Colitis and Crohn’s disease. 26 , 27 Notably, in patients with CG, ITGA4 was higher in T cell subclusters, including IFNG + CD8s, granzyme+ CTLs, Th17/T EM , cytotoxic CD4s, and activated Tregs, which indicates their improved ability to home to this mucosal tissue. Integrin α4 can form heterodimers with either integrin β1 or β7. Although we saw low expression of both ITGB1 and ITGB7 on all T cell subclusters, with no increase in patients with CG, there are likely sufficient levels of these integrins such that the increased expression of α4 integrin is meaningful. Integrin α4β1 does help T cells home to sites of inflammation, although it is not usually involved in T cell recruitment to the GI tract. However, in Crohn’s disease, the upregulation of its ligand, VCAM-1, on the endothelium due to chronic inflammation, may contribute to the recruitment of T cells to the gut via α4β1. 59 While we saw the increased expression of VCAM-1 on fibroblasts in patients with CG, we did not observe increased levels on the endothelium. Fibroblasts can express VCAM-1, especially during infection and cancer. 25 , 60 , 61 The role fibroblast-expressed VCAM-1 might play in CG is unknown, though it could help mediate interactions with T cells. We did observe increased MADCAM1 expression on endothelial cells, which implies an improved ability of T cells expressing α4β7 to attach to endothelial walls and extravasate into the gastric tissue. Blockade of α4-integrin, such as with natalizumab, which can be effective in Crohn’s disease, 62 could mitigate any T cells homing via α4β1 integrin. However, α4β7 has a stronger link to gut T cell homing, and VCAM-1 expression was limited to fibroblasts, possibly as a consequence of inflammatory mediators produced by T cells. Additionally, α4 blockade alone with natalizumab is associated with a risk of rare, but life-threatening progressive multifocal leukoencephalopathy (PML). With these combined findings, we believe blockade of α4β7 could be a promising therapeutic target for patients with CG to prevent excessive lymphocytic infiltration and potential mucosal damage. We have established a small, interventional phase 2 clinical trial to understand the effect of vedolizumab on patients with CG ( NCT06317220 ). With data presented from this current study and future clinical studies, we hope to better understand this disease and identify effective treatments. While our study is the first systems approach to describe collagenous gastritis, it does have several limitations. First, as this is a rare disease, we were limited by the number of patients from whom we could collect tissue. With a small cohort of patients, results could be driven by a variety of factors, including the age of the patients, with young and old patients having slightly different cell composition. We strove to collect tissue from roughly age- and sex-matched controls with 5/7 (71%) female patients with CG compared to 4/6 (67%) female in the control group and a respective median age of 33 and 32.5 in CG and control patients, respectively. However, the CG cohort has some outlier patients from an age perspective, with a minimum age of 14, compared to 24 in the control and a maximum age of 67, compared to 40 in the control group. Because the CG group included a pediatric and a geriatric patient, this could have introduced bias in the results. Additionally, the control group was selected from patients already undergoing esophagogastroduodenoscopies (EGDs) for other conditions, including eosinophilic esophagitis and celiac disease. Although no gastric inflammation was visible by EGD or on routine H&E staining, the presence of other gastrointestinal inflammatory diseases could have influenced the T cell repertoire and other cellular and molecular characteristics of the control group compared to true healthy controls. However, the heterogeneity of the diseases in the control group paired with the overall homogeneity of the data from the stomach suggests this limitation may not undermine the results presented. We also included both fresh and frozen tissue in our analysis. Though we tried to demonstrate that these samples were largely similar from a gene expression perspective, it could have introduced bias to the results. Furthermore, not every frozen sample was processed on the same day, such that differences in processing, staining, and gating could result in slightly different CD45 + populations.

Collagenous gastritis (CG) is a rare disease, first described in 1989, 1 that has an estimated prevalence of 13 cases per 100,000 esophagogastroduodenoscopies. 2 The disease manifests with a range of symptoms, including iron-deficiency anemia, abdominal pain, diarrhea, weight loss, nausea, and vomiting. 2 , 3 , 4 Histopathology of gastric biopsies shows characteristic thick bands of collagen, typically >10 μM. 1 CG tends to have a bimodal distribution with either pediatric or adult onset, and appears to affect females slightly more frequently than males. 2 , 5 , 6 Presentation of the disease often differs depending on the age of onset: pediatric-onset patients with CG have more abdominal pain and anemia, while adult-onset patients often present with weight loss. 6 , 7 However, overlap between the two phenotypes exists. 8 No safe and effective treatment options for CG have been clearly defined, though several small studies have tried proton-pump inhibitors, 9 , 10 H2-receptor antagonists, 11 steroids, 12 and dietary changes 12 with mixed or no responses to treatments. 5 Additionally, oral iron supplements are often administered to address the anemia commonly associated with CG, though this does not address the underlying disease process 13 , 14 Recently, in one of the largest case studies of CG, topical budesonide therapy was associated with clinical and histological improvements in patients. 6 This both implicates the immune system in pathology and suggests a potential treatment option for patients, although the single-arm, retrospective design limits the generalizability of these findings. The rarity of collagenous gastritis also makes studying and characterizing the etiology difficult. Patients with collagenous gastritis appear to be at increased risk for other gastrointestinal diseases, including celiac disease (an inflammatory response to gluten), collagenous colitis, and lymphocytic gastritis. 2 , 7 , 15 These inflammatory comorbidities are often associated with lymphocyte infiltration. In addition to the characteristic collagen damage, lamina propria lymphoplasmacytosis and increased intraepithelial lymphocytes have been proposed to be part of the histopathologic diagnostic criteria for CG. 11 Few studies have looked beyond histopathology toward characterizing this disease in molecular or cellular detail. To better understand the etiology of this disease in the hopes of identifying therapeutic targets, we collected gastric and duodenal biopsies of seven patients with CG, including one patient before and after systemic glucocorticoids, as well as control biopsies. Biopsies were analyzed with multiple modalities, including histopathology, flow cytometry, as well as single-cell RNA- and T cell receptor-sequencing (scRNA-seq). We found a striking increase in antigen-experienced α4β7 integrin + CD4 + T cells, implicating infiltrating CD4 + T cells as a potentially targetable population to treat collagenous gastritis.

MD has received consulting fees from Genentech, Regeneron, Partner Therapeutics, Eli Lilly, Mallinckrodt Pharmaceuticals, Aditum, Foghorn Therapeutics, Palleon, Sorriso Pharmaceuticals, Generate Biomedicines, Asher Bio, Neoleukin Therapeutics, Alloy Therapeutics, Third Rock Ventures, DE Shaw Research, Agenus, Astellas, and Curie Bio; and he is a member of the Scientific Advisory Board for Veravas, Monod Bio, Axxis Bio, and Cerberus Therapeutics. Following the completion of this study, MD entered into a collaborative research agreement with Takeda Pharmaceuticals for an Investigator Initiated Trial of vedolizumab for the treatment of collagenous gastritis; the vedolizumab is being provided by Takeda free of charge, and no other financial support is being provided. SKD received research funding unrelated to this project from Takeda, Novartis, and Bristol-Myers Squibb and is a founder, science advisory board member (SAB) and equity holder in Kojin, and has equity in Axxis Bio. MM-K has served as a compensated consultant or advisory board member for AstraZeneca, Pfizer, Sanofi, Daiichi-Sankyo, AbbVie, and Repare; she has received royalties from Elsevier, all unrelated to this study. MJW is an employee of and has equity in Nuvalent, Inc., unrelated to this study. Other authors declare no other competing interests.

REAGENT or RESOURCE SOURCE IDENTIFIER Antibodies FITC anti-human CD45 Antibody Biolegend 368507; RRID: AB_2566367 Brilliant Violet 785™ anti-human CD45 Antibody Biolegend 304048; RRID: AB_2563129 BD Horizon™ PE-CF594 Mouse Anti-Human CD3 BD Biosciences 562280; RRID: AB_11153674 Brilliant Violet 421™ anti-human CX3CR1 Antibody Biolegend 341620; RRID: AB_2687148 FITC anti-human CD103 (Integrin αE) Antibody Biolegend 350204; RRID: AB_10639865 PerCP/Cyanine5.5 anti-human CD69 Antibody Biolegend 310926; RRID: AB_2074956 Brilliant Violet 510™ anti-human TCR γ/δ Antibody Biolegend 331220; RRID: AB_2564275 Pacific Blue™ anti-human CD4 Antibody Biolegend 317429; RRID: AB_1595438 Brilliant Violet 711™ anti-human CD8 Antibody Biolegend 344734; RRID: AB_2565243 PE anti-human CD279 (PD-1) Antibody Biolegend 329906; RRID: AB_940483 Brilliant Violet 605™ anti-human CD14 Antibody Biolegend 301834; RRID: AB_2563798 Brilliant Violet 605™ anti-human CD15 (SSEA-1) Antibody Biolegend 323032; RRID: AB_2562132 Brilliant Violet 605™ anti-human CD19 Antibody Biolegend 302244; RRID: AB_2562015 CONFIRM anti-CD4 (SP35) Rabbit Monoclonal Primary Antibody Roche 05552737001; RRID: AB_2335982 CD8 IHC antibody Leica PA0183; RRID: AB_10555292 PD-1 (NAT105) Mouse Monoclonal Antibody Cell Marque MRQ-22; RRID: AB_1160829 PD-L1 (E1L3N®) XP® Rabbit mAb Cell Signaling Technology #13684S; RRID: AB_2687655 Gastrin IHC antibody Leica PA0681; RRID: AB_10554593 CD19 IHC antibody Leica PA0843-U; RRID: AB_10555287 CONFIRM anti-CD3 (2GV6) Rabbit Monoclonal Primary Antibody Ventana Roche 790-4341; RRID: AB_2335978 CONFIRM anti-CD68 (KP-1) Primary Antibody Ventana Roche 790-2931; RRID: AB_2335972 CD45 (Intracellular Domain) (D9M8I) XP® Rabbit mAb #13917 Cell Signaling Technology 13917S; RRID: AB_2750898 Chemicals, peptides, and recombinant proteins DISCOVERY ChromoMap DAB Kit Ventana Roche 760-159 DISCOVERY Red Kit Ventana Roche 760-228 DISCOVERY OmniMap anti-Ms HRP Ventana Roche 760-4310 DISCOVERY OmniMap anti-Rb HRP Ventana Roche 760-4311 DISCOVERY UltraMap anti-Rb Alk Phos Ventana Roche 760-4314 DISCOVERY UltraMap anti-Ms Alk Phos Ventana Roche 760-4312 DISCOVERY CC1 Ventana Roche 950-500 Cell Condition 2 (CC2) Ventana Roche 950-123 Bluing Reagent Ventana Roche 760-2037 Hematoxylin II Ventana Roche 790-2208 Critical commercial assays Zombie NIR™ Fixable Viability Kit Biolegend 423105 Tumor Dissociation Kit, human Miltenyi 130-095-729 Chromium Next GEM Single Cell 5' Kit v2 10x Genomics PN-1000263 Library Construction Kit 10x Genomics PN-1000190 Chromium Single Cell Human TCR Amplification Kit 10x Genomics PN-1000252 Chromium Next GEM Chip K Single Cell Kit 10x Genomics PN-1000286 Deposited data Single-cell RNA-sequencing This paper NCBI GEO GSE254513 Source code This paper https://github.com/mjlwalsh/collagenous-gastritis Software and algorithms FlowJo v10.10 TreeStar RRID: SCR_008520 Prism v10 GraphPad RRID: SCR_002798 Cellranger v7.0.0 10x Genomics http://software.10xgenomics.com/ ; RRID: SCR_023221 Human GRCh38 reference genome (GENCODE v32/Ensembl98) 10x Genomics https://www.10xgenomics.com/support/software/cell-ranger/latest/release-notes/cr-reference-release-notes Human V(D)J reference (GRCh38) v7.1 10x Genomics https://www.10xgenomics.com/support/software/cell-ranger/downloads R v4.2.2 R Project https://www.r-project.org/ ; RRID: SCR_001905 RStudio v2023.09.0 Posit https://posit.co/downloads/ ; RRID: SCR_000432 DropletUtils v1.18.1 Griffiths et al. 63 Bioconductor Seurat v4.3.0 Stuart et al. 64 https://satijalab.org/seurat/ fgsea v.1.24.0 https://doi.org/10.1101/060012 Bioconductor; RRID: SCR_020938 Molecular Signatures Database v7.2 Liberzon et al. 65 https://www.gsea-msigdb.org/gsea/msigdb/index.jsp ; RRID: SCR_016863 DESeq2 v1.38.3 Love et al. 66 Bioconductor; RRID: SCR_015687 ggplot2 v3.4.1 Hadley Wickham https://ggplot2.tidyverse.org/ ; RRID: SCR_014601 This study involves human participants and was conducted in accordance with the Declaration of Helsinki and the US Common Rule. Approval for this study was granted by the Dana-Farber Cancer Institute IRB (IORG #0000035, FWA #00001121). All adult participants provided written informed consent prior to sample collection. All minors and their parent(s) or guardian(s) provided written informed assent and consent, respectively, prior to sample collection. Patients with previously diagnosed collagenous gastritis were recruited for the study at the time of a clinically indicated endoscopy ( Figure 1 A). Controls patients were undergoing endoscopy for non-gastric conditions and were approximately age/sex-matched to the CG cohort ( Table S1 ; please also see limitations of the study ) Following patient consent, esophagogastroduodenoscopies (EGDs) were performed, and systematic biopsies of the gastric body/fundus ( Figure 1 B) and duodenum were taken during each procedure. Control samples were confirmed by EGD to have no visible gastric inflammation. Patient “A” had tissue collected before and after a six-week treatment course with glucocorticoids. Patient “D” had a second biopsy collected five months later at a follow-up appointment. Collagenous gastritis tends to affect females slightly more frequently than males. 2 , 5 , 6 Out of the seven patients from the collagenous gastritis cohort, five were female (5/7; 71%). Therefore, we aimed to minimize any effect of sex differences in our comparisons by collecting control samples that were roughly sex-matched. Four out of the six control samples were female (4/6; 67%). Biopsy bites were immediately collected into DMEM media on ice; put into freezing media (10% DMSO; 40% RPMI; 50% FBS) and frozen (for later CD45-enrichment); or fixed in 10% neutralized formalin. Fresh or frozen and thawed tissue was then manually dissociated with a sterile razor and transferred to pre-warmed 37°C digestion enzyme mix in RPMI (Miltenyi tumor dissociation kit, cat. 130-095-929). This kit was used as it successfully dissociates highly collagen-rich tissue, such as pancreatic adenocarcinoma, 67 and has also been used to process other gastrointestinal biopsies. 23 In a 37°C water bath, tissue digestion was agitated every 2 minutes, and pipetted up and down every 10 minutes for a total of 30 minutes. Digested tissue was filtered through a 70μM strainer, washed with RPMI containing 10% FBS, and the single-cell suspension was collected and processed for downstream assays. The formalin-fixed biopsy samples were embedded in paraffin, sectioned, and stained with Hematoxylin & Eosin (H&E) following the routine protocol. Two pathologists with expertise in Gastrointestinal pathology (YK and MM-K) reviewed the H&E-stained slides and confirmed the diagnosis of CG in the study cohort or normal gastric body/fundic mucosa in the controls. The diagnosis of CG is based on the presence of collagen bands thicker than 10 μm in the subepithelial layer and increased mononuclear cell infiltration in the lamina propria. 5 Further, five representative areas were selected for morphological and immunohistochemical evaluations. For immunohistochemical analysis, CD4, CD8, PD-1, and PD-L1 were used, and positive lymphocytes were counted as a sum of five representative high-power fields (x400 magnification) each in the epithelium and stromal compartments. Immunohistochemistry for gastrin was also performed to confirm the gastric body/fundic location of the biopsy. Dual staining to confirm cell type specificity was performed with PD-1/L1 and CD3, CD68, CD19, CD45. Dual staining was performed with the Ventana Roche DISOCVERY platform. Briefly, deparaffinization, followed by cell conditioning and retrieval with DISCOVERY CC1 reagent for 64 minutes was performed. DISCOVERY inhibitor CM (DISCOVERY ChromoMap DAB kit) incubation for 8 minutes was followed by the first antibody application for 37°C for 32 minutes. OminiMap anti-Rb or anti-Ms HRP was incubated, followed by DISCOVERY DAB CM (DISCOVERY ChromoMap DAB kit) at room temperature. Cell conditioning and retrieval was repeated with DISCOVERY CC2 reagent, followed by incubation with the antibody against the second target at room temperature. UltraMap anti-Rb or anti-Ms Alkaline Phosphatase was incubated for 16 minutes, followed by Fast Red application (DISCOVERY Red Kit). Hematoxylin II and bluing reagent were incubated for 12 minutes and then coverslips were added.Eosinophils were quantified from H&E sections with five representative high-power fields in the epithelial and stromal compartments. The details of antibodies used for immunohistochemistry are described in Tables S8 and S9 for dual stains. Single cell suspensions from fresh gastric and duodenal biopsies were generated as above and resuspended in PBS with 2% FBS containing flow cytometry antibodies. Analysis was performed on a Sony Biotechnology SP6800 Spectral Analyzer and analyzed with the Sony Biotechnology SP6800 Software and FlowJo (Tree Star, Ashland, OR). Flow cytometry antibodies used in this study were purchased from Biolegend: anti-CD45, anti-CD3, anti-CD14, anti-CD4, anti-CD8, anti-CX3CR1, anti-CD103, anti-CD69, anti-gdTCR, anti-PD1, anti-CD14, anti-CD15, anti-CD19, and Zombie-NIR. Single cell suspensions were generated from gastric biopsies as above and resuspended in PBS with 2% FBS. For CD45-enrichment, cells were stained with anti-CD45 and Zombie-NIR and sorted for CD45 + , Zombie - events into 20% FBS in RPMI on a BD FACS Aria II at the Dana Farber Flow Sorting Core. Cells from fresh (unsorted), or frozen (CD45-sorted) samples were counted on a hemacytometer and resuspended to a concentration of 1000 cells/μL in 10% FBS in RPMI. Cell suspensions were loaded onto a 10x Chromium Controller with the Single Cell K Chip, using the Chromium Next GEM Single Cell 5ʹ GEM Kit v2 reagents and beads. Library preparation was performed with the Library Construction Kit for unsorted (fresh) and CD45-sorted (frozen) samples, and the TCR Amplification Kit was also used on the CD45-sorted samples for T cell receptor (TCR) library preparation. Samples were sequenced on an Illumina NovaSeq 6000 instrument with 2x150bp sequencing (Azenta Life Sciences). The 10x Cellranger pipeline (v7.0.0) was used to align reads to the hg38 Homo sapiens genome (from 10x Genomics). The GRCh38 Human V(D)J Reference (v7.1.0) was downloaded from 10x Genomics and used with Cellranger ‘multi’ to profile the TCR repertoire of CD45-sorted samples. The count matrices were imported for downstream analysis into R using the Seurat (v4.3.0) package. The DropletUtils (v1.18.1) package was used to exclude empty droplets 63 from the count matrices with an FDR <0.01 and genes expressed in fewer than 3 cells were excluded. Cells expressing fewer than 500 transcripts and 200 total genes were removed, as were cells with mitochondrial gene expression percentages greater than 2 MADs above the median (med+2MAD). Three samples with med+2MAD > 20% had their cutoff set to 10%. Data from each sample was log-normalized, and 2000 variable features from each sample were identified. The fresh, unsorted samples were integrated together to minimize batch effects using reciprocal principal component analysis (RPCA) reduction and four reference samples (two each CG/control) with 2000 integration anchors. 64 The fresh, unsorted dataset was z-scaled and the top 30 principal components (PCs) were identified and used for Uniform Manifold Approximation and Projection (UMAP) generation. A shared nearest neighbor (SNN) graph was generated based on each cell’s 20-nearest neighbors and clusters were determined with the Louvain algorithm. ‘FindAllMarkers’ was used to find top cluster-defining genes from which to assign cell-types. A similar method was used to generate a UMAP from the frozen, CD45-sorted samples. Finally, a combined dataset and UMAP were generated with both fresh and frozen samples using the above methods. From this combined dataset, five clusters were removed that were expressing few transcripts or were doublets with expression of both canonical B and T cell genes. To compare fresh, unsorted samples to frozen, CD45-sorted samples, the average expression per gene per sample per cluster was calculated with ‘AverageExpression` function. Pearson correlations were calculated for each sample and each CD45-positive cluster (“CD8”, “CD4”, “γδ T”, “T (cycling)”, “IgA+ Plasma”, “B”, “NK”, “Mast”, and “Macrophage”) as these cells should appear in both unsorted (fresh) and CD45-sorted (frozen) samples. A combined Pearson correlation was also calculated for each gene/cluster using the ‘AverageExpression’ per cluster. Pseudobulk expression of each sample was calculated using the ‘AggregateExpression’ function and the DESeq2 (v1.38.3) package was used to generate a variance-stabilizing transformation (VST) from the samples. 66 To compare how similar each sample was to one other, a principal component analysis (PCA) was performed on these pseudobulk VST-matrices from either CD45-sorted, CD45-positive clusters or the unsorted, CD45-negative clusters for each sample. For T cell sub-clustering, cells from the “CD8”, “CD4”, “T (Cycling)” or “γδ T” main clusters were subset from the integrated object, which was then split by sample. For each sample, a new set of variable features were found, excluding TCR and IgG variable regions (e.g. TR[ABDG][VJ]), from which a new set of integration anchors were found to integrate samples together. The top 20 PCs were used to generate the UMAP and downstream clustering as described above. Three small, non-T cell clusters were removed from this T cell object based on MUC6 , PGC , or COL3A1 expression above the mean; re-clustering as above was performed. For CD4 clustering, “LTB+ CD4” and “CTLA4+ CD4” sub-clusters with no expression of CD8A or cells with exclusive expression of CD4 ( CD4 ≥ 1 and CD8A / CB8B = 0) were subset from the integrated T cell object. Three thousand new variable features were found from this CD4 subset object; common features between new variable CD4 features and original T cell integration anchors were used to generate PCs, of which the top 15 were used for UMAP and clustering as described above. The ‘FindMarkers’ function was used to compare expression of genes in clusters with greater than 10 cells per disease status (i.e. Control vs. CG) from unsorted samples for differential gene expression analysis of non-immune cells. For immune cells, this analysis was performed on CD45-sorted cells to minimize contamination of highly expressed transcripts from gastric cells. In a similar manner, activated T cell gene expression was calculated from the “CD8” or “CD4” main UMAP clusters, comparing each individual CD45-sorted patient sample to all sorted controls. For each T cell and CD4 T cell sub-cluster in which more than 10 cells were found per disease status, ‘FindMarkers’ was run comparing control vs. CG samples. For Gene Set Enrichment Analysis (GSEA), fold change gene expression between all cells from (CD45-sorted) patient A vs. cells from patient A-post steroids was calculated using the ‘FoldChange’ function. These fold changes were used with the fgsea (v1.24.0) package and the Molecular Signature (v7.2) Hallmark gene signatures 65 to calculate GSEA. Unfiltered TCR contig annotations from CellRanger were used to determine clonotypes by the matching the Vα/Jα genes with the amino acid sequence of the α-chain complementarity determining region 3 (CDR3) and the Vβ/Jβ genes with the β-chain CDR3. Cells that shared both complete Vα/Jα-CDR3 and Vβ/Jβ-CDR3 were considered of the same clonotype; in the few cells that lacked sequencing information of either α or β chains, one matching chain was considered the same clonotype. TCR clonotype information was filtered to just those cells within the total T cell sub-clustering. For shared, expanded TCR clonotype analysis, any clonotype with ≥2 cells was counted if it appeared in ≥2 subclusters. Clonotype counts across patients (or pre/post steroids for patient “A”) were summed. For differential gene expression, the Wilcoxon rank-sum test (as default for ‘FindMarkers’) was used, with P-values adjusted with the Benjamini-Hochberg (BH) procedure. Comparison of cell populations was performed with either a student’s t-test or Wilcoxon rank test. Multiple comparisons were adjusted using the BH procedure. Statistical analysis was performed with GraphPad Prism or R. All error bars are S.E.M. Data were considered significant when p ≤ 0.05; ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001, ∗∗∗∗ p < 0.0001 (or p-values displayed).

Funding: This study was funded by the 10.13039/100000002 National Institutes of Health grant R01AI169188 (MD and SKD), the Peter and Ann Lambertus Family Foundation (MD), and the National Institutes of Health grant U01CA224146 and R01AI158488 (SKD).