Gastric cancer tertiary lymphoid structures in the muscularis propria are associated with therapy resistance and promote immunosuppression

Abstract

Tertiary lymphoid structures (TLS) are local immune microenvironments within tumors housing T cells and B cells, are often coordinating anti-tumor immunity and associated with better responses to immunotherapy. Here, we use H&E morphological assessment and multiplex immunofluorescence to analyze surgical tumor specimens from 71 patients with gastric cancer (GC) after treatment with neoadjuvant anti-PD-1 therapy, chemotherapy, and subsequent gastrectomy. We find increased TLS density and area in the muscularis propria of treatment-resistant GCs. These muscularis TLS are structurally and functionally impaired, with increased naïve B cell presence but reduced germinal center B cell infiltration. Mechanistically, PLA2G2A⁺ tumor cells accumulate adjacent to disrupted muscularis TLS, and PLA2G2A treatment in vitro increases differentiation of a fibroblast cell line into PDGFRA⁺ inflammatory cancer-associated fibroblasts (iCAF), which are linked to CXCL14 secretion and recruitment of naïve B cells. In vitro, PLA2G2A also impairs B cell differentiation and induces T follicular helper cell death, potentially contributing to muscularis TLS dysfunction. Taken together, we propose a tumor cell-iCAF-B cell axis disrupting TLS function, with this axis serving as a potential therapeutic target to overcome immunotherapy resistance.

Tertiary lymphoid structures (TLS) are thought to orchestrate anti-tumor immunity and support immunotherapy. Here, the authors find dysfunctional TLS in the muscularis propria of treatment-resistant gastric cancer, where PLA2G2A⁺ tumor cells interact with B cells and T follicular helper cells to contribute to immunosuppression.

Introduction

Gastric cancer (GC) remains a leading cause of cancer-related mortality worldwide. Owing to late diagnosis and limited therapeutic options, the 5-year survival rate of patients with advanced-stage GC remains below 30%^1,2. Although immune checkpoint blockade (ICB), particularly anti-PD-1 therapy, has revolutionized oncology^3–5, more than 60% of advanced GC patients show primary or acquired resistance^6,7, underscoring an urgent need to define immunosuppressive mechanisms within the gastric tumor microenvironment.

Tertiary lymphoid structures (TLS) are ectopic immune niches that coordinate T cell-B cell interactions and germinal center formation to potentiate antitumor immunity^8–10. Across diverse cancer types, mature TLS have been generally associated with favorable immunotherapy responses and improved survival, primarily by generating tumor‑specific T cells and antibody-producing plasma cells^11–16. However, emerging spatial analyses have revealed functional stratification: the maturation state, cellular composition, and anatomic location of TLS are critical determinants of their immunogenicity^17–20. Although TLS localized to the mucosa/submucosa are linked to effector immunity, the functional role of TLS in the deeper muscularis propria remains enigmatic.

Within TLS, germinal center B (GCB) cell responses have been particularly associated with favorable prognosis and enhanced immunotherapy efficacy^16,21–23. Functionally, GCB exert anti-tumor effects via multiple mechanisms, including antibody-dependent cellular cytotoxicity, complement activation, direct neutralization, and enhanced antigen presentation capacity^24–26. Conversely, a dysfunctional GCB fails to generate effector B cells, compromising the maintenance of follicular helper T cells (Tfh) and effector T cells²⁷. In gastric cancer, B cells have been implicated in disease progression and immune regulation²⁸, with TLS-resident B cells proven to enhance anti-PD-1 therapy response. However, the functional status of GCB in gastric cancer, its relationship with anti-PD-1 treatment response, and whether GCB directly participate in antitumor effector functions remain incompletely understood.

Here, we show that functional divergence of TLS is spatially determined: the muscularis propria serves as a dysfunctional TLS niche in anti-PD-1-refractory gastric cancer, where PLA2G2A‑expressing tumor cells drive the accumulation of structurally disorganized and functionally impaired mature TLS (mTLS) through CXCL14⁺ inflammatory cancer‑associated fibroblast (iCAF), leading to defective germinal center reactions and compromised immunotherapy efficacy. These findings identify compartmental TLS heterogeneity as a potential marker for anti-PD-1 response and nominate PLA2G2A as a potential therapeutic target to restore TLS function and improve gastric cancer immunotherapy.

Results

Muscularis TLS accumulation associates with anti-PD-1 resistance in GC

To characterize clinicopathological features relevant to immunotherapy response, we enrolled 71 advanced GC patients receiving anti-PD-1 therapy. All patients received neoadjuvant anti-PD-1 agents combined with platinum‑based chemotherapy doublets before successful surgery, and all were treatment-naïve prior to neoadjuvant therapy. Patients who did not undergo surgery after neoadjuvant treatment were excluded from this cohort. Tumor tissues were surgically resected within 4-12 weeks after the last dosing. Tumor dynamics were monitored at pre-treatment and pre-surgery using MRI/PET-CT (Supplementary Fig. 1a). Patients were categorized according to Tumor Regression Grade (TRG) criteria: complete response (CR, TRG = 0, 17 patients), partial response (PR, TRG = 1 or TRG = 2, 20 patients), and progressive disease (PD, TRG = 3, 34 patients) (Fig. 1a). Demographic analysis revealed a predominantly male cohort (73.24 %) with low differentiation gastric cancer (35.21 %), diffuse type carcinoma (45.07 %), and absence of distant metastasis (74.65 %). Notably, patients in the PD group exhibited the most advanced or aggressive disease characteristics (Supplementary Table 1).

Fig. 1. Correlation of muscularis TLS with anti-PD-1 therapy resistance.

a Schematic illustration of gastric cancer patients treated with anti-PD-1, categorized by therapeutic response. Created in BioRender. Han, Y. (2026) https://BioRender.com/2ot4iue. b Histopathological images (H&E staining) of gastric tumor tissues from CR, PR and PD patients. The gastric tissue layers (Mucosa, Submucosa, Muscularis, and Serosa) were labeled. Black solid lines outline the TLS. Yellow boxes indicate the zoomed region. c Representative images of adjacent sections from the same CR, PR, and PD patient samples shown in Fig. 1b stained with CD3 (green), CD20 (red), CD23 (blue), and DAPI (gray). The leftmost column shows low-magnification overviews with dashed lines demarcating tissue layers. White boxes indicate the zoomed region. d Quantitation of total TLS density (number per mm²) across anatomical locations (n = 71 patients per group). e Quantitation of total TLS density (number per mm²) across CR, PR, and PD groups (CR: 17 patients; PR: 20 patients; PD: 34 patients). f Quantitation of total TLS density (number per mm²) in Muc & Sub versus Mus & Ser for each response group (CR: 17 patients; PR: 20 patients; PD: 34 patients). g Paired comparison of total TLS density (number per mm²) across anatomical locations. Left panel: CR patients (n = 17); Middle panel: PR patients (n = 20); Right panel: PD patients (n = 34). h Kendall’s rank correlation analysis of total TLS density and area with anti-PD-1 immunotherapy resistance across anatomical locations. Two-tailed Kendall’s tau coefficients and corresponding p-values were calculated via asymptotic normal approximation. Each dot represents one patient in (d–g). Data in (d–f) are presented as mean ± s.e.m., analyzed for significant differences by two-tailed unpaired Student’s t-test (d), two-tailed paired Student’s t-test (g) or one-way ANOVA with Tukey’s post hoc analysis for multiple comparisons (e, f). (CR complete response, PR partial response, PD progressive disease, TLS tertiary lymphoid structures, Muc & Sub Mucosa & Submucosa, Mus & Ser Muscularis & Serosa).

Of note, accumulating evidence has highlighted the critical role of tertiary lymphoid structures (TLS) in antitumor immunity^8,9,29. To assess the pathological correlation between TLS distribution and immunotherapy response, we integrated H&E-based morphological assessment (Fig. 1b) with multiplex immunofluorescence staining for CD3, CD20, and CD23 (Fig. 1c); and aggregates of T cells and B cells (at least 50 cells) were identified as TLS^8,9,30. Using these criteria, we evaluated the spatial distribution of TLS within mucosa, submucosa, muscularis propria, and serosa across CR, PR, and PD groups by quantifying TLS density (number per mm²) and total TLS area per patient (Fig. 1b, c).

Overall, TLS density and area did not differ significantly between the mucosa/submucosa and the muscularis/serosa (Fig. 1d, e), nor among CR, PR, and PD patients (Supplementary Fig. 1b, c). However, stratification by anatomical location revealed a distinct pattern: in the muscularis/serosa, both TLS density and area were higher in PD patients than in CR or PR patients, whereas no differences existed across response groups in the mucosa/submucosa (Fig. 1f and Supplementary Fig. 1d). Moreover, paired comparison within individual tissue sections further showed that in PD group, TLS density was significantly higher in the muscularis/serosa than in the mucosa/submucosa, whereas the opposite held true for CR and PR patients (Fig. 1g). Furthermore, Kendall’s rank correlation analysis confirmed a strong positive association between muscularis TLS accumulation and resistance to anti-PD-1 therapy (density: τ = 0.4790, p < 0.0001; area: τ = 0.3280, p = 0.0006; Fig. 1h). Thus, muscularis TLS accumulation is a spatial-pathological feature of anti-PD-1 resistance in GC.

Muscularis TLS exhibits disrupted architecture and impaired germinal centers

Given the established correlation between TLS maturity and immunotherapy outcomes⁸, we evaluated TLS transcriptional maturation signature using spatial transcriptome sequencing (10× Visium) on full-thickness surgically resected samples (PR, 2 samples; PD, 2 samples) from anti-PD-1-treated GC patients. Using a TLS maturity-related gene set¹⁷ (Supplementary Table 2), we found that TLS in PD sections displayed an immature state (lower mature TLS score, higher immature TLS score) compared to PR sections (Fig. 2a and Supplementary Fig. 2a). Spatially, TLS in the muscularis/serosa also exhibited an immature state (Supplementary Fig. 2b). Integrating spatial distribution with clinical outcomes, we observed that in the muscularis/serosa, PD samples showed significantly higher immature TLS scores than PR samples, whereas an opposite pattern was observed in the mucosa/submucosa (Fig. 2b).

Fig. 2. Muscularis TLS in non-responders exhibit impaired GCB zones and FDC deficiency.

a Representative images showing mature and immature TLS score in spatial transcriptomics of PR and PD patients. The gastric tissue layers (Mucosa, Submucosa, Muscularis, and Serosa) were labeled. Black solid lines outline the TLS. b Quantitation of mature and immature TLS scores across anatomical locations of PR and PD patients. c Representative images of adjacent sections from the same PR and PD patient samples shown in Fig. 1b stained with CD20 (red), Ki67 (blue), BCL6 (pink), and DAPI (gray). The leftmost column shows low-magnification overviews with dashed lines demarcating anatomical locations. White boxes indicate the zoomed region. d Proportion of GCB (BCL6⁺KI67⁺CD20⁺ cells) among total DAPI⁺ cells in TLS regions of PR and PD patients (n = 11 patients per group). e Representative images of adjacent sections from the same PR and PD patient samples shown in Fig. 1b stained with CD4 (green), CXCL13 (red), CD21 (blue), CD35 (pink), and DAPI (gray). The leftmost column shows low-magnification overviews with dashed lines demarcating anatomical locations. White boxes indicate the zoomed region. f Proportion of CD35⁺ cells among total DAPI⁺ cells in TLS regions of PR and PD patients (n = 11 patients per group). g Proportion of CD21⁺ cells among total DAPI⁺ cells in TLS regions of PR and PD patients (n = 11 patients per group). h Proportion of CD4⁺ CXCL13⁺ cells among total DAPI⁺ cells in TLS regions of PR and PD patients (n = 11 patients per group). Each dot represents one patient in (d, f, g, h). Data in (d, f, g, h) are presented as mean ± s.e.m., analyzed for significant differences by two-tailed, unpaired Student’s t-test. Data in (b) are analyzed for significant differences by two-tailed, one-way ANOVA with Tukey’s post hoc analysis for multiple comparisons. (PR partial response, PD progressive disease, Muc & Sub Mucosa & Submucosa, Mus & Ser Muscularis & Serosa, TLS tertiary lymphoid structures, mTLS mature TLS, iTLS immature TLS, GCB germinal center B cell, FDC follicular dendritic cells).

Corroborating these transcriptional findings, multiplex immunofluorescence on adjacent sections showed that, despite a larger total B cell (CD20⁺) area within muscularis TLS of PD patients (Supplementary Fig. 2c, d), proliferating germinal center B cells (GCB, CD20⁺Ki67⁺BCL6⁺) were markedly reduced (Fig. 2c, d), along with diminished follicular helper T (Tfh) cells (CD4⁺CXCL13⁺, we note CXCL13 alone is not a Tfh-specific marker³¹ and was not used to categorize TLS mature state) and follicular dendritic cells (FDC, CD21⁺ and CD35⁺) (Fig. 2e–h). Based on established criteria^17,32, we classified TLS into immature (iTLS, diffuse aggregates lacking germinal centers and FDC networks) and mature (mTLS, distinct follicular architecture with well-formed germinal centers). Quantification across the full cohort (CR, 17 patients; PR, 20 patients; PD, 34 patients) showed that mTLS density was lower in the muscularis/serosa than in the mucosa/submucosa (Supplementary Fig. 2e), with no overall difference among response groups (Supplementary Fig. 2f). However, integrating spatial distribution with clinical outcomes, PD samples displayed reduced mTLS density relative to PR patients (Supplementary Fig. 2g). In contrast, iTLS density was significantly higher in PD patients, specifically in the muscularis/serosa (Supplementary Fig. 2h–j), as confirmed by paired and correlation analyses (Supplementary Fig. 2k–m). Thus, muscularis TLS in PD samples are transcriptionally immature and structurally defective.

B cells in muscularis TLS arrest in a naïve state

To explore the apparent paradox of B cell enrichment despite a lack of germinal center B cells within muscularis TLS, we subsequently performed single-cell RNA sequencing (10× Genomics) on fresh surgical specimens from 26 GC patients (CR: n = 6, PR: n = 9, PD: n = 11) (Fig. 3a). After rigorous quality control, 137,456 single cells were assigned into 10 major lineages via canonical marker genes (Fig. 3b and Supplementary Fig. 3a, b). B cell subpopulations were subsequently clustered and annotated using signature genes (Fig. 3c and Supplementary Fig. 3c). Two subpopulations, IGHD^high SELL^high B cells and IGHD^high YBX3⁺ B cells, highly expressed naïve B cell markers (e.g., IGHM, IGHD, TCL1A) and showed higher naïve B signature scores (Fig. 3d, Supplementary Fig. 3c, and Supplementary Table 2). Notably, both subpopulations were significantly enriched in PD patients (Fig. 3e) and exhibited a gene signature suggestive of impaired B cell differentiation (Supplementary Fig. 3d), suggesting that B cells in PD patients remain in a naïve state.

Fig. 3. Single-cell and spatial transcriptomic profiling identified enrichment of naïve B cells within the muscularis TLS in non-responders.

a Schematic illustration of the single-cell RNA sequencing landscape (CR: n = 6; PR: n = 9; PD: n = 11). Created in BioRender. Han, Y. (2026) https://BioRender.com/2ot4iue. b UMAP of 26 GC patients post anti-PD-1 therapy, colored by cell types, including mast cells (Mast), neutrophils (Neu), T cells (T), B cells (B), plasma cells (PC), macrophages/dendritic cells/monocytes (Mφ&DC&Mono), epithelial cells (EP), endothelial cells (EC), and cancer-associated fibroblasts (CAF). c UMAP of B cell subsets, colored by annotated subtypes (left) and anti-PD-1 response (right). d Quantitation of naïve B scores across B subsets. Boxes: 25th, 50th, 75th percentiles; Whiskers: data minima and maxima. e Stacked bar charts illustrating the proportion of each B cell subset relative to total cells across patient groups. f Representative images showing IGHD^high SELL^high and IGHD^high YBX3⁺ B cell scores in spatial transcriptomics of PR and PD patients. g Quantitation of IGHD^high SELL^high and IGHD^high YBX3⁺ B cell scores in PR and PD patients across anatomical locations. h Representative images of adjacent sections from the same PR and PD patient samples shown in Fig. 1b, stained with CD20 (green), IGHD (red), and DAPI (gray) (upper). Proportion of IGHD⁺ CD20⁺ cells among total DAPI⁺ cells in TLS regions was calculated (lower). Each dot represents one patient (n = 11 patients per group). i Pseudotime trajectory analysis of B cell subsets, colored by B cell subtypes and anti-PD-1 response. Data in (h) are presented as mean ± s.e.m. Data in (g, h) are analyzed for significant differences by two-tailed, unpaired Student’s t-test. (CR complete response, PR partial response, PD progressive disease, TLS tertiary lymphoid structures, Muc & Sub Mucosa & Submucosa, Mus & Ser Muscularis & Serosa).

To spatially assess B cell maturity, spatial transcriptomics data were deconvoluted using scRNA-seq datasets as a reference. Region‑specific analysis revealed that signature scores of the IGHD^high SELL^high B and IGHD^high YBX3⁺ B subpopulations were elevated within muscularis TLS of PD patients (Fig. 3f, g). Co-staining for CD20 and IGHD further confirmed that B cells within muscularis TLS resided in a naïve state (Fig. 3h). Moreover, pseudotime trajectory analysis identified these two subpopulations as early developmental states that had not yet differentiated into activated effector B cells in non-responders (Fig. 3i). Together, these analyses demonstrate that B cells within muscularis TLS of PD patients are arrested in a naïve, functionally impaired state and fail to differentiate into activated effector cells.

PDGFRA⁺ iCAFs contribute to naïve B cells recruitment in the muscularis TLS

To investigate why B cells accumulate in the muscularis propria of PD patients, we first focused on B cell recruitment. Chemokine receptors CXCR4 and CCR7 were universal expressed in B cell subsets (Fig. 4a), and CXCR4 staining confirmed its presence on B cells (Supplementary Fig. 4a). To identify the cellular source of the corresponding ligands, we profiled chemokines (CXCL12, CXCL14, CCL19, CCL21) in the tumor microenvironment (TME). Cancer-associated fibroblast (CAF) emerged as the dominant source of CXCL12 and CXCL14 (Fig. 4b). Based on CAF subpopulation signature gene profiles (Fig. 4c and Supplementary Fig. 4b), PDGFRA⁺ inflammatory CAFs (iCAF) were identified, which displayed the greatest iCAF signature scores and was markedly enriched in PD patients. (Fig. 4d, Supplementary Fig. 4c, and Supplementary Table 2). Notably, cell chemotaxis pathways were significantly enriched within PDGFRA⁺ iCAFs (Supplementary Fig. 4d), which expressed the CXCR4 ligands CXCL12 and CXCL14 (Fig. 4e). However, although CXCL12 is the canonical CXCR4 ligand, its expression was lower in PD patients, whereas CXCL14 was significantly upregulated (Fig. 4f).

Fig. 4. PDGFRA⁺ iCAFs recruit naïve B cells to the muscularis TLS via CXCL14-CXCR4 signaling.

a Violin plot showing transcript expression levels of chemokine receptors in B cells from scRNA data. b Dot plot showing transcript expression levels of the CXCR4 ligand (CXCL12, CXCL14, CCL19, CCL21) across main cell types. c UMAP of CAF subsets, colored by annotated subtypes (left) and anti-PD-1 response (right). d Stacked bar charts illustrating the proportion of each CAF subset relative to total cells across patient groups. e Transcript expression levels of CXCL12 and CXCL14 across CAF subpopulations. f Violin plot showing transcript expression levels of CXCL12 and CXCL14 in CAF across CR, PR and PD patients. Boxes: 25th, 50th, 75th percentiles; Whiskers: data minima and maxima. Wilcoxon rank-sum test. g Representative images showing distribution of PDGFRA⁺ iCAFs, CXCL12, and CXCL14 via spatial transcriptomics in PR and PD patients. h Representative images of the same sections from the PR and PD patient samples shown in Fig. 3h stained with CD20 (green), PDGFRA (red), and DAPI (gray). i Experimental workflow for transwell co-culture assay of control or CXCL14-expressing GM02008 cells (human fibroblast) with sorted B cells from human peripheral blood (hPB B cell). Created in BioRender. Han, Y. (2026) https://BioRender.com/2ot4iue. j Representative flow cytometry plots of CD19⁺ B cells collected from the bottom well after co-culture. k Quantification of B cell migration, defined as the percentage of migrated B cells in total input B cells (migrated/[unmigrated + migrated] × 100%) after co-culture (n = 4 biological repeats per group). Data in (k) are presented as mean ± SD, analyzed for significant differences by two-tailed, unpaired Student’s t-test. Data in (f) are analyzed for significant differences by two-tailed, one-way ANOVA with Tukey’s post hoc analysis for multiple comparisons. (T T cells, Neu neutrophils, EP epithelial cells, B B cells, Mφ&DC&Mono macrophages/dendritic cells/monocytes, PC plasma cells, EC endothelial cells, CAF cancer-associated fibroblasts, Mast mast cells, CR complete response, PR partial response, PD progressive disease, iCAF inflammatory CAF, myCAF myofibroblastic CAF, hPB human peripheral blood).

Next, we assessed the spatial distribution of PDGFRA⁺ iCAFs and CXCL14. Spatial transcriptomics revealed that both were localized in the muscularis around TLS regions in PD patients (Fig. 4g and Supplementary Fig. 4e). Co-staining of CD20 and PDGFRA further validated the enrichment of PDGFRA⁺ cells around muscularis TLS in PD patients (Fig. 4h). To functionally confirm that CAF-derived CXCL14 acts as a potential chemoattractant for B cells, we performed a transwell co-culture assay using fibroblasts and B cells for 4 h. NIH3T3 cells (a mouse embryonic fibroblast cell line) expressing control (NIH3T3^Ctrl) or CXCL14 (NIH3T3^OE-CXCL14) cells were cultured alone for 48 h, and mouse splenic B cells were then added to the transwell inserts (3 μm pore size) for a 4 h migration assay (Supplementary Fig. 4f). Compared to the control group (5.59% migration), CXCL14-overexpressing fibroblasts significantly promoted B cell migration during 4 h co-culture (Supplementary Fig. 4g–i). Consistent results were obtained by using GM02008 cells (a human fibroblast cell line) and primary peripheral blood B cells from healthy donors (hPB) (Fig. 4i–k and Supplementary Fig. 4j).

PLA2G2A⁺ tumor cells shape an immunosuppressive niche around muscularis TLS

To investigate the cellular machinery driving PDGFRA⁺ iCAF development, we performed cell-cell interaction analysis, which identified epithelial cell-CAF crosstalk as the dominant network in the TME (Fig. 5a). Co-immunostaining demonstrated spatial proximity of KRT8⁺ tumor cells with PDGFRA⁺ iCAFs in PD patients (Supplementary Fig. 5a). This prompted an in-depth characterization of tumor cell heterogeneity. Therefore, tumor cells were subclustered into eleven distinct subsets. One PLA2G2A⁺ tumor cell subset within these populations was substantially expanded in PD patients (Fig. 5b, c). Transcriptional signatures of cell proliferation (MKI67, CCND2) and immune suppression were detected in this subset (Fig. 5d).

Fig. 5. PLA2G2A⁺ TCs were enriched within the muscularis TLS in non-responders.

a Quantified ligand-receptor interactions among distinct cell types, including epithelial cell (EP), macrophage/dendritic cell/monocyte (Mφ/DC/Mono), T cells, cancer-associated fibroblasts (CAF), and B cells. b UMAP of epithelial cell subsets, colored by annotated subtypes (upper) and anti-PD-1 response (lower). c Stacked bar charts illustrating the proportion of each epithelial cell subset relative to total cells across patient groups. d Expression levels of selected signature genes in epithelial cell subsets (left) and GSVA score showing their biological properties including immune suppression, lipid metabolism and cell proliferation (right). Wilcoxon rank-sum test. Boxes: 25th, 50th, 75th percentiles; Whiskers: data minima and maxima. e Representative images showing PLA2G2A expression and distribution of PLA2G2A⁺ TC via spatial transcriptomics in PR and PD patients. f H&E staining (upper) and immunofluorescent staining for CD20, CD3, KI67 (middle) and PLA2G2A, KRT8, αSMA (lower) on serial sections in PR and PD patients. g Proportion of KRT8⁺ PLA2G2A⁺ TC in KRT8⁺ TC (upper) and proportion of Ki67⁺ TC in PLA2G2A⁺ KRT8⁺ TC (lower) across PR and PD patients. Each dot represents one patient (n = 13 patients per group). Data in (g) are presented as mean ± s.e.m., analyzed for significant differences by two-tailed, unpaired Student’s t-test. (T T cells, EP epithelial cells, B B cells, Mφ&DC&Mono macrophages/dendritic cells/monocytes, CAF cancer-associated fibroblasts, TC tumor cells, CR complete response, PR partial response, PD progressive disease, Mus & Ser Muscularis & Serosa).

Phospholipase A2 group IIA (PLA2G2A) is a secreted enzyme belonging to the secretory phospholipase A2 (sPLA2) family, which hydrolyzes glycerophospholipids to generate lysophospholipids and free fatty acids, including arachidonic acid-a key precursor for prostaglandins, thromboxanes, and leukotrienes³³. Beyond these canonical metabolic functions, sPLA2 family members have been increasingly implicated in tumor immune modulation. PLA2G2A has been reported to exert context-dependent effects, ranging from tumor suppression in certain settings to promoting immune evasion in others^34–37. In pancreatic cancer, PLA2G2A⁺ cancer-associated fibroblasts have been shown to suppress CD8⁺ T cell function^38,39. However, the role of epithelial-derived PLA2G2A in shaping the TLS microenvironment and its impact on B cell immunity in gastric cancer has not been previously explored.

To address this, we characterized the spatial distribution of PLA2G2A⁺ TCs. Spatial transcriptomic analysis revealed that PLA2G2A⁺ TCs localized around the muscularis TLS in PD patients (Fig. 5e). H&E staining and immunostaining for αSMA, PLA2G2A and KRT8 further confirmed that PLA2G2A⁺ TCs and fibroblasts were significantly enriched around the muscularis TLS (Fig. 5f, g). Single-cell transcriptomic analysis further revealed several immunosuppressive molecules highly expressed by PLA2G2A⁺ TCs, including CD47^40–42, CEACAM1^43,44, PVR^45–47 (Supplementary Fig. 5b). Spatially, in the muscularis region of the PD group, the expression levels of PLA2G2A, CD47, CEACAM1 and PVR were significantly higher than those in the PR group in the muscularis region of PD group (Supplementary Fig. 5c). Consistently, immunosuppressive signature scores were elevated in the muscularis of PD patients (Supplementary Fig. 5d), indicating that PLA2G2A⁺ tumor cells are associated with an immunosuppressive niche around TLS in muscularis.

Tumor-secreted PLA2G2A induces iCAF phenotype and CXCL14 expression

To elucidate the putative role of PLA2G2A⁺ tumor cells in the muscularis TLS of PD patients, we first measured PLA2G2A levels in peripheral blood and tumor tissues from patients with different therapeutic responses. Indeed, both serum and tumor microenvironment PLA2G2A levels were significantly elevated in PD patients compared to CR/PR patients (PD vs CR/PR, p < 0.0001; 95% confidence interval for the difference: 190.0 to 312.3 pg/mL) (Fig. 6a and Supplementary Fig. 5e). To determine whether tumor-derived PLA2G2A induces PDGFRA⁺ iCAF phenotype, we treated NIH3T3 cells with purified recombinant Flag-tagged PLA2G2A protein for 48 h (Fig. 6b). Bulk RNA-seq analysis revealed that PLA2G2A treatment upregulated 455 genes and downregulated 60 genes compared to controls (Supplementary Fig. 6a). Gene Set enrichment Analysis (GSEA) showed that the signature of iCAF, but not the that of myCAF or apCAF, was enriched in the PLA2G2A-treated group (Fig. 6c and Supplementary Fig. 6b, c).

Fig. 6. PLA2G2A promotes PDGFRA/CXCL14 expression in fibroblasts.

a Concentration of PLA2G2A in serum measured by ELISA in CR, PR and PD patients. Each dot represents one patient (CR: n = 21; PR: n = 57; PD: n = 67). b Coomassie Brilliant Blue (CBB) staining of the purified recombinant Flag-PLA2G2A protein. c Schematic showing bulk RNA-Seq workflow of NIH3T3 cells treated with Flag-PLA2G2A for 48 h (Left). Created in BioRender. Han, Y. (2026) https://BioRender.com/2ot4iue. Subsequently, GSEA enrichment of the iCAF signature genes in NIH3T3 cells treated with Flag-PLA2G2A were analyzed (Right) (n = 3 biological repeats per group). P-value was estimated by a one-sided permutation test. d CXCL14 mRNA expression in NIH3T3 cells treated with 1 μg/mL and 10 μg/mL Flag-PLA2G2A protein (n = 3 biological repeats per group). e Western blot showing CXCL14 expression in NIH3T3 cells treated with 1 μg/mL and 10 μg/mL Flag-PLA2G2A protein. f Representative images of NIH3T3 cells treated by Flag-PLA2G2A protein, with or without Flag beads. g ACTA2, PDGFRA, CXCL14 mRNA expression in NIH3T3 cells treated by Flag-PLA2G2A protein, with or without Flag beads (n = 3 biological repeats per group). h PLA2G2A mRNA levels in patient-derived organoids (PDOs) from two gastric cancer patients (840 and 842) (n = 3 biological repeats per group). i Representative images of the IMR-90 cells co-cultured with PDOs expressing low and high levels of PLA2G2A (Left). PDGFRA and CXCL14 mRNA expression were analyzed in IMR-90 cells co-cultured with indicated PDOs (Right) (n = 3 biological repeats per group). Data in (a, d, g) are presented as mean ± s.e.m., analyzed for significant differences by two-tailed, one-way ANOVA with Tukey’s post hoc analysis for multiple comparisons. Data in (h, i) are presented as mean ± SD, analyzed for significant differences by two-tailed, unpaired Student’s t-test. Data are representative of two independent experiments in (b, d–i). (CR complete response, PR partial response, PD progressive disease, CBB Coomassie Brilliant Blue, iCAF inflammatory CAF, PDO patient‑derived organoids).

We then investigated whether PLA2G2A induces CXCL14 expression in PDGFRA⁺ iCAFs. Treatment of recombinant Flag-PLA2G2A proteins upregulated CXCL14 in NIH3T3 cells in a dose-dependent manner at both the mRNA and protein levels (Fig. 6d, e). Conversely, PLA2G2A neutralization (using anti‑Flag beads) abolished the upregulation of both Cxcl14 and Pdgfra (Fig. 6f, g). Importantly, co-culture of IMR90 (a human lung fibroblasts cell lines) with patient-derived organoids (842-PDO) from PD patients (which express high PLA2G2A) robustly induced PDGFRA and CXCL14 expression in fibroblasts, whereas organoids (840-PDO) from PR patients failed to do so (Fig. 6h, i). Together, these findings in vitro demonstrate that tumor cell‑secreted PLA2G2A directly drives the iCAF phenotype and CXCL14 expression.

Tumor-secreted PLA2G2A impairs B cell differentiation and promotes Tfh cell death

To assess a potential regulatory role of PLA2G2A⁺ TCs on muscularis TLS, we first cultured murine splenic B cells with LPS (10 μg/ml) in the presence or absence of recombinant PLA2G2A proteins (50 μg/ml) for 48 h, followed by flow cytometry and RNA-seq analysis (Fig. 7a). LPS stimulation alone induced robust B cell proliferation (Fig. 7b–e) and a marked decrease in the frequency of CD19⁺ IgM⁺ naïve B cells (Fig. 7f, g). However, when LPS was combined with PLA2G2A treatment, B cell proliferation was inhibited (Fig. 7b–e), and the percentage of CD19⁺ IgM⁺ naïve B cells significantly increased (37.2% vs 24.4%, Fig. 7f, g), indicating that PLA2G2A suppresses B cell proliferation and differentiation. RNA‑seq on B cells confirmed that PLA2G2A treatment significantly impaired B cell differentiation at the transcriptome level (Fig. 7h, i).

Fig. 7. PLA2G2A inhibits B cells differentiation.

a Schematic showing sorted CD19⁺ B cells from the murine spleen for flow cytometry and bulk RNA-Seq, treated with 10 μg/mL LPS and 50 μg/mL PLA2G2A for 48 h. Created in BioRender. Han, Y. (2026) https://BioRender.com/2ot4iue. b Representative images of the murine spleen-derived B cells treated with 50 μg/mL purified PLA2G2A protein for 48 h. c Quantitation of the number of B cell treated with PLA2G2A (n = 3 biological repeats per group). d Representative flow cytometry plots of CTV fluorescence intensity in murine splenic B cells, as shown in (b). e Quantitation of proliferative B cell based on d (n = 3 biological repeats per group). f Representative flow cytometry plots showing the proportion of CD19⁺ IgM⁺ cells in murine spleen-derived B cells treated with 50 μg/mL purified PLA2G2A protein for 48 h. g Proportions of IgM⁺ B cells after treatment based on f (n = 3 biological repeats per group). h GO term enrichment pathways of murine spleen-derived B cells following purified PLA2G2A protein treatment for 48 h. i GSEA enrichment of B cell differentiation pathways in murine spleen-derived B cells following purified PLA2G2A protein treatment for 48 h. j Representative flow cytometry plots showing the proportion of IgM⁺ IgD⁺ B cells from gastric cancer patients’ blood after 48 h treatment with supernatant from gastric cancer PDOs of PR and PD patients. k Quantitation of IgM⁺ IgD⁺ B cells based on j, with each dot representing an individual patient (n = 3 biological repeats per group). Data in (c, e, g, k) are presented as mean ± s.e.m., analyzed for significant differences by two-tailed, one-way ANOVA with Tukey’s post hoc analysis for multiple comparisons. Data are representative of two independent experiments. (LPS Lipopolysaccharide, CTV cell trace violet, PR partial response, PD progressive disease, TCM tumor conditioned medium).

Next, we treated human peripheral blood-derived B cells with conditioned medium from PD or PR patient‑derived organoids. PR organoid-conditioned medium effectively reduced naïve B cell frequency (79.3% vs 61.0%, Fig. 7j, k). In contrast, PD organoid-conditioned medium failed to promote differentiation, instead leading to an accumulation of naïve B cells (Fig. 7j, k), demonstrating that the ability to drive B cell differentiation is significantly impaired in PD patients.

Given that Tfh cells are a core driving force for mature TLS formation^9,30,48, we next examined Tfh cells in the same context. We observed that Tfh cells were reduced in PD patients (Supplementary Fig. 7a–c). Furthermore, we found that PLA2G2A promoted Tfh cell death, as demonstrated both by treatment with supernatant from PLA2G2A-overexpressing MFC cells (Supplementary Fig. 7d–f) and by direct addition of recombinant PLA2G2A protein (Supplementary Fig. 7g–i). Together, these findings in vitro reveal that PLA2G2A⁺ tumor cells keep muscularis TLS in an immature and functionally deficient state through two distinct mechanisms: directly suppressing B cell differentiation and promoting Tfh cell death, thereby disrupting the Tfh-B cell axis essential for TLS maturation.

Discussion

Our study reveals the spatial and functional heterogeneity of TLS in gastric cancer and identifies a tumor‑driven immunosuppressive niche within the muscularis propria that impairs TLS maturation and function, thereby potentially contributing to anti-PD-1 resistance. Despite mucosal TLS have been associated with favorable responses to immunotherapy²⁶, we demonstrate that muscularis TLS exhibit structural disorganization, absence of follicular dendritic cell networks, impaired B cell differentiation, and defective germinal centers. This spatial dichotomy-mucosal TLS promoting immunity versus muscularis TLS driving dysfunction-challenges the paradigm of TLS as universally beneficial and underscores anatomic location as a critical functional determinant. The robust inverse correlation between muscularis TLS burden and therapeutic efficacy further positions these structures as stromal potential markers of resistance.

While previous studies have linked sPLA2 family members to intratumoral T cell exclusion³⁷ and CAF-derived PLA2G2A to CD8⁺ T cell dysfunction in pancreatic cancer³⁹, our findings uncover a coordinated stromal‑lymphoid axis of immunosuppression orchestrated by PLA2G2A⁺ tumor cells enriched at the invasive front within the muscularis propria. PLA2G2A⁺ tumor cells simultaneously drive fibroblast differentiation into PDGFRA⁺ iCAFs, which recruit naïve B cells via CXCL14; directly suppress B cell proliferation and differentiation, arresting B cells in a naïve state; and promote Tfh cell death, thereby depriving B cells of essential T cell help. This multi‑pronged attack thus collapses the Tfh-B cell collaborative unit, abrogating the germinal center reaction and locking muscularis TLS into an arrested, immature phenotype.

Our findings prompt a reconsideration of B cell biology within the framework of anti‑PD‑1 therapy. The conventional paradigm rightly emphasizes CD8⁺ T cells as the direct effectors of PD-1/PD-L1 axis blockade. Our model does not challenge this centrality but instead proposes that a functionally mature TLS, housing active germinal center B cells, constitutes a critical permissive microenvironment for amplifying and sustaining an effective T cell response. B cells support anti-tumor T cell immunity via antibody-mediated antigen presentation, cytokine secretion, and co‑stimulation^49–56. Dysfunctional muscularis TLS, with disrupted germinal centers and arrested B cell differentiation, may deprive T cells of local support, leading to therapeutic failure even in the presence of PD-1 blockade. Thus, our results link TLS spatial heterogeneity, B cell differentiation state, and clinical response, offering a rationale for combining PD-1 axis blockade with TLS‑reprogramming strategies.

From a therapeutic perspective, our findings position PLA2G2A as an actionable target for overcoming immune checkpoint blockade resistance. Elevated serum PLA2G2A levels in non‑responders provide a potential non‑invasive stratification tool. Possible targeted interventions include pharmacological inhibition of PLA2G2A (e.g., with varespladib) to disrupt the tumor-iCAF-B cell crosstalk; CXCR4 antagonists to block iCAF-mediated recruitment of naïve B cells; or stromal reprogramming strategies to normalize iCAF function. As an emerging diagnostic innovation, AI‑driven spatial mapping of TLS maturity could guide patient selection for immunotherapy.

Our study has several limitations. First, our cohort lacks peritumoral tissues, limiting quantitative distinction of lymphoid aggregates; pan‑cancer validation is also needed. Second, the mechanistic uncertainties remain regarding the relative roles of TLS versus draining lymph nodes, PLA2G2A function, germinal-center antibody production, and the CXCL14-CXCR4 axis, all requiring in vivo and preclinical investigation. Third, our human data are largely correlative; causality and the efficacy of combinatory regimens (e.g., anti-PLA2G2A plus anti-PD-1) await validation in genetic models and independent cohorts.

Methods

Reagents

The primary antibodies and dilutions used for IHF were anti-KRT8 (ab53280, Abcam, 1:1000), anti-PDGFRα (ab203491, Abcam, 1:500), anti-KI67 (GB121141, Servicebio, 1:500), anti-CD4 (GB300601, Servicebio, 1:1000), anti-CD19 (GB300617, Servicebio, 1:1000), anti-PLA2G2A (PA5-102403, Invitrogen, 1:500), anti-IGHD (HA722433, HUABIO, 1:500), anti-CXCL13 (GT224202, Genetech, 1:500), anti-CD20 (SDT-R133, Starter-bio, 1:2000), anti-CD35 (NBP2-52667, Novus Biologicals, 1:500), anti-CD23 (GB14032-50, Servicebio, 1:1000), anti-CD21 (SDT-007-47, Starter-bio, 1:1000), anti-BCL6 (21187-1-AP, Proteintech, 1:500). The primary antibodies and dilutions used for immunoblotting were anti-Flag (M3165, Sigma, 1:1000), anti-CXCL14 (ab137541, Abcam, 1:1000). The antibodies and dilutions used for flow cytometry were: APC-anti-human CD19 (302212, BioLegend, 1:300), FITC-anti-human IgM (314506, BioLegend, 1:300), APC-anti-human IgD (69-9868-42, Thermo Fisher, 1:300), APC-anti-mouse CD19 (152410, BioLegend, 1:300), FITC-anti-mouse IgM (406506, BioLegend, 1:300), PE-anti-mouse CD4 (100528, Biolegend, 1:300), PerCP/Cyanine5.5-anti-mouse PDCD1 (135208, Biolegend, 1:300), and FITC-anti-mouse CXCR5 (145520, Biolegend, 1:300). The reagents used for in vitro cell culture were recombinant mouse IL-6 protein (RP01321, Abclonal), recombinant mouse IL-21 protein (RP00664, Abclonal). LPS (derived from Escherichia coli, serotype 055: B5, L4005, Sigma). See also Supplementary Data 1.

Plasmids

The pcDNA3.0-HA⁵⁷ and pcDNA3.1-3×Flag⁵⁸ vector have been described previously. The mouse Cxcl14 cDNA was amplified by PCR from mouse spleen cDNA followed by digestion with HindIII and XhoI and inserted into pcDNA3.0-HA to construct mouse Cxcl14 expression plasmids. The mouse Pla2g2a cDNA was amplified by PCR from mouse spleen cDNA followed by digestion with HindIII and EcoRI and inserted into pcDNA3.1-3×Flag to construct mouse Pla2g2a expression plasmids. The human PLA2G2A cDNA was amplified by PCR from AGS cell line cDNA followed by digestion with HindIII and XhoI and inserted into pcDNA3.1-3×Flag to construct human PLA2G2A expression plasmids.

Cells

MFC cells (TCM23), NIH3T3 (SCSP-515) were purchased from the Cell Resource Center of the Institute of Life Sciences, Chinese Academy of Sciences (Shanghai, China). GM02008 cells (CTCC-009-061) were purchased from MeisenCTCC (Zhejiang, China). IMR-90 cells (YC-D050) were purchased from the UBIGENE (Guangzhou, China). MFC and NIH3T3 cells were cultured in DMEM supplemented with 10% Fetal Bovine Serum (FBS) and 1% penicillin/streptomycin. IMR-90 cells were cultured in MEM supplemented with 10% FBS, 1% NEAA and 1% penicillin/streptomycin. GM02008 cells were cultured in DMEM supplemented with 10% FBS, 1% NEAA and 1% penicillin/streptomycin. All cells were cultured under standard conditions (37° and 5% CO₂). During the study, all cell cultures were periodically tested for mycoplasma using MycoAlert™ Mycoplasma Detection Kits.

Isolation of splenic T/B cell

Splenic CD4⁺ T cells and B cells were isolated from 8‑week‑old male C57BL/6 mice (Lingchang Biotechnology, Shanghai, China). Single-cell suspensions were purified by magnetic‑activated cell sorting (MACS) using isolation kits (Miltenyi Biotec, 130-104-454 and 130-090-862, respectively) according to the manufacturer’s instructions. Cell purity (>95%) was confirmed by flow cytometry prior to use. All animal were housed under specific pathogen-free conditions in automated watered and ventilated cages on a 12 hr light/dark cycle, and sacrificed via carbon dioxide inhalation. All animal procedures were approved by the Institutional Animal Care and Use Committee of Fudan University (IDM2022037).

Patient samples collection

Patients enrolled in our study were those with newly diagnosed, locally advanced gastric cancer (clinical stage II-III) who received neoadjuvant anti-PD-1 therapy combined with platinum-based doublet chemotherapy prior to radical gastrectomy. Patients who did not undergo surgery after neoadjuvant treatment were excluded. All patients were treatment‑naïve before initiating neoadjuvant therapy. Anti-PD-1 agents were restricted to Pembrolizumab, Nivolumab, Sintilimab, Camrelizumab, or Tislelizumab and were delivered once every three weeks according to their manufacturers’ instructions. The majority of patients received two to four cycles of neoadjuvant therapy and underwent surgery within 4-12 weeks after the last dose, consistent with the Chinese expert consensus on perioperative immunotherapy for advanced gastric cancer. Pre‑treatment and pre‑surgery CT scans were used as references to determine tumor location and size. Patients were categorized by TRG criteria: complete response (CR, ‌TRG = 0), partial response (PR, TRG = 1 or TRG = 2), and progressive disease (PD, TRG = 3). Tissue specimens were obtained from resected primary gastric cancers, encompassing the full thickness from the mucosa to the serosa. All human blood and tumor samples used in this study were collected from participants enrolled in a clinical trial (Protocol ID: 050432-4-2108*) approved by the Ethics Committee of Fudan University Shanghai Cancer Center. Written informed consent was obtained from all patients prior to sample collection.

TLS maturity definition

Lymphocyte aggregates containing at least 50 T and B cells were identified as TLS³⁰ based on H&E and cellular composition validated by multiplex immunofluorescence. Specifically, we used the following marker panels to identify distinct subsets: pan-T cells (CD3), pan-B cells (CD20), activated B cells (co-stained with CD23/CD20, CD23, the low-affinity receptor for IgE that mediates B cell differentiation and humoral immune responses^59,60, has been reported to be expressed in activated B cells and follicular dendritic cells within TLS^8,30,), germinal center B cells (GCB; co-stained with Ki-67/BCL6/CD20)^56,61–63, follicular dendritic cells (FDC; co-stained with CD21/CD35)^8,9,64, and follicular helper T cells (Tfh; co-stained with CD4/CXCL13)^65,66. TLS were classified as immature (iTLS) or mature (mTLS). iTLS are defined as diffuse lymphocyte aggregates with indistinct borders, lacking well-defined T/B cell zones, germinal centers, a complete follicular dendritic cell network. mTLS exhibit distinct follicular architecture with a well-formed germinal center, discernible light and dark zones, and the presence of an organized FDC network and high endothelial venules.

For quantification, TLS density and area were compared at the patient level (mean per patient), with each patient as an independent replicate. No pooling across patients was performed prior to group‑level analysis.

H&E staining

All human gastric cancer samples were performed via standard H&E staining. Briefly, tissue specimens were fixed in 4% paraformaldehyde, dehydrated through graded ethanol series, transparent in xylene, and embedded in paraffin blocks. Three-micrometer sections were mounted on adhesive slides, deparaffinized in xylene and rehydrated in descending ethanol gradients. Nuclei were stained with hematoxylin for 10 min, followed by differentiation in 1% acid ethanol. Cytoplasmic counterstaining employed eosin for 60 seconds; sections were dehydrated, cleared in xylene, and coverslipped with permount mounting medium.

ELISA

Peripheral blood samples were collected from immunotherapy-treated gastric cancer patients. Serum was isolated by centrifugation at 2000 × g for 15 min and stored at −80 °C until analysis. Serum concentrations of PLA2G2A were quantified using commercial ELISA kits (EH10703, WEIAOBIO) following the manufacturer’s optimized protocols.

In vitro Tfh differentiation

Murine naïve CD4⁺ T cells were seeded at a density of 1 × 10⁶ cells/ml into plates pre-coated with anti-CD3/CD28 antibodies (1 μg/ml) and maintained in the RPMI-1640 medium with 10% FBS supplemented with IL-6 (10 ng/ml) and IL-21 (10 ng/ml). After 5 days of induction, the cells were harvested for subsequent analysis.

Peripheral blood B cell isolation

Patient peripheral blood was incubated with red blood cell lysis buffer (B541001, Sangon Biotech) for 5 min at room temperature and washed twice with PBS, then cells were resuspended in PBS at a concentration of 2 × 10⁷ cells/mL. Cells were incubated with APC-anti-human CD19 antibody (302212, BioLegend, 1:200) for 15 min at 4 °C. Prior to sorting, doublets were excluded based on forward scatter (FSC) and side scatter (SSC) parameters, and dead cells were excluded by DAPI staining. CD19⁺ B cells were subsequently sorted from the remaining single-cell suspension by performing fluorescence-activated cell sorting (FACS) using a MoFlo XDP cell sorter (Beckman Coulter). Post-sort analysis confirmed a purity of >98% for the isolated CD19⁺ B cell population. Patient‑derived blood B cells were cultured in RPMI 1640 containing 10% FBS under standard conditions. Purified B cells were cultured in RPMI 1640 supplemented with 10% FBS and used for subsequent assays. For co-culture experiments, these B cells were incubated in RPMI 1640 supplemented with 10% FBS and 20% organoid supernatant. Following 48 h of co-culture, B cells were collected for flow cytometric analysis.

Co-culture assay

Fibroblasts (NIH3T3 or GM02008) were transfected with control vector or CXCL14-HA plasmids. After 48 h, isolated mouse or human B cells were then added to the transwell inserts (3 μm pore size). The co cultures were incubated under standard conditions (37° and 5% CO₂) for 4 h. The bottom cells were then collected for flow cytometry analysis.

Gastric cancer organoids culture

Fresh gastric cancer tissues (approximately 2-3 cm³) were rinsed with cold PBS, then minced into 1-2 mm³ fragments. After washing, the fragments were digested in Advanced DMEM/F12 containing Collagenase II (1 mg/mL), DNase I (100 μg/mL), and Y-27632 (10 μM) for 15 min at 37 °C. Digestion was stopped with Advanced DMEM/F12 containing 10% FBS, followed by vigorous pipetting, and centrifugation at 300 × g at 4 °C for 5 min. Thereafter, the pelleted cells were resuspended in Matrigel (40 μL/well) and seeded into 24‑well plates. After gel polymerization, 500 μL of gastric organoid medium (Advanced DMEM/F12 supplemented with penicillin/streptomycin (1%), GlutaMAX (1×), B27 (1×), N2 (1×), nicotinamide (10 mM), N-acetylcysteine (1 mM), A83-01 (0.5 μM), SB202190 (0.5 μM), gastrin-1 (10 nM), Y-27632 (10 μM), EGF (50 ng/mL), Wnt-Ligand (50 ng/mL), R-spondin1 (0.5 μg/mL), Noggin (0.1 μg/mL), and FGF10 (0.1 μg/mL)) was added, and cultures were maintained at 37 °C with 5% CO₂. Medium was changed every 2 days, organoids passaged every 7-8 days, and supernatant collected during culture. Detailed reagent information is available in Supplementary Data 1.

Flow cytometry analysis

All flow cytometric analyses were performed on a BD LSRFortessa™ X-20 cytometer, and data were analyzed using FlowJo software. For surface marker staining, cells were incubated with specific antibodies diluted in PBS containing 1% FBS for 30 min on ice. After staining, cells were washed twice with ice-cold PBS and analyzed immediately. For proliferation assays, cells were stained with CellTrace™ Violet (C34557, Invitrogen, 1:500) for 20 min at 37 °C, washed twice, and then cultured for 48 h at 37 °C prior to analysis. The detailed gating strategy is illustrated in Supplementary Fig. 8.

Multi-color immunohistochemistry

Paraffin-embedded gastric cancer sections were subjected to a sequential rehydration process. First, they were immersed in xylene I, followed by xylene II (1:1 mixture of xylene and 100% ethanol), and then passed through a graded series of ethanol solutions (100%, 95%, and 75%). Antigen retrieval was carried out by placing the sections in citrate buffer (pH 6.0) in a 100 °C water bath for 20 min. To block endogenous peroxidase activity, the sections were treated with 3% H₂O₂ at room temperature for 15 min. Subsequently, nonspecific binding was blocked using 5% bovine serum albumin (BSA) for 30 min. The sections were then incubated with indicated antibodies overnight at 4 °C. After that, they were incubated with corresponding secondary antibodies for 1 h at room temperature. Finally, fluorescent signals were visualized using a confocal laser scanning microscope. Quantification of multiplex immunohistochemistry staining was performed on representative tissue sections from a subset of patients within each response group. For each marker, 3-5 high-power fields were analyzed per section, and the mean value was calculated for each patient.

Immunoblotting

Cells were lysed in NETN buffer (20 mM Tris-HCl pH 8.0, 100 mM NaCl, 0.5% NP-40, 1 mM EDTA) supplemented with protease and phosphatase inhibitors. Equal amounts of protein were resolved on 8-12% SDS-PAGE gels and electrophoretically transferred to PVDF membranes. Membranes were blocked with 5% non-fat milk in PBS for 1 h at room temperature. Indicated antibodies diluted in 5% BSA (1:1000-1:5000) were applied overnight at 4 °C. After three 10-min washes with PBST, membranes were incubated with HRP-conjugated secondary antibodies (1:5000) for 1 h at room temperature. Signals were developed using ECL substrate (36208ES76, Yeasen) and captured on an Tanon 5200.

Real-time PCR

Total RNA was isolated from cellular samples using Vezol reagent (R411-01, Vazyme) following manufacturer’s protocol. Reverse transcription was performed with HiScript II Q RT SuperMix (R223-01, Vazyme) to generate cDNA. Quantitative real-time PCR (qRT-PCR) analysis was conducted on a StepTwo Real-Time PCR System (Bio-Rad, Hercules, CA) using the comparative ΔΔCt method, with β-actin serving as an endogenous control. All experiments included triplicate biological replicates with triplicate technical reactions per sample. The primers were designed in PrimerBank⁶⁷. Primers used for real-time PCR are listed in Supplementary Table 3.

Single-cell sequencing

The scRNA-seq samples were derived from fresh gastric tumor tissues which were transmural, cross-sections and including tumor. And these surgical samples were obtained from patients who had received anti-PD-1 treatment. All subsequent steps were performed following the standard manufacturer’s protocol. Single-cell RNA-seq libraries were analyzed using an Illumina Hiseq X Ten sequencer with 150-base pair (bp) paired-end reads. Raw data processing included demultiplexing with bcl2fastq v2.20, alignment using STAR v2.7.10a, and UMI counting via Cell Ranger v7.1.0 against the GRCh38-2020-A reference. Downstream analysis in Seurat v5.0.1⁶⁸ involved quality control filtering (retaining cells with 200-6000 genes and <20% mitochondrial reads, removing genes detected in <3 cells), log-normalization (scale factor 10,000), dimensionality reduction by Principal Component Analysis, graph-based clustering with the Louvain algorithm (resolution = 0.8), and visualization using Uniform Manifold Approximation and Projection (neighbors = 20). Samples with a B cell count less than 20 were not included in further subpopulation analysis.

Spatial transcriptomics analysis

The spatial RNA-seq samples were obtained as mentioned above. The samples were transmural, cross-sections that including tumor. Tissues were cut into 3-5 mm³ pieces, and then pieces were embedded in paraffin blocks. The Visium Spatial Gene Expression Slide & Reagent kit (10× Genomics) was used to generate sequencing libraries. All subsequent steps were performed following the standard manufacturer’s protocol. Spatial transcriptomics data were processed using the Space Ranger FFPE pipeline (v2.1.0). Raw sequencing reads were demultiplexed, incorporating slide barcode error correction (allowing a Hamming distance of ≤2). Probe-based alignment was performed against the pre-mRNA reference genome (GRCh38-2020-A). Degradation artifacts were digitally deconvolved using penalized matrix factorization (λ = 0.7). Gene-spot matrices underwent quality control, excluding spots with fewer than 500 detected genes or mitochondrial counts exceeding 30%. Data normalization was performed using SCTransform (v2.4.0)⁶⁹.

GSVA score analysis

Gene set variation analysis (GSVA) was performed using the GSVA package⁷⁰. Gene sets were obtained from the MSigDB database. Count data were used as the expression matrix, and the “Gaussian” kernel was selected for the kcdf parameter. GSVA scores were calculated for each spatial spot, and each spot was treated as an independent unit for statistical comparisons between groups. The same approach was applied to single-cell RNA-seq data, where GSVA scores were computed per cell and compared across groups with appropriate multiple‑testing correction.

Pseudotime trajectory reconstruction

Pseudotime trajectory reconstruction was performed using Monocle2 (v2.3.0)⁷¹. The annotated gene expression matrix was imported into R, converted to a CellDataSet object using new_cell_data_set (distribution=negative binomial), and subjected to differential gene expression analysis with differentialGeneTest (full model: ~CellType; reduced model: ~1). The top 1000 significantly differentially expressed genes were selected for dimensionality reduction using the DDRTree algorithm (max_components = 2). Trajectories were then constructed with orderCells, explicitly setting the root state, and visualized via plot_cell_trajectory with cells color-coded by pseudotime and annotated cell type.

Cell communication analysis

Ligand-receptor interaction analysis was performed using CellPhoneDB (v4.0)⁷². Single-cell RNA-seq matrices were log-normalized (scale factor = 10,000) via Seurat. Experimentally validated ligand-receptor pairs (membrane-bound, secreted, and extracellular matrix-associated proteins) were retrieved from CellPhoneDB (www.cellphonedb.org). Interaction probability was calculated through permutation testing. Mean expression values of ligand-receptor pairs were derived from normalized counts.

Statistical analysis

For comparisons between two groups, a two‑tailed unpaired Student’s t‑test was used. For comparisons among three or more groups, one‑way analysis of variance (ANOVA) was performed, followed by Tukey’s HSD test for multiple comparisons. For correlation analyses between TLS abundance and anti-PD-1 therapy response, two-tailed Kendall’s tau-b rank correlation was applied to assess correlations for non-normal variables with tied ranks. P-values were calculated based on asymptotic normal approximation.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Y.M. and M.W. performed most experiments and analysis. D.Z., W.Z. and L.S. performed co-culture assays. M.Q.L. did protein purification. M.M.L., Y.D. and J.Y. performed analysis of clinical relevance. R.Y., X.Y., Y.H., W.W., J.C., M.H., Y.T., Z.C., W.K., K.T., D.X. and H.L. discussed and analyzed results. Y.M., M.W., S.J. and Z.Z. wrote the manuscript. S.J., and Z.Z. supervised the project.

Peer review

Peer review information

Nature Communications thanks Milena Bogunovic and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

Our work was supported by the Innovative Drug Research and Development - National Science and Technology Major Project (2025ZD1800300), Noncommunicable Chronic Diseases-National Science and Technology Major Project (2025ZD0545200), National Natural Science Foundation of China Grants (82373251, 82573788, 82361168638, 82372613, U25A20113), Open Research Fund of State Key Laboratory of Genetics and Development of Complex Phenotypes (No. SKLGDP2509), Shanghai Leading Talent Program of Eastern Talent Plan (BJKJ2025031).

Data availability

Raw FASTQ files for scRNA-seq and spatial transcriptomics are deposited in GSA-Human (HRA019953, https://ngdc.cncb.ac.cn/gsa-human/browse/HRA019953) and bulk RNA-seq data under CRA045710 (https://ngdc.cncb.ac.cn/gsa/browse/CRA045710) and CRA045782 (https://ngdc.cncb.ac.cn/gsa/browse/CRA045782). All human-derived datasets are under controlled access to protect patient privacy and comply with China’s human genetic resource regulations. Non-commercial access is granted through the GSA-Human review process upon application; complete requests are processed within 4-6 weeks. All data are included in the Supplementary Information or available from the authors, as are unique reagents used in this article. The raw numbers for charts and graphs are available in the Source Data file whenever possible. Source data are provided with this paper.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77267-9.