Spatial Organization of Tumor-Infiltrating B Cells Informs Immune Checkpoint Inhibitor Response in Claudin 18.2–Expressing Gastric Cancer

Abstract

Purpose:

Claudin 18.2 (CLDN18.2), a tight junction protein normally expressed in the gastric epithelium and rendered accessible during malignant transformation, has emerged as a key therapeutic target. Here, we seek to characterize the immune microenvironment of CLDN18.2-expressing gastric cancer and identify immune features associated with CLDN18.2 expression status.

Experimental Design:

Tumor-immune features were profiled in 103 patients with HER2-neg/low advanced gastric cancer treated with first-line immune checkpoint inhibitor (ICI) using multiplex immunohistochemistry [7,423,483 cells across 2,315 regions of interest (ROI)]. The findings were validated in five published whole-transcriptome sequencing/microarray cohorts (1,521 samples), including the landmark CheckMate 649 phase III trial. Spatial immune architecture was assessed by cellular neighborhood analysis. Digital spatial profiling (DSP) was performed on 480 ROIs from a CLDN18.2-annotated tissue microarray (75 patients).

Results:

CLDN18.2^high tumors showed reproducible enrichment of humoral pathways, with higher CD20+ B-cell density and enrichment of B-cell/plasma-cell signatures. CLDN18.2 status alone was not associated with response or survival on first-line ICI therapy; in CLDN18.2^low tumors, B-cell enrichment was associated with ICI benefit, unlike CLDN18.2^high tumors. Spatial analyses indicated that CLDN18.2^high tumors preferentially harbored tumor-infiltrating B cells, rather than tertiary lymphoid structure/lymphoid aggregate–like neighborhoods, and this tumor-compartment localization was associated with diminished ICI benefit. DSP confirmed that CD20+ B-cell enrichment was most pronounced within tumor regions, alongside compartment-specific activation of humoral and extrafollicular response programs and interferon-associated and remodeling signatures.

Conclusions:

CLDN18.2^high gastric cancer exhibits a humoral-enriched microenvironment with increased tumor-infiltrating B cells and spatial immune organization associated with reduced ICI benefit. Spatial and humoral-state biomarkers may inform ongoing CLDN18.2-directed combination trials.

Translational Relevance.

Claudin 18.2 (CLDN18.2) is a therapeutic target in advanced gastric cancer, yet there remain unanswered questions about the benefit and biomarkers for immune checkpoint inhibitors (ICI) in this subgroup. In this study, we leveraged high-resolution spatial immunoprofiling (7,423,483 single cells quantified across 2,315 multiplex immunohistochemistry regions of interest (ROI) from 103 ICI-treated patients) together with digital spatial profiling of 480 ROIs from a CLDN18.2-annotated tissue microarray (75 patients). We show that CLDN18.2-high tumors are humoral-enriched, but their immune architecture is associated with reduced ICI benefit. CD20⁺ B cells are preferentially localized within tumor and myeloid-adjacent neighborhoods rather than tertiary lymphoid structure/lymphoid aggregate–like immune niches. These data support incorporating ROI-level spatial biomarkers of B-cell localization and humoral state into CLDN18.2-directed combination trials to refine patient selection and guide strategies that reprogram the immune architecture to overcome resistance.

Introduction

To date, the treatment of patients with advanced gastric cancer (1) has focused on key biomarkers including HER2 (2), PD-L1 (3–5), and microsatellite instability–high (MSI-H)/mismatch repair deficiency (dMMR) status (6–9). Chemotherapy in combination with anti–PD-1 immune checkpoint inhibitors (ICI) was established as standard of care for tumors expressing PD-L1 (3, 10, 11). Recently, claudin 18.2 (CLDN18.2), a tight junction protein expressed in malignant gastric cells, has emerged as a new target in gastric cancer (12). CLDN18.2 becomes exposed on the surface of gastroesophageal adenocarcinoma cells during malignant transformation and loss of polarity, making it a putative therapeutic target (13). Zolbetuximab, a monoclonal antibody that targets CLDN18.2, has demonstrated a statistically significant and clinically meaningful improvement in progression-free survival (PFS) and overall survival (OS) in the recent GLOW (14) and SPOTLIGHT (15) phase III trials. The addition of zolbetuximab to chemotherapy in the first-line treatment for CLDN18.2⁺/HER2⁻ advanced gastric cancer is now the standard of care (16). A considerable portion of patients with CLDN18.2-positive tumors also express PD-L1, creating a clinical challenge. Consequently, for patients positive for both biomarkers, it is currently unknown whether to prioritize the addition of zolbetuximab or an ICI to first-line chemotherapy. A recent network meta-analysis confirmed comparable efficacy between ICI agents and zolbetuximab in the first-line setting, highlighting the need for biomarker-driven treatment selection (17).

Although established predictors of ICI benefit—including PD-L1 expression, MSI-H status, tumor mutational burden, and B cell–associated immune infiltrates—have been characterized across other cancer populations (18), it remains unknown whether these parameters retain their predictive value in the distinct immune contexture of CLDN18.2-expressing tumors. We hypothesized that the specific immune contexture of CLDN18.2^high gastric cancer may play a role in guiding this strategy and, to date, has not been well characterized. In this study, we sought to elucidate the transcriptional and spatial characteristics of CLDN18.2–expressing gastric cancer and its relationship with the immune landscape. We sought to identify novel therapeutic targets or axes that could augment anti-CLDN18.2 and immunotherapy in this challenging subset of gastric cancer. By analyzing the interaction of these key markers, we aimed to shed light on their roles in shaping the tumor microenvironment (TME) and immune response, ultimately guiding more precise and personalized treatment strategies for patients with advanced gastric cancer.

Materials and Methods

Cohorts

Original cohort

A single-center cohort [Yonsei Cancer Center (YCC), Seoul, Republic of Korea] included 103 patients with HER2-neg/-low advanced gastric cancer treated with first-line ICI agent(s) (nivolumab or nivolumab with ipilimumab) combined with chemotherapy. Patients were followed for longitudinal survival outcomes, such as PFS and OS. Other patient characteristics included age, gender, Eastern Cooperative Oncology Group (ECOG) status, previous surgery, and stage at diagnosis. Tumor characteristics such as MSI status were determined in the clinical setting. The majority were primary samples (n = 74, 71.8%), whereas other samples were retrieved from metastatic sites, including the peritoneum (n = 8, 7.8%), omentum (n = 5, 4.9%), and others (n = 16, 15.5%). Patient characteristics were reported in Supplementary Table S1.

Formalin-fixed, paraffin-embedded (FFPE) pretreatment tumor tissues were collected from 103 patients with HER2-neg/-low advanced gastric cancer who received first-line 5-fluorouracil/platinum doublet chemotherapy in combination with nivolumab-containing immunotherapy between 2016 and 2023 at Yonsei Cancer Center, Korea. The study was reviewed and approved by the Institutional Review and Ethics Board of Severance Hospital (IRB No. 4-2023-0565). Written informed consent was obtained from each participant, and the study was conducted following the principles of the Declaration of Helsinki.

Validation cohorts

Whole-transcriptome sequencing/microarray

Five cohorts (3, 19–22) comprising 1,521 samples from previously published studies were included (Supplementary Table S2). For whole transcriptome sequencing (WTS), where FASTQ files were available [GCPM (19) and 3G (20)], raw reads were aligned to the GRCh38 reference genome using STAR (2.7.10a; RRID: SCR_004463; ref. 23). Gene expression was quantified with RSEM (v1.3.3; RRID: SCR_013027; ref. 24) as transcripts per million (TPM) and transformed as log₂(TPM + 1) for downstream analyses. Downstream analyses included immune cell deconvolution (25, 26) and gene set enrichment analysis (GSEA; Supplementary Table S3). For Asian Cancer Research Group (ACRG), preprocessed microarray data were retrieved, and the resulting expression matrix was log₂-transformed for downstream analyses. Batch correction was not undertaken across cohorts as they were analyzed separately.

Digital spatial profiling

The preparation of the Singapore Gastric Cancer Consortium (SGCC) tissue microarray (TMA) and details of NanoString digital spatial profiling (DSP) were previously described (27). Briefly, the SGCC TMA was constructed using 2-mm tumor and benign mucosa cores. Tumor cores included both the invasive front (“tumor edge”) and superficial or deep tumor regions as well as lymph node metastases. GeoMx DSP region of interest (ROI) selection and sample processing followed protocols previously reported in Ma and colleagues (27).

FFPE blocks from patient tissue were sectioned and stained with hematoxylin and eosin to guide ROI selection by a pathologist. Adjacent slides were processed according to the GeoMx WTA protocol (NanoString), with ROIs identified by immunostaining [DAPI (4′,6-diamidino-2-phenylindole), CD45, PanCK, and α-smooth muscle actin (α-SMA)]. Libraries were prepared using SeqCode reagents (NanoString) and sequenced on the Illumina platform.

Profiling details

Spatial proteomics

Multiplex immunohistochemistry (mIHC) staining was performed on 4-μm-thick FFPE tissue sections using a Leica Bond Max autostainer (Leica Biosystems), Bond Refine Detection Kit (Leica Biosystems), and Panovue 7-plex TSA IHC Kit (Panovue; refs. 28–31).

YCC cohort

We selected 2 to 50 fields of view per slide (each with 2,048 × 2,048 pixels, in which 1 pixel = 0.32 μm) from the multiplex tissue image of each case as ROIs using QuPath software (v0.5.0; RRID: SCR_018257; ref. 32). The ROIs were chosen depending on the size of the tissue to be representative of the overall tissue histology and with minimal to no artifacts or tissue folding based on DAPI and CK markers. Cell segmentation and feature measurements were performed using an automated and Artificial Intelligence (AI)-powered tissue image analysis software (cellXpress v2.4.0; https://cellXpress.org; Bioinformatics Institute; ref. 33).

The mean intensity of CK/EpCAM and CLDN18.2 markers in the cytoplasm regions was quantified. The mean intensity of CD68, CD8, and FOXP3 markers was measured in the whole cell, perinuclear, and nuclear regions, respectively. The 75th-percentile mean marker intensity was used for CD20 in the cytoplasm region, as well as for CD4 in the whole-cell region. Cell density was aggregated by average across ROIs for each patient.

SGCC cohort

We performed cell segmentation and feature measurement on the mIHC images of TMAs using an automated and AI-powered tissue image analysis software (cellXpress v2.4.0; https://cellXpress.org; Bioinformatics Institute; ref. 33). After all the cells from a TMA core had been identified and quantified, ROIs corresponding to the areas of interest selected for DSP analysis within the TMA core were manually matched and annotated on the mIHC images as polygons using the cellXpress software. TMA cores that were mismatched or contained artifacts were excluded from further analyses.

The mean intensity of CK/EpCAM in the cytoplasm regions was quantified using the cellXpress software. The mean cellular intensity of α-SMA, CD163, CD68, CD20, and CD3 markers was measured at the whole-cell regions.

Positive cells for each marker were identified using a web-based hyperplex tissue scoring system (ImmunoThreshold; ImmunoQs) written in R (v4.3.2). For each marker, the system used Gaussian mixture modeling (GMM; mixtools v2.0.0) for markers with bimodal or trimodal intensity distribution and the triangle algorithm (34) for markers with unimodal intensity distribution to determine a decision threshold that can separate between cells with high (“positive”) or low (“negative”) staining levels of the markers. Only for threshold determination, the system removed cells with the lowest 1-percentile mean marker intensity levels, log₂-transformed the mean intensity values, and ran the GMM algorithm up to 20,000 iterations or convergence. All the thresholds were manually checked based on the expected phenotypes, appearances, and coexpressions of other related markers on the positive cells. If a threshold was found to be inaccurate, the threshold would be manually adjusted based on the aforementioned criteria. To ensure consistency, all the ROIs from the same slides used the same thresholds. The single-cell data were further processed using custom data analysis scripts written in R. Cells with cell area (pixel number) lower or higher than the 0.1th and 99.9th percentiles of all cells were removed because they tended to be artifacts or missegmented cells. Subsequently, each cell type was identified based on either a single marker or a combination of markers (Supplementary Table S4).

Determination of CLDN18.2 status across cohorts

Given the multi-omics and integrative design of our cohorts, it was challenging to ensure that all cohorts included samples suitable for additional IHC staining, which we consider the gold standard for determining CLDN18.2 expression status. CLDN18 status was retrieved per the profiling approach described here.

YCC cohort

To retrieve CLDN18 expression, the number of tumor cells (CK/EpCAM+) and CLDN18+ tumor cells (CK/EpCAM+CLDN18+) in a ROI were aggregated by median per patient. The VENTANA 43-14A clone (Roche) was utilized for the detection of CLDN18. As both CK/EpCAM and CLDN18 can stain normal gastric tissue, only ROIs containing malignant epithelial cells, as manually annotated by a pathologist (W. Luo), were included in the analysis. To further ensure accurate identification of tumor cells and minimize contamination from infiltrating immune cells, a cell size filter of ≥9.5 μm² was applied, and ROIs with fewer than 50 total cells were excluded from analysis. To validate that CLDN18 expression levels obtained by semi-automated mIHC were consistent with manual IHC assessment, we compared the CLDN18 status of 23 clinical samples across both methods. We observed strong concordance (Pearson’s R = 0.89, P = 6.2 × 10⁻⁸), providing confidence in the reliability of our downstream analyses (Supplementary Fig. S1A and S1B).

After application of the above quality filters, n = 87 samples had available CLDN18.2 status for downstream analysis. Only 20 (23%) samples demonstrated ≥75% 2+/3+ CLDN18.2 expression, which limited the statistical power for direct comparisons between CLDN18.2-positive and CLDN18.2-negative groups. Therefore, we used a median split (25%) of the percentage of CLDN18+ tumor cells to classify samples as CLDN18.2-high or CLDN18.2-low.

WTS

As isoform-level information was not consistently available in the publicly accessible datasets, we classified samples as CLDN18-high or CLDN18-low by performing a median-split of log₂(TPM + 1) values of total CLDN18 expression. Although this approach does not distinguish between CLDN18 isoforms, it is supported by the observation that CLDN18.2 represents the primary isoform expressed in both normal gastric tissue and gastric cancer (35), providing a basis to use the overall CLDN18 transcript abundance as a reasonable surrogate for CLDN18.2 status in this setting.

SGCC TMA

CLDN18.2 (% 2+/3+) staining per TMA core was performed using the VENTANA 43-14A clone (Roche) in addition to methods previously described, with grading conducted by a pathologist (J.-Y.J. Lee). Aligned with the YCC cohort, CLDN18.2^high status was defined where CLDN18.2 2+/3+ was greater than 25%.

Determination of HER2-low status

HER2 status was classified accordingly: HER2-neg (IHC 0), HER2-low (IHC 1+ or IHC 2+ with FISH-negative confirmation), and HER2-positive (IHC 3+ or IHC 2+ with FISH-positive confirmation). Only HER2-neg and HER2-low patients were included in this analysis.

Statistical analyses

WTS

Previous reports have demonstrated strong correlation between CLDN18.2 expression on IHC and CLDN18 transcript expression (36). As such, we used CLDN18 transcript expression, stratified by median-split of log₂(TPM + 1) gene expression, to determine the CLDN18 status of gastric cancer samples. WTS data were processed to reconstruct B-cell receptor (BCR) sequences using TRUST4 (v1.1.6.1; ref. 37) from raw FASTQ files. Dimension reduction and visualization were performed using Uniform Manifold Approximation and Projection implemented in R (R package: umap, v0.2.10.0; RRID: SCR_018217). Differentially expressed genes were identified using the limma package in R (R package: limma, v3.62.2; RRID: SCR_010943), which uses a linear modeling approach to assess gene expression changes. Pathway analysis was conducted using GSEA, focusing on Gene Ontology Biological Process (GO:BP) pathways relevant to B-cell biology (R package: fgsea, v1.32.4; RRID: SCR_020938). Other gene signatures utilized are reported in Supplementary Table S3. Immune cell deconvolution was carried out in R (immunedeconv, v2.1.0), with xCell (25) used for primary analysis and MCP-counter (38) and EPIC (39) used for sensitivity analyses.

DSP

Differentially expressed genes were identified using the limma package in R (limma, v3.62.2). To determine the relative EF/GC response within ROIs, we developed a composite enrichment score (R package: singscore, v1.26.0) retrieved from gene signatures of EF/GC pathways previously reported by Ma and colleagues (Supplementary Table S3; ref. 40). Enrichment scores of B/plasma-cell subtypes were retrieved using xCell. Pathway analysis was conducted using GSEA, focusing on GO:BP pathways relevant to B-cell biology (R package: fgsea, v1.32.4).

Survival analysis

OS and PFS were assessed using the Cox proportional hazards model (R package: survival v3.7-0), with statistical significance determined by the log-rank test.

Retrieval of cellular neighborhoods

Cellular neighborhoods (CN) were retrieved through methods previously described by Schürch and colleagues (41, 42). Briefly, CN identification was performed using a custom k-nearest neighbors algorithm in Python (RRID: SCR_024202). We defined a window size of 10 cells (one center cell and nine nearest neighbors, measured by Euclidean distance) to capture local functional activity, with the 10-cell radius approximating two cell distances in each direction based on visual assessment. The windows were clustered according to the composition of their microenvironment, considering six cell types, resulting in a vector representing the frequency of each cell type in the neighborhood. Clustering was performed using MiniBatchKMeans with k = 10 (default) in Python. Each cell was then assigned to the same CN as the window it was centered in, reflecting how neighboring cell types influence the center cell’s function. Labeling of CNs was done manually based on immune cell proportions identified, regardless of CLDN18.2 expression.

Other statistical methods and reproducibility

A two-sided t test was used for continuous variables with normal distributions, and a two-sided Wilcoxon test for nonnormally distributed variables. Analysis of variance was used for comparisons between multiple groups. Fisher exact test was applied to unpaired count data, and McNemar test to paired data. P-values were adjusted for multiple testing using the Benjamini–Hochberg method, unless otherwise specified.

Results

CLDN18.2^high gastric cancer microenvironment is characterized by an enriched humoral response

We assembled a cohort of 103 FFPE biopsy samples retrieved from patients with advanced gastric cancer treated with first-line ICI combined with chemotherapy (Fig. 1A; Supplementary Table S1). Consistent with prior reports (35), a trend toward a bimodal distribution of CLDN18.2 expression was observed [Hartigans’ dip (43) test, P = 0.071; Supplementary Fig. S1A]. After quality control (Supplementary Fig. S1B), 87 samples had available CLDN18.2 expression data for downstream analysis. Of 86 patients with complete biomarker data, 4.7% (n = 4) were PD-L1 combined positive score (CPS) ≥5 and HER2-low and CLDN18.2⁺ (≥75%), whereas 11.6% (n = 10) were PD-L1 CPS ≥5 and HER2-low, 5.8% (n = 5) were CLDN18.2⁺ and PD-L1 CPS ≥5, and 2.3% (n = 2) were CLDN18.2⁺ and HER2-low (Supplementary Fig. S1C).

Figure 1.

B and plasma cells are enriched in CLDN18.2^high compared with CLDN18.2^low gastric cancers. A, Diagram illustrating a cohort of 103 patients with metastatic gastric adenocarcinoma treated with first-line ICI and chemotherapy. Biomarker expression was subsequently measured via mIHC assays, comprising CLDN18.2, CD8, CD4, CD68, CD20, and FOXP3. B, Dot plots of immune cell densities (CD20⁺ B cells, CD4⁺ T cells, CD8⁺ T cells, CD68⁺ macrophages, and FOXP3⁺ Tregs) across pMMR/MSS tumor samples (n = 90), stratified by CLDN18.2 expression status. CLDN18.2 status was determined by median-split of the percentage of CLDN18.2+ tumor cells assessed by IHC. Each dot is an individual sample. Horizontal lines indicate group medians. Statistical significance was determined using a two-sided T test. C, UMAP of WTS/microarray five cohorts (3G, ACRG, GCPM, CM649, and TCGA) consisting of 1,521 samples, stratified by CLDN18 status. CLDN18^high status was retrieved by median-split of transcript levels across all samples. Each dot is an individual sample. D, GSEA dot plot of immune and humoral GO:BP pathways between CLDN18^high vs. CLDN18^low tumor samples across WTS cohorts. E, Dot plot of enrichment levels of deconvoluted immune cell types across WTS cohorts. F, Stacked bar chart of BCR isotypes stratified by CLDN18 status across WTS cohorts. G, Dot plot of B/plasma-cell enrichment scores per sample in the ACRG cohort, stratified by CLDN18 status and Lauren classification (intestinal- and diffuse-type gastric cancer). Vertical lines indicate group means. Statistical significance was determined using a two-sided Wilcoxon test. adj, adjusted; CM, CheckMate; MSS, microsatellite stable; n.s., not significant; pMMR, proficient MMR; Tregs, regulatory T cells; UMAP, Uniform Manifold Approximation and Projection. (A, Created in BioRender. Sundar, R. [2026] https://BioRender.com/amarvna.).

To maximize statistical power and enable a balanced interrogation of immune microenvironmental differences across the CLDN18.2 expression spectrum, we dichotomized samples by median split (25.1%). We observed a unique increase in CD20⁺ B-cell density in CLDN18.2^high samples with proficient MMR (pMMR)/microsatellite stability status (P = 0.042), whereas no significant differences in cell density were observed in CD4⁺/CD8⁺ T cells, CD68⁺ macrophages, and FOXP3⁺ regulatory T cells (Fig. 1B). A similar trend of increased CD20⁺ B-cell density was also found including dMMR/MSI-H samples (Supplementary Fig. S1D). Across tumor sites, CLDN18.2 expression remained broadly consistent (Supplementary Fig. S1E). Consistent with the primary analysis, CD20⁺ B-cell density remained numerically higher in CLDN18.2^high primary tumors (P = 0.22) and other metastatic sites (P = 0.33), although these differences were not statistically significant. No difference was observed in peritoneal samples (P = 0.93), with interpretation limited by a small sample size (n = 14; Supplementary Fig. S1F; Supplementary Table S4). No significant difference in CLDN18.2 expression was observed between Helicobacter pylori–positive and H. pylori–negative samples (P = 0.81; Supplementary Fig. S1G). A numerical but statistically nonsignificant increase in CD20⁺ B-cell density was observed in H. pylori–negative, CLDN18.2^high samples (P = 0.12), and no significant difference in B-cell enrichment was observed in H. pylori–positive samples, although interpretation in this subgroup is constrained by small sample size (Supplementary Fig. S1H). These data suggest that the observed humoral immune enrichment in CLDN18.2^high tumors is unlikely to be primarily driven by H. pylori–associated inflammation.

To strengthen confidence in our observations, we initiated orthogonal validation via independent analyses of WTS and microarray data from five cohorts and trials (19–22, 44) in gastric cancer comprising 1,521 samples (Supplementary Table S2). Although CLDN18.2 IHC assessments were not available for these datasets, previous reports have demonstrated strong concordance between CLDN18 RNA transcript expression and IHC expression (36). Accordingly, we stratified samples into CLDN18^high and CLDN18^low using a median-split of CLDN18 transcript expression. Apart from The Cancer Genome Atlas (TCGA) cohort, CLDN18^high gastric cancers demonstrated a trend toward preferential co-localization on dimension reduction analysis, suggesting a degree of transcriptional separation from CLDN18^low gastric cancers (Fig. 1C). Consistent with our initial observation, CLDN18^high samples displayed a corresponding enrichment of humoral pathways consistently across the five independent cohorts, as measured through gene ontology pathway analyses (Fig. 1D). CLDN18^high samples displayed enriched B-cell signatures, particularly plasma cells and memory B cells, as measured via EPIC (39), MCP-counter (38), and xCell (25) immune deconvolution (Fig. 1E). We also observed increased levels of CD4⁺ T effector memory and CD8⁺ T central memory cells and reduced levels of cancer-associated fibroblasts within CLDN18^high samples. Akin to what we have observed at the protein level, CD4⁺ and CD8⁺ T cells showed inconsistent differences in abundance between CLDN18^high and CLDN18^low samples (Fig. 1E). We also observed enrichment of Hallmark IFNγ and IFNα response pathways in CLDN18^high samples across the five cohorts (Supplementary Fig. S1I). Given the observed enrichment of B cells and plasma cells in the CLDN18^high group, we performed BCR isotype analysis to characterize the nature of the humoral immune response and assess potential class-switching dynamics. TRUST4 (37) immune repertoire reconstruction of BCR sequences revealed expansion of IgA1 isotype in B/plasma-enriched CLDN18^high samples (P = 2.7e−16), suggesting a difference in class-switched humoral response (Fig. 1F; Supplementary Fig. S1J).

Considering that previous reports have observed an association between increased B/plasma cells and the Lauren (45) diffuse-type classification (46, 47), we examined CLDN18 status in two cohorts [ACRG (22) and TCGA (21)] with available Lauren subtype information. In both cohorts, we found increased B-cell enrichment scores in CLDN18^high versus CLDN18^low samples regardless of Lauren subtype (Fig. 1G; Supplementary Fig. S1K), suggesting that differences in humoral immunity may be associated with CLDN18 status rather than Lauren classification. Next, given the possibility that CLDN18 upregulation may directly drive B/plasma-cell enrichment, we tested whether a similar phenomenon occurs in an alternate cancer type. We examined CLDN18^high pancreatic adenocarcinoma (PDAC) using the Queensland Centre for Medical Genomics (QCMG; ref. 48) and TCGA (49) cohorts, but did not observe enrichment of B/plasma-cell signatures in CLDN18^high PDAC (Supplementary Fig. S1L), suggesting that the association seen may be unique to gastric adenocarcinoma.

Collectively, these findings suggest a unique microenvironmental landscape of enriched humoral response in CLDN18^high gastric cancer.

B-cell functional heterogeneity in the CLDN18^high microenvironment

Given enhanced humoral immunity in CLDN18^high gastric cancer samples, we hypothesized these tumors may respond better to ICIs, as increased B/plasma-cell signatures are associated with improved ICI outcomes (50, 51). However, in our cohort of patients treated with first-line ICI, CLDN18.2 status alone did not predict for response or survival [PFS, hazard ratio (HR) = 0.96; P = 0.886; OS, HR = 0.98; P = 0.947; Fig. 2A; ref. 44)]. This was likewise observed in the subgroup of patients treated with nivolumab (Supplementary Fig. S2A). CLDN18.2^high samples treated with the nivolumab–ipilimumab combination were found to have a trend toward improved OS (HR = 0.197; P = 0.101) and PFS (HR = 0.680; P = 0.598), although interpretation was limited by the small sample size (n = 12; Supplementary Fig. S2B). Compared with HER2-neg, patients with a HER2-low status (n = 26, 25.2%) had a trend toward improved PFS (HR = 0.709; P = 0.224). No significant OS differences were observed (HR = 0.826; P = 0.540; Supplementary Fig. S2C).

Figure 2.

Differential B-cell activity in a CLDN18.2^high microenvironment. A, KM plots of PFS and OS in cohort of patients treated with first-line chemoimmunotherapy stratified by CLDN18.2 status. A total of 87 patients with available CLDN18.2 status were included in this analysis. B, Forest plot of PFS and OS outcomes in patients treated with first-line chemoimmunotherapy according to PD-L1 CPS status and immune cell density [high vs. low; stratified by median-split of log₂(density + 1)]. A HR greater than 1 denotes ICI resistance, whereas a HR less than 1 denotes ICI sensitivity. C, KM plots of PFS and OS stratified by CLDN18.2 status and B-cell density (high vs. low). CLDN18.2 status was retrieved by median-split of the percentage of CLDN18.2+ tumor cells by IHC. B-cell status was retrieved through median-split of log₂(density + 1). D, Forest plot of PFS and OS outcomes in patients treated with first-line chemoimmunotherapy in the CheckMate 649 trial, stratified by B-cell subtype enrichment (high vs. low) and CLDN18 expression. E, KM plots of PFS and OS stratified by treatment arm in the CheckMate 649 trial in B cell–enriched patients, across CLDN18 expression (high vs. low). The boxplots in green denote CLDN18^high samples, whereas those in gray denote CLDN18^low samples. CI, confidence interval; KM, Kaplan–Meier; Tregs, regulatory T cells.

Regardless of CLDN18.2 status, increased density of immune cells confirmed a trend toward improved ICI sensitivity, consistent with previous work by Shitara and colleagues (44). Patients with increased CD20⁺ B-cell density was associated with a greater magnitude of benefit from first-line ICI therapy (PFS, HR = 0.80; P = 0.367; OS, HR = 0.71; P = 0.196; Fig. 2B); however, this benefit was largely observed among CLDN18.2^low tumors and not in CLDN18.2^high tumors (Fig. 2C). Sensitivity analyses of patients receiving nivolumab as their ICI agent demonstrated consistent results, whereas sensitivity analyses of patients receiving nivolumab–ipilimumab was limited by the small sample size (Supplementary Fig. S2D–S2F).

Similarly, among patients treated with first-line chemotherapy plus ICI in the CheckMate 649 trial (n = 250), we found that the survival benefit associated with increased B/plasma-cell enrichment was uniquely found in CLDN18^low expressing tumors (Fig. 2D). In contrast, this phenomenon of differential survival outcomes across CLDN18 expression and B/plasma-cell enrichment was not observed in patients treated with chemotherapy alone (Supplementary Fig. S2G). Consequently, among B cell–enriched samples, patients with high CLDN18 expression did not show any PFS or OS benefit with the addition of ICI (Fig. 2E).

Intratumoral B-cell infiltration is associated with ICI resistance in a CLDN18.2^high microenvironment

The observed interaction between CLDN18.2-status and benefit from B-cell enrichment suggests heterogeneity in B-cell states and function across CLDN18.2 microenvironments. Although cell density can provide a basic measure of immune infiltration, it may overlook key aspects of tumor-immune interactions, such as immune cell proximity and spatial arrangement (52). To better assess these features, we analyzed CNs, revealing the distribution of immune cells relative to one another. Through unsupervised clustering of 7,423,483 cells from 2,315 ROIs, we identified 10 CNs across all samples (41), uniquely characterized by differing proportions of immune cell types (Supplementary Fig. S3A). Further consolidation of these CNs revealed four key B cell–related CNs—namely B-TumorInfiltrating, B-MacrophageNiche, B-cell lymphoid aggregate (B-LA)/ B-cell tertiary lymphoid structure (TLS), and B-ImmuneNiche (Fig. 3A). B-TumorInfiltrating CN is characterized by a dispersion of B cells among tumor cells. B-MacrophageNiche CN comprises a large proportion of macrophage cells with few B and T cells. B-LA/TLS CN is composed of a large abundance of B cells with few T cells and tumor. Lastly, B-ImmuneNiche comprises a mixed population of macrophages, B cells, and T cells within the tumor. In CLDN18.2^high tumors, we observed a relative increase in average B-cell abundance within B-TumorInfiltrating and B-MacrophageNiche CNs and a relative decrease in average B-cell abundance within B-LA/TLS and B-ImmuneNiche CNs in comparison with CLDN18.2^low tumors. This suggests CLDN18.2-expressing tumors preferentially allow B-cell infiltration into tumor compartments (Fig. 3B).

Figure 3.

Intratumoral B-cell infiltration is associated with ICI resistance in CLDN18.2^high tumors. A, Cellular composition of four B cell–associated CNs. Representative ROIs are shown for each CN, alongside the cellular fractions of constituent cell types within the neighborhood. B, B-cell distribution, represented by the proportion of B cells across the B-cell CNs stratified by CLDN18.2 status. C, Forest plot of OS outcomes in patients treated with first-line chemoimmunotherapy according to absolute and relative B-cell distribution across CNs, stratified by CLDN18.2 status. A HR greater than 1 suggests ICI resistance, whereas a HR less than 1 suggests ICI sensitivity. adj, adjusted.

B-TumorInfiltrating CNs were associated with ICI resistance, whereas B-ImmuneNiche and B-LA/TLS were associated with ICI sensitivity (Fig. 3C). CLDN18.2^low tumors show preferential B-cell enrichment within the B-ImmuneNiche and B-LA/TLS neighborhoods, both of which are linked to increased sensitivity to ICIs. The relative enrichment of B cells within a B-TumorInfiltrating CN compared with a B-ImmuneNiche CN was associated with ICI resistance, particularly in CLDN18.2^high tumors (OS, HR = 2.60; P = 0.040; Fig. 3C). A similar trend was observed for PFS (Supplementary Fig. S3B). These findings collectively suggest a potentially immunosuppressive B-cell niche in a CLDN18.2^high microenvironment.

Spatial evolution of humoral response in CLDN18.2^high gastric cancer

To better characterize the spatial evolution of tumor-infiltrating B cells, we utilized DSP (NanoString DSP GeoMx) of 480 ROIs derived from a TMA assembled from 75 patients (Supplementary Fig. S4A; Supplementary Table S2; ref. 27). Tissue cores were annotated based on anatomic site—superficial tumor, deep tumor, invasive front, and lymph node metastasis—whereas ROIs were further classified by their location within the tumor, stroma, tumor–stromal interface (TSI), or intestinal metaplasia (IM). In addition to the previously reported spatial transcriptomics data in Ma and colleagues (27), for each ROI, mIHC was performed using an immune marker panel (CD3, CD20, CD163, CD68, and α-SMA) and separately with CLDN18.2 IHC using the VENTANA 43-14A clone (Roche) to determine CLDN18.2 status [≥25% 2+/3+ expression (likewise 50^th percentile), aligned with our upstream mIHC analyses; Fig. 4A]. CLDN18.2 protein expression as assessed by IHC correlated reasonably well with CLDN18 RNA transcript levels (Pearson’s R = 0.44; P = 2.7e−16; Supplementary Fig. S4B). We confirmed the observation of increased CD20⁺ B-cell infiltration within CLDN18.2^high tumor compartments, with no significant changes in CD3⁺ T-cell density. Notably, the enrichment of CD20⁺ B cells was most pronounced within tumor regions (P = 0.0061; Fig. 4B and C; Supplementary Fig. S4C). Of the 75 patients included, 20 patients had both PD-L1 22C3 CPS and CLDN18.2 IHC available. At clinically relevant thresholds, 5% (n = 1) patient was found to be PD-L1 (22C3 CPS ≥5) and CLDN18.2 (2/3+ ≥75%) coexpression (Supplementary Fig. S4D).

Figure 4.

Spatial evolution of humoral response in CLDN18.2^high gastric cancer. A, Schematic overview of DSP of samples and annotation of ROIs. B, Dot plot overview of differences in immune cell densities across ROIs. Statistical testing was undertaken utilizing a two-sided T test. C, Box plot of CD20⁺ B-cell density across ROIs illustrating enrichment of CD20⁺ B cells in TSI and tumor compartments. Statistical testing was undertaken utilizing a two-sided T test. D, Heatmap overview of mean B/plasma-cell type enrichment scores in tumor ROIs. E, Heatmap overview of GSEA normalized enrichment scores between CLDN18.2^high and CLDN18.2^low samples. LN, lymph node; LA, lymphoid aggregate; mets, metastasis; n.s., not significant; ***, adjusted P < 0.001; **, adjusted P < 0.01; *, adjusted P < 0.05. (A, Created in BioRender. Sundar, R. [2026] https://BioRender.com/imq21kx.).

B-cell enrichment in CLDN18.2^high tumors was most pronounced in superficial tumor samples closest to the mucosal surface (P = 0.012), with no significant difference observed in deeper tumor compartments (P = 0.44), suggesting that mucosal proximity may govern spatial distribution of B-cell accumulation in this subgroup (Supplementary Fig. S4E). We observed enrichment of B cells in CLDN18.2^high samples, particularly in IM, superficial tumor, and deep tumor compartments. Lymph node metastases of CLDN18.2^high tumors showed enrichment of memory and naïve B cells (Fig. 4D).

To further characterize the structural organization of this humoral response, we applied a validated transcriptomic TLS signature (53) as an orthogonal proxy for TLS enrichment across spatially resolved tumor compartments. TLS enrichment scores were highest at the TSI and lowest within the tumor compartment, consistent with the known predilection of TLS for the invasive margin (Supplementary Fig. S4F). Upon stratification by CLDN18.2 status, TLS enrichment scores were substantially higher in CLDN18.2^high tumor (P = 0.00048) and stromal (P = 0.011) compartments, whereas no significant differences were observed in the TSI (P = 0.48; Supplementary Fig. S4G). We further observed enrichment of humoral pathways in CLDN18.2^high compared with CLDN18.2^low tumors across superficial and deep compartments and invasive fronts but not in nodal metastasis (Fig. 4E).

Correspondingly, CLDN18.2^high tumor ROIs were characterized by an enrichment of an extrafollicular response pathway (40, 54) in both deep tumors (P = 0.040) and invasive fronts (P = 0.133). These evolving changes were conversely not observed in CLDN18.2^low tumor ROIs (Fig. 5A). This suggests that B cells in CLDN18.2^high gastric cancer are potentially involved in tumor progression and immune evasion.

Figure 5.

Spatial evolution of humoral response in CLDN18.2^high gastric cancer (cont’d). A, Boxplot comparison of a composite germinal center/extrafollicular signature stratified by CLDN18.2 status across sampled locations. Statistical significance was determined using a two-sided Wilcoxon test. B, Selected B/plasma-cell enrichment scores across the TSI spectrum. Statistical significance was determined using a two-sided T test. C, HALLMARK pathway enrichment analysis across TSI spectrum. Only top 10 up and down pathways are shown here. Statistical comparisons were undertaken using a two-sided Wilcoxon test. D, Selected gene expression changes across the TSI spectrum. Statistical significance was determined using a two-sided Wilcoxon test. EF, extrafollicular; GC response, germinal center response; LN, lymph node; mets, metastasis; n.s., not significant.

To further dissect the spatial dynamics of humoral immunity, we examined the localization of B- and plasma-cell subsets across the TSI. B/plasma cells were primarily observed to be enriched within stromal compartments (Supplementary Fig. S5A). Consistent with our upstream findings, we observed a relative enrichment of B cells (P = 0.101), plasma cells (P = 0.028), and class-switched memory B cells (P = 0.030) within tumor compartments of CLDN18.2^high samples (Fig. 5B).

Pathway enrichment analysis revealed that CLDN18.2^high tumor compartments exhibited upregulation of the Hallmark allograft rejection, coagulation cascade, and IFNα response pathways. Coagulation and IFNα response signatures were prominent in the tumor cell regions, suggesting an immune-activated microenvironment and a transplant rejection–like environment. Notably, the allograft rejection pathway was enriched not only in CLDN18.2^high tumor cells but also across the TSI and stromal compartments, indicating widespread immune activation within the TME. Additionally, the Hallmark epithelial–mesenchymal transition (EMT) signature was enriched in both CLDN18.2^high TSI and stroma, highlighting increased tumor plasticity and invasive potential in these regions (Fig. 5C).

Inspecting differentially expressed genes across the stromal, TSI, and tumor compartments between CLDN18.2^high and CLDN18.2^low cores reveals differential gene expression profiles particularly within tumor compartments. We observe an upregulation of CXCL17 (P = 5.24e−04), a mucosal chemokine known to recruit dendritic cells and myeloid cells and promote leukocyte homing to epithelial surfaces, toward tumor ROIs, potentially explaining the observed phenomenon of increased homing and retention of B/plasma cells within CLDN18.2^high tumor compartments. Furthermore, the TSI (P = 0.059) and tumor (P = 0.053) compartments of CLDN18.2^high tumors exhibit elevated expression of PIGR (polymeric immunoglobulin receptor), a transmembrane protein responsible for transcytosis of secretory immunoglobulins, including IgA and IgM, across epithelial surfaces, suggesting enhanced immunoglobulin transport within the tumor compartment (Fig. 5D).

Taken together, these findings suggest a complex and spatially resolved evolution of humoral immunity in CLDN18.2^high gastric cancer. The dysfunctional, extrafollicular programs, alongside enhanced B/plasma-cell homing and antibody transport at the TSI, and stromal remodeling with EMT, collectively shape a unique TME. This distinct landscape has potential implications for tumor growth, metastasis, and therapeutic resistance in this gastric cancer subtype.

Discussion

In this study, we define the immune landscape associated with CLDN18.2 expression in advanced gastric cancer using complementary protein-based profiling, bulk transcriptomic analyses, and spatially resolved interrogation of the TME. Across cohorts, CLDN18.2/CLDN18^high tumors exhibited a reproducible enrichment of humoral immune programs, characterized by increased B-cell and plasma-cell signatures. However, despite this apparently “immune-enriched” state, CLDN18.2 status alone was not associated with improved outcomes to first-line ICI-based therapy. Instead, our data suggest that the clinical relevance of humoral immunity is context-dependent: B-cell enrichment was associated with greater ICI benefit primarily in CLDN18.2^low tumors, whereas in CLDN18.2^high tumors, intratumoral B-cell localization patterns were more consistent with ICI resistance.

A key observation is the apparent dissociation between the presence of humoral immune features and therapeutic benefit in CLDN18.2^high disease. Prior studies have associated B-cell infiltration, plasma-cell signatures, and TLSs with improved responses to immunotherapy in multiple tumor types (50, 51, 55). However, in our analyses, CLDN18.2^high tumors showed preferential enrichment of B cells within tumor-infiltrating and macrophage-associated neighborhoods, alongside relative depletion of B-cell localization within TLS/LA-like and mixed immune niches—spatial contexts that are more commonly linked to productive antitumor immunity. This spatial compartmentalization provides a plausible explanation for why bulk measures of B/plasma-cell enrichment did not translate into improved ICI outcomes in CLDN18.2^high tumors: humoral immunity may be present but organized in a manner that is potentially less supportive of effective T-cell priming, antigen presentation, or coordinated cytotoxic activity. Our spatial profiling further supports compartment-specific immune remodeling in CLDN18.2^high tumors. The enrichment of humoral pathways in superficial and deep tumor regions and at invasive fronts, together with signals consistent with extrafollicular B-cell programs (40), suggests that B-cell responses in CLDN18.2^high gastric cancer may be skewed toward non-TLS, potentially dysregulated states. Such programs have been linked in other settings to inflammatory yet ineffective immunity, antibody production without durable tumor control, or immunoregulatory cross-talk with myeloid cells (54, 56). In parallel, transcriptional differences at the tumor and TSI, including trends toward increased expression of putative chemotactic and epithelial transport-associated genes (e.g., CXCL17 and PIGR), raise the possibility that CLDN18.2^high tumors may actively shape B-cell trafficking and immunoglobulin handling in specific compartments. Although these findings are hypothesis-generating, they collectively point toward a TME in which humoral activity is prominent but may not be aligned with ICI sensitivity.

Whereas these findings did not directly mirror prior reports by Lim and colleagues, whose single-cell characterization of the CLDN18.2-positive TME was limited to 11 CLDN18.2-positive and 20 CLDN18.2-negative tumors, our study is derived from a substantially larger cohort and critically incorporates spatial context, enabling resolution of immune cell localization and tumor–immune interactions that single-cell dissociation-based approaches inherently cannot capture (57). Furthermore, in contrast to pancreatic cancer, where CLDN18 expression promoted T-cell infiltration without enhanced B-cell infiltration (58), we observed the opposite pattern in our gastric cohort—predominant B-cell enrichment and TLS-like structures in CLDN18.2^high tumors without a corresponding difference in T-cell infiltration. We hypothesize that this divergence may reflect the unique epithelial barrier context of gastric cancer, where CLDN18.2-dependent tight junction integrity may influence paracellular antigen leakage and luminal antigen exposure, preferentially driving B cells over cytotoxic immune responses in a tissue-specific manner.

Emerging results from CLDN18.2-targeted chemoimmunotherapy combinations (59–61) will help dissect the implications of our spatial and transcriptomic findings. Although CLDN18.2^high tumors exhibited enriched humoral programs, this did not translate into improved benefit from ICI in our cohorts, and resistance was instead linked to B cells organized within tumor-infiltrating and macrophage-adjacent neighborhoods rather than TLS/immune niches. Early-phase data combining zolbetuximab with chemotherapy and PD-1 blockade, such as in the recent ILUSTRO trial (59, 60), suggest that CLDN18.2 engagement may help overcome this apparent “immune-present but misaligned” state, potentially by increasing immunogenic cell death and reshaping B cell–T cell interactions. Together, these observations support prioritizing combination strategies that not only add checkpoint blockade but also reprogram humoral architecture (e.g., TLS maturation) in CLDN18.2⁺ disease, perhaps with on-treatment spatial profiling in CLDN18.2-targeted trials.

The central goal of our study was to investigate the underlying biology of CLDN18.2 and its relationship with the immune microenvironment. To achieve this, we defined CLDN18.2 positivity using a median split, a method intentionally chosen for its biological relevance rather than to replicate the specific companion diagnostic criteria (≥75% IHC staining) used for zolbetuximab. Extending a single, rigid clinical cutoff across multiple, diverse international cohorts with varying assay types (IHC and transcriptomics) may have been inappropriate. Furthermore, the practical and ethical challenges of acquiring sufficient paired tissue for cross-assay validation on a large scale are prohibitive. Crucially, the therapeutic landscape for CLDN18.2-targeted agents is rapidly expanding. Whereas zolbetuximab has a high expression requirement, other strategies under investigation, such as antibody–drug conjugates and bispecific antibodies, are being developed for tumors with lower CLDN18.2 expression levels (62). By focusing on biologically meaningful, data-driven cut-points rather than a single drug-specific threshold, our study provides a broader understanding of the immunobiology across the CLDN18.2 expression spectrum, which will remain relevant as new therapies emerge.

Several limitations should be considered. First, our validation across transcriptomic datasets relied on total CLDN18 transcript expression as a surrogate for CLDN18.2 protein expression, which may introduce misclassification due to isoform biology or tumor purity effects. Second, many analyses are observational, and although they reveal robust and recurrent associations, they do not establish causality between CLDN18.2 expression and B-cell recruitment, differentiation, or spatial positioning. Third is the reliance on endoscopic biopsy specimens rather than whole tissue sections. As surgical resection is rarely performed in advanced gastric cancer, pretreatment biopsies represent the only routinely available tissue, reflecting the real-world context in which predictive biomarkers must ultimately be applied. Fourth, the use of biopsy specimens and TMA material limits spatial sampling of immune infiltrates, potentially underrepresenting heterogeneity at the invasive front and precluding a comprehensive assessment of the relationship between intratumoral inflammation and tumor margin. Finally, statistical interactions underlying subgroup effects warrant further prospective validation and formal modeling in independent datasets.

In summary, we identify CLDN18.2^high gastric cancer as a subtype with prominent humoral immune features but with distinct spatial organization of B cells that may be linked to reduced benefit from ICIs. These data challenges the prevailing view that B-cell enrichment uniformly predicts immunotherapy responsiveness by highlighting the importance of immune architecture and neighborhood context. Future work integrating CLDN18.2-targeted treatment cohorts, single-cell profiling, and functional perturbation studies will be essential not only to determine whether compartment-specific humoral programs can be leveraged therapeutically but also to inform rational combination strategies that reprogram immune architecture and unlock the full potential of CLDN18.2-directed therapy.

Supplementary Material

Supplementary Tables 1-4

Supplementary Tables

Supplementary Figures S1-S5

Figure S1: Quality-control parameters and sensitivity analyses of CLDN18.2 microenvironmental and pathway changes Figure S2: Additional sensitivity survival analyses of the YCC cohort and CheckMate 649 trial Figure S3: Features of cellular neighbourhoods and correlation with progression-free survival Figure S4: Supplementary analyses to digital spatial profiling Figure S5: B/plasma cell enrichment scores across the tumor-stroma interface

Data Availability

All data supporting the findings of this study, including the raw data underlying the figures, are available from the corresponding author (email) upon reasonable request.

Code Availability

The code utilized in this study is publicly available in the following GitHub repository: https://github.com/josephjzhao/GastricCancerTME-CLDN18.2.

Authors’ Disclosures

J.J. Zhao reports grants from National University Health System Seed Fund (NUHSRO/2024/008/RO5+6/Seed-Sep23/01), National University Hospital Junior Research Award 2023 (JRA/Sep23/002), the Chan Heng Leong Education & Research Fund 2024 awarded by the National University Hospital Singapore (CHL/Sep24/002), ExxonMobil-NUS Research Fellowship for Clinicians, and Dean’s Research Development Award awarded by the Yong Loo Lin School of Medicine, National University of Singapore during the conduct of the study. C.-k. Lee reports personal fees from Astellas Pharma, AstraZeneca, Celegene/Bristol-Myers Squibb, Ono Pharmaceutical, Servier, Yuhan, Merck Sharp & Dohme, Roche, and Daiichi-Sankyo; personal fees and non-financial support from Boryung Corporation; grants and non-financial support from GC Biopharma; non-financial support from Ono Pharmaceuticals, Idience, and Selecxine; and grants from Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea outside the submitted work. S. Blum reports other support from Two River Consulting, Kronos Bio, 76Bio, Allogene Therapeutics, and Candid Therapeutics; and grants from AstraZeneca and Astellas outside the submitted work. A. Pelissier reports employment with X Development LLC. A. Pelissier contributions to this research were conducted during previous employment and were not funded or influenced by X Development LLC. J.-Y.J. Lee reports other support from National University Hospital during the conduct of the study, and is co-founder and COO of ImmunoQs. Y. Dai reports other support from National University Hospital during the conduct of the study. L.-H. Loo reports personal fees from National University Hospital during the conduct of the study, and is co-founder and CEO of ImmunoQs. I. Nakayama reports grants from Astellas; personal fees and other support from Astellas, Ono Pharmaceutical Co., Ltd., Merck Sharp & Dohme, Daiichi Sankyo, and AstraZeneca; personal fees from Bristol-Myers Squibb; and other support from Boehringer Ingelheim and Chugai Pharmaceutical Co., Ltd. outside the submitted work. R. Walsh reports personal fees from AstraZeneca, GeneSolutions, BeiOne, Ipsen, and Roche; and non-financial support from Taiho and Daiichi-Sankyo outside the submitted work. A.D. Jeyasekharan reports personal fees from AbbVie, DKSH/BeiGene, Roche, Gilead, Antengene, MSD, IQVIA, Turbine Ltd, and KYAN Technologies outside the submitted work; and grants and personal fees from AstraZeneca and Janssen. W.P. Yong reports other support from Astellas, AstraZeneca, BMS, Ipsen, Merck, MSD, Daichii Sankyo, Taiho, and Guardant outside the submitted work. I. Vazquez-Garcia reports grants from Gastric Cancer Foundation during the conduct of the study. S.J. Klempner reports personal fees from BMS, Merck, Astellas, Natera, Novartis, AstraZeneca, Mersana, BeOne, Gilead, EsoBiotec, Eisai, Taiho, Boehringer-Ingleheim, and I-Mab; and grants from AACR, Torrey Coast Foundation, SU2C, and NCI/NIH outside the submitted work. K. Shitara reports grants and personal fees from Astellas Pharma, Ono Pharmaceutical, Daiichi-Sankyo, and AstraZeneca; grants from PRA Health Sciences, Syneos Health, PPD-SNBL K.K, and TORAY; personal fees from Bristol-Myers Squibb, Takeda, Novartis, Boehringer Ingelheim, Guardant Health Japan, Janssen, Zymeworks Biopharmaceuticals Inc., ALX Oncology Inc., Bayer, GlaxoSmithKline K.K., HEALIOS K.K., Moderna, Inc., Arcus Biosciences Inc., and Eli Lilly from outside the submitted work. Y. Janjigian reports grants and personal fees from Astellas and Bayer; grants, personal fees, and non-financial support from AstraZeneca, Arcus Biosciences, Bristol-Myers Squibb, Merck, Transcenta, and Boehringer Ingelheim; grants from Cycle for Survival, Fred’s Team, and Stand Up 2 Cancer; personal fees from AbbVie, BeOne, Debbie's Dream Foundation, 13Health, and Pfizer; personal fees and non-financial support from Daiichi-Sankyo, Gilead Sciences, HMP Global, Michael J. Hennessy Associates, OncoDaily, PeerView Institute, Physician’s Education Resource, LLC, Research to Practice, Suzhou Liangyihui Network Technology Co., Ltd, and TotalCME outside the submitted work. S.Y. Rha reports grants and other support from Amgen, Astellas, AstraZeneca, Daiichi-Sankyo, LG Biochem, MSD, Jazz, BeOne, and Boehringer Ingelheim; grants from Eisai, Roche, and Taiho outside the submitted work; and other support from Qureator. R. Sundar reports grants, personal fees, and non-financial support from Astellas; grants and personal fees from AstraZeneca and BigHat BioSciences; personal fees from BeOne Medicines, GSK, and Takeda; grants from Amgen, Genetech, MSD, Daiichi-Sankyo, and Gilead; and non-financial support from Paxman outside the submitted work. In addition, R. Sundar has a patent for Auristone pending to Auristone and a patent for Paxman pending and licensed to Paxman. No disclosures were reported by the other authors.

Authors’ Contributions

J.J. Zhao: Conceptualization, resources, software, formal analysis, funding acquisition, validation, methodology, writing–original draft, project administration, writing–review and editing. C.-k. Lee: Conceptualization, resources, data curation, methodology, project administration, writing–review and editing. A.S.Y. Chan: Conceptualization, resources, data curation, formal analysis, investigation, writing–original draft, writing–review and editing. S. Blum: Resources, data curation, writing–review and editing. C. Li: Conceptualization, data curation, writing–review and editing. H. Ma: Conceptualization, resources, data curation, writing–review and editing. W. Luo: Data curation, writing–review and editing. S. Srivastava: Data curation, writing–review and editing. H.Y.J. Lum: Data curation, writing–review and editing. S.C.N. Tay: Data curation, writing–review and editing. W.S. Kwon: Data curation, writing–review and editing. S. Park: Data curation, writing–review and editing. C.R. Joseph: Resources, data curation, writing–review and editing. L.Y. Chong: Data curation, writing–review and editing. M. Jung: Data curation, writing–review and editing. S. Sridhar: Data curation, supervision, writing–review and editing. A. Pelissier: Formal analysis, investigation. N. Bhatt: Supervision, investigation, writing–review and editing. J.-Y.J. Lee: Data curation, writing–review and editing. Y. Dai: Data curation, writing–review and editing. L.-H. Loo: Data curation, writing–review and editing. A.L.K. Tan: Data curation, writing–review and editing. T. Hagihara: Resources, writing–review and editing. F. Wee: Data curation, writing–review and editing. D.K.A. Chia: Writing–review and editing. K.M.S. Sim: Writing–review and editing. M. Teh: Writing–review and editing. I. Nakayama: Resources, writing–review and editing. J. Xu: Writing–review and editing. M.C.H. Ng: Writing–review and editing. R. Walsh: Writing–review and editing. J.P.S. Yeong: Writing–review and editing. A.D. Jeyasekharan: Data curation, writing–review and editing. H.I. Grabsch: Resources, writing–review and editing. J.B.-Y. So: Resources, writing–review and editing. W.P. Yong: Resources, writing–review and editing. I. Vazquez-Garcia: Investigation, writing–review and editing. L. Bod: Investigation, writing–review and editing. S.J. Klempner: Resources, data curation, investigation, writing–review and editing. K. Shitara: Data curation, writing–review and editing. Y. Janjigian: Data curation, writing–review and editing. P. Tan: Conceptualization, data curation, supervision, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing. M. Rodríguez Martínez: Conceptualization, data curation, software, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing. S.Y. Rha: Conceptualization, data curation, funding acquisition, investigation, writing–original draft, writing–review and editing. R. Sundar: Conceptualization, data curation, supervision, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing.