Intra- and intertumoral discordance of Claudin-18.2: a systematic review and meta-analysis Authors: TLM Vanneste, DIR Sánchez, S van der Mierden, MA Vollebergh, LL Kodach, J van Sandick, RGH Beets-Tan, P Snaebjornsson, Z Bodalal Published: 2026 ABSTRACT AbstractBackgroundClaudin-18.2 status is increasingly used to guide targeted therapy based on immunohistochemistry of limited tumor tissue. Spatial heterogeneity may produce discordant expression patterns within a primary tumor or discordant positivity classification between primary and metastatic sites, limiting the representativeness of a single specimen. No prior meta-analysis has quantified the magnitude of this discordance, either in gastric/gastroesophageal junction (GEJ) cancer or in other tumor types.Materials and methodsWe carried out a Preferred Reporting Items for Systematic Reviews and Meta-Analyses-compliant systematic review and meta-analysis. MEDLINE, Embase, and Scopus were searched for studies reporting Claudin-18.2 intratumoral discordance within primary tumors and/or intertumoral discordance between primary tumors and matched metastases. Intratumoral discordance was operationalized according to each study’s reported evidence of spatial variation in expression, whereas intertumoral discordance was defined as disagreement in positivity classification between matched primary and metastatic samples using study-specific cut-offs. Random-effects models pooled logit-transformed discordance proportions. Robustness was evaluated using leave-one-out sensitivity analyses and subgroup analyses explored potential sources of heterogeneity. Between-study heterogeneity and small-study effects were assessed using Cochran’s Q, I2, Tau2, funnel plots, and Egger’s regression.ResultsThirteen studies (n = 3097 patients) contributed to intratumoral discordance and 12 studies (n = 987 patients) contributed to intertumoral discordance in gastric/GEJ-specific cohorts. The pooled intratumoral discordance rate was 35.8% [95% confidence interval (CI) 20.5-54.7, I2 = 94.3%]. The pooled intertumoral discordance rate was 20.3% (95% CI 17.3-23.7, I2 = 17.9%).ConclusionClaudin-18.2 expression shows substantial intratumoral and clinically meaningful discordance between primary and metastatic tumors in gastric/GEJ cancer. Single-specimen testing may misclassify eligibility and multisite testing should be considered.Key words: biomarker discordance, Claudin-18.2, meta-analysis, spatial heterogeneity, systematic review Introduction Claudin-18 isoform 2 (Claudin-18.2) is a transmembrane tight junction protein that has emerged as a clinically relevant biomarker and therapeutic target in upper gastrointestinal malignancies.1,2 In healthy tissue, its expression is restricted primarily to differentiated gastric epithelial cells, reflecting a highly tissue-specific profile.3, 4, 5 This restricted distribution is therapeutically attractive because it offers a large differential between tumor and normal-tissue expression, minimizing on-target off-tumor toxicity. During malignant transformation, however, tight junction architecture is disrupted and Claudin-18.2 expression patterns become variable: some tumors retain or upregulate expression, whereas others exhibit reduced, heterogeneous, or absent staining.1,4,6 This variability has become clinically consequential with the rapid expansion of Claudin-18.2-targeted therapy. Zolbetuximab, a first-in-class anti-Claudin-18.2 monoclonal antibody, received regulatory approval in Japan, from the United States Food and Drug Administration in 2024, and from the European Commission, following the phase III SPOTLIGHT7 and GLOW8 trials. Beyond zolbetuximab, chimeric antigen receptor T-cell (CAR-T) therapies such as satricabtagene autoleucel,9 bispecific antibodies,10 and antibody–drug conjugates,11 collectively position Claudin-18.2 as a gating biomarker for an expanding therapeutic class rather than a single agent. Satricabtagene autoleucel has since become the first CAR-T therapy to show a randomized progression-free survival benefit in a solid tumor, in the phase II CT041-ST-01 trial.12 Because patient eligibility for all of these modalities depends on Claudin-18.2 immunohistochemistry (IHC), the accuracy and reproducibility of a single biomarker test now influences access to multiple treatment lines. Spatial heterogeneity of biomarker expression is a well-recognized challenge in precision oncology. For programmed death-ligand 1 (PD-L1) in gastric cancer, discordance between primary tumors and matched metastatic lesions has been extensively documented and shown to affect treatment selection.13 Whether Claudin-18.2 shows comparable spatial variability and the magnitude of discordance clinicians should expect when interpreting a single-specimen result has not been systematically quantified. In clinical practice, Claudin-18.2 IHC is interpreted as a binary positive/negative classification to guide treatment decisions. However, several factors may produce discordant results across samples from the same patient. Technical and methodological factors, including antibody clone, positivity threshold, tissue fixation, staining protocol, and assay platform, can influence classification independently of tumor biology. Biological factors, in particular spatial heterogeneity of protein expression, may lead to discordant expression patterns within the same primary tumor (intratumoral discordance) or disagreement in positivity classification between the primary tumor and metastases (intertumoral discordance). Tissue sampling characteristics further compound the problem. Endoscopic biopsies capture only a small fraction of tumor architecture, whereas surgical resection specimens provide broader spatial coverage but remain subject to block-selection effects and may not represent the expression profile of distant metastatic deposits. Interobserver scoring variability adds another layer of uncertainty. Together, these sources of error create a composite risk of misclassification that is difficult to disentangle from individual studies. Existing studies demonstrate substantial variability in both intratumoral and primary–metastatic Claudin-18.2 expression, with reported heterogeneity influenced by tissue sampling, assay methodology, and positivity criteria.14 This wide range likely reflects a combination of biological heterogeneity, including clonal evolution and microenvironment-driven adaptation, and methodological inconsistency across study designs and assay definitions. The absence of pooled estimates makes it difficult for clinicians to calibrate how much weight to place on a single Claudin-18.2 test result when making treatment decisions, including the choice of biopsy site, the need for repeat or multisite testing, and the extent to which a single specimen reliably represents patient-level biomarker status. No prior meta-analysis has quantified the magnitude of spatial Claudin-18.2 discordance or systematically identified the factors driving variation across studies. We therefore conducted a systematic review and meta-analysis to estimate pooled intratumoral and intertumoral Claudin-18.2 discordance rates, specifically in gastrointestinal malignancies, and to explore study-level characteristics, stage, metastatic location, assay type, positivity cut-offs, and geographic region, as potential sources of between-study heterogeneity. By synthesizing the available evidence, this work aims to provide a pooled quantitative framework that clinicians currently lack to calibrate the reliability of a single Claudin-18.2 test result when selecting patients for targeted therapy. Materials and methods Registration and reporting framework This systematic review was registered in PROSPERO before data extraction and analysis (CRD420251168167). Reporting followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guideline.15 Ethical approval was not required because the study synthesized data from previously published reports and did not involve direct contact with human participants or access to identifiable patient-level data. Literature search A comprehensive literature search was originally conducted on 3 June 2025 and rerun on 16 June 2026, in MEDLINE via Ovid, Embase.com, and Scopus to identify English-language studies evaluating Claudin-18.2 expression across human cancers; records retrieved in the 2026 search were deduplicated against those identified in the original search before screening. Search strings combined free-text terms, proximity operators, and Boolean operators tailored to each database. Conference abstracts and conference reviews were filtered from Embase.com via the study design tags. The full search strategy is provided in Supplementary Table S1 (available at https://doi.org/10.1016/j.esmogo.2026.100412). Eligibility criteria and study selection We included original research studies with case-control, cross-sectional, or cohort designs that assessed Claudin-18.2 expression in adults with any cancer type using IHC or other validated assays carried out on biopsy material, surgical resections, or cytology specimens. To address spatial heterogeneity, eligible studies were required to report Claudin-18.2 assessment across either (i) a single whole-slide analysis or spatially distinct samples from the same primary tumor (intratumoral comparisons) and/or (ii) paired primary tumors and matched metastatic lesions assessed using the same measurement approach (intertumoral comparisons). Studies were further required to provide sufficient information to derive a spatial discordance proportion. For intratumoral analyses, discordance was operationalized according to each study’s reported evidence of spatial variation in Claudin-18.2 expression within a whole slide or across spatially distinct regions, irrespective of whether this variation crossed the study-specific positivity threshold. For intertumoral analyses, discordance was defined as disagreement in positivity status between paired primary and metastatic samples using the study’s prespecified cut-off. Studies whose concordance analyses compared samples defined by time of collection rather than spatial origin, such as archival-versus-baseline comparisons,16 were not eligible because these conflate temporal change and prior-treatment effects with spatial heterogeneity. We excluded non-English articles, book chapters, review articles, and (study) protocols. After deduplication, two reviewers (TV and ZB) independently screened titles and abstracts for eligibility, and disagreements were resolved by consensus. Full-text assessment was then carried out by a single reader (TV) to confirm eligibility against the prespecified criteria. Data extraction and risk-of-bias assessment Full-text assessment was carried out to include articles that contributed extractable discordant outcomes. Eighteen studies contributed to the intratumoral analysis17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34 and 18 to the intertumoral analysis18,19,26,27,31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44 in the cross-tumor set, with eight contributing to both (28 unique studies). Data extraction captured study identifiers and design (including setting and recruitment approach), tumor type and stage, cohort characteristics, and details of the tissue specimen tested. We extracted sampling-related variables where reported, including sampling route, specimen type (biopsy versus resection), number of sampled blocks or cores, and whether analyses were carried out on whole sections or tissue microarrays. Assay-related variables included assay type [prescription-diagnostic (RxDx) assay or laboratory-developed test (LDT)], antibody clone, staining platform and automation, digitization and scoring workflow, and whether inter-reader agreement was assessed. We additionally extracted the definition of Claudin-18.2 positivity, including percentage thresholds, intensity requirements, and the scoring metric used. Discordance outcomes were extracted using the study-specific definitions applicable to each analysis: the proportion showing spatial variation in expression for intratumoral analyses and the proportion of matched primary–metastatic pairs with disagreement in positivity classification for intertumoral analyses. If studies explored multiple cut-off thresholds, the threshold closest to the standard ≥75% moderate-to-strong staining criterion was selected for the pooled analysis. Risk-of-bias and applicability concerns were assessed using the Quality Assessment of Diagnostic Accuracy Studies (QUADAS)-2 tool,45 adapted to the present question, which focuses on the prevalence of discordance rather than diagnostic accuracy against a reference standard. Accordingly, the reference-standard domain was deemed inapplicable and the assessment emphasized the risks of bias arising from patient selection, sampling, measurement procedures, and the completeness of outcome reporting. Applicability concerns were assessed to determine whether the included study populations and testing approaches were sufficiently aligned with the review question. Studies analyzing tissue collected in prospective trials were classified as having a retrospective analytic design if biomarker assessment and discordance analysis were carried out after the trial. Statistical analysis The primary outcomes were the proportions of intratumoral and intertumoral Claudin-18.2 discordance (P), operationalized separately because of differences in primary-study reporting. Intratumoral discordance represented the study-reported proportion of evaluated tumors or patients showing spatial variation in staining intensity and/or extent, whereas intertumoral discordance represented the proportion of matched primary–metastatic pairs with disagreement in positivity classification. Confidence intervals (CIs) for individual study proportions were calculated on the original scale using Wilson intervals and pooled estimates were calculated separately for the two outcomes. Given the current clinical implementation of Claudin-18.2 testing in gastric and gastroesophageal cancer, analyses restricted to these tumor types were treated as the primary analyses and are presented in the main text; cross-tumor analyses including other malignancies were considered secondary exploratory analyses and are reported in the Supplementary Material (available at https://doi.org/10.1016/j.esmogo.2026.100412). To address heteroscedasticity (unequal variance) and the pooling of proportions near the boundaries, study-level discordance proportions were transformed to the logit scale. This approach was used to achieve improved distributional normality and variance stability, using. Logit-transformed proportions were pooled using a random-effects model with Hartung–Knapp–Sidik–Jonkman t-based CIs and pooled estimates were backtransformed for reporting on the proportion scale. Ninety-five percent prediction intervals were calculated on the logit scale using the Hartung–Knapp variance of the pooled estimate, the estimated between-study variance, and a t-distribution with k−2 degrees of freedom and were subsequently backtransformed to the proportion scale. Between-study heterogeneity was quantified using Cochran’s Q, I2, and Tau2 (Paule–Mandel method). As a sensitivity analysis, pooled proportions were also estimated using an intercept-only binomial-normal generalized linear mixed model (GLMM) with a logit link and study-level random effects. The marginal likelihood was evaluated using adaptive Gauss–Hermite quadrature and 95% CIs were obtained using profile likelihood. This model used the study-level discordant-event counts and denominators directly and required no continuity correction. Robustness to individual studies was assessed using a leave-one-out sensitivity analysis, in which pooled estimates were recalculated after sequentially omitting each study. This approach identifies studies that materially shift the pooled discordance estimate, outside the original 95% CI, indicating potential undue influence. Additionally, publication bias was assessed using visual inspection of funnel plots and Egger’s regression test. Subgroup stratification analyses were carried out for variables that were most consistently extractable across studies. This included geographic location, sample size, disease stage, metastatic location, treatment history, and methodological characteristics such as assay type, positivity cut-offs, and staining automation. Between-group differences were assessed using z-tests on the logit scale, comparing subgroup-specific pooled logit estimates using their random-effects standard errors. These findings were interpreted as exploratory, given the limited number of studies within some strata. Sex-stratified analysis was not feasible because the included studies did not consistently report sex-disaggregated discordance outcomes. We also explored stratification by risk-of-bias ratings derived from the adapted QUADAS-2 domains. All analyses were carried out in Python v3.13.14 (Python Software Foundation, Beaverton, OR) using numpy (2.5.1), pandas (3.0.3), matplotlib (3.11.0), seaborn (0.13.2), scipy (1.18.0), and statsmodels (0.14.6). The code used for the analyses presented in this study is available in the following GitHub repository: https://github.com/nki-radiology/claudin_discordance. Results Study inclusion and risk of bias The literature search, overall, identified 2892 records across MEDLINE (n = 852), Embase (n = 1104), and Scopus (n = 936). After removal of duplicates (n = 1696), 1196 unique records underwent title and abstract screening by two independent reviewers, resulting in the exclusion of 1141 records. Fifty-five full-text reports were assessed for eligibility and 28 unique studies met the inclusion criteria. Because intratumoral and intertumoral discordance were analyzed as distinct outcomes, 18 studies contributed to the intratumoral discordance analysis and 18 to the intertumoral discordance analysis, with 8 studies contributing to both. The primary analyses were restricted to gastric and gastroesophageal cancer and included 13 studies for intratumoral discordance and 12 for intertumoral discordance; these analyses form the basis of the results presented in the main text (Figure 1). Cross-tumor analyses, including nongastric malignancies, are reported as secondary exploratory analyses in the Supplementary Material (available at https://doi.org/10.1016/j.esmogo.2026.100412). The studies contributing to each analysis are listed in Supplementary Tables S2 and S3 (available at https://doi.org/10.1016/j.esmogo.2026.100412). Flow diagram of study selection. Detailed breakdown of the identification search rounds, screening, and inclusion of studies based on intra- or intertumoral discordance, as well as all cancer types versus gastric-only focused studies. Multiple studies cover both intra- and intertumoral discordance. Risk of bias was assessed using an adapted QUADAS-2 framework (Supplementary Figure S1, available at https://doi.org/10.1016/j.esmogo.2026.100412). In intratumoral discordance studies, the greatest proportion of high-risk ratings was observed in the ‘Index Test’ domain (38.5%), followed by ‘Flow and Timing’ (30.8%), and ‘patient selection’ (7.7%). In intertumoral discordance studies, the overall risk was also highest for ‘Index Test’ (41.7%), followed by ‘Flow and Timing’ (25%) and ‘patient selection’ (16.7%). Applicability concerns were low across all included studies for both outcomes (only one article was deemed unclear for intratumoral studies for the patient selection domain). Study characteristics The intratumoral discordance meta-analysis included 3097 patients across 13 studies, with a median cohort size of 102 patients [interquartile range (IQR) 62-177], and a median age of 60 years (range 19-95). Studies were published between 2019 and 2026. Cohorts were derived from archival samples collected between 1996 and 2024, and all studies were retrospective. Studies were conducted in six countries, with the most common being South Korea (46.2%), Italy (15.4%), and Japan (15.4%), and additional cohorts from the United States, Germany, and Spain. Intratumoral comparisons were carried out within primary tumors using tissue specimens, most commonly derived from surgical resection alone (76.9%), with a minority combining sample types (23.1%). Treatment status was reported in 12 studies: 5 included treatment-naïve populations and 7 included previously treated populations. Claudin-18.2 expression was assessed using an RxDx assay in six studies and an LDT in seven. Eleven of the 13 studies used the ≥75% moderate-to-strong staining threshold; the remaining 2 used alternative positivity definitions. Of the 13 studies, 4 included only patients with stage IV disease, 4 included only patients with stage I–III disease, and 5 included patients across all disease stages. Detailed study characteristics are summarized in Table 1. Table 1Summary characteristics of included studiesParameterIntratumoral discordanceIntertumoral discordanceTotal studies1312Total pooled patients3097987Median cohort size, IQR10262-1777948-119Median age, range6019-956419-95Publication year range2019-20262019-2026Sample collection period1996-20241992-2024Retrospective analytic design100%100%Countryn%n% South Korea646.2325.0 Italy215.4216.7 Japan215.4433.3 USA17.718.3 Germany17.7216.7 Spain17.700.0Sample type Surgical resection1076.9325.0 Biopsy00.0216.7 Mixed biopsy and resection323.1541.7 Not specified00.0216.7Cut-off ≥75%1184.6975.0 Other215.4325.0Assay type RxDx646.2758.3 LDT753.8541.7Cancer stage Stage I–III430.8541.7 Stage IV430.800.0 Mixed538.5758.3Treatment history Treatment-naïve538.500.0 Previously treated753.8866.7 Unknown17.7433.3Descriptive overview of gastric and gastroesophageal cohorts included in the intratumoral and intertumoral discordance meta-analyses.IQR, interquartile range; LDT, laboratory-developed test; RxDx, prescription/diagnostic. Summary characteristics of included studies Descriptive overview of gastric and gastroesophageal cohorts included in the intratumoral and intertumoral discordance meta-analyses. IQR, interquartile range; LDT, laboratory-developed test; RxDx, prescription/diagnostic. The intertumoral discordance meta-analysis included 987 patients across 12 studies, with a median cohort size of 79 patients (IQR 48-119), and a median age of 64 years (range 19-95). Studies were published between 2019 and 2026 and included samples collected between 1992 and 2024. All studies were retrospective and conducted predominantly in Japan (33.3%) and South Korea (25%), with additional cohorts from Germany, Italy, and the United States. Intertumoral discordance was assessed by comparing Claudin-18.2 status between primary tumors and matched metastatic lesions. Tissue specimens were used in all studies. Specimens were obtained from surgical resection alone in 25% of studies, biopsy alone in 16.7%, and most commonly via a combination of biopsy and surgical resection specimens (41.7%); the specimen source was not clearly specified in 16.7%. Treatment history was reported inconsistently: eight studies included previously treated populations and four did not clearly report treatment status. Seven studies used an RxDx assay to assess Claudin-18.2 expression, whereas five used LDTs. Nine of the 12 studies used the ≥75% moderate-to-strong staining threshold; 3 used alternative positivity definitions. Seven studies included patients across all disease stages, whereas five included only patients with stage I–III disease. Detailed study characteristics are reported in Table 1. Pooled Claudin-18.2 discordance rates Across 13 studies evaluating intratumoral discordance, the pooled discordance proportion was 35.8% (95% CI 20.5-54.7; Figure 2A). Heterogeneity was high (Q = 210.6, P < 0.001, I2 = 94.3%, Tau2 = 1.55), consistent with variation in sampling strategies, assay definitions, thresholding, and sample size. Leave-one-out sensitivity analysis did not identify any study whose removal shifted the pooled estimate beyond the original 95% CI. Funnel-plot inspection and Egger’s regression test showed no evidence of small-study effects/publication bias (P = 0.298; Supplementary Figure S2A, available at https://doi.org/10.1016/j.esmogo.2026.100412). Intratumoral and intertumoral Claudin-18.2 discordance meta-analysis. (A) Forest plot showing study-specific Claudin-18.2 discordance proportions within primary tumors. (B) Forest plot showing study-specific proportions of Claudin-18.2 discordance between primary tumors and matched metastatic lesions. Points indicate individual study estimates and horizontal bars show 95% confidence intervals (CIs). The pooled discordance proportion (random-effects model) is shown in red with its 95% CI. Across 12 studies evaluating intertumoral discordance, the pooled discordance proportion was 20.3% (95% CI 17.3-23.7; Figure 2B). Between-study heterogeneity was much lower (Q = 13.4, P = 0.268, I2 = 17.9%, Tau2 = 0.02). No studies whose removal shifted the pooled estimate beyond the 95% CI were identified in the leave-one-out sensitivity analysis and the funnel-plot inspection with Egger’s regression did not detect funnel-plot asymmetry, although the assessment was limited by the small number of studies and, for the intratumoral analysis, substantial heterogeneity (P = 0.853; Supplementary Figure S2B, available at https://doi.org/10.1016/j.esmogo.2026.100412). The 95% prediction interval was ∼3.1%-90.6% for intratumoral heterogeneity and 15.3%-26.5% for intertumoral discordance. In the binomial-normal GLMM sensitivity analysis, the corresponding pooled estimates were 35.8% (95% profile-likelihood CI, 21.4%-53.7%) and 20.0% (95% profile-likelihood CI, 17.0%-23.2%), respectively. These estimates were closely aligned with the primary random-effects analyses. Discordance rates across subgroup stratifications To investigate potential sources of between-study heterogeneity, we conducted stratified analyses by study characteristics that could be consistently extracted and allowed for interpretable comparisons. The most informative stratifications were based on geographic location, sample size, cancer stage, and metastatic lesion location, as described in detail below. Additional stratifications included positivity cut-offs, assay type, and automation for both outcomes (Supplementary Tables S4 and S5, available at https://doi.org/10.1016/j.esmogo.2026.100412), and treatment history and tumor differentiation for intratumoral discordance only (Supplementary Table S4, available at https://doi.org/10.1016/j.esmogo.2026.100412). Given the limited sample sizes in several subgroups, all subgroup analyses should be interpreted as exploratory. Sample size and cohort geographic region When stratifying intratumoral studies by geographic cohort, no significant difference was observed between Asian and non-Asian cohorts (46.7% versus 21.5%, P = 0.059). Pooled estimates were numerically higher in Asian cohorts for both outcomes, although neither between-group comparison reached statistical significance (22.0% versus 18.0%, P = 0.197; Figures 3A and 4A). Studies were divided into smaller and larger cohorts using the median cohort size for each analysis. Smaller studies were found to have significantly higher intratumoral discordance (54.2% versus 23.9%, P = 0.046). This finding was not significant for intertumoral discordance (21.0% versus 19.9%, P = 0.716; Figures 3B and 4B). Stratified meta-analysis of intratumoral Claudin-18.2 discordance. Forest plots showing pooled proportions of Claudin-18.2 discordance within primary gastric tumors, stratified by (A) cohort geography, (B) sample size (median = 102), (C) cancer stage, (D) treatment history, and (E) staining automation. Points indicate pooled subgroup estimates and horizontal bars show 95% confidence intervals (CIs). Stratified meta-analysis of intertumoral Claudin-18.2 discordance. Forest plots showing pooled proportions of Claudin-18.2 discordance between primary tumors and matched metastatic lesions, stratified by (A) cohort geography, (B) sample size (median = 79), (C) lymph node metastatic locations, (D) peritoneal metastatic locations, and (E) staining automation. Points indicate pooled subgroup estimates and horizontal bars show 95% confidence intervals (CIs). Disease stage, metastatic spread, and treatment history For the intratumoral analysis, pooled estimates did not differ significantly between stage IV-only cohorts and stage I–III-only cohorts (28.9% versus 40.5%, P = 0.596; Figure 3C). Five studies with mixed-stage populations were excluded from this comparison. Stage-stratified analysis was not feasible for the intertumoral outcome because no comparable stage IV-only stratum could be formed. Intertumoral estimates also did not differ significantly between studies including versus not including liver metastases (25.2% versus 19.6%, P = 0.138), lymph node metastases (19.2% versus 23.2%, P = 0.216; Figure 4C), or peritoneal metastases (22.3% versus 17.9%, P = 0.116; Figure 4D). Treatment history was not significantly associated with intratumoral heterogeneity (44.8% versus 29.4%, P = 0.392; Figure 3D) and an equivalent intertumoral comparison was not feasible. Technical characteristics Furthermore, methodological differences between studies were assessed: positivity cut-offs, assay category, and automation of staining protocols. Intratumoral discordance studies that used automated staining methods (n = 11) had significantly lower discordance rates than studies reporting manual staining protocols (n = 2) (26.8% versus 87.6%, P < 0.001; Figure 3E). This was not evident for intertumoral studies (20.0% versus 21.2%, P = 0.709; Figure 4E). Assay category and positivity cut-off stratifications were not significant for either study type (Supplementary Tables S4 and S5, available at https://doi.org/10.1016/j.esmogo.2026.100412). Risk-of-bias and cross-tumor analyses Stratification by risk of bias per domain revealed nominal differences between groups (Supplementary Tables S6 and S7, available at https://doi.org/10.1016/j.esmogo.2026.100412). Study-characteristic stratifications were also carried out in the cross-tumor cohort and generally followed the patterns observed in the gastric cancer-specific cohorts. Significant differences were observed only among studies of intratumoral discordance when stratified by sample size and protocol automation (P = 0.005 and P = 0.027, respectively; Supplementary Table S8, available at https://doi.org/10.1016/j.esmogo.2026.100412). The treatment-naïve stratum, however, for intertumoral discordance studies comprised a single pancreatic study (Supplementary Table S9, available at https://doi.org/10.1016/j.esmogo.2026.100412). Therefore, this comparison was confounded with the primary tumor site and should not be interpreted as an effect of treatment. Discussion To our knowledge, this systematic review and meta-analysis provides the first pooled estimates of spatial Claudin-18.2 discordance in gastric and gastroesophageal malignancies. Spatial variation within the primary tumor was reported in approximately one in three cases (pooled intratumoral discordance 35.8%, 95% CI 20.5-54.7), whereas discordance between primary tumors and matched metastatic lesions affected approximately one in five patients (pooled intertumoral discordance 20.3%, 95% CI 17.3-23.7). These results highlight an important vulnerability in single-specimen testing strategies when Claudin-18.2 is used to determine eligibility for targeted therapy. Moreover, intratumoral discordance was significantly higher in studies with smaller cohorts and manually carried out stainings. Although these estimates likely reflect true biological heterogeneity, their magnitude is also shaped by methodological variation, including assay type, antibody clone, scoring criteria, and sampling strategies. Quantifying discordance is therefore central to interpreting test results and optimizing biomarker-guided treatment decisions. Beyond its overall magnitude, patterns of discordance within gastric cancer and across tumor sites provide further biological context. Intertumoral discordance was lower than intratumoral discordance, likely reflecting the clonal bottleneck model of metastasis, whereby metastatic lesions arise from a limited number of dominant tumor clones.46,47 Alternatively, higher intratumoral discordance could partly reflect selection bias, as some studies restricted their assessment to primary tumors already suspected to be Claudin-18.2-positive, compressing the range of observable discordance. Upon stratification analyses within gastric-specific cohorts, higher intratumoral discordance was observed in studies with smaller cohorts. This may reflect random instability, more intensive spatial sampling, methodological differences, or a conventional small-study effect. Because cohort size was dichotomized at the median and the analysis was exploratory, no single explanation can be established. Additionally, laboratory methodology emerged as a potentially modifiable contributor to intratumoral discordance. Studies using automated staining protocols showed substantially lower pooled discordance than the two studies using manual methods. Although this exploratory study-level comparison may be influenced by differences in assay, specimen, and cohort characteristics, the magnitude of the observed contrast supports further evaluation of staining automation as a means of reducing technical variability. For clinicians and laboratories, these findings reinforce the importance of validated, standardized workflows and of interpreting Claudin-18.2 results in the context of the staining method used. Given that Claudin-18.2 is a pan-tumor therapeutic target, we included cross-tumor analyses in the Supplementary Material (available at https://doi.org/10.1016/j.esmogo.2026.100412) to assess whether spatial discordance extends beyond gastric cancer. In non-neoplastic tissue, Claudin-18.2 expression is largely restricted to differentiated gastric epithelium; however, aberrant expression has also been reported in pancreatic neoplasms, ovarian mucinous carcinomas, and cholangiocarcinoma.48, 49, 50, 51 This likely reflects tumor plasticity, shared developmental origins, or aberrant activation of pathways involved in gastric differentiation.52,53 In contrast, gastric carcinogenesis is characterized by progressive architectural disruption and phenotypic divergence from normal gastric epithelium. Claudin-18.2 expression is lower in gastric tumor tissue than in adjacent non-neoplastic tissue, is associated with histological phenotype, and can change during chemoimmunotherapy, indicating that expression is not biologically static.54,55 Stage-stratified analyses did not show a statistically significant signal but were limited by incomplete reporting and small sample sizes; therefore, they cannot rule out stage-dependent expression dynamics. Importantly, if Claudin-18.2 loss or heterogeneity correlates with aggressive biology, this may not only affect test reliability but also treatment responsiveness, particularly if greater heterogeneity corresponds to a lower fraction of targetable cells. A recent systematic review summarized the clinical and pathological parameters that correlate with Claudin-18.2 expression in advanced gastric and gastroesophageal adenocarcinoma, including age, Lauren classification, liver and peritoneal metastasis, human epidermal growth factor receptor 2 (HER2) status, and PD-L1 coexpression, and emphasized that variable antibodies and positivity cut-offs limit firm conclusions.56 Our analysis complements that synthesis by extending the question from expression prevalence to spatial reliability. The numerically higher intertumoral discordance observed in peritoneal metastases is consistent with previous reports suggesting distinct Claudin-18.2 behavior at this metastatic site, although this finding should be interpreted cautiously given the lack of statistical significance. When interpreting these clinically relevant findings, several limitations should be considered. The number of eligible studies was modest and tumor-type-specific estimates remain underpowered and should be confirmed in dedicated analyses. Some cohorts were enriched for Claudin-18.2-positive cases by design and prevalence of positivity in the paired-sample subset was often unavailable. Because the range of possible discordance depends partly on the marginal prevalence of positivity, such enrichment may have restricted or otherwise altered the observed discordance rates. The available aggregate data did not permit reliable evaluation of this effect. The operational definition of discordance differed between the two analyses because of differences in primary-study reporting. Intratumoral discordance captured study-reported spatial variation in staining intensity and/or extent and did not necessarily indicate that different regions crossed the positivity threshold. Accordingly, the pooled intratumoral estimate should not be interpreted directly as a treatment-eligibility reclassification rate. Intertumoral discordance was analyzed as binary disagreement in positivity classification but directional data were not consistently reported. We therefore could not determine whether discordance more commonly reflected gain or loss of expression in metastatic tissue, or whether a positive or negative result from one site was the more reliable indicator of patient-level status. Lastly, subgroup analyses were constrained by incomplete reporting across cohorts and formal assessments of small-study effects are underpowered in meta-analyses with few studies. Sex-disaggregated discordance data were not available from included studies, precluding analysis of whether Claudin-18.2 spatial heterogeneity differs by sex. Beyond sex, discordance was rarely reported separately by Lauren classification, HER2 or PD-L1 status, treatment history, or claudin 18–Rho GTPase-activating protein 26 (CLDN18–ARHGAP26) fusion, precluding any test of whether spatial heterogeneity varies with these clinicopathological correlates. Future work, particularly individual-participant data meta-analyses, would enable harmonized reclassification under standardized cut-offs, modeling of continuous expression gradients across tumor sites, and evaluation of determinants of expression retention or loss during progression and dissemination. Most included studies assessed surgical resection specimens, whereas Claudin-18.2 testing for treatment eligibility is often carried out on small endoscopic biopsies, particularly in advanced disease. Because resections sample a far larger tissue area, they capture more of the underlying intratumoral heterogeneity than a single biopsy can. Small biopsies sample only a limited region and may therefore be more vulnerable to spatial sampling variation. In a recent multicenter Spanish cohort of 270 endoscopic biopsies assessed using the automated Ventana 43-14A assay, poorly cohesive tumors showed greater Claudin-18.2 staining heterogeneity, including a concentration of cases near the ≥75% treatment-eligibility threshold, while signet-ring-cell content was inversely correlated with Claudin-18.2 staining.57 However, the available studies did not always directly compare treatment-eligibility misclassification between matched biopsy and resection specimens. Our pooled estimates, consequently, cannot quantify the effect of specimen type on classification reliability and may not be directly transferable to small-biopsy testing in routine practice. Because biopsy and resection specimens differ substantially in tissue coverage, specimen adequacy, and preanalytical and analytical factors, these should be considered when interpreting Claudin-18.2 IHC; cohort-level similarity, therefore, does not establish equivalent spatial reliability at the individual-patient level.14 Clinical translation will require standardized Claudin-18.2 testing across tissue handling, scoring methods, and positivity cut-offs, given the interstudy heterogeneity driven by both biological differences and methodological variability. Differences in sampling strategies, antibody clones, assay platforms, and scoring criteria, along with subjective IHC interpretation and limited reporting of inter-reader agreement, introduce uncertainty in discordance estimates. Consequently, pooled estimates should be interpreted as average discordance across heterogeneous contexts rather than fixed clinical rates. Prospective multisite sampling studies are needed to define strategies that minimize misclassification and assess whether certain metastatic sites, such as peritoneal disease, require different sampling approaches. Integration with other actionable biomarkers, such as microsatellite instability/mismatch repair deficiency and HER2 status, may further refine patient stratification by mapping coexpression patterns. In clinical practice, these findings support prospective evaluation of repeat or multisite sampling, particularly when tissue is limited or staining lies near the eligibility threshold. Companion-diagnostic development should account for spatial sampling adequacy and evaluate whether standardized automated staining reduces technical variability. Acknowledgments Data availability This study is based on previously published data. All data are available in the cited articles. Funding None declared. Disclosure The authors have declared no conflicts of interest. Supplementary data Supplementary Figures S1-S2 and Tables S1-S9 References Associated Data Supplementary Materials Supplementary Figures S1-S2 and Tables S1-S9 Data Availability Statement This study is based on previously published data. All data are available in the cited articles.