Formalin-fixed paraffin-embedded or fresh-frozen tumor tissue from multiple clinical centers was collected retrospectively at baseline before receiving checkpoint immunotherapy. Tumor responses were evaluated according to RECIST V.1.1 criteria. Tumor specimens derived from patients with mGC (up to 90 days from treatment start) were conducted as previously described by Ayers et al.21 Of 70 specimens from five clinical centers (Nanfang Hospital of Southern Medical University, Sun Yat-sen University Cancer Center, Guangdong Provincial Hospital of Chinese Medicine, The Sixth Affiliated Hospital of Sun Yat-sen University and The First Affiliated Hospital of Sun Yat-sen University), 48 specimens were of sufficiently high quality for RNA evaluation. A minimum of approximately 80 ng of total RNA was used to measure the expression of 51 TMEscore genes, comprizing 25 TME signature A genes, 19 TME signature B genes and some checkpoint-related genes (eg, PD-L1, LAG3, PDCD1LG2, CTLA4, TIGIT, TIM3 and PDCD1), and 10 housekeeping genes (ACTB, ABCF1, B2M, G6PD, GAPDH, GUSB, PGK1, RPLPO, TFRC and TUBB) using the nCounter platform (NanoString Technologies; Seattle, Washington, USA).22 Data was normalized using the housekeeping genes.

Prospective, open-label, phase 2 trial (NCT02589496) of advanced gastric cancer was designed as a single-arm, phase 2 study at Samsung Medical Center. Immune checkpoint inhibitor (pembrolizumab) 200 mg was administered as 30 min intravenous infusion every 3 weeks until documented disease progression, unacceptable toxicity, or up to 24 months. Tumor responses were evaluated every two cycles according to RECIST V.1.1 criteria. Toxicities were graded based on the National Cancer Institute Common Terminology Criteria for Adverse Events V.4.0. Tumor sample collection, eligibility criteria, PD-L1 IHC, MSI status determination, Epstein-Barr virus (EBV) in situ hybridization, tissue genomic analysis, and RNA sequencing pipeline of this cohort were detailed in our previous research.5

Patient cohorts used in this study are summarized in online supplemental table S1. Seven genomic and transcriptomic data sets from patients with metastatic urothelial cancer treated with an anti-PD-L1 agent (NCT02951767),6 patients with metastatic melanoma and non-small-cell lung cancer treated with MAGE-3 agent-based immunotherapy (NCT00706238),23 patients with advanced melanoma treated with PD-1 blocker,24 patients with advanced melanoma treated with various types of immunotherapy from The Cancer Genome Atlas of Skin Cutaneous Melanoma (TCGA-SKCM) cohort,25 patients with melanoma treated with anti-CTLA-4 (cytotoxic T-lymphocyte-associated protein 4) or PD-1 (programmed cell death protein 1) antibody,26 and mouse model treated with anti-CTLA-427 were downloaded and analyzed to determine the predictive capacity of TMEscore and were compared with its counterparts.

For the gene expression (normalized by RMA, TPM, FPKM or housekeeping genes) matrix, the expression of each gene in a signature was standardized so that its mean expression was 0, and the SD was 1 across samples. Then, PCA was performed, and principal component 1 was extracted to serve as the gene signature score. This approach had the advantage of focusing the score on the set with the largest block of well-correlated (or anti-correlated) genes in the set, while down-weighting contributions from genes that do not track with other set members.6 15 As our previous study15 indicated, TMEscore of each patient was estimated by the formula: TMEscore = ∑ PC1_i – ∑PC1_j, where i is the signature score of clusters whose Cox coefficient is positive, and j is the expression level of the gene whose Cox coefficient is negative. The analytic code and package used to perform the TMEscore estimation are provided for non-commercial use at GitHub: https://githubcom/DongqiangZeng0808/TMEscore. To characterize the metabolism, immune microenvironment and other prevalent gene signatures activation in each tumor sample, multi-algorithms were applied to determine the pathway activity using IOBR package (https://github.com/IOBR/IOBR).28 ImmuneScore, Stromalscore, and tumor purity were assessed computationally in RNA-seq data using the ESTIMATE algorithm29 that uses gene expression signatures to infer the fraction of stromal and immune cells in tumor samples. Other computational algorithms and tools used to estimate the microenvironment were detailed in the online supplemental methods.

All differential gene analyses were conducted using the DESeq2 package.30 Differential gene expression analysis was performed using a generalized linear model with the Wald statistical test, with the assumption that underlying gene expression count data were distributed per a negative binomial distribution with DESeq2. DEGs were considered for further analysis with a q value<0.05. The adjusted p value for multiple testing was calculated using the Benjamini-Hochberg correction.31

The mutation MAF files were downloaded with TCGAbiolinks,32 and the mutation status and mutation burden were inferred from the MAF files. Mann-Whitney U test was adopted to define the significance of binary variables (wild type or mutated). We applied the Benjamini-Hochberg method to convert the p values to adjusted p values.31 The mutational signature analysis was performed using the deconstructSigs package33 in R, which selects combinations of known mutational signatures34 that account for the observed mutational profile in each sample.

Gene annotation enrichment analysis was performed with the R package clusterProfiler.35 Enrichment p values were based on 1000 permutations and subsequently adjusted for multiple testing using the Benjamini-Hochberg procedure to control the false discovery rate (FDR).31 Gene Ontology (GO) and KEGG terms were identified with a strict cut-off of p<0.01 and an FDR of less than 0.05. We also identified pathways that were up-regulated and down-regulated among groups by running a gene set enrichment analysis (GSEA)36 of the adjusted expression data for all transcripts.

To characterize the tumor processes and pathway activation status in each tumor sample, a ssGSEA algorithm37 was applied to determine the pathway activity using GO,38 KEGG39 and HALLMARK gene sets derived from MSigDB (V.6.2).40 Other prevalent gene signature scores with respect to the TME, tumor intrinsic pathway, and metabolism were calculated for each sample using the PCA algorithm by IOBR package.28

Methylation data (β values of Illumina Infinium HumanMethylation450) of The Cancer Genome Atlas of Stomach Adenocarcinoma (TCGA-STAD) patients were obtained through TCGAbiolinks.32 β values reported by the 450K Illumina platform for each probe were set as the methylation level measurement for the targeted CpG site. Methylation data quality control, normalization, and filtering of redundant probes were conducted using the pipeline of the ChAMP. Differentially methylated probes (DMP) analysis was detected by the ‘champ.DMP’ function of ChAMP package.41 DMPs were considered for further analysis with a q value <0.05. The adjusted p value for multiple testing was calculated using the Benjamini-Hochberg correction.31

The normality of the variables was tested by the Shapiro-Wilk normality test. For comparisons of two groups, statistical significance for normally distributed variables was estimated by an unpaired Student’s t-test, and non-normally distributed variables were analyzed by the Mann-Whitney U test. For comparisons of more than two groups, the Kruskal-Wallis and one-way analysis of variance tests were used for non-parametric and parametric methods, respectively. The correlation coefficient was computed by Spearman and distance correlation analyses. Χ² test and two-sided Fisher’s exact tests were used to analyze contingency tables. The cut-off values of each data set were evaluated based on the association between survival outcome and signature score in each separate data set using the survminer package. The Kaplan-Meier method was used to generate survival curves for the subgroups in each data set, and the log-rank (Mantel-Cox) test was used to determine if they were statistically different. The HRs for univariate analyses were calculated using the univariate Cox proportional hazards regression model. The sensitivity and specificity of signature scores were depicted by the receiver operating characteristic (ROC) curve and quantified by the area under the ROC using the pROC package.42 The ‘roc.test’ function of pROC package was used to compare the area under the curve (AUC) or partial AUC of two correlated or uncorrelated ROC curves. All statistical analyses were conducted using R V.3.6.3.0 (https://www.r-project.org/), and the p values were two-sided. P values of less than 0.05 were considered statistically significant.

To optimize the TME assessment for more efficient clinical translations, feature engineering (see online supplemental methods) was conducted in six ICB data sets (online supplemental table S1) and reduced TMEscore15 signature genes from 244 to 44. As previous research suggested,15 genes negatively associated with ICB response were enriched in immune exclusion phenotype (EMT/TGF-β pathway), whereas the immune relevant genes positively associated with treatment efficacy figure 1A, (online supplemental figure S1A). In several GC cohorts (online supplemental table S1), we found a consistent and closed association between the 44-gene TMEscore and the prior TMEscore measured of 244 genes (online supplemental figure S1B). Notably, the TMEscore was capable of serving as a prognostic biomarker of immunotherapy meta-cohort (GSE78220,24 IMvigor210,6 GSE93157,43 Snyder et al44 and TCGA-SKCM25) (figure 1B: TMEscore, p=0.0001; online supplemental figure S1C): TMEscoreA, p<0.0001 and online supplemental figure S1D: TMEscoreB, p=0.0396, respectively), and a predictive biomarker of ICB response in several independent cohorts (online supplemental figure S1E–L, online supplemental table S2. The AUCs of eight independent data sets indicated that the predictive value of simplifying TMEscroe (44 genes) was enhanced after dimension reduction (online supplemental figure S1E–L). In the advanced GC cohort receiving anti-PD-1 immunotherapy,5 the TMEscore yielded the highest AUC (AUC=0.891), surpassing other prevalent biomarkers, including MSI status, TMB, CPS and EBV infection (AUC=0.708, 0.672, 0.817, and 0.708, respectively) (figure 1C and online supplemental table S3), and several transcriptomic-based predictive counterparts, comprizing gene expression profile score (GEPs),18 ImmunoScore,29 CD8+ T effector score, and pan-fibroblast TGF-β response signature (Pan-F-TBRs)6(figure 1D).

We further measured expression of TMEscore genes in the tumor microenvironment, using NanoString nCounter platform22 and RNA isolated from tumor tissue obtained at baseline from 48 patients with advanced gastric cancer of multicenter before receiving ICB (table 1 and online supplemental table S4). Apparently, TMEscore achieves an overall accuracy of AUC=0.877, which is higher than other prevalent gene signature predictors6 18 21 and capturing almost all true responders (figure 1E,F). Consistent with our previous study,15 regressive tumors (complete response (CR)/partial response (PR)) were observed markedly higher TMEscoreA than stable and progressive tumors (progressive disease (PD)/stable disease (SD)), and TMEscoreB was negatively associated with the treatment efficacy of advanced GC (figure 1F, statistical p value of TMEscore, TMEscoreA and TMEscoreB were 6.1×10⁻⁶, 0.047 and 0.00046, respectively), implicating stromal activation as a critical mechanism of resistance to ICB.6 15 TMEscoreB (stromal-relevant) genes were more precise biomarker and significantly associated with treatment resistance, while TMEscoreA (immune-relevant) genes were highly expressed in a few non-responders (SD/PD) (figure 1G,H).

To provide a precise map for understanding TMEscore performance in the context of mono- and combinational immunotherapy, we further explored the NanoString result of a 48 patients gastric cancer cohort. The expression of PD-L1 is prevailingly enriched in the responsive subset (CR/PR) relative to the progressive counterparts (figure 2A–C and online supplemental table S5). Intriguingly, the PD-L2 and TIM3 were significantly higher in non-responsive tumor, suggesting that upregulations of other corresponding or bypass checkpoint pathway may contribute to the resistance of PD-1 blockades (figure 2B–D and online supplemental table S5), by which according to reports the stromal activation and T-cell exclusion were induced.6 Additionally, SYNPO was reported to be upregulated during CAF activation,45 which is the critical mechanism of ICB resistance.

The clinical benefit of ICB monotherapy for advanced gastric cancer is limited, and recent clinical trials have demonstrated that combinations of ICBs with chemotherapy, anti-vascular targeted therapy or other molecular targeted therapies significantly improve treatment outcomes such as CheckMate-649.46 47 Consequently, there will be a pressing need for biomarkers that can be applied for patient selection for anti-PD-1 immunotherapy and chemotherapy combination. Among the multicenter data of GC, 19 patients received ICB monotherapy, and 29 patients were treated with ICBs combined with chemotherapy or other inhibitors (table 1). We systematically evaluated aforementioned biomarkers in both ICB monotherapy and the combination treatment settings. The majority of ICB relevant genes and immune relevant signatures were positively related to favorable mono-immunotherapy response, corroborating former discoveries (figure 2E,F and online supplemental figure S2A,B). Whereas their predictive efficacy significantly slid in therapy combination subset, especially the signatures related with immune activation (figure 2G,H). However, the TMEscore still harbored robust predictive capacities in both settings (figure 2G,H), possibly attributing to the superiorly essential influence exerted by stromal activation during synergic treatment (online supplemental figure S2C,D). Comparable trend of PD-L2 and TIM3 expression were also exhibited in the synergic therapy. Their upregulations in progressive patients suggested the potential pivotal molecular characteristics in shaping tumor immune evasion (figure 2G,I), which also implied the existence of synchronously upregulation of immune checkpoint pertinent genes, indicating this subset of patients may be latent candidate to benefit from PD-L2 or TIM3 pathway inhibitions.

In order to assess the predictive value and underlying mechanisms of TMEscore in advanced GC systematically, we performed integrative analysis across multi-omics data of advanced GC treated with pembrolizumab as a salvage treatment (NCT02589496, N=45) (online supplemental table S3), TCGA-STAD (N=375),11 and Asian Cancer Research Group (ACRG, N=299)12 cohorts (online supplemental table S6). A combination of the TMEscore with TMB or CPS (AUC=0.964, 0.973, respectively) observed a slight elevation in the AUC compared with TMEscore alone (AUC=0.921), despite no statistically significant discrepancy observed in pairwise comparisons (online supplemental figure S2E and online supplemental table S7). Intriguingly, the TMEscore was not correlated with tumor somatic mutation burden and histology subtypes in Kim cohort (figure 2J,K). However, in markedly stratified patients, when referring to levels of some biomarkers associated with ICB responsiveness,5 48 such as tumorous PD-L1 expression evaluated using CPS, MSI status and EBV infection, respectively (figure 2L–N). Accordingly, our analyses indicated that the TME estimation might have an alternative and more amenable mechanism than that of tumor intrinsic genomic features to serve as a robust biomarker for predicting ICB responses in advanced GC.

We depicted a landscape of the TME signature score, clinicopathological features, and molecular characterization in patients with metastatic GC treated with anti-PD-1 immunotherapy5 to investigate factors potentially associated with the treatment efficacy of ICB. We observed that patients with better responses were more likely to possess EBV and MSI-H molecular subtypes but were rarely enriched in chromosomal instability (CIN), genomically stable (GS), and EMT molecular subtypes (figure 3A, EBV and MSI-H: responders (n=9), non-responders (n=0); GS and CIN: responders (n=3), non-responders (n=33); p=2.5×10⁻⁷, Fisher’s exact test). Consistent with our recent research15 in TCGA-STAD and ACRG cohorts, the TMEscore was significantly higher in patients with MSI-H and EBV subtypes, relative to CIN and GS (figure 3B, p=0.003), suggesting that the predictiveness of the TMEscore was mostly contributed to molecular phenotype stratification. We next examined the predictive capacity of gene signatures and prevalent biomarkers in stratified patients with EBV and MSI-H molecular subtypes that indicates better responses to ICBs.48 49 ROC analyses indicated that the TMEscore (AUC=0.895) was superior in predicting EBV and MSI-H molecular subtypes, compared with MSI status, TMB, CPS, EBV status, GEPs, ImmuneScore, Pan-F-TBRs, and Immune Checkpoint (AUC=0.778, 0.781, 0.797, 0.708, 0.847, 0.646, 0.764, 0.767, respectively; online supplemental figure S2F–H and online supplemental table S7).

To validate above findings, we performed the same statistical analyses in two large multi-omics GC cohorts.11 12 We next focused on TCGA-STAD cohort11 and analyzed the clinical features (figure 3C and online supplemental figure S3A). In the low TMEscore group, the MSI and EBV subtypes were largely absent, while they took the majority of the group with the high TMEscore (EBV and MSI-H: high TMEscore (n=48), low TMEscore (n=16); GS and CIN: high TMEscore (n=25), low TMEscore (n=132); p<2.2×10⁻¹⁶, χ² test; figure 3D). A similar trend was also observed in the ACRG cohort (EBV and MSI: high TMEscore (n=78), low TMEscore (n=8); other subtypes: high TMEscore (n=80), low TMEscore (n=134); p<2.2×10⁻¹⁶, χ² test; Online supplemental figure S3B–D). Intriguingly, our analyses indicated that EBV infected tumors have comparable TMEscore with MSI-H tumors in the ACRG cohort (p=0.261; figure 3E) and even possessed a higher TMEscore than that of MSI-H tumors in TCGA cohort (p=2.9×10⁻⁵; figure 3F), whereas a significantly lower tumor mutation counts than that of MSI-H tumors in both TCGA-STAD and ACRG cohorts remained (p=5.9×10⁻¹³ and 6.9×10⁻⁶, respectively; figure 3G,H). We also noted a markedly lower neoantigen load in EBV infected tumors compared with MSI-H GC in TCGA cohort (p=2.7×10⁻¹⁰, figure 3I). Correlation analysis revealed that the TMEscore was positively associated with tumor mutation burden in both data sets (TCGA-STAD: p=4.4×10⁻¹⁶; figure 3J; ACRG: p=8.6×10⁻¹¹; online supplemental figure S3D) and predicted neoantigen load in TCGA-STAD cohort (p=2.5×10⁻¹¹; online supplemental figure S3E). Collectively, the EBV subtype remained at a low level of TMB and neoantigens with a high TMEscore and immune associated signatures in Pan-Caner cohorts (online supplemental figure S4A–G and online supplemental table S8). As shown by previous research48 49 about GC cohort treated with ICBs,5 patients with EBV infection, as well as the MSI-H phenotype, had an increased potential to benefit from ICB treatment. These observations further confirmed that TMB, as a widely used predictive biomarker,50 is incapable of identifying patients with GC with EBV subtype and tumor with virus infection (online supplemental figure 1 S5A-F), which also benefit from immunotherapy. As expected, the TMEscore could identify the EBV and MSI subtypes from all patients in TCGA-STAD and ACRG cohorts with significantly higher accuracy than TMB, GEPs,18 Pan-F-TBRs,6 and Immune checkpoint score33 (DeLong test, p=2.1×10⁻⁶, 8.8×10⁻¹⁰, 1.5×10⁻³², and 2.5×10⁻⁸, respectively; figure 3K and online supplemental table S7. Sufficiently, the aforementioned data confirmed that the TMEscore might perform better in selecting candidate patients with GC that can benefit from ICB immunotherapy.

Somatic gene mutations can alter the vulnerability of cancer cells to T cells and T cell immunotherapies.44 51 52 We sought to uncover the immunogenomic determinants of therapeutic response and the tumor immune microenvironment activation of GC in two large patient cohorts (TCGA-STAD and ACRG). Mutations associated with TMEscore was identified utilizing Wilcoxon test and Fisher’s exact test (figure 4A and online supplemental table S9). Our analyses highlighted that mutation of ARID1A and PIK3CA (figure 4A), whether evaluated continuously (figure 4B,C) or binarily (online supplemental table S9), were markedly correlated with TMEscore levels in TCGA-STAD cohort, which were verified in the ACRG cohort (online supplemental figure S6A). Meanwhile, TMB was divided into high TMB group and low TMB group (cut-off=400, (online supplemental figure S6B) to analyze the relationship between TMEscore and ARID1A or PIK3CA mutations. As shown in online supplemental figure S6C,D, patients with ARID1A or PIK3CA mutations exhibited significantly higher TMEscore in the low TMB group. However, no significant trend was observed in the high TMB group. The above results suggested that in low TMB conditions, both ARID1A and PIK3CA mutations are associated with TME activation, while in high TMB conditions, the effect of ARID1A and PIK3CA mutations might be covered by the phenomenon that increasing neoantigens caused by abundant mutations further activating TME. PIK3CA is the most commonly mutated oncogene across all solid tumors.53 ARID1A deficiency, also a frequent mutation in various malignancies, has been reported to contribute to compromised mismatch repair (MMR), increased mutagenesis, and microsatellite instability genomic signature, and may cooperate with anti-PD-L1 therapy.54

Notably, we investigated further into the specific mutation locations to identify recurrent mutations with top mutation frequencies in binary TMEscore settings to visualize results by trackViewer.55 Intriguingly, p.D18550Tfs*33 and p.F2141Sfs*59 of the ARID1A mutation were highlighted in high-TMEscore tumors (figure 4D) and statistically correlated with TMEscore levels (p=0.03; figure 4E, online supplemental table S10). Gastric cancer with PIK3CA p.E545K and p.H1047R mutations were prominently enriched in the high-TMEscore group (online supplemental figure S7A, online supplemental table S10). However, limited statistical difference was observed in the continuous TMEscore despite the significant discrepancy across mutated and wild type (p=2.7×10⁻⁸; online supplemental figure S7B). Additionally, the mutation rate of ARID1A and PIK3CA in TCGA-STAD cohort were also higher in EBV and MSI molecular subtypes, which was correlated with an elevated TMEscore and immunotherapeutic response as compared with CIN and GS subtypes (ARID1A: p<2.2×10⁻¹⁶; PIK3CA: p<2.2×10⁻¹⁶; χ² test; online supplemental figure S7C,D). We further found that the ARID1A-inactivating mutation in low TMB group was correlated with an upregulated immune checkpoint, CD8+ T effector, antigen presentation process (online supplemental figure S8A), and cellular response to glutamate metabolism (online supplemental figure S8B), which collectively suggested the higher T-cell infiltration and potential benefit from the blockade of ICB. Two recent studies indicated that the mutation of signaling pathways could serve as an immunotherapy biomarker56 and suggested combination therapy opportunities.52 The current study demonstrated pathway mutations derived predominantly from MSI molecular subtype (figure 4F and online supplemental table S11) and significant mutation accumulations of almost all pathways in the high-TMEscore fraction (figure 4F and online supplemental table S11). Nevertheless, in accordance with prior results (online supplemental figure S7D), a higher PI3K pathway mutation frequency was also observed in the EBV subtype in comparison with the GS and CIN subtypes, suggesting a latent interplay between EBV infections and the PI3K signaling pathway (online supplemental figure S9A and online supplemental table S11), which may partially explain the predominant increase of the TMEscore in EBV-infected patients (figure 3F and online supplemental figure S9B). Previous studies indicated that the interaction of PIK3CA mutation and EBV protein products may activate PI3K/ATK pathway which might be an initiator in tumorigenesis and progression. PIK3CA mutation revealed high intratumoral heterogeneity characterized with three to five different PIK3CA genotypes (including wildtype) in EBV-positive gastric cancer.57 Additionally, analyzing mutation signatures in the Catalog Of Somatic Mutations In Cancer34 indicated an intimate correlation between the TMEscore and mismatch repair associated signature 6 (online supplemental figure S9C and online supplemental table S12). Collectively, large data analyses of gastric TME elucidated the estimation of ARID1A and PIK3CA mutation status as a potential biomarker for immunotherapy strategies of GC.

Given the intriguing metabolic regulations observed in the different ARID1A-mutant statuses, we further explored transcriptomic profiles and dissected the latent intrinsic mechanism contributing to the crucial predictive capacity of the TMEscore. Metabolic signatures were estimated by PCA methodology6 and comprehensively investigated in TCGA-STAD cohort. Correlation analysis highlighted that kynurenine metabolism, purine metabolism and cysteine metabolism were activated in the high-TMEscore subset, while glycogen metabolism, transsulfuration, and glycine serine metabolism were significantly upregulated in low TMEscore group (figure 5A,B). Statistical analysis suggested that kynurenine metabolism was closely correlated with a high TMEscore (p=2.0×10⁻⁵³, r=0.702; figure 5C and online supplemental table S13) and immunotherapy-favorable molecular subtypes including EBV and MSI-H (Kruskal-Wallis, p=3.3×10⁻¹⁰; figure 5C). The downregulated kynurenine metabolism was also observed to suggest T cell exclusion, which may indicate insensitivity to ICB therapy (figure 5E). Kynurenine metabolism processes may be a promising target to restore tumor-restraining T-cell immunogenicity and therefore promote ICB therapeutic efficacy in gastric cancer, such as IDO1 inhibitor.58 We observed that glycogen metabolism was significantly activated in low TMEscore tumors and immune exclusive molecular subtypes both in TCGA-STAD cohort and ACRG cohort (figure 5D and online supplemental figure S10A–F), which suggest that it may be correlated with immune exclusion phenotype (figure 5E and online supplemental figure S10G,H) and mediate treatment resistance of immunotherapy. Consistently, Curtis et al indicated that the interaction between cancer cells and CAFs supported glycogenolysis which funneled into glycolysis, leading to increased proliferation, immune evasion, and metastasis of cancer cells.59 Together, we identified a collection of metabolism characteristics and biological processes associated with TME, which reflects the intricacy of the TME and indicates potential combination therapy opportunities.

A prior study60 demonstrated that a high m6Ascore indicates an immune-exclusion TME phenotype, stromal activation, decreased survival, decreased neoantigen load, and inferior response in GC. Thereafter, we attempted to identify the epigenetic immunomodulation involved in the antitumor immunity and tumor immune editing, which may be fundamental for understanding the inflammatory reaction that occurs in the diseases. Notably, a comprehensive investigation into the DNA methylation position landscape suggested demethylation of the VAMP8, was enriched in the low-TMEscore cluster, with the demethylation of the ATG7 in the high-TMEscore cluster (figure 5F–I and online supplemental table S14). Intriguingly, further exploration of corresponding methylation regions revealed that cg04877910, cg12542933, cg05656364, cg05486094 and cg20056908 of VAMP8 methylation were consistently negatively associated with high TMEscore and MSI and EBV molecular subtypes, whereas cg23752985 of VAMP8 methylation harbored a relatively diverse distribution in molecular subtypes and correlations with the TMEscore (online supplemental figure S11A,B). Enrichment of differentially methylated genes highlighted the vital role VAMP8 methylation plays in the TME regulatory network via upregulating immune pathways, comprizing pathways of leukocyte activation regulation, protein location to the membrane, antigen processing and presentation, coated vesicle, and recycling endosome (online supplemental figure S11C), which indicated the crucial role VAMP8 plays in the complex gene interactions and crosstalk in extensive signaling pathways. Additionally, the demethylation of ATG7, as a gene marker of autophagy, is significantly correlated TMEscore (online supplemental figure S11D). Further analyses of the relationship among discovered ATG7-associated signatures (positive regulation of autophagy) indicated that demethylation of the ATG7 was contributed to the immune exclusion in TME, with elevated TMEscoreB and fibroblast infiltration in TCGA-STAD and ACRG cohort (online supplemental figure S11E,F). Collectively, DNA methylation, such as different methylation regions of VAMP8 and ATG7, may offer a lens into the complexity and diversity of the TME and immune-activity determination, thereafter might assist in optimizing immunotherapy strategies.