A tumor microenvironment-based classification of gastric cancer for more effective diagnosis and treatment

With approximately one million diagnosed cases and over 700,000 deaths recorded annually, gastric cancer (GC) is the third most common cause of cancer-related deaths worldwide. GC is a heterogeneous tumor. Thus, optimal management requires biomarkers of prognosis, treatment selection, and treatment response. The Cancer Genome Atlas program sub-classified GC into molecular subtypes, providing a framework for treatment personalization using traditional chemotherapies or biologics. Here, we report a comprehensive study of GC vascular and immune tumor microenvironment (TME)-based on stage and molecular subtypes of the disease and their correlation with outcomes. Using tissues and blood circulating biomarkers and a molecular classification, we identified cancer cell and tumor archetypes, which show that the TME evolves with the disease stage and is a major determinant of prognosis. Moreover, our TME-based subtyping strategy allowed the identification of archetype-specific prognostic biomarkers such as CDH1-mutant GC and circulating IL-6 that provided information beyond and independent of TMN staging, MSI status, and consensus molecular subtyping. The results show that integrating molecular subtyping with TME-specific biomarkers could contribute to improved patient prognostication and may provide a basis for treatment stratification, including for contemporary anti-angiogenesis and immunotherapy approaches.


Introduction
Gastric cancer (GC), with about one million new patients diagnosed yearly, is the third most common cancer with the second highest mortality rate worldwide 1,2 . The vast majority of GCs are adenocarcinomas and can be subdivided into intestinal and diffuse types according to the Lauren classi cation system 3,4 . The tumor microenvironment (TME), characterized by active angiogenesis, brosis, and chronic in ammation, is critical for the local and metastatic progression of malignant solid tumors, including GC 5 . The TME is often characterized by a structurally and functionally abnormal tumor vasculature, which limits drug delivery and suppresses the ability of the immune system to combat malignant cancer 6 . Consequently, chemotherapeutic drugs have limited e cacy in GC, and novel treatment strategies are desperately needed 5 .
Tumor growth depends on angiogenesis -the formation of new vessels -which is essential for solid tumor growth and metastasis 7 . In the absence of vascular growth, tumors cannot develop beyond a few millimeters and therefore remain dormant 8 . The vasculature is a key component of the TME, and its abnormal structure and function mediate tumor progression and treatment responses 9 .
Tumor blood vessels are highly abnormal. They have irregular shapes, diameters, and branching patterns.
As such, they cannot be classi ed as arterioles, venules, or capillaries 10,11 . The consequences of vascular abnormalities in tumors include an aberrant TME, including tissue hypoxia and immunosuppression, and poor drug and effector immune cell in ltration 12 . These abnormalities may contribute to tumor resistance to conventional chemo-, radio-, and immune-based therapies. Jain proposed that judiciously dosed anti-angiogenic treatment can normalize the tumor vasculature by reducing vascular permeability and interstitial uid pressure, thus improving blood ow and tumor perfusion 13 . A normalized vasculature can reduce hypoxia and enhance the delivery of oxygen and cytotoxic agents for radiation therapy, thereby improving the anti-tumor immune response 14 . Preclinical and clinical studies have supported the hypothesis that anti-angiogenic therapy can normalize the tumor vasculature, at least transiently 5 . Nevertheless, existing classi cation schemes do not include parameters related to the tumor vasculature or the TME, in part due to the lack of understanding of how the molecular subtype impacts the TME characteristics in GCs 15 .
Malignant cells develop a complex relationship with their TME during progression, which may be a target for enhancing treatment response 16 . Although there is abundant in vitro evidence for immune reactivity against solid tumors in patients, such responses are often ineffective due to local and systemic immunosuppression 17 . Although the immune responses of the host are critical to the success of immunotherapy, such as immune checkpoint inhibition 5 , determinants of the response are not completely understood. Tumor in ltration by immune cells, such as cytotoxic T lymphocytes, varies widely in density, composition, and clinical signi cance 6,16 . Blood vascular and lymphatic endothelial cells play important roles in the tra cking of immune cells, controlling the microenvironment and modulating the immune response. Improving access to the malignant tumor by altering the vasculature with anti-angiogenic drugs may provide an effective combinatorial strategy for immunotherapy, and might be widely applicable to many tumor types, especially in GC 18 .
Traditional staging systems are important in predicting the prognosis of patients with cancer. These systems are used to stratify patients according to prognostic variables in the setting of clinical trials, allowing the exchange of information among researchers, and nally guiding the therapeutic approach 19 .
Although surveillance protocols for patients at risk of developing GC have signi cantly improved, the clinical outcome remains poor with most patients presenting with advanced disease and not eligible for curative therapy.
At the pathologic level, GC is a morphologically heterogeneous tumor. Despite this knowledge, the staging system used worldwide is based on TMN staging, which has important limitations. Several new classi cation systems have been proposed to re ect tumor biological diversity, and their implementation may help guide classi cation, therapy, and biomarker screening for antibody-targeted therapy and immunotherapy 3 .
In this study, we focused on the relationship between our previously described protein-based classi cations of GC and the characteristics of their TMEs. This study also examined how a TME-related biomarker approach performs compared to prior approaches such as TMN staging. The results of this study may provide a classi cation system to facilitate treatment selection, for example for antiangiogenesis treatment and immunotherapy, which currently bene t only a minority of GC patients. Furthermore, these insights may be useful for designing new treatment approaches against targets in the TME of GC.

Tissue analyses
Expression of multiple GC and TME-related biomarkers was measured using EBER in situ hybridization, and immunohistochemistry (IHC) for mismatch repair proteins mucin (MUC)5AC and MUC6. EBER in situ hybridization was the gold standard for detecting Epstein-Barr virus (EBV) status and localizes the abundantly expressed long noncoding RNAs EBER1 or EBER2 in malignant cells 21,22 . IHC staining for TME markers was performed on 4µm thick sections cut from the representative formalin-xed, para nembedded tumor tissue. Sections were depara nized, and antigen retrieval was performed in a water heater with citrate buffer (pH 9) for CD31, neuron-glial 2 (NG2), carbonic anhydrase 9 (CAIX), VEGFR2, Ecadherin (CDH1), p53 or Tris ⁄ EDTA (pH 9) for programmed death ligand 1 (PD-L1) and CD8 at 97°C. This was followed by 0.03% hydrogen peroxide (H 2 O 2 ) treatment for 10 min to block the endogenous peroxidase. Mouse or rabbit monoclonal and polyclonal antibodies against these antibodies (Supplementary Table 1) were applied to the sections overnight at 4°C. The slides were then incubated with peroxidase-labeled polymer conjugated to goat anti-rabbit IgG or goat anti-mouse IgG for 30 min. The sections were stained for the appropriate time with DAB, and counterstained with hematoxylin.
MSI analysis for molecular-based classi cation MSI testing was performed using the MSI Analysis System, Version 1.2 kit (Promega), which allows for the analysis of ve mononucleotide repeat markers (BAT-25, BAT-26, NR-21, NR-24, and MONO-27) to determine MSI status, as well as two pentanucleotide repeat markers (Penta C and Penta D) used for unique sample identi cation and to detect potential sample mix-ups or contaminations. For this purpose, 1-2 ng of genomic DNA was subjected to an enzymatic ampli cation reaction using 1 µl of GoldSTR 10X Buffer and 1 µL of MSI 10X Primer Pair Mix from the kit, and 0.095 µL of Go Taq MDx Hot Start Polymerase (7.9 U/µL), in a nal volume of 10 µL. The enzymatic ampli cation conditions used were: 95ºC for 11 min; 96ºC for 1 min; 10 cycles of 94ºC for 30 sec (ramp up to 58ºC at 29% per sec), 58ºC for 30 sec (ramp up to 70ºC at 23% per sec), and 70ºC for 1 min; 20 cycles of 90ºC for 30 sec (ramp up to 58ºC at 29% per sec), 58ºC for 30 sec (ramp up to 70ºC at 23% per sec), and 70ºC for 1 min; 60ºC for 30 min; and 4ºC. The resulting reaction products were denatured with a mixture of Hi-Di formamide (Applied Biosystems) and ILS600 weight marker (included in the MSI Analysis System, Version 1.2 kit) by incubation at 95ºC for 3 minutes, followed by incubation on ice for 3 minutes, and were separated by capillary electrophoresis using the ABI PRISM® 3130 Genetic Analyzer system (Applied Biosystems). MSI status was determined using GeneMapper software (Applied Biosystems). To determine MSI status, allelic sizes of tumor and normal specimens were compared, and a marker was considered MSI unstable if there was a shift of three base pairs in the cancer samples. Specimens were classi ed as MSI-High (MSI-H) if two or more microsatellite markers were unstable, MSI-Low (MSI-L) if one marker was unstable, or microsatellite stable (MSS) if no markers were unstable.

Histopathological scoring
The immunostained sections were scored by evaluating the invasive carcinoma tissue portion. Cytoplasmic and membranous expression of CD31 (endothelial cells), CAIX (hypoxic cells), and NG2 (pericytes) was de ned in two groups (low and high) based on the median positive cell surface (%). Microvessel density (MVD) was the surface area covered by CD31-positive cells. Cytoplasmic expression of the VEGFR2 was categorized semiquantitatively based on the percentage of positive cells as follows: 0, no staining; 1, < 10% positive cells stained; 2, 10-25% positive cells stained; and 3, > 25% positive cells stained. Further analyses, we de ned VEGFR2 IHC expression in two groups (low and high) based on the median positive cell surface (%). The expression of PD-L1 was only categorized semiquantitatively based on the percentage of positive tumor cells (stroma cells) as follows: 0 (< 1% tumor cells stained); 1(≥ 1% but < 5% tumor cells stained); 2 (≥ 5% but < 10% tumor cells stained); and 3 (≥ 10% tumor cells stained) 23 . In further analyses, we de ned PD-L1 IHC expression in two groups based on the percentage of positive tumor cells: low (< 1% cells positive that include IHC score 0) and high (> 1% cells positive, that includes IHC score 1, 2 and 3). Similarly, we de ned the IHC expression of in ltrated CD8 lymphocytes in two groups: low (IHC scores 0 and 1) and high (IHC scores 2 and 3) intensity of CD8 positive in ltrate. Cytoplasmic expression of the MUC5AC and MUC6 was categorized semiquantitatively based on the percentage of positive tumor cells as follows: 0, < 10% cells positive; 1, 10-25% cells positive; and 2, > 25% cells positive. To establish a diagnostic value based on mucin expression assessed by IHC, we calculated the average of MUC5AC and MUC6 IHC scores (called MUCavg) for each GC patient. The expression of E-cadherin and p53 have been reported previously 18 . For E-cadherin, the normal expression (presence) was scored with 1, and the aberrant expression (absence) was noted as 0 (Supplementary Table 1).
The samples were classi ed according to the protein expression as previously published 18 , bringing a modi cation for group 2 (Gp2) by using an MSI analysis that involves comparing allelic pro les of microsatellite markers generated by ampli cation of DNA from matching normal and tumor samples, which may be mismatch-repair (MMR) de cient. Alleles that were present in the tumor samples but not in corresponding normal samples indicate MSI. The hierarchical clustering resulted in the determination of ve groups of gastric adenocarcinomas (Gp1: EBV-positive gastric cancers, Gp2: Microsatellite-instable gastric cancers, Gp3: Gastric cancers with aberrant E-cadherin expression, Gp4: Gastric cancers with aberrant p53 expression, Gp5: Gastric cancers with normal p53 expression).

Multiplex protein array for circulating biomarkers
We used a chemiluminescence-based multiplexed protein array (MesoScale Discovery) to measure a panel of cytokines/chemokines and angiogenic biomarkers in the serum samples collected from the GC patients in the study (n = 122) and from 51 healthy individuals. All measurements were done in duplicate in a CLIA-certi ed Core facility at MGH Boston. All samples were obtained from GC patients before surgical resection, thus allowing us to associate the circulating in ammatory factors with tissue biomarkers. The biomarkers panel used for multiplexed protein array analysis in serum samples were angiogenic biomarkers such as basic broblast growth factor (bFGF), vascular endothelial growth factor (VEGF)-A, VEGF-C, VEGF-D, soluble (s)TIE2 (TEK receptor tyrosine kinase), placental growth factor (PlGF), soluble VEGFR1 or Fms-like tyrosine kinase-1 (sFLT-1) and in ammatory biomarkers such as interleukin-8 (IL-8), IL-6, tumor necrosis factor alfa (TNF-α) and gamma interferon (IFN-γ).

Statistical analysis
Quantitative and semiquantitative analyses for tissue markers were performed with the support of experienced MGH gastrointestinal pathologists (GYL and NS). In descriptive statistics, data are presented as n (%) or median (interquartile range (IQR): Q1, Q3). Statistical signi cance of univariate analysis was determined by the Mann-Whitney-Wilcoxon test with p-values calculated by the exact method and the Kruskal-Wallis test for ordinal or continuous variables with a non-normal distribution. For normal distribution (evaluated with Shapiro test) we used an unpaired t-test with Welch's correction and one-way analysis of variance (ANOVA), a chi-square test for dichotomous variables followed by Bonferroni post hoc test for multiple comparisons. All P values have been based on 2-sided hypothesis tests. We used log-transformed serum biomarkers to normalize the protein concentration to obtain a normal distribution of data. The Kaplan-Meier method was used to calculate OS curves based on the length of time between primary surgical treatment and nal follow-up or death. The log-rank test was used to compare the survival distributions. Differences were considered signi cant when the p-value was < 0.05. A Cox proportional hazards regression was used for univariate and multivariate analyses of prognostic factors for overall survival. Multivariate survival analysis was performed on variables that were signi cant in the univariate analyses to identify independent predictors of survival. The hazard ratio (HR) and the corresponding 95% con dence interval (95% CI) were calculated. P < 0.05 was considered statistically signi cant. Overall survival (OS) was de ned as the interval between the date of surgery and the date of death or the end of follow-up. The R packages used in the survival analysis: gtsummary 1.7.1, survival 3.5.5, survminer 0.4.9. The statistical analysis used Graphpad Prism 9.5.1 for Windows (GraphPad Software, Inc) and R 4.2.3 software.

Clinicopathological characteristics
This study included 122 patients (34 females and 88 males) diagnosed with resectable GC. The median age was 65 (interquartile range, 57-70 years). The median overall survival (OS) in this cohort was 20 months. These eligible patients signed informed consent and had blood and tissue samples banked. All patients were treated by surgery. Clinical and pathologic parameters are summarized in Table 1. The median age in the control group was 40 (interquartile range, 32-48) years, and distribution by gender was unbalanced with 41 females (80%) and 10 males (20%) (Supplementary Table 2).

Stage-dependent TME characteristics of GCs
The TME-related biomarkers examined in GC surgical samples are summarized in Fig. 1a. MVD was signi cantly increased in GC tissue from patients with stage IV versus stage I (p < 0.0001) and stage II (p = 0.0016) disease, and stage III versus stage I disease (p = 0.0069) (Fig. 1b). Tissue hypoxia was signi cantly increased with GC stage, i.e., CAIX expression was higher in stage IV versus stage I (p = 0.0006) and stage II (p = 0.011) disease (Fig. 1c). In addition, pericyte (Pc) coverage of the tumor blood vessels (vessel maturity) in stage IV was signi cantly decreased compared to stage I (p < 0.0001) and stage II (p = 0.0003) disease (Fig. 1d). These results demonstrate that advanced tumors are associated with more angiogenic (structurally immature) and functionally abnormal vessels in GC. These clinicopathological parameters showed the same trend for association with lymph node status, the presence of metastasis, perineural and lymphovascular invasion, and CA19-9 tumor marker (Supplementary Fig. 1a-c). To test the correlation between these tissue biomarkers with outcomes in these GC patients, we used Kaplan-Meier survival distributions and log-rank test, based on median IHC expression of positive cell surface percentage (%), as follows: CAIX low group (< 9.9%) and CAIX high group (≥ 9.9%); MVD low group (< 0.6% ) and MVD high group (≥ 0.6% ); pericyte coverage (ratio between NG2/MVD IHC positive cell surface percentage): low group (< 1.8 median ratios) and high group (≥ 1.8 median ratios). Kaplan-Meier survival analysis revealed that the groups of patients with high MVD and high CAIX expression had signi cantly shorter survival compared to the group with low MVD (p = 0.0023) and low CAIX (p = 0.036) (Fig. 1e-f). Moreover, the group of patients with low pericyte coverage had signi cantly shorter survival compared to the group with high pericyte coverage (p = 0.0066) (Fig. 1g).
VEGFR2 is expressed across GC subtypes and is associated with tumor and immune TME biomarkers and outcome VEGF receptor 2 (VEGFR2) is a validated therapeutic target in advanced GC, based on e cacy data with the antibody ramucirumab. To determine the extent and distribution of VEGFR2 expression in GC, we used double IHC. We detected the expression of VEGFR2 in both endothelial cells (co-localized with CD31 expression), and GC cancer cells (co-localized with cytokeratin expression) (Fig. 2a). VEGFR2 expression was more pronounced intratumorally (median IHC positive cell surface percentage = 10.5%) than in peritumoral tissues (median IHC positive cell surface percentage = 1.4%) (p < 0.0001) (Fig. 2b).
Furthermore, when comparing the high versus low VEGFR2 expression groups, we found a higher expression of CAIX (p = 0.033) and lower pericyte coverage in the high VEGFR2 expression group (p = 0.026) (Fig. 2e, f). Spearman correlation con rmed that pericyte coverage was inversely correlated with VEGFR2 expression (r=-0.26, p = 0.01), but not with MVD and CAIX expression ( Supplementary Fig. 2f-i).
Correlation between TME biomarkers and PD-L1 and CD8 T cell in ltration in GC PD-L1 expression scored by IHC is frequently used for patient selection for immunotherapy. Moreover, The Cancer Genome Atlas (TCGA) project reported elevated PD-L1 expression in EBV-positive GC. Examination of the expression of PD-L1 in the tumor and in the stroma at the tumor periphery and CD8 (to detect this tumor-in ltrating lymphocyte subset) in this cohort by IHC (Fig. 3a, b) revealed that more than 50% of GC cases expressed PD-L1 in cancer and/or in in ltrating immune cells (Table 2).
Interestingly, the expression levels of PD-L1 and CD8 in GC tissues were not correlated with tumor stage (Supplementary Tables 3 and 4). However, we found that the high PD-L1 expression group had signi cantly larger tumors compared to the low-expression group (p = 0.012) (Supplementary Table 3).
Furthermore, the proportion of patients with higher PD-L1 IHC scores in the high CD8 expression group was signi cantly increased compared with the low CD8 expression group (p = 0.003) ( Table 2).  24 . Thus, we investigated whether PD-L1 and CD8 expression levels in GC were correlated with vascular-related biomarkers such as VEGFR2, MVD, and CAIX IHC expression and pericyte coverage (NG2/MVD ratio). We found that an increased MVD was signi cantly associated with the high PDL1 expression (p = 0.044) ( Table 3). However, VEGFR2, MVD, pericyte coverage, and CAIX were not signi cantly different between the GC groups with high versus low CD8 expression (Supplementary  Table 5). Circulating angiogenic and pro-in ammatory biomarkers associated with GC and its stage When comparing the levels of serum biomarkers of angiogenesis in GC patients versus healthy individuals, we found signi cantly higher levels of bFGF (p = 0.0012), PlGF (p = 0.004), sFLT-1 (p < 0.001), VEGF (p < 0.001), and VEGF-C (p < 0.001) and lower levels of sTIE2 (p < 0.001) and VEGF-D (p = 0.002) ( Supplementary Fig. 3a). Among in ammatory biomarkers, the circulating levels of IFN-γ, IL-8, and TNF-α were signi cantly higher in GC patients (p < 0.0001) (Supplementary Fig. 3b).
When we analyzed the association serum between biomarkers of angiogenesis and clinical pathological data, we found signi cantly higher sTIE2 levels for stage III and IV disease (all p < 0.05) (Fig. 4a) and higher PlGF levels for stage IV versus stage I (p = 0.0331) (Fig. 4b).
Moreover, the levels of sTIE2 directly correlated with the number of pathological metastatic lymph nodes (pN) (Fig. 4c), and sTIE2 levels directly correlated with the presence of perineural invasion (PNI+) and lymphatic vessel invasion (LVI+) (all p < 0.05) (Fig. 4d-e) and serum PlGF levels directly correlated with LVI+ (Fig. 4f). The other angiogenesis biomarkers measured did not correlate with clinical pathological data (Supplementary Table 13 and Supplementary Fig. 4).
When we analyzed the association serum between in ammatory biomarkers and clinical pathological data, we found that patients with GCs larger than the median (> 5 cm) had signi cantly higher levels of the IL-6 (p = 0.0007), IL-8 (p < 0.0001), and TNF-α (p = 0.011) compared to patients with smaller tumors (≤ 5cm) (Fig. 4g-i). Moreover, serum IL-8 showed signi cantly higher levels in the group of GC patients with LVI + compared to the group without LVI-(p = 0.046) (Fig. 4j). Interestingly, we found signi cantly higher expression levels of serum IL-6 (p = 0.011) and IL-8 (p = 0.018) in the GC patients with serum CA19-9 greater than 37U/mL (standard clinical cut-off) (Fig. 4k, l). The other serum proin ammatory biomarkers measured showed no association with the clinical-pathologic characteristics (Supplementary Table 13 and Supplementary Fig. 5).
The TIE2 receptor is expressed on endothelial cells, and together with its ligand angiopoietin (Ang)-1 and Ang-2, plays critical roles in angiogenesis and vessel maturation. The binding of Ang-1 to TIE2 maintains and stabilizes mature vessels by promoting interactions between endothelial cells and the surrounding extracellular matrix 28 .

Circulating angiogenic and pro-in ammatory biomarkers associated with vascular and immune TMEbiomarkers and with OS in GC
We next tested the associations between the circulating levels and the tissue TME biomarkers, assessed by IHC. We found signi cant direct correlations between tissue VEGFR2 expression and serum sTIE2 (p = 0.040), between MVD and serum PlGF (p = 0.023), and between MVD and serum sTIE2 (p = 0.013) (Fig. 5a-c). Interestingly, serum VEGF was inversely associated with tissue hypoxia measured by CAIX IHC (p = 0.025) (Fig. 5d). Moreover, MVD was also directly associated with the serum levels of the proin ammatory biomarkers IL-6 (p = 0.015) and IL-8 (p = 0.0073) (Fig. 5e, f). Furthermore, we observed signi cantly higher expression levels of serum IL-6 (p = 0.017), IL-8 (p = 0.013), and TNF-α (p = 0.013) in the low pericyte coverage group compared to the high pericyte coverage group (Fig. 5g-i). The relationships between the remaining serum angiogenic and proin ammatory biomarkers and the expression of the VEGFR2, MVD, and CAIX markers assessed by IHC did not indicate a signi cant correlation ( Supplementary Figs. 6 and 7).
To examine the diagnostic biomarker signi cance of serum angiogenic and proin ammatory molecules in resectable GC patients, we established the optimal cut-off value to discriminate between the cancer patients from the control group based on the Youden index and ROC curve analysis (using 'OptimalCutpoints v1.1.5' R Package for Computing Optimal Cutpoints in Diagnostic Tests). We found that serum VEGF, VEGFC, sFLT1, sTIE2, IL-8, and TNF-α levels could identify patients with GC based on the optimal cut-off value with AUC > 0.8 ( Supplementary Fig. 9).
Next, we examined the prognostic biomarker signi cance of serum angiogenic and proin ammatory molecules in resectable GC patients using optimal cut-off values and the Kaplan-Meier method for OS distributions (n = 121). We found a signi cantly shorter OS for the GC patients with high (median OS = 10 months) versus low (median = 34 months) serum sTIE2 (4,428.64pg/mL cut-off value; p < 0.0001) and for high (median OS = 11.5 months) versus low (median = 21 months) serum VEGF-D (1,116.25pg/mL cut-off value; p = 0.027) levels (Fig. 6a, b).
Tissue and circulating biomarker association with molecular-based subtypes of GC  Table 15). GC patients with Gp2 tumors had a signi cantly longer median OS (158 months) compared to the other molecular groups (median OS was 28.5 months in Gp1, 13.5 months in Gp3, 19 months in Gp4, and 13 months in Gp5) (p < 0.05) (Fig. 6c and Supplementary Table 16).
Next, we examined the levels of tissue and circulating biomarkers in the ve GC molecular classes. GCs from the Gp2 class showed non-statistically signi cant trends for higher PD-L1 expression at the tumor periphery and pericyte coverage of vessels (after the Bonferroni correction test) ( Supplementary  Tables 17 and 18).

Prognostic signi cance of clinicopathological parameters, TME-and circulating biomarkers in GC patients
We performed a Cox regression analysis of the association between OS and clinicopathological features, and tissue and blood biomarkers in these GC patients. The univariate Cox proportional hazards regression analysis con rmed the higher risk of death for patients with a higher number of metastatic lymph nodes, pN1 (HR = 2.70, p < 0.001) and pN2 (HR = 9.28, p < 0.001), as well as with more advanced stage: stage III/IV (HR = 3.08, p < 0.001). No association was detected between OS and age, sex, differentiation degree, or tumor size. In addition, a higher risk of death was associated with the molecular groups Gp3 (HR = 4.26, p = 0.004), Gp4 (HR = 2.88, p = 0.027), and Gp5 (HR = 3.28, p = 0.028) versus Gp2 (Table 4 and Supplementary Table 21). Univariate analyses of tissue biomarkers identi ed an association with a higher risk of death for elevated expression of tissue VEGFR2 (HR = 1.83, p = 0.004) and higher MVD (HR = 1.93, p = 0.003) and a non-signi cant trend for higher MUCavg (HR = 0.67, p = 0.054), (Table 4 and Supplementary Table 22). Among the serum biomarkers, a higher risk of death was associated with elevated circulating sTIE2 (HR = 2.13, p = 0.001), IL-6 (HR = 1.35, p < 0.001), and IL-8 (HR = 1.25, p = 0.012) ( Table 4 and Supplementary Table 23). In bold text, p values less than 0.05 Finally, we investigated the signi cant variables to describe how they correlate with OS. To this end, we performed a multivariate Cox regression analysis, using the proportional hazards assumption for the Cox model using statistical tests and graphical diagnostics based on the scaled Schoenfeld residuals, when including all variables that achieved statistical signi cance in the univariate analysis ( Supplementary  Fig. 10). The results showed that a higher risk of death was directly associated with pN2 (HR = 13.2, p < 0.001), pN1 (HR = 3.61, p < 0.001), molecular classi cation Gp3 (HR = 4.28, p = 0.021), and serum IL-6 (HR = 1.40, p = 0.006) ( Table 4).

Discussion
Currently, GC prognostication and treatment selection for immunotherapy is based on clinicopathological features and PD-L1 expression score. In this study, we performed an integrative analysis of tissue and blood biomarkers in GC patients with different disease stages, including a strati cation based on molecular classi cation. This approach adds to previous stage subtyping strategies that have not considered the TME and allowed the development of a model for selecting the targeted therapies in GC ( Supplementary Fig. 11).
Biomarkers for cancer treatment and diagnosis are de ned as biological variables that correlate with biological outcomes, and cancer biomarker discovery strategies that target expressed proteins are becoming increasingly popular 29 . Multiple studies have focused on identifying tumor biomarkers that could facilitate the earlier diagnosis of GC, understanding its behavior, and improving its treatment 30 .
However, no single biomarker has been proven predictive of outcomes after antiangiogenic or immunotherapeutic treatments of GC 31 . In addition, few studies have simultaneously evaluated more than one candidate biomarker to enhance the diagnostic sensitivity and speci city of GC 32 . Moreover, there is an increasing interest in developing immunotherapies for earlier GC stages, and many of the current studies are testing combination therapies (anti-angiogenic and immune checkpoint inhibitors, e.g. ClinicalTrials.gov Identi ers: NCT03995017, NCT02443324, NCT04879368, NCT04632459, NCT02999295). Therefore, a logical development to improve the effective diagnosis and treatment of GC is to simultaneously screen for multiple biomarkers, such as TME and immune biomarkers, to increase predictive and response biomarker performance across disease stages.
GCs are in ammation-induced malignancies because they often occur in a diseased stomach on the background of gastritis 33 . This includes the bacterium Helicobacter pylori and EBV 34 . These features may be relevant for the immune TME. Indeed, the TCGA project reported elevated PD-L1 expression in approximately 15% of EBV-positive GCs. Evaluation of mRNA revealed elevated expression of JAK2, PD-L1, and PD-L2 1 . In addition, Lin et al. reported that non-Asian GC was signi cantly enriched in signatures related to T-cell biology, including CTLA-4 signaling. Underlying chronic in ammation and/or viral infections create an immune suppressive TME in GC through the production of cytokines including IL-6, IL-11, TNF-α, and transforming growth factor (TGF)-ß 35,36 . We found that lymph node metastases, pathological stage, molecular classi cation, VEGFR2 expression by IHC, MVD by IHC, and circulating sTIE2, IL-6, and IL-8 were signi cantly associated with prognosis. After controlling for the confounding factors, multivariate Cox regression analysis revealed that molecular classi cation and circulating IL-6 were independent biomarkers of OS.
The in ltration of tumors by immune suppressive leukocytes such as Tregs and myeloid-derived suppressor cells is another important mechanism of immune evasion in cancer. Exhaustion of CD4 + Tcells has also been reported as a mechanism of immune evasion in patients with advanced cancer 37,38 .
Furthermore, while the immune response to speci c antigens is recognized by major histocompatibility receptors, co-stimulatory and co-inhibitory molecules regulate the intensity of the response. Immune checkpoints involve co-inhibitory molecules that are physiologically expressed for the maintenance of self-tolerance 39 . In the cancer microenvironment, including GC, immune checkpoint molecules such as CTLA-4 and PD-L1 are overexpressed and broadly induce the evasive mechanism. Based on these reports, we evaluated the relationship between the expression of PD-L1 and in ltrated CD8 + lymphocytes with the tumor microenvironment in GC samples. In our study, PD-1/PD-L1 expression was detected in clinical samples and signi cantly correlated with tumor progression. We also found that these immune system factors were associated with normalized tumor vasculature and may be important factors in classifying GCs and rationally selecting an effective treatment strategy.
Angiogenesis is a key process in the progression and metastasis of solid malignant tumors. But tumor angiogenesis results in abnormal blood vessels, which create an abnormal TME 40 . Numerous reports revealed that most tumor vessels have deformed diameters, and tumor endothelial cells have loose interconnections, intercellular openings, and abnormal pericytes that are likely to be responsible for vessel leakiness. Moreover, structural abnormalities in the basement membrane of tumor blood vessels are responsible for their relative immaturity compared with normal blood vessels. Accordingly, a tumor blood vessel has abnormal blood ow, is excessively leaky, and has aberrantly high interstitial uid pressure 9,13,41 . Insu cient perfusion of the tumor tissue in hypovascular areas leads to hypoxia and necrosis. Oncologists and cancer researchers should monitor changes in the aberrant TME for effective treatment and early diagnosis in GC patients 42 . In the current study, we demonstrate that tumor vasculature and TME become increasingly abnormal in advanced stages.
Although MVD is increased in advanced GC, the TME is characterized by hypoxia 5 . Hypoxia may promote GC growth and progression, and resistance to existing therapies. Our study demonstrates that hypoxia is increased in advanced stages of GC. Tumor hypoxia stabilizes HIF1α which induces hypoxia-responsive genes including VEGF. Pathological VEGF expression is a key factor for the abnormal structure and function of tumor vessels 40 . On the other hand, inhibition of the VEGF pathway with approved drugs in advanced GC may result over time and in a dose-dependent manner in excessive pruning of the tumor vasculature, which may also induces hypoxia 43 . Conversely, inducing vessel normalization, alleviating hypoxia, and normalizing the TME might delay GC progression and metastasis 12 .
Thus, we propose that dose titration and identifying the optimal duration of selective anti-VEGF agents are warranted to enhance anti-tumor immunity in GC 6 . This is consistent with our results in advanced liver cancer mouse models 44 and early-stage clinical studies, including in advanced GC patients 45 .
Interestingly, in GC, immunoreactivity for VEGFR2 (the main VEGF receptor responsible for angiogenesis and vascular permeability) was localized to the cytoplasm of endothelial cells but was also present in tumor cells from several histologic subtypes 46 . We also found that VEGFR2 may be useful for classifying GC, diagnosis, and potential for new combination therapy strategies. Future studies should reveal the roles of VEGFR2 on GC cells.
In conclusion, our data show that a TME-related biomarker classi cation might be potentially useful for existing and new molecularly targeted and immune therapies in GC. Growing evidence is showing that combining anti-angiogenic therapy with immunotherapy may, in certain contexts, improve the immune response in solid cancers 14,47 . A new classi cation system could also guide more effective new combination therapies, earlier in the management of GC.

Declarations
Ethical approval The retrospective study was approved by the Ethics Committee of Fundeni Clinical Institute, Bucharest, Romania.  Association between PD-L1 expression and CD8 T cell in ltration in GC tissues. a-b Representative IHC for CD8 and PD-L1 in GC tissues. c Correlation matrix for PD-L1 and CD8 IHC scores (n=98) (two-sided Spearman's correlation test).

Figure 4
Levels of circulating biomarkers of angiogenesis and in ammation and their correlation with clinical pathological data in GC patients. a Serum sTIE2 levels were signi cantly increased with tumor progression. b Serum PIGF levels were signi cantly higher in stage IV versus stage I. c The expression of serum sTIE2 levels progressively increased with the number of metastatic lymph nodules. d-e Increased serum sTIE2 levels were signi cantly correlated with perineural invasion and lymphatic vessel invasion. f Serum PIGF levels were signi cantly increased in GC patients with lymphatic vessel invasion. g-i Higher serum IL-6, IL-8, and TNF-α levels in the group of GC patients with larger tumor size (greater than the 5cm median). j Serum IL-8 levels directly correlated with lymphatic vessel invasion. k-l High levels of serum IL-6 and IL-8 were signi cantly associated with the serum tumor marker CA19-9. Comparisons between multiple groups were performed using a one-way analysis of variance (ANOVA) with Bonferroni's post hoc test. The difference between the two groups was analyzed by two-tailed unpaired t-test with Welch's correction. Each box represents the IQR and median of the normalized serum markers levels for each group, whiskers indicate 1.5 times IQR (and any values that are greater than this are plotted as individual points). Note: expression levels of serum biomarkers were normalized by log 2 .

Figure 5
Associations between tissue and circulating biomarkers in GC patients. a Correlation between tumor VEGFR2 expression and serum sTIE2 (n=117). b-c Correlation between tumor MVD and serum levels of PIGF and sTIE2 (n=110). d Correlation between tumor CAIX expression and serum VEGF (n=59). e-f Correlation between tumor MVD and serum IL-6 and IL-8 (n=110). g-i Correlation between pericyte coverage of GC vessels and serum IL-6, IL-8, and TNF-α (n=93). j-l Correlation between tumor PD-L1 expression and circulating levels of IL-8 (j) and TNF-α (k) (n=116), and between tissue CD8 expression and serum IL-6 (n=101) (l). The difference between the two groups was analyzed by two-tailed unpaired ttest with Welch's correction. Each box represents the IQR and median of the normalized serum markers levels for each group, whiskers indicate 1.5 times IQR (and any values that are greater than this are plotted as individual points). Expression levels of serum markers were normalized by log 2 .

Figure 6
Serum biomarkers association with survival and molecular classes of GC in this cohort (n=122). a-b The prognostic value of the serum sTIE2 and VEGF-C in patients with resectable GC Kaplan-Meier curve for overall survival based on molecular classi cation. c Prognostic value of molecular classi cation. d Serum VEGF levels were signi cantly higher in Gp2 compared to Gp3 and Gp4 groups. e TNF-α levels were signi cantly elevated in Gp1 compared to Gp3 and Gp4 groups. Comparisons between multiple groups were performed using a one-way analysis of variance (ANOVA) with Bonferroni's post hoc test.