Immune microenvironment of Epstein-Barr virus (EBV)-negative compared to EBV-associated gastric cancers: implications for immunotherapy.

Introduction
Gastric cancer (GC) is an aggressive malignancy and is one of the most common causes of cancer-related mortality worldwide.1 Although GC can be cured in its early stages with locoregional procedures, the median survival of patients with metastatic disease receiving chemotherapy is less than 1 year.2 Fluoropyrimidine-containing and platinum-containing chemotherapies are used in the initial management of advanced metastatic GC. Recently, nivolumab, a monoclonal antibody blocking the immune inhibitory receptor programmed cell death 1 (PD-1), received Food and Drug Administration approval for first-line management of advanced or metastatic GC and gastroesophageal junction cancer (GEJC) in combination with chemotherapy, based on improvements in overall survival (OS) and progression-free survival compared with chemotherapy alone. Subsequently, the PD-1 inhibitor pembrolizumab was authorized for use with chemotherapy in advanced GC and GEJC if human epidermal growth factor 2-negative. Although these approvals established chemoimmunotherapy as a new standard of care for advanced GC, with response rates ~50%, only 10–20% of patients respond to anti-PD-1 monotherapies.3 4 Patients with GCs expressing programmed cell death ligand 1 (PD-L1) may be somewhat more likely to benefit, but the response rate to anti-PD-(L)1 monotherapy in patients with PD-L1 positive tumors remains less than 20%.4 Notably, in GCs that are associated with Epstein-Barr virus (EBV) infection (~5–10% of GCs),5 6 response rates to anti-PD-(L)1 monotherapy exceed 50%, nominating EBV positivity as a predictive biomarker for anti-PD-(L)1 response.7 8
EBV-associated GC (EBVaGC) is a distinct subtype of GC with characteristic clinicopathological and molecular features. As compared with sporadic GC, EBVaGC occurs more often in younger male individuals, frequently arises in the gastric cardia or corpus, and typically has diffuse histology and a lower frequency of lymph node metastasis.9 The precise mechanism of oncogenesis by latent EBV infection remains unclear, but EBV latent genes have previously been shown to disrupt various cellular processes and signaling pathways.10 11 Whole-genome sequencing and comprehensive molecular profiling of EBVaGC have revealed a strong association with CpG island methylator phenotype, CDKN2A/p16 gene silencing, PIK3CA mutations, and frequent amplification of JAK2, PD-L1 and PD-L2.1215 The frequent amplification of PD-L1 and PD-L2 in EBVaGC suggests that immune evasion may be significant in EBVaGC oncogenesis.
Virus-associated cancers may have a distinct immune phenotype from their virus-negative counterparts, providing a rationale to investigate the tumor immune microenvironment (TiME) in EBVaGC and non-EBVaGC. In virus-associated cancers, tumor-specific proteins encoded by viral open reading frames may serve as strong immune stimulants, generating antigen-specific T cells but also inducing adaptive immune resistance through immune checkpoint pathways.16 For example, human papillomavirus (HPV)-positive head and neck squamous cell carcinomas (HNSCC) have a more extensive T-cell infiltrate, higher levels of immune activation, and higher expression of the immune checkpoint cytotoxic T-lymphocyte associated protein 4 than HPV-negative HNSCC.17 18 Similarly, the presence of Merkel cell polyomavirus in Merkel cell carcinoma is associated with a robust immune infiltrate and increased tumor cell PD-L1 expression.19 20 Several studies previously reported that EBVaGC has a more extensive lymphocyte infiltrate than EBV-negative (EBV(−)) GC,2126 but trends in forkhead box P3 (FOXP3)+ cell infiltration and PD-1/PD-L1 expression are less consistent.21 Such studies have been hampered by the relative rarity of EBVaGC. Identifying unique features in the TiME of EBV(−) GC, which is much less responsive to anti-PD-(L)1 than EBVaGC, may support the development of next-generation clinical trials of treatment combinations targeting specific immune inhibitory pathways. We, therefore, undertook the current study comparing the TiMEs of EBV(−) GC versus EBVaGC, finding that EBV(−) GCs are distinguished by significant overexpression of genes associated with the cyclooxygenase 2 (COX-2) pathway, which has the potential to dampen antitumor immunity and thereby offers a potential co-target with anti-PD-(L)1-based immunotherapy.

Methods

Patients and tumor specimens
After screening over 1,000 pathology records of GCs excised at Johns Hopkins Hospital in 2000–2014, 11 EBV+ and 14 EBV(−) primary GC specimens were identified from treatment-naïve patients (online supplemental table S1). EBV status was determined by EBV-encoded RNA in situ hybridization (EBER-ISH) conducted on whole tissue sections or tissue microarrays. EBV(−) cases were selected in reverse chronological order, beginning with the most recent pathology specimens. Specimens were included in this study if there was sufficient material for immunohistochemical (IHC) and gene expression analyses. An approximately equal number of EBV+ cases were then selected within the same time period.

Immunohistochemistry
Formalin-fixed paraffin-embedded (FFPE) tissue sections were stained with H&E and with antibodies specific for markers of immune cell subsets and immunomodulatory molecules, including CD3, CD4, CD8, CD20, CD68, FOXP3, PD-1, PD-L1, colony stimulating factor 1 receptor (CSF-1R), glucocorticoid-induced tumor necrosis factor receptor family-related protein (GITR), lymphocyte activating gene 3 (LAG-3), indoleamine 2,3-dioxygenase 1 (IDO-1), and COX-2. IHC for CD3, CD4, CD8, CD20, FOXP3, CSF-1R and COX-2 was performed according to standard automated methods. IHC for CD68, PD-1, PD-L1, LAG-3, IDO-127 and GITR28 was performed manually, as previously described (online supplemental table S2). For digital analysis, expression of CD8, CD20, CD68, FOXP3, PD-1, CSF-1R, and GITR was quantified as cell density (number of positive cells per mm2) using HALO software (Indica Labs, Albuquerque, New Mexico, USA) as described.18 Manual scoring for CD3, CD4, PD-L1, LAG-3, IDO-1, and COX-2 expression was conducted by a pathologist (RAA) using a compound light microscope. PD-L1 expression was determined on tumor cells (membranous cell surface staining pattern, among all tumor cells) or stromal cells (cell surface and cytoplasmic staining in non-tumor, non-lymphocyte cells, among the total number of nucleated stromal cells) and reported at 5% intervals. For COX-2, the percentage of positive tumor cells was recorded. IDO-1 staining was reported as positive tumor and normal gastric epithelial cells, with attention to prominent expression at the tumor:stromal interface. CD3 was scored semi-quantitatively as 0 (none), 1 (focal), 2 (moderate) or 3 (marked) infiltration of positively stained cells. The presence of CD4+ cells was scored as the ratio of CD4+ to CD8+ cells, reported as 0.5 (1:2), 1 (1:1) or 2 (2:1). Expression of LAG-3 was estimated as the percentage of positive cells among CD3+ tumor-infiltrating lymphocytes.

Multiplex real-time quantitative reverse transcriptase PCR
Macrodissection (manual scraping of annotated slide areas with a scalpel) was conducted on 5 µm-thick FFPE tissue sections, capturing areas of CD3+ infiltrates in the TiME and, if present, neighboring PD-L1+ stromal and/or tumor cells. Following total RNA isolation, the expression of 122 unique immune-related genes and three reference genes (18S, GUSB, PTPRC) was evaluated with quantitative reverse transcriptase PCR (qRT-PCR) using Taqman low-density array cards (Life Technologies, Carlsbad, California, USA) as previously reported.28 29 Gene expression by EBV+ versus EBV(−) tumors was normalized to either PTPRC (CD45, pan-leukocyte marker) or 18S ribosomal RNA using the 2−ΔΔCt method.30 Undetermined cycle thresholds were assigned a value of 40 for computational purposes. Volcano plots were created with GraphPad software (La Jolla, California, USA). Heat maps displaying unsupervised clustering of genes and specimens were performed by Cluster 3.031 32 using city-block distance with average linkage. Heat maps were generated using Java TreeView V.1.1.6r4.3234 The Cluster parameters EORDER (for samples) and GORDER (for genes) were used to govern the (otherwise arbitrary) respective top/bottom, left/right order at each pair of subcluster joins in the dendrograms of the heat maps.

TCGA analysis
Data from The Cancer Genome Atlas (TCGA) were downloaded from https://www.cancer.gov/tcga. Trimmed mean of M-values-normalized log2TPM (transcripts per million) tumor RNA sequencing data were used in the analysis. Clinical data pertaining to TCGA GC specimens were downloaded from cBioPortal (http://www.cbioportal.org). Among 415 GC samples in TCGA, 30 were EBV+ and 385 were EBV(−).15 Expression of PTGS2 (COX-2) as a single gene, as well as a 13-gene inflammation signature, were compared between EBV+ and EBV(−) GCs. The inflammation signature comprised CCL2, CCL3, CCL4, CD8A, CXCL9, CXCL10, GZMK, HLA-DMA, HLA-DMB, HLA-DOA, HLA-DOB, ICOS, and IRF1.35 The signature was calculated as the average log2TPM of these 13 genes. All analyses were performed using R statistics software and validated. For overall survival analysis, 407 patients with GC with available survival data were divided into four groups by expression level of PTGS2. Kaplan-Meier analysis of the lowest and highest quartiles of patients was conducted using the R survminer package (https://cran.r-project.org/web/packages/survminer/index.html), and p values were generated from the log rank test with default settings in the package. PTGS2 expression was also analyzed across 25 cancer types in TCGA having at least 100 samples each, including 9,712 primary, 394 metastatic and 42 recurrent tumors.

Statistical analysis
For analysis of qRT-PCR and IHC data, p values were obtained using the Wilcoxon rank-sum test, via the wilcox_test function from the R “coin” package (V.1.2–2)36 with distribution=“exact”. All p values are two-sided except where noted. When running the Wilcoxon rank-sum test on qRT-PCR data, all delta Ct values coming from Ct values of 40 were set to the same (tie) value, chosen larger than any other delta Ct (so that all delta Ct values coming from undetermined Ct values have the same highest rank).

Results

Characteristics of patients and tumor specimens
25 primary treatment-naïve GC specimens with sufficient tissue for protein and gene expression studies were identified by screening >1,000 pathology records from resections conducted at the Johns Hopkins Hospital from 2000 to 2014 (online supplemental table S1). Among them, 11 GCs were determined to be EBV+ and 14 EBV(−) by EBER-ISH. Patients had American Joint Committee on Cancer (AJCC) stage 1A – 4 disease at the time of resection. Patients with EBV+ GC were significantly younger than those with EBV(−) GC, as anticipated (median age 68 vs 77 years, respectively, p=0.0284). However, there were no significant differences in gender (data not shown), ethnicity, or AJCC stage.

Expression of candidate immune-related markers by IHC
To assess and compare the characteristics of the TiME in EBV+ versus EBV(−) GCs, we first conducted IHC for select markers of immune cell subsets and immunoregulatory receptors and ligands (figure 1). We found a significantly higher density of CD8+ cells infiltrating EBV+ versus EBV(−) GCs (845 cells/mm2 vs 375 cells/mm2 respectively, p=0.044). Consistent with this, we found lower CD4:CD8 ratios in EBV+ tumors using semi-quantitative IHC scoring (p=0.051; figure 2A). Although densities of tumor-infiltrating CD20+ B cells, FOXP3+ regulatory T (Treg) cells, and CD68+ macrophages were numerically higher in EBV+ versus EBV(−) GCs, these differences did not achieve statistical significance (figure 1B). Likewise, densities of cells expressing the immunomodulatory molecules PD-1, GITR, CSF-1R, and IDO-1 were not significantly different between the two groups (figure 1B), nor was the percentage of CD3+ cells expressing LAG-3 (figure 2B). The expression pattern of the immunosuppressive enzyme IDO-1 was notable in that it was observed intracellularly in macrophages, tumor cells, intratumoral endothelial cells, and/or normal gastric epithelium (when present) overlying tumor areas, regardless of EBV status (online supplemental figure S1). Heightened IDO-1 expression was observed at the tumor:stromal interface in some specimens. Consistent with previous findings in other gastrointestinal cancers, the immune checkpoint ligand PD-L1 was expressed mainly by infiltrating immune cells and not by tumor cells themselves in GC specimens37 38 (figure 2B). There was no significant difference in the proportion of stromal cells expressing PD-L1 between EBV+ and EBV(−) GCs (mean 20% vs 12%, p=0.529). Thus, both EBV+ and EBV(−) GCs were characterized by immunosuppressive features upon IHC analysis of candidate markers, although EBV+ tumors contained a significantly higher density of CD8+ T cell infiltrates, implying the potential for antitumor immune effector functions.

Expression of immune-related genes in the TiME of EBV+ versus EBV(−) GCs
We further characterized and compared the TiMEs of EBV+ versus EBV(−) GCs by performing gene expression profiling (GEP), using multiplex qRT-PCR to detect 122 unique genes of interest and two reference genes (18S, and the pan-immune cell marker PTPRC/CD45). With unsupervised hierarchical heat map clustering, specimens from each GC subtype tended to cluster together, suggesting distinct features between these transcriptomes (figure 3 and online supplemental figure S2). For some genes, robust expression was detected across all specimens regardless of EBV status; this included genes related to immune activation and antigen presentation (B2M, HLA-DRA, STAT1, UCP2) as well as others involved in immunosuppression (CD276 (B7-H3), TGFB1, VEGFA). Conversely, certain genes were poorly expressed across all samples, including those involved in immune activation (GZMB, IFNG, IL2, IL12B) or associated with a Th17 profile (IL17A, IL22, IL22RA2, IL23A, IL23R). Of particular interest were the genes that were significantly differentially expressed (figure 4 and online supplemental figure S3). When normalized to PTPRC, three genes were overexpressed in EBV+ compared with EBV(−) GCs: ITGAE (CD103, marking intraepithelial T cells and a dendritic cell subset; up threefold, p=0.036), and two interferon (IFN)-gamma inducible genes—CXCL9 (chemokine involved in cytotoxic T lymphocyte recruitment; up fourfold, p=0.005) and IDO1 (immunosuppressive enzyme; up sixfold, p=0.030). In contrast, EBV(−) tumors overexpressed several functionally-related gene groups, including those characterizing myeloid cells (CD163, IL1A, NOS2, RIGI); immunosuppressive cytokines/chemokines such as CXCL2, CXCR4, IL10 and IL32; and coinhibitory molecules such as HAVCR2 (TIM-3), VSIR (VISTA), and components of the adenosine pathway (ENTPD1 (CD39) and NT5E (CD73)). Notably, compared with EBV+ GCs, EBV(−) GCs highly overexpressed PTGS2 (the most highly upregulated single gene, 32-fold, p=0.005) encoding the COX-2 enzyme which mediates prostaglandin E2 (PGE2) synthesis and tumor-promoting inflammation; the prostaglandin receptors PTGER1 (EP1; up 21-fold, p=0.015) and PTGER4 (EP4; up twofold, p=0.022), and the major COX-2-inducing cytokine IL1B (up 11-fold, p=0.019). PTGS2, PTGER1, and IL1B were also significantly overexpressed in EBV(−) GCs when gene expression was normalized to the 18S ribosomal subunit (online supplemental figure S3).

Expression of COX-2 protein in GC
We next wanted to determine if heightened PTGS2 (COX-2) gene expression in EBV(−) versus EBV+ GCs translated into differential protein expression. With IHC for COX-2, there was a trend towards increased expression in EBV(−) compared with EBV+ tumors, with means of 56% versus 25% of tumor cells expressing COX-2, respectively, (p=0.068, one-sided) (figure 5A). Of interest, we also observed uniform COX-2 expression in the normal gastric epithelium overlying GC specimens, where PGE2 plays a normal homeostatic role. COX-2 was expressed by multiple cell types in GC, including tumor cells, myeloid cells, and non-myeloid stromal cells (figure 5B).

COX-2 analysis in TCGA data
To further explore the implications of COX-2 expression in GC, we conducted an analysis of published TCGA data from 415 GCs, including 30 EBV+ and 385 EBV(−) specimens. Examining PTGS2 gene expression, there was highly significant upregulation in EBV(−) compared with EBV+ GCs, consistent with findings from our current study (figure 6A). Of interest, EBV(−) GCs also showed significantly lower expression of a 13-gene inflammation score,35 consistent with the known immunosuppressive effects of the COX-2/PGE2 pathway (figure 6B). Furthermore, in an analysis of TCGA data across 25 different cancer types having transcriptional data from at least 100 specimens, GC was among the highest PTGS2-expressing tumor types (online supplemental figure S4). Analyzing OS among 407 patients with GC as a function of the level of tumor PTGS2 expression, we found significantly lower OS in the highest PTGS2-expressing quartile compared with the lowest expressing quartile (p=0.032; figure 6C). Thus, increased COX-2 gene expression was associated with EBV(−) tumor status and reduced OS in patients with GC.

Discussion
This comparative study of the TiME of EBV-associated versus non-EBV-associated GC is immediately relevant to the striking clinical observation that responsiveness to PD-1 pathway blockade is substantially enhanced in the EBV+ GC subtype. While the TiMEs of both GC subsets harbor a variety of immunosuppressive features, our findings of heightened inflammation in EBV+ compared with EBV(−) GC, with higher densities of infiltrating CD8+ T cells and upregulation of IFN-gamma-inducible genes such as IDO and CXCL9, are consistent with previously published observations.263941 EBV(−) GCs are heterogeneous, comprising chromosome instable, genome instable, and microsatellite instability-high (MSI-H) molecular subtypes identified by TCGA.15 Among these EBV(−) subtypes, only MSI-H GC responds favorably to PD-1 blockade, highlighting the need for improved therapeutic options for the majority of GCs.
Our GEP analysis revealed that, compared with EBV+ GCs, EBV(−) GCs differentially overexpressed certain genes associated with immunosuppression that could potentially inhibit local antitumor immunity in the TiME. Notably, PTGS2, encoding the COX-2 enzyme that catalyzes the production of the lipid-derived prostaglandin PGE2, was the most highly upregulated gene in our EBV(−) GC cohort. PGE2 mediates protumorigenic signals through the G-protein coupled EP receptors EP2 and EP4 (encoded by PTGER2 and PTGER4); importantly, PTGER4 was upregulated in EBV(−) tumors alongside PTGS2 in this study. Of interest, transient transfection of the EBV latent membrane proteins LMP1 and LMP2A into a COX-2-expressing EBV(−) cultured GC line was reported to significantly downregulate COX-2 expression, accompanied by downregulation of the anti-apoptotic signaling molecule TRAF2; these findings highlight a biological interaction between EBV infection and COX-2 expression that may potentially render EBV+ GC more susceptible than EBV(−) GC to immune attack.42
Additional upregulated genes that we identified in EBV(−) GCs, CXCL2 and IL1B, have been associated with high PTGS2 expression in human melanomas.43 We previously showed that the COX-2/PGE2 pathway is upregulated in human cancer cell lines and myeloid cells following in vitro exposure to interleukin-1 beta (IL-1B) and certain other tumor microenvironment (TME)-resident cytokines, and that PGE2 can functionally suppress several immune cell subsets including T and myeloid cells.29 44 45 Results from our current analysis of PTGS2 gene expression as well as a multigene inflammation score in TCGA data including 415 GC specimens are consistent with findings from our in-depth analysis of 25 GCs revealing increased PTGS2 expression and decreased inflammation in EBV(−) versus EBV+ GC.
Our laboratory’s previous studies of the TiMEs of two different EBV-associated cancers, classical Hodgkin’s lymphoma (CHL)28 and nasopharyngeal carcinoma (NPC),46 demonstrated that the presence of tumor-associated EBV is not in itself sufficient to generate a vigorous tumor-reactive immune milieu. While CHL is highly responsive to anti-PD-1 monotherapy,47 the NPC response rate to anti-PD-1 monotherapy is only ~20%.48 Comparing EBV+ versus EBV(−) CHLs, we found a Th1-associated IFN-driven pro-inflammatory TiME in EBV+ CHL versus a pathogenic Th17-driven TiME in EBV(−) CHL. Strikingly, NPCs, of which >95% are EBV+, were characterized by a more immunosuppressive TiME when compared with EBV+ CHL. In particular, COX-2 was highly upregulated in NPC compared with EBV+ CHL. Thus, cancers associated with the same virus may have distinct immune biologies based on tissue context,49 suggesting the need to customize anti-PD-1-based treatment strategies according to disease application.
In summary, the COX-2/PGE2 pathway may interact through multiple mechanisms to down-modulate antitumor immunity in the TiME and mitigate the therapeutic effects of immunotherapies that are currently in standard use for treating GC. Whether COX-2/PGE2 pathway inhibition as a treatment for established cancers will mediate similar beneficial effects to those observed in preventing the occurrence of new cancers is an open question.50 51 Combining PD-1 pathway blockade with inhibition of the COX-2/PGE2 pathway, by inhibiting COX-2-inducing cytokines such as IL-1B, or by blocking the PGE2 receptors EP2 and/or EP4, are strategies that are now in clinical testing across a wide array of cancer types and may offer new opportunities to improve therapeutic outcomes for patients with EBV(−) GCs.

supplementary material
10.1136/jitc-2024-010201online supplemental file 110.1136/jitc-2024-010201online supplemental file 2