Introduction

Epstein–Barr virus (EBV)-associated gastric cancer (EBVaGC), characterized by the presence of EBV in gastric cancer (GC) cells, is a distinct GC subtype found in ~ 10% of cases and with specific clinicopathologic and molecular features [1]. Although advances have been made in elucidating distinct molecular features of EBVaGC, no therapeutic modalities specific to EBVaGCs currently exist. Despite robust immune cell presence in EBVaGCs, few studies have explored immune subtyping of EBVaGCs and its association with survival [2,3,4]. Previous EBVaGC-specific studies have focused on histomorphological characteristics, such as tumor-infiltrating lymphocyte (TIL) patterns [4] or expression of PD-L1 through immunohistochemistry (IHC) [5, 6]. Large genomic studies of gastric cancer conducted by The Cancer Genome Atlas group (TCGA) [1] and the Asian Cancer Research Group (ACRG) [7] only include a small sample size of EBVaGC. There is a strong rationale for additional characterization of the immune landscape of EBVaGC as these may highlight new prognostic markers and avenues for therapy. Here, we report results from one of the largest cohorts of EBVaGC primary samples with integrated analysis of transcriptomic expression of key immune-related genes (IRG), TIL subtyping and PD-L1 IHC expression using a recently Food and Drug Administration (FDA) approved antibody and scoring algorithm for gastric cancer.

Methods

EBVaGC cohort

EBV-encoded RNA in-situ hybridization (EBER-ISH) was performed on gastric cancer samples from patients undergoing primary tumor resections at Samsung Medical Centre, South Korea, from 1996 to 2016 (supplementary methods). Only cases with a strong signal within almost all tumor cell nuclei were considered to be positive and included in this cohort. Analysis of tumor-infiltrating lymphocytes was performed and classified according to previously established histologic subtypes based on host inflammatory immune response [lymphoepithelioma-like carcinoma (LELC), Crohn’s disease-like lymphocytic reaction (CLR) and conventional adenocarcinoma (CA)] [4]. PD-L1 immunohistochemistry (IHC) was performed using the FDA approved PharmDx 22C3 Dako PD-L1 antibody and scored using the Combined Positive Score (CPS) (supplementary methods).

EBV-negative ACRG cohort

We have previously reported molecular characterization of a large cohort of gastric cancer samples as part of the Asian Cancer Research Group (ACRG) [7]. These samples were also from the same Samsung Medical Centre, South Korea, collected from 2004 to 2007. From this cohort, we selected samples of EBV-negative gastric cancer, with clinical data available as a comparative control for this study (EBVnegACRG cohort).

NanoString® analysis

To measure the transcriptomic expression of IRGs in EBVaGC samples, we used the NanoString platform, an FDA approved platform that has been shown to work with low input formalin-fixed paraffin-embedded RNA samples. Nanostring nCounter Reporter CodeSets were designed for 14 IRG corresponding to intra-tumoral cytolytic activity (CYT) [8], cytokines and immune checkpoints (supplementary methods, Supplementary Table 1). Analysis of older samples (> 10 years old) versus newer samples was performed to reduce sample-degradation bias (supplementary methods, Supplementary Fig. 1).

Statistical analysis

Associations of clinicopathologic features to histologic subclassification was performed using Fisher’s exact test. Disease-free survival (DFS) was calculated from the time of surgery to the time of disease progression or death, and overall survival (OS) was calculated from time of surgery to time of death. Kaplan–Meier (KM) curves and log rank test were used for survival analysis. The hazard ratio (HR) and its 95% confidence interval (CI) were evaluated for each analysis using Cox proportional hazards regression. All analyses were done using R (3.4.1). Samples with PD-L1 expression greater than the 34th percentile were defined as PD-L1high and remaining PD-L1low.

Results

Patients’ characteristics

We identified 71 cases of EBVaGC from a cohort study between 1996 and 2016 [4]. As a comparative control, 193 EBV-negative GC samples (of which 149 had clinical data available) were selected from the ACRG cohort [7]. Patients’ characteristics of these cohorts are described in Table 1. For the EBVaGC cohort, the median age was 56 years, with 87% being male. In general, clinicopathological features predicting poor outcome were only found in a minority of EBVaGC samples: lymphovascular invasion (42%), perineural invasion (28%) and nodal metastases (37%), consistent with the good prognosis of this GC subgroup.

Table 1 Patients’ characteristics
Full size table

Immune gene expression profiles of EBVaGC

Unsupervised clustering of the expression data revealed clear separation between EBVaGC and EBVnegACRG GCs (Fig. 1a). When analyzed as a group, EBVaGC samples collectively showed higher expression of all queried IRGs (Fig. 1b, c) compared to EBVnegACRG samples (p < 0.01, Wilcoxon one-sided test), consistent with previous reports of increased immune infiltration in EBVaGC [1, 2]. EBVaGC samples showed significantly increased expression of PD-L1 (p < 0.001, Fig. 1d) compared to EBVnegACRG samples. Notably however, not all EBVaGC samples showed higher PD-L1 expression compared to EBVnegACRG samples. We thus divided the EBVaGC samples into PD-L1high (n = 47, 66%) and PD-L1low (n = 24, 34%) expressing groups. We observed that PD-L1low samples had consistently decreased expression of all other IRG including markers of CYT, such as CD8A, GZMA and PRF1 as well as immune checkpoints, such as PD-1 (p < 0.001, Wilcoxon one-sided test) (Fig. 2a, b). When compared to the EBVnegACRG cohort, a similar trend was seen, with reduced IRG expression identified in those with reduced PD-L1 expression (Fig. 2a).

Fig. 1
figure 1

Immune gene expression profiles of EBVaGC. a Principal component analysis of ACRG (red) and EBVaGC (blue) samples showing separation of EBV-positive samples from EBV-negative ones. b Unsupervised heatmap of immune genes for EBVaGC and ACRG samples. EBVaGC samples mostly cluster together and have increased immune gene expression compared to ACRG EBV-negative samples. c Boxplot of immune genes for EBVaGC and ACRG samples showing significantly increased expression in EBVaGC samples (p value < 0.01, Wilcoxon one-sided test). d Boxplot of PD-L1 expression in EBVaGC and ACRG samples (p < 0.001, Wilcoxon one-sided test)

Full size image
Fig. 2
figure 2

Prognostic value of PD-L1 expression in EBVaGCs. a Heatmap of immune gene expression in EBVaGC and ACRG samples sorted by PD-L1 expression. b Boxplot of immune genes grouped by PD-L1 expression [p value < 0.001 for all genes except IL-10 (p = 0.02), Wilcoxon one-sided test]. Kaplan–Meier (KM) survival curves for disease-free survival of patients grouped by PD-L1 expression in c EBVaGC samples, d ACRG samples

Full size image

PD-L1 immunohistochemical expression and tumor-infiltrating lymphocyte classification

PD-L1 IHC was performed on all 71 samples. Tumor and immune PD-L1 expression was scored independently and the CPS was calculated (Table 2). PD-L1 CPS scores correlated with PD-L1 transcriptomic expression (r = 0.63, p < 0.001) (Supplementary Fig. 2). PD-L1low tumors tended to have lower CPS scores compared to PD-L1high tumors (Table 2; p = 0.004). Analysis of tumor-infiltrating lymphocytes was also performed and classified according to host inflammatory immune response: LELC, CLR, CA [4]. We observed significant differences in the expression of PD-L1 within these subtypes representative of tumor-infiltrating lymphocyte status (Table 2). PD-L1low had a higher proportion of CA subtype (25 vs 4%), while PD-L1high had a higher proportion of LELC subtype (51 vs 25%). There were no significant differences in the other clinicopathological parameters, such as T stage, N stage, lymphatic and perivascular invasion between the two groups (Table 2).

Table 2 Differences in characteristics between PD-L1high and PD-L1low groups
Full size table

PD-L1low is associated with worse survival

The EBVaGC cohort had a median follow-up of 66 months. Of the five deaths in the cohort, one of them was not cancer related. With limited OS events precluding meaningful statistical analyses, DFS was chosen as the survival outcome for further statistical and correlative analyses. PD-L1low samples were associated with worse DFS [HR 5.03 (0.97–25.92), p = 0.032, Fig. 2c]. We saw similar trends for the EBVnegACRG cohort as well, where PD-L1low samples showed worse DFS [HR 1.44 (95% CI 0.92–2.25), p = 0.11, Fig. 2d].

Discussion

Here, we analyzed a cohort of EBV-positive gastric cancers, with an integrated analysis of PD-L1 IHC, TIL subtyping and transcriptomic expression of PD-L1 and other key IRG. Through this analysis we made the following findings: While PD-L1 expression in EBVaGC is higher than non-EBVaGC, there-in lies a spectrum, with a proportion of EBVaGC having lower transcriptomic expression of PD-L1. This group of PD-L1 low expressors also have reduced expression of other IRG including immune checkpoints. This same group of PD-L1 low expressors have a poorer prognosis compared to PD-L1 high expressing EBVaGC. PD-L1 transcriptomic expression correlates with PD-L1 protein expression as measured by immunohistochemistry. PD-L1 low expressors also tend to have a different pattern of tumor-infiltrating lymphocytes.

Immune checkpoint inhibitors have recently received approval for the management of advanced gastric cancer [9, 10]. PD-L1 has emerged as one of the leading biomarkers for selecting patients who might benefit from anti-PD-(L)1 therapy [11]. However, several discrepant results from large clinical trials using immunohistochemical expression of PD-L1 have brought into question its utility and reliability as a predictive biomarker for immune checkpoint inhibition [10, 12]. Our study has analyzed PD-L1 at the transcriptomic level and correlated this with protein expression levels (IHC). Furthermore, previous studies have used various antibodies and algorithms for assessing PD-L1 IHC [5, 6, 13]. Discordance between various PD-L1 antibodies has been described [14]. Thus, it is of major clinical and scientific relevance that our study used the FDA approved PD-L1 antibody and algorithm to study PD-L1 IHC scores in our EBVaGC cohort. Notably, in both EBV-positive and EBV-negative patients, we found higher transcriptomic expression of PD-L1 to predict for better prognosis. Two previous studies have reported a tendency towards a poorer prognosis in EBVaGC with tumoral expression of PD-L1 with anti-PD-L1 antibody using clone E1L3N (Cell Signaling Technology, Danvers, MA) [5, 6]. However, in our previous studies on PD-L1 IHC in microsatellite instability-high and EBVaGC, distinct molecular subtypes of gastric cancer with high immune cell infiltration, PD-L1 expression was a good prognosticator using clone SP142 (Ventana, Tucson, AZ) [15, 16]. From these data, it is unlikely that PD-L1 either measured by IHC or other methods, such as transcriptomic analyses would be a reliable stand-alone predictive or prognostic biomarker for gastric cancer. In view of the conflicting data of using a single gene/protein (PD-L1) as an immune-biomarker, we chose to analyze a group of important IRGs instead. Of significant note, we found that those tumors with low PD-L1 expression tended to have a similarly reduced expression of all other IRGs. In another study of 12 EBVaGC samples, expression of major histocompatibility complex class II genes and chemokine activity regulating genes were found to be deregulated more frequently [2]. These findings support the hypothesis that the better prognosis of EBVaGC is likely immune-related. In the era of immunotherapy, this is of particular significance in both EBVaGC and EBV-negative tumors, and GCs that display low levels of IRG may have to possibly be treated differently to those with high IRG expression.

In conclusion, we studied transcriptomic expression of IRG in a large cohort of EBVaGC. We have demonstrated that a substantial proportion of tumors have a low transcriptomic expression of PD-L1 and tend to show a poorer survival outcome through low expression of other IRG. While large GCs landscaping groups and studies, such as the TCGA and ACRG have focused largely on tumoral genomic data, future studies must incorporate analyses of the interplay between the tumor and the immune system in order to better understand and treat this disease.