Identification of a molecular signature of prognostic subtypes in diffuse-type gastric cancer

We generated a transcriptome dataset of 150 fresh-frozen tissues including diffuse-type GC (n = 107), intestinal-type GC (n = 23), and normal gastric tissues (n = 20) obtained from the BioBank of the Chungnam and Seoul National University Hospitals (the original cohort, n = 150). We also obtained the mRNA expression or variant data of diffuse-type GC from the TCGA database (n = 61, the TCGA cohort) [7], the ACRG study (n = 135, the ACRG cohort) [8], and the Samsung Medical Center in South Korea (n = 280, the SMC cohort) [11]. Table S1 details the baseline characteristics of GC patients.

Details of the methodology are available in Supplementary Materials.

We performed gene expression profiling of 150 GC samples in the original cohort. Through unsupervised hierarchical clustering with a total of 3586 genes showing varying expression changes across GC samples [standard deviation (SD) > 0.9], three major sample clusters were observed: the normal-like (N), intestinal-type-like (INT), and core diffuse-type (COD) GC subgroups (Fig. 1). Most normal tissue samples were included in the N cluster (19 out of 20), while the vast majority of intestinal-type GC samples were included in the INT cluster (19 out of 23). Among 59 tissue samples in the COD cluster, 54 samples were diffuse-type GCs. When the histological data were compared among the sample clusters, not surprisingly, patients with poorly cohesive carcinoma (PCC), a poor prognostic histological subtype of GC, were found to be significantly plentiful in the COD cluster (χ² test, P = 0.002, Fig. 1).

Through exploring the expression patterns, we intuitively identified two distinct subsets of genes that were highly associated with the INT or COD cluster of GC (252 and 397 genes, respectively, surrounded by purple dashed lines in Fig. 1). These genes were more highly expressed in the INT or COD cluster than in the other clusters. Among the 252 genes associated with the INT cluster, many genes involved in the cell cycle or DNA repair were observed. Among the 397 genes associated with the COD cluster, on the other hand, genes involved in EMT-associated functions such as cell adhesion/migration or the TGFβ signaling pathway were significantly plentiful (Fig. 1). These results indicate that the newly discovered COD cluster may be compatible with the known poor prognosis of diffuse-type GC, which may derive from EMT activity.

To identify optimal gene sets for distinguishing patient subgroups (i.e., the N, INT, and COD clusters) in diffuse-type GC, genes that were differentially expressed between the three clusters were next identified using only diffuse-type GC samples (n = 107) from the original cohort. We selected two lists of genes that were significantly differentially expressed between the N and INT clusters or the INT and COD clusters (P < 0.001), which were compared by using a Venn diagram approach (Fig. S1a). Gene list “A” represents genes that were differentially expressed between the N and INT clusters (737 genes), and gene list “B” represents genes that were differentially expressed between the INT and COD clusters (2069 genes). When the two gene lists were compared, three different patterns were observed: only A (590 genes), A and B (147 genes), and only B (1922 genes; Fig. S1b). Diffuse-type GC samples in the N cluster showed gene expression patterns of normal stomach, which might be due to involving many normal gastric cells, even though all diffuse-type GC tissues used in the current investigation were pathologically confirmed as containing high tumor cell contents. Therefore, genes in the only-A category exhibited distinct expression patterns associated with tumorigenesis of GC in the INT subgroup. On the other hand, genes in the only-B category exhibited expression patterns associated with the progression into the COD subgroup in diffuse-type GC. Genes in both the A and B categories were common to the three subgroups of diffuse-type GC. Among the genes involved in the only-B category, many genes associated with EMT activity, such as ERG, FGF1, FGF2, FGFR1, SFRP1, SFRP2, SOX10, SOX8, TGFB1I1, and TGFB3, were highly expressed in the COD subgroup (Fig. S1b), consistent with our previous observations shown in Fig. 1.

We next tried to identify a molecular signature under the influence of 1922 genes in the only-B category and use the signature (also referred to as the COD signature) to classify risk subgroups of diffuse-type GC. On the basis of hierarchical clustering analysis of the expression patterns of these genes in the ACRG cohort (n = 135), we obtained two distinct sample clusters (i.e., INT and COD subtypes) based on the expression of genes associated with EMT (Fig. S2a). Regarding the estimated disease-free survival (DFS) of these two patient subgroups, the recurrence rate in the COD subtype was significantly higher than in the INT subtype (log-rank test, P = 0.007; Fig. 2a). The overall survival (OS) rate of the COD patients was also significantly lower than that of the INT patients (log-rank test, P = 0.003; Fig. 3b), demonstrating prognostic relevance of the COD signature in diffuse-type GC. To validate the prognostic value of the signature, we also used gene expression data for diffuse-type GC from the SMC cohort (n = 267). Using the same procedure employed for the ACRG cohort, the patients in the SMC cohort were divided into two subtypes (INT and COD) by hierarchical cluster analysis (Fig. S2b), and the DFS and OS of each group were estimated. Kaplan–Meier analyses revealed that the signature was a significant predictor of diffuse-type GC patient survival in the SMC cohort (log-rank tests, P = 0.023 for DFS and P = 0.018 for OS; Fig. 2c, d.

To determine the prognostic independence of the signature, we combined the clinical data from two patient cohorts (ACRG and SMC) and applied Cox regression analyses to the signature and known clinicopathological risk factors. In the univariate analysis, the significant prognostic indicators of OS in diffuse-type GC included age and AJCC stage, along with the COD signature (Table 1). When the multivariate test was performed on the combined cohort, the COD signature retained its statistical significance for the OS of diffuse-type GC patients even after applying a variable selection procedure (HR 2.508, 95% CI 1.53–2.766, P = 1.76 × 10^–6; Table 1), illustrating the high prognostic relevance of the signature as an independent risk factor for diffuse-type GC.

Adjuvant chemotherapy data were available for the patients from the ACRG cohort. Because adjuvant chemotherapy is the standard treatment option for GC, we investigated whether the signature could predict diffuse-type GC patients who would benefit from adjuvant chemotherapy. This analysis was performed for patients with diffuse-type GC without distant metastasis (n = 115). When the patients were divided into the INT and COD subtypes based on the signature and the difference in OS was independently assessed, adjuvant chemotherapy was found to improve the survival rate in patients with the COD subtype (P = 0.003, Fig. 2e), while patients with the INT subtype showed only a moderate benefit from adjuvant chemotherapy (P = 0.13, Fig. 2f). When the Cox regression model was applied, the interaction of the signature with adjuvant chemotherapy reached a significance level of 0.326 (Fig. 2g). However, consistent with the Kaplan–Meier and log-rank tests, the estimated HR for adjuvant chemotherapy in the COD subtype was 0.333 (95% CI 0.155–0.713; P = 0.004), retaining significant predictive value, while the HR for OS for adjuvant chemotherapy in the INT subtype was 0.576 (95% CI 0.28–1.187; P = 0.135). Taken together, the results showed that the newly identified signature exhibited significant prognostic potential as well as predictive value for chemotherapy in diffuse-type GC patients.

Functional enrichment analysis illustrated that genes involved in cellular movement, interaction, proliferation, or cell morphology in the category of molecular and cellular functions were significantly activated in the COD subtype (Fig. S3). Regulator effects analysis also showed that players involved in the EMT signature were key genetic mediators dichotomizing diffuse-type GC subgroups (Table S2; Fig. S4). Details of biological insights into the COD signature are available in Supplementary Materials.

While the COD signature showed significant predictive value for standard chemotherapy in diffuse-type GC (Fig. 2f–h), patients with the INT subtype responded moderately to this type of therapy, implying a need for additional or alternative therapeutic options for the INT subtype. To identify clues toward alternative therapies for the INT subgroup of patients, we explored mutational variants in diffuse-type GC in the TCGA cohort. A hierarchical cluster analysis based on the signature revealed stratification of patients into the INT and COD subtypes (Fig. S5). Estimation of the tumor mutation burden (TMB) in the INT and COD subtypes revealed that TMBs in the INT subtype were significantly higher than those in the COD subtype (Two-sample t test; P = 1.6 × 10^–4; Fig. 3a). Comparison of the mutation frequencies of all known genes between the INT and COD subtypes revealed that a total of 470 genes showed significantly different mutation frequencies (Fisher’s exact tests, each P < 0.05; Table S3). We searched for enriched functions using significant mutational variants (Fig. S6) and observed that genes involved in cell/focal adhesion, ECM-receptor interaction, or chromatin remodeling/modification were significantly plentiful (Fig. 3b), consistent with our previous results (Fig. 1; Fig. S3). Among the genes involved in cell adhesion, MUC16, associated with hypermutation and favorable prognosis in GC [12], and with resistance to chemotherapy in lung cancer [13], was the top discriminator of the INT and COD subtypes. The total mutation rate of MUC16 in the diffuse-type GC patients was 31.1%, and the mutation frequency of MUC16 in the COD subtype was significantly higher than that in the INT subtype (Fisher exact test; P = 0.002), suggesting that the MUC16 variant is a good indicator discriminating diffuse-type GC of the INT subtype from that of the COD subtype. Comparison of CDH1 mutations, which are well-known variants in diffuse-type GC, between the INT and COD subtypes revealed that the mutation frequency in the COD group was higher than that in the INT group; however, the difference was not statistically significant. We also observed significantly more mutations of PIK3CA [frequently found in microsatellite unstable (MSI) GCs] [14] and ARID1A (associated with MSI along with hypermutations and PD-L1 expression in cancers including GC) [15‐17] in the INT subtype compared to the COD subtype. These results suggest distinct features of the INT subtype, including high TMB and MSI, which are typical indicators predicting the response to ICI treatment [5, 18, 19].

We also compared known molecular subtypes [7] with somatic alterations illustrated by the signature (Fig. 3c). When considering molecular subtypes, all MSI (n = 7) and the majority of EBV (4 out of 6) patients exhibited the INT subtype, whereas all GS (n = 12) patients were classified as the COD subtype. The difference in molecular subtypes between the two subgroups of the signature was statistically significant (χ² test; P = 1.919 × 10^–4; Fig. 3c). These results also support a distinct feature of the INT subtype regarding the response to ICI treatment according to significant enrichment of MSI or EBV along with PIK3CA and ARID1A mutations [5, 14‐17].

We further sought to identify the predictive value of the COD signature for ICI treatment (Fig. 3d). When the expression levels of immune checkpoint genes were compared, CD274 (PD-L1) showed a significant difference in expression between the INT and COD subtypes, supporting a reported close relationship between ARID1A and PD-L1 expression in GC [15‐17]. Since activation of DNA damage response and repair (DDR) genes is significantly associated with the response to ICI [20], we also estimated the expression levels of DDR genes, revealing that the vast majority of DDR genes were significantly activated in the INT subtype. When the expression levels of members of the TGFβ pathway and its associated factors in EMT were estimated, it was found that TGFβ pathway genes were significantly downregulated in the INT subtype, and EMT genes were differentially expressed between the two subtypes, consistent with the previous report that TGFβ attenuates the response to ICI [21]. Because a dataset of ICI responsiveness in GC is publicly available, we sought to validate the predictive value of the COD signature for ICI treatment [22]. After applying the signature to the transcriptome data from GC patients and dividing them into the INT and COD subtypes, it was found that the rate of ICI responsiveness was significantly higher in the INT group than the COD group (Fig. S7), indicating a possible treatment option using ICI in patients classified into the INT subtype. Considering these findings together with gene expression and mutation profiling results, we suggest that the gene expression-based COD signature reflects ICI responsiveness. However, we also indicate that more rigorous validation steps are needed, because of insufficient GC samples availability.

To more characterize the INT and COD subtypes at copy number or epigenetic alterations, we also performed copy number variation (CNV) and methylation profiling in the TCGA cohort. When exploring CNV data, we observed 842 genes had statistically significant difference in CNV between diffuse-type patients with the INT and COD subgroups. A number of important canonical pathways associated with these genes including cell adhesion molecules were found to be enriched (Fig. S8). When a methylation profiling was carried out for comparing epigenetic alterations between two subtypes, 1412 genes had statistically significant differences in methylation between the INT and COD subtypes. A function enrichment test using these genes revealed that many important canonical pathways including focal or cell adhesion, cell adhesion molecules, and ECM-receptor interaction were found to be enriched (Fig. S9), consistent with the results of gene expression and mutation profiling.

We sought to identify the best candidates that could practically discriminate GC patients of the COD subtype from those of the INT subtype. Among upstream regulatory candidates (Table S2), we selected IGF1 as a good indicator distinguishing the two subgroups, showing the greatest fold difference in mRNA expression (5.04-fold) between the INT and COD subtypes. We also selected NXPE2 as a novel indicator whose relationship with GC has not yet been described. NXPE2 exhibited the greatest fold difference in mRNA expression (29.1-fold) between the INT and COD subtypes among the novel genes. With these two candidates, we carried out protein expression analysis using a tissue microarray, revealing that these proteins were more highly expressed in tumor cells with the COD subtype than in those with INT subtype (Fig. 4a, b). When protein expression levels were compared between subtypes, IGF1 and NXPE2 were found to show significantly higher expression in the COD subtype than in the INT subtype. Further stratification into COD subtype GC patients with histological PCC revealed significantly higher expression of the IGF1 protein than in the other subtypes, whereas no such significant elevation of NXPE2 expression was observed (Fig. 4c), suggesting IGF1 as a good indicator correlated with molecular as well as histological subtypes. By estimating the correlations between histological subtypes and protein expression levels, we found that IGF1 expression was significantly associated with histological PCC, while NXPE2 did not show a significant relationship with the histological classification (Fig. 4d). These results suggest that IGF1 may be a good clinical indicator for classifying high-risk diffuse-type GC patients and supporting histological subclassification systems.

Based on our new finding of NXPE2 characteristics, we sought to verify the effects of cell migration and proliferation of NXPE2 in diffuse-type GC cell lines. We ectopically overexpressed NXPE2 in the SNU601 and MKN45 cells, and the expression was successfully overexpressed in two cell lines using RT-PCR (Fig. S10a). Then, we performed migration assays with NXPE2 overexpressing GC and control cells. We found that ectopic NXPE2 overexpression significantly increased the migration ability in SNU601 and MKN45 cells (Two sample t tests; P = 1.48 × 10^–6 and P = 0.019 in SNU601 and MKN45, respectively; Fig. S10b). We also performed cell proliferation assays with ectopic NXPE2 overexpressing GC and control cells. Although moderate proliferation of SNU601 cells was observed (at 48 h), we found that ectopic NXPE2 overexpression significantly increased the proliferation of MKN45 cells (Fig. S10c). These results suggest that NXPE2 mediates diffuse-type GC aggressiveness promoting cell migration or proliferation, which are key determinants of malignant progression and metastasis.