The analysis of this study was supported by the Sangerbox tool (http://vip.sangerbox.com/) (Shen et al., 2022). The GDC-API was used for downloading all the Copy Number Variation (CNV), RNA-Seq, and mutation data from The Cancer Genome Atlas-Stomach Adenocarcinoma (TCGA-STAD) dataset. The expression data of GSE62254 and GSE15459 data were obtained from the Gene Expression Omnibus (GEO) cohort. In this study, TCGA-STAD was employed as a training dataset, whereas the GSE62254 and GSE15459 datasets were independent verification sets. All data without additional clinical and follow-up information, overall survival (OS) time, or status from the TCGA-STAD dataset were eliminated, while the data samples with an OS value ≥30 days were retained. The gene symbol was created from Ensembl. The annotation data for the associated chip platform was retrieved from the GEO dataset. The probe was mapped to the gene in accordance with the annotation data, and the probe that matched several genes was eliminated. If more than one probe had been found to match a particular gene, the average value was taken into account as a measure of gene expression. Subsequently, a total of 337 primary tumor samples from the TCGA-STAD dataset were included in the study. The expression data of GSE62254 and GSE15459 were obtained from the GEO database. We finally included 300 and 182 GC samples from GSE62254 and GSE15459, respectively.

We retrieved the NK cell-linked gene information from immport (https://www.immport.org/shared/home), including the data on 18 NK_ Cell-associated pathways from the Molecular Signature Database (MSigDB) (Liberzon et al., 2015). Furthermore, the NK cell module from the LM22 database (Newman et al., 2015) was also selected. We finally retained 79 NK cell-related genes for analysis.

The expression of NK cell-associated genes was acquired from the TCGA-STAD expression matrix. The Coxph function of R with threshold value p < 0.05 was used in univariate Cox regression analysis. All the NK cell genes significantly linked to the GC prognosis were obtained.

The consistent matrix was constructed by the ConsensusClusterPlus R package (Wilkerson and Hayes, 2010) based on the expression data of NK cell-linked genes. Using the “km” algorithm and “1-Pearson correlation” as the distance metric, 500 bootstraps were performed, with each bootstrap step employing 80% of all the GC patients in the training set. The optimal classification was determined after estimating the Cumulative Distribution Function (CDF) value and consistent matrix that determined the molecular subtype for each sample. The cluster number ranged between 2 and 10.

Next, various molecular subtypes were examined for the presence of differentially activated pathways. The candidate gene sets in a Hallmark database were subjected to the Gene Set Enrichment Analysis (GSEA) (Liberzon et al., 2015). A significant enrichment value was defined when the False Discovery Rate (FDR) was <0.05.

The relative abundance of 22 immune cell types was determined using CIBERSORT (https://cibersort.stanford.edu/) (Chen et al., 2018). The immune infiltration was analyzed by ESTIMATE (Yoshihara et al., 2013). In a previous report, Jiang et al. (2018) employed the Tumor Immune Dysfunction and Exclusion (TIDE) program (http://tide.dfci.harvard.edu/) to estimate the expected clinical impact of immunotherapy in the identified high and low groups. Further, it was noted when the TIDE score increased, the probability of immune escape increased, which indicated less immunotherapy benefit to patients (Pan et al., 2018). In addition, seven metagenes clusters related to inflammation were collected to evaluate the level of inflammatory activity in different subtypes. In immune regulation, interferon-gamma (IFN-γ) is an important cytokine molecule that functions in anti-cancer immunity. We acquired the GOBP_RESPONSE_ TO_INTERFERON_GAMMA gene set from the Gene Ontology (GO) database, and used it for single sample GSEA (ssGSEA) (Hänzelmann et al., 2013). The immune cytolytic activity (CYT) score was utilized to reflect the cytotoxicity of different subtypes. Additionally, the T-cell–inflamed gene expression profile (GEP) score was used for assessing various molecular subtypes in relation to cancer immunotherapy in terms of predictive significance. The ‘pRRophetic’ R software package was employed for calculating the IC50 value of various drug molecules, including cisplatin, paclitaxel, docetaxel, and 5-fluorouracil (Geeleher et al., 2014).

The differentially expressed NK cell genes among the subtypes were selected by the identified molecular subtypes (FDR < 0.05 and |log2FC|>1.5)). Limma software was employed for assessing the DEGs in the subtypes, i.e., C1 vs. other subtypes, C2 vs. other subtypes, and C3 vs. other subtypes, and the DEGs showing significant relation to GC prognosis were selected (p < 0.05). In addition, the ‘clusterProfiler’ R software package was employed for analyzing the functional enrichment of the DEGs in C1 vs. others, C2 vs. others, and C3 vs. others (Yu et al., 2012). Subsequently, the Lasso analysis was used for lowering the number of genes. Stepwise multivariate regression analysis was then carried out based on the results of the Lasso analysis. The stepwise regression uses the Akachi Information Criterion (AIC) for analysis and considers the statistical fitting of a model as well as the parameter number utilized for the fitting. The stepAIC technique in the MASS package uses the most complicated model in Step 1, successively deletes a variable for decreasing the AIC value, and finally acquires the prognostic significant genes linked to the NK cell phenotype (Zhang, 2016). The risk score (RS) of each patient was assessed using the formula: RS = Σβi × Expi). Here, Expi refers to a gene expression level associated with the genetic characteristics of the prognosis of the NK cell phenotypes, and βi indicates the Cox regression coefficient for the corresponding gene. Based on the threshold “0”, the patients were categorized into low-risk and high-risk categories. The Kaplan Meier (KM) curve was used for plotting the survival curve for analyzing the patient prognosis, while the log-rank test was employed for assessing significant differences. Further, the ‘timeROC’ R software package was utilized to analyze the ROC of the prognostic classification of the model (Blanche et al., 2013). The classification efficiency for the prognostic prediction within 1 year, 2 years, and 3 years were analyzed.

The expression matrices of TCGA-STAD were used to extract the NK-cell-linked expression levels. Univariate Cox regression analysis of the 27 NK-related genes revealed 21 NK cell genes significantly associated with the GC prognosis (Fig. 1A, p < 0.05). The relationship among these 21 NK cell-linked genes was then investigated and interactions among the 21 NK cells were detected (Fig. 1B). The gene expression patterns of the 21 NK cells with a significant link to GC prognosis were then utilized to categorize the patients using consistent clustering, with the CDF calculating the ideal number of clusters. When the Cluster number was set to 3, the CDF Delta area curve revealed that the clustering findings were mainly robust. We set k = 3 to create three unique molecular subtypes with significant prognostic characteristics (Figs. 1C–1E). In general, C3 had the best prognosis, followed by C2 and C1 (Fig. 1F). Additionally, the GSE62254 data were subjected to molecular subtyping using the same procedure, and it was shown that these three molecular types had significantly different prognoses, which was similar to the TCGA-STAD analysis (Fig. 1G).

We examined the distribution of the clinical features in each molecular subtype in the TCGA-STAD dataset to analyze clinical characteristics among the molecular subtypes. It was found that the number of patients with the N stage of N3 in the C1 subtype increased significantly, while the number of patients with the M1 stage in the M stage increased significantly. In addition, it was found that for the T stage, among C1 subtypes with poor prognosis, patients with T stages of T2, T3, and T4 were significantly higher compared to the C2 and C3 molecular subtypes. As for the pathological stage, stages II, III, and IV patients in the C1 subtype were significantly higher compared to C2 and C3 molecular subtypes. At the same time, the C1 subtype included GC patients younger than 60 years of age. More importantly, no significant difference was noted in the C1, C2, and C3 subtypes (Fig. 2A). The differences among different molecular subtypes and clinical information in GSE62254 data were also compared, and the results were consistent with the TCGA-STAD analysis outcomes (Fig. 2B).

We further studied the differences in genomic alterations among these three molecular subtypes in the TCGA-STAD dataset. The molecular features of the TCGA-STAD cohort in this study were obtained from an earlier pan-cancer publication (Thorsson et al., 2018). We observed that the C3 had the highest TMB, C1 had the highest intratumor heterogeneity, and C2 had the highest aneuploidy score, homologous recombination deficiency, purity, and ploidy (Fig. 3A). In addition, the molecular subtype of GC was acquired from a previous study, which identified five molecular subtypes (i.e., genomically stable (GS), chromosomal instability (CIN), Epstein-Barr virus (EBV), microsatellite instability (MSI), and plus tumors with elevated single nucleotide variants (HM-SNV)) (Liu et al., 2018). These two molecular subtypes were compared, and it was seen that the CIN accounted for more in the C1 and C2 subtypes, while microsatellite instability (MSI) in the C3 subtype was significantly higher than C1 and C2 subtypes. We also compared the MSI status variations in the molecular subtypes. The majority of patients in the C1 and C2 subgroups were Microsatellite stable (MSS) (Figs. 3B–3C). Furthermore, the discrepancies in gene mutations among the distinct molecular subtypes were studied. The top twenty genes with substantial mutations were displayed. The mutation frequencies of TTN, MUC16, and other genes varied significantly among the three molecular subtypes (Fig. 3D).

We next analyzed differentially activated pathways in the different molecular subtypes. When compared with GSE62254, TCGA-STAD was significantly enriched in 31 pathways (Fig. 4A). Generally, the activated pathways mainly were some epithelial-mesenchymal transition (EMT)-linked pathways, such as EPITHELIAL_MESENCHYMAL_TRANSITION, TGF_BETA_SIGNALING and several immune-related pathways. In addition, pathway differences among the TCGA-STAD subtypes were also analyzed. Generally, EMT-associated pathways in C1 patients were activated. For example, in the C1 subtype, the EPITHELIAL_MESENCHYMAL_TRANSITION pathway was significantly activated compared with C2 and C3. In addition, several immune-linked pathways were also activated in the C1 subtype, such as INTERFERON_GAMMA_RESPONSE. Therefore, these subtypes differ in pathway activation, which may affect the prognosis of different subtypes (Fig. 4B).

CIBERSORT software and ESTIMATE software were used to assess the infiltration of various immune cells to further analyze the differences in the immunological TME across various molecular subtypes. Most immunological cell types had significant differences among subtypes, in which the infiltration level of macrophages in the M2 phase in C1 was significantly increased, while macrophages in the M0 phase in C1 were significantly decreased. More importantly, there was a slight difference among activated NK cells in subtypes, but the infiltration of resting NK cells was significantly decreased (Fig. 5A) in C1 subtypes. Additionally, the immune score of the C1 subtype was noticeably higher as it showed a greater immune cell infiltration (Fig. 5B). Additionally, the same algorithm and outcomes from the TCGA-STAD were utilized to compare the infiltration levels of the immune cells using the GSE62254 data (Figs. 5C–5D). Moreover, the levels of the inflammatory activity of three molecular subtypes were calculated. Apart from MHC_I, there were significant differences among these seven meta-genes clusters, indicating an immune inflammation caused by the high degree of immune cell infiltration in the C1 subtype. Moreover, it also indicated that over-activation of immunity may promote the development of tumors instead, resulting in a poor prognosis. This observation was also observed in the GSE62254 dataset (Figs. 5E–5F).

Few representative molecules were assessed for Immune Checkpoint Blockade (ICB) of cancer immunotherapy. The results indicated that programmed death 1 (PD-1), programmed death ligand 1 (PD-L1), and cytotoxic T lymphocyte-associated antigen 4 (CTLA4) were significantly overexpressed in the C3 subtype (Fig. 6A) Further, the response of IFN-γ was significantly enhanced in C1 and C3 subtypes (Fig. 6B). Simultaneously, a comparison of the expression differences of INFG gene in the three subtypes showed that INFG was significantly overexpressed in the C3 subtype (Fig. 6C). In addition, the CYT score reflecting the cytotoxicity was significantly higher in C1 and C3 molecular subtypes than in the C2 subtype (Fig. 6D). Additionally, the T-cell-inflamed GEP score was utilized for predicting the clinical responses of different molecular subtypes and the potential to obtain the unique characteristics of neoantigenicity and T-cell activation, respectively. It was found that the T-cell-inflamed GEP score was significantly increased in C1 and C3 subtypes (Fig. 6E). More crucially, the C1 subtype had a high TIDE score, indicating a higher risk of immunological escape. Further, C1 was the molecular subtype with the greatest exclusion and dysfunction scores (Fig. 6F). We also studied the response of different molecular subtypes in TCGA-STAD to conventional chemotherapy agents such as paclitaxel, cisplatin, docetaxel, and 5-fluorouracil. Paclitaxel and 5-fluorouracil sensitivity were increased in the C1 subtype (Fig. 6G).

In the previous analysis, NK cell genes were identified through univariate Cox analyses, and three different molecular subtypes were identified. We next explored the differential genes among different subtypes. The differential gene volcano map of C1 vs. other subtypes contained 271 down-regulated genes and 1158 up-regulated genes (Suppl. Fig. S1A). In the C2 vs. other subtype differential gene volcano map, 341 down-regulated genes and 89 up-regulated genes were determined (Suppl. Fig. S1B). In the C3 vs. others differential gene volcano map, 494 down-regulated genes and 59 up-regulated genes were noted (Suppl. Fig. S1C).

Furthermore, by analyzing the functional enrichment of DEGs in C3 vs. others, C1 vs. others, and C2 vs. others, it was seen that the DEGs of the C1 subtype were significantly enriched in some signal pathways related to metastasis and invasion, such as focal adhesion, collagen-containing extracellular matrix, extracellular matrix, and organization (Suppl. Fig. S1D). In the C2 subtype, drug metabolism-cytochrome P450, antimicrobial humoral response, and other pathways related to drug metabolism and immunity were enriched (Suppl. Fig. S1E). Further, the cell cycle, chemokine receptor binding, chromosome segregation, and other pathways were enriched in the C3 subtype (Suppl. Fig. S1F).

We selected 1565 genes finally by identifying differential genes among molecular subtypes in the previous analysis. Following this, a univariate Cox regression analysis was performed to examine DEGs across various subtypes. We discovered a total of 448 genes with a larger influence on GC patient prognosis (p0.05), including 407 risk genes and 41 protective genes (Suppl. Fig. S2A). To reduce the number of genes in the suggested risk model, the 448 genes with a significant relation to GC prognosis were further compressed using Lasso regression. We first investigated the modified trajectory of each independent variable. According to the analysis, the proportion of independent variable coefficients that decreased to zero increased as the lambda value increased. The model was developed using 10-fold cross-validation, and the CI for each lambda value was explored. The model achieved an ideal level, as indicated in the figure, when lambda = 0.0697. Therefore, 13 genes were chosen as target genes for the subsequent step when the value of lambda was 0.0697 (Suppl. Fig. S2B). Furthermore, based on the 13 genes obtained in the Lasso analysis results, stepwise multivariate regression analysis determined five NK cell-related genes affecting prognosis: MATN3, SERPINE1, ARHGEF39, VSNL1, and ENTPD2 (Suppl. Figs. S2C–S2D).

Based on the formula described by the RS of these samples, the RS of NK cell-related signatures for each sample were calculated and normalized. The results of the RS distribution of patients in the TCGA-STAD dataset indicated that the RS-high samples had a poor prognosis (Fig. 7A). Additionally, the classification effectiveness of the prognosis prediction at 1 year, 2 years, and 3 years was examined, correspondingly, and the model showed a higher AUC area (Fig. 7B). Finally, groups with RS values below 0 were classified as low risk, whereas those with RS values over 0 were classified as high risk. The KM curve was then created that demonstrated the highly significant differences between the RS-high and RS-low groups. Of the 173 samples divided into the RS-high groups and the 164 samples included in the RS-low groups, patients with higher RS had worse OS in the training dataset (Fig. 7C). We then confirmed the reliability of the clinical prognostic model of the NK cell-related gene signature by testing it in the GSE62254 and GSE15459 datasets. The same formula was used to determine the RS for patients. Similar findings to those of the training set were seen in the verification analysis. While the RS-high had a bad prognosis, the prognosis of the RS-low was good (Figs. 7D–7G).

Analysis of the distribution of the RS among the clinical pathological characteristic groups in the TCGA-STAD dataset showed that patients with the late-clinical stage displayed a significantly higher RS than those at the early clinical stage. In addition, the results showed that patients younger than 60 years old had higher RS (Suppl. Fig. S3A). Further, according to the findings, the C1 molecular subtype had a much greater RS than the C3 molecular subtype, which had a better prognosis. In addition, the RS-high group included a majority of the C1 molecular subtypes (Suppl. Fig. S3B). We next compared the prognostic differences in the high- and low-risk groups with regard to different clinical pathological characteristics determined in the TCGA-STAD dataset. The results revealed that the risk group also had favorable results in different clinical groups, implying the reliability of the risk group (Suppl. Fig. S3C). At the same time, the RS differences among various clinical attributes in the GSE62254 cohort were also compared. It was seen that a higher clinical grade was related to a higher RS. In addition, the RS variations in the different molecular subtypes were compared, and the distribution of molecular subtypes and risk types were determined. The RS model proposed in this study also showed robust working in the GSE62254 cohort.

We subsequently analyzed the TME in patients of the different RS groups. In the low- and high-risk groups in the TCGA-STAD dataset, the relative abundance of the 22 different immune cells was compared. It was observed that both the RS groups showed significant differences in the abundance of immune cells (Fig. 8A). Furthermore, the relation of these 22 immune cell components with the RS was analyzed. It was found that the RS was significantly and negatively associated with the resting NK cells but was positively related to the M2 macrophages (Fig. 8B). We used ESTIMATE to assess the immune cell infiltration. The high-RS group showed a higher immune infiltration (Fig. 8C) compared to the low-RS group. For assessing the correlation between the RS and the biological functions of various samples, the expression profiles in the TCGA-STAD were selected and analyzed using the ssGSEA. The results indicated that the RS-high category was significantly enriched in several EMT-related pathways, including HALLMERK_TGF_BETA_SIGNALING, HALLMARK_IL2_STAT5_ SIGNALING, etc. (Fig. 8D). The correlation between the enrichment score of the functions and RS was then estimated and the functions with a correlation >0.4 were chosen. We noted that the RS was positively correlated with cell cycle-linked pathways (Fig. 8E).

We found that the T-cell-informed GEP score was significantly increased in the RS-high group (Fig. 9A). Simultaneously, the RS-high subtype also showed a significantly enhanced IFN-γ response (Fig. 9B). The results indicated no significant differences in the CYT scores that were used for assessing the cytotoxicity between the risk groups (Fig. 9C). We then evaluated a few representative immune checkpoint molecules, and no differential expression between CTLA4 and PD-L1 in terms of the risk resistance was found (Fig. 9D). TIDE was used for assessing the probable clinical effects of immunotherapy in the defined risk group. It was seen that the RS-high subtype had a high TIDE score, indicating that the RS-high subtype was more likely to escape from traditional immunotherapy. In addition, it was also noted that the RS-high molecular subtype had the highest exclusion and dysfunction scores (Fig. 9E). Further, the responses of different RS groups in the TCGA-STAD cohort were also analyzed against conventional chemotherapy drugs like paclitaxel, cisplatin, docetaxel, and 5-fluorouracil. We found that the RS-low was more sensitive to the chemotherapy drugs cisplatin and docetaxel on the whole (Fig. 9F).

The decision tree was constructed according to the clinical variables, including age, gender, TNM stage, pathological information, and the RS of patients in TCGA-STAD. We found that four different risk subgroups could be identified using the decision tree, among which RiskType was the most powerful prognostic factor, and the four risk groups showed significant differences in their OS rates (Figs. 10A–10B). Patients with RS-low scores were included in the risk groupings “Lowest,” “Low,” and “Mediate,” which had poorer prognoses. Additionally, disparities in the distribution of the molecular subtypes in the different risk categories were observed, with the molecular subtypes C1 and C2 of the “High” risk subgroup accounting for the majority of these variances (Figs. 10C–10D). The findings of the univariate Cox analysis of RS and clinicopathological characteristics (HR = 1.72, 95% CI: 1.45–2.04) and the multivariate Cox analysis (HR = 1.72, 95% CI: 1.44–2.05) revealed that the RS was the most important prognostic factor (Figs. 10E–10F). A nomogram combining the RS with other clinicopathological features was developed to quantify the survival probability and risk assessment of the GC patients. Again, RS showed the greatest effect on the survival rate prediction (Fig. 10G). Furthermore, the calibration curve was plotted. For three calibration points at 1 year, 2 years, and 3 years, the prediction calibration curve and the standard curve results were comparable, indicating a strong prediction by the nomogram (Fig. 10H). Further evaluation of the reliability of the model was reflected by DCA, and the RS and nomogram had advantages over the extreme curve that were much greater. When compared to other clinicopathological variables, the nomogram and the RS had the highest potential in survival prediction (Fig. 10I). The ROC analysis showed that the 1-, 2- and 3-year AUC of the nomogram had the highest AUC in the 1-, 2- and 3-year ROC analysis (Fig. 10J).