Abstract Stomach adenocarcinoma (STAD) is the most prevalent gastrointestinal malignancy and seriously threatens the life of the global population. Anoikis, a process of programmed cell death that occurs when cells detach from the extracellular matrix, is closely associated with tumor invasion and metastasis. In this study, we used the TCGA-STAD database to identify the expression patterns and prognostic relevance of anoikis-related genes (ARGs) in STAD. Functional enrichment analysis was used to explore the potential pathway. LASSO and Cox regression were used to construct anoikis-related prognostic signature. The anoikis risk score (ARS) incorporated 7 genes and stratified patients into highand low-risk subgroups by median value splitting. In addition, external validation was performed based on [36]GSE66229, [37]GSE15459, and [38]GSE84437 cohorts. Nomograms were created based on risk characteristics in combination with clinical variants and the performance of the model was validated with time-dependent AUC, calibration curves, and decision curve analysis (DCA). The prognostic signature indicated that the low-risk subgroup had better outcomes and significant correlations with tumor microenvironment, immune landscape, immunotherapy response, and drug sensitivity. In addition, single-cell analysis displayed the cell types, the subcellular localization of prognostic genes, and the cellular interaction to reveal the potential molecular communication mechanism of anoikis resistance. Finally, in vitro experiments confirmed the critical role of CRABP2 in STAD. The results indicated that CRABP2 knockdown inhibited gastric cancer cell proliferation, migration and invasion, and promoted apoptosis. In summary, ARS can serve as a biomarker for predicting survival outcomes in STAD patients, providing new tools for personalized treatment decisions for STAD patients. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-025-88882-9. Keywords: Bioinformatics, Stomach adenocarcinoma, Anoikis, Tumor microenvironment, Immune landscape Subject terms: Cancer, Computational biology and bioinformatics Introduction Gastric cancer was a prevalent cancer type globally, accounting for about 1 in every 13 cancer-related deaths. According to the GLOBOCAN2020 epidemiological data, incidence and mortality rates of gastric cancer rank 5th and 4th worldwide, respectively^[39]1. Stomach adenocarcinoma (STAD) was is one of the most common pathological types of gastric cancer. Accumulating evidence indicated that environmental, epigenetic, and genetic factors are associated with the generation and development of gastric cancer^[40]2. Most cases of STAD occur sporadically, with only a small percentage having a family history^[41]3. Nowadays, radical surgery is still the mainstay of treatment for patients with early-stage resectable STAD. Unfortunately, approximately 60% of STAD patients undergoing radical surgery might experience tumor recurrence. Meanwhile, less than 25% of STAD were suitable candidates for radical surgery due to the commonly advanced stage and distant metastasis when the STAD was detected^[42]4. The advent of targeted therapies and immune checkpoint inhibitors (ICIs) has significantly expanded the treatment options available to physicians. As a highly heterogeneous tumor, approximately 10–20% of STAD cases exhibit overexpression of the human epidermal growth factor receptor 2 (HER2)^[43]5. Consequently, trastuzumab, pertuzumab, and lapatinib have been incorporated into the treatment of HER2-overexpressing STAD, leading to notable improvements in the prognosis of patients with advanced STAD. However, despite the therapeutic success of trastuzumab as a first-line treatment for HER2-positive STAD, most treatment failures arise from primary or acquired resistance^[44]6. Although diagnostic and therapeutic advancements have been made, the 5-year overall survival rate for STAD remains around 25% ^7, and patients with advanced metastatic STAD face a dismal 5-year survival rate of less than 5% ^8. The median overall survival for these patients is only 12 months^[45]9. The lack of specific biomarkers for early diagnosis and efficacy prediction in STAD underscores the urgent need to identify potential predictive biomarkers or therapeutic targets and optimize existing therapies to improve patient outcomes. Anoikis, a programmed cell death process essential for anchorage-dependent cells, has recently attracted considerable attention. Recent research has demonstrated that tumor cells’ uncontrolled proliferation and ability to resist anoikis are fundamental to distant metastasis^[46]10. Tumor metastasis is a complex, multistep process involving local invasion, intravasation into the bloodstream, circulation, extravasation from the bloodstream, proliferation at a distant site, and angiogenesis^[47]11. The initiation of metastasis is contingent upon interactions between tumor cells and stromal cells, as well as the epithelial-to-mesenchymal transition (EMT) in individual tumor cells^[48]12. EMT and leaky blood vessel formation facilitate tumor cells’ intravasation. Circulating tumor cells that enter the bloodstream rely on anoikis resistance and the ability to evade immune cell cytotoxicity to establish distant metastases. Anoikis resistance is a critical early step in enabling tumor cells to colonize distant organs. Therefore, we aimed to establish a novel prognostic signature for STAD patients by incorporating ARGs. In our study, we performed differentially expressed genes between cancer and normal tissues. Second, we used univariate Cox regression, LASSO regression, and multivariate Cox regression to construct the anoikis risk score (ARS) for prognostic signature. Three independent external GEO datasets were used as validation cohorts to evaluate the accuracy and robustness of the model^[49]13. We stratified them into low- and high-risk subgroups using the median ARS as the cut-off. We conducted functional enrichment analyses to explore the molecular mechanisms underlying the survival differences between the subgroups. Additionally, we evaluated the correlation between the ARS, the tumor microenvironment (TME), immune cell infiltration, and sensitivity to chemotherapy and targeted drugs. Single-cell analysis was performed to validate our findings and enhance our understanding of the mechanisms involved. Finally, in vitro experiments confirmed the critical role of CRABP2 in STAD. Our study offers novel insights into personalized therapeutic strategies and prognostic prediction for STAD patients from the perspective of anoikis. Materials and methods Data collection The RNA-seq data and clinical information for STAD were obtained from the TCGA database ([50]https://portal.gdc.cancer.gov/). The RNA-seq data included 376 tumor and 36 normal samples and the clinical characteristics, including age, gender, TNM stage, survival status, and survival time. Three external validation cohorts, [51]GSE66229, [52]GSE15459, and [53]GSE84437, were acquired from the GEO database ([54]https://www.ncbi.nlm.nih.gov/geo/). The clinical information of STAD patients in training datasets (TCGA) and the validation datasets (GEO) were displayed in Table [55]S1. The list of anoikis-related genes was collected from the Genecards database ([56]https://www.genecards.org/), and details of anoikis-related genes were presented in Table [57]S2. Identification of differentially expressed genes (DEGs) and anoikis-related DEGs The R package “DESeq2” was applied to identify DEGs between normal and tumor samples in TCGA datasets or between high-risk and low-risk subgroups^[58]14. With a threshold setting at |Log2fold change| > 1.5 and p-value < 0.05, we acquired a total of 19,543 DEGs of STAD. The DEGs were performed through Wald’s test. After removing duplicated genes between DEGs and anoikis-related genes, we obtained 105 anoikis-related DEGs of STAD by taking intersection. Construction of prognostic model for anoikis-related DEGs based on TCGA datasets To investigate the association between the anoikis-related DEGs and the survival information of STAD patients, we attempted to construct a prognostic anoikis-related risk score signature. Before proceeding with subsequent analysis, we scaled the expression data of DEGs to normalize the data first. Then the anoikis-related DEGs were screened by univariate Cox regression analysis via R package “survival”. Moreover, the least absolute shrinkage and selection operator regression (LASSO) analysis via “glmnet” R package and multivariate Cox regression analysis. The risk score for each sample was calculated using the formula below: graphic file with name M1.gif According to the formula, all participating cohorts were divided into low-risk and high-risk subgroups based on the median risk score. The optimized cutoff and Kaplan-Meier (K-M) survival curve (“survival” and “survminer” R package) was conducted to estimate the difference in overall survival rates between the two subgroups. Meanwhile, the time-dependent receiver operating characteristic (ROC) curve (“timeROC” R package) was plotted to evaluate the prognostic capability of the risk model. Validation of a prognostic model for anoikis-related DEGs based on GEO datasets Three external and independent validation cohorts from GEO ([59]GSE66229, [60]GSE15459, and [61]GSE84437) were used to verify the robustness of the prognostic model. The risk score for each sample was calculated in three external validation cohorts and then divided into low-risk and high-risk subgroups. The clinical characteristics of the subjects were represented on K-M and ROC curves, respectively. Nomogram model establishment The “rms” R package was used to plot the line graphs. Moreover, we generated calibration curves for 3-, 4-, and 5-year overall survival to assess the accuracy of the nomogram model. For an ideal predictive nomogram model, the prediction results should be able to fall on the 45-degree diagonal of the calibration plot and have a higher Harrell consistency index (C-index). Decision curve analysis (DCA) was also used to assess the net clinical benefits of the nomogram model^[62]15. Functional annotation and enrichment analysis ARGs were performed with Gene Ontology (GO) and Kyoto Encyclopedia of Genes Genomes (KEGG) enrichment analysis^[63]16–[64]18 (“clusterProfiler”, “org.Hs.eg.db”, “GOplot”, R package). Moreover, we examined the alternation between the low-risk and high-risk subgroup with Gene set enrichment analysis (GSEA) to evaluate the potential biological process by using the C5 (c5.all.v2023.1.Hs.entrez) subset derived from the Molecular Signature Database (MsigDB, [65]https://www.gsea-msigdb.org/gsea/msigdb/)^[66]19. For the terms of GSEA, we selected the hallmark pathway with a p-value < 0.05 and a false discovery rate (FDR) < 0.25 so that these pathways were considered significant. The gene set variation analysis algorithm (GSVA) was also applied based on 262 hallmark signatures collected by “IOBR” R package^[67]20 to identify critical signaling pathways between the low- and high-risk subgroups (“GSVA” R package)^[68]21. Tumor microenvironment (TME) analysis ESTIMATE was an algorithm that used gene expression signatures to infer the fraction of stromal and immune cells in tumor tissue by calculating the tumor tissue stromal and immune scores to predict the level of infiltrating stromal and immune cells, thus inferring tumor purity in the tissue. Stromal, Immune, Tumor Purity, and ESTIMATE scores were calculated using the “estimate” R package. Moreover, we also investigated the relationship between risk score and TME components. Immune cell infiltration and checkpoint analysis The list of 22 tumor-infiltrating immune cells (TIICs) signatures was downloaded from TISIDB ([69]http://cis.hku.hk/TISIDB/download.php). We calculated immune cell infiltration with the single-sample gene set enrichment analysis (ssGSEA). The relative abundance of each immune cell population was also predicted between low-risk and high-risk subgroups with the CIBERSORT algorithm by R package “IOBR”. Also, the “MCPcounter”^[70]22, “xCell”^[71]23, “IPS”^[72]24, and “TIMER”^[73]25 algorithm was employed to validate the results. We compared the proportion of the TIICs with t test (p < 0.05). In addition, the expression of immune checkpoint genes (ICGs) was also selected for analysis. Therefore, we could intuitively identify the difference in immune checkpoint gene levels between low- and high-risk subgroups. Tumor immune dysfunctions and exclusions analysis Recent studies have shown that the Tumor Immune Dysfunction and Exclusion (TIDE) score was the best predictor of anti-PD1 and anti-CTLA4 therapy. In contrast, the predictive function of the TIDE score for immunotherapy response was stable regardless of the degree of cytotoxic T cells. Furthermore, a higher TIDE score was associated with poorer predictive efficacy of immune checkpoint inhibition therapy. The results of the TIDE score, dysfunction, and exclusion score were acquired from the Tumor Immune Dysfunction and Exclusion website ([74]http://tide.dfci.harvard.edu/), and the response to immunotherapy was inferred by TIDE score, dysfunction, and exclusion score^[75]26. Mutation analysis based on the predictive signature model The TCGA mutation data of STAD patients was downloaded from the cBioportal database ([76]https://www.cbioportal.org/)^[77]27,[78]28. The R package “maftools” was then used to identify the differences in somatic mutation data between the low-risk and high-risk subgroups and plotted in the form of waterfall charts. Drug sensitivity prediction analysis based on risk score With the “pRRophetic” R package, the half-maximal inhibitory concentrations (IC50) of four common chemotherapeutic drugs (Docetaxel, Cisplatin, Doxorubicin, and 5-Fluorouracil) and two targeted drugs (Lapatinib and Sunitinib) were investigated based on the Genomics of Drug Sensitivity in Cancer (GDSC) database ([79]https://www.cancerrxgene.org/) between the low-risk and high-risk subgroups^[80]29. Single-cell analysis Single-cell RNA Sequencing (scRNA-seq) data were acquired from the GEO dataset ([81]GSE206785). The expression data were normalized via R package “Seurat”^[82]30. We identify the first 2000 highly variable genes. Cellular annotation was conducted by cell marker. The R package “t-SNE” was used to map the distribution of cellular types and subcellular localization of prognostic genes. The “CellChat” R package was utilized to analyze intercellular communication networks between cell subtypes from scRNA-seq data^[83]31. Cell lines and cell culture The GES-1, AGS, HGC-27, MKN-45, and SGC-7901 cell lines were obtained from the Stem Cell Bank of the Chinese Academy of Sciences. All cell lines were cultured in RPMI 1640 medium supplemented with 10% fetal bovine serum (FBS), 100 U/mL penicillin, and 0.1 mg/mL streptomycin (Solarbio, Beijing, China) under standard cell culture conditions at 37 °C with a 5% CO[2] atmosphere. Cell transfection Small interfering RNA (siRNA) targeting human CRABP2 and negative control siRNA (siNC) were purchased from Shanghai GenePharma (Shanghai, China) and were transfected into AGS and HGC-27 cells according to the manufacturer’s instructions. AGS and HGC-27 cells were seeded the day before transfection at a density of 30–50% confluence. The siRNA duplexes were diluted in RPMI 1640 without FBS, and then Lipo8000TM Transfection Reagent (Beyotime) was added to the siRNA solution, vortex-mixed, and incubated for 15 min at room temperature. The Lipo8000-siRNA complexes were added to fresh medium and incubated with cells at 37 °C. After 48 h, the transfected cells were collected and used in further experiments to evaluate knockdown efficiency using quantitative real-time PCR and western blot analyses. Real-time qPCR assay The quantitative real-time PCR assay was designed to validate the efficiency of transfection. The primer sequences were designed as follows: CRABP2: F: 5’-CCTCTTGCAGGGTCTTGCTT-3’, R: 5’-GGGCTAGGACTGCTGACTTG-3’; GAPDH: F: 5’-CGGGAAGGAAATGAATGGGC-3’, R: 5’-GGAAAAGCATCACCCGGAGG-3’. The expression level of the GAPDH was taken asendogenous control, and the 2^−△△Ct value was used to qualify the relative gene expression levels. CCK8 assay Cell Counting Kit-8 (CCK-8; Dojindo, Japan) was used to assess cell proliferation ability according to the manufacturer’s instructions. Cells were plated into 96-well plates at 0.5 × 10^4 cells/well density and incubated for the indicated time points. Ten microliters of CCK-8 solution were added to each well, and the plates were further incubated for 1 h. Absorbance was measured at 450 nm using a microplate reader. Colony-formation assay AGS (1 × 10^3 cells per well) and HGC-27 (1 × 10^3 cells per well) cells were plated in 6-well plates and cultured for 7 days to evaluate cell proliferation and colony formation ability. The cells were then fixed and stained with 0.1% crystal violet, and colonies were counted and imaged. Wound healing assay The wound healing assay was conducted to evaluate cell migration. AGS and HGC-27 cells were wounded with a sterile 200 µL pipette tip and then washed with PBS three times. The cells were incubated for 48 h at 37 °C, and images of the initial wound and healed wound were captured using a LEICA DMIL LED microscope equipped with a digital imaging system (Germany). Transwell migration assay Cell migration was assessed using 24-well transwell chambers with 8 μm pores. Cells (1 × 10^4 cells/well) were seeded into the upper chamber in a serum-free medium, while the complete growth medium was added to the lower chamber as a chemoattractant. After 24 and 48 h of culturing at 37 °C, non-migrating cells in the upper chamber were removed with cotton swabs, and migration cells on the lower membrane surface were fixed in methanol and stained with 0.1% crystal violet (Sigma-Aldrich) for 15 min. The migrated cells were then quantified and photographed under a LEICA DMIL LED microscope. Flow cytometry Flow cytometry analysis was performed to evaluate cell apoptosis. Cells were stained with FITC Annexin V Apoptosis Detection Kit I (BD Biosciences, 556547), and non-stained cells were used as controls. One microliter of antibody was added to 100 µl of cell suspension (1 × 10^6 cells/mL) and incubated at 4℃ for 30 min. Data were acquired on a CytoFLEX flow cytometer (Beckman Coulter, Brea, USA) and analyzed using CytExpert software. Statistical analysis R (version 4.3.0) was applied for statistical analysis. Wilcoxon test was used to analyze the different functions of tumors and normal tissue. In addition, t test was used in drug sensitivity analysis. C-index was used to estimate the predictive power of risk score to overall survival. The correlation was evaluated using the Spearman method. The chi-square test was applied to assess the difference in therapeutic response. p < 0.05 was defined as statistically significant. Result Identification of the anoikis-related DEGs and functional enrichment analysis Firstly, we created the general workflow diagram of our study, which is presented in Fig. [84]1. The details of DEGs between patients with STAD and normal controls in the TCGA cohort were shown in Table [85]S3, 2652 protein-coding DEGs, including 1289 up-regulated genes and 1363 down-regulated genes, which were displayed via volcano plot and heatmap between normal and tumor groups (Fig. [86]S1A-B). Fig. 1. [87]Fig. 1 [88]Open in a new tab The flow chart of this study. A novel prognostic signature based on the anoikis-related genes in STAD In our study, we developed a novel risk model to assess the prognostic value of ARGs in STAD. The Venn diagram showed 105 anoikis-related DEGs in STAD (Fig. [89]2A, Table S4). Based on univariate Cox regression analysis (p < 0.05, Fig. [90]2B), we obtained 19 prognostic genes. LASSO regression further contributed to our study to construct a model with fewer genes. As shown in Fig. [91]2C-D, 11 prognostic genes through LASSO analysis. Finally, 7 prognostic genes were further selected through multivariate Cox regression analysis (Fig. [92]2E), including CXCL1, NOX4, SERPINE1, CRABP2, TRIM50, FRZB and PLG, to establish a novel prognostic signature for STAD patients (Table S5). graphic file with name M2.gif Fig. 2. [93]Fig. 2 [94]Open in a new tab Construction of a prognostic signature for STAD patients based on anoikis-related DEGs. (A) Venn diagram showed the overlap of 2652 DEGs and 794 anoikis-related genes, which led to 105 common genes being identified. (B) Univariate Cox regression analysis selected 19 genes in STAD. (C) 11 prognostic genes were selected by the LASSO regression. (D) Multivariate Cox regression analysis revealed that 7 genes were associated with prognosis in STAD. (E) The bar charts presented the prognostic genes and corresponding coefficients of the prognostic signature. To explore the difference in survival, we compared the expression of 7 prognostic genes between low- and high-risk subgroups (Fig. [95]S2). We found that the expression of CXCL1 was higher in the low-risk subgroup compared with the high-risk subgroup, while the expression of NOX4, SERPINE1, CRABP2, TRIM50, FRZB, and PLG was lower. Moreover, we attempted to present a positioning atlas of 7 prognostic genes in the chromosome (Fig. [96]S3). We are very interested in the upstream regulation of prognostic genes, so we analyzed the transcription factors (TFs) of prognostic genes based on the TRRUST database. The results showed 6 upstream transcription factors with statistically significant differences. The box plot indicates that the expression of SRF is significantly elevated in the high-risk group, while E2F1 and PARP1 are significantly reduced (Fig. S4). According to the median risk score, patients were divided into high-risk and low-risk subgroups. K-M curves showed patients in high-risk subgroups had worse overall survival (p = 0.0006, Fig. [97]3A). To assess the accuracy of prognostic risk models in predicting 3-, 4-, and 5-year overall survival, ROC curves were plotted with AUC values of 0.684, 0.750, and 0.810, respectively (Fig. [98]3B). Moreover, we explored the relationship between risk score and survival status, survival time, and risk ranking (Fig. [99]3C-D). The expression of the prognostic genes based on the result of multivariate Cox regression analysis was presented with a heatmap (Fig. [100]3E). Fig. 3. [101]Fig. 3 [102]Open in a new tab Construction of a prognostic signature for STAD patients based on anoikis-related DEGs. (A) K-M curves displayed the overall survival of patients in the high- and low-risk subgroups in the TCGA training cohort. (B) ROC curves for predicting 3-, 4-, and 5-year overall survival in the high- and low-risk subgroup in the TCGA training cohort. The distribution of risk score (C), survival status (D), and the expression levels of prognostic genes (E) in the signature. Validation of the prognostic risk signature via GEO datasets To evaluate the robustness of the novel prognostic risk signature, we used the [103]GSE66229, [104]GSE84437, and [105]GSE15459 datasets. In the [106]GSE66229 dataset, the AUC values for 3-, 4-, and 5-year were 0.655, 0.664, and 0.657, respectively (Fig. S5A). Meanwhile, in the [107]GSE84437 dataset, the AUC value of ROC curves for 3-, 4-, and 5-year were 0.609, 0.612, and 0.611, respectively (Fig. S5B). In [108]GSE15459, the AUC values for 3-, 4-, and 5-year were 0.632, 0.645, and 0.667, respectively (Fig. S5C). Moreover, we compared the clinical characteristics of the low-risk and high-risk subgroups, and the differences in age, race, survival status, and survival time reached statistical significance (Table [109]1). In addition, we desired to analyze the prognostic relationship between the risk score and the clinical data of STAD patients. We used univariate Cox regression analysis to access the clinical characteristics of subgroups, such as age, gender, race, TMN Stage, grade, and tumor classification. The result of the univariate Cox shows the statistical differences of age, T3, T4, N1, N3, M1, Stage III, Stage IV, risk score and multivariate Cox regression analysis showed that age, T3, T4, N1, N3 and Mucinous adenocarcinoma were recognized as statistically significant, after excluding confounders and multicollinearity (Table [110]2). Table 1. Baseline of clinical characteristics between low- and high-risk subgroups. Characteristic Levels Low-risk High-risk p Method n 188 188 Status, n (%) Alive 123 (65.4%) 97 (51.6%) 0.007 Chi-square Dead 65 (34.6%) 91 (48.4%) Age, n (%) < 60 45 (23.9%) 71 (37.8%) 0.004 Chi-square ≥ 60 143 (76.1%) 117 (62.2%) T Stage, n (%) T1 14 (7.5%) 3 (1.6%) 0.075 Fisher’s test T2 41 (21.6%) 38 (20.3%) T3 87 (46.3%) 82 (43.6%) T4 43 (22.9%) 64 (34.1%) TX 3 (1.6%) 1 (0.5%) N Stage, n (%) N0 59 (31.4%) 54 (28.7%) 0.177 Fisher’s test N1 49 (26.1%) 51 (27.1%) N2 39 (20.7%) 35 (18.6%) N3 36 (19.2%) 42 (22.4%) NX 5 (2.6%) 6 (3.2%) M Stage, n (%) M0 174 (92.6%) 165 (87.8%) 0.289 Fisher’s test M1 8 (4.2%) 14 (7.4%) MX 6 (3.2%) 9 (4.8%) Stage, n (%) Stage I 32 (17.0%) 17 (9.0%) 0.441 Fisher’s test Stage II 54 (28.7%) 65 (34.6%) Stage III 80 (42.5%) 81 (43.1%) Stage IV 15 (8.0%) 18 (9.6%) Unknown 7 (3.7%) 7 (3.7%) Gender, n (%) Female 63 (33.5%) 67 (35.6%) 0.664 Chi-square Male 125 (66.5%) 121 (64.4%) Race, n (%) White 110 (58.5%) 130 (69.2%) 0.017 Fisher’s test Asian 42 (22.4%) 41 (21.8%) Black or African American 10 (5.3%) 1 (0.5%) Native Hawaiian or other pacific islander 1 (0.5%) 0 (0.0%) Not reported 25 (13.3%) 16 (8.5%) Grade, n (%) G1 5 (2.7%) 4 (2.1%) 0.002 Fisher’s test G2 81 (43.1%) 47 (25.0%) G3 98 (52.1%) 132 (70.2%) GX 4 (2.1%) 5 (2.7%) Classification, n (%) Adenocarcinoma, intestinal type 40 (21.3%) 34 (18.1%) 0.001 Fisher’s test Adenocarcinoma, NOS 56 (29.8%) 76 (40.4%) Adenocarcinoma with mixed subtypes 0 (0.0%) 1 (0.5%) Carcinoma, diffuse type 25 (13.3%) 39 (20.7%) Mucinous adenocarcinoma 7 (3.7%) 10 (5.3%) Papillary adenocarcinoma, NOS 6 (3.2%) 0 (0.0%) Signet ring cell carcinoma 8 (4.3%) 5 (2.7%) Tubular adenocarcinoma 46 (24.5%) 23 (12.2%) Time, median (IQR) 495 (282.5, 851.5) 420 (272.25, 688.5) < 0.001 Wilcoxon test [111]Open in a new tab Table 2. Univariate and multivariate Cox regression analysis of clinical characteristics in STAD patients. Characteristic Total(N) Univariate analysis Multivariate analysis Hazard ratio (95% CI) p-value Hazard ratio (95% CI) p-value Age 376 < 60 116 Reference >=60 260 0.662 (0.478–0.918) 0.013 2.156 (1.475–3.150) < 0.001 Gender 376 Male 246 Reference Female 130 0.774 (0.549–1.090) 0.142 0.750 (0.521–1.079) 0.750 Race 376 White 240 Reference Asian 83 0.166 (0.023–1.206) 0.076 1.253 (0.755–2.081) 0.383 Black or African American 11 0.134 (0.018–1.001) 0.050 1.251 (0.518–3.023) 0.619 Native Hawaiian or other pacific islander 1 0.185 (0.025–1.398) 0.102 3.646 (0.378–35.180) 0.263 Not reported 41 0.217 (0.026–1.785) 0.155 1.640 (0.860–3.128) 0.133 T Stage 376 T1 17 Reference T2 79 3.492 (0.831–14.671) 0.088 3.611 (0.755–17.283) 0.108 T3 169 4.717 (1.157–19.223) 0.030 7.060 (1.246–39.997) 0.027 T4 107 4.985 (1.209–20.551) 0.026 6.134 (1.050-35.851) 0.044 TX 4 18.731 (3.412-102.832) 0.001 11.017 (1.327–91.439) 0.026 N Stage 376 N0 113 Reference N1 100 1.646 (1.041–2.603) 0.033 1.957 (1.004–3.816) 0.049 N2 74 1.531 (0.922–2.541) 0.100 1.583 (0.693–3.615) 0.272 N3 78 2.670 (1.691–4.216) < 0.001 2.511 (1.097–5.752) 0.029 NX 11 1.123 (0.411–3.074) 0.031 3.375 (0.917–12.430) 0.067 M Stage 376 M0 339 Reference M1 22 2.416 (1.362–4.284) 0.003 2.216 (0.016–4.426) 0.120 MX 15 1.705 (0.751–3.875) 0.202 1.184 (0.484–2.894) 0.712 Pathological Stage 376 Stage I 49 Reference Stage II 119 1.357 (0.719–2.562) 0.346 0.562 (0.209–1.515) 0.255 Stage III 161 2.029 (1.124–3.665) 0.019 0.559 (0.157-2.000) 0.372 Stage IV 33 3.874 (1.949-7.700) < 0.001 0.604 (0.135–2.695) 0.509 Unknown 14 4.442 (1.940-10.168) < 0.001 1.345 (0.338–5.357) 0.674 Grade G1 9 Reference G2 128 1.123 (0.350–3.609) 0.845 1.863 (0.460–7.542) 0.383 G3 230 1.469 (0.465–4.635) 0.512 2.357 (0.590–9.414) 0.225 GX 9 1.864 (0.417–8.338) 0.415 2.785 (0.497–15.613) 0.244 Classification Adenocarcinoma, intestinal type 74 Reference Adenocarcinoma, NOS 132 1.064 (0.693–1.634) 0.778 0.969 (0.589–1.594) 0.902 Adenocarcinoma with mixed subtypes 1 0.000 (0.000-Inf) 0.960 0.000 (0.000-Inf) 0.943 Carcinoma, diffuse type 64 0.770 (0.456-1.300) 0.328 0.672 (0.368–1.226) 0.195 Mucinous adenocarcinoma 17 0.308 (0.095–1.004) 0.051 0.211 (0.062–0.721) 0.013 Papillary adenocarcinoma, NOS 6 0.934 (0.286–3.048) 0.910 2.414 (0.592–9.845) 0.219 Signet ring cell carcinoma 13 1.976 (0.945–4.129) 0.070 1.729 (0.750–3.986) 0.199 Tubular adenocarcinoma 69 0.845 (0.509–1.404) 0.516 0.811 (0.468–1.406) 0.456 ARS 376 Low Risk 189 Reference High Risk 187 1.685 (1.224–2.320) < 0.001 2.171 (1.495–3.153) < 0.001 [112]Open in a new tab Subgroup analysis between the risk score and clinical characteristics Furthermore, subgroup analysis based on clinical characteristics showed higher risk scores in patients older than 60. The risk score increased with the deterioration of T Stage and M Stage, but the risk score did not show significant differences in gender, N Stage, and Stage subgroups (Fig. S6A). In addition, considering the effect of histologic classification, we analyzed the relationship between histologic classification and ARS, and the results showed that Papillary adenocarcinoma, NOS had the lowest ARS. In contrast, Mucinous adenocarcinoma had the highest ARS (Fig. S6B). In addition to this, we modeled subgroups to explore the model’s generalization ability in subgroups, and the results presented that the model had encouraging predictive ability in the remaining subgroups classified on the basis of clinical characteristics, except for the Stage III/IV subgroup (Fig. S7A-F). Construction of nomogram This nomogram chart integrates ARS and clinical characteristics of STAD to predict prognosis and provide a quantitative reference for clinical decision-making. In our study, univariate and multivariate Cox regression analyses were used to determine the ARS (p < 0.0001), N stage (p < 0.05), M stage (p < 0.01), and age (p < 0.001) for constructing the nomogram (Fig. [113]4A). Therefore, we constructed a prognostic nomogram chart (100 points) that integrated the comprehensive ARS and age. The ROC curve of clinical characteristics showed that the top three AUC values of overall survival were ARS (0.810), gender (0.514), age (0.626), T Stage (0.716), N Stage (0.456), and M Stage (0.508) as shown in Fig. [114]4B. The calibration curve showed that the actual survival probabilities at 3-year, 4-year, and 5-year were almost consistent with the predicted survival probabilities of the nomogram chart model (Fig. [115]4C-E). Decision curve analyses also revealed the tremendous clinical utility of the nomogram model, further confirming the reliability of the signature (Fig. [116]4F-H). Fig. 4. [117]Fig. 4 [118]Open in a new tab Construction of nomogram model of prognostic signature and ARS. (A) A nomogram model was constructed to predict STAD patients’ 3-year, 4-year, and 5-year overall survival. (B) ROC curves for ARS, gender, age, T, N, and M Stage of STAD patients. (C-E) Calibration curves of the nomogram model for 3-year, 4-year, and 5-year overall survival. Decision curve analysis for 3-year (F), 4-year (G), and 5-year (H) overall survival of the nomogram model. Functional enrichment analysis GO analysis and KEGG enrichment analysis were included in our study to evaluate the differences in signaling pathways between low-risk and high-risk subgroups, as Fig. [119]5A shows the results of the GO enrichment analysis and KEGG enrichment analysis demonstrating eight signaling pathways with definable biological processes and highly ranked enrichment scores (Fig. [120]5B). GO analysis was mainly enriched in collagen fibril organization, fibroblast growth factor receptor signaling pathway, extracellular matrix organization, extracellular structure organization, fibroblast growth factor receptor signaling pathway regulation, collagen-containing extracellular matrix, and extracellular matrix structural constituent conferring compression resistance. The eight signaling pathways obtained by KEGG enrichment analysis were Cytokine-cytokine receptor interaction, Calcium signaling pathway, Cholesterol metabolism, Renin-angiotensin system, IL-17 signaling pathway, Malaria, beta-alanine metabolism, Protein digestion and absorption. To deepen our understanding of the mechanisms inherent in the differences in survival, we performed a gene set enrichment analysis (GSEA). The results revealed that ARS was significantly associated with Degradation of the Extracellular Matrix, Activation of Matrix Metalloproteinases, ECM Proteoglycans, and Extracellular Matrix Organization in the high-risk subgroup (Fig. [121]5C). The results of gene set variation analysis (GSVA) indicated that various CAFs, NK cells, TGF-β, and EMT pathways were enriched in the high-risk subgroup (Fig. [122]5D). Fig. 5. [123]Fig. 5 [124]Open in a new tab Enrichment analysis between high-risk and low-risk subgroups. (A) Gene Ontology enrichment analysis. (B) KEGG pathway enrichment analysis. (C) The gene set enrichment analysis. (D) The gene set variation analysis. Correlation analysis between ARS and TME Since the signaling pathways associated with ECM, CAFs, and EMT were enriched by functional enrichment analysis, we further explored the relationship between ARS and TME signature. In STAD, the ARS positively correlated with the stromal score (R = 0.47, p < 0.0001, Fig. S8A). The ARS did not correlate positively with the immune score (R = 0.13, p = 0.01, Fig. S8B). Meanwhile, a significant negative correlation was presented between ARS and tumor purity (R = -0.34, p < 0.0001, Fig. S8C). Overall, we found that the higher the ARS, the higher the ESTIMATE score (R = 0.33, p < 0.0001, Fig. S8D). ESTIMATE results indicated that the high-risk subgroup had higher stromal scores and significantly lower tumor purity (p < 0.001, Fig. S8E-H). These results demonstrated that the ARS significantly correlated with the TME signature. Based on the results, we compared the expression of integrin, CXCL, and CCL between low- and high-risk subgroups. The results revealed that integrin and CCL expression in high-risk subgroups was higher than in low-risk subgroups. However, the expression of CXCL was lower in the high-risk subgroup compared with the low-risk subgroup (Fig. S9A-C). Immune cell infiltration analysis We evaluated the proportion of immune cell infiltration in the TME between low-risk and high-risk subgroups (Fig. [125]6A). To access the clinical significance and prognostic value of the feature, we also plotted the correlation between prognostic genes and immune cells (Fig. [126]6B). The high-risk subgroups showed more significant immune cell infiltration and specific cell types than the low-risk subgroup (p < 0.05), as determined by the ssGSEA algorithm. Myeloid-derived suppressor cells (MDSCs), regulatory T Cells (Tregs), macrophage cells, natural killer cells, natural killer T cells, and plasmacytoid dendritic cells were all identified in the TME (Fig. [127]6C). The heatmap of immune cell infiltration via ssGSEA is presented in Fig. S10. These results strongly suggested significant differences in immune cell infiltration between high and low-risk subgroups in the TME. Taken together, ssGSEA algorithm findings indicated a higher level of MDSCs infiltration compared with the low-risk subgroup, and the results of the CIBERSORT algorithm demonstrated a higher proportion of M2 macrophages in the high-risk subgroup compared with the low-risk subgroup (Fig. S11). Meanwhile, we analyzed the correlation between the 22 immune cell subtypes (Fig. S12), demonstrating a potential cross-talk between immune cells in the TME. Therefore, the immunosuppressive microenvironment might be one of the critical reasons for the poorer prognosis of STAD patients in the high-risk subgroup, which requires further study. Moreover, the detailed results of the “MCPcounter”, “xCell”, “IPS”, and “TIMER” algorithms were displayed in Fig. S13. The results displayed the differences in tumor immune microenvironment between high and low-risk subgroups using a heatmap for four algorithms. The IPS algorithm analysis shows that in the low-risk subgroup, Effector cells (EC), MHC molecules (MHC), and Immune Checkpoints (CP) are higher, while the high-risk subgroup has lower scores, indicating issues such as low antigen expression, exhaustion, or dysfunction of immune effector cells in the high-risk subgroup. Meanwhile, the TIMER algorithm reveals that the stromal score is higher in the high-risk subgroup, which is consistent with the results from the ESTIMATE algorithm mentioned above. Fig. 6. [128]Fig. 6 [129]Open in a new tab The proportion of immune cells between low-risk and high-risk subgroups. (A) CIBERSORT analysis. (B) Correlation between signature genes and immune cells. (C) Comparison of immune cell infiltration in the TCGA cohort via ssGSEA algorithm. P values were showed as: *p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001. Immunotherapy response prediction based on the prognostic signature We compared the immune signature between high- and low-risk subgroups. The results indicated that the expression of HLA, TGF-β, TNF, chemokines, interleukins, interferons, and immunotherapeutic response in high-risk subgroups was higher than in low-risk subgroups (Fig. S14A). Immune checkpoint analysis was employed to investigate the immunotherapy response. Our study analyzed 48 immune checkpoints and detected that 36 were significantly regulated in the high-risk subgroups (Fig. S14B), including 33 up-regulated and 3 down-regulated immune checkpoints between the high-risk and low-risk subgroups. In common immune checkpoints such as PD1 (PDCD1), PDL1 (CD274), and CTLA4, the high-risk subgroups did not significantly differ from the low-risk subgroup. Still, the ARS was positively correlated with the expression of 5 immune checkpoints, including SELPLG, TNFSF18, CD200, NRP1, TNFSF4(R > 0.2, Fig. S15A-E), and negatively correlated with the expression levels of TNFRSF14(R = -0.24, Fig. S15F). The level of immune cell infiltration and expression of immune checkpoints were closely related to immunotherapy response. Based on the findings mentioned above, we further utilized the TIDE algorithm to assess the ARS’s ability to predict immunotherapy response in STAD. Our results revealed a significant positive correlation between ARS and TIDE scores (R = 0.4, Fig. S16), and we presented the results in a proportional bar chart to visualize that the low-risk subgroups exhibited a better response to immunotherapy compared to the high-risk subgroups (chi-squarep < 0.001, Fig. S14C) according to the TIDE. Among them, 52.1% of patients in the low-risk subgroup achieved CR/PR to immunotherapy, whereas only 22.0% achieved a response in the high-risk subgroup. The remaining 78% of patients experienced SD/PD. Meanwhile, we found that the high-risk subgroup demonstrated significantly higher TIDE scores, Dysfunction, and Exclusion scores than the low-risk subgroups (p < 0.0001, Fig. S14D). However, no significant differences were observed between the two subgroups regarding MSI scores. Tumor mutation analysis of STAD between low-risk and high-risk subgroups Genomic sequencing provided researchers with insights into the somatic mutation landscape, offering a detailed understanding of the mutational processes and genes driving cancer^[130]32. Gene mutation profiles have been created for high-risk and low-risk STAD subgroups. Waterfall plots were employed to depict the types and frequencies of somatic mutations in the high- and low-risk subgroups, respectively. Furthermore, we analyzed the gene mutation landscape in the high-risk and low-risk subgroups and examined the relationship between risk scoring and the gene mutation profiles. The top 15 frequently mutated genes in the high-risk and low-risk subgroups were presented in Fig. S17A-B. Both subgroups identified TTN, TP53, MUC16, LRP1B, ARID1A, CSMD3, SYNE1, FAT4, CSMD1, and FLG as commonly mutated genes. Among these ten commonly mutated genes, the mutation rates significantly decreased in all genes in the high-risk subgroup. Furthermore, evidence suggests that TMB and microsatellite instability (MSI) could be underlying predictive biomarkers for various applications, including immunotherapeutic response in malignant tumors^[131]33–[132]36. Our data revealed that the ARS exhibited a significant negative correlation with the MSI score (R = -0.26) and TMB (R = -0.38) in STAD (Fig. S17C-D). Additionally, the MSI score and TMB in the high-risk subgroup were significantly lower than those in the low-risk subgroup (Fig. S17E-F). As expected, we visualized the proportion of low- and high-risk subgroups in MSS, MSI, Low-TMB, and High-TMB by proportional bar charts (chi-square p < 0.001, Fig. S17G-H). Moreover, to investigate the relationship between TMB and survival, we divided the TCGA cohort into High-TMB and Low-TMB subgroups based on the median TMB value and performed survival analysis. Interestingly, individuals with higher TMB tended to have better overall survival than those with lower TMB, with a statistically significant difference (p = 0.002, Fig. S17I). Moreover, in the High-TMB subgroup, the low-risk subgroup had better overall survival than the high-risk subgroup (p < 0.0001, Fig. S17J). Similar results were obtained in the Low-TMB subgroup but without statistically significant differences (p = 0.609, Fig. S17K). These results suggest that the combination of ARS and TMB is a valuable prognostic biomarker in STAD. Drug sensitivity analysis To further investigate the relationship between the risk model and the sensitivity to chemotherapy and targeted drugs, we compared the sensitivity of STAD patients in different risk groups to common chemotherapeutic and targeted drugs according to the GDSC database. The results showed that the estimated IC50 values of the four chemotherapeutic agents (Docetaxel, Cisplatin, Doxorubicin, and 5-Fluorouracil) were significantly higher in the high-risk group than in the low-risk group (p < 0.05), which suggested that STAD patients in the low-risk group were more sensitive to chemotherapeutic agents (Fig. S18A-D). We also investigated the drug sensitivity analysis of the two subgroups in terms of targeted therapy. The estimated IC50 values of two targeted drugs (Lapatinib and Sunitinib) were significantly higher in the high-risk subgroup compared to the low-risk subgroup (p < 0.05, Fig. S18E-F). Single-cell analysis We performed single-cell analysis to explore the potential molecular mechanisms behind the role of prognostic genes. A total of 18 cell types were identified by single-cell sequencing data annotation. t-SNE plots (Fig. [133]7A) depicted the cell subtypes in the STAD samples, and t-SNE plots showed the prognostic genes localization (Fig. [134]7B-H). The results showed that epithelial, endothelial, fibroblast, mural, and pericyte cells were the major cell subtypes with significant prognostic gene expression. We also investigated cellular communication and demonstrated the degree of correlation and strength of signaling incomes and outcomes between these cell subtypes (Fig. [135]7I). According to the analysis above, our attention was focused on stromal and immune cells in the TME, so we selected two vital signaling pathways (CXCL and CCL) for cell communication network analysis (Fig. [136]7J). The results showed that endothelial, fibroblast and the communication between them were the core factors that drove tumor gain the anoikis resistance. Fig. 7. [137]Fig. 7 [138]Open in a new tab Atlas of single-cell sequencing in STAD. (A) The t-SNE plot displayed cell subtypes in STAD. The subcellular localization of prognostic genes, including CXCL1 (B), NOX4 (C), SERPINE1 (D), CRABP2 (E), TRIM50 (F), FRZB (G), and PLG (H). (I) Incoming and outcoming signaling patterns of cellular interaction. (J) The cellular interaction between cell subtypes in STAD is based on CCL and CXCL signaling pathways. Knockdown of CRABP2 inhibits the malignant biological behavior First, we identified the mRNA expression levels of prognostic genes, and the results showed that the expression of CRABP2 was higher in AGS and HGC-27 cells compared to GES-1 (Fig. [139]8A). The expression of other prognostic genes is basically consistent with the results of bioinformatics analysis (Fig. S19A-G). A similar result was presented by western blot (Fig. [140]8B). The original images were shown in Fig. S20. Therefore, we knocked down CRABP2 in AGS and HGC-27 cells to explore the role of CRABP2 in STAD. We validated the knockdown level of CRABP2 by western blot (Fig. [141]8C). CCK8, Wound healing and Transwell migration assay showed that CRABP2 knockdown reduced the proliferation and migration ability of STAD cells (Fig. [142]8D-G). In addition, we investigated the effect of CRABP2 on the apoptosis of STAD. Compared to the NC, CRABP2 knockdown dramatically increased the percentage of apoptosis in AGS and HGC-27 cells (Fig. [143]8H). These results suggested that CRABP2 knockdown might inhibit the malignant biological behavior of STAD cells and promote the apoptosis of STAD cells. Fig. 8. [144]Fig. 8 [145]Open in a new tab The knockdown of CRABP2 inhibited the malignant biological behavior of STAD cells. (A) Identification of CRABP2 expression in GES-1 and STAD cells by RT-qPCR. (B) Identification of CRABP2 expression in GES-1 and STAD cells by western blot. (C) Detection of knockdown of CRABP2 protein expression levels in STAD cells. (D) Knockdown of CRABP2 inhibits the cloning of STAD cells. (E) Knockdown of CRABP2 suppressed the proliferation of STAD cells. (F) Knockdown of CRABP2 suppressed the migration ability of STAD cells. (G) Knockdown of CRABP2 suppressed the invasion ability of STAD cells. (H) Apoptotic detection. Data were shown as mean ± SD. n = 3, P values were showed as: *p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001. Discussion Currently, several biomarkers have been applied to predict the prognosis of STAD. However, most of these studies focused on a single biomarker. While previous research has identified several biomarkers for STAD prognosis, most studies focus on single markers, such as LAYN^[146]37, ITGB1^38, THBS2^39, and GGT5^40, limiting their predictive power and clinical utility. Therefore, we sought to develop a novel prognostic signature incorporating multiple ARGs and explore its potential implications for personalized treatment and understanding the mechanisms of STAD progression. We utilized TCGA and GEO databases datasets to identify DEGs between tumor and normal tissues. Subsequently, we integrated clinical prognostic data and employed univariate Cox regression, LASSO, and multivariate Cox regression to construct a prognostic signature based on 7 key ARGs: CXCL1, NOX4, SERPINE1, CRABP2, TRIM50, FRZB, and PLG. CXCL1 played an essential role in STAD progression and migration. CXCL1 is bound explicitly to CXCR2, and activation of the CXCL1-CXCR2 axis was a potential mechanism for enhancing STAD invasion^[147]41. Meanwhile, CXCL1 enhances the expression of vascular endothelial growth factor (VEGF). It leads to angiogenesis within the tumor via the VEGF pathway, which is associated with lousy survival in STAD^[148]42. NOX4 was a vital effector of oxidative stress. Recent studies suggested that NOX4 promoted anoikis resistance^[149]43. Furthermore, NOX4 enhanced tumor cell proliferation. Conversely, down-regulation of NOX4 inhibited cell proliferation, and knockdown of NOX4 significantly inhibited invasion, proliferation, and EMT in STAD cells^[150]44. SERPINE1 was screened as an independent prognostic biomarker of STAD in recent studies. SERPINE1 plays an essential role in developing vascular diseases, obesity, metabolic syndrome, and several types of cancer^[151]45. Notably, stromal cells in the TME were significant sources of SERPINE1, which might enhance the pro-tumorigenic effects of SERPINE1 through complex crosstalk between the tumor and the TME^[152]46. CRABP2 was significantly up-regulated in chemotherapy-resistant STAD tissues and silenced CRABP2 by RNA interference reversed oxaliplatin resistance in vivo^[153]47. PLG circulated in plasma as an inactive zymogen capable of being activated by uPA or tPA, promoting fibrinolysis and mediating cancer metastasis. PLA, a product of PLG activation by uPA and tPA^[154]48, could encourage tumor invasion through catabolizing ECM. Studies have confirmed that PLG deficiency significantly reduced the number of spontaneous lung metastases^[155]49. Moreover, PLA activated MMP, TGF-β, and VEGF, which were equally capable of mediating tumor metastasis^[156]50. TRIM50 had been suggested to be a protective factor in pancreatic cancer^[157]51, and overexpression of TRIM50 was able to reverse EMT by degrading Snail1, as well as confirmed by studies in hepatocellular carcinoma (HCC)^[158]52. In STAD, overexpression of FRZB inhibits proliferation and induces differentiation of STAD cells^[159]53, and knockdown of FRZB could enhance STAD cell proliferation by up-regulating β-catenin activity^[160]54. In our study, TRIM50 and FRZB were independent risk factors for STAD patients. Nevertheless, it was not entirely consistent with the results of the previous literature, and its underlying mechanism deserves to be further explored in the follow-up studies. It has been found that CRABP2 is highly expressed in lung cancer metastasis and suppresses anoikis through the integrin pathway. Conversely, silencing can inhibit its metastasis^[161]55,[162]56, while the role of this gene in STAD is not well understood. For further validation, we explored the role of CRABP2 in the progression of gastric cancer through in vitro experiments. The results have demonstrated that downregulation of CRABP2 could attenuate the proliferation, migration and invasion ability of STAD cells. In addition, flow cytometry analysis showed that CRABP2 could promote apoptosis of STAD cells. However, the details and steps of this apoptosis mechanism need further investigation, and we will explore the potential mechanism in our future research. To explore the potential mechanisms underlying the prognostic differences, we performed functional enrichment analysis and TME analysis. The high-risk subgroup was enriched in pathways related to extracellular matrix remodeling, cancer-associated fibroblasts (CAFs), natural killer (NK) cells, TGF-β signaling, and EMT. TME analysis revealed a positive correlation between the risk and stromal scores and a negative correlation with tumor purity. Additionally, we observed a negative correlation between ARS and immune score. To address this discrepancy, we delved deeper into the subtypes of immune cell infiltration. Immune cell infiltration analysis using CIBERSORT and other algorithms revealed a complex immune landscape in the TME. While the high-risk subgroup showed increased infiltration of CD[8]^+ T cells (Exhausted), it also exhibited higher levels of immunosuppressive cells, such as M2 macrophages, MDSCs, and Tregs, creating a “cold” tumor microenvironment that hinders effective anti-tumor immunity. Despite the higher abundance and expression of immune cells and immune-related factors in the high-risk subgroup, we observed a reverse pattern regarding the immune checkpoint pathways. There were no significant differences inPD-L1 and CTLA4 expression but differences in emerging immune checkpoints like TIGIT. This suggests potential differential response to immunotherapy between the two subgroups. Based on the correlation between feature genes and immune infiltration, we attempt to infer the sensitivity of patients in different risk subgroups to immunotherapy. We confirmed poorer responses in the high-risk subgroup, likely due to T cell exhaustion and immune evasion mechanisms within the immunosuppressive TME. To further investigate the potential mechanisms influencing survival differences, we compared the somatic mutation rates in STAD. We found that the TMB and MSI scores in the high-risk subgroup were lower compared to the low-risk subgroup, suggesting reduced sensitivity to immunotherapy. As expected, patients in the low-risk subgroup were more sensitive to immune checkpoint inhibitors (ICIs). Subsequently, we conducted subgroup analyses on TMB and MSI. We found significant survival differences between the high TMB and low TMB groups, indicating that TMB is a prognostic factor for gastric cancer patients. In contrast, MSI did not show survival differences across the entire TCGA cohort. Therefore, we further explored the predictive ability of ARS within the TMB subgroups. Survival analysis showed that in the high TMB subgroup, the low-risk subgroup had better prognosis than the high-risk subgroup. However, this difference was not observed in the low TMB subgroup. Our previous research results emphasized the significant differences in immunotherapy response between high-risk and low-risk subgroups. Notably, the potential benefit of ICIs in the high-risk subgroup is reduced. In the low TMB group, the high-risk population constituted a significant majority, which may obscure any observable differences in prognostic outcomes. The above evidence may indicate differences in immunotherapy response between the two subgroups. In clinical practice, chemotherapy remains the first-line treatment for gastric cancer. Drug sensitivity analysis shows that the low-risk subgroup has a higher sensitivity to chemotherapy and targeted therapy, highlighting the potential treatment pathways for these patients. Therefore, ARS could identify populations that benefit from chemotherapy, targeted therapy, and immunotherapy. The field of cancer research still has more questions than answers. The complex heterogeneity of the TME (spatial, temporal, and intratumoral) created an additional layer of complexity. Therefore, single-cell analysis was needed to comprehensively and visually display the mapping of various cell types in the TME. This could greatly expand our knowledge of the roles of different cell types in tumor generation and progression. From the above research, we gradually clarified the focus and goals of our following study: stromal cells such as CAFs, immune suppressor cells, and molecules such as MDSCs, Tregs, and TGF-β in the TME. Due to the complexity of spatial and temporal characteristics of the TME, we speculated that the intercellular crosstalk between stromal cells and immune cells may be a key factor leading to the current effect on the efficacy of STAD high-risk subgroups. Single-cell analysis provided insights into cellular interactions and molecular mechanisms underlying anoikis resistance. The expression of prognostic genes was concentrated in epithelial, endothelial, fibroblast, mural, and pericyte cells, suggesting their involvement in tumor progression. We also established a cell communication network focusing on the CXCL and CCL signaling pathways, which play a role in cell migration and immune cell recruitment. The analysis identified a central relationship between CAFs and endothelial cells, suggesting a potential molecular communication mechanism associated with anoikis resistance. In conclusion, our study presents a novel anoikis-related prognostic signature in STAD. The findings provide new insights into the complex interactions between anoikis resistance, TME and immune cell infiltration. The signature may help to personalize therapeutic decisions and develop new therapeutic strategies against anoikis resistance in STAD. However, our study had certain limitations: (1) All the samples involved were based on RNA sequencing and microarray data from online databases. The accuracy of the results might have been affected by data heterogeneity, platform differences, and the lack or inconsistency of clinical information. Therefore, further prospective studies with large sample sizes were needed to confirm our findings. (2) The association between the screened anoikis-related prognostic gene expression, immune cell infiltration, and the direct molecular mechanism of anoikis resistance involved in STAD must be further verified. (3) The sample sizes of our study still needed to be improved, particularly regarding the results of some subgroup analyses, which might have been due to chance. (4) These findings must be validated through future studies with larger sample sizes. Conclusion The study constructed and validated a novel prognostic signature, which could serve as a biomarker for predicting survival outcomes in STAD patients, providing new tools for personalized treatment decisions for STAD patients. In addition, the in vitro experiments indicated that knockdown of CRABP2 could the malignant biological behavior in STAD cells. These observations provide preliminary evidence of the prognostic value of ARGs and provide a theoretical basis for future research. Electronic supplementary material Below is the link to the electronic supplementary material. [163]Supplementary Material 1^ (208.7KB, xlsx) [164]Supplementary Material 2^ (3.4MB, docx) [165]Supplementary Material 3^ (84.6KB, docx) Acknowledgements