This study utilized publicly available TCGA RNA-seq data (375 tumor and 32 normal samples) as the training set, sourced from the official database (https://www.cancer.gov/tcga). This cohort included clinical data (age, gender, T stage, N stage, and M stage) along with survival information for 350 patients with GC. Additionally, the GSE66229 dataset, available through the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/gds) and sequenced by microarray on the GPL570 platform, was downloaded. A total of 300 GC tumor tissue samples were selected from this dataset to form the validation cohort. In this study, 3,569 TME-RGs were obtained from two publicly available databases: the Molecular Signatures Database (MSigDB) (http://www.gsea-msigdb.org/gsea/msigdb/index.jsp) and Reactome Pathway (https://reactome.org/PathwayBrowser/). Furthermore, 10 PA-RGs were sourced from previous literature (16), specifically PARP-1, PAEP, MIF, AIFM1, HSP70, PAAN, ARH3, RNF146, ADPRHL2, and OGG1. The flowchart for this study was shown in Supplementary Figure 1.

In the training cohort, differential expression analysis was performed using the DESeq2 package (v 1.34.0) (17), identifying differentially expressed genes (DEGs) between tumor and control samples. A stringent threshold of |log2Fold Change (FC)| > 1 and P < 0.05 was applied, and these genes were designated as DEGs1. A volcano plot was generated using the ggplot2 package (v 3.4.1) (18) to visualize the distribution of DEGs1, highlighting the top 10 upregulated and downregulated genes in GC samples based on their log2FC values. Additionally, a heatmap was created with the pheatmap package (v 1.0.12) (19) to visually represent gene expression patterns. Further analysis of the expression profiles of the 10 PA-RGs in the training cohort was conducted using the single-sample gene set enrichment analysis (ssGSEA) function from the GSVA package (v 1.42.0) (20) to calculate the enrichment scores of PA-RGs across GC samples. Samples were stratified into high and low PA-RG score groups based on the optimal cutoff point, and differences in PA-RG scores between groups were assessed (P < 0.05). Kaplan-Meier (K-M) survival curves were generated using the survminer package (v 0.4.9) (21) to compare overall survival (OS) between the two groups (P < 0.05). To further explore the molecular features associated with PA-RGs scores, differential expression analysis was repeated in the training cohort to identify DEGs between the PA-RG score groups, using the same criteria of |log2FC| > 1.2 and P < 0.05, and these genes were designated as DEGs2. Volcano plots and heatmaps were again used to visually represent these DEGs2.

Following the aforementioned analyses, the intersection of DEGs1, DEGs2, and 3,569 TME-RGs was utilized to identify candidate genes associated with both TME and PA in GC. The VennDiagram package (v 1.7.1) (22) was employed to visualize these candidate genes. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were then performed to explore the biological functions and pathways of the candidate genes. Enrichment was considered significant for P < 0.05, with analyses conducted via the clusterProfiler package (v 4.6.0) (23). To effectively present the results, the GOplot package (v 1.0.2) (24) was used to graphically depict the top 5 most significant GO terms from each of the 3 GO domains, as well as the top 15 KEGG pathways, ranked by their statistical significance.

Expression profiles of the candidate genes were integrated with survival data from 350 patients with GC in the training cohort. Univariate Cox regression analysis was conducted using the survival package (v 3.3-1) (21), with a threshold set at P < 0.05 and Hazard Ratio (HR) ≠ 1. Genes meeting these criteria were further tested for proportional hazards (PH) assumptions using the cox.zph function. Genes that passed the threshold of P > 0.05 were classified as survival-associated genes for subsequent analysis. A forest plot was generated using the forestplot package (v 2.0.1) (25) to visually represent the results.

Subsequently, a good predictive framework was constructed within the training and validation cohorts by integrating 10 distinct machine learning algorithms to create 101 potential algorithmic combinations. This analytical pipeline included multiple machine learning approaches, such as random survival forest (RSF), elastic net (Enet), Least Absolute Shrinkage and Selection Operator (LASSO), Ridge regression, stepwise Cox, CoxBoost, plsRcox, SuperPC, GBM, and survival-SVM. To evaluate the reliability and generalizability of the model, a rigorous 10-fold cross-validation process was implemented. Each model’s predictive accuracy was quantified using the C-index, calculated across both the training and validation cohorts. The objective was to identify algorithms consistently exhibiting a high C-index in both cohorts, indicating their robustness and predictive power. This comprehensive approach led to the identification of a set of prognostic genes for GC, along with their corresponding risk coefficients. The specific algorithm combinations and descriptions of their key parameters are provided in the Appendix.

Based on the expression levels of the prognostic genes and their respective risk coefficients, a risk model was constructed following the formula:

where Xi represents the expression level of prognostic genes, and coefi denotes the risk coefficient of the corresponding gene. Using the median risk score as the cutoff, the 350 patients with GC in the training cohort were stratified into high- and low-risk groups. The risk model was validated through multiple approaches (1): risk score distribution and survival status visualization using the survminer and ggrisk (v1.3) (26) packages (2), K-M analysis to demonstrate significant OS differences (P < 0.05), and (3) time-dependent ROC analysis (survivalROC v1.0.3) with AUC calculations at 1-, 3-, and 5-year intervals. AUC between 0.6 and 0.7 indicated that the prognostic model had certain predictive value (27). The same validation framework was applied to the validation cohort. Additionally, based on data from the training and validation sets, KM analysis for TNM staging was incorporated.

To deepen the understanding of the correlation between risk scores and clinical characteristics, and to assess the prognostic capabilities of risk scores in predicting clinical outcomes, clinical data (age, gender, T stage, N stage, and M stage) from 350 patients with GC with complete survival information in the training cohort were analyzed. Further validation of the GSE66229 dataset was performed. Distribution patterns of risk scores across different clinical subgroups in both high-risk and low-risk categories were delineated, providing insights into the heterogeneity of risk among patient populations. A comparative analysis was then performed to examine the variance in risk scores across distinct clinical subgroups (P < 0.05). This analysis identified potential disparities in the prognostic significance of risk scores across various stages and characteristics of GC.

The study integrated risk scores with clinical parameters from 350 patients with GC in the training cohort, all possessing complete survival data. A comprehensive statistical analysis was conducted using the survival package, which included (1): preliminary screening via univariate Cox regression (P < 0.05) (2), validation of PH assumptions (P > 0.05), and (3) final selection through multivariate Cox analysis (P < 0.05). This sequential approach identified good independent predictors of patient outcomes.

The identified prognostic factors were incorporated into a predictive nomogram developed with the rms package (v6.2-0) (28). This visual tool assigned weighted point values to each variable, with the total score (Total Points) allowing for the estimation of 1-, 3-, and 5-year survival probabilities. Notably, the scoring system demonstrated a positive correlation between higher total points and improved clinical prognosis.

The nomogram’s predictive performance was rigorously validated using two complementary methods (1): calibration analysis, comparing predicted versus observed survival outcomes, and (2) time-dependent ROC curve analysis (timeROC v0.4 (29)), which quantified discrimination accuracy at clinically relevant intervals. These validation procedures confirmed the model’s reliability in prognostic stratification. Decision curve analysis (DCA) was further performed to assess the clinical net benefit of the nomogram, gene risk score, and conventional clinical indicators across a threshold probability range of 0 to 0.6.

Further investigation focused on delineating the immune landscape within patients with GC. The CIBERSORT algorithm (v 1.03) (30) was applied to quantify the infiltration levels of 22 distinct immune cell types (31) across GC samples in the training cohort, with particular attention to comparing these levels between high-risk and low-risk categories (P < 0.05). To explore the relationship between immune cell infiltration and prognostic genes, the correlation function was used to assess their associations. Additionally, 47 immune checkpoints from published literature (32) were curated, and their expression levels were compared between high-risk and low-risk groups in the training cohort (P < 0.05). To evaluate potential responses to immunotherapy, TIDE scores for GC samples were calculated using the TIDE database (http://tide.dfci.harvard.edu/), and differences between risk groups were assessed (P < 0.05). This approach provided insights into the potential efficacy of immunotherapy across distinct risk stratifications.

To assess the correlation between risk categories and chemotherapy responsiveness, the pRRophetic package (v 0.5) (33) was employed to compute the IC50 values for 138 commonly used chemotherapeutic and molecular targeting agents in the training cohort. IC50 values reflect drug cytotoxicity or cell resistance, with lower IC50 values indicating greater drug efficacy. After calculating the IC50 values, comparisons were made between the higher-risk and lower-risk categories to identify significant differences (P < 0.05). The ggplot2 package was utilized to graphically display the top 5 drugs with the most pronounced differences in IC50 values through box plots.

Five GC tumor tissue samples and five adjacent tumor normal tissue samples were collected at Shaanxi Provincial People’s Hospital. All samples underwent quantitative real-time polymerase chain reaction (RT-qPCR) analysis, approved by the Medical Ethics Committee of Shaanxi Provincial People’s Hospital. Total RNA was extracted from approximately 50 mg of each tissue sample using TRIzol reagent (Ambion, USA) according to the manufacturer’s instructions. RNA concentration and purity were assessed using a NanoPhotometer N50, and samples with A260/A280 ratios between 1.8 and 2.0 were considered suitable for downstream applications. First-strand cDNA was synthesized from 2 µg of total RNA using the SureScript First-Strand cDNA Synthesis Kit (Servicebio, China) in a 20 µL reaction volume. The reverse transcription reaction was carried out under the following conditions: 25 °C for 5 min, 50 °C for 15 min, and 85 °C for 5 sec, followed by hold at 4 °C. RT-qPCR was performed using 2× Universal Blue SYBR Green qPCR Master Mix (Servicebio, China) on a CFX Connect Real-Time PCR System (Bio-Rad, USA). Each 10 µL reaction contained 3 µL of diluted cDNA, 5 µL of master mix, and 0.5 µM each of forward and reverse primers. The amplification protocol consisted of an initial denaturation at 95 °C for 1 min, followed by 40 cycles of 95 °C for 20 sec, 55 °C for 20 sec, and 72 °C for 30 sec. Melting curve analysis was performed to confirm primer specificity. Primer sequences (Supplementary Table 1) and GAPDH (internal control) were used for amplification. Relative gene expression was quantified using the 2^−ΔΔCt method (34).

GC tumor tissue samples and adjacent tumor normal tissue samples were collected at Shaanxi Provincial People’s Hospital. Paraffin sections of gastric cancer tissues and adjacent tumor normal tissues were dewaxed, followed by antigen retrieval for 15 minutes. Endogenous peroxidase was blocked with 3% hydrogen peroxide, and non-specific binding sites were blocked with normal goat serum. Subsequently, primary antibodies against CD36 (18836-1-AP, Sanying, China) or KIT (AF6153, Affinity, USA) were added, and the sections were incubated overnight at 4°C followed by incubation with HRP-conjugated secondary antibody (K5007, Dako, Denmark) at 37 °C for 30 minutes. After DAB staining and hematoxylin counterstaining, the sections were sequentially dehydrated, cleared, mounted with neutral balsam, and observed under a microscope. The experiment was repeated 3 times.

Paraffin sections of gastric cancer tissues and adjacent tumor normal tissues were dewaxed, followed by antigen retrieval for 15 minutes. Endogenous peroxidase was blocked with 3% hydrogen peroxide, and non-specific binding sites were blocked with normal goat serum. Subsequently, primary antibodies against CD3 (16-0032-81, invitrogen, China), CD4 (67786-1-Ig, Sanying, China), FOXP3 (98377, CST, China) or CD68 (ab213363, abcam, UK), CD86(bs-1035R, bioss, USA), CD206 (18704-1-AP, Sanying, China) or AKAP12(ab198895, abcam, UK), CD14(ab133335, abcam, UK) were applied, and the sections were incubated overnight at 4°C. After returning to room temperature, fluorescent secondary antibodies were added and incubated at 37°C for 30 minutes in the dark. Nuclei were stained with DAPI. Finally, the sections were mounted with an anti-fade mounting medium and imaged under a fluorescence microscope. The experiment was repeated 3 times.

Cell proliferation was measured using the CCK-8 assay kit (APExBIO, Houston, USA). The MKN45 gastric cancer cell line was kindly provided by the Shaanxi Provincial People’s Hospital. MKN-45 cells were inoculated into 96-well cell culture plates at a density of 3×10^3/well and cultured overnight at 37°C in a 5% CO2 incubator. The cells were then treated with CD36 inhibitor (SMS121, HY-163541, MCE, USA), KIT inhibitor (c-Kit-IN-5-1, HY-18302, MCE, USA).10ul of CCK8 was added to each well and cultured at 37°C for 2h, the enzyme marker (BioTek, Shanghai, China) was used to determine the absorbance (A450nm) values. Cell viability (%) was calculated using the formula: 100×(experimental absorbance)/(control absorbance). The experiment was repeated 3 times.

Cell migration was measured using the Transwell assay kit (353097, FALCON, USA). After MKN45 cells were treated with CD36/KIT inhibitors, the cells were digested with trypsin, cells were adjusted to a concentration of 3×10⁵ cells/ml. Subsequently, the cell suspension was added to Transwell inserts and cultured in a 37°C, 5% CO₂ incubator for 24 h. Following incubation, cells were fixed with 70% ice-cold ethanol for 1 h, then stained with 0.5% crystal violet staining solution at room temperature. Finally, cells were observed and photographed under a microscope, and the number of migrated cells was counted and the experiment was repeated 3 times.

Proteins were extracted from MKN45 cells treated with CD36/KIT inhibitors. The protein concentration was determined and standardized by the BCA method. Equal amounts of protein samples were separated by SDS-PAGE gel electrophoresis and then transferred onto PVDF membranes. The membranes were blocked with 5% non-fat milk on a shaker at room temperature for 2 h, followed by incubation with primary antibodies overnight at 4°C: mouse monoclonal antibody GAPDH (36KD, 60004-1-Ig, Sanying, China); rabbit polyclonal antibody CD36 (88KD, 18836-1-AP, Sanying, China); rabbit polyclonal antibody KIT (110KD, AF6153, Affinity, USA). On the next day, the membranes were washed 3 times, incubated with HRP-conjugated secondary antibodies at room temperature for 1 h, and then washed 3 times. ECL chemiluminescence kit was used for development, the gray value of the target proteins was quantitatively analyzed, and the experiment was repeated 3 times.

The secretion levels of cytokines were detected using the Human IFN-γ ELISA Kit (USEA049Hu, Yunkelong, China), Human TNF-α ELISA Kit (USEA133Hu, Yunkelong, China) and Human IL-10 ELISA Kit (USEA056Hu, Yunkelong, China). Supernatants were collected from the culture medium of MKN45 cells treated with CD36/KIT inhibitors after centrifugation, the experiment was performed following the manufacturer’s instructions of the ELISA kit. Standard substances were serially diluted and added to the pre-coated microplate together with the samples, followed by incubation at 37°C for 1 h. The liquid was discarded and the microplate was washed with washing buffer. Detection Solution A was added and incubated at 37°C for 1 h; after washing, Detection Solution B was added and incubated at room temperature for 30 min in the dark, followed by thorough washing again. Chromogenic substrate was added for color development in the dark for 20 min, and the reaction was terminated with stop solution. The absorbance was immediately measured at 450 nm using a microplate reader, and the target protein concentration in samples was calculated according to the standard curve. The experiment was repeated three times.

MKN45 cells were seeded on cell climbing sheets and cultured until adherent. After treatment with CD36/KIT inhibitors, the cells were fixed with 4% paraformaldehyde at room temperature for 15 min, permeabilized with 0.1% Triton X-100 for 5 min, and rinsed three times. The cells were blocked with normal goat serum at room temperature for 30 min, followed by the addition of primary antibodies: rabbit polyclonal antibody PARP-1 (DF7198, Affinity, USA) and rabbit polyclonal antibody AIFM1 (DF7021, Affinity, USA), and incubated overnight at 4°C. On the next day, after rinsing, secondary antibody: CY3-conjugated goat anti-rabbit IgG (BA1032, BOSTER, USA) was added, and the cells were incubated in the dark at 37°C for 1 h. The nuclei were stained with DAPI in the dark for 5 min after rinsing. After thorough rinsing, the cells were mounted with anti-fluorescence quenching mounting medium, observed and imaged under a fluorescence microscope, and the experiment was repeated 3 times.

All statistical analyses were conducted using R software (version 4.2.2; R Foundation for Statistical Computing, Vienna, Austria). Specifically, the clusterProfiler package was used for GO and KEGG enrichment analysis, the DESeq2 package for gene differential expression analysis, the glmnet package for constructing Lasso-Logistic regression model, the rms package for plotting nomogram, calibration curves, and decision curves, and the pROC package for ROC analysis. All assays were performed in triplicate, and the data are expressed as the mean ± standard deviation (SD). Statistical analysis was performed using GraphPad Prism 25.0((Version 25.0, San Diego, CA, USA).). Statistical analyses were conducted using GraphPad Prism software Differences between two groups were evaluated by Student’s t-test or the Wilcoxon rank-sum test (Mann-Whitney U test), while one-way analysis of variance (ANOVA) was applied for comparisons among multiple groups. All data were expressed as mean ± SD, where *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001 were considered statistically significant A two-sided P < 0.05 was considered statistically significant.

In the training cohort, differential expression analysis between tumor and control samples revealed 4,570 DEGs, comprising 2,208 upregulated and 2,362 downregulated genes in GC (Figures 1A, B). PA-RG scores were then calculated for GC samples, which were stratified into high and low score groups. The high PA-RG score group exhibited significantly higher scores (P < 0.0001, Figure 1C), while K-M analysis indicated worse OS in the low score group (P = 0.0054, Figure 1D), suggesting the potential of PA-RGs as prognostic markers in GC. Additionally, comparisons between PA-RG score groups revealed 1,417 DEGs (1,230 upregulated and 187 downregulated, Figures 1E, F).

By intersecting 4,570 DEGs1, 1,417 DEGs2, and 3,569 TME-RGs, 205 candidate genes associated with both TME and PA in GC were identified (Figure 2A). These 205 genes were subjected to enrichment analysis, resulting in 990 GO terms, including 878 biological processes (BP), 40 cellular components (CC), and 72 molecular functions (MF). Key biological processes involved the ERK1/2 signaling cascade, ECM remodeling, mesenchymal differentiation, mesenchyme development, collagen-rich ECM, cell adhesion-associated protein complexes, heparin binding, and glycosaminoglycan interactions, all of which contribute to GC pathogenesis (Figures 2B–D). Furthermore, the candidate genes were enriched in 36 KEGG pathways, such as arachidonic acid metabolism, ECM-receptor crosstalk, PI3K-Akt signaling, focal adhesion dynamics, PPAR signaling, and adipocyte lipolysis regulation (Figure 2E), all of which are likely involved in tumor cell proliferation, invasion, metastatic potential, and the establishment of a pro-inflammatory TME in GC.

A univariate Cox regression analysis was performed on the 205 candidate genes within the training cohort, identifying 31 genes that met the established criteria (P < 0.05 and HR ≠ 1) (Figure 3A). All 31 genes passed the PH assumption test (P > 0.05) (Supplementary Figure 2) and were thus classified as survival-associated genes for further selection in machine learning algorithms.

Subsequent analyses were conducted in both the training and validation cohorts, utilizing 10 machine learning algorithms to generate 101 algorithmic combinations, incorporating survival-associated genes for further refinement. Of particular note was the RSF + plsRcox algorithm, which achieved a C-index of 0.622 in the training cohort and 0.611 in the validation cohort, exceeding the critical threshold of 0.6 in both instances (Figure 3B). The C-index reflects the model’s ability to distinguish between patients with different prognoses. When the C-index is between 0.6 and 0.7, it indicates that the model can somewhat differentiate between patients with better and worse prognoses, although its discriminatory power is moderate (35). This suggests that the model has potential clinical application value. Given its good performance, this model was selected for generating risk coefficients and identifying the 7 prognostic genes integral to the model’s framework: EGF, PCOLCE2, CD36, ADAMTS8, CIDEC, KIT, and AKAP12.

A risk model was developed using the expression profiles of prognostic genes and their associated risk coefficients. In the training cohort, 350 samples with survival data were stratified into two distinct groups: 175 higher-risk and 175 lower-risk categories. The risk curve showed a clear increase in mortality as the risk score rose (Figures 4A, B). As expected, the higher-risk group exhibited significantly poorer OS compared to the lower-risk group (P = 0.0025) (Figure 4C). KM analysis of TNM staging in the training cohort revealed no significant differences between T stages, whereas significant differences were observed between N and M stages (N: P = 0.0012, M: P = 0.012) (Supplementary Figure 3A). The ROC curve further validated the model’s predictive accuracy, with AUC values of 0.65, 0.68, and 0.60 for 1, 3, and 5-year predictions, respectively (Figure 4D).

This predictive efficacy was corroborated in the validation cohort, which included 300 samples with survival data, divided into 150 higher-risk and 150 lower-risk categories (Figures 4E, F). Similar to the training cohort, the high-risk group was associated with worse OS (P < 0.0001) (Figure 4G). In the KM analysis of the validation cohort, significant differences were observed across TNM stages (T: P < 0.0001, N: P < 0.0001, M: P < 0.0001) (Supplementary Figure 3B). The AUC values for 1, 3, and 5-year predictions in the validation cohort were 0.63, 0.65, and 0.62, respectively (Figure 4H). These results underscore the reliability of the model and its potential clinical utility in prognostic assessment for patients with GC.

The distribution patterns of different clinical subgroups within the higher-risk and lower-risk categories were visualized in the training set, including age, gender, T stage, N stage, and M stage. In the higher-risk category, 40.5% were over 60 years old, and 59.5% were 60 or younger; 38.9% were female, and 61.1% were male; T stage distributions were 2.3% for T1, 23.3% for T2, 45.3% for T3, and 29.1% for T4; N stage distributions were 28% for N0, 28.6% for N1, 21.4% for N2, and 22% for N3; M stage distributions were 91.6% for M0 and 8.4% for M1 (Figure 5A). In the lower-risk category, 27% were over 60 years old, and 73% were 60 or younger; 32% were female, and 68% were male; T stage distributions were 6.9% for T1, 19.5% for T2, 47.7% for T3, and 25.9% for T4; N stage distributions were 32.7% for N0, 26.3% for N1, 21.1% for N2, and 19.9% for N3; M stage distributions were 94.7% for M0 and 5.3% for M1 (Figure 5B). These findings suggest that older male patients may be at higher risk of disease. In terms of TNM staging, the higher-risk category had a larger proportion in later T stages (T3 and T4) and N stages (N2 and N3), while the lower-risk category had higher proportions in earlier T stages (T1 and T2) and N stages (N0 and N1). The higher-risk category also had a greater proportion of distant metastasis (M1), indicating a strong association between distant metastasis and increased disease risk. These results highlight that TNM staging is closely related to disease risk, prompting further investigation of risk score variations across these clinical subtypes. Notably, significant differences in risk scores were observed based on age and T stage (P < 0.05) (Figure 5C). In the validation set, clinical subgroup analyses were also conducted for high- and low-risk groups. Results indicated consistent age and gender distributions across both risk groups, with similar TNM staging distributions. Compared to the training set, although T1 stage was absent in T staging, outcomes for other stages largely aligned with those observed in the training set (Supplementary Figure 4A). This further validated the conclusions drawn from the training set. Additionally, significant differences in the risk scores were observed based on age, T stage, and N stage (P < 0.05) (Supplementary Figure 4B). These observations suggest that risk scores correlate with clinical characteristics and cancer staging in patients with GC, with more advanced stages likely associated with higher risk scores. Patients in the high-risk category may require more intensive monitoring and potentially more aggressive treatment, while those in the low-risk category may have a better prognosis.

A series of analyses were performed by integrating risk scores with clinical characteristics of patients with GC, ultimately identifying risk scores, age, N stage, and M stage as independent prognostic factors for GC (P < 0.05) (Figures 6A, B; Supplementary Figure 5). To leverage the prognostic implications of these factors, a comprehensive nomogram was developed to refine clinical prognostication for patients with GC (Figure 6C). This tool calculated the 1-, 3-, and 5-year survival probabilities for patients with GC, assigning a weight to each factor’s contribution to the outcome. Calibration curves demonstrated a strong alignment between predicted survival probabilities and the reference line, highlighting the high predictive accuracy of the nomogram (Figure 6D). ROC curves further confirmed the nomogram’s superior predictive performance, with AUC values of 0.72, 0.75, and 0.71 for 1-, 3-, and 5-year predictions, respectively (Figure 6E), these values notably outperformed those obtained using risk scores alone. In addition, the DCA results demonstrated that the nomogram established in this study yielded a sustained positive net benefit within the clinically common threshold range of 0 to 0.6, and its clinical decision-making value was significantly superior to that of the single genetic risk score and TNM staging-related indicators (Supplementary Figure 6), further validating the favorable predictive performance of this nomogram.

This study further evaluated the infiltration levels of 22 distinct immune cell types across the two risk categories of patients with GC (Figure 7A). A significant difference in the abundance of 11 immune cell types was observed between the groups (P < 0.05). Notably, memory B cells, naive B cells, M2 macrophages, resting mast cells, and monocytes were more abundant in the higher-risk category (P < 0.05), suggesting their potential role in facilitating tumor progression and suppressing immune responses. In contrast, M0 macrophages, activated mast cells, resting NK cells, activated memory CD4 T cells, T follicular helper cells, and regulatory T cells (Tregs) showed significantly reduced abundance in the higher-risk category (P < 0.05) (Figure 7B). Also, we selected gastric cancer tissues and their paired adjacent normal tissues to perform immunofluorescence staining for immune cells. The results showed that the positive expression levels of CD3⁺ T cells and CD3⁺CD4⁺ T helper cells in adjacent normal tissues were significantly higher than those in gastric cancer tissues (Figure 8A, C, D), suggesting that the number of total T cells and T helper cells in adjacent normal tissues was markedly greater than that in gastric cancer tissues. The proportion of CD68⁺CD86⁺ M1 macrophages in adjacent normal tissues was significantly higher than that in gastric cancer tissues (Figures 8B, E), while the proportion of CD68⁺CD206⁺ M2 macrophages was significantly lower (Figures 8B, F), indicating that adjacent normal tissues were enriched in M1-type macrophages with a markedly reduced proportion of M2 macrophages. Furthermore, the proportion of CD14⁺ monocytes in gastric cancer tissues was significantly higher than that in adjacent normal tissues (Figures 8G, H).

Furthermore, a correlation between prognostic genes and immune cells was observed; specifically, AKAP12 exhibited a significant positive correlation with monocytes, with a correlation coefficient greater than 0.3 (P < 0.05) (Figure 7C). Immunofluorescent co-staining assays for AKAP12 and CD14 (a surface marker of monocytes) showed that the number of AKAP12/CD14 double-positive stained cells was significantly higher in cancer tissues than in adjacent normal tissues. This experimental finding directly corroborated the existence of a correlation between AKAP12 and monocytes at the tissue level (Figures 8G, I).

Subsequent analysis of immune checkpoint expression profiles in higher-risk and lower-risk GC categories revealed significant differences in the expression levels of 31 unique immune checkpoints (P < 0.05) (Figure 7D). These differences in immune checkpoint expression may indicate a relationship between disease severity and the modulation of immune responses.

Finally, TIDE scores were assessed for the higher-risk and lower-risk categories of patients with GC. The TIDE model predicts whether a tumor can resist current immunotherapies, such as checkpoint inhibitors, through immune evasion mechanisms. Consequently, a high TIDE score typically indicates a tumor with greater immune evasion capacity (36). TIDE scores were significantly higher in the higher-risk category (P < 0.001), suggesting that patients with a higher risk may have a poorer response to immunotherapy. Correlation analysis revealed a significant positive correlation between risk scores and TIDE scores (cor = 0.29, P < 0.0001) (Figure 7E), indicating that as the risk score increases, the tumor’s immune evasion capability is enhanced, potentially reducing the effectiveness of immunotherapy.

Further analysis of sensitivity to 138 commonly used chemotherapeutic and molecular targeting agents revealed significant IC50 differences between the higher-risk and lower-risk categories for 116 of these drugs (P < 0.05) (Supplementary Table 2). Imatinib and PLX4720 exhibited lower IC50 values in the higher-risk category (P < 0.0001), suggesting enhanced responsiveness to these agents in patients with higher risk profiles. In contrast, GW.441756, BAY.61.3606, and JNK.9L showed higher IC50 values in the higher-risk category (P < 0.0001), indicating a reduced therapeutic response to these drugs in patients at greater risk of disease progression (Figure 7F).

Expression validation analysis demonstrated significant differences in the expression of prognostic genes (PCOLCE2, CD36, ADAMTS8, KIT, and AKAP12) between control and GC groups. The expression of PCOLCE2, CD36, ADAMTS8, and KIT was notably elevated in the control group compared to the GC group, while AKAP12 exhibited higher expression in the GC group (Supplementary Figure 7A). All five prognostic genes showed statistically significant differences between the two groups (P < 0.05).We further selected CD36 and KIT for protein-level validation via immunohistochemical (IHC) assays. The results showed that the protein expression levels of both CD36 and KIT in gastric cancer tissues were significantly higher than those in the control group (Figures 9A, B), which was contrary to the findings of qPCR assays. Based on the above protein-level detection results, we performed subsequent functional validation using CD36 and KIT inhibitors in gastric cancer MKN45 cells, confirming that the two inhibitors could downregulate the protein expression of CD36 and KIT in MKN45 cells, respectively (Supplementary Figure 8). Furthermore, the downregulation of CD36/KIT protein expression significantly inhibited the migration (Figures 9C, D) and proliferation (Figure 9E) of MKN45 cells, and promoted their apoptosis (Figure 9F). Collectively, these results suggest that the prognosis-related genes CD36 and KIT can promote the initiation and progression of gastric cancer when highly expressed in gastric cancer cells.

To clarify the correlations between CD36, KIT and TME as well as the PA pathway, we performed ELISA-based cytokine detection and immunofluorescence staining for PARP-1 and AIFM1 in MKN45 cells following treatment with CD36 or KIT inhibitors. ELISA results showed that the secretion levels of TNF-α and IFN-γ in the CD36 and KIT inhibitor groups were significantly higher than those in the control group, while the secretion level of IL-10 was significantly lower (Figures 10A–C). Immunofluorescence staining results revealed that PARP-1 was predominantly localized in the cell nucleus with a uniform diffuse distribution in gastric cancer MKN45 cells of the control group. After separate treatment with CD36 or KIT inhibitors, the nuclear fluorescence intensity of PARP-1 was significantly increased, and its distribution pattern was converted to a typical punctate aggregation associated with DNA damage foci (Figures 10D, F). Meanwhile, the ratio of nuclear to total fluorescence intensity of AIFM1 was markedly elevated (Figures 10E, G).