mTranscriptomic profiles (32 normal and 375 tumor samples), along with somatic mutation and clinical annotation data for stomach adenocarcinoma (STAD), were obtained from TCGA repository. To reduce potential bias from non–cancer-related deaths or perioperative complications, patients with an overall survival of less than 30 days were excluded, resulting in 335 eligible cases for downstream analyses. The Gene Expression Omnibus (GEO) dataset GSE26901 (n=109) served as an external validation cohort. Single-cell transcriptomic profiles were retrieved from the GSE183904 dataset. A total of 59 genes associated with polyamine metabolism (PMRGs) were selected based on the REACTOME_METABOLISM_OF_POLYAMINES pathway defined in the MSigDB database.

The GSE183904 dataset was analyzed using the Seurat package, with the top 2,000 highly variable genes selected through the application of the FindVariableFeatures() function. Data integration and batch correction were performed using Harmony R package on the top 20 principal components with default parameters. UMAP reduction was applied using Harmony-corrected embeddings with dims = 1:20 and default settings. Cells with >250 genes, <20% mitochondrial content, and ≥5 reads per gene were retained, yielding 73,846 cells. Data were normalized and scaled; PCA was performed followed by clustering at resolution 0.3. Annotation was manually performed using canonical cell markers. Scissor R package identified cells associated with high- and low-risk patients by integrating single-cell expression profiles, bulk TCGA expression, and risk-group labels. AUCell was employed to quantify the activity of PMRGs at the single-cell level, and cells were subsequently stratified into high- and low-score groups according to the median module activity. Differential PMRGs expression was assessed using FindAllMarkers(). For CellChat analysis, interactions were computed using computeCommunProb() with min.cells = 10, and interactions with a communication probability below 0.05 were filtered out. Single-cell pseudotime trajectories were inferred via Monocle2 to explore differentiation dynamics.

A total of 408 PMRGs-derived differentially expressed genes (PMRDEGs) were subjected to prognostic feature selection across 10 machine learning methods and 101 model combinations, including Support Vector Machine (SVM), Least Absolute Shrinkage and Selection Operator (LASSO), Gradient Boosting Machine (GBM), Random Forest (RF), Elastic Net, Stepwise Cox Proportional Hazards Regression, Ridge Regression, CoxBoost Algorithm, Supervised Principal Components (SuperPC), and Partial Least Squares Regression for Cox Model (plsRcox). Model development was implemented using the ML.Dev.Prog.Sig() function from the Mime1 R package. The Cancer Genome Atlas - Stomach Adenocarcinoma (TCGA-STAD) cohort served as the training dataset, while GSE26901 was used for external validation. Internal validation within the training set was performed via 10-fold cross-validation to evaluate model stability and prevent overfitting. The random seed was set to 5201314 to ensure reproducibility. Candidate prognostic genes were first screened by univariate Cox regression (p < 0.05), and model selection was based on the highest Harrell’s concordance index (C-index) across all algorithmic combinations. Based on the median risk score, patients were stratified into high- and low-risk subgroups. Kaplan-Meier survival analysis was conducted to assess prognostic differences, while the model’s predictive accuracy was validated through calibration plots and time-dependent receiver operating characteristic (ROC) curve analysis.

To determine the independent prognostic significance of the risk score, both univariate and multivariate Cox regression models were applied, integrating clinical features. A nomogram predicting 3- and 5-year overall survival probabilities was developed using the rms R package. Its performance was assessed through calibration curve consistency and decision curve analysis (DCA) to evaluate clinical utility.

Genes differentially expressed between high- and low-risk cohorts (threshold: |log₂FC| > 1, adjusted p < 0.05) were identified via the limma package. Gene Set Enrichment Analysis (GSEA) was subsequently conducted using both the standalone GSEA tool (v4.3.3) and the clusterProfiler R package to explore enriched biological processes (GO) and pathways (KEGG).

Immune infiltration was evaluated using ssGSEA and CIBERSORT algorithms. Immune checkpoint gene expression, Immunophenoscore (IPS) scores from The Cancer Immunome Atlas (TCIA), and Tumor Immune Dysfunction and Exclusion (TIDE) scores were compared between risk groups. Immunotherapy response was predicted in the IMvigor210 anti-PD-L1 cohort by categorizing responders (R) and nonresponders (NR) relative to risk scores.

STAD mutation data from TCGA were used to calculate tumor mutational burden (TMB), comparing high- and low-risk groups via Wilcoxon test. The top 20 mutated genes were visualized using GenVisR.

For drug sensitivity analysis, gene-drug interaction data were obtained from the DGIdb database. Correlations between model gene expression and compound activity (IC50) were evaluated using CellMiner, which integrates NCI-60 pharmacogenomic profiles. Gene expression and drug sensitivity data were log-transformed, normalized, and Spearman correlations were computed. To further predict chemotherapeutic responses, the pRRophetic R package was employed to estimate drug IC50 values. The model utilized ridge regression trained on the GDSC pharmacogenomic dataset, assuming consistent gene-drug associations between GDSC and TCGA-STAD transcriptomic profiles. Batch effects were corrected using the “Empirical Bayes” method, and the average expression of duplicated genes was computed using the avereps() function. Patients were stratified into high- and low-risk groups to compare predicted IC50 values.

This study employed normal gastric epithelial cells (GES-1) and human gastric cancer cells (AGS) for experimentation. Under strict aseptic conditions, both cell lines were cultured in RPMI-1640 medium supplemented with 10% fetal bovine serum. Total RNA was extracted from the cells using TRIzol reagent, and cDNA was synthesized via reverse transcription with PrimeScript™ RT Master Mix. RNA and cDNA concentrations were quantified using a NanoDrop 2000 spectrophotometer. RT-qPCR was performed on a QuantStudio 5 real-time fluorescence quantitative PCR system, with primer sequences detailed in Table 1. GAPDH served as the internal reference gene, and relative expression levels were calculated using the 2^(-ΔΔCT) method.

All statistical analyses were conducted in R (v.4.1.0) unless otherwise noted. P-value < 0.05 was considered significant. Survival curves were compared via log-rank test, and other comparisons used Wilcoxon or appropriate tests.

A total of 73,846 cells were obtained from GC patients in the GSE183904 dataset. Following batch effect removal, normalization, dimensionality reduction, and clustering, cells were annotated into 11 distinct types: T cells, B cells, endothelial cells, epithelial cells, fibroblasts, macrophages, mast cells, monocytes, pericytes, plasma cells, and cancer cells (Figures 1A, B; Supplementary Figure 1). Representative marker genes for each cell type were visualized using a bubble plot (Figure 1C). To further validate the accuracy of cell-type annotations, UMAP plots showing the expression of representative marker genes were generated for each cell population and provided in Supplementary Figure 2. Using a curated set of 59 polyamine metabolism-related genes, AUCell scoring stratified cells into high- and low-scoring groups (Figures 1D, E). Differentially expressed genes between these two groups were identified as PMRDEGs.

Cell-type composition was assessed across individual samples (Figure 1F), followed by functional enrichment analyses of differentially expressed genes (DEGs) identified within each cell population. GO and KEGG pathway analyses indicated that DEGs from T cells were predominantly enriched in T cell-associated signaling cascades—such as the T cell receptor signaling pathway—providing additional support for the accuracy of our cell-type annotations (Supplementary Figure 3). Hallmark analysis revealed enrichment of epithelial-mesenchymal transition signatures in fibroblasts and pericytes, while monocytes and macrophages showed enrichment in inflammatory response and TNFα signaling via NF-κB pathways (Figure 1G).

To construct a prognostic model, we evaluated the expression profiles of 408 PMRDEGs using 10 machine learning algorithms, including CoxBoost, Enet, GBM, Lasso, plsRcox, Ridge, RSF, stepwise Cox, SuperPC, and survival-SVM, as implemented in the Mime1 R package. A 10-fold cross-validation framework was used within the TCGA training cohort to assess model robustness and minimize overfitting, and the best-performing model was selected based on the highest C-index. (Figure 2A). Univariate Cox regression identified 17 prognostically relevant PMRDEGs (Figure 2B), which were further optimized using a combination of stepwise Cox (forward) selection and the Enet model (α = 0.1). 13 genes (ANXA5, CD59, CXCR4, SLC2A3, ZFP36, PER1, RGS2, ARGLU1, CDK5RAP3, GLA, DNM2, TAP1, TCIRG1) were ultimately selected to construct the final model (Figure 2C). Survival analysis revealed that high-risk patients exhibited significantly worse outcomes in both cohorts (Figure 2D). The time-dependent AUCs for 1-, 3-, and 5-year survival were 0.67, 0.69, and 0.70 in TCGA, and 0.64, 0.67, and 0.66 in GSE26901, respectively (Figure 2E). Risk score distribution, survival status, and heatmaps of gene expression levels consistently distinguished high- from low-risk groups (Figures 2F, G). PCA confirmed clear separation between the two risk groups (Figures 2H, I). Additionally, expression levels of the 11 genes significantly differed between tumor and normal tissues, demonstrating their diagnostic potential (Figure 2J), RT-qPCR experiments further confirmed these analytical results (Figure 2K). To further validate the clinical utility of our model, we compared its predictive performance with ten previously published gastric cancer prognostic models. As shown in Supplementary Figure 4, our model achieved higher time-dependent AUC values at 1-, 3-, and 5-year survival, demonstrating superior prognostic accuracy and robustness.

Univariate and multivariate Cox regression analyses indicated that both risk score and age were independent prognostic factors in gastric cancer (Figures 3A, B). A nomogram integrating clinical features and risk score was constructed to predict individual survival probabilities (Figure 3C). Calibration curves across 1-, 3-, and 5-year temporal points revealed superb alignment of nomogram-predicted probabilities with actual observations (Figure 3D). Decision curve and clinical impact curve analyses consistently demonstrated favorable net benefit and predictive utility across a range of threshold probabilities (Figure 3E; Supplementary Figure 5). Furthermore, subgroup survival analyses showed that the risk model retained significant prognostic value in stage I–III patients but not in stage IV, suggesting that its predictive power is more pronounced in earlier disease stages (Supplementary Figure 6). Notably, patients with the genomically stable (GS) subtype exhibited significantly higher risk scores (Figure 3F). Single-cell analysis revealed that GS subtype patients had a higher proportion of cancer and mast cells, consistent with their poor prognosis (Supplementary Figure 7).

SCISSOR analysis was used to associate specific cell populations with patient risk stratification (Figure 4A). CellChat analysis showed that high-risk cells exhibited a greater number and intensity of intercellular interactions compared to low-risk cells (Figures 4B–D; Supplementary Figures 8, 9). Interaction heatmaps emonstrated that fibroblasts and macrophages in the high-risk group had significantly enhanced signaling activity (Figure 4E). Outgoing signals were primarily driven by cancer cells in the low-risk group and by fibroblasts and macrophages in the high-risk group, whereas T cells exhibited strong incoming signals in both groups (Figure 4F). Signaling pathway comparison revealed that high-risk groups were enriched in SPP1, TNF, and ANNEXIN pathways, while low-risk groups showed enrichment in PTN, ncWNT, and AGT pathways (Figure 4G). Key ligand-receptor interactions were further statistically validated to confirm their robustness (Supplementary Tables 1, 2).

Transcriptomic profiling identified 139 genes with significant upregulation in the high-risk cohort (|log₂FC| > 1, adjusted p < 0.05), indicative of an aggressive molecular phenotype (Figure 5A). GO enrichment highlighted terms related to extracellular matrix organization, while KEGG pathways included TGF-beta signaling, focal adhesion, and proteoglycans in cancer (Figures 5B, C). GSEA further revealed that high-risk groups were enriched in vascular smooth muscle contraction and ECM receptor interaction, whereas low-risk groups showed enrichment in spliceosome, base excision repair, and RNA degradation pathways (Figures 5D, E). These results demonstrate that high-risk tumors are characterized by extracellular matrix remodeling and stromal activation, suggesting dysregulated polyamine metabolism drives tumor-promoting microenvironments. Conversely, low-risk tumors exhibit enhanced nucleic acid maintenance mechanisms, indicating preserved genomic integrity.

ssGSEA quantification demonstrated augmented immune effector activity and intensified leukocyte infiltration within high-risk cohorts (Figure 6A). Specifically, the high-risk stratum exhibited markedly elevated enrichment scores for B lymphocytes, cytotoxic CD8+ T cells, dendritic cells, macrophages, mast cells, neutrophils, and regulatory T cells (Tregs), whereas major histocompatibility complex class I (MHC-I) molecules showed preferential expression in low-risk counterparts (Figures 6B, C). ESTIMATE computational deconvolution revealed augmented stromal/immune compartment abundance and composite microenvironment scores, concurrent with diminished tumor purity in high-risk cohorts (Figure 6D). CIBERSORT analysis indicated that activated NK cells, resting dendritic cells, monocytes, and resting mast cells were more prevalent in the high-risk group, whereas activated CD4+ memory T cells, M0 macrophages, resting NK cells, and follicular helper T cells were higher in the low-risk group (Figure 6E). Dysregulated overexpression of immune checkpoint molecules predominated in high-risk cohorts (Figure 6F). Tumor IPS quantification from TCIA demonstrated significantly enhanced scores in low-risk patients, correlating with improved immunotherapy responsiveness (Figure 6G). Complementing these findings, TIDE algorithms revealed substantially elevated immune evasion potential in high-risk versus low-risk strata (Figure 6H). Building upon immune infiltration profiling, we assessed anti-PD-L1 therapeutic efficacy in the IMvigor210 cohort (n=348). Treatment responses were categorized into four clinically defined groups: progressive disease (PD), stable disease (SD), partial response (PR), and complete response (CR). For analysis, SD and PD were grouped as NR, while CR and PR were grouped as R. The proportion of responders was significantly higher in the low-risk group, and responder patients exhibited significantly lower risk scores compared to non-responders (Figures 6I, J).

Based on ssGSEA using T cell markers from a previously published study (16), we observed significantly higher T cell infiltration in high-risk patients from the TCGA cohort (Figure 7A). To further characterize T cell dynamics, we re-annotated 27,703 single T cells into CD4+ and CD8+ subsets (Figures 7B, C), and confirmed marker expression at the single-cell level (Figure 7D). In CD8+ T cells, early-stage markers such as CCR7 and TCF7 were highly expressed at the beginning of the trajectory, while genes associated with effector function and exhaustion like GZMB, PRF1, PDCD1 and TOX showed increasing expression toward the terminal states (Figures 7E, F). Pseudotime distributions confirmed a progression from CD8+ TN (naïve) cells to CD8+ Tex (exhausted) cells, with Tex cells occupying more advanced positions along the trajectory (Figure 7G). A similar pattern was observed in CD4+ T cells. CCR7 was predominantly expressed in early-stage CD4+ TN cells, while FOXP3, a hallmark of regulatory T cells (Tregs), was enriched in cells at later pseudotime stages (Figures 7H–J), reflecting a transition toward immunosuppressive phenotypes. These results highlight the continuous differentiation of tumor-infiltrating T cells in gastric cancer, offering insights into their functional states within the immune microenvironment.

Analysis of TMB demonstrated that the low-risk cohort exhibited a significantly elevated total mutational burden. (Figure 8A, Supplementary Figure 10). The top nine significantly mutated genes, including MKI67, CDH6 and GOLGB1 were more frequently mutated in low-risk patients (Figure 8B). Potential drug-gene interactions involving the identified signature genes were explored using DGIdb (Figure 8C). Furthermore, analysis using the CellMiner database revealed several negative correlations between key genes and drug sensitivities. ANXA5 and CD59 negatively correlated with Oxaliplatin and Docetaxel; CD59 also correlated with Fluorouracil, and DNM2 with Epirubicin (Figure 8D). Consistently, drug-response prediction using pRRophetic estimated lower IC50 values for 5-Fluorouracil, Docetaxel, Doxorubicin, and Paclitaxel in the low-risk group (Figure 8E), indicating a potential increased chemosensitivity in this cohort. These computational inferences derive from established in vitro pharmacogenomic datasets and thus represent hypotheses about differential drug response that warrant follow-up experimental and clinical validation to determine their translational relevance.