Acrylamide-associated gastric cancer progression involves RRM2-related nucleotide metabolic dysregulation and immune evasion Authors: Peiru Si, Tianlong Zhang, Xuefeng Lv Published: 2026 ABSTRACT Highlights•RRM2 identified as acrylamide-associated hub gene in gastric cancer.•Multi-cohort integration reveals nucleotide metabolism dysregulation.•Machine learning prognostic model stratifies GC risk groups.•Single-cell analysis uncovers immune evasion signaling networks.•Experimental validation confirms RRM2 oncogenic roles in GC. Introduction Acrylamide (AA), an environmental contaminant and exposure factor, is classified as a Group 2A probable human carcinogen by the International Agency for Research on Cancer (IARC) [1]. AA is formed during high-temperature processing of certain foods (e.g., French fries, potato chips, and baked goods) via the Maillard reaction [2]. Epidemiological evidence has examined associations between long-term dietary AA exposure and several cancers, including renal, endometrial, and ovarian cancers [3,4]. Mechanistically, orally ingested parent AA may directly contact the gastric epithelium before absorption, whereas absorbed AA is metabolized predominantly by cytochrome P450 2E1 (CYP2E1) to glycidamide, an epoxide that forms DNA adducts and induces genotoxic damage [5,6]. Because CYP2E1 abundance and metabolic capacity vary across gastrointestinal and hepatic tissues, the relative contributions of local parent-AA toxicity and systemic or locally generated glycidamide cannot be assumed to be equivalent. Accordingly, we hypothesize that gastric carcinogenic stress may reflect two non-exclusive routes: direct luminal exposure of gastric epithelial cells to parent AA and genotoxic exposure to glycidamide formed after absorption or by local CYP2E1 activity. The present study does not quantify tissue-specific AA/glycidamide exposure or CYP2E1 activity; therefore, it evaluates AA-associated molecular networks rather than assigning causality to either exposure route. Gastric cancer (GC) remains a major global health burden and a leading cause of cancer-related death [7,8]. Its development reflects genetic susceptibility, chronic infection such as Helicobacter pylori, environmental exposure, and metabolic reprogramming. Nucleotide metabolism is especially relevant because it supports DNA and RNA synthesis. It also influences epigenetic regulation, redox balance, and energy metabolism [9]. In GC, enzymes in purine and pyrimidine pathways are often overexpressed and may support rapid proliferation and resistance to apoptosis [[10], [11], [12]]. Environmental exposure to AA has been reported to disrupt several cellular metabolic pathways, although much of the available mechanistic evidence derives from hepatic, renal, or neural models rather than gastric epithelium. In vivo studies have described decreased ATP levels, adenylate-pool disruption, and inhibition of mitochondrial respiratory-chain complexes in liver and kidney tissues [13,14], while neural-cell studies indicate interference with glycolysis and, indirectly, nucleotide-precursor supply [15]. These observations provide biological context but cannot be directly extrapolated to the gastric mucosa because local parent-compound exposure and CYP2E1-dependent glycidamide formation differ by tissue. Separately, AA/glycidamide-induced DNA damage may activate the ATM/ATR-Chk1 pathway and increase nucleotide demand for repair [16]. Although AA-related genotoxicity has been widely documented, its relationship to nucleotide metabolic dysregulation in GC remains poorly understood. Two mechanistically distinct possibilities should be distinguished. First, as an electrophile, parent AA or glycidamide could directly modify nucleophilic protein residues and thereby alter enzyme activity. Second, chronic DNA damage, oxidative stress, and checkpoint signaling could indirectly change the transcription or abundance of nucleotide-metabolic enzymes. The network and transcriptomic analyses used here primarily assess the second, downstream regulatory scenario, whereas the docking analysis only explores a possible non-covalent pose and cannot establish direct biochemical modification. This study therefore aimed to identify AA-associated nucleotide-metabolic candidates in GC and to evaluate the functional and clinical relevance of the prioritized candidate, without presuming that AA directly covalently modifies that protein [17]. Materials and methods Data acquisition and processing Transcriptomic data and associated clinical information for stomach adenocarcinoma (STAD) were obtained from The Cancer Genome Atlas (TCGA-STAD, comprising 375 tumor cases and 32 normal controls, RNA-seq data, HTSeq-FPKM expression unit) and the Gene Expression Omnibus (GEO) datasets: GSE15459 (n = 200, platform GPL570, Affymetrix HG-U133 Plus 2.0 microarray), GSE62254 (n = 300, platform GPL570, Affymetrix HG-U133 Plus 2.0 microarray), GSE84433 (n = 357, platform GPL6947, Illumina HumanHT-12 V3.0 microarray), GSE26899 (n = 108, platform GPL570, Affymetrix HG-U133 Plus 2.0 microarray), GSE26253 (n = 432, platform GPL8432, Illumina HumanRef-8 WG-DASL v3.0 microarray), and GSE27342 (n = 160, platform GPL5175, Affymetrix Human Exon 1.0 ST Array). Each dataset was processed individually according to its platform-specific requirements. For Affymetrix datasets, raw CEL files were processed using the robust multi-array average (RMA) algorithm for background correction, quantile normalization, and log2 transformation via the affy R package. For Illumina datasets, preprocessed expression matrices were directly downloaded from GEO and further quantile-normalized. For TCGA-STAD RNA-seq data, HTSeq-FPKM values were log2(x + 1) transformed. To minimize cross-platform batch effects, each dataset was independently processed and analyzed, with individual dataset results subsequently compared for consistency rather than merging raw data across platforms. Single-cell RNA sequencing (scRNA-seq) data were acquired from the GSE163558 dataset (n = 4, comprising 3 primary gastric cancer samples and 1 normal control). Baseline information for all datasets is provided in Supplementary File 1. Genes potentially affected by acrylamide (AA) were retrieved from the Comparative Toxicogenomics Database (CTD; http://ctdbase.org/). Nucleotide metabolism-related gene sets were obtained from the MSigDB database (https://www.gsea-msigdb.org). GC-associated genes were acquired from the GeneCards database (https://www.genecards.org/, version 5.12, accessed on March 15, 2024), using a relevance score threshold of >10. This threshold was chosen to balance sensitivity and specificity: a lower cutoff would introduce excessive noise from weakly associated genes, whereas a more stringent cutoff might exclude biologically relevant candidates. An overlapping gene set was identified by intersecting these three gene lists (Supplementary File 2) for downstream analyses. Molecular docking analysis Three-dimensional structures of target proteins were retrieved from the RCSB Protein Data Bank (PDB). The molecular structure of acrylamide was obtained from the PubChem database and energy-minimized prior to docking. Molecular docking was performed using the CB-Dock2 web server, which integrates AutoDock Vina to predict standard non-covalent protein-ligand binding modes and affinities. Protein structures were prepared by removing water molecules and adding polar hydrogen atoms. Docking was carried out using default parameters with consistent grid-box settings for all targets. No covalent docking protocol, Michael-addition reaction, covalent bond constraint, or cysteine-adduct model was implemented. Thus, the analysis evaluates only a geometrically plausible non-covalent encounter complex and cannot predict or demonstrate irreversible AA-protein adduct formation. For each protein-ligand complex, the pose with the lowest AutoDock Vina score (kcal/mol) was selected for visualization and interaction analysis. The Vina score is a computational estimate for a non-covalent pose and should not be interpreted as an experimentally measured binding free energy, covalent reaction energy, or evidence of target engagement. Protein-ligand interactions were visualized using PyMOL software, and key amino acid residues involved in binding were identified based on hydrogen bonding and hydrophobic interactions. Potential ligand-binding pockets of target proteins were predicted using CurPockets. CurPockets evaluates protein surface cavities based on geometric and physicochemical features and assigns a pocket score reflecting predicted pocket potential. These CurPockets scores are independent of molecular docking binding energies and should not be interpreted as thermodynamic binding free energies; instead, they were used for relative ranking of binding pocket potential among candidate proteins. Pathway analysis Functional states were analyzed using CancerSEA [18]. Gene set variation analysis (GSVA) with the GSVA R package was used to calculate enrichment scores for functional states and KEGG metabolic pathways. Correlations with gene expression were assessed using appropriate correlation analyses. Drug sensitivity analysis The XSum algorithm compared key gene signatures against CMap to identify potential inhibitory compounds [19,20]. Correlations between gene expression and drug sensitivity (AUC from CTRP, IC50 from GDSC) were evaluated via Spearman analysis. Construction of a clinical prognostic model Protein-protein interaction data for RRM2 were retrieved from the BioGRID database to identify interacting partner genes. A total of 10 RRM2-interacting genes were identified and subjected to univariate Cox proportional hazards regression in the TCGA-STAD cohort (n = 375) to identify genes associated with overall survival (p < 0.05). The prognostic signature was constructed using an integrative framework comprising Stepwise Cox, Lasso, Ridge, plsRcox, CoxBoost, Random Survival Forest, Gradient Boosting Machine, Elastic Net, Supervised Principal Components, and Survival-Support Vector Machine. Ten-fold cross-validation was used to select the signature with the highest C-index. For the final Lasso plus bidirectional Stepwise Cox model, each patient's risk score was calculated as the Cox linear predictor: Risk score = Σ(i = 1 to n) βi × Expri, where Expri is the expression value of retained gene i and βi is its fixed coefficient estimated in the TCGA-STAD training cohort. The same coefficient set was applied to the validation cohorts without outcome-based re-fitting. Discriminative performance was evaluated using PCA, time-dependent ROC analysis, and C-index comparison. Independent prognostic value was assessed by multivariable Cox regression incorporating age, AJCC stage, and gender. Within each evaluated cohort, patients were stratified into high- and low-risk groups using that cohort's median calculated risk score as the prespecified cutoff [21]. Analysis of drug sensitivity, immune response, and metabolic pathways based on the prognostic model The prognostic model was used to divide the TCGA-STAD cohort into high-risk and low-risk groups. The pRRophetic R package was used to predict drug sensitivity by comparing estimated responses to different chemotherapeutic drugs between risk groups. Potential responses to immune checkpoint inhibitors, namely anti-PD-1 and anti-CTLA-4 therapies, were predicted using the Submap method (http://cloud.genepattern.org/gp) [22]. To investigate alterations in metabolic activity, enrichment scores for KEGG metabolic pathways were calculated using GSVA, and differential metabolic pathway activity between the high-risk and low-risk groups was assessed. Single-cell RNA sequencing data processing and quality control The Seurat R package (v4.4.0) was used to process raw scRNA-seq gene expression matrices. Quality control (QC) retained genes detected in at least 10 cells and cells with 200–5000 detected genes. Cells with mitochondrial gene content greater than 10% of UMIs were excluded. Data were normalized using the standard Seurat workflow. The Harmony algorithm was used to correct batch effects between samples. After PCA on highly variable genes, t-distributed stochastic neighbor embedding (t-SNE) visualization was performed using the top 30 principal components. The FindAllMarkers function was used to compare gene expression across identified cell clusters. Cell type annotation was assigned based on canonical lineage-specific marker genes: T cells (CD3D, CD3E, CD2), B cells (CD79A, MS4A1/CD20, CD19), neutrophils (CSF3R, FCGR3B, S100A8), plasma cells (MZB1, SDC1/CD138, JCHAIN), epithelial cells (EPCAM, KRT8, KRT18), macrophages (CD68, CD163, CD14), fibroblasts (COL1A1, DCN, LUM), endothelial cells (PECAM1, CDH5, VWF), and mast cells (KIT, TPSAB1, CPA3). Cell-cell communication network analysis We employed the CellChat algorithm to analyze intercellular communication networks in the tumor microenvironment [23]. This computational approach systematically deduces biologically relevant signaling relationships by amalgamating scRNA-seq data with a curated repository of ligand-receptor interactions (CellChatDB). Potential communication events were deduced from the coordinated overexpression of ligands and their corresponding receptors inside particular cell subpopulations, adhering to the established CellChat procedure. Beyondcell analysis for drug sensitivity prediction at single-cell resolution The Beyondcell algorithm was employed to predict drug sensitivity at single-cell resolution [24]. This computational tool calculates a Beyondcell Score (BCS) for individual cells by comparing their transcriptomic profiles against a large collection of drug perturbation signatures. We applied Beyondcell to the scRNA-seq data to identify potential therapeutic agents selectively targeting specific cell subpopulations of interest. Cell culture Hs746T, MKN-45, and MKN-74 cell lines were maintained in DMEM or RPMI-1640 medium, supplemented with 10% fetal bovine serum and 1% penicillin/streptomycin. KATO III cells were cultured in IMDM medium supplemented with 20% FBS and 1% penicillin/streptomycin. All cells were incubated at 37 °C in a humidified environment containing 5% CO₂. RRM2 knockdown and overexpression Lentiviral vectors carrying shRNA targeting RRM2 were used to establish stable RRM2-knockdown Hs746T cells. For overexpression, a plasmid containing the RRM2 gene was transfected into Hs746T cells. Stable cells were selected using antibiotics, and modulation of RRM2 expression was confirmed by Western blot. Transwell invasion assay Cell invasion was evaluated using Transwell chambers that were coated with Corning Matrigel® Matrix at a dilution of 1:50. Cells deprived of serum were placed in the upper chamber, while the lower chamber was filled with medium supplemented with 10% FBS. Following a 12-hour incubation, invaded cells were fixed using 4% formaldehyde, stained with 0.1% crystal violet, and subsequently counted via microscopy. Western blot A BCA assay was used for total protein quantification after RIPA lysis buffer was used for extraction. After running the proteins through SDS-PAGE, they were transferred to NC membranes and then blocked with 5% non-fat milk. Prior to using an HRP-conjugated secondary antibody, the membranes were treated with primary antibodies against RRM2 and β-actin (Proteintech). In order to examine the protein bands, we used ImageJ and ECL for visualization. Flow cytometry Apoptosis was assessed utilizing an Annexin V-FITC/PI apoptosis kit. Cell proliferation was evaluated utilizing an EdU imaging kit in accordance with the manufacturer's guidelines. Data were collected using a flow cytometer and subsequently analyzed with FlowJo software. Statistical analysis Data are presented as mean ± SD. For in vitro experiments, comparisons between two groups were performed using two-tailed Student's t-test; comparisons among three or more groups were analyzed by one-way ANOVA followed by Dunnett's post-hoc test. All experiments were performed with at least three independent biological replicates (n ≥ 3). Specific statistical tests and sample sizes for each experiment are reported in the corresponding figure legends. Statistical analyses were performed using GraphPad Prism software (version 10.0). A p-value of < 0.05 was considered statistically significant. Public datasets were processed and visualized using R software (version 4.4.0). Statistical significance was denoted as *P < 0.05, **P < 0.01, and ***P < 0.001. Results Identification of ANGRGs Through integrative analysis of genes retrieved from CTD, MSigDB, and GeneCards, we identified 11 overlapping genes: ADA, CDC45, CENPF, GPSM2, PLK4, RRM2, GSR, HPRT1, IER3, NT5C2, and TXNRD1 (Fig. 1A). This gene set was defined as Acrylamide-Nucleotide Metabolism-Gastric Cancer-Related Genes (ANGRGs). CurPockets analysis prioritized RRM2 on the basis of its pocket score (Fig. 1B), and standard AutoDock Vina docking was used to visualize possible non-covalent encounter poses between AA and the proteins encoded by the 11 ANGRGs (Fig. 1C). These simulations did not model covalent binding to cysteine or any other residue and therefore do not demonstrate AA-protein adduct formation.Fig. 1Identification of ANGRGs and molecular docking analysis.(A) Venn diagram showing the intersection of genes associated with acrylamide exposure, nucleotide metabolism, and gastric cancer, resulting in 11 candidate genes.(B) Binding pocket scores of candidate proteins evaluated by CurPockets analysis. More negative values indicate stronger predicted pocket features.(C) Representative 3D structural models depicting standard non-covalent docking poses between acrylamide (stick representation) and the proteins encoded by the 11 ANGRGs. Highlighted residues indicate predicted non-covalent contacts; no covalent bond or cysteine-adduct formation was simulated. Identification of ANGRGs and molecular docking analysis. (A) Venn diagram showing the intersection of genes associated with acrylamide exposure, nucleotide metabolism, and gastric cancer, resulting in 11 candidate genes. (B) Binding pocket scores of candidate proteins evaluated by CurPockets analysis. More negative values indicate stronger predicted pocket features. (C) Representative 3D structural models depicting standard non-covalent docking poses between acrylamide (stick representation) and the proteins encoded by the 11 ANGRGs. Highlighted residues indicate predicted non-covalent contacts; no covalent bond or cysteine-adduct formation was simulated. RRM2 expression, functional associations, and therapeutic implications RRM2 mRNA was significantly overexpressed in GC tissues compared with adjacent normal tissues in the TCGA-STAD cohort and GSE27342 dataset (Supplementary Figure 1A). Functionally, high RRM2 expression showed positive correlations with nucleotide-metabolism pathways (including pyrimidine and purine metabolism) and functional states such as cell cycle, DNA damage, and DNA repair, while showing a negative correlation with tumor immune cytolytic activity (Supplementary Figure 1B-D). These findings support an association of RRM2 with metabolic reprogramming and immune suppression but do not independently establish its prognostic effect. Elevated RRM2 expression was also associated with resistance to multiple tyrosine kinase inhibitors and chemotherapeutic agents (Supplementary Figure 2A-C). CMap analysis nominated butein, imatinib, and MK-886 as candidate agents for RRM2-high tumors (Supplementary Figure 2D); these computational predictions require experimental validation. Experimental validation of RRM2 expression and function Western blot analysis confirmed that RRM2 protein was highly expressed in multiple gastric cancer cell lines, including Hs746T, KATO III, and MKN45 (Fig. 2A). Given its prominent expression, the Hs746T cell line was selected for functional characterization. We successfully established stable RRM2-knockdown (using shRNA, specifically the sh2 construct) and RRM2-overexpressing (OE) Hs746T cell lines, as confirmed by Western blot (Fig. 2B, C). Flow cytometric analysis revealed that RRM2 exerted both anti-apoptotic and pro-proliferative effects. Overexpression of RRM2 significantly decreased the proportion of apoptotic cells (Annexin V-positive), whereas RRM2 knockdown increased apoptosis (Fig. 2D). Conversely, RRM2 overexpression increased the percentage of proliferating cells (EdU-positive), while its knockdown reduced it (Fig. 2E). Furthermore, migration and invasion assays demonstrated that RRM2 is associated with metastatic behavior: RRM2 overexpression enhanced the migratory and invasive capacities of Hs746T cells, whereas RRM2 knockdown impaired their ability to penetrate the extracellular matrix (Fig. 2F, G). In summary, our functional experiments indicate that RRM2, frequently overexpressed in GC, promotes invasion and proliferation while inhibiting apoptosis.Fig. 2Functional validation of RRM2 in gastric cancer cells.(A) Western blot analysis confirming high RRM2 protein expression in a panel of gastric cancer cell lines (Hs746T, KATO III, MKN45).(B, C) Western blot validation of efficient RRM2 knockdown (shRRM2) and overexpression (OE-RRM2) in Hs746T cells.(D) Flow cytometry analysis (Annexin V/PI staining) showing the percentage of apoptotic cells under RRM2 knockdown and overexpression conditions (*P < 0.05, **P < 0.01).(E) EdU incorporation assay quantifying the proportion of proliferating cells following RRM2 manipulation (*P < 0.05, **P < 0.01).(F, G) Representative images and quantification of Transwell migration and Matrigel invasion assays in Hs746T cells (scale bar: 100 μm; *P < 0.05, **P < 0.01). Functional validation of RRM2 in gastric cancer cells. (A) Western blot analysis confirming high RRM2 protein expression in a panel of gastric cancer cell lines (Hs746T, KATO III, MKN45). (B, C) Western blot validation of efficient RRM2 knockdown (shRRM2) and overexpression (OE-RRM2) in Hs746T cells. (D) Flow cytometry analysis (Annexin V/PI staining) showing the percentage of apoptotic cells under RRM2 knockdown and overexpression conditions (*P < 0.05, **P < 0.01). (E) EdU incorporation assay quantifying the proportion of proliferating cells following RRM2 manipulation (*P < 0.05, **P < 0.01). (F, G) Representative images and quantification of Transwell migration and Matrigel invasion assays in Hs746T cells (scale bar: 100 μm; *P < 0.05, **P < 0.01). Construction of an RRM2-related prognostic model Proteins that directly interact with RRM2 were obtained from the BioGRID database, and a protein-protein interaction (PPI) network centered on RRM2 was constructed using the igraph R package (Fig. 3A). Pathway enrichment analysis of the interacting partners indicated significant participation in biological processes, including processing of capped intron-containing pre-mRNA and spliceosomal snRNP assembly (Fig. 3B), implying a role for the RRM2 interactome in RNA processing. Univariate Cox regression analysis of the genes encoding the interacting proteins revealed 10 genes significantly associated with overall survival in gastric cancer patients, including 4 protective factors and 6 risk factors (Fig. 3C). A prognostic model was subsequently built using these 10 genes. Among the algorithmic combinations tested, the model generated by Lasso regression combined with Stepwise Cox (both-direction selection) demonstrated the optimal predictive performance based on C-index comparison (Fig. 3D, Supplementary File 3). Patients in the high-risk group showed significantly poorer overall survival than those in the low-risk group across the evaluated cohorts (Fig. 3E-H). Time-dependent ROC curves were used to evaluate the predictive performance of the model. In the GSE15459 cohort, the AUC values for 1-, 3-, and 5-year survival were 0.59, 0.64, and 0.61, respectively (Fig. 3E). In the GSE62254 cohort, the corresponding AUC values were 0.54, 0.57, and 0.59 (Fig. 3F). In the GSE84433 cohort, the AUC values were 0.54, 0.59, and 0.61, respectively (Fig. 3G). In the TCGA-STAD cohort, the corresponding AUC values were 0.67, 0.62, and 0.66 (Fig. 3H).Fig. 3Construction and validation of an RRM2-centric prognostic model.(A) A protein-protein interaction (PPI) network centered on RRM2 (target node), built using data from the BioGRID database.(B) Bar plot of the top significantly enriched Gene Ontology (GO) biological processes among RRM2-interacting proteins.(C) Forest plot displaying the 10 RRM2-interacting genes with significant prognostic value (hazard ratios) in GC.(D) Concordance index (C-index) comparison of prognostic models built using different machine-learning algorithms. The Lasso+StepCox (both-direction selection) model showed superior performance.(E-H) Kaplan-Meier survival curves and time-dependent ROC curves comparing high-risk and low-risk patients across validation cohorts. Construction and validation of an RRM2-centric prognostic model. (A) A protein-protein interaction (PPI) network centered on RRM2 (target node), built using data from the BioGRID database. (B) Bar plot of the top significantly enriched Gene Ontology (GO) biological processes among RRM2-interacting proteins. (C) Forest plot displaying the 10 RRM2-interacting genes with significant prognostic value (hazard ratios) in GC. (D) Concordance index (C-index) comparison of prognostic models built using different machine-learning algorithms. The Lasso+StepCox (both-direction selection) model showed superior performance. (E-H) Kaplan-Meier survival curves and time-dependent ROC curves comparing high-risk and low-risk patients across validation cohorts. Metabolic profile, chemotherapy response, and immunotherapy prediction of the prognostic model To maintain platform consistency, GSVA of KEGG metabolic pathways was performed on the three GEO microarray cohorts with the largest sample sizes and complete clinical annotations (GSE62254, GSE26253, and GSE84433). The high-risk group exhibited lower enrichment scores for nucleotide-metabolism pathways than the low-risk group (Fig. 4A), a pattern that may be associated with altered DNA-synthesis or repair capacity but does not itself measure genomic instability. Drug-sensitivity analysis indicated higher predicted IC50 values for cisplatin and 5-fluorouracil (5-FU) in the high-risk group (Fig. 4B). Submap analysis also generated a predicted association between the high-risk expression profile and anti-CTLA-4 response (Fig. 4C). However, patient-level MSI and EBV subtype annotations were not incorporated into this analysis; consequently, the result cannot establish that the high-risk group is MSI-high, EBV-positive, or clinically sensitive to immune-checkpoint blockade. It should be interpreted as a hypothesis-generating transcriptomic similarity signal requiring validation in an independent, molecularly annotated immunotherapy cohort.Fig. 4Metabolic and therapeutic characteristics of the prognostic risk groups.(A) Heatmap of GSVA scores for KEGG metabolic pathways, showing significantly lower nucleotide metabolism activity in the high-risk group.(B) Box plots comparing the predicted IC50 values for cisplatin and 5-FU between risk groups, indicating higher chemoresistance in the high-risk group (*P < 0.05, **P < 0.01, ***P < 0.001).(C) Submap analysis showing a predicted transcriptomic association between the high-risk group and anti-CTLA-4 response. This computational result is hypothesis-generating and does not establish clinical sensitivity. Metabolic and therapeutic characteristics of the prognostic risk groups. (A) Heatmap of GSVA scores for KEGG metabolic pathways, showing significantly lower nucleotide metabolism activity in the high-risk group. (B) Box plots comparing the predicted IC50 values for cisplatin and 5-FU between risk groups, indicating higher chemoresistance in the high-risk group (*P < 0.05, **P < 0.01, ***P < 0.001). (C) Submap analysis showing a predicted transcriptomic association between the high-risk group and anti-CTLA-4 response. This computational result is hypothesis-generating and does not establish clinical sensitivity. Defining the ANGScore in the single-cell transcriptome Cell annotation of the scRNA-seq data revealed nine primary cell types: T cells, B cells, neutrophils, plasma cells, epithelial cells, macrophages, fibroblasts, endothelial cells, and mast cells (Supplementary Figure 3). The t-SNE visualization depicts the expression patterns of the 11 ANGRGs among these cell populations (Fig. 5A). We calculated an ANGScore for each individual cell using ssGSEA based on the expression levels of the 11 ANGRGs, and cells were categorized into high- and low-ANGScore groups. In the epithelial (tumor) cell compartment, CellChat analysis indicated that tumor cells with a high ANGScore demonstrated increased outgoing communication, predominantly facilitated by signaling pathways including MIF, APP, and CXCL (Fig. 5B). Chord diagrams and heatmaps detailed these specific interactions, highlighting enhanced ligand-receptor-mediated communication between high-ANGScore tumor cells and key immune populations such as macrophages, T cells, and B cells via the MIF, APP, and CXCL pathways (Fig. 5C, D). This amplified crosstalk suggests that these tumor cells may actively remodel the immune microenvironment toward an immunosuppressive state favorable for tumor progression. The ANGScore was computed using all 11 ANGRGs to capture the collective transcriptional signature of the AA-nucleotide metabolism-GC axis; this integrative approach provides a more comprehensive representation of the biological pathway than any single gene alone. To specifically evaluate the contribution of RRM2, an RRM2-focused sub-analysis is provided in Supplementary Figure 4.Fig. 5Single-cell transcriptomic analysis and intercellular communication.(A) t-SNE projection of the gastric cancer scRNA-seq dataset, colored by cell type, and feature plots showing the expression distribution of representative ANGRGs across cells.(B) Bubble plot from CellChat analysis comparing outgoing communication patterns of high- versus low-ANGScore epithelial (tumor) cells. Key signaling pathways are highlighted.(C) Chord diagram visualizing specific ligand-receptor interactions between high-ANGScore tumor cells and recipient immune cells.(D) Heatmap summarizing the differential communication probability of key ligand-receptor pairs involving high-ANGScore tumor cells. Single-cell transcriptomic analysis and intercellular communication. (A) t-SNE projection of the gastric cancer scRNA-seq dataset, colored by cell type, and feature plots showing the expression distribution of representative ANGRGs across cells. (B) Bubble plot from CellChat analysis comparing outgoing communication patterns of high- versus low-ANGScore epithelial (tumor) cells. Key signaling pathways are highlighted. (C) Chord diagram visualizing specific ligand-receptor interactions between high-ANGScore tumor cells and recipient immune cells. (D) Heatmap summarizing the differential communication probability of key ligand-receptor pairs involving high-ANGScore tumor cells. Drug sensitivity evaluation of ANGScore subtypes via beyondcell Beyondcell analysis revealed distinct drug sensitivity profiles between the ANGScore-defined subtypes. Patients/cells characterized by a low ANGScore were predicted to exhibit higher sensitivity to compounds such as CHIR-99,021 (a GSK-3β inhibitor) and l-701,252 (a dopamine receptor antagonist) (Fig. 6A). Conversely, the high ANGScore subtype was predicted to be more vulnerable to agents including TYRPHOSTIN-A9 (a PDGFR inhibitor), RAZOXANE (a topoisomerase II inhibitor), THAPSIGARGIN (a SERCA pump inhibitor), and CCCP (a mitochondrial uncoupler) (Fig. 6B). These predictions, derived from single-cell transcriptomic data, suggest that tailored therapeutic strategies based on the ANGScore subtype could provide more effective, personalized treatment options beyond standard chemotherapy, potentially overcoming drug resistance and improving clinical outcomes.Fig. 6Drug sensitivity prediction for ANGScore subtypes at single-cell resolution.(A) Bar plot of top Beyondcell-predicted sensitive compounds for the low-ANGScore subtype, ranked by normalized Beyondcell Score (BCS).(B) Bar plot of top Beyondcell-predicted sensitive compounds for the high-ANGScore subtype, ranked by normalized BCS. Drug sensitivity prediction for ANGScore subtypes at single-cell resolution. (A) Bar plot of top Beyondcell-predicted sensitive compounds for the low-ANGScore subtype, ranked by normalized Beyondcell Score (BCS). (B) Bar plot of top Beyondcell-predicted sensitive compounds for the high-ANGScore subtype, ranked by normalized BCS. Discussion This study identifies RRM2 as a candidate molecular factor connecting AA-associated stress with nucleotide-metabolic dysregulation in GC. The study combines database integration, machine learning, functional assays, and single-cell analysis. Together, these approaches place RRM2 at the intersection of tumor-cell behavior, metabolic adaptation, and immune communication. However, they do not prove that AA directly targets RRM2. Our screening prioritized RRM2 because it had a favorable CurPockets score and a plausible non-covalent docking pose. This result must be interpreted carefully. AA is an electrophile and can form covalent adducts, but AutoDock Vina does not simulate covalent reactions with cysteine or other nucleophilic residues. The docking result therefore describes only a possible encounter geometry. Direct target-engagement assays and adduct mapping are needed before RRM2 can be considered a biochemical target of AA. RRM2 supplies deoxyribonucleotides for DNA synthesis and repair [25,26]. Its overexpression in GC is therefore biologically consistent with rapid cell division and replication-stress tolerance [27,28]. In Hs746T cells, RRM2 overexpression increased proliferation, migration, and invasion and reduced apoptosis. Knockdown produced the opposite effects. These results support an oncogenic role for RRM2 in this model. The GSVA results also linked RRM2 expression with purine and pyrimidine metabolism. This association fits the established role of metabolic reprogramming in cancer [29,30]. We also observed an inverse association between RRM2 expression and immune cytolytic activity. This finding raises the possibility that nucleotide demand and immune escape are coordinated rather than independent features of GC progression [31,32]. The present data show association, however, and do not define the causal mediator between these processes. The RRM2-centered prognostic model separated patients into groups with different survival outcomes. The high-risk group had lower nucleotide-metabolism GSVA scores. In our view, this apparent paradox may reflect a stress-adapted state rather than simple metabolic inactivity. Pathway scores do not directly measure dNTP pools, replication stress, or chromosomal instability. Experimental studies in other systems nevertheless show that nucleotide insufficiency and abnormal ribonucleotide-reductase activity can promote genome instability [33,34]. Predicted treatment responses also require cautious interpretation. Cisplatin resistance is multifactorial [35]. Likewise, the Submap association with anti-CTLA-4 response is not evidence of clinical sensitivity. MSI-high and EBV-positive gastric cancers are often more immunogenic, whereas chromosomally unstable tumors may be immune-cold and treatment resistant. Because our analysis lacked patient-level MSI and EBV annotations, molecular subtype could explain part of the predicted signal. Validation in a treated and molecularly annotated GC cohort is essential [36]. At single-cell resolution, high-ANGScore epithelial cells showed stronger outgoing communication with macrophages, T cells, and B cells. MIF, APP, and CXCL pathways contributed to this pattern. These pathways can influence macrophage polarization, T-cell function, and B-cell recruitment in other cancers [[37], [38], [39]]. We therefore propose that high-ANGScore tumor cells may help organize an immunosuppressive niche. This proposal is testable in GC co-culture models and spatially resolved tissues. Beyondcell analysis provided a translational extension of this framework. It nominated CHIR-99,021 and l-701,252 for the low-ANGScore state and TYRPHOSTIN-A9, razoxane, thapsigargin, and CCCP for the high-ANGScore state. These predictions support a subtype-guided drug-repurposing strategy [40]. The known antiproliferative activity of compounds such as imatinib and butein adds biological plausibility [41,42], but efficacy and safety must be established experimentally. Our perspective is that the main contribution of this work is not a claim of a completed linear mechanism. Instead, it is a testable model in which environmental electrophilic stress, nucleotide homeostasis, and tumor-immune communication converge around an RRM2-centered state. This model generates three priorities for future work: define whether AA or glycidamide engages RRM2 directly; determine whether RRM2-dependent nucleotide imbalance alters immune signaling; and test whether ANGScore-defined states predict therapeutic response in molecularly stratified GC. This interpretation is more conservative than a direct-target claim, but it better reflects the evidence and provides a clearer path for validation. This study has several limitations. First, direct RRM2 engagement by AA requires SPR or CETSA [43]. Because AA is electrophilic, mass-spectrometric adduct mapping is also needed. Second, we did not resolve the relative effects of parent AA and glycidamide or measure tissue-specific CYP2E1 activity. Third, controlled in vivo exposure models are needed to establish causality. Fourth, the single-cell analysis included only four samples. Its communication and drug-response findings are therefore preliminary. Fifth, functional validation relied mainly on Hs746T cells and should be repeated in additional GC models. Sixth, the lower nucleotide-metabolism score in the high-risk group requires mechanistic study. Seventh, the predicted anti-CTLA-4 association was not stratified by MSI, EBV, chromosomal-instability status, or measured immune phenotype. It cannot be considered clinical evidence of checkpoint-blockade sensitivity. Finally, all predicted drugs require preclinical and prospective validation. Evidence from other tumor systems provides context: p53 can shape responses to environmental toxicants [44], and XRCC2 and the HOTAIR/EZH2 axis can affect DNA repair and treatment resistance [45,46]. These studies support the broader rationale but do not directly validate the AA-RRM2 axis in GC. In summary, we identify an AA-associated, RRM2-related nucleotide-metabolism network in GC. RRM2 manipulation altered malignant phenotypes in vitro. The current evidence does not show that parent AA or glycidamide directly and covalently targets RRM2. The prognostic signature and predicted therapeutic vulnerabilities should therefore be viewed as hypothesis-generating. Their clinical value requires validation in cohorts annotated for MSI, EBV, chromosomal-instability status, treatment, and response. Limitations and future directions This study has several limitations. First, the direct binding and functional modulation of RRM2 by AA require experimental validation using target-engagement methods such As spr or CETSA [43] and, because AA is electrophilic, covalent-adduct mapping by mass spectrometry. The present non-covalent docking cannot model or prove cysteine modification. Second, the relative contributions of direct gastric exposure to parent AA and systemic or locally generated glycidamide were not resolved, and tissue-specific CYP2E1 activity was not measured. Third, the causal role of AA in gastric carcinogenesis in vivo remains to be established in controlled exposure models. Fourth, the single-cell RNA-seq analysis included only four samples, so cell-cell communication and drug-sensitivity findings are preliminary. Fifth, functional validation was performed primarily in Hs746T cells and should be extended to additional GC models. Sixth, the reduction in nucleotide-metabolism pathway activity in the high-risk group requires mechanistic investigation. Seventh, the predicted anti-CTLA-4 association was derived from Submap and was not adjusted or stratified by MSI, EBV, chromosomal-instability status, or measured immune phenotype; it therefore cannot be interpreted as clinical evidence of ICB sensitivity. Finally, all drug-response predictions require preclinical and prospective clinical validation. Future work should combine controlled AA/glycidamide exposure, direct target-engagement and adduct assays, molecularly annotated GC cohorts, and in vivo or co-culture models to test the proposed relationships. Evidence from other tumor systems supports the broader biological rationale: environmental toxicant responses can be shaped by p53-dependent transcriptional regulation [44], whereas the homologous-recombination factor XRCC2 and the epigenetic HOTAIR/EZH2 axis can modulate DNA-repair capacity and treatment resistance [45,46]. These studies provide mechanistic context but are not direct evidence for the AA-RRM2 axis in GC. Conclusion In summary, our findings identify an AA-associated, RRM2-related nucleotide-metabolism network in GC and demonstrate that RRM2 manipulation affects malignant phenotypes in vitro. The data do not yet establish whether parent AA or glycidamide directly and covalently targets RRM2. The RRM2-centered prognostic signature and predicted therapeutic vulnerabilities are hypothesis-generating tools that require validation, particularly in cohorts annotated for MSI, EBV, chromosomal-instability status, and treatment response. Funding Not applicable. Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. CRediT authorship contribution statement Peiru Si: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Tianlong Zhang: Writing – review & editing, Writing – original draft, Visualization, Validation, Formal analysis, Data curation, Conceptualization. Xuefeng Lv: Writing – review & editing, Writing – original draft, Supervision, Resources, Project administration, Funding acquisition, Formal analysis, Data curation, Conceptualization. Declaration of competing interest The authors declare that they have no conflicts of interest.