Development of an ADME gene signature for prognostic and therapeutic stratification in gastric cancer.

Introduction
Cancer remains the most lethal disease globally. According to statistics, gastric cancer (GC) holds the fifth position in global cancer incidence. In 2020 alone around 1.09 million new GC cases were diagnosed globally, and 769,000 patients died of GC [1]. Alarmingly, the incidence of GC in young people is increasing. It is noteworthy that the morbidity of GC in Asia (highest in East Asia) is significantly higher than that in Europe and the United States [2]. As is widely known, the prognosis of GC patients largely depends on timely diagnosis. Although advanced diagnostic techniques and treatment methods have improved the survival rate of GC patients, early diagnosis remains a critical challenge. To address this issue, researchers have increasingly focused on the tumor microenvironment (TME), which is crucial for the survival and progression of tumor cells [3]. TME is composed of tumor cells, neighboring stromal cells, immune cells, vascular architecture, extracellular matrix, and metabolic byproducts. Quail’s study [4] has primarily focused on inhibiting tumor growth by modulating the cellular and non-cellular components of the TME. Among these components, immune cells are crucial in tumor development. Therefore, investigating tumor-immune interactions can provide important references for developing next-generation immunotherapies and enhancing the predictive value of GC prognosis.
Conventional chemotherapeutic agents, such as fluorouracil and platinum, are limited by chemoresistance and toxicity, which hinders their further development. Targeted therapies, mainly including anti-HER2 therapies and antiangiogenic drugs, have notably prolonged the survival of GC patients [5, 6]. However, Kang et al. [7] reported that these treatment methods were only applicable to specific subtypes, which accounted for less than 5% of all GC cases. Further research has shown that immune checkpoint inhibitors have broad clinical applicability and long-lasting efficacy, overcoming the limitations of targeted therapies to some extent. They are particularly suitable for GC tumors characterized by microsatellite instability-high or high expression of PD-L1. Unfortunately, these treatment methods still have low treatment response rates and lack precise predictive biomarkers [8].
Therefore, exploring new biomarkers to optimize clinical decision-making and improve the prognosis of GC patients is of great significance. The latest results of core gene research in pancreatic cancer have shown that ADME-related genes are associated with tumorigenesis and prognosis. ADME-related genes regulate the in vivo pharmacokinetic processes of drugs, including systemic drug metabolism (mediated by enzymes) and hepatic metabolism. The ADME gene family is relatively large, comprising 266 extended genes and 32 core genes. These genes are primarily involved in two key drug metabolism phases: Phase I reactions (mediated by drug-metabolizing enzymes) and Phase II conjugation reactions (catalyzed by transferases). Additionally, drug transporters and drug-metabolizing enzymes contribute to the ADME pharmacokinetic process. Crucially, the activity of these drug metabolizing systems is a key factor determining the bioavailability of drug therapy [9, 10]. Moreover, drug metabolism can trigger immune responses, which may modulate tumor progression [11]. The interplay between ADME-related genes and malignancies is mediated by dual mechanisms of pharmacodynamics and pharmacokinetics. Firstly, ADME-related genes regulate the pharmacokinetics of anticancer drugs, including metabolism, transport, and activation. Secondly, ADME-related genes modulate signaling pathways, thereby promoting or inhibiting the growth of cancer cells. In these signaling pathways, endogenous molecules and exogenous regulators play critical roles in network modulation. Subsequently, investigating the relationship between ADME gene polymorphisms and pharmacokinetics can optimize tumor drug targeting and lay the foundation for personalized therapy. The tumor-modulatory roles of ADME-related genes have been confirmed. For instance, ABCB5 is involved in chemoresistance in breast cancer. Wu et al. [12] found that upregulation of NAT1, ADH1B, CYP1A1, ABCC9, and CYP46A1 expression promoted tumor aggressiveness in breast cancer, while overexpression of CYP21A2 inhibited the malignant progression. Seven ADME-related genes with prognostic significance have been discovered in research on non-small cell lung cancer, including SLC16A1. Notably, patient prognosis may be improved by regulating in vivo drug metabolism mediated by these genes [13]. It can be inferred that ADME-related genes influence the efficacy of GC treatment and the composition of the TME. Unfortunately, there have been an insufficient number of studies regarding the prognostic relevance of ADME-related genes to GC, making it challenging to exploit ADME-related genes for the treatment of GC. We aim to identify the characteristics of ADME-related genes in GC to optimize personalized treatment plans for patients.

Materials and methods

Data collection and analysis
Based on complete cases, we retrieved genomic data and corresponding clinical metadata of 371 GC patients from the Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov, accessed on: 05 June 2025). A random 1:1 split was applied to allocate cases to the training set and the internal validation set. For the external validation set, data were derived from the GSE84433 dataset within the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/gds, accessed on: 05 June 2025). Samples with complete clinical and survival information were included in the study. Gene expression data were converted to FPKM for subsequent analysis. To eliminate batch effects, batch correction was performed using the Combat method. In total, 298 ADME-related genes were sourced from previous academic works (Supplementary Table S1) [14].

Differentially expressed genes and network analysis of ADME-related genes
Initial differential expression analysis was conducted using the ‘’limma’’ package to detect significantly dysregulated genes between gastric tumor tissue and paired normal tissue (|log2-fold change|>1 and p < 0.01). The ‘’ggplot2’’ package was employed to draw a volcano map based on the analysis results. Subsequently, we used the STRING database (http://string-db.org/, accessed on: 05 June 2025) to create a protein-protein interaction (PPI) network. We measured node-node correlations and analyzed the prognostic value of ADME gene co-expression. The correlation coefficient of network nodes was computed, and the co-expression of ADME genes associated with prognosis was examined.

Development of a risk model on ADME-related genes
We first performed univariate Cox regression analysis on ADME-related differentially expressed genes (DEGs), followed by LASSO regression analysis for dimensionality reduction and screening of non-redundant genes. The threshold was set at P < 0.05. Both analyses were implemented using the ‘’glmnet’’ package. We constructed a prognostic model for ADME-associated signatures. Subsequently, multivariate Cox regression analysis was used to estimate the regression coefficients of each gene. We multiplied gene expression levels by their corresponding regression coefficients and then summed the weighted values to calculate the risk score. The training set was divided into low-risk and high-risk strata based on the optimal cutoff point for survival analysis. To further verify the predictive performance, the same algorithm was applied to the TCGA-test, TCGA-all, and GSE84433 datasets to calculate risk scores. The predictive capacity was assessed via receiver operating characteristic (ROC) curve analysis.

Nomogram construction
In the construction of our prognostic model, the genetic signature was incorporated alongside key clinical variables into a multivariable Cox regression model. To enhance clinical translatability and provide a user-friendly tool for individualized prediction, a nomogram was subsequently developed using the ‘’rms’’ package. This nomogram visually integrates the genetic signature with established clinical factors by featuring the risk score as a distinct line item, positioned alongside age, gender, pathological stage, and tumor-node-metastasis (TNM) stage for scoring purposes. Furthermore, to rigorously validate the robustness and generalizability of the genetic signature across different patient populations, we performed comprehensive subgroup analyses. The predictive reliability was assessed using calibration curves and ROC curves at 1-year, 3-year, and 5-year follow-ups. Decision curve analysis further evaluated net clinical benefit at different decision thresholds.

Immune microenvironment analysis
The ESTIMATE algorithm [15] exhibited high sensitivity in differentiating immune cell infiltration types and evaluating immune function scores in GC. To investigate the disparities in immune cell infiltration between the two risk strata, we compared cell abundance using multiple computational algorithms, namely TIMER, CIBERSORT-ABS, QUANTISEQ, EPIC, MCPCOUNTER, CIBERSORT, and XCELL. Single-sample gene set enrichment analysis (ssGSEA) was employed to quantify and compare the extent of immune cell infiltration and the functional characteristics of diverse immune cell populations in the two strata. The tumor immune dysfunction and rejection (TIDE) calculation model was further employed to forecast the differences in immunotherapy responses in risk-stratified groups. Data were derived from the TIDEweb (http://tide.dfci.harvard.edu/). Fisher’s exact test was employed to evaluate intergroup differences in treatment response based on drug efficacy outcomes.

ADME-related gene enrichment analysis
The DAVID bioinformatics platform (v2022q4 version; https://david.ncifcrf.gov/, accessed on: 05 June 2025) was utilized. The Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway and Gene Ontology (GO) database were used for functional annotation to characterize gene functions. The enriched GO terms included cellular component (CC), biological process (BP), and molecular function (MF) terms. We further conducted gene set variation analysis (GSVA) to compare functional pathway activities between different sample strata.

Drug sensitivity prediction
Depending on the Genomics of Drug Sensitivity in Cancer database (https://www.cancerrxgene.org/), the half-maximal inhibitory concentrations (IC50) of 17-AAG, AKT inhibitor VIII, EKB-569, FK866, GSK690693, LY317615, IXO2, BX-795, and CGP-60,474 in GC patients from TCGA database were computed to evaluate the predictive performance of models for chemosensitivity. The pRRophetic algorithm [16] with ridge regression was used to predict drug sensitivity.

In vitro culture
AGS and HGC-27 cell lines (Invitrogen) were incubated in RPMI 1640 medium supplemented with 10% fetal bovine serum and 1% antibiotics at 37 °C. Cells were allowed to adhere for 24 h. Small interfering RNA targeting CYP2A6 (si-CYP2A6) was obtained from iGeneBio company. Cells were transfected using Lipofectamine 3000 reagent (Invitrogen) for 6 h. Then, cells were incubated for 24 h.

Cell viability assay
Two cell lines were inoculated into 96-well plates at a density of 5,000 cells per well and cultured for 48 h. Afterwards, 10 µL of cell counting kit-8 solution was introduced into each well. The plate was incubated in the dark for 2 h. The absorbance of each well was determined with a microplate reader at a wavelength of 450 nm. The corresponding cellular bioactivity was calculated based on optical density values.

Cell scratch assay
Two cell lines were seeded into 6-well plates at a density of 1 × 105 cells per well and cultured for 24 h. A sterile 200 µL pipette tip was used to make a scratch on each cell monolayer, resulting in a uniform linear incision. Bright-field images were taken at 0 and 48 h after scratching, respectively.

Colony formation assay
Two cell lines were inoculated into 6-well plates at a density of 1,000 cells per well. After 14 days, the culture medium was carefully removed. The colonies were fixed with 4% paraformaldehyde at room temperature for 15 min and then stained with 0.1% crystal violet solution for 30 min. Subsequently, digital images of the colonies were acquired and quantitatively analyzed using ImageJ software to determine the number and size of the colonies.

Data statistics
Bioinformatics analysis was performed using R software. Survival curves were drawn using Kaplan-Meier (K-M) survival analysis. Differences between groups were statistically analyzed using the log-rank test. Cell-based experimental data were statistically analyzed using GraphPad Prism 9.4. Student’s t-test was used for two-sample data, and ANOVA method was used for multiple-sample data. All experiments were performed with at least three biological replicates. The criterion for statistically significant differences was defined as P ≤ 0.05.

Results

Differentially expressed gene analysis of gastric cancer
A total of 78 ADME-related DEGs associated with GC were screened (Fig. 1A). Among these, 45 were upregulated DEGs and 33 were downregulated DEGs (Fig. 1B). Subsequently, a PPI analysis was conducted on these 78 DEGs, and an interaction network diagram was constructed, with key nodes highlighted (the number of adjacent nodes ranged from 6 to 72) (Fig. 1C). The five genes with the greatest number of linkages were CYP3A4 (72 adjacent nodes), CYP2B6 (60 adjacent nodes), ABCG2 (56 adjacent nodes), CYP1A1 (54 adjacent nodes) and CYP2A6 (48 adjacent nodes) (Fig. 1D).

Establishment and analysis of ADME-associated signatures
Univariate Cox regression analysis was initially employed to screen 12 ADME-related genes with prognostic value. These included 9 risk predictors (CYP2A6, ADH1B, CYP19A1, FMO1, FMO2, GPX3, GSTM5, PON1, and SULF1) and 3 protective factors (KCNJ11, SLC29A2, and SLC5A6) (Fig. 2A). The LASSO algorithm was utilized to assist in constructing five optimal ADME-related risk characteristics (CYP2A6, CYP19A1, GPX3, SLC5A6, and SULF1) (Fig. 2B and C).
Risk score = CYP2A6 × (0.940429248002504) + CYP19A1 × (10.807285861356969) + GPX3 × (0.143080771069433) + SLC5A6 × (-0.338003981667798) + SULF1 × (0.187258576372905).
Figure 2D illustrates the correlation between survival outcomes and the distribution of risk scores in the entire training set. Patients in the low-risk stratum exhibited a higher survival rate (Fig. 2E). As illustrated in Fig. 2F, different expression patterns of these five prognostic genes were observed between the defined risk strata. The expression levels of CYP2A6, CYP19A1, GPX3, and SULF1 were markedly increased in the high-risk stratum, whereas SLC5A6 displayed an opposite expression pattern. The results of principal component analysis demonstrated that individuals could be clearly divided into two clusters (Fig. 2G). K-M survival curves indicated that, compared with the high-risk stratum in the TCGA training cohort, patients in the low-risk stratum showed significantly longer survival times (Fig. 2H). The accuracies of 1-year, 3-year, and 5-year survival predictions were 0.685, 0.738, and 0.838, respectively (Fig. 2I). This finding further validated the reliability of our model in predicting survival outcomes of GC patients.

Verification of ADME-associated signatures
The prognostic model we constructed showed robust predictive ability, which was validated in the test set. K-M survival analysis consistently indicated that the overall survival in the low-risk stratum was notably longer than that in the high-risk stratum in TCGA-test (p < 0.001), TCGA-all (p < 0.001), and GSE84433 datasets (p = 0.006) (Fig. 3A-C). The area under the receiver operating characteristic curve (AUC) values corresponding to 1-year, 3-year, and 5-year overall survival in each dataset were as follows: In the TCGA-test dataset, the AUC values were 0.661, 0.742, and 0.794, respectively (Fig. 3D); in the TCGA-all dataset, they were 0.705, 0.698, 0.738, respectively (Fig. 3E); and in the GSE8443 dataset, they were 0.673, 0.786, 0.830, respectively (Fig. 3F). The 95% confidence intervals for each dataset can be observed in Supplementary Table S2.

Survival analysis of clinical parameters in gastric cancer patients
We further examined the association between clinical parameters and survival time in different risk strata of GC patients. Key variables included age (with 65 years as the cutoff), gender, pathological stage, and TNM stage (Supplementary Table S3). K-M survival analysis verified that these variables were linked to clinical outcomes of GC patients. More specifically, among all the analyzed variables, the prognosis outcomes of patients in the low-risk stratum were relatively consistent (Fig. 4).

Prognostic efficacy of clinical parameters in gastric cancer patients
The concordance index of our model exceeded the values reported in previous studies, indicating that the model had excellent predictive performance (Fig. 5A). Cox regression analysis provided strong evidence that the risk score could precisely forecast the survival time. Statistical data further revealed that the risk score could be a robust clinical prognostic indicator (p < 0.001) (Fig. 5B and C). The verification set also provided strong evidence for this conclusion (Supplementary Figure S1). To further investigate the prognostic ability of ADME-associated signatures, Cox regression analysis was conducted by integrating the risk score with clinical parameters to construct a new nomogram (Fig. 5D). The calibration curve further provided robust evidence for the predictive accuracy of the model (Fig. 5E). The study found that the nomogram had good and satisfactory clinical net benefit in predicting clinical outcomes of GC patients (Fig. 5F). The AUC values of parameters (risk score, nomogram, age, gender, pathological stage, and TNM stage) for the 1-year overall survival were 0.639, 0.713, 0.580, 0.544, 0.563, and 0.607, respectively (Fig. 5G). The AUC values of parameters for the 3-year overall survival were 0.628, 0.730, 0.586, 0.508, 0.548, and 0.641, respectively (Fig. 5H). The AUC values of parameters for the 5-year overall survival were 0.732, 0.741, 0.600, 0.574, 0.539, and 0.623, respectively (Fig. 5I).

Immune profile of gastric cancer patients
This study employed the CIBERSORT and ssGSEA to characterize the immune microenvironment of GC. Specifically, it focused on exploring the alterations in 16 immune cell types and 13 immune functions across different risk-stratified subgroups. The analysis results revealed that, in comparison to the low-risk stratum, the high-risk stratum showed a greater number of infiltrating immune cells. These included B cells, CD8 + T cells, DCs, iDCs, macrophages, mast cells, neutrophils, pDCs, T helper cells, Tfh, TILs, and Tregs (Fig. 6A). Immune functions of the high-risk stratum were enhanced, as evidenced by the upregulation of APC costimulation, CCR, immune checkpoint, HLA, parainflammation, and IFN-I and IFN-II responses (Fig. 6B). The high-risk stratum exhibited both elevated cytotoxic T-cell infiltration and a significant enrichment of immunosuppressive markers (Supplementary Figure S2). This indicated that there were disparities in the tumor immune microenvironment between different prognostic strata. Multimodel evaluation findings suggested that the high-risk stratum demonstrated a greater degree of immune cell enrichment (Fig. 6C). Figure 6D illustrates distinct distribution patterns of different immune subtypes in individuals at different risk levels. To assess the immune evasion capacity of GC cells, this study calculated the TIDE scores for different risk strata. The TIDE score was lower in the high-risk stratum (Fig. 6E). In a clinical immunotherapy cohort, Fig. 6F illustrated the higher risk score in the CR/PR group (p = 0.0012). This suggests that our hypothesis that the high-risk group would respond better to immunotherapy is correct.

Functional and pathway analysis of differentially expressed genes
Functional enrichment analysis revealed that ADME-related DEGs were predominantly associated with the terms ‘’collagen-containing extracellular matrix’’, ‘’contractile fiber’’, and ‘’cell-substrate junction’’ in terms of CC. In BP, the most significantly enriched terms included ‘’organization of external encapsulating structures’’, ‘’extracellular structure assembly’’, and ‘’extracellular matrix’’. In terms of MF, DEGs were notably enriched in ‘’extracellular matrix structural constituent’’, ‘’glycosaminoglycan binding’’, and ‘’sulfur compound binding’’ (Fig. 7A, B). Based on the analysis results of KEGG, it was discovered that pathways related to ‘’focal adhesion’’, ‘’PI3K-Akt signaling pathway’’, and ‘’human papillomavirus infection’’ were activated (Fig. 7C and D).
Figure 7E indicates that the distribution of signaling pathways in GC altered dramatically across different risk strata. Among them, 14 signaling pathways were upregulated in the high-risk stratum (such as REGULATION_OF_ACTIN_CYTOSKELETON and ECM_RECEPTOR_INTERACTION), while 36 signaling pathways were upregulated in the low-risk stratum (such as BASAL_TRANSCRIPTION_FACTORS and SPLICEOSOME). Evidence strongly supports an association between these pathways and the ADME-related genes of GC that we have identified. For instance, CYP19A1 induces EMT in GC by regulating stromal markers, while SULF1 binds to TGFBR3 to activate the TGF-β signaling pathway, consequently triggering EMT in gastric cancer [17, 18]. Moreover, GPX3 promotes the development of GC by activating the PI3K-Akt signaling pathway [19]. Although no studies have directly linked CYP2A6 regulatory pathways to GC, the metabolism mediated by CYP2A6 can induce polarization of macrophages, such as activating the NF-κB pathway [20].

Tumor inhibitor efficacy prediction
Figure 8 depicts the results on the sensitivity analysis of tumor inhibitors. When compared to the low-risk stratum, the IC50 values for 7 drugs (17-AAG, AKT inhibitor 8, EKB-569, FK866, GSK690693, IXO2 and LY317615) were significantly elevated in the high-risk stratum (all p < 0.01). This suggests that patients in low-risk stratum may benefit from these drugs. Compared with the low-risk stratum, the high-risk stratum exhibited lower IC50 values for BX-795 and CGP-60,474 (p = 0.013 and p < 0.01, respectively), which indicates that the high-risk stratum is less sensitive to these two drugs.

Effect of CYP2A6 in gastric cancer
Based on a review of previous studies and preliminary investigations conducted by our team, it was discovered that the role of CYP2A6 in GC has not been previously documented. To elucidate the function of CYP2A6 in GC, this study designated it as the key regulatory gene for further exploration. This study successfully transfected si-CYP2A6 into AGS and HGC27 cell lines and assessed cellular activity via cell proliferation and colony formation assays. qPCR results showed good knockout efficiency (Fig. 9A). The proliferation of cells transfected with si-CYP2A6 was distinctly inhibited (Fig. 9B), and the number of colonies formed decreased (Fig. 9C). The findings of the cell migration analysis were consistent in both the 48-hour cell scratch assay (Fig. 9D) and the transwell assay (Fig. 9E), suggesting that knocking down CYP2A6 hindered cell migration. The carcinogenic effect of CYP2A6 was also confirmed in the apoptosis experiment (Supplementary Figure S3). In addition, we also explored the downstream pathways of CYP2A6. The results showed that the PI3K/AKT signaling pathway was significantly downregulated after CYP2A6 knockdown (Supplementary Figure S4), and this pathway played a key role in the development of GC. These findings implied that CYP2A6 might facilitate the progression of GC. This finding further suggests that CYP2A6 is a potential novel target for inhibiting the progression of GC.

Discussion
This study analyzed the correlation between ADME-associated genes in GC and the development and prognosis of GC. After merging the data from TCGA and GEO and performing quality control, a total of 371 GC patients were finally included. Through cross-database analysis, overlapping genes were screened between TCGA database and the known ADME gene set, identifying 45 upregulated genes and 33 downregulated genes. A prognostic model was created for dimensionality reduction and feature selection. The model included five ADME-associated signatures. Subsequently, the model was validated using TCGA-test and GSE84433 datasets to verify its accuracy. Then, the training set was divided into low-risk and high-risk strata based on the median score. The risk score is a reliable and independent predictor of survival. K-M survival analysis demonstrated significant improvement in survival outcomes in the low-risk stratum across all cohorts, while ROC analysis confirmed the good predictive performance of the model. The further development of a clinical nomogram that combined risk scores with key clinical features was validated as an effective tool for assessing the prognosis of GC patients. To further explore differences in the immune characteristics of TME and treatment responses among patients with different risk stratifications, an analysis of immune cell infiltration and function revealed that low-risk GC patients had better sensitivity to immunotherapy. Subsequently, this study conducted an enrichment analysis on GC patients to determine the correlation between these co-expressed genes and the TME. Finally, this study selected CYP2A6 as a key regulatory gene and further validated its role in GC. Cell viability and cell scratch assays were conducted in GC cells transfected with si-CYP2A6. The conclusions indicated that CYP2A6 could promote the progression of GC and was a potential novel target for inhibiting GC progression.
ADME is defined as the quantitative study of pharmacokinetic processes in vivo. Drug-metabolizing enzymes and transport proteins are the main regulatory determinants in this study. Cytochrome families are crucial for pharmacokinetics [21]. This finding inspired us to conduct a study exploring the potential of drug metabolism as a novel direction for targeted cancer therapy. This study identified five DEGs, including four upregulated genes (CYP2A6, CYP19A1, GPX3, and SULF1) and one downregulated gene (SLC5A6). These genes are linked to the onset and development of various tumors and display tumor-type-specific mechanisms of action. The functional enrichment analysis revealed compelling insights into the collective role of these genes in GC pathogenesis, particularly through modulation of the TME and key oncogenic pathways. The significant enrichment of terms related to the extracellular matrix (ECM), cell-substrate junctions, and focal adhesion underscores their central role in remodeling tissue architecture and promoting tumor-stroma crosstalk. CYP19A1, a key metabolic enzyme, accelerates GC development by simultaneously modulating fatty acid and glycerophospholipid biosynthesis pathways [22]. As essential constituents of cellular membranes and signaling molecules, glycerophospholipids play a crucial role in ECM remodeling, consistent with the enrichment of “extracellular matrix structural constituent” and “extracellular structure assembly”. SULF1, a member of the extracellular enzymes of the sulfatase family, crucially modulates HSPG signaling [23]. Fang et al. [18] reported that SULF1 facilitates GC progression by promoting the invasion and migration of tumor cells. Mechanistically, SULF1 secreted by cancer-associated fibroblasts binds to TGFBR3 on the surface of tumor cells, enhancing the phosphorylation and nuclear translocation of SMAD2/3, thereby activating canonical TGF-β signaling. The subsequent upregulation of EMT-related genes alters ECM organization and modulates the bioavailability of growth factors, including FGF and VEGF. These findings are strongly supported by enrichment of MF terms including “extracellular matrix structural constituent” and “glycosaminoglycan binding”, and BP terms such as “extracellular structure assembly”. Similarly, GPX3, an extracellular peroxidase, scavenges reactive oxygen species (ROS) via enzymatic activity to maintain oxidative homeostasis. Impaired GPX3 function may lead to ROS accumulation, particularly in collagen-rich regions, disrupting ECM organization and dysregulating integrin- and focal adhesion-mediated signaling [19]. This aligns with the significant enrichment of KEGG pathways such as “Focal adhesion” and “PI3K-Akt signaling pathway” in our analysis. Furthermore, SLC5A5 primarily facilitates the transport of biotin (vitamin B7) and pantothenic acid (vitamin B5) into mitochondria, where they serve as essential coenzymes in mitochondrial metabolic processes, including fatty acid oxidation and coenzyme A synthesis. By enhancing mitochondrial function and reducing ROS, SLC5A5 ultimately inhibits apoptosis [24]. These mechanistic insights are consistent with the enrichment of “Focal adhesion” and “PI3K-Akt signaling pathway”.However, CYP2A6 was ultimately selected as the target gene in this study. CYP2A6 demonstrated the highest regression coefficient weight, making the greatest contribution to the risk score model. Even minor alterations in its expression exerted substantial effects on patient risk stratification. PPI network analysis further identified CYP2A6 as a hub gene with the highest connectivity, suggesting its pivotal role in this functional cluster. CYP2A6 participates in tumor-related processes through its roles in drug metabolism and carcinogen activation. CYP2A6-mediated metabolism activates aflatoxin B1 (AFB1) and generates ROS, leading to DNA damage and oxidative stress. These events activate NF-κB signaling and induce the release of pro-inflammatory cytokines, including TNF-α and IL-6, which recruit immune cells and drive macrophage polarization toward a tumor-promoting phenotype. The polarized macrophages subsequently secrete immunosuppressive and pro-tumorigenic factors. Collectively, the resulting chronic inflammatory microenvironment facilitates tumor initiation, progression, and metastasis. Furthermore, ROS generated through CYP2A6-mediated metabolism act as second messengers that activate the PI3K-Akt signaling pathway. Persistent oxidative stress also oxidizes proteins involved in intercellular and cell-matrix junctions, impairing their function and enhancing cell motility [20, 25, 26]. These mechanisms are consistent with the enrichment of the terms “cell-substrate junctions” and “focal adhesions”. More importantly, tegafur, a core component of first-line GC chemotherapy regimens, such as S-1 tegafur capsules, requires metabolic activation by CYP2A6 in the liver. This enzyme influences the conversion of tegafur to 5-fluorouracil (5-FU) [27, 28], thereby affecting the pharmacological efficacy of chemotherapy in GC patients. TME plays an indispensable role in cancer immunotherapy [29]. Immune cells combine with antigen-presenting cells (such as dendritic cells) to form immunological synapses, which promote the specific killing of tumor antigens. Consequently, this study examined the quantity and functional activity of immune cell infiltration in GC. The TME demonstrates dynamic changes corresponding to the risk stratification of GC patients, which directly result in alterations in the composition of tumor-infiltrating immune cells [30, 31]. Genes highly expressed in the high-risk stratum include CYP2A6, CYP19A1, GPX3, and SULF1. On one hand, the TME in this stratum exhibits features of active anti-tumor immunity, such as increased infiltration of CD8⁺ T cells, T helper cells, macrophages, and mast cells—processes in which CYP19A1 and SULF1 may play a regulatory role [22, 32]. On the other hand, immunosuppressive mechanisms are also enhanced. For example, elevated GPX3 expression has been associated with the expression of immune checkpoint inhibitors [33], while CYP2A6 promotes polarization of macrophages toward the M2 phenotype, thereby reinforcing the immunosuppressive state of the TME [26]. The coexistence of such pro-tumor and anti-tumor effects creates a complex and dynamically balanced microenvironment that may contribute to the aggressiveness of the high-risk stratum. In contrast, SLC5A6 is highly expressed in the low-risk stratum, consistent with the findings reported by Chen [34]. It is hypothesized that SLC5A6 helps maintain an indolent TME in the low-risk stratum, potentially through its metabolic effects that shape an immunosuppressive microenvironment. Furthermore, a lower TIDE score in the high-risk stratum indicates that patients in this group may derive relatively greater clinical benefit from immunotherapeutic strategies, particularly immune checkpoint inhibitors.
Drug sensitivity analysis based on risk stratification revealed that BX-795 and CGP-60,474 exhibited greater therapeutic sensitivity in the high-risk stratum, findings that have also been corroborated by other studies [35, 36]. Conversely, GC in the low-risk stratum showed higher sensitivity to most targeted agents, including 17-AAG, AKT inhibitors, EKB-569, FK866, GSK690693, LY317615, and IXO2 [37–41]. The therapeutic efficacy of these agents against GC has been confirmed in previous studies. Nevertheless, the precise molecular mechanisms underlying their differential effects across risk-stratified patient subgroups remain incompletely elucidated. It is speculated that drug molecular properties, pathway-specific differences, and variations in the immune microenvironment may contribute to these observations [36].
Smoking is a well-established risk factor for GC [42, 43]. CYP2A6, which possesses strong metabolic ability toward procarcinogens, notably tobacco-specific nitrosamines, may play an additional role in the development of gastric precancerous lesions. Accordingly, CYP2A6 may also serve as a marker to identify GC subgroups with a history of smoking exposure.
Although the prediction model constructed in this study demonstrated good predictive performance in both the training and validation sets, it still has several limitations. First, this study was based on retrospective data analysis, which is subject to inherent selection bias and information bias. Future prospective studies are needed to further validate the model’s clinical applicability. Second, the relatively limited sample size, particularly in some subgroups, may affect the model’s stability and generalization ability. Although statistical methods such as cross-validation were used to mitigate the risk of overfitting, larger-scale multicenter validation is still necessary. Third, the biological interpretability of the model remains to be further explored. Although bioinformatics analysis identified signature genes and functional enrichment analysis was performed, the specific molecular mechanisms underlying these genes in the development and progression of gastric cancer require further experimental verification. Future research directions include expanding the sample size, conducting multicenter validation, integrating multi-omics data, and further exploring the biological functions of the signature genes through basic experiments to promote the translation of this prediction model into clinical practice.

Conclusion
In conclusion, this study successfully developed a clinical predictive model for GC incorporating five ADME-related genes. Moreover, the superior clinical predictive ability of this model was verified in multiple independent cohort studies. Subsequently, this study stratified GC patients according to their risk levels, offering a reference for immunotherapy regimen selection and prognostic evaluation. This model preliminarily clarified the influence of ADME-related genes on GC progression. More importantly, it emphasized their dual value as biomarkers for prognostic prediction and treatment strategy optimization. In vitro experiments confirmed that the key gene CYP2A6 serves as a new biomarker for the treatment and management of GC.

Supplementary Information
Below is the link to the electronic supplementary material.