Bioinformatics analysis of macrophage-associated genes reveals prognostic signatures and immune landscape in gastric cancer

Abstract

Gastric cancer (GC) remains one of the leading causes of cancer-related mortality worldwide. The interaction between macrophages and the tumor immune microenvironment (TME) plays a critical role in disease progression and patient prognosis. In this study, we conducted a comprehensive bioinformatics analysis to identify macrophage-associated prognostic genes and construct a predictive risk model in GC. Using transcriptome data from TCGA (n = 350 tumors, 31 controls) and GEO datasets (GSE84437, n = 483; GSE183904), we applied differential expression analysis (DESeq2), weighted gene co-expression network analysis (WGCNA), single-cell RNA sequencing (Seurat), and Cox-LASSO regression to screen for key prognostic markers. Three genes—GPX3, SERPINE1, and SPARC—were identified and used to build a risk score model. Patients were stratified into high- and low-risk groups. Kaplan-Meier analysis showed significantly shorter survival in the high-risk group (HR = 2.35, p < 0.001). The model achieved strong predictive performance with area under the curve (AUC) values of 0.73, 0.70, and 0.68 at 1, 3, and 5 years, respectively. Immune infiltration analysis using CIBERSORT revealed that GPX3 and SPARC were positively correlated with plasma cells and negatively with M0 macrophages. A nomogram incorporating risk score, age, and N/M stage further improved prognostic accuracy. Drug sensitivity analysis (pRRophetic) identified 27 compounds with differential predicted IC50 values between risk groups.Our study demonstrates that macrophage-associated gene signatures are robust predictors of GC prognosis. These findings provide novel insights into immune regulation and potential therapeutic targets in gastric cancer.

Introduction

Gastric cancer (GC) remains a significant global health challenge, representing the fifth most commonly diagnosed malignancy and the third leading cause of cancer-related mortality worldwide [1, 2]. Despite advances in early detection and treatment, the overall prognosis for patients with advanced GC remains poor, with a 5-year survival rate of less than 30% [3, 4]. This underscores the urgent need for more precise prognostic markers and therapeutic targets, that can aid in better patient stratification and personalized treatment approaches.

The molecular landscape of GC is highly heterogeneous, with diverse genetic and epigenetic alterations driving tumor progression and response to therapy [5]. Over the years, efforts have been made to characterize the genetic mutations and molecular subtypes of GC, leading to significant insights into key pathways such as cell cycle regulation, DNA repair, and p53 signaling [6–8]. However, despite these advancements, current clinical models do not sufficiently account for complex interactions between tumor cells and the surrounding tumor microenvironment (TME), particularly immune cells, which play a pivotal role in GC development and progression [9, 10].

Understanding the role of tumor-associated macrophages (TAMs) in the TME is crucial for advancing therapeutic strategies, particularly in gastric cancer (GC) [11, 12]. TAMs can adopt either a pro-inflammatory (M1) or immunosuppressive (M2) phenotype, and their polarization status directly influences tumor behaviors such as angiogenesis, immune evasion, and metastasis [13]. These processes are essential in determining patient prognosis, underscoring the importance of understanding macrophage polarization and their interactions with other immune cells. Despite the well-established impact of TAMs on tumor progression, the precise mechanisms underlying macrophage differentiation in GC remain unclear, limiting the development of targeted therapies [14]. Additionally, the immune microenvironment in GC has attracted significant attention, as immune checkpoint inhibitors and other immunotherapies have shown promise in treating subsets of patients [15]. Identifying prognostic immune-related biomarkers is vital for personalizing immunotherapy strategies, as they offer insights into immune cell infiltration patterns that may predict treatment responsiveness [16, 17]. Recent studies have clarified the current focus on tumor biology, the immune environment correlates, and the clinical need for biomarkers with robust basic biology [18, 19]. Given the high degree of molecular complexity, including crosstalk between immune cells and tumor cells, a deeper investigation into the regulatory networks governing these processes is essential for optimizing immunotherapy and improving patient outcomes.

Building upon these findings, our project aims to combine data from public databases and bioinformatics tools to delineate the molecular characteristics of GC, with a particular focus on the role of macrophages and other immune cells within the TME. Similar TCGA-based analytical frameworks have been successfully applied in previous studies to characterize oncogenic signaling and identify biomarkers across cancers, including oral squamous cell carcinoma and bladder cancer [20, 21]. These precedents provide methodological support for our integrative design. By integrating data from bulk RNA sequencing, single-cell RNA sequencing (scRNA-seq), and public cancer databases such as The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) datasets, this study aims to identify key prognostic gene signatures, construct a risk model for survival prediction, and elucidate the regulatory networks influencing immune cell behavior in GC. This research not only provides novel insights into the molecular subtyping of GC but also contributes to the growing body of evidence that highlights the critical importance of the immune microenvironment in cancer progression and therapy response.

Materials and methods

Data extraction

TCGA-GC dataset which included transcriptome, miRNA, lncRNA, survival information and clinical features data was obtained from TCGA database as training set. This dataset involved 350 GC tumor and 31 control tissue samples. The GC-related transcriptome dataset GSE84437 and single-cell RNA sequencing dataset GSE183904 were downloaded from the GEO. The GSE84437 dataset consisted of 483 tissue samples from patients with GC as the validation set. GSE183904 dataset included 30 GC and 10 control samples. Then 69 chemokines-related genes (CRGs) were acquired from published literature [22]. This research is not a clinical study. Ethical approval for this study was granted by the Ethics Committee of the Fourth Affiliated Hospital of Nanjing Medical University, Ethics approval number (20241024-K113).

Differential expression analysis

Differential gene expression analysis was performed on the TCGA-GC dataset (350 tumor vs. 31 control samples) using the DESeq2 package (v1.38.0). Raw count data were normalized using the default estimateSizeFactors method. Differentially expressed genes (DEGs) were identified with criteria of adjusted p-value (Benjamini-Hochberg) < 0.05 and |log2FoldChange| >0.5. Gene expression dispersion was estimated per gene using shrinkage estimators, and significance was assessed via the Wald test. A volcano plot was generated using ggplot2 (v3.4.3) and a heatmap of the top 20 DEGs was visualized using the ComplexHeatmap package (v2.14.0), ranked by absolute log2FoldChange.

Weighted gene co-expression network analysis (WGCNA)

To construct a gene co-expression network, we applied the WGCNA package (v1.72-1) to the DEGs identified from TCGA-GC. First, genes with the lowest 50% median absolute deviation (MAD) were filtered out to reduce noise. Outlier samples were identified via hierarchical clustering (hclust) using Euclidean distance and removed if the height threshold exceeded the average inter-sample distance. The soft-thresholding power (β) was selected using the pickSoftThreshold function, aiming for a scale-free topology fit index (R²) ≥ 0.85 while maintaining low mean connectivity. The network was constructed using the signed hybrid method. Modules were identified via dynamic tree cutting with a minimum module size of 50 genes and merged using a module eigengene (ME) correlation threshold of 0.75 (cut height = 0.25). Modules with significant correlation to clinical traits (Pearson’s r > 0.3, p < 0.05) were selected for further analysis.

Identification and analysis of candidate genes

Intersection of DEGs and key module genes yielded candidate genes. To understand biological functions and pathways involved in candidate genes, GO and KEGG enrichment analyses were performed using clusterProfiler package (p < 0.05) (version 4.7.1.003).

Consensus cluster analysis

To analyze the influence of various factors on GC survival, univariate and multivariate Cox regression analyses were conducted to screen for survival-related gene 1 in CRGs (p < 0.05). Based on expression of prognosis genes 1, consensus cluster analysis was performed on GC samples in TCGA-GC dataset by ConsensusClusterPlus package (version 2.6). GC samples were divided into different subtypes, and PCA was carried out in different subtypes. DEGs 2 were further selected between different subtypes.

Single cell analysis

We analyzed the GSE183904 scRNA-seq dataset using the Seurat package (v4.3.0). Initial filtering retained cells expressing > 200 and < 4200 genes and with total UMI counts between 550 and 25,000. Cells with > 10% mitochondrial gene expression were excluded to remove low-quality or apoptotic cells. The dataset was normalized using LogNormalize and highly variable genes were identified via vst method (top 2,000 genes). Principal component analysis (PCA) was conducted, and significant PCs (p < 0.05, JackStrawPlot) were used for dimensionality reduction and clustering (FindNeighbors, FindClusters). Clustering resolution was optimized using the clustree package. UMAP was used for visualization. Cell types were annotated via SingleR (v2.0.0) using the Human Primary Cell Atlas reference and confirmed with marker gene expression from the literature. Differential gene expression between tumor and control macrophages was performed using FindMarkers with thresholds of adjusted p < 0.05 and |log2FoldChange| >0.1.

Construction of risk model

Based on candidate signature genes, univariate Cox analysis was used to screen candidate prognosis genes-related survival of GC (HR ≠ 1, p < 0.05). PH assumption test was then carried out on candidate prognosis genes-related survival of GC. Furthermore, LASSO regression analysis was utilized to further select prognosis genes using glmnet package (version 4.1-4). After obtaining prognosis genes, a risk model was constructed, and riskscore of GC patients was calculated by following formula: Riskscore = × (coef indicated coefficients of prognosis genes, expr indicated expression of prognosis genes). All patients were divided into high and low risk groups relied on median value of riskscore. KM survival analysis was performed to observe survival difference of two risk groups (p < 0.05), and KM survival curve was drafted by survminer package (version 0.4.9). Then efficiency of risk model was evaluated by ROC curve which was drafted by pROC package (version 1.18.0). In addition, performance of risk model was validated in GSE84437 dataset.

Construction of nomogram

Univariate and multivariate Cox regression analysis, and PH assumption test were conducted to screen independent prognostic factors based on riskscore and clinical features (age, gender, grade, T/N/M stage). After obtaining independent prognostic factors, a nomogram was constructed to evaluate survival probability of GC patients by rms package (version 6.5-0). Ability of nomogram was validated by calibration and ROC curves.

Gene set variation analysis (GSVA)

Based on two risk groups, gsva function was utilized to calculate expression of all samples in TCGA-GC. Reference gene set c2.cp.kegg.v7.4.symbols.gmt was acquired from MSigDB. Then differential analysis was performed in different expression groups, and trends of gene set were found.

Immune infiltration analysis

To determine condition of immune infiltration in GC and control groups in TCGA-GC dataset, CIBERSORT algorithm was performed to filter out samples with poor quality (p > 0.05). Difference of immune cells between GC and control groups was compared by wilcox test (p < 0.05). Furthermore, correlation of differential immune cells with prognosis genes was analyzed by spearman correlation analysis, and correlation diagram was drafted by corrplot package (p < 0.05) (version 0.92). Finally, in the GC samples with survival information from TCGA-GC, ESTIMATE algorithm in estimate package was used to comparedifference of Stromal score, immune score and ESTIMATE score between high and low risk groups (p < 0.05).

Drug sensitivity analysis

Drug sensitivity data from the Genomics of Drug Sensitivity in Cancer (GDSC) database (https://www.cancerrxgene.org/) were used to assess chemotherapy response in GC patients in the TCGA-GC. Half-maximal inhibitory concentration (IC50) values of GC samples with survival information were estimated via “pRRophetic” package (v 0.1) [23]. The difference in IC50 of two risk groups was evaluated using the Wilcoxon test with a significance level of p < 0.05.

Regulatory network and drug prediction

To identify potential regulatory mechanism, TFs of prognosis genes were obtained from ChEA3 database. miRNAs-targeting prognosis genes were predicted by DIANA-microT and miRDB databases. Intersection of miRNAs was identified as key miRNAs. According to key miRNAs, lncRNAs-targeting key miRNAs were predicted by starBase database. Based on key miRNAs, lncRNAs and prognosis genes, a lncRNA-key miRNA-prognosis gene regulatory network was constructed. To explore potential drugs for treatment of GC, CTD was carried out. After that, cytoscape software (version 3.9.1) was utilized to visualize above network. In addition, expression of prognosis genes in TCGA-GC dataset was analyzed between GC and control groups (p < 0.05).

Cell communication and pseudotime analysis

Expression of prognosis genes in different cell subpopulations was analyzed and compared, and cells with differences between GC and control groups were defined as key cell clusters (p < 0.05). Cell communication analysis of key cell clusters was applied to enable a systematic analysis of cell-cell communication molecules using CellChat package (p < 0.05) (version 1.6.1). Pseudotime analysis of key cell clusters was performed to explore their differentiation directions using Monocle package (version 2.26.0). Single cell trajectory was constructed by DDRTree package (version 0.1.5).

Statistical analysis

R software (version: 4.2.2) was used to process and analyze data. Statistical significance of two group was performed via Wilcoxon rank-sum test. P value < 0.05 was considered statistically significant.

Results

Differential expression and co-expression analysis in GC

A differential expression analysis identified 8,387 differentially expressed genes (DEGs), including 4,117 up-regulated and 4,270 down-regulated, between the gastric cancer (GC) and control groups in the TCGA-GC dataset (p < 0.05) (Fig. 1a). Consequently, 15 co-expression modules were derived through similarity analysis, with a minimum of 50 genes set for each gene module (Fig. 1b). The intersection of the 8,387 DEGs and 6,030 key module genes yielded 1,802 candidate genes, enriched in processes such as nuclear division and chromosomal regions, with KEGG pathways involving the cell cycle, DNA replication, and p53 signaling (p < 0.05) (Supplementary Fig. 1a-c). Significantly different pathways between risk groups were explored using GSVA, and the results indicated that glycosaminoglycan biosynthesis (chondroitin sulfate), extracellular matrix (ECM)-receptor interaction, and focal adhesion were up-regulated, while base excision repair and peroxisome pathways were down-regulated (Supplementary Fig. 1d). Univariate Cox analysis of CRCs revealed that the differential expression of CXCR6, CXCL3, CXCR4, and ACKR317 influenced the survival of GC patients (Fig. 1c). Multivariate Cox regression further identified that the differential expression of CXCR6 and CXCR4 had a direct impact on GC patient survival (Fig. 1d). Therefore, a consensus cluster analysis was performed on GC samples from the TCGA-GC dataset based on the expression of CXCR6 and CXCR4. All GC patients were subsequently divided into two clusters. Kaplan-Meier (KM) survival analysis demonstrated that patients in cluster 1 exhibited a longer life span (Fig. 1e). Differential expression analysis between the clusters identified 3,364 DEGs, including 901 up-regulated and 2,463 down-regulated genes (Fig. 1f).

Fig. 1.

Identification of DEGs in the TCGA-GC datasets. a Volcano plot showing DEGs between GC and normal tissues. Red dots represent upregulated genes, green dots represent downregulated genes, and gray dots indicate non-significant genes. b Cluster dendrogram showing the grouping of genes into modules based on expression similarities using WGCNA. Different colors represent different gene modules. c Univariate Cox regression analysis of the differential expression of CXCR6, CXCL3, CXCR4, and ACKR3 on the survival of GC patients. d Multivariate Cox regression analysis of CXCR6, CXCR3, CXCR4, and ACKR3. e Kaplan-Meier survival curve showing significant differences in overall survival between cluster 1 and cluster 2. f Volcano plot illustrating the DEGs between cluster 1 and cluster 2. Red dots represent upregulated genes, and green dots represent downregulated genes. 3.3 Totally 3,364 DEGs 2 were identified

Establishment of macrophage-related core genes

Quality control of the single-cell data from the GSE183904 dataset is presented in Supplementary Fig. 2, where 110,809 cells and 25,504 genes were retained for further analysis. From the 2,000 highly variable genes, the top 10 principal components (PCs) are displayed in Supplementary Fig. 3, and the first 30 PCs were selected for subsequent analysis (Supplementary Figs. 3b and 4). Subsequently, 37 cell clusters were identified, and the differences between cell clusters in the GC and control groups were compared (Fig. 2a, Supplementary Fig. 5). A total of 12 cell subpopulations, including epithelial cells, T cells, plasma cells, B cells, natural killer (NK) cells, macrophages, fibroblasts, endothelial cells, mast cells, dendritic cells, pericytes, and chief cells, were annotated, and their expression between GC and control groups was displayed in a proportional diagram (Fig. 2b, Supplementary Fig. 6). Differential expression analysis in macrophages identified 125 DEGs, including 87 up-regulated and 38 down-regulated genes (Fig. 2c). By intersecting the 1,802 candidate genes, 3,364 DEGs, and 125 DEGs, GPX3, HBB, SPARC, SERPINE1, COL1A1, and TIMP1 were identified as candidate signature genes (Fig. 2d).

Fig. 2.

Single-cell RNA sequencing analysis focusing on DEGs in gastric cancer macrophages. a UMAP clustering plot showing 37 cell clusters identified from single-cell data, with different colors representing distinct clusters. b UMAP plot showing the 12 annotated cell subtypes, including T cells, macrophages, epithelial cells, and fibroblasts. Different colors represent distinct cell types. c Volcano plot showing DEGs in macrophages. Red dots indicate upregulated genes, and green dots indicate downregulated genes. d Venn diagram showing the overlap of genes from WGCNA, subtype, and macrophage differential expression analyses

Prognostic gene signature and risk model construction for survival prediction in gastric cancer patients

Based on univariate Cox regression, proportional hazards (PH) assumption testing, and LASSO analysis of candidate signature genes, GPX3, SERPINE1, and SPARC were selected as prognostic genes (Fig. 3a, b). A risk model was subsequently constructed to evaluate the survival outcomes of GC patients, and the risk scores of GC patients were calculated. In the TCGA-GC dataset, GC patients were divided into high- and low-risk groups based on the median risk score (Fig. 3c). The prognostic genes exhibited higher expression in the high-risk group (Fig. 3d). Kaplan-Meier (KM) survival curves showed that high-risk patients had shorter lifespans (Fig. 3e), and the area under the curve (AUC) value was greater than 0.6 at 1, 3, and 5 years (Fig. 3f). Additionally, the risk model was validated using the GSE84437 dataset, where all GC patients were similarly divided into two risk groups (Fig. 3g). The KM survival curves showed similar results to the TCGA-GC dataset (Fig. 3h), with 1, 3, and 5 years of the AUC value again exceeding 0.6 (Fig. 3i). These results collectively demonstrate that the risk model possesses a strong predictive ability for GC patient outcomes.

Fig. 3.

Survival analysis for GC patients based on the expression of key prognostic genes. a Univariate Cox regression analysis of candidate genes. b LASSO regression analysis for the selection of prognostic genes. The coefficient profiles (top) and the partial likelihood deviance (bottom) were plotted against log lambda. The optimal lambda value was chosen based on cross-validation. c Risk score distribution and survival status of GC patients in the training cohort. Patients were classified into high-risk and low-risk groups based on the median risk score. Survival status is shown below, with blue representing alive and red representing deceased. d Heatmap displaying the expression levels of the selected prognostic genes (GPX3, SERPINE1, and SPARC) in the training cohort. Red indicates high expression, and blue indicates low expression. e Kaplan-Meier survival curves for GC patients in the training cohort. f Time-dependent ROC curves for the risk model in the training cohort, assessing the model’s predictive accuracy at 1, 3, and 5 years. g Risk score distribution and survival status of gastric cancer patients in the validation cohort. h Kaplan-Meier survival curves for the validation cohort. i Time-dependent ROC curves for the validation cohort, assessing the model’s predictive performance at 1, 3, and 5 years

Integrating prognostic factors into a nomogram

Univariate Cox analysis and PH assumption test showed that risk score, age, and T/N/M stage had a significant impact on GC patients (p < 0.05) (Fig. 4a). Multivariate Cox analysis subsequently identified risk score, age and N/M stage as independent prognostic factors, and all four factors passed the PH assumption test (Fig. 4b). Based on these independent prognostic factors, a nomogram was constructed to predict the survival probability of GC patients (Fig. 4c). The slope of the nomogram’s calibration curve was close to 1, indicating that the nomogram had a reliable predictive ability (Fig. 4d). Finally, the AUC value of the nomogram was greater than 0.6 at 1, 3, and 5 years (Fig. 4e). Although we did not directly compare the nomogram’s performance with the standalone TNM staging system, the inclusion of N and M stages within the model suggests that it complements the clinical standard. The integration of the risk score further enhances predictive accuracy.

Fig. 4.

Univariate and multivariate Cox regression analyses, and the construction of a nomogram to predict the 1-year, 3-year, and 5-year survival probabilities for GC patients. a Univariate Cox regression analysis and PH assumption test results showing the impact of risk score, age, differentiation and T/N/M staging on the prognosis of GC patients. b Multivariate Cox regression analysis results, identifying age, N stage, and M stage as independent prognostic factors for GC patients. c Nomogram illustrating the impact of risk score, age, N stage, and M stage on survival prediction for GC patients. d Calibration curve for the nomogram showing the agreement between predicted and actual survival probabilities at 1 year, 3 years, and 5 years. e ROC curve for the nomogram displaying the AUC values for 1-year, 3-year, and 5-year survival predictions

Differential immune cell infiltration and prognostic gene associations

After removing low-quality samples from the TCGA-GC dataset, the relative abundance of 22 immune cell types between GC and control samples is shown in Fig. 5a. Among the 22 immune cell types, 6 types, including M0, M1, and M2 macrophages, monocytes, plasma cells, and activated memory CD4 + T cells, showed significant differences between the GC and control groups (p < 0.05) (Fig. 5b). To assess the relationship between differentially infiltrating immune cells and prognostic genes, a correlation analysis was performed. The results indicated that GPX3 and SPARC had a significant positive correlation with plasma cells (r > 0, p < 0.05), and a negative correlation with M0 macrophages (r < 0, p < 0.05) (Fig. 5c). Furthermore, the stromal score, immune score, and ESTIMATE score were higher in the high-risk group (Fig. 5d). Finally, expression analysis of the prognostic genes revealed that GPX3 was expressed at higher levels in the control group, while SERPINE1 and SPARC were more highly expressed in the GC group (Supplementary Fig. 7).

Fig. 5.

Correlation analysis between infiltration cells and prognosis genes. a Relative abundance of 22 immune cell types in GC and normal control samples. Different colors represent different immune cell types. b Violin plot depicting the significant differences in immune cell types between GC and normal control samples. c Correlation heatmap between immune infiltration cells and prognostic genes. d Violin plot comparing stromal score, immune score, and ESTIMATE score between high-risk and low-risk groups

Differences in drug sensitivity between high and low risk groups

We queried the Comparative Toxicogenomics Database (CTD) for gastric-cancer–related chemicals and chemical–gene interactions involving GPX3, SERPINE1, or SPARC, and visualized the network in Cytoscape (v3.9.1). Predicted IC50 values for 138 GDSC agents were obtained using pRRophetic (v0.5). Across 138 agents modeled from GDSC, 27 drugs showed significant differences in predicted IC50 between risk groups after BH–FDR correction. In 25/27 agents, the high-risk group exhibited higher IC50 (i.e., lower predicted sensitivity), indicating a broad chemoresistant/stroma-enriched phenotype. Only two agents—[AS601245], [XMD8-85]—showed lower IC50 in the high-risk group, suggesting pathway-specific vulnerabilities within the high-risk state.

Interactions of transcription factors, miRNAs, and LncRNAs in gastric cancer

Analysis of the TFs-prognosis gene regulatory network revealed that HOXB6, ZEB1, and others regulated SERPINE1 and SPARC, while SOX7, AHR, and others regulated SERPINE1 and GPX3. Additionally, HOXC10, TP53, and others were found to regulate SPARC and GPX3 (Supplementary Fig. 9). Based on the prognostic genes, 15 key miRNAs were predicted using two databases, and 54 lncRNAs were identified through these miRNAs. Subsequently, a lncRNA-miRNA-prognostic gene regulatory network was constructed, incorporating 3 prognostic genes, 15 miRNAs, and 54 lncRNAs. In this network, OIP5-AS1, AL050341.2, LRRC75A-AS1, and others regulated SERPINE1 via hsa-miR-1277-5p. NEAT1, MALAT1, and AC130462.1 regulated GPX3 through hsa-miR-185-5p, while TUG1, OIP5-AS1, and MIRLET7BHG regulated SPARC via hsa-miR-29a-3p (Supplementary Fig. 10). From the CTD database, atrazine, cadmium chloride, arsenite, and others were predicted to be associated with the three prognostic genes simultaneously (Supplementary Fig. 11).

Cell subpopulation-specific expression of prognostic genes in gastric cancer

To explore the expression of prognostic genes in cell subpopulations, their distribution was analyzed (Supplementary Fig. 12). The results showed that the expression of prognostic genes differed between the GC and control groups in B cells, dendritic cells, endothelial cells, epithelial cells, fibroblasts, macrophages, mast cells, NK cells, pericytes, and plasma cells, and these cells were identified as key cell types. In GC samples, GPX3, SERPINE1, and SPARC were up-regulated in macrophages, fibroblasts, and pericytes (Fig. 6a–c). Cell communication analysis indicated that macrophages had the strongest and most frequent interactions with other key cell types (Fig. 6d). Pseudotime analysis of macrophages revealed five distinct differentiation states. Macrophages were distributed across nearly all stages of both the GC and control groups. However, the number of macrophages was significantly higher in the GC group (Fig. 6e and f).

Fig. 6.

Expression of prognostic gene (GPX3, SERPINE1, SPARC) in various cell subpopulations, cell communication networks, and macrophage differentiation trajectories in GC. a–c Violin plots representing the expression levels of GPX3(a), SERPINE1(b) and SPARC(c) in various cell subpopulations. Red indicates the GC group, and green represents the normal group. d Cell communication network analysis of different cell subpopulations in GC tissue. e Pseudotime analysis showing the differentiation trajectories of macrophages in GC and normal tissues. f Expression of GPX3, SERPINE1, and SPARC along the pseudotime trajectory. The panel shows the relationship between gene expression and pseudotime

Discussion

GC is characterized by significant heterogeneity at both the molecular and microenvironmental levels, which complicates prognosis and treatment. While prior studies have highlighted the role of TAMs in shaping the tumor immune landscape, the precise molecular mediators linking macrophage phenotypes to clinical outcomes remain insufficiently defined. It is important to acknowledge that reliance on bulk transcriptomic repositories such as TCGA introduces potential biases in sample composition and technical processing. As emphasized in recent reviews, these factors can influence gene expression interpretation and must be considered when drawing biological conclusions [24, 25]. In this study, we employed integrative bioinformatics approaches—combining bulk RNA-seq, single-cell transcriptomics, WGCNA, and Cox-LASSO modeling—to identify macrophage-associated prognostic markers in GC. Our model, composed of GPX3, SERPINE1, and SPARC, robustly stratified patients into high- and low-risk groups with distinct immune features and drug sensitivity profiles.

The prognostic model is biologically anchored in the immunoregulatory and metabolic functions of its three components. GPX3 (Glutathione Peroxidase 3), an antioxidant enzyme, has dual and context-specific roles in cancer [26]. While it typically functions as a tumor suppressor by reducing oxidative stress, recent evidence indicates that GPX3 may also support M2-like polarization of TAMs by minimizing reactive oxygen species (ROS), thereby fostering an immunosuppressive environment [27]. In GC, GPX3 has been reported as a tumor suppressor, with reduced expression correlating with advanced tumor stages and poor prognosis [28]. Consistent with previous findings, our study shows that downregulation of GPX3 correlates with poor prognosis in GC, suggesting that oxidative stress contributes to tumor progression. Interestingly, the expression of GPX3 was increased in the macrophages of patients with GC. Chang et al. showed in their study that GPX3 also plays a dichotomous role in different cancers [29], and some gastric cancer patients with high GPX3 expression have poor prognosis [30]. We hypothesized that GPX3 may act as a chemotactic factor to induce to M2 polarization of macrophages, thereby promoting tumor progression.

SERPINE1 (Serpin Family E Member 1), also known as plasminogen activator inhibitor-1 (PAI-1), is involved in extracellular matrix remodeling and is frequently upregulated in various cancers, including GC [31, 32]. It has been shown to be induced by TGF-β, a key cytokine in immune suppression and fibrosis [33]. Guo et al. noted that SERPINE1 expression in tumors may be associated with immune cell infiltration and that its promotion of CD8 + T cell infiltration is only part of its role [34]. However, this partial effect on overall prognosis is not sufficient to offset other effects that may be detrimental to prognosis, such as infiltration of tumor-associated macrophages. In GC, SERPINE1 overexpression correlates with and poor prognosis, potentially via LRP1-mediated signaling that enhances M2 polarization and limits CD8 + T cell infiltration [35]. Our study shows that SERPINE1 is highly expressed in GC macrophages and associated with poor patient outcomes, possibly indicating a shift to a more tumoral supportive environment in which TAM is polarized toward an M2 phenotype.

SPARC (Secreted Protein Acidic and Rich in Cysteine), a matricellular protein, is involved in cell-matrix interactions and has been implicated in cancer progression and metastasis [36, 37]. SPARC’s interaction with integrins and growth factor receptors can alter immune cell trafficking, particularly by creating an immune-excluded or desmoplastic TME [38]. In GC, SPARC expression has been associated with tumor aggressiveness and poor outcomes [39, 40]. In this study, immune infiltration analysis revealed that GPX3 and SPARC were positively correlated with plasma cell infiltration, whereas they were negatively correlated with M0 macrophages. These findings suggest that chemokine-related genes, particularly SPARC, may influence macrophage polarization within the TME, potentially driving either pro- or anti-tumor responses.

Our findings underscore the importance of the tumor microenvironment (TME) as a determinant of GC progression. High-risk tumors exhibited elevated stromal and immune scores, indicative of a densely interactive and potentially immunosuppressive niche. This aligns with the emerging paradigm that macrophages, fibroblasts, and tumor cells communicate through extracellular vehicles (EVs) to coordinate local immune responses and metabolic programs. As discussed by Kowal et al. [41], EVs serve as couriers of oncogenic signals, transferring proteins, metabolites, and regulatory RNAs between immune cells and tumor cells to promote metabolic reprogramming and immune escape. The presence of macrophage-specific gene signatures such as GPX3 and SERPINE1 in high-risk tumors suggests that macrophage-derived EVs may be modulating tumor cell behavior by enhancing antioxidant defenses and ECM remodeling. This EV-mediated communication could also contribute to the observed resistance to chemotherapy and immune checkpoint inhibitors. Furthermore, the stromal enrichment observed in high-risk groups may reflect a fibrotic response facilitated by SPARC and TGF-β signaling, which creates a physical and immunological barrier to T-cell infiltration.

An important dimension of our findings is their relevance to cancer cell plasticity, a hallmark of tumor evolution and treatment resistance. Tumor cells can undergo phenotypic switching to transiently adopt drug-tolerant states, particularly under selective pressure from the immune system or chemotherapy. The review “Unraveling the Dangerous Duet between Cancer Cell Plasticity and Drug Resistance” articulates how phenotypic plasticity—driven by stress-adaptive pathways like PI3K-AKT, EMT, or oxidative stress response—permits tumor cells to escape pathway inhibition and reprogram transcriptionally [42]. In this light, the upregulation of GPX3 and SERPINE1 in high-risk GC may reflect an adaptive response to oxidative and immune stress, potentially enabling resistance through metabolic reprogramming and macrophage-mediated immunosuppression. The dynamic nature of macrophage polarization and tumor cell feedback loops highlights the need to consider combination therapeutic strategies that address not only the tumor cells but also the plasticity of surrounding immune and stromal components.

Recent insights into the non-classical roles of ion channels in cancer biology further contextualize our findings. Voltage-gated sodium channels (Nav), for example, influence tumor cell migration, invasion, and metastasis via regulation of membrane potential, intracellular calcium, and protease activity [43]. These bioelectrical cues may intersect with redox-sensitive gene networks such as GPX3, suggesting a feedback mechanism between oxidative stress and ion channel signaling. Moreover, ion channels may function as transducers for extracellular vesicle uptake or TGF-β responsiveness, potentially impacting genes like SPARC and SERPINE1. In Traditional Chinese Medicine (TCM), many pharmacologically active compounds exert their effects through ion channel modulation, as highlighted by Engrui [44]. This connection warrants investigation into whether natural ion channel-targeting compounds influence the macrophage-related gene expression patterns we identified in GC.

Traditional medicine, particularly TCM, represents a promising adjunct to standard GC therapy. Liu et al. emphasize TCM’s potential to regulate immune balance, inflammation, and microbial burden [45]. For instance, Chebulinic acid, a polyphenol extracted from Terminalia chebula, has been shown to disrupt Helicobacter pylori infection by binding to the CagA protein and inhibiting epithelial adhesion [46]. This mechanism may indirectly modulate TME inflammation, macrophage activation, and barrier integrity—thereby altering risk stratification and therapeutic responses. Given that H. pylori infection plays a central role in GC etiology, future studies should explore TCM-derived anti-inflammatory agents as modulators of macrophage polarization and immune infiltration, particularly in high-risk patients with TME activation.

Our findings may also converge with known oncogenic pathways, such as TOP2A overexpression, which has been implicated in genomic instability and tumor proliferation [47]. While direct interactions between GPX3, SPARC, SERPINE1, and TOP2A remain unclear, gene co-expression analysis suggests possible convergence within stress and repair signaling modules. Integrating these into a combined risk signature may improve clinical outcome prediction and therapeutic targeting. Additionally, though our current model is based on protein-coding genes, lncRNAs are increasingly recognized as key regulators in GC biology. Ghorbani et al. [48] emphasized the role of lncRNAs in modulating immune signaling, EMT, and chemoresistance. In our regulatory network analysis, NEAT1, MALAT1, and TUG1 emerged as upstream lncRNAs of GPX3, SPARC, and SERPINE1 via miRNA intermediates. This highlights the lncRNA–miRNA–mRNA axis as a potential extension of our model, and a valuable direction for future research.

This study found significant differences in the half-maximal inhibitory concentration (IC50) of 27 therapeutic drugs between high-risk and low-risk groups. ABT-888 and BIRB-0796 showed higher sensitivity in the low-risk group, suggesting they could achieve therapeutic effects at lower concentrations in low-risk disease states. Conversely, 25 drugs, including AP-24,534, AS601245, and DMOG, exhibited higher sensitivity in the high-risk group, highlighting their potential efficacy in these patients. AS601245, a JNK inhibitor, enhances the activity of PPARα and PPARγ ligands, such as clofibrate and rosiglitazone, in cancer cells. It synergistically reduces proliferation, migration, and adhesion by modulating pathways like β-PIX and STAT3 [49, 50]. Similarly, DMOG, a PHD inhibitor, stabilizes HIF to mitigate hypoxia-induced damage, as demonstrated in CKD models where it reduced renal injury and hypertension compared to rHuEPO [51, 52]. These findings from other disease contexts underscore the therapeutic potential of these drugs in high-risk patients, where their increased sensitivity may translate into improved efficacy. However, the heightened drug sensitivity in the high-risk group may also carry a greater risk of adverse effects, necessitating careful dose adjustments. These results emphasize the importance of tailoring drug treatment plans based on a patient’s risk profile while leveraging existing evidence on the broader therapeutic applications of these drugs to guide clinical decision-making.

Limitations

This study has several limitations that must be acknowledged. First, the analyses were conducted using retrospective, publicly available datasets (TCGA and GEO), which may introduce sample heterogeneity, batch effects, and variations in clinical annotations. These factors can influence the reproducibility of the gene expression signatures and immune profiling results. Second, the lack of experimental validation of the identified prognostic genes (GPX3, SERPINE1, SPARC) limits the mechanistic insight into their roles in the tumor microenvironment. This study is the absence of a direct performance comparison between the constructed nomogram and the conventional TNM staging system using metrics such as C-index or AUC difference. Future studies using larger, independent clinical cohorts should evaluate whether our integrated model offers superior predictive accuracy compared to existing clinical tools. To enhance clinical translation, future work will focus on collecting gastric cancer tissues from independent clinical cohorts. We plan to validate gene expression levels and immune correlations using quantitative PCR, Western blot, and immunohistochemistry (IHC). Moreover, multi-center, prospective studies will be designed to test the generalizability and real-world utility of the proposed risk model. Integration of this model into clinical workflows will require comparison against standard staging systems and prospective validation using survival outcomes.

Conclusion

In conclusion, this study identified and validated key prognostic genes (GPX3, SERPINE1, and SPARC) associated with macrophages in GC. The findings provide novel insights into the role of the tumor immune microenvironment in GC progression, particularly regarding the interactions between macrophages, the ECM, and tumor cells. Although this study identified key genes through machine learning algorithms and functional enrichment analysis, there was a lack of validation and evaluation of these key genes in clinical applications. Therefore, further experimental and clinical studies are necessary.

Author contributions

Rongbo Han contributed to the conception and design; Fei Wang and Xiujuan Wang contributed to data analysis and interpretation. Benxin Zhao contributed to manuscript language revision. All authors read and approved the final manuscript.

Funding

This research was supported by the “Promoting New Life” Public Welfare Project of Zhongguancun Precision Medicine Foundation, Project Number (GYLZH64).

Data availability

Data were obtained from public databases and the relevant data sets and codes are available with the first author.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of the Fourth Affiliated Hospital of Nanjing Medical University (Approval number: 20241024-K113).

Consent for publication

All authors have read and approved the final version of the manuscript and consent to its publication.

Competing interests

The authors declare no competing interests.