2.1. Data collection

The TCGA-STAD dataset functioned as the training cohort, providing transcriptomic, prognostic, and clinical data for gastric cancer (GC) from The Cancer Genome Atlas database (https://portal.gdc.cancer.gov/) (accessed March 20th, 2025). From TCGA-STAD, 448 samples were collected, including 420 tumor tissue samples from GC (361 samples with survival information) and 35 normal tissue samples [14]. In addition, the GSE66229 dataset (GPL570 platform) was downloaded from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) to serve as the validation set. This dataset comprised 299 gastric cancer tissue samples with corresponding survival information [14].

To systematically identify tolerogenic dendritic cell-related genes, we further collected four independent tolDCs datasets from the GEO database:

1. tolDCs dataset 1 (GSE23371, GPL570 platform) comprised 3 samples of activated tolerogenic monocyte-derived dendritic cells (moDCs) (experimental group) and 3 control tolerogenic moDCs samples.
2. tolDCs dataset 2 (GSE52894, GPL10558 platform) comprised 5 lipopolysaccharide-tolerized moDCs samples (experimental group) and 4 control tolerogenic moDCs samples.
3. tolDCs dataset 3 (GSE56017, GPL570 platform) comprised 8 tolerogenic moDCs samples (experimental group) and 7 mature moDCs samples (control group).
4. tolDCs dataset 4 (GSE182528, GPL570 platform) comprised 5 samples of in vitro-induced dexamethasone-tolerized dendritic cells (experimental group) and 5 samples of in vitro-induced moDCs (control group).

Finally, the single-cell RNA sequencing dataset GSE183904 (Platform: GPL24676) was obtained from the GEO database. This dataset comprised 36 samples, including 10 normal gastric tissues and 26 gastric cancer tissues. All samples were included in the single-cell analysis for this study. The data access date was October 14, 2025.

2.2. Differential expression analysis in the tolDCs-related genes (tolDCs-RGs)

To systematically identify genes significantly associated with tolerogenic dendritic cells (tolDCs), differential expression analysis was performed across the four independent tolDCs datasets. The R package “limma” [15] was utilized to identify differentially expressed genes (DEGs) between the experimental and control groups in each dataset. The filtering criteria were set at an absolute log2 fold change (FC) > 2 and an adjusted P-value < 0.01. This analysis sequentially identified four distinct gene sets: tolDCs-DEGs1, tolDCs-DEGs2, tolDCs-DEGs3, and tolDCs-DEGs4. Subsequently, these four sets of DEGs were merged, and duplicate genes were removed to obtain the final set of tolDCs-RGs. The volcano maps of 4 groups of DEGs were plotted with ‘ggplot2’ (v 3.5.1) in R (v 3.5.1) [16] to visualize these DEGs, and the heatmaps of the expression of 4 groups of DEGs were plotted using the R package “ComplexHeatmap” (v 2.16.0) [17], with labeling and displaying top 10 up- and down-regulated genes sorted based on the differential ploidy log₂FC.

2.3. Obtainment of candidate genes associated with tolDCs-RGs in GC

To attain the DEGs comparing GC samples and control samples in the training cohort, the R package “DESeq2” (v 3.56.2) [18] was used under the threshold of |log2FC| > 2, P-value < 0.01. The volcano of the top 10 DEGs was drawn via the R package “ggVolcano” (v 0.0.2) [19]. A heat map visualizing the top 10 differentially expressed genes ranked by log₂FC from largest to smallest was drawn by using R packages “ComplexHeatmap” (v 2.16.0) [17]. The R package “ggvenn” (v 0.1.9) (https://CRAN.R-project.org/package=ggvenn) was employed to derive the intersection of the DEGs and the tolDCs-RGs. In this way, the candidate genes related to tolDCs-RGs in GC were identified.

2.4. Enrichment analysis of candidate genes

To systematically uncover the biological roles and pathway networks linked to candidate genes, we performed functional enrichment analyses through two complementary approaches. Pathway enrichment analysis based on the kyoto encyclopedia of genes and genomes (KEGG) database was performed to systematically detect metabolic and signaling pathways with significant enrichment. Gene Ontology (GO) functional annotation analysis was employed to systematically characterize the molecular roles of candidate genes. The GO analysis encompassed three primary categories: biological process (BP) to identify cellular and physiological processes, cellular component (CC) to determine subcellular localization and protein complexes, and molecular function (MF) to characterize biochemical activities and binding properties. These comprehensive functional analyses were undertaken through the R package “Clusterprofiler” (v 4.15.0.3) [20] with statistical significance determined at P < 0.05.

To investigate protein-protein interactions and construct functional networks at the molecular level, candidate genes were systematically submitted to the Search Tool for the Retrieval of Interacting Genes (STRING) database (http://string-db.org), a comprehensive platform for protein association networks. The analysis was performed with a confidence level threshold of 0.4 to ensure reliable interaction predictions while maintaining network complexity. The resulting protein-protein interaction (PPI) network, including both direct physical interactions and functional associations, was subsequently imported and visualized using Cytoscape software (v 3.1.0) [21], allowing for detailed network topology analysis and identification of key hub proteins and functional modules within the interaction network.

2.5. Identification of prognostic genes

Multiple analytical approaches were implemented to screen prognostic genes with prognostic significance from candidate genes and build a risk model. First, the previously obtained candidate genes were subjected to univariate Cox regression analysis (P < 0.05, Hazard Ratio (HR) ≠ 1) with proportional hazards (PH) assumption test (P > 0.05) via “survival” package (v 3.7−0) [22]. The genes that met these criteria were recorded as potential prognostic genes. The results were visualized by drawing a forest plot with the R package ‘forestplot’ (v 3.1.5) [23]. Least absolute shrinkage and selection operator (LASSO) regression analysis was carried out for potential prognostic genes using the R package “glmnet” (v 4.3.2) in the training set to identify candidate prognostic genes [24] (family = cox, nlambda = 1,000, alpha = 1, nfolds = 10). Genes were finally selected based on lambda. min and regression coefficients not equal to 0. Ultimately, a multivariate Cox proportional hazards model was developed via the coxph function from the ‘survival’ package (v 3.7−0), with the PH assumption validated by a statistical test (P > 0.05). The model was optimized through stepwise regression analysis using the step function. The final model coefficients were used for prognostic genes with statistical significance.

2.6. Construction of risk model

The prognostic genes were employed to develop the risk prediction model for GC. The risk score was computed in the TCGA-STAD dataset using the formula:

[Note: For each prognostic gene, the transcriptome expression value is represented by expi, while the coefficient in the stepwise analysis is indicated by β]. Based on comprehensive survival data and clinical follow-up information, GC patients from TCGA-STAD and validation datasets were systematically categorized into high-risk group (HRG) and low-risk group (LRG) according to the optimal risk score threshold determined through statistical analysis.

Survival analysis was performed through the Kaplan-Meier estimation approach implemented in the R package “survival” (v 3.7−0) for assessing survival disparities across different cohorts, employing log-rank testing to determine statistical significance (P < 0.05). Risk score distribution curves and scatter plots were constructed to display the patterns of risk stratification and clinical outcome representation in each risk-defined patient subgroup.

To evaluate the discriminative performance of the prognostic model, time-dependent receiver operating characteristic (ROC) analysis was performed using the R package “timeROC” (v 0.4). This analysis covered clinically meaningful time points (1, 3, and 5 years). To ensure the reliability of the results, the bootstrap method was employed to calculate confidence intervals. The area under the curve (AUC) was computed to assess the predictive ability (AUC > 0.6).

2.7. Construction of regulatory networks of GC

To explore the upstream regulatory mechanisms of the prognostic genes, we constructed both transcription factor-mRNA (TF-mRNA) and microRNA-mRNA (miRNA-mRNA) regulatory networks. Potential transcription factors (TFs) of prognostic genes were found by the Human Transcription Factor Targets database (htfTARGET, http://bioinfo.life.hust.edu.cn/hTFtarget#!/). After conditional filtering, TFs that appeared more than 5 times in the htfTARGET database were selected as the final results. A TF-mRNA molecular regulatory network was assembled utilizing Cytoscape software (v 3.1.0).

The mRNA-microRNA (miRNA) interactions of the prognostic genes were forecasted with the R package “multiMiR” (v 1.24.0) [25]. This R package contains 8 relevant databases: DIANA-microT-CDS (http://www.microrna.gr/microT), ElMMo, MicroCosm (https://www.ebi.ac.uk/enright-srv/microcosm/htdocs/targets/v5/), miRanda, miRDB (http://www.mirdb.org/), PicTar (http://pictar.mdc-berlin.de/), PITA (http://genie.weizmann.ac.il/pubs/mir07/mir07_data.html/), and TargetScan (http://www.targetscan.org/). Finally, the miRNAs predicted in four of the databases for the prognostic genes were selected as the final miRNAs, and the miRNA-mRNA regulatory network was assembled through the use of the Cytoscape software (v 3.1.0).

2.8. Clinical feature analysis

In the clinical characteristic analysis, to investigate the correlation between the risk score and various clinical parameters of gastric cancer patients, this study utilized data from GC patients with complete clinical information in the TCGA-STAD dataset. The risk scores were compared across subgroups of different clinical parameters. Specifically, age was stratified based on the median value (67 years) of the entire cohort into two groups: > 67 years and ≥67 years, to ensure balanced sample sizes between groups and enhance the statistical power. Other clinical features were categorized according to their standard classification systems, including gender (male, female), pathological T stage (T1, T2, T3, T4), N stage (N0, N1, N2, N3), and M stage (M0, M1). The Wilcoxon test (P < 0.05) was applied to compare risk score variations among subgroups with distinct clinical characteristics, thereby investigating risk score associations with GC patient clinical features. And the box plot was drawn employing the R package “ggplot2” (v 3.5.1) to illustrate the results.

2.9. Immune infiltration analysis

First, to examine the infiltration levels of immune cells in the HRG and LRG, the relative abundance of 28 immune cells [26] in both groups was assessed using the R package “GSVA” (v 1.53) [27]. Then, the Wilcoxon test was conducted with the R package “rstatix” (v 0.7.2) [28] to detect divergences in immune cell abundance when comparing HRG with LRG, and to identify significantly different immune cells (P < 0.05). Subsequently, Spearman’s rank correlation analysis was performed on the differentially expressed immune cells using the R package “psych” (v2.1.6) [29]. The p-values were adjusted using the Benjamini-Hochberg method, with significance thresholds set at a correlation coefficient absolute value |cor| > 0.30 and an adjusted P-value < 0.05. Then, integrating the prognostic gene set and differentially abundant immune cell populations from TCGA-STAD, the relationship between them was evaluated using the Spearman analysis, and heatmaps were constructed by means of the R package ‘ggplot2’ (v 3.5.1) (|cor| > 0.30, adj.P < 0.05).

2.10. Gene set enrichment analysis (GSEA)

To identify the underlying biological functions and pathways that differed between the high-risk and low-risk groups, we performed Gene Set Enrichment Analysis (GSEA) on the samples from both groups. Comparative analysis of HRG and LRG in TCGA-STAD was analyzed using the R package “DESeq2” (v 3.56.2), a robust statistical framework specifically designed for RNA-sequencing count data normalization and differential expression detection. Following differential expression profiling, the resulting gene list was systematically arranged in descending order based on log2 fold change values to establish a pre-ranked gene signature for subsequent pathway enrichment analysis. The genes were sorted (ranked by log₂FC from high to low) and then enriched using the GSEA function in the R package “clusterProfiler” (v 4.15.0.3) [20]. which enables comprehensive functional annotation and pathway enrichment assessment. The reference gene set collection, designated as “c2.cp.kegg.v7.4.symbols.gmt”, was obtained from the Molecular Signatures Database (MSigDB) portal (https://www.gsea-msigdb.org/gsea/index.jsp). Stringent selection criteria were applied to identify statistically significant enriched pathways, including adj.P < 0.05 for multiple testing correction and Normalized Enrichment Score (|NES|) > 0 to ensure biological relevance. The most significantly enriched pathways (top 5) were visualized and presented through sophisticated graphical representations generated using the R package “ggplot2” (v 3.5.1), providing clear visual interpretation of the functional enrichment results.

2.11. Drug sensitivity analysis

To probe the therapeutic effects of drugs on patients in HRG and LRG, 297 commonly used chemotherapeutic agents for gastric cancer were retrieved from the Genomics of Drug Sensitivity in Cancer (GDSC) database (https://www.cancerrxgene.org). In the TCGA-STAD dataset, by utilizing the R package “pRRophetic” (v 0.5) [30], the IC50-based drug response predictions for the standard chemotherapeutic agents in HRG and LRG were computed. Furthermore, IC50 differences across risk differences across groups were assessed via the Wilcoxon rank-sum test (P < 0.05). Generally, the IC50 value was negatively correlated with drug sensitivity. Specifically, the lower the IC50 value, the stronger the inhibitory effect of the drug can be exerted at a lower concentration, and the higher the drug sensitivity; conversely, the higher the IC50 value, the lower the drug sensitivity. Spearman’s correlation was employed to evaluate the association between the risk score and drug sensitivity, with Benjamini-Hochberg correction applied (|correlation (cor)| > 0.30, adj.P < 0.05).

2.12. Expression analysis

To examine prognostic gene expression patterns in GC patients versus normal controls, the Wilcoxon rank test (P < 0.05) was applied to determine prognostic gene expression disparities in GC compared to normal groups in the training datasets, with box plots generated via “ggplot2” (3.5.1) for visualization.

2.13. Single-cell RNA sequencing data characterized the gene expression patterns of cells within the gastric cancer microenvironment

To characterize gene expression patterns within the gastric cancer microenvironment at single-cell resolution, a systematic analysis was performed using the GSE183904 dataset from the GEO database and the Seurat software package (v5.1.0) [31]. Data quality control was first conducted by calculating metrics including gene counts, cell counts, and the percentage of mitochondrial genes. Genes covered by fewer than 30,000 cells were filtered out, and high-quality cells retaining between 600 and 4,000 expressed genes with less than 20% mitochondrial gene content were preserved. Violin plots generated by the VlnPlot function were used to visualize the quality metrics before and after filtering.

During the dimensionality reduction and clustering phase, data integration was performed using the IntegrateData function to mitigate batch effects. Normalization was carried out via the NormalizeData function with a scale factor of 10,000. The FindVariableFeatures function with the “vst” method was then employed to identify the top 2,000 highly variable genes (HVGs) for downstream analysis. Principal component analysis (PCA) was conducted using the RunPCA function based on these HVGs. The Jackstraw function with 100 permutations (P < 0.05) was applied alongside the ElbowPlot to determine the optimal number of principal components. Subsequently, a k-nearest neighbor graph was constructed using the FindNeighbors function, and cell clustering was performed with the FindClusters function at a resolution of 0.4. Finally, Uniform Manifold Approximation and Projection (UMAP) was used for nonlinear dimensionality reduction, and cell distribution patterns were visualized with the DimPlot function.

For cell type annotation, the FindAllMarkers function was first utilized to identify cluster-specific differentially expressed genes (thresholds: logfc.threshold = 0.25, min.pct = 0.25, only.pos = TRUE). Marker genes from the literature [32] were also referenced. Cell subpopulations were systematically annotated using the SingleR package (v2.4.0) and the CellMarker 2.0 database. Dot plots were generated to display marker gene expression patterns, UMAP plots were used to present the annotation results, and bar plots were created to compare the distribution of different cell types between the normal and gastric cancer groups.

Finally, based on the cell annotation results, the expression characteristics of prognosis-associated genes across various cell types were analyzed. Wilcoxon tests were employed to compare the expression differences of key genes between the normal and gastric cancer groups, and box plots were generated for visualization. Particular attention was paid to the expression patterns of key genes in dendritic cells, providing single-cell-level evidence for a deeper understanding of the biological functions of these genes within the gastric cancer microenvironment.

2.14. Statistical analysis

All statistical analyses were performed in the R programming environment (v 4.2.2), which provides comprehensive support for data processing, statistical modeling, and result visualization. The specific analytical methods and corresponding tools were as follows: differential expression analysis was conducted using the “limma” package (v 3.56); gene function and pathway enrichment analyses were performed with “ClusterProfiler” (v 4.15.0.3); univariate and multivariate Cox regression analyses were carried out via the “survival” package (v 3.7−0); group comparisons were assessed using the Wilcoxon rank-sum test, implemented with the “rstatix” package (v 0.7.2); correlations between variables were evaluated by Spearman’s rank correlation analysis using the “psych” package (v 2.1.6); in drug sensitivity analysis, the “pRRophetic” package (v 0.5) was utilized to predict the IC50 of chemotherapeutic agents. Unless otherwise specified, all hypothesis tests were two-sided, and statistical significance was defined as P < 0.05.

2.15. Ethics approval and consent to participate

Not applicable.

3.1. Identification of tolDCs-RGs

To obtain a comprehensive set of tolerogenic dendritic cell-related genes, differential expression analysis was performed across four independent datasets. The results showed that 553 (179 down-regulated and 374 up-regulated), 607 (260 down-regulated and 347 up-regulated), and 15 (1 down-regulated and 14 up-regulated) DEGs were identified in tolDCs-DEGs1, tolDCs-DEGs2, and tolDCs-DEGs3, respectively (Figs 1A and S1A-B), while no significant DEGs were detected in tolDCs-DEGs4 (S1C Fig). By integrating the DEGs from the first three datasets and removing redundancies, a final set of 1003 tolDCs-RGs (S1 Table) was established, providing a critical foundation for subsequent analyses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Identification and enrichment analysis of candidate genes.

(A) Volcano map and expression heat map of differentially expressed genes of ToldC-DEGS1. (B) DEGs distribution volcano maps of GC. (C) DEGs expression heat maps of GC. (D) 46 candidate genes associated with tolDCs in GC were identified. (E) Association of candidate genes with biological processes, cell components, and molecular function. (F) The top 10 KEGG pathways with significant enrichment of candidate genes. (G) The network interaction of candidate genes at the protein level.

https://doi.org/10.1371/journal.pone.0343688.g001

3.2. Identification and enrichment analysis of candidate genes

To identify candidate genes associated with the function of tolerogenic dendritic cells in gastric cancer, 1,464 differentially expressed genes (DEGs) were screened between the GC and control groups. Among these, 680 genes were up-regulated and 784 were down-regulated in the GC group (Fig 1B–1C). The intersection of the above DEGs and tolDCs-RGs was taken to obtain the 46 candidate genes related to tolDCs in GC (Fig 2D). Further functional enrichment analysis of these candidate genes revealed that the GO analysis yielded 510 biological processes, 20 cellular components, and 67 molecular functions that were significantly enriched (P < 0.05). (S2 Table), of which the top 5 most significant pathways (ranked by P value from lowest to highest) in each section were displayed. Specifically, for BP, the top pathways included chemokine-mediated signaling pathway, response to chemokine, etc; for CC, included specific granule, specific granule membrane, cluster of actin-based cell projections, and so on; for MF, included CXCR chemokine receptor binding, chemokine activity, and so on. (Fig 2E). Furthermore, the candidate genes confirmed substantial enrichment in 18 KEGG pathways (P < 0.05) (S3 Table). The top results included IL-17 signaling pathway, Toll-like receptor signaling pathway, and cytokine-cytokine receptor interaction, among others (Fig 2F). Subsequently, we investigated the protein-protein interactions. After excluding 12 unconnected target genes, the PPI network results showed that the proteins encoded by the 34 remaining genes shared 116 connecting edges. Among these, CXCL8 exhibited the most prevalent associations at the protein level with other genes like SCIN and GDF15 (Fig 2G). These results suggested that the identified candidate genes might be linked to immune function, thereby providing clues for further exploration of the mechanisms by which tolerogenic dendritic cells function in the gastric cancer microenvironment.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Potential prognostic genes were determined by univariate Cox regression analysis, LASSO regression analysis, and stepwise regression analysis.

(A) The forest map of 12 potential prognostic genes obtained by univariate Cox regression analysis. (B) LASSO regression analysis was performed on 12 potential prognostic genes. (C) The forest map of 5 prognostic genes obtained by stepwise regression analysis was finally obtained.

https://doi.org/10.1371/journal.pone.0343688.g002

3.3. Screening of prognostic genes

To construct a prognostic risk assessment model for gastric cancer, we performed a multi-stage screening process based on 46 candidate genes. First, 12 potential prognostic genes significantly associated with survival were identified through univariable Cox regression (HR ≠ 1 and P < 0.05) in combination with the proportional hazards assumption test (P > 0.05) (Fig 2A). The 8 LASSO genes were obtained by LASSO (RNASE1, PDK4, CXCL1, GPX3, INHBA, NFE2L3, ASCL2, CD36) (Fig 2B) at log(lambda.min) = 0.019. The 5 prognostic genes (RNASE1, CXCL1, GPX3, INHBA, and ASCL2) were finally obtained by multivariate stepwise regression (Fig 2C). Based on the expression levels of these genes and their regression coefficients, the following risk score formula was established: Risk Score = (0.11550940) × RNASE1 + (−0.9954886) × CXCL1 + (0.15820155) × GPX3 + (0.12561823) × INHBA + (0.08923591) × ASCL2. This laid the foundation for the subsequent construction of the risk model.

3.4. Construction of risk assessment

Based on the determined risk score, we first stratified the gastric cancer patients in the TCGA-STAD training set. Using the optimal cut-off value of the risk score (0.9974999) as the threshold, the GC patient samples in the TCGA-STAD training set were classified into a high-risk group (n = 198) and a low-risk group (n = 163). Fig 3A demonstrated the survival times and risk scores for each sample in the HRG and LRG; the results showed an obvious excess of deaths in the HRG over the LRG. The K-M survival curves demonstrated that individuals in the HRG had significantly poorer survival compared to those in the LRG (Fig 3B). This strongly suggested that there was a strong link between risk grouping and survival outcomes. The ROC curves for survival at 1, 3, and 5 years showed that AUC values were > 0.6 in 1, 3, and 5 years, suggesting that the model’s ability to discriminate between survival statuses at the 1, 3, and 5-year time points was excellent (Fig 3C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The construction and evaluation of risk assessment.

(A) Survival status of GC patients in high and low risk groups in TCGA-STAD training set. (B) Survival curve of GC patients in high and low risk groups in TCGA-STAD training set. (C) ROC curve of 1-, 3-, and 5-year survival rates in the TCGA-STAD training set. (D) The survival status of GC patients in the high and low risk groups in the validation set. (E) Survival curve of GC patients in high and low risk groups in the validation set. (F) ROC curve of 1-, 3-, and 5-year survival rates in the validation set.

https://doi.org/10.1371/journal.pone.0343688.g003

To evaluate the model’s robustness, we conducted an identical analysis on the independent validation set GSE66229. Each patient in the validation set received a risk score calculated through the same formula, leading to the classification of GC patients into HRG (53) and LRG (246) using the best cutoff value (0.6149928). There was a statistically significant disparities between HRG and LRG for survival status in GC patients (Fig 3D). The K-M survival curves in the validation set had the same effect as the training set; the LRG maintained a relatively high survival probability over a longer period (P < 0.0001) (Fig 3E). Finally, the ROC curves for the OS of GC patients at 1, 3, and 5 years, and the AUC values were all above 0.6, suggesting that the prediction accuracy was satisfactory (Fig 3F). These results collectively demonstrate that the five-gene-based risk model could effectively distinguish the prognostic risk of gastric cancer patients to a certain extent, and possessed a degree of reliability and generalizability.

3.5. Exploring the potential regulatory mechanisms of GC

We analyzed the potential transcriptional factors (TFs) and miRNA regulatory network using the hTFtarget database. The results identified 18 TFs that may regulate these genes, such as KDM5B, FOXA2, and GATA3. Among them, 5 TFs corresponded to RNASE1, 7 to ASCL2, 12 to INHBA, 7 to CXCL1, and 9 to GPX3 (Fig 4A). At the post-transcriptional level, we further predicted miRNAs targeting these genes. The results indicated that 86 miRNAs might be involved in the regulation of these prognostic genes. The distribution included 1 miRNA for ASCL2, 15 for GPX3, 23 for CXCL1, 2 for RNASE1, and 45 for INHBA (Fig 4B). Examples include hsa-miR-1233-3p for INHBA and hsa-miR-582-5p for CXCL1. These findings shed initial light on the multi-level regulation that these five prognostic genes may be subjected to, providing direction for a deeper understanding of their molecular mechanisms in gastric cancer progression.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. The potential regulatory mechanisms of GC and the correlation of risk score in different clinical features.

(A) TF-mRNA network. (B) miRNA-mRNA network. (C) Risk score and age difference analysis. (D) Risk score and gender difference analysis. (E) Analysis of the difference between risk score and M stage. (F) Analysis of the difference between risk score and N stage. (G) Analysis of the difference between risk score and T stage.

https://doi.org/10.1371/journal.pone.0343688.g004

3.6. Clinical feature

The P-value of 0.018 (P < 0.05) for age group less than 67 years and age group more than 67 years indicated that risk scores differed significantly across age groups, while the risk score value for gender was 0.56, suggested that there was no significant disparity between genders (male and female) (Fig 4C-4D).

Moreover, significant risk score disparities were observed between some of the levels in the M, N, and T stages (P < 0.05). The risk score value of 0.013 between M0 and M1 demonstrated significant differences between these groups, similarly, there was a notable disparity among N1 and N2 (P = 0.024), T1 and T2 (P = 0.00088), T3 (P = 0.0042), T4 (P = 6.4 × 10⁻⁵⁾ (Fig 4E-4G). This suggested that the clinical stage was potentially important in assessing the risk and could assist in determining the degree of risk of the patient.

3.7. Exploring immune infiltration differences

The distribution of immune cells was illustrated through heatmap visualization (Fig 5A). Employing the Wilcoxon test for statistical analysis, 20 significantly different immune cells were identified with statistical significance set at adj.P < 0.05. In particular, activated B cells in HRG demonstrated substantially elevated infiltration levels (P < 0.0001), while activated CD4 T cells in LRG showed markedly enhanced infiltration (P < 0.01). According to the box plot analysis results, there were notable marked variations in the tumor-infiltrating immune cell patterns between the two groups. The box-and-whisker plot visualization indicated that the degree of cell infiltration was predominantly higher in the HRG compared to the LRG (Fig 5B). Via correlation analysis methodology, the majority of differential immune cells showed statistically significant positive associations (cor > 0.3, with significance threshold at P < 0.05). As an illustrative example, immature B cells and activated B cells exhibited a strong positive correlation (cor = 0.92, achieving statistical significance at P < 0.001) (Fig 5C). These results revealed a complex interrelationship among differential immune cells, which might reflect changes in the GC immune microenvironment.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Difference of immune infiltration.

(A) Heat map of immune cell infiltration in high and low risk groups of patients. (B) Difference analysis of immune cells in GC high and low risk group samples. (C) Correlation analysis between different immune cells. (D) Correlation analysis between differential immune cells and prognostic genes.

https://doi.org/10.1371/journal.pone.0343688.g005

The association of prognostic genes with differential immune cells in the training set was evaluated by Spearman correlation analysis, with bubble plots illustrating the correlation coefficients and significance levels. The findings indicated that GPX3 demonstrated a direct positive relationship with the majority of cells, encompassing regulatory T cells. ASCL2 was weakly correlated with most immune cells (Fig 5D). These results suggested a complex correlation between prognostic genes and immune cells, which might reveal their potential interactions in the GC microenvironment.

3.8. GSEA result

To explore the biological characteristics of the different risk groups, we performed GSEA to examine the pathway enrichment between the high-risk and low-risk groups.GSEA showed that HRG was enriched in a total of 30 pathways. Five pathways were significantly enriched, including hypertrophic cardiomyopathy and dilated cardiomyopathy, among others (Fig 6A) (S4 Table). LRG was enriched in a total of 17 pathways. The top 5 significantly enriched pathways were the peroxisome pathway, pyrimidine metabolism pathway, RNA polymerase pathway, and steroid hormone synthesis pathway (Fig 6B) (S4 Table).The divergent enrichment of pathways between the two groups suggested distinct biological characteristics in high-risk versus low-risk gastric cancer patients: the high-risk group was associated with certain cardiovascular disease pathways, whereas the low-risk group exhibited more active metabolic and biosynthetic processes. These findings provide new clues for understanding the molecular mechanisms underlying the different risk subgroups.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. GSEA enrichment analysis and drug sensitivity analysis.

(A) The top five most enriched KEGG pathways in high-risk group. (B) The top five most enriched KEGG pathways in low-risk group. (C) The difference between the top 5 drugs with IC50 up-regulation in the high (red) and low (blue) risk group. The difference between the top 5 drugs with IC50 down-regulation in the high and low risk group. (D) Correlation between risk score and drugs.

https://doi.org/10.1371/journal.pone.0343688.g006

3.9. Exploration of the effectiveness of drug therapy

Drug sensitivity analysis results showed that the difference in IC50 between HRG and LRG for each drug; the lower the IC50, the greater the sensitivity to the patient. The top 5 drugs with down-regulation of IC50 and the top 5 drugs with up-regulation of IC50 were collated separately (P < 0.05). BIBW2992, GW.441756, and VX.680, etc., 5 drugs had higher IC50 in the HRG than in the LRG, and there were significant differences between all 5 drugs in both the HRG and LRG (P < 0.0001). While dasatinib, DMOG, etc., 5 drugs had higher IC50 in the LRG than in the HRG (P < 0.0001) (Fig 6C). BIBW2992 displayed a significant positive association with risk score (cor > 0.3, P < 0.05) by Spearman correlation analysis (Fig 6D). These findings suggested that the classification based on the risk model could provide a reference for drug selection in gastric cancer patients, and that patient groups with distinct risk profiles might be suited for differential treatment strategies.

3.10. Single-cell RNA sequencing data were analyzed and expression differences of key genes across different cell types were identified

Single-cell RNA sequencing analysis of the GSE183904 dataset yielded 134,563 high-quality cells after quality control (retention rate 84.8%) (S2A-B Fig), with the gene expression matrix remaining intact. Through identification of 2,000 highly variable genes and UMAP dimensionality reduction clustering (S2C-F Fig), cells were successfully partitioned into 22 distinct subpopulations, which were further annotated into 11 cell types including T cells, dendritic cells, and epithelial cells (S2G-I Fig). Cell proportion analysis showed epithelial cells as the most abundant population, while dendritic cells showed an increased proportion in the gastric cancer group, suggesting their potential involvement in tumor microenvironment regulation (S2J Fig).

Analysis of the key gene expression patterns (Fig 7) showed that the expression levels of INHBA, ASCL2, and CD36 in dendritic cells of the gastric cancer group were significantly upregulated. Such significant changes in expression indicate that these genes may play important roles in the regulation of dendritic cell function. In addition, ASCL2 and CD36 were specifically overexpressed in chief cells, while ASCL2 and INHBA were significantly upregulated in neutrophils. These results suggest that although the prognostic signature we constructed was derived from the screening of tolDC - related genes, its role in the gastric cancer microenvironment may not be limited to dendritic cells, but rather is achieved through the coordinated interaction of multiple cell types. These findings validate the expression specificity of prognostic genes at the single – cell resolution, especially suggesting the potential regulatory functions of INHBA, ASCL2, and CD36 in gastric cancer, thus providing a new perspective for understanding how these genes affect gastric cancer progression through the regulation of the immune microenvironment.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Expression differences of key genes across different cell types.

The blue color represents the normal control group, and the red color represents the gastric cancer group. ns, not significant; *, p < 0.05; **, p < 0.01; ***, p < 0.001.

https://doi.org/10.1371/journal.pone.0343688.g007