A non-invasive secreted protein-based gene signature for prognostic stratification and tumor microenvironment assessment in gastric cancer

Data acquisition and differentially expressed gene screening

RNA sequencing profiles and clinical data for 375 GC tissues and 32 adjacent normal tissues were retrieved from The Cancer Genome Atlas Stomach Adenocarcinoma (TCGA-STAD; https://portal.gdc.cancer.gov/). An independent validation cohort (GSE15459, n = 200 GC cases) was obtained from the Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/). To focus on blood-detectable secreted proteins, a list of 731 SPCGs was sourced from the Human Protein Atlas (HPA; https://www.proteinatlas.org/humanproteome/blood+protein/secreted+to+blood) (Chen et al., 2022c). Differential expression analysis between tumor and normal tissues was performed using the “limma” R package, with significance thresholds set at |log2 fold change| > 1 and false discovery rate (FDR)-adjusted p < 0.05. SPCGs overlapping with DEGs were visualized via Venn diagram.

Prognostic signature construction

Univariate Cox proportional hazard model analysis was conducted on each SPCGs to identify genes significantly associated with overall survival (OS) in the training dataset, using a log-rank P-value of less than 0.05 as the threshold. This analysis identified 57 genes with prognostic significance for OS prediction. Subsequently, the least absolute shrinkage and selection operator (LASSO) regression method was employed to reduce the number of genes, using the “glmnet” and “survival” packages. The SPCGs risk signature for calculating individual risk scores was developed based on nonzero coefficients from the LASSO regression analysis. The formula was established as follows: risk score = Σ (coefi×expi), where coefi represents the coefficient and expi denotes the expression level of each gene.

Model validation and nomogram development

Patients were stratified into high-risk and low-risk groups based on their median risk scores derived from the training set. Differences in OS time between the groups were evaluated using log-rank tests and Cox regression analyses. Kaplan–Meier survival analysis was employed to compare survival rates between the two groups, using the “survival” and “survminer” packages in R software. The receiver operating characteristic (ROC) curve was used to evaluate the model’s predictive accuracy for 1-, 2-, 3-, and 5-year OS survival, using the “survival”, “survminer”, and “timeROC” packages. The predictive value of the prognostic model was further validated using the GEO database as a validation set, following the same methodologies. Age, sex, grade, TNM stage, and risk score were incorporated into both univariate and multivariate Cox regression analyses using the “survival” package to identify independent variables associated with OS. A nomogram based on the 8-SPCG signature was developed using the R packages “rms” and “regplot”, incorporating variables such as risk score, age, gender, grade, and stage. A calibration plot was used to validate the predictive accuracy of the nomogram for 1-, 3-, and 5-year OS in GC patients.

Functional enrichment analysis

Gene set enrichment analysis (GSEA) was performed using the “clusterProfiler” R package, based on the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway and Gene Ontology (GO) analyses. Additionally, gene set variation analysis (GSVA) was conducted using the “GSVA” R package to explore the relationship between the expression of SPCGs, risk score, and KEGG pathways.

TMB and immune microenvironment analysis

The TME landscape was comprehensively analyzed, focusing on tumor mutation burden (TMB), immune cell infiltration, immune checkpoint expression, and immune-associated scores within the TCGA-STAD database. TMB was assessed using the “maftools” package. The ESTIMATE algorithm was employed to calculate stromal, immune, and ESTIMATE scores using the “estimate” package. The composition of tumor-infiltrating immune cells within the TME was elucidated using multiple computational tools, including XCELL, TIMER, QUANTISEQ, MCPcounter, EPIC, CIBERSORT-ABS, and the CIBERSORT algorithm, along with the LM22 signature matrix file downloaded from (http://cibersort.stanford.edu/). Spearman’s correlation analysis was carried out to examine the relationship between the risk score, signature genes, immune cell composition, and immune checkpoint gene expression. The immunophenoscore (IPS) was used to predict the response to immune checkpoint inhibitors in high-risk and low-risk groups, based on the positive or negative status of programmed cell death ligand 1 (PD-L1) and cytotoxic T-lymphocyte-associated protein 4 (CTLA-4). IPS data were sourced from the Cancer Immunome Database (TCIA) (https://tcia.at/home).

Clinical samples

Clinical samples were obtained from Shaoxing People’s Hospital. This study utilized 56 formalin-fixed, paraffin-embedded GC tissue samples (52 for SERPINE1 and 47 for α-smooth muscle actin (α-SMA) staining), peripheral blood samples from GC patients, and effusion fluids. Additionally, fresh GC and paired adjacent non-tumorous tissues were collected during surgery for transcriptome sequencing, western blot and primary fibroblast isolation. The study protocol was approved by the Hospital’s Academic Ethics Committee (Approval No: 2023 Scientific research project 104-Y-01). Due to the retrospective data analysis and use of leftover clinical specimens, informed consent for collecting clinical samples and corresponding clinical data was waived. However, informed consent for use of patient biospecimens and health information was still obtained from all patients providing fresh tissues.

Cell lines and culture

Normal gastric epithelial cells (GES-1) and three human GC cell lines (AGS, HGC27, and MKN7) were used. AGS and MKN7 cell lines were obtained from Procell Life Science & Technology Co., Ltd (Wuhan, China), while the GES-1 and HGC27 cell lines were sourced from Cellverse Co., Ltd (Shanghai, China). All cell lines had the Short Tandem Repeat (STR) authentication. AGS and HGC27 cells were cultured in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10% fetal bovine serum (FBS), whereas GES-1 and MKN7 cells were maintained in Roswell Park Memorial Institute (RPMI) 1640 medium with 10% FBS. All cells were incubated at 37 °C in a humidified atmosphere with 5% CO₂.

Isolation and cultivation of CAFs and NFs

CAFs and NFs were isolated from surgically resected gastric cancer tissues and adjacent non-cancerous tissues, respectively. The tissues were minced and digested with 0.15% collagenase type I for 4–6 h. The resulting cell suspension was filtered and cultured in DMEM containing 20% FBS. The cells were further isolated and purified through differential enzymatic digestion using Accutase cell detachment solution (Gibco, Billings, MT, USA).

GC-CAF Co-culture system

An indirect co-culture was established using 0.4-µm pore-size Transwells. CAFs were plated in the upper chamber and AGS or HGC27 cells (1.5 × 10⁵) in the lower chamber. Cells were co-cultured in complete DMEM for 48 h before collection for analysis.

Establishment and validation of stable SERPINE1 knockdown and overexpression cell lines

Lentiviral vectors from Sangon Biotech (Shanghai, China) were used to establish stable SERPINE1 knockdown and overexpression cell lines. Transduction efficiency was verified by reverse transcription quantitative polymerase chain reaction (RT-qPCR) and western blot. The functional impact was then assessed using plate colony formation and Transwell assays to evaluate clonogenic and migratory capacities, respectively.

Plate colony formation assay and transwell assay

Cell clonogenicity was assessed by plating 400 cells per well in 6-well plates for 2 weeks. Resulting colonies were fixed with 4% formaldehyde, stained with 0.2% crystal violet, and quantified using an Enzyme-linked Spot Image Automatic Analyzer (CTL, USA). Cell migration was evaluated using 24-well Transwell chambers (Corning, NY, USA). GC cells (5 × 10⁴ for AGS/HGC27; 1 × 10⁵ for MKN7) in serum-free medium were seeded into the upper chamber, with complete medium as a chemoattractant in the lower chamber. After 24 h, migrated cells on the membrane were fixed, stained, and quantified. To assess CAF-induced migration, 1 × 10⁵ CAFs were plated in the lower chamber.

mRNA sequencing and RT-qPCR

mRNA sequencing was performed on three pairs of fresh GC and adjacent normal tissues by Sangon Biotech using the MGISEQ-T7 platform.

For RT-qPCR, total RNA was extracted from cells, reverse transcribed into cDNA, and amplified using SYBR Green Master Mix on a QuantStudio 5 system. Gene expression was normalized to GAPDH and calculated using the ΔΔCt method. The primer sequences for SERPINE1 were as follows: forward primer, ACTGCCCTCTATTCAACGGC; reverse primer, GGGCGTGGTGAACTCAGTAT.

Western blot analysis

Proteins were separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE), transferred to polyvinylidene fluoride (PVDF) membranes, and blocked with 5% non-fat milk. Membranes were incubated overnight at 4 °C with primary antibodies against SERPINE1 (1:1000, Abcam, Cambridge, UK) and GAPDH (1:3000, ZENBIO, China), followed by a 2-hour incubation with a goat anti-rabbit secondary antibody (1:100,000, HUABIO, China) at room temperature. Signals were detected using an ECL kit (Beyotime, China) on a Bio-Rad ChemiDoc system. The original western blot images can be found in Fig. S1.

Immunohistochemistry assay

IHC was performed on paraffin-embedded sections using standard protocols. Sections were incubated with primary antibodies against SERPINE1 (1:800, Abcam, Cambridge, UK) and α-SMA (1:800, Abcam, Cambridge, UK) overnight at 4 °C, followed by a secondary antibody (DAKO, Germany) and 3,3′-diaminobenzidine (DAB) visualization. Staining intensity was quantified using ImageJ by calculating the average optical density (AOD = Integrated Optical Density/Area) from five random fields per section.

Enzyme-linked immunosorbent assay (ELISA)

SERPINE1 levels in peripheral blood, pleural/peritoneal fluids, and culture supernatants from cell lines/fibroblasts were quantified using Enzyme-linked immunosorbent assay (ELISA) kits from MultiSciences (Hangzhou, China) according to the manufacturer’s instructions. All fluid samples were centrifuged prior to analysis.

Immunofluorescence staining

CAFs and NFs were plated on coverslips, fixed with 4% paraformaldehyde, and permeabilized with 0.1% Triton X-100. After blocking with 5% bovine serum albumin (BSA), cells were incubated with an α-SMA primary antibody (1:400, Abcam, Cambridge, UK) overnight at 4 °C, followed by an Alexa Fluor 488-conjugated secondary antibody (1:400, Solarbio, Beijing, China) for 1 h at room temperature. Nuclei were counterstained with 4′,6-diamidino-2-phenylindole (DAPI), and images were captured using a confocal microscope.

Statistical analysis

Data from clinical and cellular experiments are expressed as mean ± standard deviation (SD) and were analyzed using GraphPad Prism 10 software (San Diego, CA, USA). T test or Wilcoxon-test were used for comparisons between two groups, while one-way ANOVA was used for comparisons among multiple groups. The relationship between the IHC score of SERPINE1 and α-SMA was assessed using Spearman’s correlation analysis, considering the non-normal distribution of the variables. Outliers were determined by the box plot method or Grubbs test with an α-value ≤ 0.05 (3≥ n ≤30). All data were included and outliers were not excluded. Statistical significance was defined as a p-value below 0.05, with significance levels indicated as: * for p < 0.05, ** for p < 0.01, and *** for p < 0.001.

Identification of an 8-SPCG signature

The flow diagram of this study is shown in Fig. 1. A list of 233 differentially expressed SPCGs was compiled by intersecting the DEGs from the TCGA-STAD dataset with those encoding secreted proteins in the HPA (Fig. 2A). Among them, 57 SPCGs were found to be significantly associated with OS, as determined via univariate Cox regression analysis (Fig. 2B). Subsequent LASSO regression analysis identified an 8-gene signature comprising SERPINE1, C6, GRP, GCG, IL1F10, IGFBP1, ITIH2, and APOD (Figs. 2C, 2D). Analysis of differential expression between GC tumor tissues and normal tissues revealed upregulated SERPINE1, IL1F10, IGFBP1, and ITIH2, whereas C6, GCG, GRP, and APOD were downregulated (Fig. 2E). In summary, building upon data from the TCGA and HPA databases, an 8-SPCG signature was hereby identified through univariate Cox regression and LASSO regression analyses.

Construction of an 8-SPCG prognostic risk model

Following the identification of the SPCG signature, we employed LASSO Cox regression analysis to establish a prognostic risk model using the TCGA-STAD cohort as the training set. This model was subsequently validated through independent analysis of the GEO database. Risk scores for individual patients were computed by integrating the expression profiles of the 8-SPCG with their respective regression coefficients derived from the LASSO analysis. Using the median risk score as the stratification threshold, GC patients were effectively stratified into distinct high-risk and low-risk subgroups. Visual representation of risk score distributions revealed clear bimodal patterns in both the training cohort (Fig. 3A) and validation cohort (Fig. 3B). Survival analysis demonstrated significant divergence between risk groups, with corresponding patient survival status and risk score gradients illustrated through scatter plots in the training set (Fig. 3C) and validation set (Fig. 3D). Comparative expression profiling via heatmap visualization confirmed differential expression patterns of all 8-SPCG between risk subgroups, with consistent patterns observed in both training (Fig. 3E) and validation cohorts (Fig. 3F). Collectively, results from the training and validation cohorts provide preliminary support for the ability of the 8-SPCG-derived risk model to distinguish patient outcomes, suggesting its exploratory value for prognostic assessment in GC.

Validation the predictive value of 8-SPCG prognostic model

To comprehensively evaluate the prognostic efficacy of the 8-SPCG signature, Kaplan–Meier survival analysis was performed. The resulting curves revealed a significant divergence in OS between high-risk and low-risk groups in both the training and validation cohorts. Specifically, in the TCGA-STAD training set (Fig. 4A), patients assigned to the high-risk group exhibited a markedly shorter OS compared to their low-risk counterparts (p < 0.001), a trend that was consistently replicated in the GEO validation set (Fig. 4B, p < 0.001). These findings strongly underscore the discriminatory power of the 8-SPCG-based risk model in stratifying patients with varying survival outcomes. Subsequently, ROC curve analysis was conducted to quantitatively assess the sensitivity and specificity of the prognostic model. Using data from the TCGA database, the time-dependent ROC curves demonstrated favorable discriminatory performance, with area under the curve (AUC) values of 0.632 for 1-year survival, 0.709 for 2-year survival, 0.729 for 3-year survival, and 0.745 for 5-year survival (Fig. 4C). These results were further validated in the GEO dataset (Fig. 4D), reinforcing the robustness of the model’s predictive capabilities across independent cohorts. To identify independent prognostic factors associated with OS, univariate and multivariate Cox regression analyses were carried out, incorporating the risk score and a comprehensive set of clinicopathological variables, including age, gender, tumor grade, T stage, N stage, M stage, and tumor stage. Univariate Cox regression analysis (Fig. 4E) initially screened potential predictors, while multivariate Cox regression analysis (Fig. 4F) adjusted for confounding factors to determine independent prognostic indicators. The results of these analyses unequivocally identified age and the 8-SPCG-derived risk score as independent predictors of OS in GC patients. Leveraging these findings, a nomogram was constructed, integrating the selected clinicopathological parameters from the regression analyses to facilitate individualized OS prediction for GC patients (Fig. 4G). The nomogram’s predictive accuracy was further evaluated using a calibration plot, which demonstrated excellent agreement between the predicted and observed survival probabilities at 1, 3, and 5 years (Fig. 4H). Therefore, results from multiple analytical approaches collectively suggest that the 8-SPCG signature may hold predictive value, supporting its further investigation as a potential tool for clinical decision-making and patient management in GC.

Correlation of the risk score with TMB and immune landscape of GC

ESTIMATE algorithm was employed to explore the differences of the TME scores (stromal score, immune score and ESTIMATE score) between these two risk score groups. The results revealed that high-risk patients exhibited elevated stromal and ESTIMATE scores, indicative of an immunosuppressive phenotype, whereas low-risk patients showed immune-active features (Fig. 5A). Immune cell components analysis suggested the risk score was remarkably correlated with CAFs, hematopoietic stem cells, endothelial cells, macrophages, B cells, monocytes, myeloid dendritic cells, CD4+T cells, CD8+T cells, γδT cells, Tregs etc. Specifically, the risk score showed a negative correlation with CD8+ T cells (indicated by XCELL and CIBERSORT algorithms) and positive correlations with M2 macrophages (CIBERSORT-ABS and QUANTISEQ algorithms) and Treg cells (QUANTISEQ and CIBERSORT algorithms) (Fig. 5B), further supporting immune evasion in high-risk subgroups. Among SPCGs, APOD, SERPINE1, and GRP were specifically associated with immune cell subsets (Fig. S2). Notably, among them, CAFs emerged as the only cell type exhibiting significant correlations with all the three genes (Fig. 5C). Additionally, the risk score and key SPCGs (e.g., SERPINE1, C6, APOD) positively correlated with the majority of immune checkpoint genes (Fig. 5D), suggesting their role in immune tolerance regulation. TMB, a potential biomarker for immunotherapy response, showed an inverse correlation with risk scores (Figs. 5E, 5F). Survival analysis confirmed TMB’s prognostic value, with high-TMB patients demonstrating superior outcomes (Fig. 5G). Integration of TMB with risk scores enhanced prognostic stratification, identifying the high-TMB/low-risk subgroup with optimal survival and low-TMB/high-risk patients with the poorest prognosis (Fig. 5H). High-risk patients exhibited significantly reduced mutation frequencies in top 20 driver genes compared to low-risk counterparts (Figs. 5I, 5J). Moreover, it has been verified that IPS value, ranged from 0 to 10, was positively bound up with tumor immunogenicity and considered to be a new predictor of immunotherapy response to anti-CTLA-4 and anti-PD-1. Our results of IPS analysis revealed that the patients in the low-risk group had higher IPS values, demonstrating a better benefit potential in immunotherapy (Fig. 5K). In general, the SPCG-based risk score system deciphers TME heterogeneity in GC, synergizes with TMB for prognostic stratification, and provides a framework for personalized immunotherapy strategies.

Functional enrichment analysis

GSEA revealed distinct molecular pathway activation patterns between high- and low-risk GC patients. In the high-risk group, significant enrichment was observed in extracellular matrix (ECM)-related biological processes and pro-fibrotic signaling pathways. Specifically, GO analysis highlighted activation of digestion-related processes (GOBP_DIGESTION), collagen-containing ECM organization (GOCC_COLLAGEN_CONTAINING_EXTRACELLULAR_MATRIX), and hormone activity (GOME_HORMONE_ACTIVITY) (Fig. 6A). KEGG pathway analysis further corroborated these findings, showing pronounced activation of ECM-receptor interaction, complement and coagulation cascades, and hypertrophic cardiomyopathy pathways, suggesting stromal remodeling and pathological tissue fibrosis as key features of high-risk tumors (Fig. 6B). In contrast, the low-risk group exhibited enrichment in cellular metabolism and translational regulation. GO analysis demonstrated strong involvement in mitochondrial energy production, including NADH dehydrogenase complex assembly (GOBP_NADH_DEHYDROGENASE_COMPLEX_ASSEMBLY) and ribosomal biogenesis (GOBP_RIBOSOMAL_LARGE_SUBUNIT_BIOGENESIS), with cytosolic and organellar ribosomal subunits prominently represented (Fig. 6C). KEGG analysis revealed activation of aminoacyl-tRNA biosynthesis, the citrate cycle (TCA cycle), and DNA replication pathways, indicative of heightened metabolic and proliferative homeostasis (Fig. 6D). Mechanistically, GSVA uncovered the associations between risk scores and oncogenic signaling cascades. Pro-tumorigenic pathways including TGF-β, MAPK, and Hedgehog exhibited progressive activation with increasing risk scores (Fig. 6E). Notably, SERPINE1 within the SPCG signature emerged as a multi-pathway regulatory hub, while ITIH2 exhibited inhibitory effects. These findings collectively underscore that high-risk GC is characterized by ECM-driven stromal activation, whereas low-risk tumors maintain metabolic and translational fidelity, providing mechanistic insights into their differential clinical outcomes.

Experimental validation of SERPINE1’s clinical significance and biological function in GC

To assess the expression levels and clinical relevance of SERPINE1 in GC, we utilized the Gene Expression Profiling Interactive Analysis (GEPIA) platform (http://gepia.cancer-pku.cn/index.html). Our analysis revealed significantly higher expression of SERPINE1 in GC tissues compared to normal tissues (Fig. 7A). Kaplan–Meier survival analysis further demonstrated that elevated SERPINE1 expression was associated with poorer prognosis in GC patients (Fig. 7B). To further validate the clinical relevance of SERPINE1, we analyzed clinical specimens from our institutional cohort. Next-generation sequencing of paired post-surgical tissues demonstrated substantially higher SERPINE1 mRNA levels in tumor tissues compared to adjacent precancerous lesions (Fig. 7C; Table S1). Circulating SERPINE1 protein levels, as determined by ELISA, showed progressive elevation in advanced GC patients compared to post-operative tumor-free controls (Fig. 7D; clinical details in Table S2). Notably, malignant effusions exhibited significantly higher SERPINE1 concentrations than benign effusions (Fig. 7E; Table S3). Immunohistochemical analysis of 52 GC specimens (including 16 matched tumor-adjacent pairs) revealed pronounced SERPINE1 overexpression in malignant tissues (Figs. 7F, 7G; Table S4). Subgroup analysis demonstrated significantly elevated SERPINE1 immunoreactivity scores in advanced pathological stages (III vs. I–II), deeper tumor invasion (T4 vs. T1-3), and extensive nodal metastasis (N2-3 vs. N0-1) (Figs. 7H–7J). However, no significant associations were observed with tumor differentiation status or histological subtypes (Fig. S3). The robustness of these findings was confirmed by sensitivity analyses, which showed that all key associations remained statistically significant after excluding outliers (Fig. S4).

We assessed the basal expression of SERPINE1 by RT-qPCR and measured its secretory levels in the cell culture supernatant by western blot and ELISA in the gastric cancer cell lines and normal gastric epithelial cells used in this study. The results confirmed that cell lines with high SERPINE1 expression secreted higher levels of the protein, while those with low expression exhibited lower secretory levels (Fig. S5). To elucidate SERPINE1’s functional role, we established genetic manipulation models based on its expression patterns across cell lines. Given the lower baseline expression of SERPINE1 in AGS and HGC27 cells and its higher expression in MKN7 cells, we generated SERPINE1-knockdown MKN7 cells (shSERPINE1) and SERPINE1-overexpressing AGS/HGC27 lines (SERPINE1-OE). Successful modulation was confirmed at both transcriptional and protein levels (Fig. 7K; Fig. S6). Functional assays demonstrated that SERPINE1 overexpression significantly enhanced colony formation capacity and transwell invasion, while SERPINE1 knockdown produced opposing effects (Figs. 7L–7O).

Overall, these findings suggest that SERPINE1 may promote GC cell proliferation and invasion in vitro. Elevated SERPINE1 levels appear to be associated with advanced clinicopathological stages and poorer prognosis in GC patients, supporting its further investigation as a potential therapeutic target and biomarker for this disease.

Interplay between CAFs and SERPINE1 in GC progression

As principal components of tumor stroma, CAFs were identified using α-SMA—a canonical marker for CAFs, which is encoded by the ACTA2 gene. Immunofluorescence analysis confirmed significantly enhanced α-SMA expression in CAFs compared to NFs derived from our primary cultures (Figs. 8A, 8B). Functional characterization through Transwell assays demonstrated that CAFs substantially potentiated the invasive capacity of HGC27 and AGS cells (Figs. 8C, 8D). Clinical validation using 47 GC specimens revealed pronounced α-SMA elevation in tumor tissues versus matched adjacent normal tissues (Figs. 8E, 8F; Table S5). Subgroup analyses identified significant associations between α-SMA overexpression and advanced disease stage (III vs. I–II) as well as lymph node metastasis (N1-3 vs. N0) (Figs. 8G, 8H). However, no correlations were observed with T-stage invasion depth or histopathological subtypes (Fig. S7). Correlative analysis of 40 GC specimens demonstrated significant positive correlation between SERPINE1 and α-SMA expression (r = 0.3395, p = 0.0321; Fig. 8I), which intensified in stage III patients (r = 0.4972, p = 0.0158; Fig. 8J). Exclusion of outliers further strengthened the association between SERPINE1 and α-SMA IHC scores, yielding higher correlation coefficients (r = 0.4054 for all patients; r = 0.4972 for stage III) and greater statistical significance (p = 0.0142 and p = 0.0158, respectively; Fig. S4). These findings were corroborated by GEPIA analysis showing SERPINE1-ACTA2 co-expression in both TCGA (Fig. 8K) and GTEx cohorts (Fig. 8L). ELISA quantification revealed that CAFs secreted higher SERPINE1 compared to NFs and GC cell lines (Fig. 8M). Furthermore, when HGC27 or AGS cells were co-cultured with CAFs, the SERPINE1 concentration in the supernatant significantly exceeded the theoretical additive concentration derived from the corresponding monocultures (Fig. 8N), indicating a synergistic enhancement. Notably, co-culture experiments uncovered a dose-dependent upregulation of SERPINE1 in HGC27 and AGS cells upon CAFs exposure (Figs. 8O, 8P). Intriguingly, exogenous epidermal growth factor (EGF) stimulation—a canonical ERBB pathway activator–mimicked this effect, significantly elevating SERPINE1 expression in both cell lines (Figs. 8Q, 8R). This observation aligns with prior KEGG pathway analysis linking SERPINE1 to ERBB signaling (Fig. 6E). These findings suggest that CAFs may serve as potential contributors to SERPINE1-driven malignancy, possibly through direct secretion and EGF/ERBB-mediated paracrine mechanisms. This stromal-epithelial interaction represents an avenue worthy of further exploration for future therapeutic strategies.