A stromal-derived five-gene signature predicts gastric cancer recurrence through integrated bioinformatics and single-cell analysis Authors: Yi Wang, Rui Mao Published: 2026 ABSTRACT BackgroundGastric cancer (GC) remains a major cause of cancer-related mortality worldwide, with recurrence rates exceeding 40% following curative resection. While molecular profiling has revealed GC heterogeneity, computational approaches integrating tumor microenvironment (TME) characteristics for prognostic assessment remain underexplored. Therefore, this study aimed to develop and validate a stromal-derived gene signature for predicting gastric cancer recurrence and to characterize its tumor-microenvironmental and cell-type specificity through integrated bulk transcriptomic, machine-learning, and single-cell analyses.MethodsUsing publicly available datasets, we developed a prognostic signature through bootstrap-based feature selection (1,000 iterations) using 279 GC patients from GSE62254, divided into training (n=196) and validation (n=83) cohorts. Weighted gene co-expression network analysis (WGCNA) identified recurrence-associated modules. Single-cell RNA sequencing (scRNA-seq) of 43,560 cells determined cell-type specificity. Six machine learning algorithms assessed predictive performance. External validation used The Cancer Genome Atlas Stomach Adenocarcinoma (TCGA-STAD), including 419 patients with complete risk-score data, and pan-cancer cohorts (n=9,328 across 33 cancer types).ResultsThe five-gene signature (FBXL2, MLLT11, NES, HEYL, AKAP12) achieved significant risk stratification [training: hazard ratio (HR) =3.074, 95% confidence interval (CI): 2.283–4.140, P=1.44×10−13; validation: HR =3.07, 95% CI: 1.56–6.05, P=7.93×10−4] with area under the curve (AUC) of 0.792 (95% CI: 0.721–0.863) at 12 months. Single-cell analysis revealed cell-type-enriched expression patterns: AKAP12 in fibroblasts (18.3%), NES in myeloid cells (11.3%), HEYL in smooth muscle cells (7.2%), and MLLT11 in secretory cells (4.6%). WGCNA localized 80% of signature genes to extracellular matrix (ECM) remodeling pathways. High-risk patients exhibited elevated predicted chemotherapy resistance across eight agents (all P<0.001) and distinct immune correlation patterns, with MLLT11 correlating with M2 macrophages (r=0.224, P=1.73×10−4). The signature remained prognostic independent of tumor mutational burden (TMB) (P=0.78). Pan-cancer analysis across 33 TCGA cohorts (n=9,328) demonstrated tissue-specific prognostic value in 4 cancer types after false discovery rate (FDR) correction (FDR <0.05), including the primary GC cohort (STAD, HR =1.6, FDR =0.041), indicating limited cross-cancer applicability.ConclusionsThis stromal-derived five-gene signature suggests that microenvironment composition is strongly associated with GC recurrence, independent of tumor-intrinsic mutational burden. The predominant stromal expression and independence from TMB indicate that prognostic value derives from stromal-tumor interactions rather than tumor cell-autonomous features. While experimental validation is required to establish causality, these findings identify the stromal compartment as a potential therapeutic target and highlight the importance of integrating microenvironment characteristics into prognostic assessment. Highlight box Key findings • A bootstrap-derived five-gene signature (FBXL2, MLLT11, NES, HEYL, AKAP12) stratified gastric cancer (GC) recurrence risk in two independent cohorts [training hazard ratio (HR) =3.07, P=1.4×10−13; external The Cancer Genome Atlas (TCGA) Stomach Adenocarcinoma validation false discovery rate =0.041] and remained independent of tumor mutational burden. • Single-cell RNA sequencing of 43,560 cells localized signature expression predominantly to stromal compartments (cancer-associated fibroblasts, smooth-muscle cells, myeloid cells), and the risk score retained within-subtype prognostic value across all four Asian Cancer Research Group (ACRG) molecular subtypes (HR =2.18–5.27; all Cox P<0.001). What is known and what is new? • Existing GC prognostic signatures and the established ACRG and TCGA molecular classifications are dominated by tumor-cell-intrinsic features and capture only a limited fraction of microenvironment-driven recurrence risk. • This work shows at single-cell resolution that a stable, predominantly stromal-derived five-gene signature provides prognostic information that is complementary to—rather than redundant with—the ACRG/TCGA classifications, with the highest median risk scores converging on the two most stromal-rich subtypes (ACRG microsatellite stable/epithelial-mesenchymal transition and TCGA GS) across both cohorts. What is the implication, and what should change now? • The signature identifies the stromal compartment as a candidate therapeutic and surveillance target in GC and supports the integration of microenvironment-aware features into future prognostic models. Prospective validation in treatment-stratified cohorts and in patient-derived organoid or explant systems that preserve stromal-tumor crosstalk is required before the signature can inform individual treatment decisions. Introduction Gastric cancer (GC) ranks as the fifth most common malignancy and fifth leading cause of cancer-related mortality worldwide, with approximately 968,000 new cases and 660,000 deaths annually (1,2). Despite advances in surgical techniques and perioperative therapy, approximately 30–60% of patients with locally advanced GC experience recurrence within five years following curative-intent resection (3-5). This persistent high recurrence rate underscores the critical need for prognostic biomarkers capable of identifying high-risk patients who would benefit from intensified surveillance or adjuvant therapy. Current prognostic assessment relies primarily on the tumor-node-metastasis (TNM) staging system, which, although providing essential anatomical information, cannot adequately capture the complex biological features including immune characteristics and cellular heterogeneity that drive GC progression and recurrence (6,7). Large-scale genomic initiatives, including The Cancer Genome Atlas (TCGA) and Asian Cancer Research Group (ACRG) classifications, have delineated distinct molecular subtypes with varying clinical outcomes (8,9). However, these classifications have limited capacity to capture the complex interactions within the tumor microenvironment (TME), which plays a crucial role in disease progression (10,11). The TME has emerged as a critical determinant of GC progression, metastasis, and therapeutic response (12-14). Single-cell genomic characterization has revealed the gastric TME as a complex ecosystem comprising stromal cells, diverse immune cell populations including macrophages and T cells, and extracellular matrix (ECM) components, with extensive intercellular crosstalk driving tumor progression (15). Cancer-associated fibroblasts (CAFs), in particular, orchestrate the pro-tumorigenic microenvironment through secretion of growth factors, cytokines, and matrix proteins that facilitate tumor growth, invasion, and treatment resistance (16,17). This stromal heterogeneity in the gastric TME suggests that capturing TME cellular composition may be essential for accurate prognostic assessment and patient stratification (18,19). Despite numerous proposed gene signatures for GC prognosis (20), the clinical reliability and validation of these signatures remain to be further established. Notably, the TME exhibits substantial spatiotemporal heterogeneity (21), which may not be adequately captured by conventional gene signatures, potentially limiting their clinical applicability. Recent advances in spatial transcriptomics and single-cell technologies have revolutionized our understanding of tumor-TME interactions at fine-grained resolution (22), offering opportunities to develop mechanistically informed prognostic models that account for spatiotemporal heterogeneity. To address these limitations, we employed bootstrap-based feature selection, a stable feature selection approach specifically designed to overcome the instability issues inherent in traditional feature selection methods when applied to high-dimensional genomic data (23). We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0518/rc). Methods Study design and data sources This retrospective multi-cohort study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study involved only secondary analysis of de-identified data from publicly available repositories [Gene Expression Omnibus (GEO) and TCGA]; the original studies that generated these datasets obtained appropriate institutional review board approval and informed consent from all participants at the time of sample collection, and additional ethical approval and informed consent for the present secondary analysis were therefore not required by our institution. Data collection Three independent datasets were analyzed in this study. The discovery cohort was obtained from the GEO database (accession No. GSE62254), containing gene expression profiles and clinical follow-up data for 300 GC patients. The external validation cohort was derived from TCGA Stomach Adenocarcinoma (TCGA-STAD) project; among patients with documented progression-free survival data, 419 had complete data required for risk-score calculation and external validation. For single-cell validation, we utilized dataset GSE134520 from GEO, comprising 43,560 high-quality cells derived from gastric tissue specimens across various disease progression stages. Patient selection and stratification Patient selection followed strict inclusion criteria: histologically confirmed gastric adenocarcinoma, complete gene expression data, and available recurrence outcome information. After quality control, 279 patients from the discovery cohort were retained for analysis. This cohort was randomly partitioned into training (70%, n=196) and validation sets (30%, n=83) using stratified random sampling to achieve comparable recurrence-event distributions between the training and validation sets. Development of the prognostic signature Differential expression analysis Differentially expressed genes between recurrent and non-recurrent cases were identified using the limma package (version 3.54.2) (24). Genes with Benjamini-Hochberg false discovery rate (FDR) <0.01 and absolute log2 fold change >1.0 were retained as high-confidence candidates for feature selection. Feature selection strategy Bootstrap-based feature selection was performed with 1,000 iterations using least absolute shrinkage and selection operator (LASSO)-Cox regression implemented in the glmnet package (version 4.1-6) (25). In each iteration, the training set was resampled with replacement and subjected to LASSO with 10-fold cross-validation. Selection frequency was recorded for each gene across all iterations. The five genes with the highest selection frequencies were retained for signature construction. Risk score calculation Risk scores were calculated as Risk Score = Σ (βi × Zi), where βi represents the Cox regression coefficient and Zi the standardized expression value (Z-score) of gene i. Standardization parameters (mean and standard deviation) were derived from the training set and applied to all validation cohorts. For internal validation (GSE62254), the optimal cutoff value was determined in the training set using the surv_cutpoint function (survminer package, version 0.4.9) (26) by maximizing the log-rank statistic, then applied unchanged to the validation cohort. For external validation (TCGA-STAD), direct cutoff transfer was not feasible due to platform differences between microarray and RNA sequencing (RNA-seq) data. A dual validation strategy was therefore employed: (I) Cox regression with risk score as a continuous variable to assess discriminative ability; and (II) median-based stratification using equivalent percentile cutoffs to evaluate categorical risk classification. Gene expression in TCGA was converted to percentile ranks to normalize platform differences while preserving relative expression patterns. Cox coefficients derived from the training set were kept fixed and never re-fitted. Prognostic performance assessment Survival analysis Survival differences between risk groups were evaluated using Kaplan-Meier analysis with log-rank test. Time-dependent receiver operating characteristic (ROC) curves were constructed using the timeROC package (version 0.4) (27) to assess predictive accuracy at 12, 24, and 36 months. Kaplan-Meier curves were visualized using the survminer package (version 0.4.9). Independent prognostic value Multivariate Cox proportional hazards regression models incorporating age, sex, and American Joint Committee on Cancer (AJCC) stage as covariates were constructed to assess whether the risk score provided independent prognostic value beyond conventional clinical factors. The proportional hazards assumption was formally tested using Schoenfeld residuals (cox.zph function in R). Calibration curves were generated to evaluate the agreement between predicted and observed survival probabilities, with calibration slope and intercept reported with 95% confidence intervals (CIs). Decision curve analysis (DCA) (28) was performed to assess the net clinical benefit of the prognostic model across a range of threshold probabilities. Machine learning model development Algorithm implementation Six machine learning algorithms were implemented: Cox proportional hazards model with log-rank splitting; Random Survival Forest (29) (1,000 trees, minimum node size =5); XGBoost (30) (Cox objective, η =0.01, max depth =3, 100 rounds); Gradient Boosting Machine (31) (1,000 trees, 5-fold cross-validation); neural network [two-layer architecture with 10-5 hidden units, rectified linear unit (ReLU) activation]; and ensemble model (weighted average based on C-index performance). Pan-cancer validation and multiple testing correction Prognostic value was evaluated across 33 TCGA cancer types (n=9,328) using GEPIA2 (32). Disease-free survival analysis was performed with median expression cutoff. Hazard ratios (HRs) with 95% CIs were calculated, and Benjamini-Hochberg FDR correction was applied (significance threshold: FDR <0.05). Tumor mutational burden (TMB) analysis TMB calculation Somatic mutation data from TCGA-STAD were retrieved and processed using the maftools package (version 2.18.0) (33). TMB was calculated as the total number of non-synonymous mutations per megabase of genome examined. Patients were stratified into high and low TMB groups using the median TMB value as the cutoff threshold. Mutational landscape Mutation landscapes were visualized using the oncoplot function in maftools. Transition-to-transversion (Ti/Tv) ratios were calculated to assess mutational signatures. Statistical comparisons of TMB between risk groups were performed using Wilcoxon rank-sum tests. Co-expression network and functional analysis Network construction Co-expression networks were constructed using the weighted gene co-expression network analysis (WGCNA) package (version 1.72-5) (34). The top 50% most variable genes based on median absolute deviation were selected for network construction. A soft-thresholding power of β=4 was selected to achieve scale-free topology (R2>0.9). Module identification Gene modules were identified using dynamic tree cutting with minimum module size of 30 genes and merge height threshold of 0.25. Module eigengenes were calculated as the first principal component of each module’s expression matrix. Hub genes were defined as genes with module membership >0.6 and gene significance >0.1. Gene co-expression network analysis Correlations between signature gene expression and other genes were assessed using Spearman correlation analysis (|r|>0.3, P<0.05) to identify potential functional relationships. Functional enrichment analysis was performed using the clusterProfiler package (version 4.8.0) (35) with Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway databases (adjusted P<0.05 by Benjamini-Hochberg method). Immune microenvironment characterization Immune cell infiltration Immune cell infiltration levels were quantified using single-sample gene set enrichment analysis (ssGSEA) implemented in the GSVA package (version 1.48.0) (36). Established immune cell type-specific gene signatures from Bindea et al. (37) and Charoentong et al. (38) were used to calculate enrichment scores, which were subsequently z-score normalized across samples. Correlations between signature gene expression and immune cell infiltration were assessed using Spearman’s rank correlation. Differences in immune infiltration between risk groups were compared using Wilcoxon rank-sum tests. Drug sensitivity prediction In silico drug sensitivity predictions were generated using the pRRophetic algorithm (version 0.5) (39), which was trained on pharmacogenomic data from the Cancer Cell Line Encyclopedia (CCLE) and Genomics of Drug Sensitivity in Cancer (GDSC) databases. Estimated half-maximal inhibitory concentration (IC50) values were predicted for eight commonly used chemotherapeutic agents in GC. Differences in predicted IC50 values between risk groups were evaluated using Wilcoxon rank-sum tests with Benjamini-Hochberg correction for multiple testing. Single-cell RNA-seq (scRNA-seq) analysis Data processing scRNA-seq data from GSE134520 (43,560 cells) were processed using the Seurat package (version 5.3.0) (40). Quality control filtering removed cells with mitochondrial gene content >20% or gene detection outside the range of 200–4,000 genes per cell. Potential doublets were identified using a conservative statistical threshold approach [98th percentile of gene and unique molecular identifier (UMI) counts], consistent with computational methods such as DoubletFinder (41). Gene expression data were log-normalized with a scale factor of 10,000. Dimensionality reduction was performed using principal component analysis (PCA), and the first 30 principal components were selected based on elbow plot analysis. Cell clustering was conducted using the Louvain algorithm with a resolution of 0.5. As all samples were processed in the same experimental batch and sample-type separation primarily reflected biological disease progression rather than technical artifacts (confirmed by Shannon diversity analysis of sample mixing within cell type clusters), batch effect correction was not necessary. Cell type annotation Cell types were annotated based on canonical marker gene expression. Epithelial cells were identified by expression of EPCAM, KRT8, KRT18, KRT19, and CDH1. Mucous cells were distinguished by MUC5AC, TFF1, TFF2, and MUC6. Stromal populations were annotated as fibroblasts (COL1A1, DCN, LUM) and smooth muscle cells (ACTA2, MYH11, TAGLN). Secretory cell types were identified using PGC, LIPF, CHGA, and CHGB. Immune cell populations were annotated using established lineage-specific markers: T cells (CD3D, CD3E, CD2, CD3G), B cells (CD79A, BANK1, IGHA1, JCHAIN), myeloid cells (LYZ, CD14, CD68, APOE), and natural killer (NK) cells (NKG7, NCAM1). Endothelial cells (PECAM1, PLVAP) and proliferating cells (MKI67, TOP2A, PCNA) were also identified. Cell type assignment was performed using the AddModuleScore function with automated scoring of marker gene expression patterns. Statistical analysis All statistical analyses were performed using R software (version 4.3.1) (42). Continuous variables were compared between groups using Wilcoxon rank-sum test for non-normally distributed data or Student’s t-test for normally distributed data. Categorical variables were compared using Fisher’s exact test or chi-square test as appropriate. Correlations were assessed using Spearman’s rank correlation coefficient. For multiple hypothesis testing, the Benjamini-Hochberg FDR method was applied with significance threshold set at FDR <0.05. All random processes used a fixed seed value of 42 to ensure reproducibility. Two-sided P<0.05 was considered statistically significant unless otherwise specified. Molecular subtype stratification analysis To evaluate whether the continuous five-gene risk score captures information beyond established GC molecular classifications, two pre-specified stratified analyses were performed in independent cohorts (results in section of risk-score stratification across established molecular subtypes). For GSE62254 (n=279), patients were stratified by the four-class ACRG classification (9) [Cristescu et al., Nat Med 2015; microsatellite instability (MSI), microsatellite stable (MSS)/epithelial-mesenchymal transition (EMT), MSS/TP53⁺, MSS/TP53−] using the Molecular_Subtype_Simple field provided with the original dataset, with progression-free recurrence as the survival endpoint. For TCGA-STAD, patients were stratified by the four-class TCGA 2014 molecular classification (8) [also referred to as the Bass 2014 integrative classification: chromosomal instability (CIN), GS, MSI, EBV] obtained from the TCGA PanCancerAtlas integrative classification via the TCGAbiolinks Bioconductor R package [PanCancerAtlas_subtypes() function]; hypermutated samples [hypermutated (HM)-single nucleotide variant (SNV)/indel and POLE] were grouped with MSI per standard convention. Three pre-specified survival endpoints were evaluated separately in TCGA-STAD: progression-free interval (PFI) (n=377), disease-free interval (DFI) (n=221), and overall survival (OS) (n=375). Cross-subtype differences in the continuous risk score were tested with the non-parametric Kruskal-Wallis test. Within-subtype prognostic value was assessed using univariable Cox proportional-hazards regression with the continuous risk score as the predictor; HRs are reported per unit increase on the z-scored risk-score scale together with two-sided 95% CIs and Wald P values. To prevent unstable estimates in small subtypes, a pre-specified stability rule was applied (n≥15 patients and ≥5 events per subtype); subtypes failing this rule are reported as not estimable (NA). For GSE62254, the pre-specified GSE62254-derived risk-score cutpoint (−0.1748) was used to dichotomize patients into high-/low-risk groups for Kaplan-Meier visualization; for TCGA-STAD, a within-cohort median split of the continuous risk score was used because direct cutpoint transfer between microarray and RNA-seq platforms is not appropriate, consistent with the dual validation strategy described in section of risk score calculation. Cox coefficients derived from the training set were kept fixed throughout these analyses and were not re-fitted within subtypes. Subtype merging, Kruskal-Wallis tests, within-subtype Cox regressions, and stability-rule application were performed in R, as described in the “Results” section. Results Patient characteristics and signature development A total of 279 GC patients from the GSE62254 dataset with complete clinical and survival data were included in this study. The cohort was divided into a training set (n=196, 86 recurrence events, 43.9%) and an internal validation set (n=83, 39 recurrence events, 47.0%), with no significant difference in recurrence rates between the two subsets (P=0.69). Baseline clinical characteristics, including age, sex, T stage, N stage, M stage, and AJCC stage, were comparable between the two subsets (Table 1). Table 1Clinical characteristics of patients in the training and validation cohorts (GSE62254, n=279)CharacteristicTraining cohort (n=196)Validation cohort (n=83)P valueAge, years63.5 (28.0–86.0)63.0 (24.0–82.0)0.53Sex>0.99 Male129 (65.8)55 (66.3) Female67 (34.2)28 (33.7)T stage0.79 T120 (10.2)12 (14.5) T241 (20.9)17 (20.5) T378 (39.8)31 (37.3) T457 (29.1)23 (27.7)N stage0.21 N056 (28.6)26 (31.3) N164 (32.7)17 (20.5) N241 (20.9)23 (27.7) N335 (17.9)17 (20.5)M stage0.30 M0169 (86.2)76 (91.6) M127 (13.8)7 (8.4)AJCC stage0.90 Stage I0 (0.0)0 (0.0) Stage II110 (56.1)44 (53.0) Stage III72 (36.7)33 (39.8) Stage IV14 (7.1)6 (7.2)Recurrence events86 (43.9)39 (47.0)0.69Continuous variables are summarized as median (range) and compared using the Wilcoxon rank-sum test; categorical variables are summarized as n (%) and compared using the Chi-squared or Fisher’s exact test, as appropriate. AJCC, American Joint Committee on Cancer; M, metastasis; N, node; T, tumor. Clinical characteristics of patients in the training and validation cohorts (GSE62254, n=279) Continuous variables are summarized as median (range) and compared using the Wilcoxon rank-sum test; categorical variables are summarized as n (%) and compared using the Chi-squared or Fisher’s exact test, as appropriate. AJCC, American Joint Committee on Cancer; M, metastasis; N, node; T, tumor. Differential expression analysis between recurrence and non-recurrence groups identified 23 genes significantly associated with GC recurrence, comprising 20 upregulated and 3 downregulated genes in the recurrence group (Figure 1A; Table S1). These genes demonstrated substantial expression differences and served as candidate features for prognostic model development. Development and validation of the five-gene prognostic signature in gastric cancer. (A) Volcano plot of differentially expressed genes between recurrent and non-recurrent cases. (B) Bootstrap selection frequencies of five signature genes (FBXL2, MLLT11, NES, HEYL, AKAP12). (C) Kaplan-Meier curves for training (n=196) and validation (n=83) cohorts. (D) Time-dependent ROC curves at 12, 24, and 36 months. (E) Forest plot of multivariate Cox regression analysis. (F) Heatmap of signature gene expression across samples. **, P<0.01; ***, P<0.001. AJCC, American Joint Committee on Cancer; AUC, area under the curve; CI, confidence interval; FC, fold change; HR, hazard ratio; M, metastasis; N, node; ROC, receiver operating characteristic; T, tumor. Bootstrap resampling analysis with 1,000 iterations was performed to identify the most stable prognostic genes. Five genes demonstrated the highest selection frequencies and were selected for signature construction: FBXL2 (22.3%, 223/1,000 iterations), MLLT11 (21.1%, 211/1,000), NES (18.9%, 189/1,000), HEYL (17.7%, 177/1,000), and AKAP12 (10.2%, 102/1,000) (Figure 1B). These five genes showed clear separation from the sixth-ranked gene (selection frequency <10%), indicating distinct prognostic stability. Bootstrap coefficient distributions demonstrated consistent effect directions across iterations (Figure S1A). The five-gene risk score was calculated using standardized expression values and Cox regression-derived coefficients (Table 2): Risk Score =0.373 × Z [FBXL2] + 0.010 × Z [MLLT11] + 0.484 × Z [NES] + 0.183 × Z [HEYL] + 0.120 × Z [AKAP12], where Z [gene] represents the standardized expression (z-score) calculated as Z = (Expression − μ)/σ, with μ and σ derived from the training set (Table 2). Among the five genes, NES contributed the largest effect (β=0.484), followed by FBXL2 (β=0.373), suggesting these two genes are the primary drivers of the prognostic signal. The optimal risk stratification cutoff was determined as −0.1748 (approximately the 51st percentile in the training cohort). Table 2Risk-score model specifications: Cox regression coefficients and standardization parameters of the five-gene signature derived from the training cohort (GSE62254, n=196)GeneCox βμ (training mean)σ (training SD)Predominant cell-type expression (scRNA-seq, GSE134520) FBXL2 0.3731.7810.262Broadly expressed; epithelial and stromal compartments MLLT11 0.0102.1600.328Secretory cells (4.6%); positive correlation with M2 macrophages (r=0.224, P=1.7×10−4) NES 0.4841.8330.166Myeloid cells (11.3%) HEYL 0.1831.9360.137Smooth-muscle cells (7.2%) AKAP12 0.1201.7070.299CAFs (18.3%); smooth-muscle cells (18.1%)Risk Score = Σ (βi × Zi), where Zi = (Expression − μi)/σi. Standardization parameters (μ and σ) were derived from the training set (n=196) and applied unchanged to all validation cohorts. The optimal risk-stratification cutpoint determined in the training set by maximization of the log-rank statistic was −0.1748 (≈51st percentile in the training cohort). Cell-type localization summaries are taken from section single-cell validation and stromal enrichment of the manuscript and Table S2. CAF, cancer-associated fibroblast; scRNA-seq, single-cell RNA sequencing; SD, standard deviation. Risk-score model specifications: Cox regression coefficients and standardization parameters of the five-gene signature derived from the training cohort (GSE62254, n=196) Risk Score = Σ (βi × Zi), where Zi = (Expression − μi)/σi. Standardization parameters (μ and σ) were derived from the training set (n=196) and applied unchanged to all validation cohorts. The optimal risk-stratification cutpoint determined in the training set by maximization of the log-rank statistic was −0.1748 (≈51st percentile in the training cohort). Cell-type localization summaries are taken from section single-cell validation and stromal enrichment of the manuscript and Table S2. CAF, cancer-associated fibroblast; scRNA-seq, single-cell RNA sequencing; SD, standard deviation. Prognostic performance and model validation Six machine learning algorithms were compared for prognostic model construction. The Cox proportional hazards model achieved a C-index of 0.681 in the training cohort and 0.625 in the validation cohort. Random Survival Forest (training: 0.796, validation: 0.618), Gradient Boosting Machine (training: 0.800, validation: 0.624), Neural Network (training: 0.681, validation: 0.625), Ensemble (training: 0.817, validation: 0.616), and XGBoost (training: 0.947, validation: 0.600) showed varied performance patterns across training and validation sets (Figure S1B). The Cox model was selected for final signature development due to its balanced performance and model interpretability. Based on the five-gene signature, patients were stratified into high- and low-risk groups. Kaplan-Meier analysis showed significantly shorter recurrence-free survival (RFS) in the high-risk group compared to the low-risk group in both the training cohort (log-rank P=3.12×10−12, HR =4.85, 95% CI: 3.12–7.54) and validation cohort (log-rank P=7.93×10−4, HR =3.07, 95% CI: 1.56–6.05) (Figure 1C). Time-dependent ROC analysis evaluated the signature’s predictive performance at multiple time points (Figure 1D; Table S3). In the training cohort, area under the curve (AUC) values were 0.792 (95% CI: 0.721–0.863) at 12 months, 0.780 (95% CI: 0.711–0.850) at 24 months, and 0.777 (95% CI: 0.708–0.847) at 36 months. In the validation cohort, AUC values were 0.725 (95% CI: 0.596–0.855) at 12 months, 0.697 (95% CI: 0.572–0.821) at 24 months, and 0.698 (95% CI: 0.574–0.823) at 36 months. Multivariate Cox regression analysis was performed to evaluate the independent prognostic value of the risk score. The proportional hazards assumption was satisfied for the final model (Schoenfeld global test P=0.055), with no significant violations detected for individual covariates (all P>0.05). The risk score was identified as the strongest independent prognostic factor (HR =3.074, 95% CI: 2.283–4.140, P=1.44×10−13), outperforming AJCC stage (HR =1.707, 95% CI: 1.256–2.321, P=0.001) and age (HR =1.033, 95% CI: 1.010–1.057, P=0.005) (Figure 1E; detailed statistics in Table S4). Heatmap visualization revealed distinct expression patterns that correlated with risk stratification, recurrence events, and clinical stage (Figure 1F). High-risk patients exhibited coordinated upregulation of FBXL2, MLLT11, NES, HEYL, and AKAP12. A comprehensive nomogram integrating age, gender, risk score, and TNM staging was developed for individualized recurrence prediction at 12, 24, and 36 months (Figure S2A). Calibration curves demonstrated good agreement between predicted and observed recurrence probabilities at all time points (Figure S2B). DCA showed superior net benefit of the risk model across threshold probabilities of 10–50%, with particular utility in the 20–40% range (Figure S2C). External and pan-cancer validation The TCGA-STAD cohort included 419 patients with complete data (94.6% of original cohort), comprising 142 progression events (33.9%) with a median follow-up of 1.1 years (range: 0.003–10.2 years). Due to substantial expression scale differences between microarray (GSE62254) and RNA-seq (TCGA) platforms, a dual validation strategy was employed. Gene expression was converted to percentile ranks, and Cox coefficients from the training set were applied unchanged. Primary analysis-continuous risk score: Cox regression analysis showed risk score as a predictor of progression (HR =1.615, 95% CI: 0.881–2.960, P=0.12; C-index =0.531±0.026). Time-dependent ROC analysis yielded AUC values of 0.522 (95% CI: 0.452–0.592) at 1 year and 0.551 (95% CI: 0.475–0.627) at 2 years (Figure S3B). Secondary analysis-median stratification: patients were stratified into low-risk (n=210) and high-risk (n=209) groups using median split. Event rates were 31.9% (67/210) and 35.9% (75/209), respectively. Kaplan-Meier analysis showed a trend toward shorter progression-free survival in the high-risk group (log-rank P=0.24; HR =1.219, 95% CI: 0.876–1.696; Figure S3A). Pan-cancer expression analysis across 33 TCGA cancer types (n=9,328) revealed tissue-selective patterns of the five-gene signature (Figure S4A). After Benjamini-Hochberg FDR correction for multiple testing, 4 of 33 cancer types demonstrated significant prognostic associations (FDR <0.05): STAD (HR =1.6, FDR =0.041), bladder urothelial carcinoma (BLCA, HR =1.6, FDR =0.037), kidney renal papillary cell carcinoma (KIRP, HR =2.7, FDR =0.037), and mesothelioma (MESO, HR =2.1, FDR =0.037). The initially hypothesized gastrointestinal enrichment was not statistically significant after FDR correction (Fisher’s P=1.0), indicating tissue-specific rather than pan-cancer applicability. External validation in aggregated cohorts showed GC-specific performance (n=384, HR =2.1, P=2.3×10−4; Figure S4B), pooled gastrointestinal cancers (n=930, HR =1.8, P=9.7×10−7; Figure S4C), and pan-cancer meta-analysis (n=9,328, HR =1.2, P=1.2×10−4; Figure S4D). The modest pan-cancer effect size compared to gastric-specific validation reinforces tissue-selective utility. Co-expression network and functional analysis WGCNA of 6,892 genes identified 12 distinct co-expression modules (Tables S5-S7). Soft-thresholding power β=4 achieved scale-free topology (R2=0.904) (Figure 2A). Gene dendrogram with module color assignments showed hierarchical clustering patterns (Figure 2B). Module-trait correlation analysis demonstrated that the turquoise module exhibited the strongest associations with recurrence (r=0.329, P<0.001) and AJCC stage (r=0.251, P<0.001) (Figure 2C). WGCNA network analysis and functional enrichment. (A) Network topology analysis for soft-thresholding power selection (β=4, R2=0.904). (B) Hierarchical clustering dendrogram showing 12 co-expression modules. (C) Module-trait correlation heatmap. (D) Venn diagram of signature genes and turquoise module overlap. (E,F) GO biological process and KEGG pathway enrichment of the turquoise module. GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; M, metastasis; N, node; T, tumor; WGCNA, weighted gene co-expression network analysis. Four of the five signature genes (FBXL2, MLLT11, NES, AKAP12; 80%) were assigned to the turquoise module, the largest co-expression module containing 2,106 genes, while HEYL was not included in the WGCNA network (Figure 2D). Among the four turquoise-localized genes, MLLT11 and AKAP12 were identified as hub genes (module membership >0.8, gene significance >0.2), suggesting central regulatory roles within the network. GO enrichment analysis of the turquoise module revealed 536 significantly enriched biological processes (adjusted P<0.05; Table S8), predominantly including ECM organization, extracellular structure organization, cell-substrate adhesion, and wound healing (Figure 2E). KEGG pathway analysis identified 31 significantly enriched pathways (Table S9), with cytoskeleton in muscle cells, focal adhesion, ECM-receptor interaction, and phosphatidylinositol 3-kinase (PI3K)-Akt signaling pathway among the top-ranked (Figure 2F). These enrichment patterns indicate that the turquoise module reflects stromal cell programs involving ECM remodeling and cell adhesion machinery, consistent with single-cell analysis showing signature genes predominantly expressed in stromal populations including CAFs, smooth muscle cells, and myeloid cells. Mechanistic characterization of FBXL2 (22.3% selection frequency) and MLLT11 (12.0% selection frequency) revealed no differential expression between risk groups (P=0.36 and P=0.44, respectively), confirming that multi-gene integration drives predictive power rather than individual gene expression (Figure S5A,S5C). Immune correlation analysis revealed opposing patterns: FBXL2 expression negatively correlated with anti-tumor M1 macrophages (r=−0.192, P=0.001) and CD8+ T cells (r=−0.145, P=0.02), while MLLT11 expression positively correlated with immunosuppressive M2 macrophages (r=0.224, P=1.7×10−4) (Figure S5B,S5D). Tumor immune microenvironment and drug sensitivity Gene-immune correlation heatmap (Figure 3A) demonstrated distinct association patterns: AKAP12 and HEYL showed strong positive correlations with M2 macrophages and mast cells, while MLLT11 and NES exhibited moderate positive associations with M2 macrophages. All five signature genes displayed predominantly negative or neutral correlations with effector immune populations (CD8+ T cells, NK cells, M1 macrophages), suggesting that the signature reflects gene-immune correlation patterns rather than absolute immune cell abundance differences. Comprehensive immune infiltration analysis across ten cell populations revealed no significant differences in individual immune cell abundances between high-risk and low-risk groups, including B cells, CD8+ T cells, regulatory T cells (Tregs), NK cells, macrophages, and M1 macrophages (all P>0.05, Figure 3B). Immune microenvironment and drug sensitivity analysis. (A) Correlation heatmap between signature genes and immune cells. (B) Comparison of immune cell infiltration between risk groups. (C) Correlation between MLLT11 expression and M2 macrophage infiltration (r=0.224, P=1.73×10−4). (D) Correlation network of MLLT11 and M2 macrophage markers. (E) Predicted IC50 values of eight chemotherapeutic agents. (F) Immune functional score comparison between risk groups. IC50, half-maximal inhibitory concentration; 5-FU, 5-fluorouracil. Despite the absence of differential M2 macrophage infiltration between risk groups, correlation analysis revealed that MLLT11 expression was significantly associated with M2 macrophage scores (Spearman r=0.224, P=1.73×10−4; Table S10) (Figure 3C), suggesting potential involvement in M2 polarization. MLLT11-M2 marker gene correlation network demonstrated strong interconnections among M2-associated genes including TGFB1, MSR1, CD163, and MRC1 (Figure 3D), with MSR1, CD163, and MRC1 showing particularly robust inter-correlations (r>0.8). Analysis of eight chemotherapy agents revealed that high-risk patients showed significantly elevated predicted IC50 values for all tested drugs (all P<0.001) based on pRRophetic algorithm (Figure 3E; Table S11). Predicted resistance ratios ranged from 1.083 (doxorubicin) to 1.137 (irinotecan), with irinotecan (ratio =1.137, P=2.8×10−8), paclitaxel (ratio =1.132, P=1.5×10−7), cisplatin (ratio =1.127, P=3.2×10−6), and 5-fluorouracil (5-FU) (ratio =1.121, P=8.7×10−7) showing the strongest predicted resistance patterns. This coordinated multi-drug resistance pattern suggests potential shared underlying mechanisms linking the five-gene signature to chemotherapy response. However, these predictions are derived from cell line pharmacogenomics (CCLE/GDSC) and require prospective clinical validation, as in vitro IC50 values correlate only modestly with patient outcomes. Comprehensive immune functional assessment revealed selective alterations in immune activity (Figure 3F; Table S12). Among multiple immune scores evaluated, only the overall immune score showed significant difference between risk groups (P=0.01), with high-risk patients exhibiting lower scores. Individual functional components including T cell score (P=0.59), immunosuppression score (P=0.30), NK activity (P=0.59), and M1/M2 ratio (P=0.84), showed no significant differences, suggesting that the prognostic signature reflects subtle coordinated immune alterations rather than dysfunction of specific immune cell populations. Mutational landscape and TMB independence Comprehensive somatic mutation profiling revealed overlapping genomic landscapes between high-risk (n=266) and low-risk (n=165) groups (Figure 4A,4B). The three most frequently mutated genes were shared: TTN (51% in high-risk vs. 52% in low-risk), TP53 (47% vs. 44%), and MUC16 (31% in both groups). Additional frequently mutated genes included LRP1B, MUC16A, SYNE1, CSMD3, FLG, and HMCN1, with similar mutation frequencies across risk groups. Oncoplot visualization revealed complex mutation patterns with substantial inter-patient heterogeneity in both groups, with no individual genes showing significantly different mutation frequencies between risk stratifications. Mutational landscape and tumor mutational burden analysis. (A,B) Oncoplot of somatic mutations in high-risk (n=266) and low-risk (n=165) groups. (C) TMB comparison between risk groups (P=0.78). (D,E) Mutational signature characteristics showing variant classification, type distribution, and SNV classes. SNV, single nucleotide variant; Ti, transition; TMB, tumor mutational burden; Tv, transversion. TMB analysis demonstrated no significant difference between high-risk and low-risk groups (Wilcoxon rank-sum test, P=0.78) (Figure 4C). Both groups exhibited similar TMB distributions with median values of approximately 1.5 log2 (TMB + 1) mutations per megabase and comparable ranges (0 to >6), indicating that the five-gene signature’s prognostic value operates independently of overall mutational load. Detailed mutational signature analysis revealed typical gastric adenocarcinoma patterns shared between risk groups (Figure 4D,4E). Variant classification showed missense mutations as the dominant category, followed by frameshift deletions, nonsense mutations, and splice site alterations. Variant type distribution demonstrated single nucleotide polymorphism (SNP) predominance (>80% of variants), with deletions and insertions comprising smaller fractions. SNV class analysis revealed C>T transitions as the most frequent mutation type (51.2%), characteristic of aging-related cytosine deamination (Figure 4E, upper panels). Ti/Tv ratio was approximately 1.89 (median), within the expected range for solid tumors (Figure 4E, upper right). Mutations per sample showed wide variability (50 to >300), with right-skewed distribution typical of cancer genomics and a subset of hypermutated samples (Figure 4D, lower left; Figure 4E, lower panel). Single-cell validation and stromal enrichment scRNA-seq analysis of 43,560 high-quality cells spanning gastric carcinogenesis stages, including non-atrophic gastritis (NAG), chronic atrophic gastritis (CAG), intestinal metaplasia (IM), and early GC (EGC), identified 24 distinct cell populations through unsupervised clustering (Figure 5A). Uniform Manifold Approximation and Projection (UMAP) visualization demonstrated clear spatial segregation by disease stage (Figure 5B), with cells from different pathological stages occupying distinct but partially overlapping regions. Single-cell RNA sequencing validation of stromal cell-type specificity. (A) UMAP plot of 24 cell populations from 43,560 cells. (B) UMAP visualization colored by disease stage. (C) Feature plots showing expression of five signature genes. (D) Expression frequency of each gene across all cells. (E) Cell type-specific expression analysis. CAG, chronic atrophic gastritis; EGC, early gastric cancer; IM, intestinal metaplasia; NAG, non-atrophic gastritis; UMAP, Uniform Manifold Approximation and Projection. Feature plot analysis revealed remarkably sparse but specific expression patterns for all five signature genes (Figure 5C). Quantification across the entire dataset showed: AKAP12 (1.28% of cells), MLLT11 (1.00%), NES (0.81%), FBXL2 (0.31%), and HEYL (0.20%) (Figure 5D; Tables S2,S13,S14). Despite low overall frequencies, expression remained detectable across all disease stages with stable patterns (all P>0.05 between stages), suggesting that these genes mark constitutive rare cell populations present throughout disease progression. Cell type annotation revealed striking stromal enrichment of the five-gene signature (Figure 5E). AKAP12 showed highest expression in fibroblasts (18.3% of fibroblasts), HEYL in smooth muscle cells (7.2%), NES in myeloid cells (11.3%), and MLLT11 in secretory cells (4.6%). Minimal expression was detected in T cells, B cells, and epithelial cells (<1%), confirming that the signature predominantly reflects stromal rather than tumor-intrinsic programs. This cell type-specific distribution pattern validates that high-risk bulk tissue profiles represent increased proportions of specific stromal populations rather than uniform upregulation. Integration of bulk tissue, WGCNA, and single-cell data enabled construction of a comprehensive mechanistic model (Figure 6A-6C). The five-gene signature captures four distinct stromal cell types that collectively influence the TME: (I) CAFs (AKAP12-expressing) mediating ECM remodeling, (II) smooth muscle cells (HEYL-expressing) regulating vascular stability, (III) myeloid cells (NES-expressing) promoting immune modulation, and (IV) secretory cells (MLLT11-expressing) associated with M2 macrophage polarization signatures (Figure 6A,6B). These four cell populations may converge on tumor cells through complementary mechanisms—ECM remodeling, angiogenesis, immune modulation, and growth factor provision—collectively establishing a pro-tumorigenic microenvironment that may facilitate recurrence (Figure 6C). This stromal-centric model explains why the signature maintains prognostic value independent of TMB and may contribute to treatment resistance through altered tumor-stroma interactions. Integrated mechanistic model of stromal-driven gastric cancer recurrence. (A) Schematic of stromal cell types expressing signature genes. (B) Cell-cell interaction networks. (C) Integrated model depicting convergent mechanisms promoting tumor microenvironment remodeling. CAF, cancer-associated fibroblast; ECM, extracellular matrix. Risk-score stratification across established molecular subtypes To evaluate whether the 5-gene risk score captures information beyond established molecular classifications, we stratified patients by the ACRG classification (9) (Cristescu et al., Nat Med 2015) in GSE62254 (n=279) and by the TCGA 2014 molecular classification (8) (also referred to as the Bass 2014 integrative classification) in TCGA-STAD (n=377 for PFI, 221 for DFI, 375 for OS). In GSE62254, the continuous risk score differed significantly across the four ACRG subtypes (Kruskal-Wallis P=3.47×10−15; Figure 7A). The MSS/EMT subtype—the molecular class characterized by active stromal reprogramming—showed the highest scores (median=1.03; 93.0% classified as high-risk; n=43), whereas MSI tumors showed the lowest (median=−0.59; 33.3% high-risk; n=60); intermediate scores were observed for MSS/TP53+ (median=−0.45; 32.0%; n=75) and MSS/TP53− (median=−0.29; 47.5%; n=101). Within every ACRG subtype, the continuous risk score remained a statistically significant predictor of RFS in within-subtype Cox models (HR range 2.18–5.27; all Cox P<0.001; Figure 7B,7C; Table S15), indicating that the signature provides prognostic information complementary to—rather than redundant with—the ACRG classification. Risk-score stratification across established gastric cancer molecular subtypes. (A) Distribution of the 5-gene risk score across the four ACRG molecular subtypes in GSE62254 (n=279; MSI 60, MSS/EMT 43, MSS/TP53+ 75, MSS/TP53− 101). Kruskal-Wallis P=3.5×10−15. (B) Kaplan-Meier curves for RFS within each ACRG subtype, stratified by the pre-specified risk group; log-rank P<0.001 in MSI, P<0.001 in MSS/TP53−, P=0.003 in MSS/TP53+, and P=0.054 in the MSS/EMT subset. (C) Within-subtype Cox-regression forest plot for the ACRG cohort using the continuous risk score; all four subtypes show significant prognostic value of the risk score (HR range, 2.18–5.27; all Cox P<0.001). The numerical contrast between the log-rank P=0.054 and the Cox P=5.1×10−4 in the MSS/EMT subset reflects the different statistical models used—an unadjusted log-rank test on the pre-specified dichotomized risk groups in (B) versus a continuous-risk-score Cox model in (C)—both of which are directionally consistent. (D) Distribution of the 5-gene risk score across the four TCGA 2014 molecular subtypes in TCGA-STAD (n=377, PFI cohort; CIN 220, GS 49, MSI 78, EBV 30). Kruskal-Wallis P=1.9×10−10. (E) Kaplan-Meier curves for PFI within each TCGA subtype, stratified by within-cohort median split of the continuous risk score. (F) Within-subtype Cox-regression forest plot for the TCGA cohort (PFI endpoint, continuous risk score); HRs are numerically greater than 1 in CIN (1.08), MSI (1.69) and EBV (1.89) and below 1 in GS (0.86, 95% CI: 0.52–1.42), consistent with the attenuation discussed in Section of risk-score stratification across established molecular subtypes (small within-subtype N and saturation of the GS stromal-rich phenotype). At the OS endpoint (Table S15) all four within-subtype HRs are numerically greater than 1 (CIN 1.08, GS 1.34, MSI 1.88, EBV 1.10), restoring directional consistency at the longest follow-up endpoint. Molecular-subtype annotations for TCGA-STAD were obtained from the TCGA PanCancerAtlas integrative classification via the TCGAbiolinks R package; hypermutated samples (HM-SNV/indel, POLE) were grouped with MSI per standard convention. Box plots show median and interquartile range with individual points jittered. ACRG, Asian Cancer Research Group; CI, confidence interval; CIN, chromosomal instability; EBV, Epstein-Barr virus; EMT, epithelial-mesenchymal transition; GS, genomically stable; HM-SNV, hypermutated-single nucleotide variant; HR, hazard ratio; KM, Kaplan-Meier curve; MSI, microsatellite instability; MSS, microsatellite stable; OS, overall survival; PFI, progression-free interval; RFS, recurrence-free survival; STAD, stomach adenocarcinoma; TCGA, The Cancer Genome Atlas. External validation in TCGA-STAD recapitulated the cross-subtype pattern at every evaluated survival endpoint (Kruskal-Wallis P=1.92×10−10 for PFI, 1.55×10−8 for DFI, 2.51×10−10 for OS; Figure 7D). Notably, the TCGA GS subtype—which, like ACRG MSS/EMT, is enriched for diffuse-type, stromal-rich histology—again displayed the highest risk scores (median=0.37; 75.5% high-risk at PFI), while MSI tumors again showed the lowest (median=−0.54; 26.9% high-risk). Across both cohorts, the convergence of the highest median risk scores in the two most stromal-rich subtypes (ACRG MSS/EMT and TCGA GS) provides consistent cross-cohort biological support for a stromal-driven interpretation of the signature. Within-subtype Cox regressions in TCGA were directionally consistent at the overall-survival endpoint, with HRs numerically greater than 1 in all four subtypes (CIN: HR =1.08; GS: HR =1.34; MSI: HR =1.88; EBV: HR =1.10; Table S15), although none reached statistical significance after stratification. At the PFI endpoint (Figure 7E,7F), within-subtype HRs were greater than 1 in CIN (1.08), MSI (1.69) and EBV (1.89), whereas the GS subtype showed an attenuated and inverted point estimate (HR =0.86, 95% CI: 0.52–1.42, P=0.55). We interpret the GS finding as reflecting the joint effect of (I) a small subtype size with limited events (n=49, 20 events); (II) saturation of the categorical stromal-rich phenotype that defines the GS class; and (III) endpoint heterogeneity, since TCGA PFI/DFI capture progression rather than the recurrence endpoint on which the model was trained. Importantly, the cross-subtype effect (Kruskal-Wallis P≈1.9×10−10) and the cross-cohort convergence on the stromal-rich classes are marked and remain the principal findings of this stratified analysis. Discussion Stromal-centric prognostic biomarkers: novel insights into GC recurrence Our study reveals a notable shift in understanding GC prognosis by demonstrating that a five-gene signature (FBXL2, MLLT11, NES, HEYL, AKAP12) predominantly expressed in stromal cells, rather than tumor epithelial cells, is associated with recurrence prediction. This finding complements the prevailing tumor-centric approach by demonstrating that stromal composition is strongly associated with recurrence risk, independent of tumor-intrinsic mutational features. Recent studies have begun to recognize the heterogeneity and functional importance of CAFs in GC, with different CAF subsets contributing to distinct aspects of tumor progression. However, our work extends beyond CAF characterization by demonstrating that stromal reprogramming across multiple cell types—including CAFs, myeloid cells, and vascular smooth muscle cells—collectively correlates with clinical outcomes. The stromal enrichment of our prognostic genes aligns with emerging evidence that CAFs extensively participate in the malignant development of GC through secretion of various bioactive molecules, release of exosomes, and direct cell-cell interactions (43,44). Recent studies have demonstrated that CAFs promote proliferation, angiogenesis, metastasis, and immunosuppression in GC (44), and are recognized as key protagonists of the GC TME (45). However, unlike previous studies focusing on individual CAF markers such as FAP or ACTA2 (α-SMA), our signature captures a broader stromal reprogramming program that transcends single cell-type boundaries. This multi-cellular stromal signature may explain its superior prognostic performance compared to existing tumor-intrinsic signatures. NES expression in myeloid cells: a novel mechanism of immunosuppression Single-cell analysis revealed NES expression predominantly in myeloid cells (11.3%). This finding was validated by co-expression with canonical myeloid markers (C1QA, C1QB, AIF1). The identification of NES (Nestin) expression specifically in myeloid cells (11.3%) represents an unexpected observation with significant implications for understanding tumor-immune interactions. While Nestin has been traditionally recognized as a neural stem cell marker, recent evidence shows that Nestin-positive cells with progenitor and/or regulatory functions exist in multiple tissues, including bone marrow. Our finding that NES expression correlates with M2 macrophage infiltration suggests a previously unrecognized association with myeloid cell plasticity within the GC microenvironment. Studies in acute myeloid leukemia have demonstrated that Nestin-positive bone marrow mesenchymal stem cells support cancer cell survival through increased oxidative phosphorylation and glutathione-mediated antioxidant defense (46). Nestin-positive cells are recognized as markers of multilineage stem/progenitor cells and contribute to tissue microenvironment homeostasis (47). The spatial localization of Nestin-positive cells in perivascular regions has been documented in the bone marrow niche (48). We propose that a similar mechanism may operate in GC, where NES-expressing myeloid cells may create a metabolically supportive niche for tumor cells. The correlation between NES expression and predicted chemotherapy resistance across eight agents (P<0.001) is consistent with this metabolic coupling hypothesis. Furthermore, tissue-resident Nestin-positive cells have been shown to differentiate into pericytes and smooth muscle cells through tumor-cell-secreted factors, contributing to vessel stabilization and resistance to anti-angiogenic therapy (49). This dual association of NES with both immune modulation and vascular remodeling, combined with its enrichment in cell migration and ECM remodeling pathways identified through co-expression analysis, suggests it may contribute importantly to the overall prognostic signature. AKAP12: beyond tumor suppression to stromal activation AKAP12, traditionally viewed as a tumor suppressor (50), shows paradoxically high expression in CAFs of high-risk patients. Recent research provides mechanistic insight into this apparent contradiction. AKAP12-expressing CAFs in triple-negative breast cancer promote formation of an immunosuppressive TME by spatially mediating macrophage M2 polarization via interleukin-34 (IL-34)/macrophage colony-stimulating factor receptor (CSF1R) signaling (51). While M2 macrophage abundance showed no significant difference between risk groups (P>0.05), MLLT11 expression demonstrated significant correlation with M2 macrophage scores (r=0.224, P=1.73×10−4), suggesting that a possible IL-34/CSF1R axis may operate through M2 polarization states rather than abundance changes. AKAP12-null cells exhibit robust stress fiber formation and upregulated focal adhesion complexes, indicating its crucial role in cytoskeletal architecture regulation (52). The high expression of AKAP12 in CAFs (18.34%) may represent a compensatory mechanism attempting to normalize the aberrant ECM architecture characteristic of aggressive tumors. Pan-cancer analysis has revealed that AKAP12 expression correlates with resistance to anti-vascular endothelial growth factor (anti-VEGF) inhibitors and is positively associated with pro-tumor CAF abundance, consistent with our finding of poor prognosis in AKAP12-high patients (53). The scaffold function of AKAP12 for protein kinase A (PKA) and protein kinase C (PKC) signaling suggests it may serve as a central hub coordinating multiple pro-tumorigenic signaling pathways in CAFs. AKAP12 controls cell cycle progression by regulating cyclin D1 localization and influences cytokinesis through PKC scaffolding. These functions in CAFs could potentially promote a pro-proliferative microenvironment supporting tumor cell expansion. HEYL and vascular remodeling: implications for metastatic dissemination The specific expression of HEYL in smooth muscle cells (7.2%) implicates aberrant vascular maturation in GC progression. HEYL, a Notch signaling effector, plays crucial roles in vascular development and arterial-venous specification. In tumors, defective smooth muscle cell investment of vessels correlates with increased tumor cell shedding and metastasis, as these poorly invested vessels are readily penetrated by tumor cells. The structurally and functionally aberrant tumor vasculature contributes to the pro-tumorigenic and immunosuppressive microenvironment by maintaining hypoxia, acidosis, and high interstitial pressure, while generating a physical barrier to T cell infiltration (54). Vascular normalization strategies have demonstrated the potential to reprogram the immunosuppressive TME and enhance immunotherapy efficacy (55). Vascular smooth muscle cell plasticity has been recently recognized as a key determinant of TME composition and function (56). The enrichment of vascular smooth muscle contraction pathways (P=1.8×10−8) in our WGCNA analysis, coupled with HEYL expression in smooth muscle cells, suggests that vascular dysfunction is associated with the high-risk phenotype. Moreover, vascular smooth muscle cell plasticity in the TME is associated with various pathological processes. HEYL may regulate this plasticity, determining whether smooth muscle cells maintain their contractile phenotype or undergo transdifferentiation to synthetic or even macrophage-like phenotypes, each with distinct impacts on tumor progression. Integration with TME heterogeneity The heterogeneity of CAFs has emerged as a major challenge in developing stromal-targeted therapies (57,58). Single-cell studies have revealed multiple CAF subpopulations with distinct functions, including inflammatory CAFs (iCAFs) that promote chemoresistance through CXCL12 secretion and M2 macrophage polarization (59). Recent advances in single-cell and spatial transcriptomics have begun to elucidate the functional diversity of CAF subsets and their differential contributions to tumor progression and therapy resistance. Our five-gene signature appears to capture this heterogeneity by incorporating markers from different stromal populations, potentially explaining its reliable prognostic performance. Recent evidence shows that the GC TME composition changes with disease progression, with increases in immunosuppressive cells and angiogenesis-promoting cytokines (60,61). The temporal dynamics of our five genes during tumor progression remain to be elucidated, but their collective elevation in high-risk patients suggests they may mark an advanced state of stromal reprogramming associated with aggressive disease. The multifaceted functions of CAFs in the TME include cytokine secretion, ECM remodeling, metabolic reprogramming, and immune modulation (62). Our signature genes are associated with each of these processes: AKAP12 in cytokine signaling, FBXL2 in protein degradation and ECM turnover, NES in metabolic support, HEYL in vascular function, and MLLT11 in secretory processes. This comprehensive coverage of CAF functions may underlie the signature’s prognostic power. The stratified analyses across established molecular classifications (section of risk-score stratification across established molecular subtypes) provide additional cross-cohort biological support for this interpretation: the ACRG MSS/EMT subtype in GSE62254 and the TCGA GS subtype in TCGA-STAD—both defined by activation of stromal and mesenchymal programs rather than by CIN or mismatch-repair deficiency—consistently displayed the highest 5-gene risk scores, in line with recent GC prognostic and CAF-focused studies emphasizing the importance of stromal heterogeneity beyond classical molecular classification (63-65). Therapeutic implications: targeting the stromal network Disrupting the IL-34/CSF1R axis CAFs have emerged as promising therapeutic targets with significant advances in understanding their biology, despite challenges posed by their heterogeneity and complex functional roles (66). Many preclinical studies have identified promising targets on CAFs, demonstrating the efficacy of CAF-targeted treatments in preclinical models that have now advanced into clinical evaluation (66). Given our finding of stromal enrichment in high-risk GC patients, therapeutic strategies targeting CAFs—either through direct targeting, reprogramming, or blocking CAF-derived factors—may be warranted in combination with existing therapies, particularly for patients with high stromal signature scores. Metabolic interventions targeting NES-expressing cells The metabolic support provided by NES-expressing stromal cells suggests vulnerability to metabolic interventions. Recent studies in soft-tissue sarcomas have demonstrated that CAFs, particularly glycolytic CAFs (GlyCAFs), rely on GLUT1 (also known as SLC2A1)-dependent metabolism to impede cytotoxic T-cell infiltration into the tumor parenchyma, and that targeting glycolysis can decrease T-cell-restrictive CAF accumulation at the tumor margin, thereby enhancing T-cell infiltration and augmenting the efficacy of chemotherapy (67). Similarly, strategies targeting specific metabolic pathways in NES-expressing stromal cells, such as oxidative phosphorylation or glutathione synthesis, could potentially disrupt the metabolic symbiosis between stroma and tumor cells and overcome therapeutic resistance in GC. The HEYL-associated vascular abnormalities present opportunities for vascular normalization therapy. Combination of anti-angiogenic therapies with immunotherapy has shown promise in overcoming the immunosuppressive vascular microenvironment (68). Tumor endothelial cells are associated with immunosuppression through multiple mechanisms, including acting as gatekeepers of immune cell trafficking into the tumor compartment, metabolic competition via hypoxia and reactive oxygen species, and altered expression of immunomodulatory cytokines and cell-surface receptors (69). However, our finding of HEYL expression in smooth muscle cells suggests that targeting pericyte-endothelial cell interactions, rather than endothelial cells alone, may be more effective. Tissue-specific applicability and lack of pan-cancer universality Pan-cancer validation across 33 TCGA cohorts with FDR correction revealed tissue-specific rather than universal prognostic value. After Benjamini-Hochberg correction, only 4 of 33 cancer types retained statistical significance (FDR <0.05): GC (STAD, HR =1.6, FDR =0.041), BLCA, KIRP, and MESO. Notably, the initially observed enrichment in gastrointestinal malignancies was not statistically significant after multiple testing correction (Fisher’s P=1.0). This tissue-specific pattern likely reflects fundamental biological differences in stromal composition across cancer types. GC is characterized by extensive desmoplastic stroma with abundant CAFs, which aligns with the stromal-enriched nature of our signature genes. In contrast, hematologic malignancies and brain tumors have markedly different microenvironmental architectures, potentially explaining the lack of prognostic association in these cancer types. The limited cross-cancer applicability underscores that our signature captures GC-specific stromal programs rather than universal cancer hallmarks. This specificity, while limiting broad applicability, may actually enhance the potential translational relevance in GC by capturing disease-specific microenvironmental features that generic pan-cancer signatures might miss. Comparison with existing prognostic signatures Unlike previously published GC signatures that focus on EMT, cell cycle, or immune genes, our signature uniquely captures stromal reprogramming. Recent single-cell studies have identified SLIT2 as a CAF-derived factor promoting GC metastasis through NEK9 activation (70). Understanding CAF biology requires appreciation of their diverse origins and functional heterogeneity, as well as the recognition that some CAF functions may be antitumorigenic and should be preserved (71). While SLIT2 was not included in our signature, the identification of multiple stromal factors suggests that comprehensive stromal profiling, rather than single markers, is necessary for accurate prognosis. The pan-cancer validation showing particular strength in gastrointestinal cancers (HR =1.8, 95% CI: 1.46–2.22, P=9.7×10−7) suggests common stromal reprogramming mechanisms across the digestive system. The conservation of our signature across gastrointestinal cancers implies fundamental stromal responses to epithelial transformation that transcend organ boundaries. Limitations and future directions Several limitations should be acknowledged when interpreting these findings. First, because perioperative and adjuvant treatment information was not uniformly available at the patient level in the discovery cohort, an elevated risk score likely reflects a composite of two partially overlapping phenomena: a tumor-intrinsic propensity to recur, and a potentially less treatment-responsive stromal–myeloid microenvironment. The latter interpretation is supported by the coordinated pattern of higher predicted IC50 values across all eight agents tested (section of tumor immune microenvironment and drug sensitivity; all P<0.001) and by the observed association between signature components and an immunosuppressive stromal-myeloid context, including the positive MLLT11-M2 macrophage correlation (r=0.224, P=1.7×10−4; section of tumor immune microenvironment and drug sensitivity). This biological context is also compatible with the AKAP12⁺ fibroblast-IL-34/CSF1R-M2 macrophage axis reported in other solid tumors (51), although direct treatment-response validation in GC remains necessary. This dual interpretation does not diminish the potential translational relevance of the signature; rather, it suggests that high-risk patients may warrant closer postoperative surveillance and prospective evaluation in treatment-stratified or stromal-context-aware cohorts. Spatial context and cell-cell interactions While single-cell analysis revealed cell-type-specific expression patterns of our signature genes, we lack spatial information about the organization of these signature-expressing cells within the TME. Recent advances in spatial multi-omics have demonstrated that characterizing the spatial organization and interactions of CAFs within complex tissue environments provides critical insights into tumor biology and immunobiology, with CAF spatial subtypes exhibiting distinct spatial organizational patterns, neighboring cell compositions, and interaction networks (72). Integration of spatial transcriptomics with scRNA-seq can reveal how CAFs spatially coordinate with other cell populations to promote cancer progression (73). Future studies employing spatial transcriptomics or multiplex immunofluorescence are needed to map the spatial architecture of NES-expressing myeloid cells, AKAP12-expressing CAFs, and HEYL-expressing smooth muscle cells, and to understand how their spatial arrangement creates pro-tumorigenic microenvironmental niches. Temporal dynamics and clonal evolution The cross-sectional nature of our analysis cannot capture the temporal evolution of stromal reprogramming. Longitudinal studies tracking signature expression from early to advanced disease stages would clarify whether stromal changes precede or follow tumor progression. The heterogeneous nature of CAFs, arising from diverse cellular origins and shaped by tumor-stromal crosstalk, presents serious challenges for therapeutic targeting (74). Understanding the architecture of CAFs in the TME, including mapping their diverse origins and heterogeneity, is essential for developing effective strategies to overcome cancer therapy resistance (75). Lack of functional validation The most significant limitation is the absence of experimental validation establishing causal relationships between signature genes and clinical outcomes. While our comprehensive computational approach—bootstrap feature selection, WGCNA analysis, single-cell validation, and pan-cancer analysis—provides substantial correlative evidence, functional studies are essential to confirm mechanistic hypotheses. Future work should employ clustered regularly interspaced short palindromic repeats (CRISPR) screening in organoid-stromal co-cultures, genetically engineered mouse models with stromal-specific gene manipulation, and in vivo validation to establish causality. Clinical implementation Translation to clinical practice requires addressing practical challenges including RNA quality requirements, cost considerations, and integration with existing molecular classifications. Development of immunohistochemistry (IHC)-based scoring systems or targeted RNA panels would facilitate broader adoption. Prospective validation studies are needed to determine optimal assessment timing and potential clinical role in treatment decision-making. Tumor-microenvironment limitation of cell-line-based drug-sensitivity prediction A further methodological caveat concerns the pRRophetic drug-sensitivity predictions. The framework is trained on pharmacogenomic profiles derived from two-dimensional cancer cell lines (CCLE and GDSC), which by design lack the TME. Because our 5-gene signature is predominantly stromal-derived—as confirmed at single-cell resolution, with AKAP12 enriched in fibroblasts, HEYL in smooth-muscle cells, and NES in myeloid cells (section of single-cell validation and stromal enrichment)—predicted IC50 values should be interpreted as hypothesis-generating rather than as quantitative clinical estimates. Prospective validation in patient-derived organoid co-cultures, explant models, or in vivo systems that preserve stromal-tumor crosstalk will be required before these predictions can inform treatment decisions. The transparent within-subtype TCGA analyses (section of risk-score stratification across established molecular subtypes) further highlight that additional large, matched-endpoint external cohorts on RNA-seq platforms—particularly with recurrence-anchored follow-up rather than PFI/DFI surrogates—will be needed to establish within-subtype prognostic utility. Conclusions This study establishes that a five-gene signature predominantly expressed in stromal cells—CAFs (AKAP12), myeloid cells (NES), smooth muscle cells (HEYL), and secretory cells (MLLT11, FBXL2)—is associated with GC recurrence independent of TMB and traditional clinical factors (HR =3.074, 95% CI: 2.283–4.140, P=1.44×10−13). The predominant stromal expression of all signature genes indicates that microenvironment composition may contribute significantly to recurrence risk. The signature’s independence from TMB, coordinated associations with predicted chemotherapy resistance across eight agents, and distinct immune correlation patterns suggest that stromal-tumor interactions represent important determinants of clinical outcomes. Integration of bulk tissue, WGCNA, and single-cell analyses revealed that signature genes localize to specific stromal populations and are associated with ECM remodeling, vascular function, immune modulation, and metabolic support. While experimental validation is needed to establish causal mechanisms and therapeutic potential, this work provides evidence that comprehensive microenvironment characterization complements tumor-centric profiling for prognostic assessment. The identification of specific stromal cell populations associated with poor outcomes—AKAP12-expressing CAFs, NES-expressing myeloid cells, HEYL-expressing smooth muscle cells, and MLLT11-expressing secretory cells—provides targets for future mechanistic studies and potential therapeutic intervention. As precision oncology evolves, integration of both tumor genomic features and microenvironment characteristics may improve risk stratification and guide therapeutic strategies. Future work should focus on functional validation of stromal-tumor interactions, spatial characterization of signature-expressing cells, and prospective clinical validation to realize the translational potential of this stromal-derived prognostic signature. Supplementary The article’s supplementary files as Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-0518/prf