Disulfidptosis-related signatures identify SPARC+/CYR61+ stromal cells mediating immunosuppression in gastric cancer Authors: Yuxuan Xu, Xianming Kong, Jin Li, Dongxi Xiang Published: 2026 ABSTRACT AbstractBackgroundDisulfidptosis is a novel form of glucose starvation-induced cell death, yet its prognostic implications in gastric cancer (GC) remain largely undefined.MethodsWe integrated single-cell and bulk multi-omics data to identify core disulfidptosis-related genes. A robust Cox regression-based prognostic model was constructed. Machine learning combined with single-cell interaction analyses was used to delineate the mechanistic roles of key genes across different patient strata. Multi‑omics profiling was performed to characterize mutation landscapes and immune microenvironment features. Cell‑cell interaction analysis was conducted to explore ligand‑receptor pathways.ResultsSPARC, CYR61, and TIMP1 were identified as core disulfidptosis-related genes. High disulfidptosis scores correlated with IGHA1/IGHA2 mutations and a depleted immune microenvironment, whereas low scores were driven by LYZ amplification and active immune infiltration. Machine learning identified SPARC and CYR61 as core genes, predominantly expressed in endothelial cells and fibroblasts. Cell‑cell interaction analysis revealed that SPARC+/CYR61+ endothelial cells inhibited B‑cell function via the SELPLG‑SELL signaling pathway, while SPARC+/CYR61+ fibroblasts suppressed multiple immune populations (T cells, macrophages, mast cells, and B cells) through the SPP1‑CD44 signaling.ConclusionsDisulfidptosis features serve as critical prognostic biomarkers in gastric cancer and actively orchestrate immunosuppression within the tumor microenvironment. These insights present promising translational avenues for patient specific immunotherapeutic strategies.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s13402-026-01231-4.Keywords: Disulfidptosis, Prognostic signature, Immunosuppression, Stromal cells, Gastric cancer Introduction Gastric cancer (GC) is one of the most common malignancies worldwide and remains the third leading cause of cancer‑related mortality, accounting for approximately 770,000 deaths annually [1, 2]. Despite advances in diagnostic techniques and therapeutic strategies, including radical gastrectomy, perioperative chemotherapy, targeted therapy, and immune checkpoint inhibitors, the prognosis of patients with advanced GC remains unsatisfactory, with a 5‑year survival rate below 30% [3, 4]. The high degree of intertumoral and intratumoral heterogeneity in GC leads to variable treatment responses and highlights the urgent need for reliable prognostic biomarkers and novel therapeutic targets [5, 6]. Disulfidptosis is a recently identified form of regulated cell death that specifically occurs under glucose starvation or deprivation conditions [7]. This process is characterized by the abnormal accumulation of disulfide bonds in proteins, particularly within the actin cytoskeleton, leading to cytoskeletal collapse, cell contraction, and eventual plasma membrane detachment [8, 9]. Mechanistically, glucose starvation reduces NADPH production, which impairs the reduction of disulfides, causing excessive disulfide stress and F‑actin aggregation [9]. Unlike apoptosis, necroptosis, and ferroptosis, disulfidptosis represents a distinct cell death pathway that is highly sensitive to changes in NADPH levels and does not rely on known cell death executors [10]. Emerging evidence has implicated disulfidptosis in the pathogenesis and progression of various cancers, including colorectal cancer, lung adenocarcinoma, and hepatocellular carcinoma [11–14]. However, the precise roles of disulfidptosis‑related genes in the development, tumor microenvironment (TME) modulation, and prognosis of GC remain largely unexplored. Recent studies have demonstrated that disulfidptosis‑related gene signatures can serve as effective prognostic tools and predictors of immunotherapy response in multiple cancer types [15, 16]. For instance, in hepatocellular carcinoma, integrative analyses combining single‑cell RNA sequencing (scRNA‑seq) and bulk transcriptomic data have successfully established disulfidptosis‑based prognostic models [17]. Nevertheless, similar systematic investigations in GC are still lacking. The cellular heterogeneity of the GC TME, particularly the crosstalk between stromal cells and immune cells, may play a critical role in disulfidptosis‑mediated tumor progression and immune evasion [18]. Therefore, a comprehensive approach that integrates single‑cell and bulk omics data is urgently needed to uncover disulfidptosis‑related molecular signatures and their functional mechanisms in GC. In this study, we systematically analyzed single‑cell transcriptomic data combined with genomic and transcriptomic datasets from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) cohorts to identify disulfidptosis‑related prognostic genes in GC. Using Cox regression analysis, we constructed a robust prognostic model based on three genes (SPARC, CYR61, and TIMP1). Multi‑omics analyses revealed distinct characteristics between high- and low-disulfidptosis score groups. Machine learning algorithms integrated with single‑cell interaction analysis pinpointed SPARC and CYR61 as core disulfidptosis genes, which were specifically highly expressed in endothelial cells and fibroblasts. Moreover, we elucidated that SPARC⁺/CYR61⁺ endothelial cells suppress B‑cell immune function via the SELPLG‑SELL pathway, while SPARC⁺/CYR61⁺ fibroblasts inhibit multiple immune cell populations through the SPP1‑CD44 axis. Collectively, our findings highlight the critical role of disulfidptosis features in predicting GC prognosis and reveal a novel mechanism of disulfidptosis‑driven immunosuppression, offering new perspectives for GC therapy. Results Single‑cell transcriptomic landscape of disulfidptosis in GC To investigate the cellular distribution and prognostic significance of disulfidptosis in GC, we first analyzed single‑cell transcriptomic data from GC patients. After quality control and dimensionality reduction using t‑distributed stochastic neighbor embedding (t‑SNE), all cells were visualized according to their sample origins and tissue origins (Fig. 1A and B). Based on canonical cell surface markers, we annotated eight major cell types, including epithelial cells, fibroblasts, endothelial cells, T cells, B cells, plasma cells, macrophages, mast cells (Fig. 1C). The expression patterns of representative genes across different cell types are shown in Fig. 1D. We then calculated a disulfidptosis signature score for each single cell using a set of previously reported disulfidptosis-related genes [8]. The t-SNE plot revealed a heterogeneous distribution of disulfidptosis scores among individual cells (Fig. 1E, left). By applying the median score as a cutoff, we classified cells into high- and low-disulfidptosis groups (Fig. 1E, right). Notably, fibroblasts and endothelial cells exhibited significantly higher disulfidptosis scores compared to other cell types, whereas T cells and B cells showed relatively lower scores (Fig. 1F). These results suggest that disulfidptosis features are enriched in stromal cells within the GC tumor microenvironment. To evaluate the prognostic relevance of disulfidptosis at the bulk transcriptome level, we stratified GC patients from the TCGA and GEO cohorts based on the median disulfidptosis score derived from bulk RNA-seq data. Kaplan-Meier survival analysis demonstrated that patients with high disulfidptosis scores had worse overall survival compared to those with low scores, both in the TCGA cohort (Fig. 1G) and in the independent GEO validation cohort (Fig. 1G and H). Collectively, these findings indicate that disulfidptosis is not only enriched in specific stromal cell populations but also serves as a robust prognostic indicator in GC. Disulfidptosis features in single-cell transcriptomic data of GC. (A) t-SNE plot showing sample origins. (B) t-SNE plot showing tissue origins (tumor/normal). (C) t-SNE plot showing annotated cell types. (D) t-SNE plot showing gene expression levels in cells. (E) Left: t-SNE plot showing disulfidptosis signature scores in cells. Right: Classification of all cells into high- and low- disulfidptosis cells based on the scores in the left panel (red: high disulfidptosis score, blue: low disulfidptosis score). (F) Boxplot displaying disulfidptosis scores for each of the eight cell types. (G) Survival analysis comparing high- and low- disulfidptosis patients in the TCGA GC cohort. P value was calculated using the log-rank test. (H) Survival analysis comparing high- and low- disulfidptosis patients in the GEO GC cohort. P value was calculated using the log-rank test Identification of a disulfidptosis‑related prognostic signature To construct a disulfidptosis‑related prognostic signature for GC, we first identified differentially expressed genes (DEGs) between high- and low-disulfidptosis patients in the TCGA cohort. A total of 3,897 DEGs were identified, with 1,781 upregulated and 2,116 downregulated genes in the high‑disulfidptosis group (Fig. 2A, Supplementary Table 1). To refine the gene list, we intersected these DEGs with genes that exhibited differential expression between high- and low‑disulfidptosis cells from the single‑cell transcriptomic analysis. This intersection yielded 22 overlapping genes (Fig. 2B, Supplementary Table 1). Among these 22 genes, we further evaluated their prognostic value by performing univariate Cox regression analysis using the TCGA GC dataset. Three genes (SPARC, CYR61, and TIMP1) that were significantly associated with overall survival (highlighted in lavender, Fig. 2C). The forest plot displayed the hazard ratios (HRs) of all 22 genes ranked from highest to lowest, with SPARC (HR = 1.279, P = 0.0008), CYR61 (HR = 1.182, P = 0.0041), and TIMP1 (HR = 1.173, P = 0.0281) showing the most prominent risk effects (Fig. 2D). We then validated the prognostic performance of these three genes individually. Kaplan‑Meier survival analysis in the TCGA cohort demonstrated that patients with high expression of SPARC, CYR61, or TIMP1 had worse overall survival compared to those with low expression (Fig. 2E). These findings were further confirmed in the independent GEO validation cohort (Fig. 2F). To determine whether these three genes serve as independent prognostic factors, we performed univariate and multivariate Cox regression analyses. The univariate Cox regression forest plot confirmed that SPARC, CYR61, and TIMP1 were each significantly associated with poor overall survival (Fig. 2G). In the multivariate Cox regression model, SPARC and CYR61 remained relatively independent prognostic factors, whereas TIMP1 did not retain significance after adjusting for the other two genes (Fig. 2H). Construction of a disulfidptosis-related prognostic model. (A) Volcano plot showing differentially expressed genes between high- and low- disulfidptosis patients in the TCGA GC cohort. (B) Venn diagram showing overlapping differentially expressed genes between bulk transcriptomic and single-cell transcriptomic datasets. (C) Volcano plot showing prognosis-associated genes among the 22 overlapping genes (highlighted in lavender). (D) Forest plot showing hazard ratios of the 22 genes (ranked from highest to lowest). (E) Survival analysis comparing GC patients with high vs. low expression of SPARC, CYR61, and TIMP1 in the TCGA cohort. P value was calculated using the log-rank test. (F) Survival analysis comparing GC patients with high vs. low expression of SPARC, CYR61, and TIMP1 in the GEO cohort. P value was calculated using the log-rank test. (G) Forest plot shows the univariate Cox regression based on SPARC, CYR61, and TIMP1. (H) Forest plot shows the multivariate Cox regression based on SPARC, CYR61, and TIMP1 Validation of the disulfidptosis‑related prognostic model To evaluate the predictive performance of the three‑gene prognostic signature (SPARC, CYR61, and TIMP1), we calculated a risk score for each patient using the formula derived from the multivariate Cox regression model. Patients in the TCGA cohort were stratified into high‑risk and low‑risk groups based on the median risk score. As shown in Fig. 3A, patients with higher risk scores exhibited shorter overall survival times and a greater proportion of deaths compared to those with lower risk scores. This pattern was consistently observed in the independent GEO validation cohort (Fig. 3B). Kaplan-Meier survival analysis further demonstrated that high-risk patients had significantly worse overall survival than low-risk patients in both the TCGA cohort (Fig. 3C, P = 0.02) and the GEO cohort (Fig. 3D, P = 0.02). These results indicate that the disulfidptosis-related risk score effectively discriminates between patients with favorable and poor prognoses. We next assessed the predictive accuracy of the risk score using time‑dependent receiver operating characteristic (ROC) curves. In the TCGA cohort, the area under the curve (AUC) values for 1‑, 3‑, and 5‑year overall survival were 0.600, 0.610, and 0.727, respectively (Fig. 3E). In the GEO validation cohort, the AUC values for 1‑, 3‑, and 5‑year survival were 0.602, 0.611, and 0.617, respectively (Fig. 3F). These findings suggest that the prognostic model has moderate to good predictive power for long‑term survival outcomes in GC. To determine whether the risk score provides independent prognostic value beyond conventional clinical characteristics, we performed univariate and multivariate Cox regression analyses incorporating six clinical variables (risk score, disulfidptosis, age, gender, stage, MSI). The univariate analysis revealed that risk score, age and stage were significantly associated with overall survival (Fig. 3G, top). In the multivariate analysis, the risk score remained an independent prognostic factor after adjusting for all clinical covariates (Fig. 3G, bottom, HR = 2.501, P = 0.004). Finally, we compared the 5‑year survival predictive performance of the risk score against each clinical characteristic using ROC curves. The risk score achieved the highest AUC (0.727) among all individual clinical variables (Fig. 3H), further supporting its superior prognostic capability. Collectively, these results validate that the disulfidptosis‑related three‑gene signature is a robust and independent predictor of overall survival in GC, outperforming conventional clinical parameters. Validation of the disulfidptosis‑related prognostic model. (A-B) Risk scores (top) and survival time (bottom) in the TCGA and GEO cohorts. Black dotted lines represent the median risk score cutoff dividing patients into high‑ and low‑risk groups. Yellow and blue dots represent high‑ and low‑risk patients, respectively. (C-D) Survival analysis comparing high‑ and low‑risk patients in the TCGA (C) and GEO (D) cohorts. P value was calculated using the log-rank test. (E-F) ROC curves showing AUCs for 1‑, 3‑, and 5‑year overall survival (OS) for the risk score in the TCGA (E) and GEO (F) cohorts. (G) Forest plot showing univariate (top) and multivariate (bottom) Cox regression analyses based on six clinical characteristics. P < 0.05 was considered statistically significant. (H) ROC curves showing AUCs for 5‑year OS based on the six clinical characteristics in the TCGA cohort. AUC values are as follows: risk score = 0.727, disulfidptosis score = 0.545, age = 0.590, gender = 0.564, stage = 0.579, MSI = 0.496 Transcriptomic and genomic features of disulfidptosis‑related genes To further characterize the 22 disulfidptosis‑related overlapping genes (identified in Fig. 2B) in GC, we examined their expression patterns, functional enrichment, genomic alterations, and protein‑protein interactions. In the TCGA cohort, most of those genes were significantly upregulated in GC patients with high-disulfidptosis score compared with low-disulfidptosis score patients (Fig. 4A). This finding was consistently validated in the independent GEO cohort (Fig. 4B). To gain insight into the biological processes associated with these genes, we performed transcriptomic enrichment analysis. The results revealed that these genes were significantly enriched in immune-related signaling, such as antigen receptor-mediated signaling pathway, T cell receptor binding, PD-L1 expression and PD-1 checkpoint pathway in cancer (Fig. 4C), indicating that the disulfidptosis‑related signature may play a critical role in modulating the immune microenvironment of GC. We next investigated the genomic alteration profiles of key genes between high‑ and low‑disulfidptosis groups. Patients in the high-disulfidptosis score group exhibited point mutations in IGHA1 and IGHA2, which encode the constant regions of immunoglobulin A heavy chains and are essential for mucosal immunity, B-cell receptor signaling, and antibody-dependent cellular functions [19, 20]. Mutations in these genes may impair B-cell maturation and antibody production, thereby contributing to the reduced immune infiltration observed in this group. Conversely, the low-disulfidptosis score group displayed high-level amplification of LYZ, which encodes lysozyme, a key antimicrobial enzyme involved in innate immune defense, macrophage activation, and the regulation of antitumor immunity [21]. Lysozyme overexpression has been associated with enhanced immune cell recruitment and favorable prognosis in several cancers. These genomic alterations may therefore underlie the divergent immune landscapes and clinical outcomes between high- and low-disulfidptosis patients. Finally, we constructed a protein‑protein interaction (PPI) network to explore potential functional connections among these disulfidptosis‑related proteins and their interacting partners. As shown in Fig. 4F, these proteins formed a highly interconnected network, with central hubs such as LYZ and SPARC, connecting to immune modulators and extracellular matrix components respectively. This network supports the cooperative functions of disulfidptosis‑related genes in shaping the GC tumor microenvironment. Transcriptomic and genomic features of disulfidptosis genes. (A) Expression of disulfidptosis genes in the TCGA cohort. (B) Expression of disulfidptosis genes in the GEO cohort. (C) Transcriptomic enrichment results of disulfidptosis genes. (D) Genomic mutation landscape of disulfidptosis genes. (E) Copy number variation of disulfidptosis genes. (F) Protein-protein interaction network of disulfidptosis genes Immune landscape differences between high- and low-disulfidptosis groups Given that the disulfidptosis-related signature was enriched in immune-related pathways (Fig. 4C), we next sought to comprehensively characterize the immune microenvironment of GC patients stratified by their disulfidptosis scores. In the TCGA cohort, high-disulfidptosis patients exhibited lower immune scores and tumor purity, along with higher stromal scores, compared to low-disulfidptosis patients (Fig. 5A). Additionally, the stemness score was slightly depressed in the high-disulfidptosis group. Notably, the expression levels of multiple immune checkpoint genes, including CD274 and PDCD1, were slightly decreased in the high-disulfidptosis group. These findings were consistently validated in the independent GEO cohort (Fig. 5B). The decreased immune checkpoint expression in the high-disulfidptosis group suggests a potential immunosuppressive state that may contribute to poor prognosis. We next performed single‑sample gene set enrichment analysis (ssGSEA) to evaluate the relative abundance of various immune and stromal cell types. The high‑disulfidptosis group showed lower enrichment scores for most immune cell types, including CD4⁺ T cells, CD8⁺ T cells, and plasma cells, whereas the low‑disulfidptosis group exhibited higher enrichment for these populations (Fig. 5C and D). To quantify the differences in immune cell composition, we estimated the proportions of 22 immune cell subtypes using the CIBERSORT algorithm. The high‑disulfidptosis group had significantly lower proportions of CD8⁺ T cells and plasma cells, but higher proportions of M2 macrophages compared to the low‑disulfidptosis group in both TCGA and GEO cohorts (Fig. 5E and F). The shift towards an immunosuppressive cell profile further supports the notion that high disulfidptosis scores are linked to an immune‑excluded or immune‑suppressed tumor microenvironment. Overview of immune infiltration in high- vs. low-disulfidptosis groups. (A) Boxplots showing immune score, stromal score, tumor purity, stemness score, and immune checkpoint expression in high- vs. low-disulfidptosis patients in the TCGA cohort. (B) Boxplots showing the same parameters in high- vs. low-disulfidptosis patients in the GEO cohort. (C) Heatmap showing enrichment of immune and stromal cells in high- vs. low-disulfidptosis patients in the TCGA cohort. (D) Heatmap showing enrichment of immune and stromal cells in high- vs. low- disulfidptosis patients in the GEO cohort. (E) Boxplot showing differences in immune cell proportions between high- and low- disulfidptosis patients in the TCGA cohort. (F) Boxplot showing differences in immune cell proportions between high- and low‑disulfidptosis patients in the GEO cohort Machine learning identifies SPARC and CYR61 as core disulfidptosis-related genes To further pinpoint the most critical disulfidptosis-related genes driving the prognostic and immunological features observed above, we applied three independent machine learning algorithms to the 22 overlapping genes. The random forest algorithm identified five characteristic genes (Fig. 6A), the Surv-SVM method selected fourteen characteristic genes (Fig. 6B), and the Lasso regression identified four characteristic genes (Fig. 6C). By intersecting the results from these three approaches, we obtained two overlapping genes, SPARC and CYR61, as the optimal characteristic genes (Fig. 6D). These two genes were therefore considered the core disulfidptosis-related genes in GC. SPARC was significantly upregulated in tumor tissues compared to adjacent normal tissues (Fig. 6E). Moreover, their expression levels were markedly higher in the high-disulfidptosis group than in the low-disulfidptosis group, as demonstrated in both the TCGA and GEO cohorts (Fig. 6F and G). Patients with high expression of SPARC and CYR61 had significantly worse overall survival compared to those with low expression in both the TCGA and GEO cohorts (Fig. 6H-K), confirming that both genes serve as robust prognostic biomarkers. We next investigated the cellular localization of SPARC and CYR61 using single-cell transcriptomic data. A bubble plot revealed that both genes were predominantly expressed in fibroblasts and endothelial cells, with relatively low expression in epithelial cells and immune cells (Fig. 6L). The t-SNE plots further illustrated that SPARC and CYR61 were highly enriched in stromal cell populations (Fig. 6M and N), consistent with earlier observations that disulfidptosis features are concentrated in fibroblasts and endothelial cells (Fig. 1F). These findings suggest that SPARC and CYR61 may exert their prognostic effects by modulating the tumor microenvironment through stromal-immune cell crosstalk. Machine learning identifies characteristic genes of high- vs. low-disulfidptosis patients. (A) Random forest identified five characteristic genes. (B) Surv-SVM identified fourteen characteristic genes. (C) Lasso identified four characteristic genes. (D) Venn diagram showing the optimal characteristic genes. (E) Boxplot showing expression differences of SPARC and CYR61 between tumor and adjacent normal tissues in the TCGA cohort. (F) Boxplot showing expression of SPARC and CYR61 in high- vs. low-disulfidptosis patients in the TCGA cohort. (G) Boxplot showing expression of SPARC and CYR61 in high- vs. low-disulfidptosis patients in the GEO cohort. (H-I) Survival analysis comparing GC patients with high vs. low SPARC expression in the TCGA (H) and GEO (I) cohorts. P value was calculated using the log-rank test. (J-K) Survival analysis comparing GC patients with high vs. low CYR61 expression in the TCGA (J) and GEO (K) cohorts. P value was calculated using the log-rank test. (L) Bubble plot showing expression of SPARC and CYR61 across eight cell types. (M) t-SNE plot showing SPARC expression levels in cells. (N) t-SNE plot showing CYR61 expression levels in cells SPARC+/CYR61+ endothelial cells and fibroblasts suppress immune cells via distinct signaling pathways Given that SPARC and CYR61 were predominantly expressed in fibroblasts and endothelial cells (Fig. 6L-N), we next sought to investigate the potential cell-cell communication mediated by these two core disulfidptosis-related genes. We first reclustered endothelial cells and fibroblasts from the single-cell transcriptomic data to identify subpopulations with high expression of SPARC and CYR61. As shown in Fig. 7A, endothelial cells were classified into 11 subclusters, and the expression levels of SPARC and CYR61 were particularly enriched in cluster 5. Similarly, fibroblasts were reclustered into 11 subclusters, with SPARC and CYR61 co-expressed predominantly in cluster 9 (Fig. 7B). These two subclusters were designated as SPARC⁺/CYR61⁺ endothelial cells and SPARC⁺/CYR61⁺ fibroblasts for subsequent analyses. We then performed cell-cell interaction analysis using a ligand-receptor database to explore the communication networks among different cell types. The bubble plot in Fig. 7C shows the number of outgoing (horizontal axis) and incoming (vertical axis) ligand-receptor pairs for each cell type, revealing that SPARC⁺/CYR61⁺ endothelial cells and fibroblasts exhibited a high capacity for both sending and receiving signals. The pairwise interaction strength between cell clusters also demonstrated that SPARC⁺/CYR61⁺ endothelial cells and fibroblasts strongly interacted with other cell populations, with particularly robust homotypic interactions among themselves (Fig. 7D). To identify the specific signaling pathways involved, we analyzed the outgoing signaling patterns of SPARC⁺/CYR61⁺ cells. We found SPARC⁺/CYR61⁺ endothelial cells specifically emitted the SELPLG and SELE signaling pathways, while SPARC⁺/CYR61⁺ fibroblasts specifically emitted the SPP1 and NT signaling pathways. The reception of these signals by different cell types is illustrated in Fig. 7F, confirming that these pathways were actively received by various immune cells. Focusing on endothelial cell‑derived signals, we found that the SELPLG signaling pathway was predominantly transmitted from SPARC⁺/CYR61⁺ endothelial cells to B cells (Fig. 7G). The corresponding ligand‑receptor pair was identified as SELPLG‑SELL (Fig. 7H). The SELPLG-SELL axis is known to mediate leukocyte adhesion and trafficking, and its aberrant activation has been associated with B-cell immunosuppression in the tumor microenvironment [22]. Similarly, the SELE signaling pathway was also emitted by these endothelial cells and received by multiple cell types, particularly mast cells, with the ligand‑receptor pair SELE‑CD44 (Fig. 7I and J). SELE (E‑selectin) promotes immune evasion by facilitating the recruitment of immunosuppressive cells, including mast cells, to the tumor site [23]. Figure 7K further demonstrates that multiple ligands derived from SPARC⁺/CYR61⁺ endothelial cells, including SELPLG and SELE, were received by B cells and, to a lesser extent, by other immune cell types. Regarding fibroblast‑derived signals, the SPP1 signaling pathway was predominantly emitted by SPARC⁺/CYR61⁺ fibroblasts and received by a variety of immune cell populations, most notably mast cells, macrophages, B cells, and T cells. Additionally, macrophages also activated the SPP1 signaling pathway to act on other cells (Fig. 7L and N). The underlying ligand‑receptor interaction was SPP1‑CD44 (Fig. 7N). The SPP1‑CD44 axis is a well‑characterized immunosuppressive pathway that promotes M2 macrophage polarization, inhibits T cell activation, and supports immune evasion in multiple cancers [24]. Additionally, the NT signaling pathway (NTS‑NTRK1) was also identified as a fibroblast‑specific outgoing signal, which primarily targeted mast cells (Fig. 7M and N). Interactions of SPARC+/CYR61+ endothelial cells/fibroblasts with other cells. (A) Left: t-SNE plot showing reclustering of endothelial cells. Right: Expression of SPARC and CYR61 in different endothelial cell subclusters. (B) Left: t-SNE plot showing reclustering of fibroblasts. Right: Expression of SPARC and CYR61 in different fibroblast subclusters. (C) Bubble plot showing the number of ligand-receptor pairs for outgoing (horizontal axis) and incoming (vertical axis) signals in different cell subsets. (D) Heatmap showing interaction strength among different cell clusters. (E) Heatmap showing signaling pathways specifically emitted by SPARC+/CYR61+ endothelial cells/fibroblasts. (F) Heatmap showing reception of the signaling pathways shown in (E) by different cell types. (G) SELPLG signaling pathway interactions among different cell types. (H) Ligand-receptor interactions in the SELPLG signaling pathway. (I) SELE signaling pathway interactions among different cell types. (J) Ligand-receptor interactions in the SELE signaling pathway. (K) Bubble plot showing reception of ligands emitted by SPARC+/CYR61+ endothelial cells in different cell types. (L) SPP1 signaling pathway interactions among different cell types. (M) NT signaling pathway interactions among different cell types. (N) Bubble plot showing reception of ligands emitted by SPARC+/CYR61+ fibroblasts in different cell types Discussion In this study, we systematically dissected the clinical and immunological roles of disulfidptosis‑related genes in GC by integrating single‑cell transcriptomics, bulk multi‑omics data, and machine learning approaches. We identified three prognosis‑associated disulfidptosis genes (SPARC, CYR61, and TIMP1) and constructed a robust prognostic model that effectively stratified GC patients into high‑ and low‑risk groups with distinct survival outcomes. Multi‑omics analyses revealed that the high‑disulfidptosis group harbored specific genomic alterations (IGHA1 and IGHA2 mutations) and exhibited an immunosuppressive tumor microenvironment, characterized by reduced immune cell infiltration, lower immune scores, and enrichment of M2 macrophages. Machine learning further pinpointed SPARC and CYR61 as core disulfidptosis genes, predominantly expressed in endothelial cells and fibroblasts. Cell‑cell interaction analysis uncovered that SPARC⁺/CYR61⁺ endothelial cells suppress B cells and mast cells via the SELPLG‑SELL and SELE‑CD44 pathways, while SPARC⁺/CYR61⁺ fibroblasts inhibit multiple immune cell populations through the SPP1‑CD44 and NT‑NTRK1 axes. Collectively, our findings establish a disulfidptosis‑related prognostic signature and reveal a novel mechanism of stromal‑driven immunosuppression in GC. Disulfidptosis is an emerging form of regulated cell death distinct from apoptosis, ferroptosis, and necroptosis [8]. Although recent studies have explored disulfidptosis‑related signatures in various cancers, including lung adenocarcinoma and hepatocellular carcinoma [11, 13], its role in GC has remained largely unexplored. Our work is the first to systematically characterize the prognostic and immunological landscape of disulfidptosis in GC using an integrative multi‑omics strategy. Compared to previous studies that focused solely on bulk transcriptomic signatures [15, 16], our single‑cell resolution analysis allowed us to pinpoint the cellular sources of disulfidptosis‑related genes, identifying endothelial cells and fibroblasts as the primary expressors of SPARC and CYR61. This cellular precision is critical, as it shifts the paradigm from viewing disulfidptosis as merely a tumor‑cell‑autonomous process to recognizing its role in stromal‑immune crosstalk. The identification of SPARC, CYR61, and TIMP1 as prognostic biomarkers is consistent with their established functions in cancer progression. SPARC (secreted protein acidic and rich in cysteine) promotes extracellular matrix remodeling and has been linked to poor prognosis in multiple cancers [25]. CYR61 (cysteine‑rich angiogenic inducer 61) is involved in angiogenesis and metastasis, and its high expression often correlates with chemoresistance [26]. TIMP1 (tissue inhibitor of metalloproteinases 1) modulates matrix degradation and has been implicated in immune evasion [27]. However, our study goes beyond mere association by demonstrating that these genes are enriched in specific stromal subsets and that their expression is tightly linked to an immunosuppressive TME. The multivariate Cox regression analysis indicated that SPARC and CYR61 independently predict poor survival, whereas TIMP1 lost significance after adjustment, suggesting that the former two may be the key drivers of the disulfidptosis‑related prognostic signature. One of the most striking findings is the genomic alteration dichotomy between high- and low-disulfidptosis groups. High‑disulfidptosis patients exhibited point mutations in IGHA1 and IGHA2, which encode immunoglobulin heavy chain constant regions. Mutations in these genes may affect B‑cell receptor signaling and antibody production, potentially contributing to the observed B‑cell dysfunction and low immune infiltration [19]. Conversely, low‑disulfidptosis patients showed high‑level amplification of LYZ (lysozyme), a gene associated with innate immune activation and favorable prognosis [21]. These genomic differences provide a plausible explanation for the divergent immune landscapes and clinical outcomes between the two groups. Our cell‑cell interaction analysis revealed that SPARC⁺/CYR61⁺ endothelial cells suppress B cells through SELPLG‑SELL and SELE‑CD44 axes. SELPLG binding to SELL is classically involved in leukocyte homing, but recent evidence indicates that aberrant activation of this axis in the TME can induce B‑cell anergy and promote immune escape [22]. Similarly, SELE facilitates recruitment of immunosuppressive mast cells, which we observed as the primary recipients of SELE signaling [23]. Fibroblast‑derived SPP1 acted on a broad range of immune cells via CD44, promoting M2 macrophage polarization, inhibiting T‑cell activation, and supporting immune evasion [24]. These mechanistic insights highlight that disulfidptosis‑related genes orchestrate a multi‑cellular communication network that collectively suppresses anti‑tumor immunity. Despite these promising findings, several limitations should be acknowledged. Our analyses were primarily based on publicly available datasets and computational predictions, which may be subject to inherent biases and require experimental validation. Additionally, the prognostic model and the proposed cell‑cell communication axes need to be further verified in prospective cohorts and through functional assays. The precise molecular mechanisms by which SPARC and CYR61 regulate disulfidptosis and immunosuppression also warrant deeper investigation. Addressing these limitations in future studies will help consolidate the clinical and biological relevance of our findings. In conclusion, our study establishes a comprehensive disulfidptosis‑related prognostic signature for GC and reveals a novel paradigm wherein stromal SPARC⁺/CYR61⁺ endothelial cells and fibroblasts suppress immune cells through SELPLG, SELE, SPP1, and NT signaling pathways. These findings not only improve risk stratification for GC patients but also offer potential therapeutic targets to reverse immunosuppression. Future work should focus on experimental validation of the identified ligand‑receptor axes, exploration of therapeutic combinations (e.g., blocking SELPLG or SPP1 in conjunction with immune checkpoint inhibitors), and prospective clinical trials to assess the utility of the prognostic model in guiding personalized immunotherapy. Methods Data acquisition and processing Bulk RNA‑seq data and corresponding clinical information for GC patients were obtained from TCGA stomach adenocarcinoma (STAD) cohort and the GEO database (accession numbers: GSE15459). The TCGA dataset served as the training cohort, and the GEO dataset was used as an independent validation cohort. scRNA‑seq data of GC were downloaded from the GEO database (accession number: GSE183904). Single‑cell RNA‑seq data processing and analysis The scRNA-seq gene expression matrices of GC samples were converted into a Seurat object by the R package Seurat (version 4.3.0). Cells with < 200 detected genes and > 20% mitochondrial reads were filtered out. After quality control, the data were normalized using the NormalizeData function and mitochondrial read-counts by linear regression implemented using the ScaleData function. Dimensionality reduction was performed using principal component analysis (PCA). Batch effects among GC samples were removed using the harmony algorithm [28], and cells were visualized with t-distributed stochastic neighbor embedding (t-SNE). Cell types were annotated based on canonical marker genes. Calculation of disulfidptosis scores A disulfidptosis-related gene set was compiled from previously published studies [8]. For each single cell and bulk sample, ssGSEA was applied to compute disulfidptosis scores using the GSVA package (version 1.48.3) and the disulfidptosis-related gene set provided in Supplementary Table 2. Patients or cells were stratified into high- and low-disulfidptosis groups based on the median score. Identification of differentially expressed genes Differentially expressed genes between high- and low-disulfidptosis groups in the TCGA cohort were identified using the limma package (version 3.56.2). Genes with |log2fold change| > 0.5 and an adjusted P value < 0.05 were considered significantly differentially expressed. Overlapping genes between bulk DEGs and scRNA-seq derived DEGs (high vs. low disulfidptosis cells) were identified using Venn diagram analysis. Prognostic model construction and validation Univariate Cox regression analysis was performed on the 22 overlapping genes using the survival package (version 3.5.7) to identify genes significantly associated with overall survival. Multivariate Cox regression was then applied to the selected genes to construct a prognostic risk score formula: Risk score = Σ(coefficienti × expressioni). Patients were divided into high-risk and low-risk groups according to the median risk score. Kaplan-Meier survival curves were generated using the survminer package (version 0.4.9), and the log-rank test was used to compare survival differences. ROC curves were plotted to evaluate the predictive accuracy for 1-, 3-, and 5- year OS using the timeROC package (version 0.4). Univariate and multivariate Cox regression analyses were also performed to assess whether the risk score was an independent prognostic factor after adjusting for clinical covariates (disulfidptosis, age, gender, stage, and MSI status). Transcriptomic enrichment analysis Functional enrichment analysis of the 22 overlapping genes was performed using the clusterProfiler package (version 4.8.3). Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways with an adjusted P value < 0.05 were considered significantly enriched. Genomic alteration analysis Mutation data from TCGA were analized using maftools package (version 2.16.0). Only gene mutations present in at least one sample were retained and synonymous mutations were excluded. Copy number variation (CNV) data from TCGA were analized using gistic2 software (https://broadinstitute.github.io/gistic2/) with a q-value threshold of 0.25 to define significant gain and loss. Protein-protein interaction network A protein-protein interaction (PPI) network of the identified disulfidptosis-related genes was constructed using the STRING database (https://cn.string-db.org/) with a confidence score > 0.4. No additional experimental data (e.g., from BioGRID) were included. The network was visualized using Cytoscape software (version 3.7.2). Immune landscape analysis Immune scores, stromal scores, and tumor purity were calculated using the ESTIMATE algorithm and xCell package (version 1.1.0). Stemness scores were computed using the TCGAbiolinks software (version 2.28.4). The expression levels of immune checkpoint genes (e.g., CD274, PDCD1, CTLA4) were extracted from the bulk RNA-seq data. ssGSEA was performed to evaluate the enrichment scores of various immune and stromal cell types using gene sets from the xCell database (https://comphealth.ucsf.edu/app/xcell). The relative abundances of 22 immune cell types were estimated using the CIBERSORT algorithm. Machine learning for feature gene selection Three machine learning algorithms were applied to the 22 overlapping genes to identify core disulfidptosis-related genes. Random forest analysis was performed using the randomForest package (version 4.7.1.2). Surv-SVM (support vector machine for survival data) was implemented using the survivalSVM package (version 0.0.6). Lasso regression was conducted using the glmnet package (version 4.1.8) with 10-fold cross-validation. The overlapping genes selected by all three methods were considered optimal characteristic genes. Single-cell cell-cell interaction analysis To investigate cell-cell communication, we first reclustered endothelial cells and fibroblasts from the scRNA-seq data using the FindClusters function in Seurat. Subclusters with high expression of SPARC and CYR61 were identified. Cell-cell interaction analysis was performed using the CellChat package (version 2.2.0.9001) with the human ligand-receptor database. The number of outgoing and incoming signaling pairs, interaction strength, and specific signaling pathways (e.g., SELPLG, SELE, SPP1, NT) were analyzed. Statistical analysis All statistical analyses were performed using R software (version 4.3.1). Comparisons between two groups were performed using the Student’s t‑test. Survival differences were evaluated using the log‑rank test. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Acknowledgements We acknowledge the National Natural Science Foundation of China (82573430), Science and Technology Commission of Shanghai Municipality (23141901000), and 2024 Natural Science Foundation of Shanghai Science & Technology Commission (24ZR1412500) for supporting our research. Author contributions Y. X., and D. X. conceived the project. X. K. and J. L. joint the project discussion, J.L. supervised the clinical interpretation and translational relevance Y. X. did the bioinformatic analysis; Y. X., and D. X. wrote the manuscript. All authors read and were invited to make comments on the manuscript. Funding There was no funding. Data availability No datasets were generated or analysed during the current study. Declarations Ethical approval Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Jin Li, Email: 081106258@fudan.edu.cn. Dongxi Xiang, Email: dxiang@shsmu.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement No datasets were generated or analysed during the current study.