Single-cell analysis identifies a tumor-specific T-cell metabolic signature: prognostic model and association with immunosuppressive microenvironment in ovarian cancer

Introduction

Ovarian cancer (OC) is the leading cause of death among gynecologic malignancies, with major challenges including difficulty in early diagnosis and poor prognosis in advanced-stage patients. Although surgery combined with platinum-based chemotherapy remains standard treatment, most patients experience recurrence and develop platinum resistance, resulting in unfavorable outcomes (1). The advent of precision medicine and immunotherapy has revolutionized cancer treatment; however, OC demonstrates relatively low overall response rates to immune checkpoint inhibitors (ICIs). Despite the integration of multimodal therapies including anti-angiogenic agents and PARP inhibitors, clinical benefits remain highly variable across the patient population, emphasizing the need for precise biomarkers. This highlights the urgent need to understand the tumor microenvironment (TME) complexity to develop more effective therapeutic strategies and identify precise biomarkers (2). The TME is a multifaceted network of tumor cells, immune cells, stromal cells, and extracellular matrix (ECM), of which the composition and functional state critically determine tumor progression, metastasis, and therapeutic response (3). Within the TME, immune cells, particularly T cells, play central roles in antitumor immunity (4). However, tumors frequently induce T-cell dysfunction or exclude them from tumor core regions through various mechanisms, achieving immune evasion. While traditional bulk transcriptomic sequencing provides global gene expression profiles, it cannot resolve the high cellular heterogeneity within the TME, particularly rare subsets with distinct functional or prognostic significance (5). Consequently, applying single-cell RNA sequencing (scRNA-seq) to delineate T-cell subset composition and functional states in TME of OC, and investigating their prognostic associations (6), is an important direction in tumor immunology research.

Beyond cellular composition, T-cell metabolic status has been increasingly recognized as a critical determinant of antitumor function and therapeutic responsiveness. The TME typically features nutrient deprivation and hypoxia, affecting T-cell metabolic reprogramming and consequently influencing proliferation, differentiation, and effector functions (7). However, understanding of how tumor-specific T-cell metabolic characteristics in TME of OC impact disease progression and immunotherapy response remains insufficient (8). Identifying T-cell metabolic features closely associated with OC prognosis and elucidating their roles in shaping immunosuppressive TME will provide crucial theoretical foundations for developing novel prognostic biomarkers and targeted therapeutic strategies (9).

This study aimed to comprehensively characterize T-cell subset heterogeneity and metabolic profiles in TME of OC by integrating scRNA-seq, machine learning, and multi-omics immune analysis. We identified a unique tumor-specific T-cell subset (CA-T cells) and established a novel multi-gene risk scoring model (riskScore) based on its core genes. This model not only facilitates OC prognostic prediction but also demonstrates associations with immunosuppressive TME characteristics, offering new perspectives and potential directions for OC prognostic evaluation and individualized immunotherapy strategies. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0211/rc).

Methods

Processing and analysis of scRNA-seq data

This study analyzed the single-cell transcriptome dataset GSE184880, which is available from the Gene Expression Omnibus (GEO) repository. Rigorous quality control filtered cells based on gene count, UMI count, and mitochondrial gene proportion, removing low-quality cells. Quality-controlled data underwent normalization and identification of 3,000 highly variable genes (HVGs) using the Seurat package. Batch effects were corrected using the Harmony algorithm on principal component analysis (PCA) results, followed by UMAP dimensionality reduction and visualization. Cells were clustered at 0.3 resolution on Harmony-corrected data (10) and manually annotated using established cell-type markers: T cells (CD3D, CD3E, CD8A), fibroblasts (DCN, OGN), endothelial cells (PECAM1, CLDN5), epithelial cells (KRT18, EPCAM, KRT19, CD24), B cells/plasma cells (JCHAIN, CD79A), monocytes (CD14, C1QA, C1QB), smooth muscle cells/myofibroblasts (ACTA2, MYH11, TAGLN), and cycling cells (MKI67, TOP2A). Differentially expressed genes (DEGs) in T-cell subsets were detected with FindAllMarkers under strict thresholds (P<0.05, log₂FC >0.25, expression proportion >0.1). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Weighted gene co-expression network analysis (WGCNA)

WGCNA was applied to a metacell matrix from single-cell data in R to detect co-expression modules associated with CA-T cells and their functions. Soft threshold β=0.9 was selected according to the scale-free topology fit index (R²>0.85) and mean connectivity assessment to construct the co-expression network. Multiple gene co-expression modules were identified using the blockwiseModules function based on topological overlap matrix (TOM) (11). In module-trait correlation analysis, cell types were defined as binary traits, and Pearson correlation coefficients were calculated between module eigengenes (MEs) and each cell type. Hub genes within the core module significantly positively correlated with CA-T cells were identified. Finally, intersection analysis between genes from this core module and DEGs from the TCGA-OV cohort yielded core genes strongly associated with OC prognosis for riskScore construction.

Cox regression analysis

Cox regression analysis utilized the survival and survminer packages. Clinical survival data, including overall survival time (OS_TIME) and survival status (OS_STATUS), were obtained from The Cancer Genome Atlas Ovarian Cancer (TCGA-OV) cohort (12), accessible through the Genomic Data Commons (GDC) portal. Expression profiles of WGCNA-identified core genes were integrated with clinical survival data. Key statistics extracted included hazard ratio (HR), 95% confidence interval (CI), and P value, all reported to three decimal places or in scientific notation. Genes with P<0.05 were deemed statistically significant. Forest plots visualized univariate Cox regression results, clearly displaying HR and 95% CI for each significant gene. P values below 0.001 were reported as “<0.001”, while others were presented to three decimal places.

Machine learning analysis

To identify core prognostic genes, this study employed two machine learning approaches. First, the least absolute shrinkage and selection operator (LASSO) Cox regression with ten-fold cross-validation and partial likelihood deviance plots was implemented, selecting the optimal penalty parameter based on minimum cross-validation error (λ.min) to retain genes with non-zero regression coefficients as key predictors (13). Second, a random survival forest model with 3,000 trees was constructed to assess gene importance, extracting the top 25 genes by importance ranking (14). Finally, intersection analysis using Venn diagrams identified the core gene set from both methods for subsequent analyses.

Construction and evaluation of gene-based risk scoring model

To establish the riskScore and evaluate its predictive performance, expression data of eight core genes from the TCGA-OV cohort were integrated with clinical survival information. Multivariate Cox regression was applied to build the riskScore. Using the median riskScore, patients were separated into high- and low-risk groups, and survival outcomes were compared by Kaplan-Meier analysis with log-rank testing (15). Time-dependent receiver operating characteristic (ROC) analysis was applied to evaluate predictive accuracy, and area under the curve (AUC) values were determined for 1-, 3-, and 5-year OS. Additionally, individual Kaplan-Meier survival analyses validated the respective prognostic value of each gene within the model. All analyses were conducted in R statistical environment.

Construction and evaluation of OC prognostic nomogram

To develop a practical prognostic prediction tool for OC and evaluate its performance, a nomogram was constructed incorporating the riskScore and clinicopathological features. This nomogram provides intuitive predictions of 1-, 3-, and 5-year OS probability (16). Prediction accuracy and consistency were assessed through calibration curves with internal validation using 1,000 bootstrap resamples. Time-dependent ROC analyses quantified the nomogram’s discriminative ability by calculating AUC values for 1-, 3-, and 5-year OS predictions. To explore associations between riskScore and disease progression, boxplots comparing riskScore distributions across clinical stages were generated, with statistical comparisons performed using Wilcoxon rank-sum testing.

Tumor immune microenvironment analysis

To comprehensively characterize the TME of OC, multidimensional analyses of TCGA-OV gene expression data were performed. First, the ESTIMATE algorithm assessed stromal and immune infiltration levels within the TME, comparing differences between high- and low-risk groups (17). Second, CIBERSORT algorithm quantified relative abundances of 22 immune cell subsets in the TME, with stacked bar plots illustrating cellular composition and Spearman correlation analysis evaluating quantitative relationships between riskScore and immune cell abundances (18). Finally, to predict immune evasion and potential responsiveness to ICI therapy, TIDE algorithm analysis of TCGA-OV gene expression data assessed differences in TIDE metrics between high- and low-risk groups (19).

Prediction of immunotherapeutic response using immunophenoscore (IPS)

To predict immunotherapeutic responsiveness across risk strata, we employed the IPS algorithm. The IPS is a machine-learning-based framework that quantifies tumor immunogenicity to predict responsiveness to anti-PD-1/PD-L1 and anti-CTLA-4 therapies. The score integrates gene expression profiles across four determinant categories: major histocompatibility complex (MHC) molecules, immunomodulators, effector cells, and suppressor cells. IPS values for the TCGA-OV cohort were obtained from The Cancer Immunome Atlas (TCIA) database (https://tcia.at/). The IPS ranges from 0 to 10, with higher scores indicating greater immunogenicity and more favorable predicted immunotherapy response. IPS values were compared between the high- and low-risk groups across four clinical scenarios: baseline immunity (CTLA4-negative/PD1-negative), anti-PD-1 monotherapy (CTLA4-negative/PD1-positive), anti-CTLA-4 monotherapy (CTLA4-positive/PD1-negative), and combined checkpoint blockade (CTLA4-positive/PD1-positive). Between-group differences were assessed by the Wilcoxon rank-sum test.

Prediction of therapeutic sensitivity and gene-drug correlation analysis

To explore the therapeutic implications of the riskScore, we estimated drug sensitivity profiles for each patient in the TCGA-OV cohort. The oncoPredict R package was used to calculate estimated half-maximal inhibitory concentration (IC₅₀) values, trained on in vitro gene expression and drug response data from the Genomics of Drug Sensitivity in Cancer (GDSC1) dataset. A lower IC₅₀ value indicates greater drug sensitivity. We focused on standard chemotherapeutics (cisplatin, paclitaxel), metabolic inhibitors (phenformin, CPI-613), and epigenetic modulators (JQ1, vorinostat). IC₅₀ differences between risk groups were assessed by the Wilcoxon rank-sum test. Additionally, Spearman correlation analysis was performed to examine relationships between the expression of the 8 signature genes and predicted IC₅₀ values.

Statistical analysis

All statistical analyses were performed using R software (v4.4.0). Group differences were evaluated using Wilcoxon testing, and Pearson correlation coefficients assessed linear relationships. Statistical significance levels were defined as: *P<0.05, **P<0.01, ***P<0.001, ****P<0.0001, with “ns” indicating non-significant differences. These statistical methods and significance thresholds ensured analytical robustness and enhanced reliability of study conclusions (20).

Results

Identification of T-cell-associated gene expression profiles

Based on single-cell transcriptome data from GSE184880, batch effects were effectively corrected using the Harmony algorithm. Unsupervised clustering and UMAP dimensionality reduction identified ten distinct cell clusters (Figure 1A), which were categorized into seven major cell types using established marker genes (Figure 1B): T cells, fibroblasts, monocytes, epithelial cells, endothelial cells, cycling cells, and smooth muscle cells/myofibroblasts. Comparative analysis of cell composition revealed that T cells comprised a significantly higher proportion in tumor tissues (37.79%) versus normal tissues (13.66%), indicating their crucial role in the TME (Figure 1C). Further T-cell subset analysis identified a unique population, CA-T cells (cluster 6) (Figure 1D,1E), highly enriched in tumor tissues but virtually absent in normal tissues. This distinct distribution pattern suggests that the TME exerts specific recruitment and regulatory effects on T-cell subsets.

Figure 1 Single-cell transcriptomic landscape of the ovarian cancer tumor microenvironment. (A) UMAP visualization of 48,649 cells from the GSE184880 dataset, color-coded by eight cell lineages: T cells, fibroblasts, monocytes, epithelial cells, endothelial cells, cycling cells, B cells/plasma cells, and SMCs/myofibroblasts. (B) Dot plot of canonical marker gene expression across cell clusters. Dot size indicates percentage of expressing cells; color gradient represents average expression level. (C) Quantitative comparison of cellular composition between normal and cancer tissues. (D) T-cell sub-cluster distribution in cancer versus normal tissues. (E) Differential cell-type distribution across tissue origins. (F) GO enrichment analysis of CA-T cell-associated genes. Terms are categorized by BP, CC, and MF. Bubble size indicates gene count; color intensity represents adjusted P value. BP, biological process; CC, cellular component; GO, Gene Ontology; MF, molecular function; SMC; UMAP, Uniform Manifold Approximation and Projection.

Tumor-specific T-cell metabolic features drive OC prognosis and shape immunosuppressive microenvironment

WGCNA was performed on single-cell data, and a high-quality network was constructed using soft thresholding (Figure 2A), which identified 24 gene co-expression modules (Figure 2B). A strong association was observed between the green module and tumor-specific CA-T cells. Cross-analysis between genes in this module and T-cell DEGs yielded 218 core genes (Figure 2C). GO enrichment analysis of the 218 core genes revealed significant enrichment of mitochondrial processes (Figure 1F). The biological process (BP) category was dominated by electron transport chain and oxidative phosphorylation, while cellular component (CC) analysis confirmed predominant localization to mitochondrial complexes and ribosomes. Molecular function (MF) analysis revealed primary involvement in redox-driven transmembrane transport and electron transfer activity (7). These findings indicate that CA-T cells in the TME exist in a highly metabolically active state, with functional status closely linked to mitochondrial energy metabolism and redox regulation (21).

Figure 2 Weighted gene co-expression network analysis and key prognostic gene selection. (A) Analysis of the scale-free fit index for various soft-thresholding powers. The left panel shows the relationship between the soft-threshold power and the scale-free topology fit index (R²); the right panel displays the mean connectivity across different powers. (B) Module-trait correlation heatmap showing associations between co-expression modules and cell types. (C) Venn diagram of the intersection between green module genes and T-cell DEGs, yielding 218 core genes. (D) Forest plot of univariate Cox regression analysis identifying prognostic candidates from the CA-T cell-associated core genes. Genes with HR >1 (e.g., SFT2D1, CCNDBP1) represent risk factors; those with HR <1 (e.g., SNRPA1, SLAMF1) indicate favorable prognostic markers. CI, confidence interval; DEG, differentially expressed gene; HR, hazard ratio; WGCNA, weighted gene co-expression network analysis.

Cox regression analysis and machine learning for key gene identification

To identify core genes associated with OS in OC patients, univariate Cox proportional hazards regression analysis was performed. Results showed that high expression of five genes, including SFT2D1 (HR =1.302, 95% CI: 1.046–1.621, P=0.02) and CCNDBP1 (HR =1.243, 95% CI: 1.010–1.530, P=0.04), with HR>1 and P<0.05, was significantly associated with poor prognosis. Conversely, high expression of 20 genes, including SNRPA1 (HR =0.732, 95% CI: 0.599–0.894, P=0.002) and SLAMF1 (HR =0.711, 95% CI: 0.554–0.911, P=0.007), with HR <1 and P<0.05, correlated with prolonged survival (Figure 2D). These findings suggest that these genes may serve as potential prognostic biomarkers for OC.

Machine learning approaches were subsequently applied for further key gene selection. LASSO regression successfully identified 14 genes closely associated with OV patient prognosis (Figure 3A,3B). Random survival forest analysis with 10-fold cross-validation evaluated all genes, selecting the top 25 genes by importance ranking (Figure 3C). Finally, intersection analysis using Venn diagrams identified eight significant genes from both methods: SLAMF1, CXCR4, SFT2D1, SH3KBP1, SPOCK2, CDKN1B, GNPTAB, and SNRPA1 (Figure 3D).

Figure 3 Identification of core prognostic genes in ovarian cancer using integrated machine learning frameworks. (A,B) Identification of candidate predictors via LASSO Cox regression analysis. (A) Selection of the optimal penalty parameter λ was performed using 10-fold cross-validation based on partial likelihood deviance. The vertical dashed line on the left indicates the minimum cross-validation error (λ.min), used to select 14 genes with non-zero coefficients. (B) LASSO coefficient profiles of the candidate genes are plotted against the log(λ) sequence. (C) Assessment of gene importance using the RSF algorithm. The left panel illustrates the relationship between the OOB error rate and the number of survival trees (set to 3,000), demonstrating the convergence and stability of the model. The right panel displays the top 25 genes ranked by their relative variable importance scores, reflecting their contribution to the predictive accuracy of the survival model. (D) Venn diagram showing the consensus set of 8 genes identified by both LASSO and RSF. LASSO, least absolute shrinkage and selection operator; OOB, out-of-bag; RF; RSF, Random Survival Forest.

Prognostic analysis of OC based on multigene risk scoring model

Using selected key genes, the riskScore was developed to divide OC patients into high- and low-risk groups by median score. The high-risk group exhibited significantly reduced OS compared with the low-risk group, as demonstrated by Kaplan-Meier survival analysis (P<0.0001) (Figure 4B-4D). Risk tables and scatter plots confirmed the negative association between risk score and patient outcomes: higher risk scores corresponded with increased mortality and shorter survival times. The model’s predictive ability was tested through time-dependent ROC analysis, yielding AUCs of 0.707, 0.698, and 0.724 for 1-, 3-, and 5-year OS (Figure 4A), respectively, indicating high predictive performance. High expression of the eight model genes (SFT2D1, CDKN1B, CXCR4, GNPTAB, SH3KBP1, SLAMF1, SNRPA1, SPOCK2) was significantly associated with worse prognosis in OC, as revealed by individual Kaplan-Meier analysis (all P<0.05) (Figure 4E-4L).

Figure 4 Prognostic analysis of the multi-gene risk score model and core prognostic genes. (A) Time-dependent ROC curves for 1-, 3-, and 5-year OS prediction. (B) Risk score distribution across TCGA-OV patients ranked by ascending riskScore. Dashed lines indicate the median threshold separating high-risk (red) and low-risk (blue) groups. (C) Scatter plot of patient survival status versus risk score ranking. (D) Kaplan-Meier survival analysis; high-risk group exhibits significantly reduced OS (P<0.0001, log-rank test). Shaded areas represent 95% confidence intervals. (E-L) Kaplan-Meier survival curves for the eight core prognostic genes. Patients from the TCGA-OV cohort were stratified into high-expression (red) and low-expression (blue) groups based on the median expression level of each gene: GNPTAB (E), CDKN1B (F), SFT2D1 (G), SH3KBP1 (H), SLAMF1 (I), SNRPA1 (J), SPOCK2 (K), and CXCR4 (L). High expression of each gene is significantly associated with shorter OS (all P<0.05, log-rank test). AUC, area under the curve; FP, OS, overall survival; ROC, receiver operating characteristic; TCGA-OV, The Cancer Genome Atlas Ovarian Cancer; TIME; TP.

Integration of risk score and clinical features for OC prognostic nomogram construction

Univariate and multivariate Cox proportional hazards regression analyses integrating TCGA-OV clinical data and the riskScore were performed to evaluate its independent prognostic significance. Univariate analysis identified riskScore (HR =2.108, 95% CI: 1.721–2.583, P<0.001) and age (HR =1.024, 95% CI: 1.012–1.036, P<0.001) as risk factors. Multivariate analysis confirmed riskScore (HR =2.042, 95% CI: 1.666–2.503, P<0.001), age (HR =1.022, 95% CI: 1.010–1.034, P<0.001), and clinical stage (HR =1.344, 95% CI: 1.015–1.779, P=0.04) as independent prognostic factors for OS (Figure 5G,5H). Based on these independent factors, a nomogram was constructed to predict 1-, 3-, and 5-year OS probabilities. Model calibration was verified by calibration curve analysis, showing high concordance of predicted and actual survival (Figure 5A-5D). The nomogram’s discriminatory performance was supported by time-dependent ROC analysis, yielding AUCs of 0.729, 0.715, and 0.678 for 1-, 3-, and 5-year OS (Figure 5F). Additionally, riskScore correlated closely with disease progression. Boxplot analysis showed riskScore increased with advancing clinical stage, with statistically significant differences across all stages (all P<0.05) (Figure 5E).

Figure 5 Construction and validation of the integrated prognostic nomogram for ovarian cancer. (A) Nomogram integrating riskScore, age, and clinical stage for predicting 1-, 3-, and 5-year OS probabilities. (B-D) Calibration curves comparing predicted versus observed survival at 1 year (B), 3 years (C), and 5 years (D). The dashed diagonal represents ideal prediction. (E) Distribution of riskScore across FIGO stages (I–IV); significance assessed by Wilcoxon rank-sum test. (F) Time-dependent ROC curves for the nomogram (AUC: 0.729, 0.715, and 0.678 for 1-, 3-, and 5-year OS). (G,H) Forest plots of univariate (G) and multivariate (H) Cox regression identifying the riskScore as an independent prognostic factor (HR =2.042, P<0.001) after adjusting for age and clinical stage. AUC, area under the curve; CI, confidence interval; FIGO, International Federation of Gynecology and Obstetrics; FP, HR, hazard ratio; OS, overall survival; ROC, receiver operating characteristic; TP.

Tumor immune microenvironment analysis

Comparison of high- and low-risk patients through transcriptome-based TME analysis revealed significantly lower stromal, immune, and ESTIMATE scores in the high-risk group, as identified by ESTIMATE (all P<0.0001), indicating a “cold” tumor phenotype (Figure 6A). CIBERSORT analysis further refined these findings, revealing a trend toward reduced infiltration of multiple effector immune cells in the high-risk group, including CD8⁺ T cells, naive/memory B cells, and activated dendritic cells (Figure 6B-6D). As visualized in Figure 6C, where patients are ordered by increasing riskScore, the TME progressively transitions from an effector-rich state in low-risk patients to an immunosuppressive phenotype in high-risk patients, characterized by effector cell depletion and enrichment of M2-TAMs and myeloid-derived suppressor cells (MDSCs). TIDE analysis showed significantly elevated immune dysfunction and immune exclusion scores in the high-risk group (P<0.05), accompanied by increased immunosuppressive cell scores, including MDSCs and M2-type tumor-associated macrophages (M2-TAMs) (P<0.01) (Figure 6D). At the molecular level, the high-risk group exhibited decreased CD8⁺ T cell abundance and interferon-γ (IFN-γ) expression (P<0.001) (Figure 6E-6M). Notably, while high PD-L1 expression typically indicates poor prognosis, this study observed lower PD-L1 expression in the high-risk group (Figure 6M), suggesting that immune evasion occurs primarily through immune desert formation or exclusion rather than T-cell exhaustion. The TME in high-risk OC patients was characterized by inadequate effector immune cell infiltration, immunosuppressive cell enrichment, and diminished antitumor immune signaling. These findings provide mechanistic insights into the unfavorable prognosis of this population and may guide individualized therapeutic strategy development.

Figure 6 Characterization of the tumor immune microenvironment by the multi-gene risk score model. (A) Association between risk score and ESTIMATE Immune and Stromal scores. (B) Correlation between Immune score and Immune cell subtype abundance. (C) Overview of the tumor immune cell subtype composition. The x-axis represents individual patient samples from the TCGA-OV cohort, which are ordered from left to right according to their increasing riskScore within the high- and low-risk groups. Each vertical bar represents the fractional composition of 22 immune cell types for a single patient. (D) Differential infiltration of immune cell subtypes between risk groups. (E-G) TIDE analysis: overall TIDE score (E), immune dysfunction (F), and immune exclusion (G). (H) IFN-γ expression. (I) MDSC scores. (J) TIDE subscore. (K) M2-TAM infiltration. (L) CD8⁺ T-cell infiltration. (M) PD-L1 (CD274) expression. All comparisons between high- and low-risk groups; Wilcoxon rank-sum test. ns, non-significant differences; *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001. MDSC, myeloid-derived suppressor cell; TIDE, TCGA-OV, The Cancer Genome Atlas Ovarian Cancer.

Prediction of immunotherapeutic response via IPS analysis

Building on the observed “cold” tumor phenotype in high-risk patients, we assessed the predictive potential of the riskScore for immunotherapy response using the IPS. The low-risk group exhibited significantly higher IPS values across all four clinical scenarios: (I) baseline immunity (CTLA4-neg/PD1-neg; Figure 7A); (II) anti-PD-1 blockade (CTLA4-neg/PD1-pos; Figure 7B); (III) anti-CTLA-4 blockade (CTLA4-pos/PD1-neg; Figure 7C); and (IV) combined blockade (CTLA4-pos/PD1-pos; Figure 7D) (all P<0.01 or P<0.001, Wilcoxon rank-sum test). These results indicate that low-risk patients possess a more immunologically active tumor phenotype with greater potential for ICI benefit. Conversely, the markedly lower IPS in high-risk patients provides a mechanistic basis for their poor prognosis and limited ICI responsiveness, consistent with the impaired immune infiltration identified in our TME analysis.

Figure 7 Clinical therapeutic potential of the riskScore model. (A-D) Prediction of immunotherapeutic response via IPS analysis. Comparison of IPS values between high- and low-risk groups across four scenarios: (A) CTLA4-neg/PD-1-neg, (B) CTLA4-neg/PD-1-pos, (C) CTLA4-pos/PD-1-neg, and (D) CTLA4-pos/PD-1-pos. Higher IPS indicates greater tumor immunogenicity and predicted ICI responsiveness. (E-J) Estimated IC₅₀ values for standard chemotherapeutics [cisplatin (E), paclitaxel (F)], metabolic inhibitors [phenformin (G), CPI-613 (H)], and epigenetic modulators [JQ1 (I), vorinostat (J)]; Wilcoxon rank-sum test. (K) Spearman correlation matrix showing associations between expression of the eight core genes and estimated IC₅₀ values. Bubble size indicates correlation magnitude; color intensity represents statistical significance. *, P<0.05; **, P<0.01; ***, P<0.001. CTLA4, cytotoxic T-lymphocyte associated protein 4; IC₅₀, half-maximal inhibitory concentration; ICI, immune checkpoint inhibitor; IPS, immunophenoscore; PD-1, programmed cell death protein 1.

Prediction of therapeutic sensitivity to chemotherapy and targeted agents

We next estimated the IC₅₀ values for representative therapeutic agents using the oncoPredict algorithm to assess the riskScore’s utility in guiding treatment selection. The high-risk group showed significantly elevated IC₅₀ for Cisplatin (Figure 7E), indicating platinum resistance, whereas the estimated IC₅₀ for Paclitaxel showed no significant difference between groups (Figure 7F). Given the metabolic underpinning of our signature, we evaluated responses to metabolic inhibitors. The high-risk group demonstrated significantly lower IC₅₀ values for Phenformin (Figure 7G) and the TCA cycle inhibitor CPI-613 (Figure 7H), suggesting that mitochondrial metabolism may represent a therapeutic vulnerability in this population. Among epigenetic modulators, JQ1 sensitivity was comparable between groups (Figure 7I), while the high-risk group trended toward greater sensitivity to the HDAC inhibitor Vorinostat (Figure 7J). Spearman correlation analysis was performed between the expression of the 8 signature genes (SPOCK2, SNRPA1, SLAMF1, SH3KBP1, SFT2D1, GNPTAB, CXCR4, and CDKN1B) and the predicted drug sensitivities. The correlation matrix (Figure 7K) revealed significant associations between core gene expression and IC₅₀ values across evaluated agents. In particular, CXCR4, CDKN1B, and SPOCK2 exhibited strong correlations with sensitivity to both metabolic and epigenetic inhibitors. These findings validate the biological relevance of the model and suggest that these core genes may serve as candidate biomarkers for drug response in OC.

Discussion

OC is the most lethal gynecologic malignancy, with major challenges including difficulty in early diagnosis and poor prognosis in advanced-stage patients. Identification of reliable prognostic biomarkers and elucidation of TME complexity remain crucial for improving patient outcomes. This study constructed the riskScore through single-cell transcriptome sequencing combined with machine learning algorithms. Based on a tumor-specific T-cell subset (CA-T cells) highly enriched in OC single-cell data, eight key prognostic genes were identified through machine learning approaches. These genes primarily participate in mitochondrial energy metabolism processes. The riskScore based on their expression levels demonstrated robust predictive performance for OS at 1, 3, and 5 years, as confirmed by time-dependent ROC analysis. Integration of the multigene risk score with important clinical prognostic factors, including age and clinical stage, enabled nomogram construction. By quantifying each predictive factor’s contribution to patient prognosis, this nomogram provides clinicians with an intuitive and reliable tool for individualized prediction of 1-, 3-, and 5-year survival probabilities. High concordance between predicted and actual survival at all time points was demonstrated by calibration curve analyses, validating the tool’s accuracy, reliability, and clinical utility.

Eight core genes associated with OC prognosis were identified. High CXCR4 expression correlated with unfavorable outcomes in OC patients, consistent with previous findings. Based on published studies, this mechanism likely involves the CXCR4-CXCL12 signaling axis within the TME (22). Activation of this axis may promote cancer cell proliferation, migration, and angiogenesis via the MAPK pathway while recruiting and promoting M2-type macrophage maturation, thereby suppressing anti-tumor immune responses and facilitating tumor progression.

SPOCK2, an ECM proteoglycan, showed association with poorer OS when highly expressed, consistent with prior observations in high-grade serous OC (23). SPOCK2 may exert oncogenic effects by modulating matrix metalloproteinase (MMP) activity, promoting epithelial-mesenchymal transition (EMT), and activating PI3K/AKT signaling pathways, thereby contributing to tumor growth and metastasis (24).

CDKN1B encodes the classical tumor suppressor protein p27; however, our results demonstrated that its high expression in high-risk patients predicted unfavorable prognosis (25). This seemingly paradoxical observation aligns with reports indicating that cytoplasmic p27 expression in OC predicts poor prognosis. We propose that CDKN1B/p27 in CA-T cells may undergo subcellular relocalization or acquire non-classical tumor-promoting functions, leading to T-cell dysfunction or exhaustion. This finding provides important insights for further investigation of p27’s role in immune cell dysregulation.

High SLAMF1 expression in CA-T cells also correlated with unfavorable prognosis, potentially due to SLAMF1 functional inhibition within specific cell types and immunosuppressive microenvironments (26). In the immunosuppressive TME of high-risk OC, elevated SLAMF1 expression may impair anti-tumor immune responses by contributing to T-cell dysfunction. SFT2D1 has been associated with cuproptosis-dependent angiogenesis pathways and aggressive immunosuppressive phenotypes in cervical cancer, with high expression predicting poor prognosis (27). Given OC’s similar dependence on angiogenesis and metastasis, often accompanied by immunosuppressive TME, SFT2D1 likely exerts comparable tumor-promoting functions in OC. SH3KBP1 encodes an adaptor protein involved in essential cellular processes, including apoptosis, cytoskeletal reorganization, cell adhesion, and clathrin-mediated endocytosis. Recent studies demonstrated that SH3KBP1 knockout inhibited melanoma growth in mice and enhanced intratumoral CD8⁺ T-cell abundance and functionality (28). These findings suggest SH3KBP1 may promote OC progression by impairing CD8⁺ T-cell function or enhancing immunosuppression. GNPTAB, which regulates lysosomal enzyme trafficking, plays a recognized role in cancer progression. Lysosomal proteases, including cathepsins B and D, are highly active in malignant ovarian tumors and correlate with tumor burden and invasiveness (29). High GNPTAB expression may cause aberrant lysosomal function, including enhanced secretion of tumor-promoting lysosomal enzymes into the TME or metabolic reprogramming within CA-T cells, thereby facilitating tumor growth, invasion, and metastasis, ultimately resulting in poor prognosis. SNRPA1, a key spliceosome complex subunit, plays an essential role in RNA processing (30). Its overexpression has been documented across multiple cancers. In clear cell renal cell carcinoma (ccRCC), elevated SNRPA1 correlated with enhanced migration, invasion, poor prognosis, and altered immune regulation and drug sensitivity (31). In this study, high SNRPA1 expression may directly promote tumor cell proliferation and invasion while fostering immunosuppressive TME formation, contributing to unfavorable outcomes (32).

Our comprehensive TME analysis revealed that high-risk patients exhibited distinct immunosuppressive features consistent with “cold” tumor phenotypes, providing mechanistic insights into the biology underlying poor OC prognosis (33). In the high-risk group, stromal, immune, and ESTIMATE scores were significantly lower (all P<0.0001) according to ESTIMATE analysis, indicating decreased stromal and immune infiltration (34). Consistently, CIBERSORT analysis showed reduced infiltration of effector immune subsets such as CD8⁺ T cells, memory B cells, plasma cells, and activated dendritic cells. This broad effector immune cell depletion aligns with previous reports explaining limited tumor responsiveness to ICIs. “Cold” tumors with poor cytotoxic lymphocyte infiltration are generally associated with worse outcomes and reduced ICI sensitivity (33). TIDE analysis revealed profound immunosuppression in the high-risk group, showing significantly elevated immune dysfunction and exclusion scores, plus increased infiltration of MDSCs and M2-TAMs. M2-TAMs are tumor-promoting cells that suppress T-cell anti-tumor activity by secreting immunosuppressive cytokines and enzymes, including IL-10, TGF-β, and arginase-1, while facilitating regulatory T-cell (Treg) recruitment and differentiation (35,36). MDSCs are a heterogeneous population of immature myeloid cells with potent immunosuppressive capacity, contributing through multiple mechanisms to immunosuppressive environment establishment, effectively counteracting potential anti-tumor immune responses and driving the “cold” and immune-evasive phenotype observed in high-risk patients (37). The dominance of these immunosuppressive cell populations strongly correlates with poor prognosis. Therefore, therapeutic strategies targeting MDSCs and M2-TAMs—such as reprogramming toward M1 phenotypes or inhibiting their recruitment and function—show promise for reversing immunosuppression in high-risk OC (35). These findings highlight the potential of riskScore for identifying such patients and guiding treatment selection. Overcoming widespread immune cell deficiency may require multifaceted therapeutic approaches, such as enhancing immune cell infiltration through neoadjuvant chemotherapy to convert “cold” tumors into “hot” tumors. Consistent with this rationale, recent clinical trials have demonstrated the efficacy of combining ICIs with chemotherapy, PARP inhibitors, and anti-angiogenic agents (38,39). Anti-angiogenic therapy normalizes aberrant tumor vasculature and alleviates hypoxia, thereby remodeling the metabolic microenvironment to facilitate cytotoxic T-cell infiltration (40). Similarly, chemotherapy and PARP inhibitors augment immunogenicity through immunogenic cell death and cGAS-STING pathway activation, synergizing with immune checkpoint blockade (41). Given the metabolic basis of our 8-gene signature, the riskScore may help identify patients most likely to benefit from such combination regimens. Prospective validation in combination therapy cohorts is warranted to confirm the predictive utility of this metabolic risk model.

These findings have important clinical implications for precision diagnosis and treatment of OC. The riskScore effectively identifies patients with “cold” and highly immunosuppressive TMEs, who likely respond poorly to ICI monotherapy. Observed reductions in CD8⁺ T cells, PD-L1 expression, and IFN-γ levels in the high-risk group further support this conclusion. The counterintuitive low PD-L1 expression suggests that immune evasion in high-risk tumors occurs primarily through immune desert formation or exclusion rather than active T-cell exhaustion (42). This indicates that patients with “cold” tumors and low PD-L1 expression may benefit from strategies enhancing immune cell infiltration and overcoming immune exclusion (42,43). The model therefore enables precise risk stratification, providing clinicians with evidence for selecting multimodal combination therapies focused on reversing immunosuppressive states in high-risk patients. Furthermore, these findings address the clinical challenge of heterogeneous responses to ICI-based combination therapies. Although multimodal regimens have reshaped OC treatment, outcomes remain variable. Our IPS analysis demonstrated that the low-risk group consistently exhibits superior immunogenicity across multiple treatment scenarios, supporting the riskScore as a tool for identifying immunotherapy-responsive patients. For high-risk patients with ‘cold’ tumors, drug sensitivity analysis revealed selective therapeutic vulnerabilities. The observed sensitivity to metabolic inhibitors (CPI-613, phenformin) and a trend toward sensitivity to the HDAC inhibitor vorinostat suggest that these agents could be incorporated into combination regimens to convert immunosuppressive TMEs toward an immune-permissive state. Together, these findings underscore the potential of the riskScore for guiding individualized therapeutic strategies in OC.

Despite these advances, certain limitations should be acknowledged. First, this study relied primarily on retrospective bioinformatics analyses of public datasets, lacking independent external validation cohorts and prospective clinical validation, which limits the generalizability and robustness of riskScore. Second, computational approaches including CIBERSORT and TIDE are derived from transcriptomic profiles and may not fully capture complex cellular heterogeneity, functional states, and spatial architecture within the TME. Future research should validate the inferred immune cell subsets, functional states, and spatial interactions using experimental approaches such as flow cytometry, multiplex immunofluorescence (mIF), or spatial transcriptomics (44). Additionally, this study relied solely on gene expression data. Future investigations should integrate proteomic and metabolomic data to comprehensively characterize key gene functions and regulatory networks, providing stronger biological evidence. Furthermore, the current riskScore relies on bulk RNA-sequencing data from tumor biopsies. Given the invasive nature and spatiotemporal limitations of tissue sampling, adapting this 8-gene signature for liquid biopsy platforms represents a promising direction. Quantifying these metabolic gene transcripts in blood samples—via circulating tumor RNA (ctRNA) or tumor-derived exosomes—could enable dynamic riskScore monitoring and real-time assessment of therapeutic response. Finally, the clinical value of the riskScore and its TME associations requires confirmation through large-scale, multicenter prospective studies before clinical implementation.

Conclusions

By integrating single-cell transcriptomics, machine learning, and multi-omics immune infiltration analyses, this study revealed the complex heterogeneity of TME of OC and its prognostic impact. A tumor-specific T-cell subset (CA-T cells) was identified from single-cell data, and based on its core genes, the riskScore was developed. This model demonstrated robust accuracy and independent prognostic value for predicting the OS of OC and positively correlated with clinical stage.

High-risk group tumors exhibited a characteristic “cold” phenotype with reduced stromal and immune cell infiltration. Immune analyses revealed that high-risk tumors were characterized by markedly decreased effector immune cell abundance and enrichment of immunosuppressive cells, including MDSCs and M2-TAMs, along with elevated immune dysfunction and exclusion scores. Additionally, reduced CD8+ T-cell levels and decreased IFN-γ expression in the high-risk group were consistent with poor prognosis and indicated limited responsiveness to ICI therapy.

This study provides novel prognostic biomarkers for OC and offers mechanistic insights into poor outcomes and immune evasion mechanisms. These findings may help refine therapeutic strategies by identifying patient populations less likely to respond to current immunotherapies while informing development of combination strategies targeting immunosuppressive TME. Such approaches hold promise for improving clinical outcomes in OC patients.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0211/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0211/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0211/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Yoshino Y, Furusawa A, Nara K, et al. Optimizing treatment for platinum-resistant ovarian clear cell carcinoma: Efficacy of gemcitabine and combination therapy with bevacizumab. Cancer 2025;131:e70071. [Crossref] [PubMed]
2. Yi L, Pan H, Ning Z, et al. Clinical and biomarker analyses of SHR-1701 combined with famitinib in patients with previously treated advanced biliary tract cancer or pancreatic ductal adenocarcinoma: a phase II trial. Signal Transduct Target Ther 2024;9:347. [Crossref] [PubMed]
3. Tang L, Mei Y, Shen Y, et al. Nanoparticle-Mediated Targeted Drug Delivery to Remodel Tumor Microenvironment for Cancer Therapy. Int J Nanomedicine 2021;16:5811-29. [Crossref] [PubMed]
4. Zhang Y, Zhang Z. The history and advances in cancer immunotherapy: understanding the characteristics of tumor-infiltrating immune cells and their therapeutic implications. Cell Mol Immunol 2020;17:807-21. [Crossref] [PubMed]
5. Lee H, Kim B, Park J, et al. Cancer stem cells: landscape, challenges and emerging therapeutic innovations. Signal Transduct Target Ther 2025;10:248. [Crossref] [PubMed]
6. Zhao Q, Shao H, Zhang T. Single-cell RNA sequencing in ovarian cancer: revealing new perspectives in the tumor microenvironment. Am J Transl Res 2024;16:3338-54. [Crossref] [PubMed]
7. Liu X, Hoft DF, Peng G. Tumor microenvironment metabolites directing T cell differentiation and function. Trends Immunol 2022;43:132-47. [Crossref] [PubMed]
8. Gao F, Shah R, Xin G, et al. Metabolic Dialogue Shapes Immune Response in the Tumor Microenvironment. Eur J Immunol 2025;55:e202451102. [Crossref] [PubMed]
9. Du H, Xu T, Yu S, et al. Mitochondrial metabolism and cancer therapeutic innovation. Signal Transduct Target Ther 2025;10:245. [Crossref] [PubMed]
10. Korsunsky I, Millard N, Fan J, et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat Methods 2019;16:1289-96. [Crossref] [PubMed]
11. Ding Y, Niu X, Guo P, et al. Single-Cell Data and Weighted Correlation Network Analysis Revealed the Regulatory Mechanisms of Macrophages in Carotid Plaques. J Immunol Res 2025;2025:9987367. [Crossref] [PubMed]
12. Liu K, Xu Y, Mao Y, et al. Systematic pan-cancer analysis identifies PKNOX1 as a potential prognostic and immunological biomarker and its functional validation. Front Immunol 2025;16:1533690. [Crossref] [PubMed]
13. Luo B, Yang M, Han Z, et al. Establishment of a Nomogram-Based Prognostic Model (LASSO-COX Regression) for Predicting Progression-Free Survival of Primary Non-Small Cell Lung Cancer Patients Treated with Adjuvant Chinese Herbal Medicines Therapy: A Retrospective Study of Case Series. Front Oncol 2022;12:882278. [Crossref] [PubMed]
14. Shi Q, Zeng Y, Xue C, et al. Development of a promising PPAR signaling pathway-related prognostic prediction model for hepatocellular carcinoma. Sci Rep 2024;14:4926. [Crossref] [PubMed]
15. Li W, Ding Z, Wang D, et al. Ten-gene signature reveals the significance of clinical prognosis and immuno-correlation of osteosarcoma and study on novel skeleton inhibitors regarding MMP9. Cancer Cell Int 2021;21:377. [Crossref] [PubMed]
16. Zhang X, Huang Y, Li Q, et al. Senescence risk score: a multifaceted prognostic tool predicting outcomes, stemness, and immune responses in colorectal cancer. Front Immunol 2023;14:1265911. [Crossref] [PubMed]
17. Zhang D, Wang Y, Yang Q. A High Epigenetic Risk Score Shapes the Non-Inflamed Tumor Microenvironment in Breast Cancer. Front Mol Biosci 2021;8:675198. [Crossref] [PubMed]
18. Lin MJ, Tang XX, Yao GS, et al. A novel 7-chemokine-genes predictive signature for prognosis and therapeutic response in renal clear cell carcinoma. Front Pharmacol 2023;14:1120562. [Crossref] [PubMed]
19. Ravensbergen CJ, Polack M, Roelands J, et al. Combined Assessment of the Tumor-Stroma Ratio and Tumor Immune Cell Infiltrate for Immune Checkpoint Inhibitor Therapy Response Prediction in Colon Cancer. Cells 2021;10:2935. [Crossref] [PubMed]
20. Fefeu M, Blatzer M, Kneppers A, et al. Serotonin reuptake inhibitors improve muscle stem cell function and muscle regeneration in male mice. Nat Commun 2024;15:6457. [Crossref] [PubMed]
21. Jiang S. Mitochondrial oxidative phosphorylation is linked to T-cell exhaustion. Aging (Albany NY) 2020;12:16665-6. [Crossref] [PubMed]
22. Guo Q, Ding H, Zheng Q. Single-cell transcriptomic construction of fibroblast score for analysis of immune infiltration in primary and metastatic ovarian cancer. Front Genet 2025;16:1549541. [Crossref] [PubMed]
23. Zheng X, Wu H, Zhu W. SPOCK2 promotes the invasion and migration of ovarian cancer cells through FAK signaling pathway. J Gynecol Oncol 2025;36:e98. [Crossref] [PubMed]
24. Moghbeli M. PI3K/AKT pathway as a pivotal regulator of epithelial-mesenchymal transition in lung tumor cells. Cancer Cell Int 2024;24:165. [Crossref] [PubMed]
25. Nowosad A, Besson A. CDKN1B/p27 regulates autophagy via the control of Ragulator and MTOR activity in amino acid-deprived cells. Autophagy 2020;16:2297-8. [Crossref] [PubMed]
26. Li J, Fan T, Wang D, et al. SLAMF receptors: key regulators of tumor progression and emerging targets for cancer immunotherapy. Mol Cancer 2025;24:145. [Crossref] [PubMed]
27. Kang J, Jiang J, Xiang X, et al. Identification of a new gene signature for prognostic evaluation in cervical cancer: based on cuproptosis-associated angiogenesis and multi-omics analysis. Cancer Cell Int 2024;24:23. [Crossref] [PubMed]
28. Sun X, Guo J, Zhao N, et al. Clathrin Light Chain B Drives Hepatocellular Carcinoma Progression Through Dual Mechanisms: Small Extracellular Vesicle-Mediated Angiogenesis and the NF-κB-PCLAF Signaling Axis. Adv Sci (Weinh) 2025;12:e08613. [Crossref] [PubMed]
29. Skidmore ZL, Kunisaki J, Lin Y, et al. Genomic and transcriptomic somatic alterations of hepatocellular carcinoma in non-cirrhotic livers. Cancer Genet 2022;264-265:90-9. [Crossref] [PubMed]
30. Yang JJ, Yang YJ, Gu YL, et al. High SNRPA1 expression leads to poor prognosis in patients with lung adenocarcinoma. Clin Respir J 2023;17:719-32. [Crossref] [PubMed]
31. Fish L, Khoroshkin M, Navickas A, et al. A prometastatic splicing program regulated by SNRPA1 interactions with structured RNA elements. Science 2021;372:eabc7531. [Crossref] [PubMed]
32. Ye L, Long C, Xu B, et al. Multi-omics identification of a novel signature for serous ovarian carcinoma in the context of 3P medicine and based on twelve programmed cell death patterns: a multi-cohort machine learning study. Mol Med 2025;31:5. [Crossref] [PubMed]
33. Arcieri M, Capezzali E, Restaino S, et al. Study of the Role of the Tumor Microenvironment in Ovarian Cancer (MICO): A Prospective Monocentric Trial. Cancer Rep (Hoboken) 2025;8:e70242. [Crossref] [PubMed]
34. Song Z, Gui S, Xiao S, et al. A novel anoikis-related gene signature identifies LYPD1 as a novel therapy target for bladder cancer. Sci Rep 2024;14:3198. [Crossref] [PubMed]
35. Xu C, Chen J, Tan M, et al. The role of macrophage polarization in ovarian cancer: from molecular mechanism to therapeutic potentials. Front Immunol 2025;16:1543096. [Crossref] [PubMed]
36. Wang Y, Ma C, Li X, et al. Unraveling the role of M2 TAMs in ovarian cancer dynamics: a systematic review. J Transl Med 2025;23:623. [Crossref] [PubMed]
37. Chow A, Perica K, Klebanoff CA, et al. Clinical implications of T cell exhaustion for cancer immunotherapy. Nat Rev Clin Oncol 2022;19:775-90. [Crossref] [PubMed]
38. Li C, Qi X, Yan M. Chemotherapy-induced immunogenic cell death in combination with ICIs: a brief review of mechanisms, clinical insights, and therapeutic implications. Front Pharmacol 2025;16:1572195. [Crossref] [PubMed]
39. Finn RS, Qin S, Ikeda M, et al. Atezolizumab plus Bevacizumab in Unresectable Hepatocellular Carcinoma. N Engl J Med 2020;382:1894-905. [Crossref] [PubMed]
40. Cabezón-Gutiérrez L, Custodio-Cabello S, Palka-Kotlowska M, et al. Immune-checkpoint inhibitor resistance in patients with cancer. Explor Immunol 2025;5:1003196.
41. Ding L, Kim HJ, Wang Q, et al. PARP Inhibition Elicits STING-Dependent Antitumor Immunity in Brca1-Deficient Ovarian Cancer. Cell Rep 2018;25:2972-2980.e5. [Crossref] [PubMed]
42. Zhong T, Sun S, Zhao M, et al. The mechanisms and clinical significance of CD8(+) T cell exhaustion in anti-tumor immunity. Cancer Biol Med 2025;22:460-80. [Crossref] [PubMed]
43. Liang H, Zhang S. Thrombospondin-1 induces CD8(+) T cell exhaustion and immune suppression within the tumor microenvironment of ovarian cancer. J Ovarian Res 2025;18:99. [Crossref] [PubMed]
44. Li S, Jiang B, Zhou H, et al. Development of a prognostic immune cell-based model for ovarian cancer using multiplex immunofluorescence. J Transl Med 2025;23:688. [Crossref] [PubMed]