Pan cohort immune biomarker of CD8 lymphocyte activation enabling HGSOC outcome prediction and treatment response

Abstract

Background

High-grade serous ovarian cancer (HGSOC) exhibits poor prognosis due to late diagnosis, chemoresistance, and limited responses to immune checkpoint inhibitors. Although tumor‐infiltrating CD8⁺ T cells correlate with improved survival, current prognostic models remain inadequate. Thus, robust biomarkers linked to CD8⁺ T cell activation are urgently needed to guide clinical management.

Methods

Transcriptomic and clinical profiles from 874 late-stage HGSOC patients were analyzed via single-sample gene set enrichment analysis for immune infiltration and weighted gene co-expression network analysis to identify CD8⁺ T cell-associated genes. An integrative machine learning approach was employed to develop a CD8⁺ T cell-associated immune prognostic signature (CIPS), which was then validated across multiple independent cohorts and benchmarked against 56 published models. CIPS was further characterized using single-cell RNA-seq analysis.

Results

The resulting 10-gene signature independently predicted overall survival in all cohorts and consistently surpassed most clinicopathological variables and comparator models. Low-risk patients exhibited significantly enhanced CD8⁺ T cell and cytotoxic gene scores, correlating with better responses to chemotherapy and immunotherapy. CIPS inversely correlated with tumor-mutation burden, BRCA1/2 mutations and homologous-recombination deficiency. Single-cell analysis localized signature genes to T lymphocyte and myeloid compartments and linked elevated CIPS activity to augmented intercellular communication in platinum-resistant tumors.

Conclusion

CIPS captures a CD8⁺ T cell activation program that powerfully stratifies late-stage HGSOC, forecasts therapeutic benefit and offers a practicable biomarker for personalized immuno-oncology strategies.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-04363-5.

Background

Epithelial ovarian cancer is widely recognized as a highly lethal form of gynecologic malignancy and occupies the fifth position among the main drivers of cancer-induced mortality in women [1, 2]. High-grade serous ovarian cancer (HGSOC) stands out as the prevailing histopathological subtype of epithelial ovarian cancer, comprising 60–80% of all documented cases [2, 3]. A substantial portion of patients are already in advanced stages at the time of diagnosis, and subsequent treatment resistance or relapse is a frequent occurrence following initial chemotherapy, significantly impacting the 5-year survival rate which stands at approximately 30% [4, 5]. Currently, the recommended approach for managing HGSOC patients remains tumor-cell surgical debulking followed by a chemotherapy regimen utilizing platinum and taxane-based agents [2, 6]. Despite the potential of HGSOC to exhibit a positive response to immunotherapy based on its endogenous immunity at the molecular or T cell level, the effectiveness of immunotherapy in treating this specific malignancy has not lived up to expectations thus far [2].

CD8⁺ cytotoxic T lymphocytes (CTLs) assume an essential role as the primary effector cells in antitumor immune responses by directly killing tumor cells, and CD8⁺ intraepithelial tumor-infiltrating lymphocytes (TILs) have been convincingly established as a hallmark of immune attack [7, 8]. Among all types of ovarian cancers, HGSOC exhibits the highest propensity for significant infiltration of CD8⁺ T cells. Moreover, irrespective of the surgical cytoreduction extent, chemotherapy regime, or germline BRCA1 mutation status, the presence of CD8⁺ intraepithelial TILs has been identified as a favorable prognostic factor [9, 10]. Single-cell transcriptomic analysis of HGSOC reveals the presence of a substantial number of T lymphocytes in the primary tumor lesions along with the metastatic sites [11–13]. Nevertheless, the documented outcomes of PD-1 or PD-L1 blockade, along with the utilization of anti-PD-L1 antibodies in combination with chemotherapy, have consistently yielded unsatisfactory results in patients diagnosed with ovarian cancer [2, 6, 14–16]. Despite the observed existence and prognostic implications of T cell infiltration in HGSOC, the specific mechanisms responsible for the insufficient response to immune checkpoint blockade (ICB) remain predominantly obscure [11]. Consequently, the discovery of promising prognostic markers associated with CD8⁺ T cell activation in HGSOC holds paramount importance in the pursuit of efficacious novel therapies, particularly those capable of fostering robust immune engagement.

The study sought to construct and substantiate a risk assessment signature in 874 late-stage HGSOC patients across four large independent public datasets based on CD8⁺ T cell activation status-related genes, with the aim of assessing the prognosis of late-stage HGSOC. This work holds promise in terms of enhancing the precision of treatment options based on CD8⁺ T cell activation, thus potentially having a positive impact on the prognostic results of late-stage HGSOC.

Methods

Transcriptomic data collection and processing

In this study, four independent ovarian cancer cohorts from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) were retrospectively incorporated. The data included in the study satisfied the designated conditions: (1) Primary HGSOC; (2) Availability of transcriptomic data and patient information; (3) FIGO stage Ⅲ/Ⅳ; (4) Absence of prior administration of chemotherapy or radiotherapy; (5) Each cohort comprises more than 100 HGSOC patients with an overall survival (OS) duration exceeding one month; (6) Utilization of whole transcriptomic platforms. A sum of 874 HGSOC patients with intact survival records were retained, separately included in TCGA-OV (n = 364), GSE32062 [17] (n = 260), GSE9891 [18] (n = 140) and GSE17260 [19] (n = 110). The TCGA-OV dataset was used for signature development, while other GEO datasets were utilized for validation. Detailed baseline information was outlined in Table S1.

Transcriptomic data from the TCGA-OV cohort in the format of Fragments Per Kilobase Million (FPKM) were obtained using TCGAbiolinks [20], while a curated set of clinical traits was acquired from the TCGA-Clinical Data Resource (TCGA-CDR) [21]. Data retrieved from the GEO database for this study originated from either the Affymetrix GPL570 platform or GPL6480 platform. For data generated from GPL570, the raw samples underwent processing utilizing the Robust Multi-array Average (RMA) method within the Affy package [22], and for data from GPL6480, the same RMA algorithm embedded in the limma package [23] was employed for processing the raw samples. Considering the utilization of the GENCODE reference genome (GRCh38, GENCODE Release 36) in TCGA data, we performed a reannotation of probe sets on both GPL570 and GPL6480 arrays by aligning all probes to the same reference genome employing Rsubread [24].

Immune cell infiltration

To assess the relative infiltration of designated immune cell categories within the tumor microenvironment (TME), feature gene sets comprising multiple immune cell subsets [25] were utilized. The embedded single-sample gene set enrichment analysis (ssGSEA) algorithm from the GSVA package [26] was employed to execute this quantification. Two gene lists, namely the activated CD8⁺ T cell [25] and cytotoxic gene list [27], were employed to appraise the activation status of CD8⁺ T cells in HGSOC patients. To substantiate the credibility and stability of the ssGSEA results, six alternative algorithms, including ESTIMATE [28], ConsensusTME [29], MCP-counter [30], EPIC [31], quanTIseq [32], and TIMER [33], were additionally executed. The prognostic value of CD8⁺ T cell activation status was assessed through Kaplan-Meier survival analysis on the four HGSOC datasets.

Consensus clustering

An unsupervised consensus clustering method that incorporated Euclidean distance and Ward’s linkage was applied for cluster identification based on the infiltration profiles of various immune cell categories. This calculation entailed the utilization of the ConsensusClusterPlus package [34], with 1,000 resampling iterations carried out to ensure the robustness of the classification outcomes. The determination of the optimal number of clusters involved a combination of strategies, including the consensus matrix, cumulative distribution function (CDF) curve, proportion of ambiguous clustering (PAC) statistic, and the NbClust package [35].

Weighted gene correlation network analysis (WGCNA)

The dataset used for WGCNA [36] analysis comprised 336 TCGA-OV samples (after removing outliers) with complete traits, focusing on 9798 genes that exhibited a median absolute deviation surpassing 0.5. The computation of an appropriate soft threshold β was performed to achieve a scale-free network. Additionally, the weighted adjacency matrix was converted into a topological overlap matrix (TOM), resulting in the generation of its dissimilarity (1-TOM). The dynamic tree cutting method was performed for module identification. In order to pinpoint gene modules showing notable correlations with immune clusters, activated CD8⁺ T cell score, and cytotoxic signature score, the hub modules were determined as the top three modules with the highest correlation values (P < 0.05). Genes within these modules displaying both gene significance (GS) and module membership (MM) exceeding 0.3 were considered trait-related and chosen for subsequent analysis.

Signature generated from machine learning

Prognostic gene identification in the TCGA-OV cohort was accomplished through Kaplan-Meier survival analysis. The gene expression values of all patients were standardized using the following formula:

In this equation, x_i signifies the expression level of gene i, represents the average expression level of the respective gene across the entire patient cohort, and sd(x) signifies the standard deviation of the gene’s expression level across all patients. The least absolute shrinkage and selection operator (LASSO) regression, embedded in the glmnet package [37], was utilized to develop a CD8⁺ T cell activation status-related gene signature in TCGA-OV. This approach incorporated 10-fold cross-validation for model fitting. The variables corresponding to the penalty parameter λ that produced the lowest mean cross-validated error were then incorporated into an optimal subset regression model using the leaps package. Lastly, the selected genes were incorporated into a multivariate Cox proportional hazards model, where the prognostic risk signature for CD8⁺ T cell activation status was computed as the summation of each gene’s expression value weighted by its respective regression coefficient. The formula used was as follows:

where t denotes the survival time, h₀(t) represents the baseline hazard, n refers to the total number of genes, x_i indicates the expression level of the ith gene, and β_i denotes the corresponding regression coefficient linked to the ith gene. The immune prognostic signature developed for HGSOC has been designated as CIPS (CD8⁺ T cell-associated Immune Prognostic Signature).

Signature evaluation

Late-stage HGSOC patients were classified into low-risk and high-risk groups based on the optimal threshold identified using the survminer package. The dissimilarity in OS between the low-risk and high-risk patients was evaluated through Kaplan-Meier survival analysis. Multivariate Cox analysis was conducted on multiple HGSOC datasets to investigate the independent predictive capacity of CIPS for prognosis. The clinical applicability of this signature was evaluated through various metrics, including Harrell’s concordance index (C-index), receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). Additionally, time-dependent C-index and ROC curves were generated with bootstrap cross-validation repeated 1,000 times to ensure robustness of the outcomes. To investigate the clinical relevance of the signature, the SubMap module of GenePattern [38] and the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm [39] were implemented to forecast the ICB responses of each HGSOC patient stratified into low-risk and high-risk groups. The pRRophetic package [40], which includes a linear ridge regression model, was implemented to forecast drug response using gene expression data. Moreover, the correlation of this signature with tumor mutation burden (TMB), BRCA1/2 mutations, and homologous recombination deficiency (HRD) in late-stage HGSOC was also evaluated. Mutation data were analyzed and summarized utilizing the Maftools package [41], while HRD data were acquired from the TCGA PanCan Atlas [42].

To rigorously evaluate the predictive performance of CIPS in comparison with other existing models, we systematically identified and retrieved previously published OC prognostic signatures through a comprehensive literature review. We initially compiled all currently available ovarian cancer models and subsequently retained only those with less than 20% missing genes in the ovarian cancer expression profiles considered in this study for comparative analysis. Considering the limited clinical utility of models containing an excessive number of genes, we exclusively selected published models comprising fewer than 20 genes for comparative evaluation. Ultimately, 56 published ovarian cancer prognostic models were included in our analysis and compared with our developed model across four large-scale HGSOC cohorts and a meta-cohort comprising all test sets. For models with published regression coefficients, we directly applied these coefficients to calculate risk scores. For models lacking regression coefficients or those with missing genes (< 20%), we reconstructed multivariate Cox regression models to derive risk scores. The comparative evaluation of model performance was conducted using the C-index as the primary metric. Following data retrieval of the Pan-cancer TCGA dataset from the UCSC Cancer Genome Browser, our study focused on investigating the relationship between gene expression within the signature and the relative abundance of CD8⁺ T cells across 32 different cancer types.

Single-cell RNA-sequencing (scRNA-seq) analysis

The scRNA-seq data from 10 primary tumor samples of patients diagnosed with HGSOC were collected [43], and subsequent data analysis was conducted using the Seurat package [44] within the R software environment. During the quality control procedures, genes with detection in fewer than 10 cells were initially excluded. Cells meeting the following criteria were retained for downstream analyses to ensure inclusion of high-quality cells: > 200 expressed genes, > 500 unique molecular identifiers (UMIs), < 10% alignment to mitochondrial genes, and < 1% alignment to hemoglobin genes. Normalization of the UMI count data was accomplished through regularized negative binomial regression [45], taking into account the proportions of mitochondrial genes and cell cycle influences. The fast mutual nearest neighbors (MNN) correction method [46] was applied to achieve the integration of individual Seurat objects, with a focus on batch correction by choosing the top 3,000 genes which displayed the highest variability for each sample. Dimensional reduction and clustering were carried out using the Uniform Manifold Approximation and Projection (UMAP) approach, with the 30 most informative principal components (PCs) taken into account. The Louvain algorithm with a resolution parameter set to 0.8 was employed to identify cell clusters, implemented through the “FindClusters” function provided by the Seurat package. Scoring of gene signatures was performed in the HGSOC single-cell dataset using the UCell package [47]. CellChat [48] was employed to quantitatively infer and analyze intercellular communication networks using scRNA-seq data.

Statistical analysis

All data processing, statistical analysis, and visualization were conducted by employing R (https://www.r-project.org/). The normality of variables was assessed utilizing the Shapiro-Wilk test prior to comparison. The Wilcoxon rank-sum test was utilized for comparing non-normally distributed variables between two groups. The Kruskal-Wallis test as a non-parametric method was employed for the comparison among more than two groups. Categorical variables were compared using the Fisher test. The assessment of correlations between two continuous variables was performed using Spearman correlation coefficients (ρ). The determination of the optimal cut-off value for survival was achieved by employing the maximally selected rank statistics method provided by the survminer package. To prevent issues arising from too few patients in certain groups, a minimum proportion of observations per group was set at 30%. Cox proportional hazards regression and Kaplan-Meier analyses were conducted using the survival package. Survival differences were evaluated through the Kaplan-Meier curve and log-rank test. The timeROC package was utilized to assess the area under curve (AUC) of receiver operating characteristic (ROC) curve for survival variables. Statistical significance was determined with a two-tailed P value threshold of < 0.05.

Results

Construction of immune infiltration consensus clusters

Following the estimation of cell infiltration for 28 distinct immune cell populations, a consensus cluster analysis was conducted. Initially, all HGSOC samples were classified into k clusters (k = 2–9). The optimal number of clusters was determined by examining the CDF curves of the consensus matrix and the PAC statistic, which highlighted that k = 3 was the preferred choice (Fig. 1A-B). This result was further supported by Nbclust testing (Fig. 1C). The principal component analysis (PCA) revealed the differences among the three clusters of HGSOC patients (Fig. 1D). The three consensus clusters (C1, C2, and C3) exhibited notable variations in immune infiltration (Fig. 1E). Notably, C3 displayed a substantially higher overall infiltration abundance compared to C1 and C2 (Fig. 1F). Therefore, we classified C1 as the low infiltration group (n = 69), C2 as the median infiltration group (n = 196), and C3 as the high infiltration group (n = 99). Patients with a high level of infiltration status exhibited the highest scores for both cytotoxic signature and immune function, suggesting that these patients efficiently activated cytotoxic responses and overall immune function (Fig. 1G, H). Corresponding to TCGA’s four-category system of serous ovarian cancer based on gene expression (differentiated, immunoreactive, mesenchymal, and proliferative), a substantial proportion of patients in the immunoreactive subtype were predominantly represented by the C2 (medium infiltration) and C3 subgroups (high infiltration) in this study’s TME-based classification of serous ovarian cancer patients (Fig. 1I).

Fig. 1.

Identification of robust immune phenotyping in late-stage HGSOC patients. (A) Consensus score matrix representing all samples at k = 3. (B) PAC scores. (C) Determination of the optimum cluster count guided by 26 criteria using the Nbclust package. (D) PCA for dimensionality reduction of all samples, illustrating the three clusters identified by consensus clustering. (E) Heatmap of the immune infiltration landscape generated using ssGSEA, ESTIMATE, and five additional algorithms. (F) Significant differences in the relative abundance of 28 distinct immune cell categories across the three immune clusters. (G) Comparison of cytotoxic signature scores across the three clusters. (H) Comparison of immune scores across the three clusters. (I) Distribution of the four molecular subtypes within each of the three clusters. ***P < 0.001. ns, non-significant

The relationship between CD8⁺ T cells and prognosis

For the late-stage HGSOC patients in the TCGA-OV dataset, the activated CD8⁺ T cell score showed no significant correlation with OS in both the entire patient cohort and the C1 subgroup (P > 0.05). However, in the patients belonging to the C2 + C3 subgroups, the activated CD8⁺ T cell score and cytotoxic score exhibited a significant positive relation with OS (P < 0.05) (Fig. S1A). The same results were observed for disease-specific survival (DSS) and progression-free interval (PFI) among the advanced HGSOC patients in the TCGA-OV cohort (Fig. S1B-C). In the patients with HGSOC from the ICGC database, both the activated CD8⁺ T cell score and cytotoxic score showed a remarkable positive relation with OS and recurrence-free survival (RFS) (P < 0.05) (Fig. S1D). In the patients from the GSE32062 and GSE9891 datasets, the activated CD8⁺ T cell score exhibited a noticeable positive correlation with OS (P < 0.05) (Fig. S1E-F). In the patients from the GSE17260 dataset, the activated CD8⁺ T cell score demonstrated a marked positive correlation with both OS and RFS (P < 0.05) (Fig. S1G).

Identification of CD8⁺ T cell-associated genes

The WGCNA approach was implemented to detect cohesive gene modules linked to CD8⁺ T cell activation, specifically observed in late-stage HGSOC patients. In the WGCNA analysis, a soft threshold (β) of 4 was selected, resulting in a scale-free network (no scale R² = 0.910), which ensured an appropriate power value for constructing the co-expression network (Fig. S2A). Subsequently, a total of 37 modules were recognized and visualized by different colors (Fig. S2B). To represent each module, the eigengene was determined as the principal component reflecting the predominant variance in gene expression within that particular module. The adjacency between eigengenes within modules was depicted in a heatmap (Fig. S2C). Moreover, the associations between modules and various clinical features encompassing age, histological grade, FIGO stage, mRNA subtype, immune cluster, activated CD8⁺ T cell score, cytotoxic signature score, immune score, and tumor purity were assessed. The turquoise, tan, and pink modules were observed to cluster together within a specific branch in the clustering analysis (Fig. S2C), and these three modules displayed the top three coefficients (> 0.4) in relation to the immune cluster (Fig. 2A). In terms of the activation status of CD8⁺ T cells, the turquoise, tan, and pink modules exhibited the strongest correlations in the module-trait relationships, exhibiting correlation coefficients of 0.73, 0.64, and 0.51, respectively (Fig. 2A). Meanwhile, concerning the correlation coefficient between GS and MM, all the three modules exhibited coefficients greater than 0.7 in relation to cytotoxic score, activated CD8⁺ T cell score, and immune score (Fig. 2B-D, Fig. S3). To identify hub genes derived from CD8⁺ T cell activation patterns within the three modules, 663 genes meeting the threshold of GS > 0.3 and MM > 0.3 were deemed to be hub CD8⁺ T cell activation-related genes. Subsequently, Kaplan-Meier analysis was performed on the expression patterns of the 663 genes, resulting in the recognition of 214 prognostic genes (Fig. 2E). Among these, 148 genes were detected in multiple datasets including TCGA-OV, GSE32062, GSE9891, and GSE17260. Therefore, these 148 genes were chosen for further investigation using a machine learning-based integrative approach in order to establish a signature associated with CD8⁺ T cell activation specifically in late-stage HGSOC. Based on the over representation analysis (ORA) using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG), these genes were found to be significantly enriched in immune-relevant pathways, encompassing T cell receptor signaling pathway, T cell differentiation, T cell proliferation, and T cell mediated immunity (Fig. 2F, G).

Fig. 2.

Screening of CD8⁺ T cell activation-associated genes in late-stage HGSOC patients via WGCNA. (A) Investigation of the correlation between module eigengenes and clinical characteristics. (B) The significant relation between GS for cluster and MM in the turquoise module. (C) The significant relation between GS for activated CD8⁺ T cell and MM in the turquoise module. (D) The significant relation between GS for cytotoxic score and MM in the turquoise module. (E) Overlapping genes identified through integration of WGCNA and Kaplan-Meier analysis. (F) ORA based on GO biological processes. (G) ORA based on KEGG pathways. The biological processes or pathways associated with T cells were presented, with adjusted P < 0.05

Construction of a prognostic signature

To further narrow down candidate genes, we employed a LASSO-penalized Cox proportional hazards model with 10-fold cross-validation to determine the optimal penalty parameter (λ). The model identified 23 prognostic genes at the minimum criteria (logₑλ = −3.52, λ = 0.0537; Fig. S4). Subsequent optimal subset regression further refined the gene set to 10 key prognostic markers: GBP2, CCL18, CD69, SLC28A3, ISG20, CTSC, PLEK2, NFKB1, IFNGR1, and VSIG4 (Fig. S5). These genes were incorporated into a multivariate Cox regression model, ultimately constructing a prognostic signature significantly associated with CD8⁺ T cell activation in late-stage HGSOC.

Each patient’s risk score was determined by calculating a weighted sum of the expression values of 10 genes, based on their respective regression coefficients in a multivariate Cox proportional hazards model (Fig. 3A). The optimal cutoff value for survival, ascertained by the survminer package, was employed to dichotomize all patients into high- and low-risk subgroups (Fig. 3B). As illustrated in Fig. 3C, the Kaplan-Meier survival analysis performed on the training cohort TCGA demonstrated a notably higher mortality rate for high-risk patients relative to low-risk patients (n = 364, P < 0.001). Similar results were observed in the validation cohorts GSE32062 (n = 260, P < 0.001), GSE9891 (n = 140, P < 0.001), and GSE17260 (n = 110, P = 0.0014). The risk score demonstrated independent prognostic significance across the four HGSOC datasets, regardless of age, histological grade, clinical stage, and surgery status (Fig. 3D).

Fig. 3.

Development of the prognostic signature based on CD8⁺ T cell-associated genes in late-stage HGSOC. (A) The regression coefficients of the 10 genes in the multivariate Cox regression analysis. (B) Distribution of risk scores, survival time, and expression values of prognostic genes among patients in the TCGA-OV, GSE32062, GSE9891, and GSE17260 datasets. (C) Kaplan-Meier curves depicting OS based on the prognostic signature in the TCGA-OV, GSE32062, GSE9891, and GSE17260 datasets. (D) Multivariate Cox proportional hazards regression analysis carried out on the TCGA-OV, GSE32062, GSE9891, and GSE17260 datasets

Assessment of the prognostic signature

The AUC of the ROC curve for the 1–5 year OS is not only consistently greater than 0.6 in the training set (Fig. S6A) but also shows similar results across the three test sets (Fig. S6C, E, G), demonstrating the signature’s good efficacy to predict the prognosis of late-stage HGSOC. The calibration curves exhibited a satisfactory concordance between the expected and actual survival rates (Fig. S7). The DCA was used to evaluate the efficacy of the signature, and the standardized net benefit substantiated the predictive value of the signature in OS (Fig. S8). In addition, the prognostic signature exhibited superior predictive ability in contrast to other clinical characteristics. When integrated with these clinical features, the model demonstrated enhanced predictive power surpassing that of any individual clinical factor used independently (Fig. S6B, D, F, H).

Comparative analysis of gene expression signatures

The rapid improvements in large-scale sequencing technologies and computational bioinformatics have catalyzed the development of diverse predictive and prognostic signatures across multiple tumor types, including ovarian cancer. We systematically compiled and analyzed 56 ovarian cancer prognostic signatures associated with a wide spectrum of biological features, including TME, cuproptosis, necroptosis, PANoptosis, chromatin remodeling, and neutrophil extracellular traps, among others (Table S2). To rigorously evaluate the comparative performance of CIPS against these published signatures, we calculated the C-index for each model. Remarkably, CIPS demonstrated robust prognostic performance across multiple independent datasets, including TCGA (Fig. S9A), GSE32062 (Fig. S9B), GSE9891 (Fig. S9C), GSE17260 (Fig. S9D), and the meta-cohort (Fig. S9E). While one model exhibited a marginally higher C-index than CIPS in the training cohort (TCGA), CIPS consistently outperformed this model across multiple independent test sets. Similarly, although another model showed a slightly superior C-index in the GSE17260 dataset, CIPS demonstrated significantly better performance in the remaining three datasets and the meta-cohort. These findings collectively underscore the superior generalizability and robust predictive capacity of the CIPS model across diverse patient populations and datasets.

Clinical relevance of CIPS

A significant positive relationship was found between the expression of 10 genes and the relative abundance of most immune cells, along with cytotoxic effector gene signature scores and immune scores. Meanwhile, the expression of 10 genes demonstrated a significant negative relation with tumor purity scores. The risk score presented a pronounced negative relation with activated CD8⁺ T cell and cytotoxic scores, while showing no significant correlation with the relative proportions of most other cell categories (Fig. 4A). Patients classified as low-risk exhibited significant elevation in activated CD8⁺ T cell scores and cytotoxic signature scores when contrasted with those categorized as high-risk (Fig. 4B). Additionally, the expression of most T lymphocyte and cytotoxic effector-related marker genes was significantly higher in the low-risk patients in comparison to the high-risk patients (Fig. 4C, Fig. S10). Meanwhile, low-risk patients also exhibited significantly elevated expression of most dysfunctional T cell marker genes (e.g., PDCD1, CTLA4, LAG3, TIGIT) compared to high-risk patients (Fig. 4C, Fig. S10). In the late-stage HGSOC patients from the TCGA cohort, a noticeable discrepancy in chemotherapy response was observed between the high- and low-risk patients, with a higher proportion of low-risk patients achieving complete remission (CR) (Fig. 4D). Through a comparison of the expression profile similarity between late-stage HGSOC patients and another published dataset comprising melanoma patients who responded to immunotherapies [49], the SubMap analysis revealed that the low-risk group exhibited potentially increased sensitivity to PD-1 treatment (adjusted P = 0.008, Fig. 4E). In the late-stage HGSOC patients from the GSE32062, GSE9891, and GSE17260 cohorts, the low-risk patients displayed lower TIDE scores, suggesting a higher likelihood of benefiting from immunotherapy in comparison to the high-risk patients (Fig. 4F). The drug sensitivity prediction analysis revealed that the low-risk late-stage HGSOC patients exhibited a higher sensitivity to small molecule drugs, including commonly used chemotherapeutic agents such as paclitaxel and cisplatin in the treatment of ovarian cancer (Fig. 4G).

Fig. 4.

Clinical relevance of CIPS in late-stage HGSOC. (A) Associations between the expression levels of 10 prognostic genes and the relative abundance of different immune cell subsets. Risk score, PDCD1, CD274, CTLA4, cytotoxic score, immune score, and tumor purity were also included. (B) Differential enrichment scores of activated CD8⁺ T cell and cytotoxic effector gene signatures between patients classified as low-risk and high-risk. (C) Differential expression levels of canonical marker genes for T lymphocytes, cytotoxic T lymphocytes, and exhausted T lymphocytes between the low-risk and high-risk patients using the GSE32062 dataset. **0.001 < P < 0.01, ***P < 0.001. (D) Differential response to chemotherapy between the high- and low-risk patients in the TCGA-OV dataset. CR: complete remission/response; PR: partial remission/response; PD: progressive disease; SD: stable disease. (E) Comparison of the responsiveness to CTLA-4 inhibitors and PD-1 inhibitors between the low-risk and high-risk patients in the TCGA-OV dataset. R: response; NR: no response. (F) Differential TIDE scores between patients classified as low-risk and high-risk in the GSE32062, GSE9891, and GSE17260 datasets. (G) Distribution of predicted IC₅₀ response values between patients classified as low-risk and high-risk in the TCGA-OV dataset. The drugs with P < 0.05 were shown

We explored the association of this signature with somatic mutations and HRD in the TCGA late-stage HGSOC. The somatic mutation analysis revealed a high frequency of mutations in TP53 (92%), TTN (25%), and USH2A (11%) within the low-risk subgroup (Fig. S11A), and in TP53 (93%), TTN (31%), and CSMD3 (13%) within the high-risk subgroup (Fig. S11B). The patients with high TMB (Fig. S11C), BRCA1/2 mutations (Fig. S11D), and high HRD scores (Fig. S11E) exhibited lower risk scores, suggesting longer survival and greater likelihood of benefiting from treatment, which was consistent with previous findings in this study.

The correlations between these 10 genes and survival across multiple cancer types exhibited inconsistent patterns, indicating their prognostic value and emphasizing the need for further in-depth investigations (Fig. S12A). The expression levels of GBP2, CCL18, CD69, SLC28A3, ISG20, CTSC, and VSIG4 genes demonstrated a noticeable positive association with the relative quantity of CD8⁺ T cells in at least 15 cancer types (median ρ > 0.4 and P < 0.05, Fig. S12B). Conversely, PLEK2, NFKB1, and IFNGR1 genes did not demonstrate a noticeable positive relation with the relative abundance of CD8⁺ T cells in more than half of the analyzed cancer types (median ρ < 0.1 and P > 0.05, Fig. S12B).

Expression distribution of prognostic genes

With the aim of examining the expression distribution of prognostic genes at the single-cell level, the study collected the scRNA-seq data from 10 primary tumor samples of HGSOC patients and performed subsequent analyses. After conducting quality control measures, we retained a total of 24,105 genes and 43,567 cells, with a median of 2,854 genes and 10,832 UMIs detected per cell. The integration of multiple samples was presented in Fig. 5A. By assessing the expression of canonical cell type gene markers, we annotated various cell types in the TME of HGSOC (Fig. 5B-C, Fig. S13A), revealing five major cell clusters consisting of epithelial cells (23,331 cells), macrophages (6,380 cells), fibroblasts (3,288 cells), CD8⁺ T cells (2,724 cells), and proliferative cells (2,163 cells). Meanwhile, seven minor cell clusters were identified: CD4⁺ T cells, other stromal cells, NK cells, endothelial cells, mesothelial cells, B cells and dendritic cells. In this study, the expression distribution of the 10 prognostic genes was observed in primary tumor tissues from HGSOC. GBP2, CTSC, ISG20, and CD69 exhibited significantly higher expression on T lymphocytes; IFNGR1, VSIG4, NFKB1, GBP2, CTSC, and CCL18 showed significantly higher expression on monocytes and macrophages; SLC28A3 demonstrated significant expression in epithelial cells and fibroblasts (Fig. 5D, Fig. S13B).

Fig. 5.

Evaluation of CIPS in single-cell transcriptomic data from primary HGSOC tissues. (A) UMAP plot color-coded by 10 HGSOC samples. (B) UMAP plot color-coded by 12 cell types. (C) Comparison of cell types in primary tumor tissues between platinum-resistant and platinum-sensitive patients. (D) Density distribution of the 10 prognostic genes. (E) Density distribution of CIPS scored by UCell. (F) Comparison of gene signature scores between platinum-resistant and platinum-sensitive patients. (G) Comparison of differences in the number and strength of cell-cell interactions between low- and high-risk patient groups. (H) Comparison of differential interaction counts and strengths among annotated cell types between low- and high-risk patient groups. Red denotes increased interactions/strength; blue denotes decreased interactions/strength. *P < 0.05, **P < 0.01, ***P < 0.001. ns, non-significant

Based on this CD8⁺ T cell-associated signature, single-cell transcriptomic data from primary HGSOC tumor tissues were scored, and the signature was found to be predominantly enriched in myeloid cells and T lymphocytes (Fig. 5E). Compared to platinum-resistant patients, platinum-sensitive patients exhibited significantly decreased enrichment scores for CD4⁺ T cells, CD8⁺ T cells, NK cells, and macrophages, which was consistent with the findings of this study indicating that patients in the low-risk group of HGSOC were more likely to benefit from treatment (Fig. 5F). Cellular communication analysis demonstrated an overall increase in the intensity of intercellular communication within the TME of platinum-resistant patients compared to platinum-sensitive group (Fig. S14). Using the median value of enrichment scores for the CD8⁺ T cell-associated signature, the cells within the TME were stratified into low-risk and high-risk groups. We found that the high-risk group exhibited an elevated overall intensity of intercellular communication compared to the low-risk group (Fig. 5G-H), indicating a potential link between high enrichment scores and treatment resistance.

In the TCGA late-stage HGSOC cohort, the 10 model genes demonstrated significant positive associations with two immune‑related biological processes: (1) T cell activation involved in immune response (GO:0002286; Fig. 6A) and (2) positive regulation of T cell activation (GO:0050870; Fig. 6B). In scRNA‑seq data from primary HGSOC tumors, four of the ten genes (CD69, ISG20, CTSC, and GBP2) were highly expressed in CD8⁺ T cells and exhibited significant positive correlations with the cytotoxic effector markers GZMB (Fig. 6C) and PRF1 (Fig. 6D). In contrast, the remaining six genes demonstrated no significant associations. Notably, all genes except SLC28A3 were significantly positively correlated with the two T‑cell‑activation processes above (GO:0002286 and GO:0050870; Fig. 6E-F).

Fig. 6.

Association of CIPS with T cell activation-related genes or gene sets in transcriptomic data from primary HGSOC tissues. (A) Correlation between model genes and enrichment scores for T cell activation involved in immune response (TCGA bulk RNA-seq data). (B) Correlation between model genes and enrichment scores for positive regulation of T cell activation (TCGA bulk RNA-seq data). (C) Correlation between model genes and GZMB (scRNA-seq data). (D) Correlation between model genes and PRF1 (scRNA-seq data). (E) Correlation between model genes and enrichment scores for T cell activation involved in immune response (scRNA-seq data). (F) Correlation between model genes and enrichment scores for positive regulation of T cell activation (scRNA-seq data)

Discussion

While some ovarian cancer patients receiving ICB therapy have shown long-lasting clinical responses, only a limited proportion of patients derive clinical benefit. Recent years have witnessed substantial progress in utilizing ICB therapy for ovarian cancer [50–52]; however, inconsistencies or contradictions persist in the findings concerning the significance of PD-L1 expression as a predictive marker for both ICB response and clinical outcomes in this malignancy [53]. Therefore, relying exclusively on PD-L1 status might be inadequate for determining the appropriateness of ICB therapy for ovarian cancer patients [54]. Hence, there is a pressing demand for the development of predictive biomarkers that can enhance patient benefits. Fortunately, substantial evidence has compellingly demonstrated that the presence of tumor-infiltrating CD3⁺ and CD8⁺ T cells is strongly correlated with a favorable prognosis among patients with ovarian cancer, including HGSOC [10, 55, 56]. Consequently, it becomes imperative to establish dependable tumor markers that are intricately linked with CD8⁺ T cell activation, facilitating the investigation of precise prognostic strategies for the management of HGSOC.

In our study, we performed a bioinformatics-based retrospective analysis on 874 late-stage HGSOC patients from four independent cohorts. Multiple algorithms, primarily encompassing ssGSEA, consensus clustering, and WGCNA, were employed to discern immune heterogeneity phenotypes and CD8⁺ T cell activation-associated genes in late-stage HGSOC. A 10-gene prognostic signature, linked to CD8⁺ T cell activation in late-stage HGSOC, was constructed using three regression algorithms. The identified signature demonstrated an association with an unfavorable prognosis in late-stage HGSOC patients and functioned as an independent prognostic factor for OS. Further assessment using time-dependent C-index and ROC analysis consistently revealed that CIPS exhibited sustained high accuracy and stable performance across four publicly available large cohorts. These findings underscored the significant potential of this signature for future clinical applications in late-stage HGSOC. In addition, compared with the 56 published signatures encompassing diverse functional gene combinations that we retrieved, CIPS consistently outperformed nearly all models in every dataset, indicating its superior predictive ability. It was observed that the majority of models displayed favorable performance within their respective training datasets and limited external datasets, but exhibited weaker performance in other datasets, implying a potential lack of generalizability due to overfitting. Our signature underwent dimensionality reduction using multiple machine learning algorithms, thereby enhancing its potential for extrapolation.

The 10-gene signature developed in this study has shown significant clinical relevance. Low-risk patients exhibited elevated activated CD8⁺ T cell scores and cytotoxic signature scores compared to high-risk patients, along with higher expression of T lymphocyte and cytotoxic effector-related marker genes. Low-risk patients exhibited co-elevated cytotoxic and inhibitory markers, reflecting chronic antigen exposure driving CD8 + T cells toward an “activated but exhausted” phenotype. This paradoxical co-expression signifies a tumor microenvironment enriched with antigen-experienced cytotoxic T cells concomitantly upregulating inhibitory receptors. Notably, low-risk patients with late-stage HGSOC demonstrated better chemotherapy response, characterized by a higher rate of complete remission. Additionally, low-risk patients were predicted to possess heightened sensitivity to small molecule drugs, including commonly used chemotherapy agents for ovarian cancer treatment. Furthermore, the low-risk group showed potentially increased sensitivity to immunotherapy. Analysis of somatic mutations and HRD in late-stage HGSOC revealed that patients with high TMB, BRCA1/2 mutations, and high HRD scores exhibited lower risk scores, indicating better survival and response to treatment. Moreover, the 10 model genes demonstrated showed significant positive associations with the “T cell activation involved in immune response” and “positive regulation of T cell activation” pathways. These findings support that CIPS not only serves as a statistical prognostic model but indeed reflects an immune activation program closely related to CD8⁺ cytotoxic function.

Single-cell resolution analysis provided cellular-level evidence for these population-based correlations: Analysis of scRNA-seq data from primary HGSOC tumor tissues revealed that the CD8⁺ T cell-associated signature was predominantly enriched in myeloid cells and T lymphocytes. Immune cells from platinum-sensitive patients exhibited significantly lower enrichment scores compared to platinum-resistant patients, consistent with the better treatment response observed in the low-risk group. CD69, ISG20, CTSC, and GBP2 showed high expression in CD8⁺ T cells and demonstrated significant positive correlations with GZMB/PRF1 in single-cell data, suggesting these four genes may represent or participate in the activation or effector functions of intratumoral CD8⁺ T cells. In contrast, the remaining six genes were primarily localized to myeloid or stromal cell populations. Although not directly enriched in CD8⁺ T cells, all except SLC28A3 still maintained positive correlations with the aforementioned T cell activation-related biological processes. This suggests that CIPS simultaneously integrates both intrinsic T cell activation signals and microenvironmental regulatory information from myeloid and stromal compartments. In other words, the signature reflects not only the presence and functional state of effector T cells but also the intercellular interactions or immunoregulatory context that influences their function. This dual characteristic may explain the model’s superior performance in predicting responses to both chemotherapy and immunotherapy. One additional point warrants attention: SLC28A3 shows only a weak association with T cell activation at the single‑cell level and is predominantly expressed in epithelial and fibroblast compartments, suggesting that its contribution to the signature may reflect indirect effects of tumor cells or stroma on the immune milieu. This hypothesis likewise merits experimental validation.

Among these 10 genes, three genes (GBP2, ISG20, and VSIG4) were already found to be included in established prognostic prediction models for ovarian cancer. GBP2 was reported to play a pivotal role in promoting protective immunity against microorganisms [57], and it was positively correlated with improved prognosis, antigen processing, CD8⁺ T cell infiltration, and the efficacy of ICB in cancers [58, 59]. Furthermore, previous studies have developed network-pathway based and TME-related signatures for predicting outcomes in ovarian cancer patients, which incorporated the GBP2 gene [60, 61]. As for ISG20, its involvement in augmenting the activation of anti-tumor immunity through a double-stranded RNA-triggered interferon response has been well-documented in ovarian cancer [62]. Additionally, ISG20 has been identified within two gene signatures associated with glycolysis [63, 64] and one gene signature associated with necroptosis [65], which possess the potential to serve as prognostic indicators for patients diagnosed with ovarian cancer. VSIG4 has been identified as an inhibitor of T cell activation, with its expression predominantly limited to quiescent tissue macrophages [66]. Previous studies have established prognostic gene signatures for ovarian cancer by incorporating invasion-, tumor associated macrophage-, or TME-related genes, with the inclusion of VSIG4 in these signatures [61, 67, 68]. Among the remaining seven genes, CCL18, CD69, NFKB1, and IFNGR1 have been extensively documented in the literature for their significant involvement in ovarian cancer. The cytokine encoded by CCL18 attracts naive T cells, CD4⁺ T cells, CD8⁺ T cells, and nonactivated lymphocytes through chemotaxis, while also facilitating the migration and epithelial-mesenchymal transition of diverse tumor cells, including ovarian cancer [69, 70]. Additionally, CCL18 is incorporated in a prognostic signature for glioblastoma that is linked to M2 macrophages and impaired T cell function [71]. CD69 is a well-recognized activation marker of T lymphocytes, and in ovarian cancer research, CD69⁺ is also considered a marker for activated cytotoxic T cells [72, 73]. NFKB1 has been extensively implicated in carcinogenesis, displaying a dualistic role characterized by promoting cancer progression in certain cases and acting as a tumor suppressor in others [74]. IFNGR1, as a receptor subunit for IFN-γ, assumed pivotal functions in antimicrobial, antiviral, and antitumor responses by facilitating the activation of effector immune cells and augmenting antigen presentation mechanisms [75]. The relationship of SLC28A3, CTSC, and PLEK2 in human ovarian cancer remains incompletely elucidated. For SLC28A3, the association between multiple genetic variants of this gene and cardiotoxicity has been extensively reported [76–78]. CTSC plays a crucial role in the regulation of inflammation and immune responses linked to polymorphonuclear neutrophils [79], while simultaneously impacting cancer progression and acting as a potential predictive marker for brain- and lung-specific metastasis [80]. PLEK2 assumes critical functions not only in cell spreading but also in diverse cellular processes like erythropoiesis, tumorigenesis, and metastasis, thereby establishing its potential as a candidate for diagnostic and prognostic biomarkers and an attractive prospect for cancer therapy [81–84]. Nonetheless, the precise involvement of the three genes in human ovarian cancer necessitates further exploration and investigation. To sum up, among the 10 candidate prognostic genes identified for HGSOC, some genes have already shown significant prognostic value, while the prognostic potential of other genes remains to be fully elucidated. The consistency of our findings with existing research highlights the reliability of our integrative machine learning analysis approach. Hence, given the pivotal contributions of these 10 genes in the prognostic signature, it emphasizes that in-depth investigation is warranted to further explore their prognostic significance in ovarian cancer.

It should be noted that our results are primarily based on transcriptomic and bioinformatics computational analyses and therefore do not establish causality. Accordingly, future work should include functional validation: for example, use flow cytometry and immunohistochemistry in larger patient cohorts to validate co‑localization of candidate genes with activation and cytotoxic phenotypes; perform in vitro and in vivo perturbation experiments to evaluate the effects of CTSC, GBP2, ISG20, CD69 and other genes on CD8⁺ T cell cytotoxicity; and apply spatial transcriptomics or multiplexed in‑situ staining to further resolve the functional significance of these genes’ proximity to immune cells within the tumor microenvironment. Collectively, CIPS is not a unidimensional immune cell counting metric but rather a multi‑layered readout of immune programs that integrates intrinsic T cell activation signals with microenvironmental regulatory cues. With further mechanistic studies and validation in larger prospective clinical cohorts, CIPS has the potential not only to serve as a prognostic and treatment‑response biomarker but also to highlight actionable immune or microenvironmental targets, thereby providing a stronger biological rationale for individualized immunotherapy strategies in late-stage HGSOC.

In summary, we conducted a bioinformatics-based retrospective analysis on 874 late-stage HGSOC patients, identifying immune heterogeneity phenotypes and CD8⁺ T cell activation-associated genes using multiple algorithms. A 10-gene prognostic signature, linked to CD8⁺ T cell activation, demonstrated consistent high accuracy and stable performance for prognostic prediction across large cohorts, highlighting its potential for clinical applications in late-stage HGSOC. Our study incorporated multiple relatively large and high-quality HGSOC cohorts, utilizing diverse algorithms for the screening of CD8⁺ T cell activation-related genes and the construction of prognostic signatures, thereby guaranteeing the reliability and generalizability of the results. While this study did not incorporate broader omics data and subsequent validation through larger-scale multi-center retrospective studies is warranted, the current findings offer promising evidence that could guide future prospective clinical trials.

Conclusions

In conclusion, this study focuses on addressing the clinical challenge of low response rates to immunotherapy in ovarian cancer and highlights the critical necessity for developing novel prognostic and therapeutic biomarkers that are relevant to CD8⁺ T cell activation. Specifically, we investigated the TME of late-stage HGSOC using transcriptomic data and multiple machine learning algorithms, which led to the identification of crucial prognostic genes linked to CD8⁺ T cells. The resulting CD8⁺ T cell-related signature comprised 10 genes, enabling accurate identification of high-risk mortality patients in late-stage HGSOC. The prognostic signature demonstrated sustained high accuracy and stable performance across distinct late-stage HGSOC cohorts, outperforming the majority of previously reported signatures. Therefore, this study could offer valuable evidence and innovative insights for the establishment of prognostic biomarkers and potential therapeutic targets to enhance clinical outcomes of late-stage HGSOC.

Author contributions

X.Liu, X.Li, and J.W. designed the study. X.Liu, Z.Z., and B.W. performed the bioinformatics analysis. X.Liu prepared the figures and wrote the manuscript. X.Li and J.W. supervised the research and revised the manuscript. L.Q., N.F.,R.Wang, X.Y., Q.H., Z.L., S. K., F.S., C.C., Y.Z., L.N., and R.Wei contributed to the performance of the *in silico* experiments. All authors reviewed the manuscript and approved the final version.

Funding

This study was supported by National Natural Science Foundation of China (82274575, 82405614, 82474564, 82471704), China Postdoctoral Science Foundation (2024M761879), Major Basic Research Project of Shandong Natural Science Foundation (ZR2023ZD56), “Youth Innovation Program” endorsed by the Department of Education of Shandong Province (2024KJJ057), Natural Science Foundation of Shandong Province (ZR2025MS1212, ZR2022LZY011, ZR2023QC167, ZR2023QH452), Co-construction project of State Administration of TCM (GZY-KJS-SD-2023-034, GZY-KJS-SD-2023-046), Taishan Scholars Program (tstp20240513), Central Government Guides Local Science and Technology Development Fund Projects of Shandong Province (YDZX20203700001407). Key Laboratory of Traditional Chinese Medicine Classical Theory, Ministry of Education, Open Research Project. National Youth Qihuang Scholar Training Program and Shandong Province Traditional Chinese Medicine High Level Talent Cultivation Project.

Data availability

Publicly available datasets analyzed in this study can be accessed as follows: TCGA ovarian cancer (TCGA-OV) data were obtained from the Genomic Data Commons (GDC) portal: [https://portal.gdc.cancer.gov/](https:/portal.gdc.cancer.gov). Gene expression profiles for GSE32062, GSE9891, and GSE17260 are available from the Gene Expression Omnibus (GEO) repository under accession numbers GSE32062, GSE9891, and GSE17260 ([https://www.ncbi.nlm.nih.gov/geo/](https:/www.ncbi.nlm.nih.gov/geo)). Single-cell RNA-seq data for HGSOC were downloaded from Mendeley Data ([https://doi.org/10.17632/rc47y6m9mp.1](https:/doi.org/10.17632/rc47y6m9mp.1)). No new sequencing data were generated in this study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.