The single-cell sequencing data of 5 OV patients (GSE154600) was retrieved from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) database. The RNA-seq data and accompanying clinical information of 378 OV patients were downloaded from The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/). In addition, the GSE140082 dataset including the RNA-seq and clinical data of 379 OV patients was also downloaded from the GEO database as the validation cohort. All available datasets used in this study were based on published reports that had received ethical approval. Dataset inclusion criteria: (1) Pathological diagnosis of ovarian cancer; (2) Complete transcriptome and clinical prognostic data; Exclusion criteria: (1) The pathological diagnosis was unclear; (2) Survival time less than 1 day; (3) The transcriptome information did not contain T cell exhaustion related genes.

The scRNA-seq data was analyzed using the “Seurat”, “SingleR” software packages. The cells with unique feature counts >5000 or <200, or with mitochondrial counts >5% were filtered out. The data of each retained cell was normalized using the default parameters of Seurat’s ‘NormalizeData’ function on feature-expression measurements, and then transferred to a combined Seurat object via the Harmony package. The variable genes were scaled and subjected to principal component analysis (PCA). The significant principal components (PCs) were selected by the “RunUMAP” function (min. dist = 0.2, n. neighbors = 20) and the “FindClusters” function (resolution = 0.5) for umap analysis and cluster analysis.

The individual cell types were annotated by two modalities. The T-cell subsets were initially clustered using SingleR, an automated annotation method for scRNA-seq data [26], and the identity of each cluster was further determined by manually searching the cell markers (http://biocc.hrbmu.edu.cn/CellMarker/). The differentially expressed genes (DEGs) in the CD8+ exhausted T cell subset were screened using the FindAllMarker function in the R “Seurat” package, with logFC > 0.3 and p < 0.05 as the thresholds.

The differences in the biological functions of the T cell-related subsets were determined by gene set enrichment analysis (GSEA), gene ontology (GO) and KEGG pathway analyses using R ‘ABGSEase’ and ‘clusterprofiler’ packages with Hallmark, GO and KEGG reference gene sets. Maximum gene set size (maxGSSize) was set to 100, minimum gene set size (minGSSize) to 50, and p-value < 0.05 was considered statistically significant.

The DEGs in the CD8+ exhausted T cell subsets in the TCGA and GEO cohorts were corrected to the same level, and the prognostic genes were identified through univariate Cox regression analysis (p < 0.01). A random forest model was then built using the R package “randomForest”. The mean model inaccuracies for all genes were first estimated. The optimal number of trees in the random forest model was set at 500, and the dimensional effect sizes of the model were determined using the Gini coefficient method. Genes with importance indices higher than 0.6 were selected as disease hub genes. TRS was calculated using the following formula: TRS= Coef1×Geneexpression1+Coef2×Geneexpression2+⋯Coefn×Geneexpressionn

The Cox regression coefficient represents the prognostic value of each gene. Expression values for genes correspond to those of their modeled counterparts. Based on the median TRS, OV patients were classified as high-risk or low-risk. The R “survival” package [27,28] was used to plot Kaplan-Meier survival curves. Based on clinical features, the “rms” package of R software was used to develop a nomogram to predict survival of individual patients [29].

The infiltration of different immune cell populations was quantified based on the transcriptional profiles of patients using MCPcounter [30] and xCell [31] algorithms. In addition, the absolute content of the four T cell subsets were derived from the reference single-cell dataset. The data was visualized using boxplots, heat maps, and scatter plots.

The sensitivity score of the drugs in the Genomics of Drug Sensitivity in Cancer (GDSC) database was calculated with the R package ‘oncoPredict’ [32]. Wilcox’s test was used to identify the candidate drugs with p < 0.05 as the threshold for statistical significance.

The A2780 OV cell line and IOSE80 normal ovarian cell line were obtained from the American Type Culture Collection (ATCC, Rockville, MD, USA). The cells were cultured in RPMI-1640 medium supplemented with 10% FBS at 37°C and 5% CO₂.

RNA was extracted from the cells using TRIzol reagent (Takara, Japan) and reverse transcribed into cDNA using SuperScript II Reverse Transcriptase (Invitrogen, USA), The relative expression level of CD38 was analyzed by qRT-PCR and normalized to GAPDH. The primers are listed in Suppl. Table S1.

The CD38 siRNA was mixed with the RNAiMAX transfection reagent in Opti-MEM for 20 min at room temperature, and then transfected into the OV cells. After 12 h, the medium was discarded and RPMI-1640 medium was added. The procedure was repeated after 12 h. The CD38-siRNA target sequences were as follows: sense 5′-GCCUGAUCUAUACUCAAAUTT-3′ and antisense 5′-AUUUGAGUAUAGAUCAGGCTT-3′. CD38 knockdown was verified by qRT-PCR.

The transfected cells were washed with PBS, and harvested with trypsin digestion solution containing no EDTA (Solarbio, China). After centrifuging at 1000 rpm for 5 min, the cells were stained with 7-AAD (BD Biosciences, USA) and annexin-APC (BD Biosciences, USA) for 15 min.

The transfected cells were seeded on matrix gel-coated upper chamber of Transwell inserts in serum-free RPMI-1640 medium, and the lower chambers were filled with complete medium. After culturing for 24 h, the cells remaining on the upper membrane surface were removed with cotton swabs, and those that had adhered to the bottom surface were fixed with 4% paraformaldehyde and then stained with crystal violet. The cells were observed under a light microscope, and counted using ImageJ.

R (version 4.1.0) program was used for statistical analyses. Overall survival was evaluated using Kaplan-Meier (KM) and log-rank tests, and differences among subgroups were determined with Wilcox test and Kruskal test.

The study design is outlined in the flow chart in Fig. 1. The scRNA-seq data of obtained from 5 OV patients was screened, and the cells were annotated in terms of spatial distribution (Fig. 2A, left), as well as characteristic markers. Six cell types were annotated, including fibroblasts, T cells, myeloid cells, epithelial cells, endothelial cells, and B cells (Fig. 2A, right). The expression profiles of the eight classical marker genes are shown in Fig. 2B, and their expression in each cell type is shown in the bubble plots in Fig. 2C. Based on the absolute counts, the T cell subsets were predominant in almost all samples, suggesting a crucial role of these cells in OV progression (Fig. 2D). In addition, the content of the other cell types in each cell subset was significantly different, which confirms the cellular heterogeneity of OV.

Following dimensionality reduction cluster analysis of T cell subsets, we annotated four subsets, including CD8+ T cells, CD4+ T conv cells, regulatory T cells (Tregs) and CD8+ exhausted T cells (Fig. 3A). The expression of signature genes in each cell subset is shown in the bubble plots in Fig. 3B. The CD8+ T cells were the most abundant subset among the samples (Fig. 3C), whereas the CD8+ exhausted T cells were the least abundant in each sample. Analysis of the differentially expressed genes in the T cell subsets indicated that CXCL13 and HAVCR2 were significantly upregulated in the CD8+ exhausted T cells (Fig. 3D).

GSEA of tumor-associated pathways was performed to further explore the differences in the biological functions of the T cell subsets. The myogenesis and epithelial mesenchymal transition pathways were significantly activated in CD8+ T cells, and TNFA signaling via NFKB pathway in the CD4+ T conv cells. The Tregs were significantly associated with the oxidative phosphorylation, pancreatic beta cells and peroxidase pathways, whereas the CD8+ exhausted T cells showed significant enrichment of the interferon alpha response, interferon gamma response, MYC target V1, oxidative phosphorylation, E2F target and G2M checkpoint pathways (Figs. 4A and 4B). Subsequently, we analyzed tumor-associated pathways in different cell populations based on PROGENy score. As shown in Fig. 4C, The JAK-STAT and MAPK pathways were activated, and the p53 and PI3K pathways were inhibited in the CD8+ exhausted T cells. Receptor ligand pair analysis between CD8+ exhausted T cells and other T cell subsets revealed high enrichment of the HLA−A−CD8A, HLA−B−CD8A, and HLA−C−CD8A pairs (Fig. 4D). Given the central role of HLA molecules in immunity, our findings indicate that this subset may influence the response to immunotherapy. Furthermore, GO analysis of the DEGs between CD8+ exhausted T cells and controls showed significant enrichment of ATP metabolic process, aerobic respiration, cellular respiration and oxidative phosphorylation in biological processes (BP), redox-driven activity transmembrane transporter activity and NADH dehydrogenase activity in molecular function (MF), and mitochondrial inner membrane, mitochondrial protein complex and mitochondrial inner membrane protein complex in cellular compartment (CC) terms (Fig. 4E). Finally, the KEGG pathways significantly associated with the DEGs included those related to Huntington’s disease, prion disease, thermogenesis, amyotrophic lateral sclerosis, and other neurodegenerative diseases (Fig. 4F).

To identify the prognostic genes associated with CD8+ T cell exhaustion, we performed univariate Cox regression analysis of the DEGs in this subset. We screened 47 prognostic genes that were incorporated into a random forest classifier. Seven genes, including MOB1A, TIMM8B, CD38, HINT2, NAA38, DLEU2 and TXNDC17, were finally identified and prognostic models were constructed using a random forest classifier (Fig. 5A). The TRS was calculated as follows:

TRS= 0.25516×MOB1A+(−0.07000)×TIMM8B+(−0.12049)×HINT2+(−0.06368)×NAA38+(−0.15835)×CD38+(−0.04413)×TXNDC17+0.04650×DLEU2

According to the median risk score, the patients in TCGA cohort were divided into the low-risk and high-risk groups. As shown in Fig. 5B, the patients with low-risk score had significantly better OS. Using the median TRS of the training set, we classified patients in the external GSE140082 set as high- or low-risk. As expected, the low-risk group had significantly better OS compared to the high-risk group, thus validating our prognostic model (Fig. 5C). To accurately estimate the survival status of individual OV patients, we developed a predictive nomogram based on TRS, age and tumor stage. As shown in Fig. 5D, the TRS-based nomogram can accurately predict the survival of OV patients.

Except for CD38, the other genes in the prognostic model showed significant correlation with each other (Fig. 5E). In addition, TIMM8B, CD38, HINT2, NAA38 and TXNDC17 were significantly upregulated in the low-risk patients, whereas higher levels of MOB1A were detected in the high-risk group (Fig. 5F).

Since the immune landscape plays a key role in tumor development and progression, we compared the immune infiltration in the high-risk and low-risk groups using MCPcounter and xCell. According to the MCPcounter algorithm, patients in the low-risk group had significantly higher number of infiltrating CD8+ T cells, cytotoxic lymphocytes, myeloid dendritic cells, and NK cells compared to patients in the high-risk group, whereas the latter had significantly higher abundance of neutrophils (Fig. 6A). We also detected a significant association between CD38 and most immune-related genes (Fig. 6B). In addition, the B lineage cells and cytotoxic lymphocytes showed a significant negative correlation with the TRS (Fig. 6C). The xCell algorithm also indicated significant differences between the immune cell composition of the high- and low-risk groups. As shown in Fig. 6D, the aDC cells, M1 macrophages, M2 macrophages, Th1 cells and Th2 cells were significantly more abundant in the high-risk vs. the low-risk patients. Therefore, the poor prognosis in the high-risk group may be due to the distinct immune landscape. In addition, we found a significant negative correlation between CD4+ memory T cells and TRS, and a significant positive correlation between eosinophils and mast cells and TRS (Fig. 6E).

CD38 was the only model gene that was significantly associated with prognostic indicators in OV patients. In addition, CD38 expression levels were significantly higher in the OV tissues compared to that in the normal ovarian tissues (Fig. 7A). Consistent with this, the OV cell line A2780 expressed higher levels of CD38 compared to the normal ovarian cells IOSE80 (Fig. 7B). To further evaluate the biological function of CD38 in OV, we silenced the gene in A2780 cells. As shown in Fig. 7C, CD38 was significantly downregulated in the OV cells following gene knockdown. In addition, knocking down CD38 significantly increased the apoptosis rates in the OV cells (Fig. 7D), and inhibited their invasion ability in the Transwell assay (Fig. 7E). Thus, CD38 is a potential therapeutic target for OV.

In addition, we examined the relationship between TRS and the expression of immunomodulators. the TRS was negatively correlated with most immune checkpoints which suggests that TRS may be used to predict the efficacy of immunotherapy (Fig. 8A). To further evaluate the chemosensitivity of OV patients, we assessed the sensitivity scores for each compound. The sensitivity of six drugs was significantly different between the high and low risk groups. The patients in the high-risk group were more sensitive to AZD6738, Ribociclib, AZD3759, LCL161 and Navitoclax, whereas the low-risk group showed greater sensitivity to BI−2536 (Fig. 8B). Furthermore, there was a significant positive correlation between MOB1A and most of these drugs, and CD38 showed correlation with BI−2536, RO−3306 and Tozasertib (Fig. 8C).