2.1. Datasets

Bulk RNA-seq data, mutation data, and clinical characteristics of TCGA-OV were retrieved from the UCSC Xena website. There were 360 samples that contained survival information, and the expression profiling data were in TPM format. The validation model used gene expression profiling information of 278 OC patients from GSE9891. The “limma” and “sva” software tools were applied to combine the two data and correct for batch effect [10]. TISCH gathered 76 sets of tumor single-cell transcriptome data from the Gene Expression Omnibus (GEO) database and ArrayExpress together with associated patient data. The data spanned 27 different cancer types. The EMTAB8107 cohort of TISCH, which has a total of 24,781 cells from cancer and paracancerous tissue, is where the OC single-cell sequencing data came from [11]. From GSE203612, the primary OC spatial transcriptome data for the 10x Visium platform was retrieved [12]. The term “purine metabolism” was used to derive 163 purine metabolism–related genes (PMRGs) from the Molecular Signatures Database (MSigDB) (Table S1).

2.2. Analysis With Single-Cell RNA Sequencing Data

The scRNA-seq data were analyzed using the “Seurat” and “SingleR” R packages [13]. We only retained genes present in more than two cells and filtered out cells with less than 200 or more than 10,000 genes, less than 1000 molecules, and more than 20% of nuclear and translational genes. Normalization was conducted for the data using “NormalizeData” from the R package “Seurat.” With the “FindVariableFeatures” function, the first 2000 highly variable genes were identified. After that, principal component analysis (PCA) was applied to minimize the dimensionality of the data. Essential principal components (PCs) were determined using the JackStraw analysis, with the top 20 PCs selected. For cell clustering, we utilized the “FindNeighbors” and “FindClusters” functions on the integrated data and visualized cell clusters by UMAP and tSNE. Marker genes for various cell types were identified through a Wilcoxon test, carried out using the “FindAllMarkers” and “FindMarkers” routines. We annotated different cell subpopulations by merging data from the original text with annotations from the TISCH database, relying on common marker genes.

Package “Monocle” was applied for the pseudotemporal analysis of cellular developmental trajectories in epithelial cells [14], and unsupervised analysis was utilized to choose developmental differential genes. Differential gene testing for the pseudotime analysis was based on a cutoff for significance q-value < 0.01. DDRTree was used to reduce the dimensionality of the data. We scored the gene set for single-cell data using five widely used techniques, including AddModuleScore, single-sample Gene Set Enrichment Analysis (ssGSEA), AUCell, UCell, and singscore, based on the expression of 163 PMRGs. With the “AddModuleScore,” we computed the average expression value for each gene, divided the expression matrix into segments accordingly, and randomly selected control genes as background values from each segment [15]. “AUCell” employed the area under the curve to determine whether an input gene collection is enriched for expressed genes in the samples [16]. Based on the gene expression patterns in single-cell transcriptome data, we clustered cells using techniques like dimensionality reduction and assigned a predicted cell type label to each cell [17]. A cell state evaluation technique called a “singscore” was used to measure how active a specific biological function or process is in a single sample or cell. Taking into account the weights and orientations of the genes, it determines the cell state score of a sample or cell on the basis of the set of genes in a gene expression profile.

Finally, the “CellCall” tool combined intracellular and intercellular data to infer intercellular communication networks [18]. The ability to combine intracellular transcriptional factor expression with intercellular ligand–receptor communication, which enables the analysis of changes in receptor cellular pathways brought on by the communication between two specific cell types, is its most significant feature.

2.3. Analysis With Spatial Transcriptomic Data

Additionally, Seurat in R was used to analyze spatial transcriptome data. The function “SCTransform” was used to normalize, scale, and identify the most variable features in UMI counts. After that, the UMI counts were downscaled, and unsupervised cluster analysis was performed using “RunPCA.” Subgroups and genes were visualized using the SpatialFeaturePlot tool. The “scMetabolism” R package is in view of a standard single-cell matrix file and scores each cell using the VISION algorithm to determine how active the cell is in each metabolic pathway [19].

Scanpy is a Python-based software suite that includes preprocessing, visualization, clustering, and differential expression analysis. The University of Queensland’s Institute of Molecular Biosciences produced the stlearn software, which was used to combine data from ligand pairings and gene expression to infer relationships.

Reverse Compositional Transcriptomics Deconvolution (RCTD) calculates the geographic distribution and relative abundance of cell types [20]. Multiview Intercellular SpaTial Modelling Framework is an interpretable multiview framework based on machine learning for parsing intercellular interactions in highly multiplexed data. It allows researchers to analyze spatial transcriptomic data without the necessity for cell type identification, and it offers fresh insights and hypotheses that help in the understanding of cellular interaction patterns and mechanisms [21]. MISTy is now available as a R package named mistyR, complete with full documentation and instructions.

2.4. Construction of Prognostic Models

2.5. Analysis of Nomograms and Mutations

We built a nomogram to visualize the chance of survival. We inferred the net advantage of the nomogram and clinical features alone using consistency index analysis and decision curve analysis (DCA). The R “oncoplot” function was applied to investigate detailed mutation profiles.

2.6. The Evaluation of Immunotherapy and Immunological Microenvironment

The “estimate” was applied to assess malignant tissue mesenchymal and immune cell abundance [22]. The amount of immune infiltration was then evaluated using the TIMER 2.0, which incorporates the outcomes of seven assessment approaches. In the form of a heat map, these data were used to quantify the relative fraction of immune cell infiltration. ssGSEA can assess the proportions of immunological-related activities. We assessed the two gene sets by the GSVA technique to analyze the connection with risk scores. The association between selected genes and immune genes was explored further [23]. GSEA was carried out comparing two groups through the MSigDB database “c2.cp.kegg.v6.2.symbols.gmts” file in order to explore the differences in biological functions between various populations. Tumor immune dysfunction and exclusion (TIDE) stands for tumor immune dysfunction and rejection. It is a computational framework for analyzing the gene expression profiles of tumor samples to determine the possibility of tumor immune escape. Comprehensive immunogenomic analysis results are conducted via the Cancer Immunome Atlas (TCIA) web application. Immunophenotype score (IPS) ranges from 0 to 10 [24].

2.7. qRT-PCR to Verify the Expression of NME6 in Samples

Three families of ovarian cells were obtained. The acquisition process and sample storage and processing are the same as before [25]. The extraction of RNA and the implementation of qRT-PCR were also operated strictly in accordance with the protocols. The forward sequence primer is TCCAGCTCACTCTAGCCCTG, and the reverse sequence primer is CGGTAAAACCTCTGGCAATCTT.

2.8. Statistical Analysis

R 4.1.3 and Python 3.9 were used to conduct all of the analyses. The Kaplan–Meier survival analyses and log-rank tests were applied to each dataset in order to assess predict values and compare the survival of patients in various subgroups. The connection of continuous variables was examined by the nonparametric Wilcoxon rank sum test. p < 0.05 was regarded as statistically significant.

3.1. Single-Cell Enrichment Scores for Genes Involved in Purine Metabolism

In order to explore the heterogeneity of OC and to evaluate variations between cells within and near the tumor, we analyzed cell types and gene expression within them. The PCA was used to perform dimensionality reduction using 2000 highly variable genes, and the top 20 PCs were selected for additional analysis. Following the identification of 33 distinct cell clusters, we integrated cell annotations from the original literature using established marker genes to annotate these clusters (Figures 1a, 1b, 1c, and 1d). The percentage of various cell types in the five samples is shown in Figure 1e. Using several well-known cell marker genes, we demonstrated the expression of these genes in each type of cell (Figure 1f). Additionally, a heat map displays the relative expression of the estimated marker gene in each cell subpopulation (Figure 1g). We employed the five most popular methods to score the gene sets based on the expression of 163 PMRGs. Purine metabolism scores (PMSs) were comparatively greater in epithelial or tumor cells, as seen in Figure 1h,i. We also found that the PMSs were comparatively higher in tumor-associated monocyte macrophages, fibroblasts, T/NK cells, and epithelial cells (Figure 1j). We identified 44 genes correlated with purine metabolism between epithelial cells in tumors and normal tissues (Table S2).

3.2. Pseudotime Analysis and Intercellular Communication Analysis

In total, we counted three distinct cell states in developing epithelial cells. Cluster 4 is an illustration of the early stage of cell formation (Figure 2a). The heat map in Figure 2b shows the expression of 44 PMRGs that are differently expressed at various developmental stages. Figure 2c displays the analysis results for the investigation of signaling pathway activity. For instance, fibroblasts displayed higher levels of TGF-β signaling pathway activity in epithelial cells with high scores. The TGF-β signaling pathway is essential for the growth and development of cells and tissues. It also has a vital regulatory correlation with the regulation of cell proliferation, interstitial cell production, differentiation, and death, as well as immune function, inflammatory response, organ formation, and the healing of wounds. The loops in Figure 2d show the strength of ligand receptor signaling between different cell types, with high-scoring epithelial cells exhibiting stronger cellular interactions with fibroblasts. High-scoring epithelial cells interacted with fibroblasts through ligand–receptor pairs such as WNT7A-FZD1 more strongly (Figure 2f). We then focused on ligand–receptor pairs and associated transcription factors between epithelial cells and fibroblasts with higher PMS (Figure 2e).

3.3. Purine Metabolism–Related Characteristics in stRNA-Seq

A total of seven cellular subpopulations were obtained in space after downscaling and clustering the spatial transcriptome sequencing data of an OC patient. The SCTransform’s method was used to correct the sequencing depth and a series of normalization processes (Figure 3a,b). Subpopulation 0 was predominately found in the tumor core of the OC, as indicated by annotations in the original literature, and 44 differentially expressed PMRGs in the prior study were strongly expressed in Subpopulation 0 (Figure 3c). Among the metabolic activity of several cell subpopulations, Subpopulation 0 was intimately linked to metabolic processes involving purines and pyrimidines (Figure 3d). Purine and pyrimidine metabolic activities were primarily abundant in the tumor’s core, as seen in Figure 3e,f. After normalizing and clustering the spatial transcriptome data, we obtained nine cell subpopulations (Figure 3g). It is interesting to note that Cluster 0, which is located in the peripheral region of the tumor, differentiates toward Cluster 2, which is located in the tumor’s core, to carry out cell differentiation (Figure 3h). Finally, we deduced the dominating cell type at each position in the space. The core part of the tumor was where epithelial cells with low and high PMSs were concentrated, whereas the tumor’s periphery was where fibroblasts were concentrated (Figure 3i). The MISTy results revealed that epithelial cells with higher PMS interacted with fibroblasts more frequently in space (Figure 3j,k).

3.4. Building and Evaluating Prognostic Models

One hundred and sixty-three PMRGs were applied to build predictive signatures in bulk-seq for OC in order to assist clinical decision-making. Three hundred and sixty OC samples from the TCGA and 278 OC samples from the GSE9891 were used as the training set and the external validation set, respectively. Firstly, 26 PMRGs affecting the OS of OC patients were screened out (Table S3). A 12-gene signature was created to prevent overfitting of prognostic characteristics and to focus the genes that predict OS (Figure 4a,b). Based on the formula, risk ratings were computed for each patient (Figure 4c). Our model was utilized to determine the risk score using the previously mentioned formula (Table S4). For each patient, risk maps display specific survival results (Figure 4h,i). Participants with high risk had lower OS compared to those in the low-risk group, according to the survival curves (Figure 4d,e). It also showed great performance in inferring OS, according to the ROC curves (Figure 4f,g).

3.5. Nomograms and Clinical Features

The risk score is an independent predict factor when compared to other features (age, grade, and stage) (Figure 5a,b). In view of previously mentioned link between clinical characteristics and risk ratings, we constructed a nomogram to predict the OS (Figure 5c). The calibration curves also demonstrated that it could generate accurate predictions (Figure 5d). Furthermore, the risk score and nomogram score had a substantially higher AUC at 3 years than other clinicopathological variables (Figure 5e). The 3-year DCA curves and C-index revealed that our risk score had the biggest net benefit (Figure 5f,g), and risk grouping was connected with patient STAGE (Figure 5h). The high-risk group was classified at a higher level (Figure 5i). We are more certain that the risk score and nomogram are a reliable clinical prediction scoring system based on the outcome analysis.

3.6. Mutational Landscape

Given that genetic mutations have great association with a patients’ personalized treatment. The most often mutated type of these genes, as indicated in Figure 6a, was missense-mutation. However, the overall mutation frequency was exceedingly low. The distribution of the most often altered genes in OC in risk score groupings was studied (Figure 6b). TMB, however, did not differ across individuals in the groups (Figure 6c,d). The patients were then divided into four groups in view of their median TMB values and risk scores. The results revealed that patients with low risk scores and high mutations had a relatively better prognosis (Figure 6e).

3.7. Analysis of Immune Infiltration and Immune Function

Clinical results and treatment response of patients are influenced by the TME. Among them, tumor-infiltrating immune cells (TIICs) have a substantial effect on the advance of tumors and the effectiveness of antitumor treatments. TIIC’s distribution and composition are directly connected to the tumor development [26]. Therefore, we evaluated the immune landscape of the high- and low-risk score groups (Figure 7a). According to the findings, immune cell infiltration and immune pathway scores were higher in low-risk groupings (Figure 7b,c). Nearly all ICs were more expressed in the grouping with low risk scores (Figure 7e). Additionally, the immune complexes (ICs), immunological scores, and immune cell infiltration in various risk categories were shown using heat maps (Figure 7d). We used GSEA to investigate any potential changes in biological function and chose the four most crucial signaling pathways for each due to the differing prognoses and immune infiltration profiles of patients in subgroups (Figure 7f,g). The low-risk subgroups were primarily enriched in immune response and cytokine pathways, whereas the high-risk subgroups were primarily linked to signaling pathways in various types of tumors. The immunosuppressive microenvironment of OC tumors may correspond to patients with higher risk scores, which could lead to a decreased incidence of immunotherapy response.

3.8. Prediction of the Outcomes of Immunotherapy

With greater IPSs indicating stronger responses to ICB, violin plots demonstrate the link between risk groups and IPSs. Individuals from the low-risk group were thought to respond favorably to immune checkpoint inhibitors (Figure 8a). We then examined the associations between risk scores and ICB response profiles because the immune microenvironment mediates ICB responses and discovered that there was a substantial inverse relationship between risk scores and these ICB response profiles (Figure 8b). We conducted correlation analyses utilizing 12 model genes and traditional immune-related genes to further explore the variations in immune responses between groups (Figure 8c). The gene expression profile of tumor samples is reflected by TIDE as a measure of the likelihood of tumor immune escape. Poorer ICB response and patient survival were linked to higher tumor TIDE prediction scores. Above all, the high-risk group exhibited greater dysfunction and exclusion scores as well as TIDE scores that were relatively higher (Figure 8d).

3.9. NME6+ Epithelial Cells as a Potential Immunotherapy Target

According to the intersection of the genes with the 44 model genes in Table S2, three PMRGs were considered to affect the progression and treatment of OC (Figure 9a). The spatial map shows the expression of GMPR, NME6, and POLR3H (Figure 9b). We categorized single-cell sequenced epithelial cells into the expression-positive and expression-negative groups. There was no significant difference between the TIDE results and survival analysis results for GMPR+ and POLR3H+ epithelial cells (Figure 9c,d). The high NME6+ epithelial cell expression group had a better prognosis and statistically significant difference in the TIDE scores (Figure 9e). The number of NME6+ epithelial cells showed significant differences in the TIDE scores and the amount of cancer-associated fibroblast content, and they were all significantly negatively correlated (Figure 9f,g). Therefore, we speculated that NME6-expressing positive epithelial cells would be a protective factor for OC patients. We also confirm the expression of NME6 in IOSE and OVCAR3 which are normal ovarian epithelial cell lines and OC cell lines, respectively (Figure S1, Table S5). What can we see is that normal ovarian cell lines have higher NME6 levels than those in OC cell lines.