A novel defined programmed cell death related gene signature for predicting the prognosis of serous ovarian cancer

We collected 375 tumor tissues samples (TCGA-OV) (Homo sapiens) from the University of California Santa Cruz Xena (UCSC Xena, https://​xenabrowser.​net/​datapages/​) [19], ensuring all selected samples had complete survival data. These samples provided transcriptome sequencing data with fragments per kilobase million (FPKM) expression values, along with relevant clinical information, such as age, histologic grade, and clinical stage (Table S1). Additionally, we sourced 88 healthy ovarian tissue samples (Homo sapiens) from the Genotype-Tissue Expression database (GTEx, https://​gtexportal.​org/​home/​) [20]. These samples provided the transcriptome sequencing FPKM expression profile and count matrix. In addition, we downloaded the RNA-seq dataset GSE63885 [21] from the Gene Expression Omnibus database(GEO, https://​www.​ncbi.​nlm.​nih.​gov/​geo/​) through the “GEOquery” R package [22]. This dataset was derived from the GPL570 [HG-U133_Plus_2] Affymetrix Human Genome U133 Plus 2.0 Array, which focuses on Homo sapiens and comprises a total of 75 ovary tumor samples after removing missing survival data. We further downloaded scRNA-seq data GSE184880 from the GEO database [23], which was derived from the platform GPL24676 Illumina NovaSeq 6000 (Homo sapiens) and contained seven SOC samples without treatment and five control samples. The inclusion and exclusion criteria of this study were defined as follows: (1) The inclusion criteria: ① Patients diagnosed and treated for SOC initially, excluding those with recurrent SOC; ② Complete clinical and pathological data. (2) The exclusion criteria: ① Excluded patients with incomplete pathologica or clinical data; ② Patients with incomplete follow-up time, other causes of death and unknown death status; ③ Patients with multiple tumors and non-primary tumors.

Initially, we collected 268 programmed cell death genes (PCD genes) from existing literature sources [24, 25]. Using scRNA-seq datasets, we identified 3,000 cell Differential genes (cellDiffgenes) intersecting with PCD genes to obtain our target gene. Subsequently, we obtained a combined dataset by emerging OC samples from TCGA databases with healthy ovarian samples from GTEx databases using the “ComBat” R package [26]. We successfully identified differentially expressed programmed cell death genes (DEPCDGs) in SOC by analyzing these target genes in the combined dataset using the “limma” R package [27] (p < 0.05 and | logFC (Fold Change) |> 1). Among the DEPCDGs, those with a p-value less than 0.05 and logFC more than 1 were classified as up-regulated. Conversely, DEPCDGs with a p-value less than 0.05 and logFC less than -1 were categorized as down-regulated. The “heatmap” and “ggplot2” R packages were employed to generate visual representations of heat and volcano maps.

Gene Ontology (GO) [28] is a commonly used approach in conducting comprehensive investigations of functional enrichment studies. This method encompasses the examination of cell composition (CC), biological process (BP), and molecular function (MF). Similarly, the Kyoto Encyclopedia of Genes and Genomes (KEGG) [29] is an extensively utilized database encompassing comprehensive data on genomes, biological processes, diseases, and pharmaceuticals. To analyze the functional characteristics and pathway enrichment of the DEPCDGs, we performed GO and KEGG analyses using the “ClusterProfler” R package [30] and graphically represented using the “ggplot2” R package. For statistical significance, we defined enrichment as a function or pathway term with a false discovery rate (FDR) less than 0.25 and a p-value less than 0.05. The p-value adjustment used the Benjamini-Hochberg (BH) approach [31].

We performed survival analysis on DEPCDGs utilizing the “survival” R package [32] and identified key prognostic genes with statistical significance (p < 0.05). These key prognostic genes were then selected as key DEPCDGs for further analysis.

We imported raw data from SOC samples in the scRNA-seq dataset utilizing the “Seurat” R package (version 4.0) [33] and created Seurat objects for subsequent analysis. We applied gene < 200 or > 3,000 filtration conditions to remove low-quality cells [34]. The proportion of mitochondrial genes in relation to the total genetic material can indicate cellular homeostasis. Cells with mitochondrial gene content > 10% were excluded from further analysis due to potential stress. Consequently, we obtained a final set of 3,555 cells for subsequent analysis.

The scRNA-seq data was normalized using the LogNormalize method. We identified cellDiffgenes in individual cells after controlling for the relationship between average expression and dispersion. Next, we employed Principal Component Analysis (PCA) to decentralize all genes and cluster all cells. Subsequently, we displayed the resulting cell subclusters utilizing Uniform Manifold Approximation and Projection (UMAP) [35]. The cell type of each cluster was determined by referencing the Human Primary Cell Atlas (HPCA) dataset using the singleR method [36].

CIBERSORT (https://​cibersortx.​stanford.​edu/​) is a computational tool that employs linear support vector regression to deconvolute the transcriptome expression matrix. Its purpose is to estimate the composition and number of immune cells within a mixture of cells [37]. We utilized the CIBERSORT algorithm to determine the fraction of 22 immune cell types, exploring the association between key DEPCDGs and the immunological microenvironment. The relative abundance of immune cells in a dataset sample can be calculated using single-sample gene set enrichment analysis (ssGSEA) [38]. The immune cell enrichment scores of combined datasets were assessed by using ssGSEA with the “GSVA” R package. This analysis was performed based on the relative abundance of each immunocyte infiltrate in every sample. Samples with a p-value less than 0.05 were filtered and included in the output. Finally, the correlation analysis results between key DEPCDGs and infiltrating immune cells in combined datasets were visually represented using the “pheatmap” R package. The core genes for further analysis were selected based on the most relevant key DEPCDGs.

Within the TCGA-OV dataset, SOC patients were classified into high- and low-expression groups by utilizing the median value of the core gene. Differential analysis was performed using the “limma” R package to identify the differentially expressed genes (DEGs) with statistical significance (p < 0.05 and | logFC |> 1). To determine the biologically significant pathways mediated by the hub gene, we performed GO and KEGG enrichment analyses using the “ClusterProfiler” R package [39] and visualized using the “ggplot2” R package.

In order to evaluate the contribution of DEGs to the phenotype, we employed Gene Set Enrichment Analysis (GSEA) [40]. GSEA is a computational methodology in which genes in a predetermined genetic set are analyzed within the gene list ordered by phenotypic correlation. We performed enrichment analysis on all DEGs with high and low phenotype correlations in both groups using the “clusterProfiler” R package. The parameters employed for GSEA were as follows: a seed value of 2020 was utilized for random number generation, 10,000 computations were performed, and each gene set contained a minimum of 10 genes and a maximum of 500 genes. Enrichment analysis was conducted using the “c2.cp.v7.2.symbols.gmt” gene set obtained from the Molecular Signatures Database (MSigDB) [41] via the GSEA method. We defined statistically significant enrichment as a pathway or function term with an FDR less than 0.25 and a p-value less than 0.05. The p-value correction was conducted using the BH method.

DEGs identified from hub gene grouping were selected as candidate genes. To investigate their prognostic value for SOC, we assess the correlation between these candidate genes and survival outcomes via univariate Cox regression analysis using “survival” and “forestplot” R packages. Based on the DEGs with noteworthy prognostic value (p < 0.05), we conducted multivariate Cox regression analysis to calculate regression coefficients and develop a risk model. This model enabled us to assign a risk score (RS) to each tumor sample using the following formula,

Where N denotes the number of genes, coef(k) represents the multivariate Cox regression coefficient, and x(k) represents the expression value of each gene.

Our data analysis identified seven prognosis-related feature genes: CD3E, CD2, IL2RG, FCGBP, RARRES1, UBD, VSIG4, and STAB1. Subsequently, the patients were classified into high- and low-risk groups according to the median of the RS. To evaluate the predictive effectiveness of our risk model, we employed several methods: the Risk Triptych, time-dependent receiver operator characteristic (time-ROC) curve analysis [42], Kaplan–Meier (K-M) curve analysis [43], and decision curve analysis (DCA).

In order to determine the potential independence of the prognostic factor, we assess the correlation between survival outcomes and variables such as RS, age, stage, and grade via univariate Cox regression using the “survival” and “forestplot” R packages. Furthermore, we explored independent influencing factors through multivariate Cox regression and visualized them in forest plots. A nomogram was ultimately constructed utilizing the RS and clinical characteristics in order to forecast the prognosis of SOC. The performance of this nomogram was subsequently assessed through the use of the Calibration curve and ROC curve.

We performed statistical analysis in this study using RStudio (version 4.2). The Kruskal–Wallis test was employed to compare groups consisting of three or more, while the Wilcoxon rank sum test was utilized for the comparison of two groups. Spearman’s method was employed for correlation analysis. The “survival” R package was employed to conduct univariate and multivariate Cox analyses. Additionally, survival differences were displayed using K-M survival curves. The Log-rank test was employed to evaluate the extent of the disparity in survival durations among the various groups of patients. All statistical tests were conducted with bilateral p-values, and a significance level of p < 0.05 was employed.

In order to provide a clearer understanding of the research process, we presented the workflow of our study in Fig. 1.

We identified a total of 97 PCD-related cellDiffgenes by intersecting 3,000 cellDiffgenes from the scRNA-seq dataset with 268 PCD genes from existing literature sources. Moreover, the 171 genes are uniquely associated with the PCD gene set, which were not present among the cellDiffgenes., hinting at their potential involvement in specific PCD pathways relevant to ovarian cancer (Fig. 2A). A combined dataset was obtained by emerging OC samples from TCGA databases with healthy ovarian samples from GTEx databases. We identified 48 DEPCDGs by performing differential analysis on the expression of PCD-related cellDiffgenes in the combined dataset using the “limma” R package ( |logFC|> 1 and p < 0.05). Among them, 18 genes were up-regulated ( logFC > 1 and p < 0.05), and 30 genes were down-regulated ( logFC < -1 and p < 0.05). The volcano plot visualized these DEPCDGs (Fig. 2). The differential expression of DEPCDGs between various sample groups in combined datasets was analyzed. The results of this analysis were visualized using a heatmap plot generated by the “pheatmap” R package (Fig. 2C).

Forty-eight DEPCDGs were analyzed by GO and KEGG pathway enrichment analysis (Table 1). GO analysis revealed enrichment of DEPCDGs in pathways such as extrinsic apoptotic signaling pathway, cellular response to chemical stress, regulation of extrinsic apoptotic signaling pathway, response to oxidative stress, and regulation of apoptotic signaling pathway (Fig. 2D). In the KEGG pathway analysis, several enriched pathways of DEPCDGs were identified, including legionellosis, proteoglycans in cancer, and salmonella infection (Fig. 2E).

We conducted survival analysis on 48 DEPCDGs in the TCGA-OV group and identified seven key DEPCDGs with prognostic significance in SOC (p < 0.05). These genes were the third comparative assessment of techniques of protein structure prediction (CASP3), growth arrest and DNA-damage-inducible protein 45 beta (GADD45B), GNA15, Granzyme B (GZMB), cytokine interleukin-1β (IL1B), Interferon-stimulated gene 20 (ISG20), and RhoB (RHOB) (p < 0.05) (Fig. 3A-G). The differential expression of these seven genes in the combined dataset shown in Fig. 3H, with GASP3, GNA15, GZMB, and IL1B strongly expressed in the tumor group, while GADD45B, ISG20, and RHOB were lowly expressed (p < 0.001). Additionally, correlation analysis revealed that GNA15 and IL1B were relatively highly correlated (Fig. 3I). These findings suggest the potential prognostic significance of these key DEPCDGs in SOC.

In our study, RNA sequencing was performed on single cells from seven ovarian cancer samples. To ensure the overall quality of single-cell data, we implemented filtering conditions to eliminate low-quality cells and batch effects. Specifically, we set the filtering condition as follows: the number of RNA features (nFeature_RNA) had to be between 200 and 3,000, and the percentage of mitochondrial genes (percent.mito) had to be below 20%. Through this filtering process, we successfully identified and retained a total of 3,555 high-quality cells.

Following data normalization and gene centralization, we performed PCA dimensionality reduction by extracting the top 3, 000 cellDiffgenes at the single-cell level. To identify distinct groups of cells with similar gene expression profiles, we employed the top 50 principal components for clustering. This clustering analysis yielded 13 independent clusters, which were subsequently visualized using UMAP (Fig. 4A).

Using the HPCA data to identify cell type of each cluster, we found eight cell subtypes after annotating, including intermediate monocytes, myeloid dendritic cells (mDCs), NK cells, plasmablasts, progenitor cells, switched memory B cells, Tregs, and Vdelta2 gamma-delta (Vδ2 gδ) T cells (Fig. 4B). Expression levels of seven key DEPCDGs in various cell types visually showed by the UMAP (Fig. 4D-J). Our findings revealed that GADD45B exhibited a high expression level in intermediate monocytes, GZMB was highly expressed in NK cells, and IL1B showed significant expression in intermediate monocytes. Interestingly, GNA15 displayed low expression across all cell subtypes. These results were further validated by the violin diagram (Fig. 4K-Q) and heatmap (Fig. 4C).

In order to examine the relationship between key DEPCDGs and immune infiltration in the transcriptome, we conducted an analysis of immune cell infiltrates in the combined datasets utilizing CIBERSOFT and ssGSEA methods. The CIBERSORT algorithm analyzed immune infiltrates abundance of Tregs, gδT cells, NK cells activated, monocytes, dendritic cells resting, dendritic cells activated, and B cells memory. The results revealed that the tumor group exhibited higher immune infiltration levels of Tregs, gδT cells, dendritic cells resting, and dendritic cells activated compared to the normal group. Additionally, the infiltration abundance of monocytes displayed a statistically significant decrease in the tumor group (Fig. 5A, p < 0.001). Using the ssGSEA algorithm, we analyzed immune infiltrates’ abundance of gδT cells, plasmacytoid dendritic cells, NK cells, Tregs, monocytes, activated B cells, activated dendritic cells, immature B cells, and immature dendritic cells. The results indicated that gδT cells, NK cells, Tregs, monocytes, and activated dendritic cells exhibited higher immune infiltration levels in the tumor group compared to the normal group (Fig. 5C, p < 0.01). We screened out GNA15 as a core DEPCDG in SOC, as it showed a strong correlation with dendritic cells resting in the CIBERSORT algorithm (R = 0.35), none of the other genes were highly correlated with cells (Fig. 5B). Similarly, GNA15 demonstrated a high correlation with all identified immune cell subtypes in the ssGSEA algorithm(all R ≥ 0.58) (Fig. 5D). Specific high correlation results were shown in Fig. 5E-H, where GNA15 exhibited a correlation with activated dendritic cells (R = 0.80, p < 0.001, Fig. 5E), monocytes (R = 0.70, p < 0.001, Fig. 5F), NK cells (R = 0.73, p < 0.001, Fig. 5G), and Tregs (R = 0.82, p < 0.001, Fig. 5H).

Based on our analysis results, we specifically focused on the core gene GNA15 for single-gene bioinformatics analysis. The HGSOC patients were categorized into high- and low-expression groups using the median value of the GNA15 in the TCGA-OV dataset. Through differential analysis, we identified 5 down-regulated DEGs and 180 up-regulated DEGs (|logFC|> 1, p < 0.05) (Fig. 6A). To gain insights into biological pathways modulated by GNA15, we conducted GO and KEGG enrichment analyses on these DEGs. The GO enrichment analysis revealed a predominant involvement of DEGs in leukocyte-mediated immunity, B cell-mediated immunity, antigen-binding pathways, and other relevant pathways (Fig. 6B). Meanwhile, the KEGG enrichment analysis discovered that DEGs enrichment in pathways associated with staphylococcus aureus infection, phagosome, and other related pathways (Fig. 6C). Additionally, we employed GSEA analysis to investigate the implications of GNA15 expression further (Table 2). The results of our study indicate a significant association between the high expression of GNA15 and B cell receptor signaling pathway, T cell receptor signaling pathway, and Toll-like receptor signaling pathway (Fig. 6D-F). Conversely, low expression of GNA15 was enriched in the ribosome, spliceosome, and RNA polymerase pathways (Fig. 6G-I).

A predictive model was created utilizing the core gene GNA15. Initially, a univariate Cox regression analysis was performed on DEGs between high and low GNA15 expression. Our results revealed 11 genes with significant prognostic value (p < 0.05) in SOC (Fig. 7A). Subsequently, we performed multivariate Cox regression analysis on these 11 genes to construct the predictive model consisting of eight genes: CD3E, CD2, IL2RG, FCGBP, RARRES1, UBD, VSIG4, and STAB1 (Fig. 7B). We categorized patients into high- and low-risk groups based on the RS median. The Risk Triptych showed the strong predictive capacity of the model in both TCGA-OV and GSE63885 datasets (Fig. 7D-E, H-I). Furthermore, the K-M survival curve indicated that the high-risk group had worse prognoses compared to the low-risk group in both TCGA-OV and GSE63885 datasets (p < 0.0001, p = 0.048) (Fig. 7C, G). The timeROC curve showed the RS’s strong predictive ability for overall survival (OS) in SOC patients, with AUCs of 0.690, 0.694, and 0.713 for 1-year, 3-year, and 5-year respectively (Fig. 7F). Finally, the DCA confirmed the substantial predictive ability of the RS signature (Fig. 7J).

We conducted a comprehensive analysis to determine whether RS could used as an independent prognostic factor. Firstly, a univariate Cox regression analysis was performed to assess the relationship between RS, age, stage, grade, and OS. The forest plots revealed that age and RS are significantly related to OS (Fig. 8A). Further analysis was undertaken using multivariate Cox regression, considering the aforementioned variables. Remarkably, both RS and age emerged as independent prognostic factors for predicting patients’ OS without relying on other clinical features (Fig. 8B). Subsequently, we constructed a nomogram model that incorporates RS along with three other clinical features for forecasting SOC patient outcomes (Fig. 8C). The good predictive power of this model was demonstrated by the Calibration curve (Fig. 8D). Moreover, the ROC curve analysis of the nomogram displayed its precise predictive ability for OS of SOC patients, with AUC values of 0.670, 0.650, and 0.653 for the 1-year, 3-year, and 5-year OS predictions, respectively (Fig. 8E).