Gene expression profiling of ovarian cancer reveals candidate diagnostic and prognostic biomarkers: An integrated bioinformatics approach

Data acquisition

In this study, microarray datasets (GSE14407, GSE18520, GSE26712 & GSE27651) were obtained from Gene Expression Omnibus (GEO). In selecting datasets for this study, we prioritized those generated using the Affymetrix Human Genome U133 Plus 2.0 Array (HG-U133_Plus_2) and the Illumina HiSeq 2000 platform. These platforms were chosen because they provide high-throughput, high-resolution sequencing data, ensuring data quality, consistency, and comparability. Datasets from these platforms were further selected based on the presence of both tumor and normal tissue samples. The overall workflow of the bioinformatics analysis for this study is summarized in Figure 1.OPEN IN VIEWER

Differential expression analysis

DEGs between ovarian tumor and normal tissue samples were identified using the R software (version 4.5.1) with the limma package (version 3.66.0). ¹⁹ For each dataset, raw expression data were retrieved from GEO in Series Matrix format and imported into R using the GEOquery package. ²⁰ Expression matrices were examined for scale consistency; datasets on a linear scale were log₂ transformed before analysis, while those already on a log₂ scale were used directly. Probe-level expression values were mapped to official gene symbols using platform annotation files. For probes mapping to multiple genes, indicated by “///” separators in the annotation, were excluded to ensure unambiguous gene-level quantification. Where multiple probes mapped to the same gene symbol, the probe with the highest mean expression across all samples was retained.

For each dataset, samples were assigned to two groups: normal ovarian tissue (Group 1, reference) and ovarian tumor tissue (Group 2, experimental). A design matrix was constructed accordingly, and differential expression was assessed using empirical Bayes moderated t-statistics. Multiple testing correction was applied using the Benjamini-Hochberg method to control the false discovery rate (FDR). Genes with an absolute log₂ fold change (|log₂FC|) ≥ 1.2 and an adjusted p-value < 0.05 were considered significantly differentially expressed. This threshold was selected to prioritize genes with biologically meaningful expression differences while minimizing noise-driven signals. Importantly, both upregulated (log₂FC ≥ +1.2) and downregulated (log₂FC ≤ −1.2) genes were retained for downstream analysis to ensure a comprehensive and biologically complete representation of the transcriptomic landscape of OC. Additionally, to identify genes consistently dysregulated across multiple independent datasets, overlapping DEGs were determined separately for upregulated and downregulated gene sets using Venn diagram analysis implemented in the ggvenn package in R. ²¹ Only genes present in all four datasets were carried forward for downstream functional enrichment, network construction, and clinical relevance evaluation. A total of 96 common upregulated and 113 common downregulated DEGs were identified, yielding a final set of 209 overlapping DEGs. Volcano plots were generated using ggplot2 to visualize the distribution of DEGs across each dataset.

Functional enrichment analysis

To explore the biological significance of the overlapping DEGs, functional enrichment analysis was performed separately for the common upregulated and common downregulated DEG sets using the clusterProfiler package (version 4.18.4) in R. ²² Before analysis, gene symbols were converted to Entrez IDs using the bitr () function with the human genome annotation database org. Hs.eg.db. ²³ Genes that could not be mapped using current official symbols were updated to their most recent HGNC-approved gene symbols before remapping to ensure maximum gene coverage.

Gene Ontology (GO) enrichment analysis was performed independently for each DEG direction across three ontology categories: Biological Process (BP), Molecular Function (MF), and Cellular Component (CC). Similarly, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis^24,25 was conducted separately for upregulated and downregulated DEG sets to identify significant metabolic and signalling pathways associated with each expression direction. For both GO and KEGG analyses, the Benjamini-Hochberg method was applied to correct for multiple testing, and a false discovery rate (FDR) adjusted p-value threshold of < 0.05 was used to define statistically significant enrichment. For KEGG analysis, the organism code “hsa” (Homo sapiens) was specified, and gene symbols were converted from Entrez IDs to official gene symbols using the setReadable () function to improve biological interpretability. For visualization, the top 15 most significantly enriched terms were displayed for GO BP and CC categories. In contrast, all significant terms were displayed for GO MF and KEGG categories, given the smaller number of significant hits. Dot plots were generated using ggplot2, where dot size represents the number of enriched genes (Gene Count) and color intensity reflects the -Log₁₀ transformed adjusted p-value. All figures were assembled using the cowplot package in R.

Protein-protein interaction (PPI) network construction

PPI networks were constructed separately for upregulated and downregulated gene sets using the STRING database (https://string-db.org) ²⁶ with a minimum interaction confidence score of 0.4. Unconnected nodes were removed to retain only genes with at least one significant interaction. The resulting networks were imported into Cytoscape v3.10.4 for visualization and topological analysis.

To identify key regulatory genes (hub genes) within each network, three centrality-based algorithms, Maximal Clique Centrality (MCC), Degree, and Closeness, were applied independently using the CytoHubba plugin in Cytoscape. Each algorithm ranked genes based on their topological importance within the network: MCC identifies nodes belonging to densely connected subnetworks, Degree identifies the most highly connected nodes, and Closeness identifies genes most central to information flow across the network. The top 20-ranked genes from each algorithm were identified for both the upregulated and downregulated networks separately. Hub genes were defined as those appearing in the top 20 rankings of all three centrality measures simultaneously. The intersection of the three top 20 gene lists was determined using the Merge Networks function in Cytoscape with the Intersection operation, yielding a final hub gene panel for each expression direction. This multi-algorithm approach was adopted to reduce bias associated with any single centrality measure and to ensure the robustness and reproducibility of hub gene selection.

Gene expression validation and clinicopathological analysis

Principal component analysis (PCA) and hierarchical clustering

PCA and hierarchical clustering were performed to evaluate the ability of the identified hub genes to discriminate between OC and normal tissue samples. Expression data for the hub genes were extracted from all four GEO datasets and combined into a single integrated expression matrix comprising 331 samples, of which 293 were tumor samples, and 38 were normal samples. Before analysis, expression values were log2-transformed where necessary, and the highest-expressing probe per gene was retained to ensure one-to-one gene-to-probe mapping. Only genes present across all four datasets were included in the final analysis. PCA was performed using the FactoMineR package in R, ³² with expression values standardized by unit variance scaling before analysis. Three PCA scenarios were evaluated: all 35 hub genes combined, the 17 upregulated hub genes only, and the 18 downregulated hub genes only. For each scenario, principal component 1 (PC1) and principal component 2 (PC2) were plotted against each other with samples colored by group: tumor samples in red and normal samples in grey. 95% confidence ellipses were overlaid for each group to visualize the degree of separation between tumor and normal samples. Again, hierarchical clustering heatmaps were generated using the ComplexHeatmap package ³¹ in R to visualize gene-level expression patterns across all samples. Expression values were z-score scaled per gene before clustering to normalize for differences in expression magnitude across genes and datasets. Separate heatmaps were generated for the upregulated and downregulated hub gene panels. Both rows and columns were clustered using the Ward. D2 agglomeration method with Euclidean distance as the distance metric. Columns were split by sample group, tumor, and normal, to facilitate visual comparison of expression patterns between the two groups. A color scale ranging from blue (low expression) to white (intermediate) to red (high expression) was used to represent z-score normalized expression values.

Receiver operating characteristic (ROC) Analysis

ROC analysis was also performed to evaluate the diagnostic potential of the identified hub genes in discriminating OC samples from normal tissue samples. The analysis was conducted using the pROC package in R ³³ on the integrated expression dataset comprising all four GEO datasets. Sample labels were encoded as a binary response variable with tumor samples assigned a value of 1 and normal samples assigned a value of 0. ROC analysis was performed separately for the 17 upregulated and 18 downregulated hub gene panels. For each hub gene, the area under the ROC curve (AUC) was computed as a measure of overall diagnostic performance, with higher AUC values indicating greater discriminatory ability. The 95% confidence interval for each AUC was calculated using the DeLong method. The optimal classification threshold for each gene was determined using the Youden Index, which maximizes the sum of sensitivity and specificity. Sensitivity and specificity values at the optimal threshold were recorded for each gene. Hub genes with AUC ≥ 0.9 were considered to have excellent diagnostic value, while those with AUC between 0.8 and 0.9 were considered to have good diagnostic value. Individual ROC curves were visualized in a grid layout with AUC, 95% CI, sensitivity, and specificity displayed within each plot. A summary table of all ROC statistics was compiled for all 35 hub genes.

External validation of the diagnostic performance of the hub genes

To further strengthen the robustness of our diagnostic findings and address the potential concern of overfitting inherent in analyses performed on differential expression analysis datasets, the diagnostic performance of the identified hub genes was independently validated using GSE69428, a publicly available gene expression dataset that was not used at any stage of the DEG analysis or primary validation pipeline. GSE69428 comprises laser capture microdissection (LCM) expression profiles of high-grade serous ovarian carcinoma (HGSOC) tissue samples and matched normal oviduct tissue samples generated on the Affymetrix HG-U133 Plus 2.0 platform (GPL570). Twenty samples were selected for external validation, comprising 10 HGSOC tumor samples and 10 normal oviduct tissue samples, while cell culture, immortalized, and xenograft samples present in the dataset were excluded to ensure biological relevance and consistency with the whole tissue design of the discovery cohort. To begin with, probe-level expression data were extracted from the GSE69428 series matrix file using the GEOquery package in R. Probe identifiers were mapped to official gene symbols using the hgu133plus2. db Bioconductor annotation package, which provides complete probe-to-gene mappings for the Affymetrix HG-U133 Plus 2.0 platform. Where multiple probes mapped to the same gene, the probe with the highest mean expression across all samples was retained. Expression values were log2-transformed where necessary before analysis. Again, differential expression of the hub genes between HGSOC tumor and normal oviduct samples was assessed using the Student's t-test, with Benjamini-Hochberg correction applied to control the false discovery rate. Hub genes with adjusted p-value < 0.05 were considered statistically significant. Directional consistency was evaluated by comparing the observed direction of expression change in GSE69428 with the expected direction based on the initial DEG analysis. Finally, ROC analysis was subsequently performed on the GSE69428 cohort using the same methodology applied in the initial ROC analysis. The AUC values obtained from the external validation cohort were compared with those from the initial cohort (datasets) to assess the consistency of diagnostic performance and evaluate the presence of overfitting. Hub genes with AUC ≥ 0.9 were considered to demonstrate excellent diagnostic performance in the external validation cohort.

Dataset characteristics

Based on our inclusion and exclusion criteria, four datasets (GSE14407, GSE18520, GSE26712 & GSE27651) were included in this study. Key information of the datasets, including sample size, platform ID, etc., is summarized in Table 1.

OPEN IN VIEWER

Table 1.

Dataset characteristics.

Dataset ID	Platform	Tumor samples	Normal samples	Total samples	Year	Reference (PMID)
GSE14407	Affymetrix Human Genome U133 Plus 2.0 Array	12	12	24	2019	20040092
GSE18520	Affymetrix Human Genome U133 Plus 2.0 Array	53	10	63	2019	25944803, 18593951
GSE26712	Affymetrix Human Genome U133 Plus 2.0 Array	185	10	195	2019	19962670
GSE27651	Affymetrix Human Genome U133 Plus 2.0 Array	43	6	49	2019	21451362

Identification of DEGs

In GSE14407, a total of 3,599 DEGs were identified, comprising 2,231 upregulated and 1,368 downregulated genes. In GSE18520, 2,366 DEGs were identified, of which 1,437 were upregulated, and 929 were downregulated. GSE26712 yielded 1,037 DEGs, with 359 upregulated and 678 downregulated genes. Again, GSE27651, 3,726 DEGs were identified, comprising 2,244 upregulated and 1,482 downregulated genes. The distribution of upregulated and downregulated DEGs across all four datasets is illustrated in the volcano plots presented in Figure 2.OPEN IN VIEWER

More importantly, to identify genes consistently dysregulated across all four independent datasets, overlapping DEGs were determined separately for upregulated and downregulated gene sets using Venn diagram analysis. A total of 96 genes were found to be commonly upregulated, and 113 genes were commonly downregulated across all four datasets, yielding a final panel of 209 overlapping DEGs for downstream functional enrichment and network analyses (Figure 2).

Functional enrichment analysis of overlapping genes

The GO enrichment analysis of the upregulated genes identified 240 significant BP terms, 33 CC terms, and 5 MF terms (adjusted p < 0.05). The top 15 most significantly enriched terms for BP and CC, and all significant MF terms, are presented in Figure 3. In the BP category, the most significantly enriched terms were predominantly related to chromosome segregation and cell cycle regulation, including chromosome segregation, nuclear chromosome segregation, nuclear division, and sister chromatid segregation (Figure 3(a)). Additional enriched terms included mitotic nuclear division, mitotic sister chromatid segregation, and metaphase/anaphase transition of the cell cycle, collectively pointing to dysregulation of mitotic processes as a hallmark of the upregulated transcriptomic signature in OC. In the CC category, the upregulated DEGs were predominantly enriched in components of the mitotic spindle apparatus and chromosomal structural elements, including the spindle, chromosomal region, condensed chromosome, chromosome centromeric region, midbody, and kinetochore (Figure 3(b)). The enrichment of genes within these cellular compartments is consistent with the BP findings and further supports active mitotic dysregulation in ovarian tumor cells. In the MF category, the most prominent enriched terms were microtubule binding, cyclin-dependent protein kinase regulator activity, and cyclin-dependent protein serine/threonine kinase regulator activity (Figure 3(c)). These molecular functions are directly implicated in cell cycle progression and mitotic spindle assembly, providing mechanistic support for the observed BP and CC enrichments.OPEN IN VIEWER

For the downregulated genes, the most significantly enriched term was hormone metabolic process, which showed the highest gene count and strongest statistical significance among all downregulated BP terms in the BP category (Figure 4(a)). Other highly enriched terms included connective tissue development, respiratory system development, and lung development. Notably, multiple terms related to BMP signalling were enriched, including cellular response to BMP stimulus, response to BMP, regulation of BMP signalling pathway, and negative regulation of BMP signalling pathway. Additionally, enrichment of steroid biosynthetic process and multiple metabolic terms, including isoprenoid, terpenoid, diterpenoid, retinoid, and retinol metabolic processes, highlights significant suppression of steroid hormone biosynthesis and lipid metabolism, which are known to play protective roles in normal ovarian tissue physiology. ³⁴ Also, in the CC category, only 2 significant terms were identified (Figure 4(b)). Both terms had a notably high gene count of 14 (Figure 4(c)). Furthermore, in the MF category, the most prominently enriched terms were glycosaminoglycan binding, heparin binding, and sulfur compound binding, followed by extracellular matrix structural constituent, alcohol dehydrogenase (NAD+) activity, and all-trans-retinol dehydrogenase (NAD+) activity (Figure 4(c)).OPEN IN VIEWER

Moving on, KEGG pathway enrichment analysis identified 7 significant pathways among the upregulated DEGs and 4 significant pathways among the downregulated DEGs (adjusted p < 0.05) (Figure 5). For the upregulated DEGs, the most significantly enriched pathway was the cell cycle pathway, which exhibited the highest gene count (n = 11) and the strongest statistical significance (−Log₁₀ adjusted p > 6) among all identified pathways (Figure 5(a)). Additional significantly enriched pathways included Human T-cell leukemia virus 1 infection and oocyte meiosis, both of which share mechanistic overlap with cell cycle regulatory genes and reflect the hijacking of proliferative machinery in tumor cells. The biosynthesis of amino acids and the p53 signalling pathway were also significantly enriched, with the p53 pathway being particularly noteworthy given its central role as a tumor suppressor. Furthermore, glycine, serine, and threonine metabolism and the one-carbon pool by folate pathways were enriched, which could reflect active metabolic reprogramming to support rapid tumor cell proliferation. Among the downregulated DEGs, the most significantly enriched pathway was retinol metabolism, which had the highest gene count and the strongest statistical significance among the downregulated pathways (Figure 5(b)). Drug metabolism via cytochrome P450 was the second most enriched pathway, reflecting suppression of xenobiotic-metabolizing enzymes. Tyrosine and histidine metabolism were also significantly enriched among the downregulated DEGs, consistent with the broader suppression of amino acid catabolic pathways observed in the GO enrichment analysis.OPEN IN VIEWER

PPI network and hub gene identification

The STRING network of all the 96 upregulated and 113 downregulated DEGs is shown in Figure 6. For the upregulated DEG network, three centrality measures, MCC, Degree, and Closeness, were applied independently using the CytoHubba plugin, with the top 20 genes identified from each method (Figure 7). The MCC analysis identified genes predominantly associated with spindle assembly checkpoint regulation and mitotic progression, with CDK1, UBE2C, MAD2L1, CCNB1, TTK, CCNB2, KIF2C, BUB1B, TRIP13, MCM4, CENPF, and TOP2A among the highest ranked (Figure 7(a)). The Degree analysis identified CDK1 as the most highly connected node, followed by UBE2C, CCNB2, BUB1B, CCNB1, TOP2A, MCM4, TRIP13, KIF2C and TTK (Figure 7(b)). The Closeness analysis similarly ranked CDK1, BUB1B, UBE2C, TOP2A, CCNB1, and MCM4 among the top genes (Figure 7(c)). The intersection of all three methods yielded a final panel of 17 upregulated hub genes: BUB1B, CCNB1, CCNB2, CDK1, CENPF, KIF2C, KIF20A, MAD2L1, MCM4, NEK2, PBK, RACGAP1, TOP2A, TRIP13, TTK, UBE2C, and ZWINT (Figure 8(a)). These genes form a densely interconnected subnetwork, reflecting their functional cohesion within cell cycle regulatory pathways.OPEN IN VIEWER

OPEN IN VIEWER

Figure 7.

Hub gene identification from PPI networks of upregulated and downregulated DEGs using CytoHubba centrality measures. Hub genes identified from the upregulated (a–c) and downregulated (d–f) PPI networks using three centrality-based algorithms in cytoHubba: Maximal Clique Centrality MCC (a, d), degree (b, e), and closeness (c, f). The top 20-ranked genes from each method are highlighted. Node color intensity reflects the centrality score, with darker red indicating a higher ranking.

OPEN IN VIEWER

Figure 8.

Merged hub gene networks derived from the intersection of MCC, degree, and closeness centrality measures for upregulated and downregulated DEGs in OC. Merged PPI subnetworks showing hub genes identified consistently across all three CytoHubba centrality measures for (a) upregulated DEGs (17 hub genes, shown in green) and (b) downregulated DEGs (18 hub genes, shown in blue).

For the downregulated DEG network, the same three centrality measures were applied (Figure 7). The MCC analysis ranked ALDH1A1, AOX1, ADH1B, ALDH1A3, ADH1C, MAOB, and MAOA as the highest-scoring genes, followed by CAV1, KDR, PDGFRA, WNT5A, SNAI2, LYVE1, PDPN, CDH11, FGF13, GATA6, HSD17B6, NPY1R, and ZFPM2 (Figure 7(d)). The Degree analysis identified PDGFRA and KDR as the most highly connected nodes, with ALDH1A1, SNAI2, WNT5A, ADH1B, MAOA, AOX1, CAV1, and MAOB also among the top-ranked genes (Figure 7(e)). The Closeness analysis similarly identified SNAI2, PDGFRA, KDR, CAV1, WNT5A, ALDH1A1, ADH1B, FGF13, GATA6 and LYVE1 as the most centrally positioned genes (Figure 7(f)). The intersection of all three methods yielded a final panel of 18 downregulated hub genes: ADH1B, ADH1C, ALDH1A1, ALDH1A3, AOX1, CAV1, CDH11, FGF13, GATA6, KDR, LYVE1, MAOA, MAOB, PDGFRA, PDPN, SNAI2, WNT5A, and ZFPM2 (Figure 8(b)). The downregulated hub gene network comprised two distinct functional clusters: a metabolic enzyme cluster centered on aldehyde dehydrogenases and monoamine oxidases, and a signalling and ECM cluster centered on growth factor receptors, EMT regulators, and lymphangiogenic markers.

Gene expression validation

All eight selected upregulated hub genes, BUB1B, KIF2C, MAD2L1, MCM4, NEK2, TOP2A, TRIP13, and UBE2C, demonstrated significantly higher expression in OC samples compared to normal tissue (p < 0.01) (Figure 9). Among these, UBE2C showed the most pronounced upregulation with the greatest expression difference between tumor and normal samples, followed by MCM4, NEK2, and TOP2A. The consistent upregulation of these cell cycle regulatory genes across both the GEO microarray datasets and the TCGA RNA-seq cohort strongly supports their biological relevance in OC pathogenesis.OPEN IN VIEWER

Among the eight selected downregulated hub genes, five, ADH1B, ALDH1A1, GATA6, LYVE1, and MAOA, showed significantly lower expression in OC samples compared to normal tissue (p < 0.01). ADH1B and MAOA demonstrated the most pronounced downregulation with near-complete loss of expression in tumor samples relative to normal tissue. Although KDR , FGF13, and WNT5A showed a consistent trend of lower expression in tumor samples, these differences did not reach statistical significance in the GEPIA analysis (p ≥ 0.01) (Figure 10).OPEN IN VIEWER

OS and DFS analysis

Kaplan-Meier survival analysis was performed to assess the prognostic significance of the 35 hub genes for both survival outcomes (OS and DFS). Among the 17 upregulated hub genes, only TRIP13 demonstrated a statistically significant association with OS (HR = 0.72, p = 0.0079), with high TRIP13 expression correlating with improved OS in OC patients (Figure 11(a)). None of the upregulated hub genes showed significant associations with DFS. Among the 18 downregulated hub genes, two genes, FGF13 and LYVE1, were significantly associated with OS. High FGF13 expression was associated with favourable OS (HR = 0.68, p = 0.0022), while high LYVE1 expression was associated with poor OS (HR = 1.4, p = 0.0076) (Figure 11(b)). For DFS, three downregulated hub genes demonstrated significant prognostic associations (Figure 11(c)). High ALDH1A1 expression was associated with improved DFS (HR = 0.78, p = 0.044), while high GATA6 expression was associated with shorter DFS intervals (HR = 1.3, p = 0.035). Additionally, high WNT5A expression was associated with improved DFS (HR = 0.77, p = 0.04).OPEN IN VIEWER

Association of hub genes with pathological stage, race, age, tumor grade, and TP53 mutation status

ROC analysis

All 18 downregulated hub genes demonstrated excellent diagnostic performance with AUC values ranging from 0.859 to 0.964 (Figure 15). The top five performing genes were KDR (AUC = 0.964, 95% CI: 0.941-0.987), FGF13 (AUC = 0.963, 95% CI: 0.944-0.983), ADH1B (AUC = 0.962, 95% CI: 0.938-0.986), MAOA (AUC = 0.960, 95% CI: 0.939-0.981) and ADH1C (AUC = 0.953, 95% CI: 0.923-0.983). Notably, MAOA and PDPN achieved perfect specificity (specificity = 1.000) at their optimal Youden Index thresholds. The remaining downregulated hub genes, ALDH1A1 (AUC = 0.952), GATA6 (AUC = 0.930), LYVE1 (AUC = 0.928), CAV1 (AUC = 0.926), PDPN (AUC = 0.926), AOX1 (AUC = 0.943), ZFPM2 (AUC = 0.941), WNT5A (AUC = 0.910), MAOB (AUC = 0.905), SNAI2 (AUC = 0.906), ALDH1A3 (AUC = 0.902), PDGFRA (AUC = 0.914) and CDH11 (AUC = 0.859), all exceeded the AUC threshold of 0.85. The mean AUC across all 18 downregulated hub genes was 0.934.OPEN IN VIEWER

The upregulated hub genes demonstrated more heterogeneous diagnostic performance compared to the downregulated panel. UBE2C emerged as the strongest diagnostic biomarker among the upregulated hub genes with an AUC of 0.933 (95% CI: 0.903-0.962), sensitivity of 0.802, and specificity of 0.974, values comparable to the top-performing downregulated hub genes. NEK2 (AUC = 0.805, 95% CI: 0.732–0.879) and TRIP13 (AUC = 0.766, 95% CI: 0.677–0.855) also demonstrated good diagnostic performance. The remaining upregulated hub genes showed moderate to poor diagnostic performance with AUC values ranging from 0.422 to 0.710. Notably, CDK1 (AUC = 0.422), MCM4 (AUC = 0.450), and PBK (AUC = 0.451) showed AUC values below 0.5. The mean AUC across all 17 upregulated hub genes was 0.630. The combined ROC results are summarized in Figure 16.OPEN IN VIEWER

External validation of the diagnostic performance of the hub genes

The external validation ROC analysis demonstrated outstanding diagnostic performance for the upregulated hub gene panel in the independent external cohort (GSE69428) (Figure 17). Seven genes (CCNB2, KIF2C, KIF20A, TOP2A, TTK, UBE2C, and ZWINT) achieved perfect AUC values of 1.0 (95% CI: 1-1), indicating complete discrimination between HGSOC tumor and normal oviduct samples. Additionally, BUB1B (AUC = 0.99, 95% CI: 0.962-1), CCNB1 (AUC = 0.98, 95% CI: 0.933-1), RACGAP1 (AUC = 0.98, 95% CI: 0.933-1), PBK (AUC = 0.96, 95% CI: 0.875-1), MCM4 (AUC = 0.93, 95% CI: 0.82-1), CDK1 (AUC = 0.94, 95% CI: 0.841-1) and NEK2 (AUC = 0.91, 95% CI: 0.764-1) also demonstrated excellent diagnostic performance. CENPF (AUC = 0.76, 95% CI: 0.529-0.991) and MAD2L1 (AUC = 0.78, 95% CI: 0.564-0.996) showed good diagnostic performance, while TRIP13 demonstrated moderate performance (AUC = 0.69, 95% CI: 0.441-0.939). The mean AUC across all 17 upregulated hub genes in the external validation dataset was 0.938, a substantial improvement compared to the mean AUC of 0.630 observed in the primary ROC analysis.OPEN IN VIEWER

Among the downregulated hub genes, ALDH1A3 achieved a perfect AUC of 1.0 (95% CI: 1-1) in the external validation dataset, while ALDH1A1 (AUC = 0.99, 95% CI: 0.962-1), ADH1B (AUC = 0.95, 95% CI: 0.86-1), GATA6 (AUC = 0.92, 95% CI: 0.761-1) and ZFPM2 (AUC = 0.86, 95% CI: 0.681-1) demonstrated excellent diagnostic performance. MAOA (AUC = 0.82, 95% CI: 0.612-1), PDPN (AUC = 0.81, 95% CI: 0.612-1), SNAI2 (AUC = 0.75, 95% CI: 0.515-0.985), and ADH1C (AUC = 0.745, 95% CI: 0.523-0.967) showed good diagnostic performance. The remaining downregulated hub genes demonstrated moderate to poor diagnostic performance in the external validation cohort (Figure 18). The mean AUC across all 18 downregulated hub genes was 0.742.OPEN IN VIEWER