Large-scale transcriptome analysis reveals KIF20A as a novel indicator of cisplatin resistance in high-grade serous ovarian cancer

Abstract

Background

Accurately assessing chemosensitivity to drugs such as cisplatin is crucial for optimizing outcomes of high-grade serous ovarian carcinoma (HGSOC) patients. Developing prediction models and identifying biomarkers can assist differentiate chemotherapy-sensitive from chemotherapy-resistant subtypes.

Methods

We integrated 908 transcriptome samples from 12 HGSOC datasets of GEO and TCGA across four detection platforms. Differentially ranked genes (DRGs) were identified by applying intra-sample rank transformation and performing T-test. To refine the selection of informative DRGs, a recurrent logistic regression (RLR) model was employed, followed by model optimization using the Pairwise Analysis of Gene Expression (PAGE) algorithm to establish clinically applicable gene pair signature.

Results

Kinesin Family Member 20 A (KIF20A) and Chromogranin B (CHGB) were paired to establish a robust gene pair signature, KIF20A::CHGB Pair (KCP), which exhibited remarkable discriminant accuracy (Accuracy = 0.9397; AUROC = 0.9668; AUPRC = 0.9799) in distinguishing HGSOC from controls in eight independent validation cohorts. Further survival and functional analyses revealed that altered rank relationships between KIF20A and CHGB were significantly associated with prognosis and platinum resistance in HGSOC. Following exposure to standard chemotherapy regimens, we conducted RT-qPCR, CCK-8(Cell Counting Kit-8) viability assays, EdU proliferation assays, flow cytometry, and SynergyFinder, which collectively assessed cellular proliferation and cell-cycle responses, demonstrating that both genes contribute comparably to the malignant phenotype.

Conclusions

Elevated KIF20A or CHGB expression drives chemoresistance and correlates with poor prognosis in HGSOC. Silencing either gene significantly enhances the growth‑inhibitory effects of cisplatin and other standard chemotherapeutic agents. Our findings establish KCP as a robust diagnostic and prognostic biomarker and highlight KIF20A and CHGB as synergistic targets for overcoming platinum resistance in ovarian cancer.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13048-026-02173-z.

Key points

• In silico experiments identified KIF20A::CHGB (KCP) as a cross-platform gene pair signature for distinguishing high-grade serous ovarian carcinoma (HGSOC) from normal tissues, achieving exceptional accuracy (AUROC = 0.9715) across eight independent cohorts.

• In vitro experiments demonstrated that KIF20A and CHGB knockdown enhanced the sensitivity of ovarian cancer cells to cisplatin and suppressed tumor cell proliferation.

• We developed a novel Recurrent Logistic Regression (RLR) model combined with Pairwise Analysis of Gene Expression (PAGE) to overcome batch effects and improve biomarker generalizability, offering a framework for multi-cohort transcriptomic studies.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13048-026-02173-z.

Introduction

High-grade serous ovarian carcinoma (HGSOC) is the most prevalent histological subtype of ovarian cancer, constituting approximately 74% of epithelial ovarian cancers [1].The majority of HGSOC patients are diagnosed at advanced stages (FIGO stage III or IV), by which time extensive intraperitoneal dissemination of malignant cells has typically occurred. In this situation, cytoreductive surgery combined with platinum-based chemotherapy remains the cornerstone of treatment [2]. Despite therapeutic advances, most individuals with advanced-stage HGSOC still face disease recurrence, resulting in low cure rates and dismal 5-year survival rate below 40% [3]. Two major challenges in HGSOC management are early diagnosis of ovarian cancer and the mitigation of chemoresistance, which remain urgent clinical priorities. Therefore, the early identification of reliable biomarkers is critical for guiding risk stratification and therapeutic decision-making in HGSOC patients.

With the rapid development of high-throughput sequencing technologies and bioinformatics algorithms [4–6], numerous studies sought to identify potential biomarkers for HGSOC recently [7, 8]. However, existing methodologies face two major limitations: First, most biomarkers are derived from absolute gene expression levels and have only been validated in only a limited number of datasets, lacking large-scale, multi-data evaluation [9, 10]. Second, inconsistencies in data normalization methods across different datasets introduce pronounced batch effects, which undermine the generalizability of these prediction models.

Recent advances in the diagnosis of HGSOC have been facilitated by the integration of liquid biopsy, multi-omics profiling, and advanced pattern recognition techniques in medical imaging and histopathology [1].However, effective long-term treatment remains a major clinical challenge. Current first-line chemotherapeutic regimens, including cisplatin plus paclitaxel or carboplatin plus paclitaxel, have demonstrated clinical efficacy [11]. However, primary resistance occurs in 10–20% of patients at treatment onset, and over 70% eventually relapse, most with acquired chemoresistance [12]. Once resistance develops, median overall survival declines sharply to around 12 months [13, 14].Therefore, accurately assess chemosensitivity to drugs such as cisplatin and paclitaxel is crucial for guiding personalized therapeutic strategies.

Platinum resistance is a major cause of treatment failure in patients with advanced or recurrent HGSOC. Developing prediction models and identifying biomarkers can assist differentiate chemotherapy-sensitive from chemotherapy-resistant subtypes. In this study, we collected a total of 908 transcriptomic samples from 12 independent datasets. After detecting Differentially Ranked Genes (DRGs) and modeling via Reverse recursive Logistic Regression (RLR), we identified gene signatures that differentiate HGSOC and normal samples. These biomarkers were further optimized using the Pairwise Analysis of Gene Expression (PAGE) algorithm and validated across eight independent cohorts to assess their classification performance [15, 16]. Function and network analysis showed one marker, Kinesin Family Member 20 A (KIF20A), is associated with Cisplatin resistance. To assess the functional roles of KIF20A::CHGB Pair (KCP), we conducted a series of in vitro experiments with clinically relevant chemotherapeutic agents.

Materials and methods

Data collection and processing

We collected 12 independent gene expression cohorts from public gene expression repositories, including the Gene Expression Omnibus (GEO), The Cancer Genome Atlas (TCGA), and Li et al., comprising 817 HGSOC samples and 91 healthy controls (Table 1). Among these, the three largest datasets—GSE6008, GSE18520, and TCGA (Discovery Set I)—were integrated for model training. GSE27651 (Discovery Set II) was used for parameter tuning. The remaining eight datasets, including GSE14407, GSE36668, GSE40595, GSE54388, GSE69428, Li et al., GSE38666, and GSE23391, served as external validation cohorts. In total, we analyzed 908 transcriptomic profiles generated from four distinct gene expression platforms: the Affymetrix HG-U133 Plus 2.0 Array, Affymetrix HG-U133A Array, the Affymetrix HG-U133A Array, and the Agilent-Arraystar. A total of 9,575 genes were commonly profiled across these platforms, and all subsequent analyses were conducted based on this shared gene set.

Table 1.

Datasets used in this study

Cohorts	Dataset	Platform	OC	NC	Total
GSE6008	Discovery set Ⅰ	Affymetrix HG-U133A Array	41	4	45
GSE18520		Affymetrix HG-U133 Plus 2.0 Array	53	10	63
TGCA		Affymetrix HG-U133A Array	585	8	593
GSE27651	Discovery set Ⅱ	Affymetrix HG-U133 Plus 2.0 Array	35	6	41
GSE14407	External validation	Affymetrix HG-U133 Plus 2.0 Array	12	12	24
GSE36668		Affymetrix HG-U133 Plus 2.0 Array	4	4	8
GSE40595		Affymetrix HG-U133 Plus 2.0 Array	32	6	38
GSE54388		Affymetrix HG-U133 Plus 2.0 Array	16	6	22
GSE69428		Affymetrix HG-U133 Plus 2.0 Array	10	10	20
Li et al.		Agilent-Arraystar	8	8	16
GSE38666		Affymetrix HG-U133 Plus 2.0 Array	18	12	30
GSE23391		Affymetrix HG-U133 Plus 2.0 Array	3	5	8

Identification of differentially ranked gene

We applied the relative ranking method to normalize gene expression profiles across all cohorts [17]. This approach focuses on the within-sample relative rank of gene expression, disregarding absolute expression abundance and requiring no external reference [18–20]. Differential rank analysis was performed by comparing the OC and NC groups within each dataset of Discovery Set I (GSE6008, GSE18520, and TCGA). P-values were calculated using independent t-tests, with a significance threshold set at P < 0.05. Genes were defined as up-regulated if the change in their average rank exceeded 10% of the global rank distribution; conversely, genes were considered down-regulated if the change in their average rank decreased by 10% of the global rank distribution.

Functional enrichment analysis

Based on the differentially ranked genes (DRGs) identified in HGSOC, we performed Gene Ontology (GO) analysis using the ‘clusterProfiler’ package in the R environment [21]. The analysis was conducted separately for up-regulated and down-regulated genes across three GO categories: Cellular Component (CC), Biological Process (BP), and Molecular Function (MF). Significantly enriched GO terms were identified using a Benjamini–Hochberg adjusted P-value threshold of 0.05 (adjusted P < 0.05). Beyond GO enrichment analysis, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis was conducted based on the KEGG database to explore the biological pathways that are mainly associated with the differentially expressed genes. Initially, we enumerated the number of DRGs involved in each pathway. Subsequently, the hypergeometric test was employed to evaluate the degree of enrichment of DRGs in individual signaling pathways, using the entire genome as the background. This approach enabled the identification of biological pathways significantly enriched among the DRGs. Pathways with a P-value less than 0.05 were considered statistically significant, suggesting a potential biological relevance of the enriched signaling pathways in the context of HGSOC [22].

Recurrent logistic regression

Recurrent logistic regression contains multiple iterations with model optimization and automatic feature selection, since each iteration involves regression step and elimination step.

Regression step

Logistic regression is applied to the candidate gene set (initially 517 DRGs). The expression abundance of genes in each sample, is represented by a vector denoted as.

where is the i-th gene expression.

To construct a classifier involving as fewer genes as possible based on the expression vector of a sample, a function is built, where  0 represents healthy controls and 1 represents ovarian cancer. The logistic model applied in RLR is a binary classifier expressed by

where is an expression vector and is a diagnostic risk score used to predict the probability of having ovarian cancer. are parameters optimized by the cost function with regularization,

where  X is the collection of samples in discovery set and is the label for each sample.

Elimination step

After optimizing the regression model, the minor genes regarded as less significant are eliminated. A gene is defined as minor gene if the absolute value of its corresponding weight is less than a proportion of the absolute maximum weight, i.e.,

where is a hyperparameter. Instead of using the traditional way that chooses a fixed threshold such as P value < 0.01, this step selects features adaptively based on the maximum weight trained by the model.

The regression step and elimination step are repeated iteratively until it converges, specifically, no more minor gene remained. In this sense, the algorithm is named the recurrent logistic regression (RLR).

The RLR is first trained on the discovery setⅠand evaluated by the Performance Curve on discovery setⅡ. We exhaustedly tried possible values of the hyperparameter  C with the search space between 0.80 and 0.90 and each interval equaling to 0.001. The hyperparameter  C can therefore be determined by the optimal performance on discovery set Ⅱ.

To obtain the optimal parameters, we computed the ratio of AUROC to the number of genes and get the :

2.5 Model optimization

To enhance model reliability and attenuate batch effects introduced by temporal variation, operator differences, reagent lots, and instrumentation discrepancies, we employed the PAGE algorithm into the modeling pipeline. This adjustment effectively improves model robustness and supports generalizability across heterogeneous experimental conditions. The feature richness of each sample is defined as follows:

Let denote the gene expression ranking, a value of 1 is assigned if  Gi>Gj, and 0 otherwise.

Biomarker performance evaluation

Accuracy is a metric used to evaluate the overall performance of a classification model. It represents the proportion of correctly predicted instances to the total number of instances. In binary classification tasks, accuracy can be calculated using the following formula:

In the formula, TP denotes True Positives, TN denotes True Negatives, FP represents False Positives, and FN represents False Negatives.

We employed Risk Score to quantify the degree of differential variation between gene pairs:

We predicted the class of each sample as positive (Risk Score > 0) or negative (Risk Score ≤ 0) based on a threshold of zero for the risk score and calculated the classification accuracy accordingly. To visualize the distribution of risk scores, we used the ‘ggplot2’ package in the R environment to create scatter plots. To minimize data point overlap and enhance clarity, we specifically applied the ‘geom_jitter’ function.

The Receiver Operating Characteristic (ROC) curve is a widely used tool for evaluating the performance of classification models. It illustrates the trade-off between the true positive rate (TPR) and the false positive rate (FPR). A higher area under the curve (AUC) indicates a stronger ability of the model to discriminate between positive and negative samples.

The Precision-Recall (PR) curve is another important evaluation metric, especially for imbalanced datasets. It presents the relationship between precision and recall. A higher area under the PR curve (AUPRC) suggests that the model can maintain high precision while achieving a high recall.

We assessed the diagnostic performance of the model across eight independent validation sets using three metrics: accuracy, AUROC, and AURPC.

Interaction network construction

Protein–protein interaction (PPI) networks represent a network in which genes or proteins serve as nodes, and the edges denote their interactions—such as co-participation in metabolic pathways, biological processes, structural complexes, functional associations, or physical binding [23–25]. Understanding PPI networks is crucial for elucidating cellular network architecture and functions, as well as the molecular mechanisms underlying disease development and progression [26, 27]. To investigate the interaction network of HGSOC biomarkers and the KEGG pathways they involved, we searched for KIF20A-related genes in the STRING database (https://string-db.org), and genes with a confidence score ≥ 0.15 were selected. We then calculated the correlation between these genes and the 9,572 common genes from the four platform and genes with a Spearman correlation coefficient |ρ| ≥ 0.4 and p < 0.001 were retained. Based on the interaction network constructed by these genes, we subsequently performed KEGG pathway enrichment analysis to explore the potential functional mechanisms.

Survival analysis

Survival analysis was performed on the TCGA and GSE18520 cohort using the R package survival. Patients were stratified into high-risk and low-risk groups based on the optimal threshold of biomarker expression levels, which employed the ‘survminer’ package in the R environment. Kaplan–Meier survival curves were generated to visualize the differences in 5-year survival between the high-risk and low-risk groups, and statistical significance was assessed using the log-rank test [19].

Cell culture and treatment

The human ovarian cancer cell lines SKOV3 and OVCAR3 were used in this study. Cells were cultured in RPMI 1640 medium (Gibco, Grand Island, NY, USA) supplemented with 10% fetal bovine serum (FBS, Gibco Life Technologies, NY, USA) and 1% penicillin-streptomycin (Gibco, Grand Island, NY, USA) in a 37 °C, 5% CO2 incubator. The experimental groups included: NC (negative control), siKIF20A knockdown group, chemotherapy drug treatment group, and siKIF20A knockdown combined with chemotherapy treatment group, which were used for subsequent experiments.

siRNA transfection

SKOV3 and OVCAR3 cells were seeded at a density of 2.5 × 10^5 cells/well in 6-well plates and cultured for 24 h before transfection. According to the manufacturer’s instructions, 1 pmol of siNC and siKIF20A were transfected into the cells using Lipofectamine RNAiMAX transfection reagent (Thermo Fisher Scientific). Three different siKIF20A sequences were used:

1. siKIF20A-1: sense 5’-GUUCCUGCAUGAUUGUCAATT-3’, antisense 5’-UUGACAAUCAUGCAGGAACGG-3’;

2. siKIF20A-2: sense 5’-GAUGACGAUGUCGUAGUUUTT-3’, antisense 5’-AAACUACGACAUCGUCAUCGG-3’;

3. siKIF20A-3: sense 5’-UGAAGAAAUAGGUCUCUUUTT-3’, antisense 5’-AAAGAGACCUAUUUCUUCAGG-3’ (Sequences were synthesized by Suzhou GenePharma Biotechnology Co., Ltd.).

Quantitative Real-Time PCR (qRT-PCR)

Total RNA was extracted from the cultured cells using the Total RNA Isolation Kit (Foregene, Chengdu, China). The RNA was then reverse-transcribed into cDNA using the PrimeScript RT reagent Kit (TaKaRa, Shiga, Japan) according to the manufacturer’s instructions. The cDNA was used as a template for qRT-PCR to measure the expression level of KIF20A. SYBR Green PCR Master Mix (Selleck Chemicals, Texas, USA) was used for the qRT-PCR experiments. GAPDH was used as the internal control, with the forward primer 5’-GGAGCGAGATCCCTCCAAAAT-3’ and the reverse primer 5’-GGCTGTTGTCATACTTCTCATGG-3’. The relative gene expression levels were calculated using the 2-ΔΔCT method. All primers were synthesized by Shanghai Shenggong Biotech Co., Ltd. (Shanghai, China). The forward primer for KIF20A was 5’-TGCTGTCCGATGACGACGATGTC-3’, and the reverse primer was 5’-AGGTTCTTGCGTACCACAGAC-3’. The mRNA expression levels of the experimental groups and control group were plotted using GraphPad Prism 10 software.

CCK-8 assay

Cell viability was assessed using the Cell Counting Kit-8 (CCK-8, Beyotime Institute of Biotechnology, China). SKOV3 and OVCAR3 cells were seeded at 5000 cells/well in 96-well plates and cultured for 24 h before transfection and drug treatment. The concentration ranges for the drug treatment groups were: cisplatin (0–24 µM), paclitaxel (0-0.068 µM), gemcitabine (0-0.2 µM), docetaxel (0-0.032 µM), and olaparib (0–40 µM). After 48 h of drug treatment, the old medium was discarded, and CCK-8 solution was mixed with serum-free medium (1:9) and added to each well. The cells were incubated for 1–2 h. The absorbance at 450 nm was measured using an automated microplate reader (Bio-Rad Laboratories Inc., Hercules, CA). Cell viability was calculated using the following formula: Cell viability (%) = O.D. treatment / O.D. control × 100%. Cell viability curves were plotted, and the half-maximal inhibitory concentration (IC₅₀) was calculated using GraphPad Prism 10. The synergy between gene knockdown and drug sensitivity was evaluated using the SynergyFinder tool, where a score greater than 10 indicates potential synergy, and a score less than − 10 suggests antagonism (https://synergyfinder.org).

EdU assay

The EdU-594 assay kit (Beyotime Biotechnology, Shanghai, China) was used to evaluate cell proliferation. After treatment, the culture medium was replaced, and 100 µL of 20 µM EdU was added to each well. After incubation at 37 °C for 2 h to label proliferating cells the cells were fixed with 4% paraformaldehyde and permeabilized with 0.3% Triton X-100. After washing with PBS containing 3% BSA, the cells were stained with Azide 594 for 30 min and Hoechst 33,342 for 10 min. Fluorescence microscopy (Cytation 5, Bio-Tek, USA) was used to capture images. The cell proliferation rate was determined by calculating the ratio of EdU-positive nuclei (red) to Hoechst 33,342-stained nuclei (blue), and image analysis was performed using ImageJ Pro 6.0.

Cell cycle analysis

SKOV3 cells were seeded at a density of 4.0 × 10^5 cells/well in 6-well plates, treated with drugs, and then collected for analysis. The cells were fixed overnight at 4 °C with 75% cold ethanol. Cell cycle analysis was performed using the Cell Cycle Detection Kit (Beyotime, Shanghai, China) and stained with propidium iodide (PI). After staining, the cells were filtered through a mesh to remove cell clumps and analyzed using a FACSCalibur flow cytometer (BD Biosciences) equipped with CellQuest Pro software.

Statistical analysis

Statistical analyses were performed using R v4.2 and GraphPad Prism 10 software [28]. The experimental data are presented as the mean ± standard deviation (mean ± SD) of at least three independent experiments. Statistical comparisons were made using unpaired T-test or two-way ANOVA, and P < 0.05 was considered statistically significant.

Results

Workflow of this study

We developed an integrated transcriptomic analysis framework to systematically identify robust gene markers and potential therapeutic targets for HGSOC (Fig. 1). Initially, we integrated 12 publicly available transcriptomic datasets across four distinct gene expression platforms. Differentially ranked genes (DRGs) were identified by performing intra-sample rank transformation and T-test across three discovery cohorts. To obtain more informative DRGs, a recurrent logistic regression (RLR) model was employed, followed by model optimization using the Pairwise Analysis of Gene Expression (PAGE) algorithm to establish a clinically applicable gene pair signature, Kinesin Family Member 20 A (KIF20A) and Chromogranin B (CHGB). The developed gene pair signature, KIF20A::CHGB (KCP), demonstrated robust prediction performance in eight external HGSOC cohorts, showing consistent performance across diverse clinical settings. Functional annotation and protein-protein interaction analysis further revealed that KIF20A is primarily involved in platinum resistance related pathways. Survival analysis and in vitro experiments validated KIF20A as a potential marker for platinum resistance and a promising chemosensitization target in ovarian cancer cells.

Fig. 1.

Workflow diagram.①Identification of intersecting genes across platforms; ②Transforming gene expression matrices of individual HGSOC samples into rank-based matrices according to intra-sample gene expression levels; ③Performing independent t-tests to extract intersecting up-regulated and down-regulated rank-differential genes from Discovery Set I as candidate genes; ④Merging Discovery Set I to generate the input for model construction; ⑤Constructing the model using recursive logistic regression and optimize model parameters using Discovery Set II; ⑥Validating the model using an external validation set; ⑦Platinum drug resistance assay

Differential expression and functional enrichment analysis

To address platforms heterogeneity, we identified a total of 9,572 common genes through cross-platform integration of four platforms, including: Affymetrix HG-U133 Plus 2.0 Array (GPL570), Affymetrix HG-U133A Array (GPL96), Affymetrix HG-U133A Array (TCGA), and Agilent-Arraystar (Li et al.) (Fig. 2A). After rank transforming, differential rank analysis was performed on the Discovery Set I, comprising GSE18520, GSE6008, and TCGA datasets. 1,327, 1,735, and 987 up-regulated differentially ranked genes (DRGs) were identified in GSE18520, GSE6008, and TCGA, respectively. Also, 1,559, 1,770, and 978 down-regulated DRGs were detected (Fig. 2B-D). In total, we identified 254 up-regulated and 263 down-regulated DRGs among the three datasets (Fig. 2E-F). Notably, KIF20A was consistently up-regulated across all three datasets, with marked rank increases observed in GSE18520 (RankDiff = 4897.160), GSE6008 (RankDiff = 4154.152), and TCGA (RankDiff = 5216.337). Conversely, CHGB exhibited a persistent downregulation pattern, with substantial declines in rankings detected in GSE18520 (RankDiff=–6350.098), GSE6008 (RankDiff=–3853.390), and TCGA (RankDiff=–2429.667).

Fig. 2.

Differential and functional analysis. A Venn diagram showing the detected genes of the four sequencing platforms. B-D Volcano plots of rank-based differential genes in Discovery Set I. E, F Intersection of up-regulated and down-regulated genes in Discovery Set I. G GO enrichment analysis of differentially up-regulated and down-regulated genes. H KEGG enrichment analysis of differentially ranked genes

Enrichment analysis was performed separately for up-regulated and down-regulated DRGs. The top five significant terms from each GO category were enriched and visualized to illustrate functional differences (Fig. 2G). The up-regulated DRGs were significantly enriched in cell proliferation and division-related pathways (such as microtubule motor activity, cytoskeletal motor activity and nuclear division), suggesting that the abnormal activation of these genes in the context of HGSOC may lead to chromosomal missegregation and genomic instability [29]. Conversely, down-regulated DRGs were enriched in TGF-β–related functions, including cellular response to transforming growth factor beta stimulus, response to transforming growth factor beta, and growth factor binding. While TGF-β is considered a cytokine that inhibits tumor proliferation, the observed pathway suppression implies that TGF-β may be inactivated by genetic mutations and epigenetic modifications, thereby losing its inhibitory effect on tumor cell growth [30]. Further KEGG enrichment analysis revealed that these DRGs were involved in oncogenic pathways such as cell cycle, p53 signaling pathway, oocyte meiosis, DNA replication, and motor proteins, highlighting the potential role of these genes in the occurrence, progression, and metastasis of ovarian cancer cells (Fig. 2H) [31].

Identification and characterization of gene pair signature

The recurrent logistic regression (RLR) was then applied to the DRGs of Discovery Set I (GSE18520, GSE6008, and TCGA). To identify optimal gene panel and the hyperparameter, threshold optimization was performed with independent Discovery Set II (GSE27651) (Fig. 3A). Maximum performance of model (AUROC = 0.9836) was achieved with the parameters C = 0.803 and 0.804. KIF20A and CHGB were identified as discriminant gene markers for HGSOC in the model. To address batch effects arising from multi-platform and multi-cohort integration, PAGE algorithm was implemented for model robustness and reliability. The gene pair signature KIF20A::CHGB was established: Samples with KIF20A expression exceeding CHGB were classified as tumorous; others as normal.

Fig. 3.

Identification of drug-target biomarkers and prognostic analysis. A The red solid line represents the AUROC values corresponding to varying hyperparameter C values ranging from 0.800 to 0.900 (step size = 0.001). The blue curve indicates the number of genes selected under each C value within the same interval (step size = 0.001). The green dashed line denotes the model performance curve, reflecting a comprehensive evaluation metric across different C values. Red dots highlight two performance peaks at C = 0.803 and C = 0.804, where the model achieved optimal trade-offs between classification performance and gene selection. B Forest plot showing the rank-based expression of KIF20A and CHGB across multiple datasets. The x-axis indicates the standardized mean difference (SMD) between HGSOC patients and healthy controls. The blue square is the average value of the difference and the size corresponds to the concentration of the data. Red horizontal lines denote the 95% confidence intervals for each study. The purple diamond represents the pooled standardized mean difference across all cohorts, and its width reflects the 95% confidence interval of the overall estimate. C Rank distribution of KIF20A and CHGB between HGSOC and normal samples in the discovery set. D Distribution of KCP risk scores and classification accuracy in eight external validation cohorts. E, F AUROC and AUPRC of KCP in the eight validation cohorts. G Performance comparison of KCP and other models on all datasets in terms of AUROC and AUPRC

Subsequently, ranking differences between the two genes in HGSOC versus normal controls were calculated separately in different datasets (Fig. 3B). Forest plot results demonstrated a consistent pattern: KIF20A ranked significantly higher in HGSOC patients, whereas CHGB showed inverse trends (P < 0.05, Student’s T-test). Using the PAGE strategy, KIF20A and CHGB were paired to construct an optimal gene pair signature, named KIF20A::CHGB Pair (KCP). The decision criterion is solely based on intra-sample ranking of gene expression between KIF20A and CHGB, equipping KCP to cross-platform comparison which is independent of batch effects [18, 20, 32].

Further investigation of the expression ranking relationship between KIF20A and CHGB in the four discovery sets revealed a consistent pattern, KIF20A exhibited significantly higher ranks than CHGB in HGSOC samples, while normal controls exhibited the inverse (Fig. 3C). For instance, in dataset GSE18520 (63 samples), all 53 HGSOC samples displayed a higher rank of KIF20A compared to CHGB. In contrast, all of the 10 control samples showed CHGB predominance. Collectively, these findings demonstrate the robustness and discriminative power of KCP in distinguish HGSOC from normal samples.

Evaluation of the discriminant performance of KIF20A and CHGB

The discriminant ability of KCP is highlighted by its consistent accuracy in identifying HGSOC, with exceptional universality and reproducibility across diverse platforms and independent datasets. A risk prediction model was constructed based on the relative expression ranks of the KIF20A and CHGB in Discovery Set I and II (Supplementary Fig. 1). Risk scores were subsequently calculated for each sample in the validation cohorts, and patients across all cohorts were classified using a decision threshold of 0. Scatter plot revealed that the average classification accuracy of the model reached 0.9397 across eight independent validation datasets (Fig. 3D). The average AUROC of KCP in the eight validation cohorts is as high as 0.9668 and all AUPRC values exceeded 0.9799, illustrating high precision of KCP in HGSOC identification (Fig. 3E-F). Comparative analyses of AUROC and AUROC against other models across all datasets further supported these conclusions (Fig. 3H). Additionally, PAGE plot demonstrated the consistent rank-based separation of KIF20A::CHGB expression between HGSOC patients and controls across all external validation cohorts (Supplementary Fig. 2). Cumulatively, these multi-cohort validations confirm the robustness and accuracy of KCP in the classification of HGSOC.

KIF20A as a therapeutic target to restore chemosensitivity in cisplatin-resistant HGSOC

We constructed the interaction network of the biomarker KIF20A and performed KEGG pathway enrichment analysis on its interacting genes. KIF20A-related genes clustered in cell proliferation and carcinogenesis pathways (Fig. 4A). Notably, multiple pathways are primarily involved in platinum-based drug mechanisms, including the p53 signaling pathway, Platinum drug resistance, Homologous recombination (HR), Base excision repair (BER), Mismatch repair (MMR), and Nucleotide excision repair (NER), suggesting KIF20A’s coordination of platinum response mechanisms (Fig. 4B). Moreover, genes in these pathways associated with KIF20A showed significantly higher expression pattern in HGSOC samples (Fig. 4C), indicating these KIF20A-associated pathways were activated in HGSOC (Fig. 4D).

Fig. 4.

KIF20A as a potential platinum drug target for HGSOC. A KEGG pathways enriched by the intersection of two gene sets: (1) genes interacting with KIF20A in the STRING database with a combined interaction score ≥ 0.15, and (2) genes showing a Spearman correlation coefficient |ρ| > 0.4 and p < 0.001 with KIF20A in the validation cohorts. B KEGG pathways associated with KIF20A and its interacting genes involved in the mechanisms of platinum-based chemotherapy. C Heatmap of rank-based expression of platinum-related pathway genes in HGSOC and healthy control groups, along with the corresponding mean expression ranks between the two groups. D Mean rank values of KIF20A-interacting genes within platinum resistance-related pathways in HGSOC and normal controls. E Kaplan–Meier plot comparing the five-year survival rates of high-risk and low-risk groups in the combined TCGA and GSE18520 cohort, stratified by the optimal threshold of KIF20A expression

Collectively, these findings validate KIF20A’s prognostic biomarker value in HGSOC. We further evaluated the correlation between KIF20A expression and five-year overall survival using data from the TCGA and GSE18520 cohort. Patients were stratified into high-risk and low-risk groups based on the optimal cutoff value (threshold = 7304) for Kaplan–Meier survival analysis. The results indicated that patients in the high-risk group (KIF20A > 7304) exhibited significantly poorer clinical outcomes than the low-risk counterparts (KIF20A < 7304) (P < 0.028, log-rank test) (Fig. 4E). Overall, these findings suggest that elevated expression of KIF20A may enhance HGSOC resistance to platinum-based drugs, reducing treatment efficacy and highlighting its clinical utility as a potential platinum resistance biomarker.

Targeting KIF20A and CHGB reverses chemoresistance in ovarian cancer cells

Based on the above analytical results, we hypothesized that KIF20A and CHGB may be candidate genes affecting the sensitivity of ovarian cancer cells to chemotherapeutic drugs. To test this hypothesis, we separately knocked down KIF20A and CHGB and examined their effects on drug sensitivity in ovarian cancer cells. First, qRT-PCR was used to detect KIF20A mRNA expression after treatment with three different siKIF20A sequences, and all three siRNAs significantly reduced KIF20A expression (Fig. 5A). Similarly, CHGB mRNA expression after treatment with three different siCHGB sequences was detected, and all three siRNAs also significantly reduced CHGB expression (Fig. 5B). The siKIF20A with the highest knockdown efficiency (over 70%) and the siCHGB with the highest knockdown efficiency (over 70%) were selected for subsequent experiments. Next, CCK-8 assays were performed to examine the effects of KIF20A and CHGB knockdown on the IC50 values of commonly used ovarian-cancer chemotherapeutic drugs, including cisplatin, paclitaxel, gemcitabine, docetaxel, and olaparib. The results showed that KIF20A knockdown reduced the IC50 values of cisplatin, paclitaxel, gemcitabine, and docetaxel to varying degrees (Fig. 5C, Supplementary Fig. 3). CHGB knockdown also reduced the IC50 values of cisplatin, paclitaxel, gemcitabine, and docetaxel to varying degrees (Fig. 5D, Supplementary Fig. 3). Notably, after siKIF20A treatment, the cisplatin IC50 decreased from 8.787 μm to 4.084 μm in SKOV3 cells and from 3.640 μm to 1.393 μm in OVCAR3 cells (Fig. 5C-D). After siCHGB treatment, the cisplatin IC50 decreased from 8.787 μm to 1.981 μm in SKOV3 cells and from 3.640 μm to 1.751 μm in OVCAR3 cells (Fig. 5C-D).

Fig. 5.

Effect of KIF20A and CHGB knockdown on chemosensitivity in ovarian cancer cells. A Knockdown efficiency of KIF20A in SKOV3 and OVCAR3 cells after treatment with three different siKIF20A sequences. B Knockdown efficiency of CHGB in SKOV3 and OVCAR3 cells after treatment with three different siCHGB sequences. C Changes in IC50 values of five commonly used chemotherapeutic drugs (cisplatin, paclitaxel, gemcitabine, docetaxel, and olaparib) in SKOV3 cells after siKIF20A and siCHGB treatment. D Changes in IC50 values of the five commonly used chemotherapeutic drugs in OVCAR3 cells after siKIF20A or siCHGB treatment. E Synergistic effects (HSA values) of siKIF20A combined with the five chemotherapeutic drugs. F Synergistic effects (HSA values) of siCHGB combined with the five chemotherapeutic drugs. G Synergy-score plot of siKIF20A combined with cisplatin in SKOV3 and OVCAR3 cells. H Synergy-score plot of siCHGB combined with cisplatin in SKOV3 and OVCAR3 cells. *P < 0.05, **P < 0.01, ***P < 0.001. Data are from three independent experiments, and representative images are shown

In addition, we used the SynergyFinder tool to evaluate the gene-silencing–drug interaction for KIF20A or CHGB knockdown combined with drug treatment; higher synergy scores indicate stronger enhancement. Positive HSA scores indicated that KIF20A knockdown increased sensitivity to cisplatin in both SKOV3 and OVCAR3 cells (Fig. 5E-G) These results suggest that both KIF20A and CHGB have potential as chemosensitizing targets in ovarian cancer, and their knockdown can improve the efficacy of chemotherapeutic drugs.

Knockdown of KIF20A or CHGB synergizes with cisplatin to suppress growth and induce G2/M arrest in ovarian cancer cells

Because cisplatin showed the strongest and most consistent interaction with KIF20A and CHGB knockdown, we further examined the cellular effects of the combined treatments. Morphological observation and cell counting showed that siKIF20A alone, siCHGB alone, or cisplatin alone only modestly reduced cell numbers, whereas the combination treatments (siKIF20A + cisplatin and siCHGB + cisplatin) markedly decreased cell numbers in both SKOV3 and OVCAR3 cells and increased the proportion of enlarged cells (Fig. 6A-D).

Fig. 6.

Effect of KIF20A and CHGB knockdown combined with cisplatin treatment on growth, proliferation, and cell cycle in ovarian cancer cells. A Effects of siKIF20A or siCHGB combined with cisplatin treatment on SKOV3 cell morphology and number. B Effects of siKIF20A or siCHGB combined with cisplatin treatment on OVCAR3 cell morphology and number. C Bar chart showing relative changes in cell number in SKOV3 cells. D Bar chart showing relative changes in cell number in OVCAR3 cells. E Fluorescence images of EdU-positive cell proliferation in SKOV3 cells. F Fluorescence images of EdU-positive cell proliferation in OVCAR3 cells. G EdU-positive cell proportion in SKOV3. H EdU-positive cell proportion in OVCAR3. I Flow-cytometric analysis of cell-cycle distribution in SKOV3 cells. J Statistical plot of the proportions of each cell-cycle phase in SKOV3 cells. *P < 0.05, **P < 0.01, ***P < 0.001. Scale bar: 200 μm. Data are from three independent experiments, and representative images are shown

In EdU assays, the number of proliferating tumor cells was significantly reduced in the combination-treatment groups (siKIF20A + cisplatin and siCHGB + cisplatin), suggesting that knockdown of both KIF20A and CHGB can significantly enhance the inhibitory effect of cisplatin on ovarian-cancer cell proliferation (Fig. 6E-H). In addition, flow-cytometric cell-cycle analysis showed that siKIF20A alone, siCHGB alone, or cisplatin alone increased the proportion of cells in the G2/M phase. When KIF20A inhibition was combined with cisplatin, or CHGB inhibition was combined with cisplatin, the proportion of cells in the G2/M phase further increased significantly, suggesting that the enhanced therapeutic effect of combination treatment may be attributable to altered G2/M cell-cycle arrest in ovarian cancer cells (Fig. 6I-J).

Discussion

We identified KIF20A and CHGB as potential gene markers for HGSOC using the RLR algorithm, constructed KCP as a classification model for differentiating HGSOC from normal samples via the PAGE method, and validated the generalizability of this model based on eight external cohorts. In vitro experiments demonstrated that KIF20A and CHGB knockdown enhanced the sensitivity of ovarian cancer cells to chemotherapeutic agents, especially cisplatin. These results suggest that KIF20A and CHGB may serve not only as candidate biomarkers for HGSOC classification but also as potential targets for improving cisplatin response in ovarian cancer.

Although high-throughput data have been widely used for molecular biomarker screening, there is still no standardized protocol for preprocessing and integration of multi-source datasets [33–36]. A key advantage of the PAGE algorithm lies in its robustness and applicability without requiring data cleaning, normalization, or standardization, which makes it highly generalizable [37–39]. However, this pairwise approach results in a factorial increase in dimensionality (up to n! features), which poses substantial computational challenges. To address these challenges, we proposed a novel strategy that integrates reverse logistic regression model with PAGE [32, 40]. The reverse logistic regression assigns weights to individual features, quantifying gene importance and automatically eliminating irrelevant genes [41]. This approach improves classification accuracy while reducing model complexity. By leveraging the feature selection capability of reverse logistic regression and the robustness of PAGE, we aimed to maximize predictive performance while minimizing model complexity and enhancing translational applicability.

While the integration of RLR and PAGE introduces additional complexity in feature selection, we implemented several safeguards to mitigate the risk of overfitting. Importantly, feature extraction via RLR was performed exclusively on Discovery Set I, whereas all subsequent model optimization and validation were strictly conducted on the independent Discovery Set II, ensuring strict separation between training and validation data. To further validate the stability of the feature selection process, we conducted extensive robustness testing: we repeatedly subsampled 80% of the samples and ran the entire RLR workflow 50 times. Across all iterations, KIF20A and CHGB consistently ranked highest in selection frequency, confirming their robustness as candidate biomarkers (Supplementary Table 1). Moreover, the application of PAGE transforms raw expression data into rank-based features, which inherently reduces sensitivity to batch effects and variations that often confound molecular signature studies. The generalizability and stability of the resulting signature are further evidenced by the consistent high performance of the KCP score across eight independent external cohorts (average AUROC = 0.95), demonstrating that the KCP signature captures biologically reproducible signals rather than dataset specific noise.

KIF20A, a kinesin superfamily member, plays a critical role in chromosome transport during cell division [42]. KEGG enrichment analysis of DRGs in HGSOC and KIF20A-related gene interaction networks revealed that KIF20A participates in pathways of Cell cycle, Platinum resistance-related pathways, p53 signaling pathway, Oocyte meiosis, DNA replication, and Motor proteins [43, 44]. Platinum drug resistance significantly limits the clinical efficacy of platinum-based chemotherapeutic agents in cancer treatment. Previous studies have demonstrated that among these pathways, the p53 signaling pathway and the Mismatch Repair (MMR) pathway critically mediate platinum drug resistance mechanisms [45]. Our analysis indicates that KIF20A is significantly up-regulated in ovarian cancer and may modulate platinum drug efficacy through multimodal regulation. Aberrant overexpression of KIF20A has been consistently associated with cancer aggressiveness and poor clinical prognosis in various malignancies [46–49], including ovarian cancer [43, 50]. Further in vitro experiments supported this hypothesis. Knockdown of KIF20A markedly enhanced the sensitivity of ovarian cancer cells to cisplatin, with a reduced cisplatin IC50; combined treatment further suppressed cell proliferation and exacerbated G2/M phase cell cycle arrest. Cisplatin primarily exerts its anti‑tumor effects by inducing DNA damage, and both G2/M phase arrest and alterations in DNA damage repair capacity can influence cellular responses to platinum agents [51]. Therefore, high KIF20A expression may help tumor cells tolerate cisplatin‑induced damage by sustaining mitotic progression and DNA damage response capacity. Conversely, KIF20A inhibition may impair cell cycle regulation and damage repair capability, thereby enhancing the cytotoxicity of cisplatin. These findings suggest that KIF20A holds promise as a candidate marker for assessing platinum resistance risk in HGSOC and as a potential intervention target to increase cisplatin sensitivity.

Meanwhile, CHGB is a secretory granule protein primarily expressed in neuroendocrine cells and peptidergic neurons. It serves as a precursor for various regulatory peptides and participates in secretory pathway regulation, extracellular signaling, and maintenance of neuroendocrine functions. Unlike KIF20A, CHGB has been rarely reported in HGSOC, and its biological significance in ovarian cancer initiation and progression remains unclear. In the KIF20A::CHGB gene pair model constructed in this study, CHGB was not identified as a single high‑expression oncogene, but rather as a paired gene exhibiting an inverse ranking relationship with KIF20A, together forming a stable intra‑sample relative expression signature. This signature demonstrated good discriminative ability across multiple external validation cohorts, indicating that CHGB contributes significantly to improving cross‑platform stability and reducing batch effects. Notably, although direct associations between CHGB and HGSOC are still underreported, our survival analyses suggested a certain correlation between CHGB expression levels and patient 5‑year overall survival (Supplementary Fig. 4); cellular experiments also showed that CHGB knockdown similarly enhanced cisplatin sensitivity in ovarian cancer cells and aggravated cisplatin‑induced proliferation inhibition and G2/M phase arrest. Thus, CHGB knockdown can increase the sensitivity of ovarian cancer cells to cisplatin and other chemotherapeutic agents to some extent. This result suggests that CHGB may not only serve as a pairing signature in the KCP model but also participate in regulating drug response in ovarian cancer cells.

While the diagnostic performance of our model for ovarian cancer was high, it relies on a limited number of genes, and the potential contribution of additional chemoresistance-associated genes cannot be ruled out. Moreover, the current experimental data mainly demonstrate the phenotypic effects of KIF20A and CHGB knockdown on chemosensitivity, cell proliferation, and G2/M-phase arrest, whereas the specific downstream molecular pathways remain to be fully characterized.Future studies will further examine key molecules involved in DNA damage response, cell-cycle checkpoint regulation, and apoptosis, such as γ-H2AX, p53, p21, Cyclin B1, CDK1, cleaved PARP, and cleaved caspase-3. In addition, RNA sequencing, protein-level validation, and rescue experiments will be required to clarify the causal relationship between KIF20A/CHGB and platinum resistance-related pathways. In summary, we identified KIF20A and CHGB as potential HGSOC biomarkers. Subsequently, we paired them to develop a classification model with enhanced robustness and generalizability. In vitro assays demonstrated that KIF20A and CHGB knockdown enhanced the sensitivity of ovarian cancer cells to cisplatin and suppressed tumor cell proliferation, suggesting their potential value as chemosensitization targets for cisplatin-resistant HGSOC. However, the detailed downstream mechanisms underlying these effects remain to be further elucidated.

Authors’ contributions

LC, ML, and HL conceived and designed the algorithms and experiments. XW, NZ, HW, QC, JZ, and SM performed the analysis and experiments. JL and DH preprocessed the data and provided suggestions on improving the analysis. XW, SM, and SW performed the experimental validation. LC, JZ, CZ, JB, PY, and YZ drafted and revised the manuscript. LC, ML, and HL supervised the study. All authors read and approved the final manuscript.

Funding

This research was supported by National Natural Science Foundation of China (32370711, 32470804, 82373126), Shenzhen Medical Research Fund (A2303033), and Shenzhen People’s Hospital Physician Scientist Training “Five Three Program” (SYWGSJCYJ202401); This research was also supported by Sanming Project of Medicine in Shenzhen (SZSM202311011).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.