Construction of a prognostic model based on disulfidptosis-related genes and identification of CCNA2 as a novel biomarker for hepatocellular carcinoma

Abstract

Background

Disulfidptosis, identified as an innovative form of cellular death subsequent to cuproptosis, is currently under investigation for its mechanisms in oncological contexts. In-depth analyses exploring the relationship between disulfidptosis-related genes (DRGs) and hepatocellular carcinoma (HCC) are currently limited.

Methods

Transcriptomic data and clinical information were retrieved from the TCGA and GEO databases (GSE76427 and GSE54236), concentrating on the expression levels of 24 DRGs. Subsequently, multifactor and LASSO regression analyses were utilized to construct the 5-DRG prognostic signature. Immunohistochemistry (IHC) was employed to assess Cyclin A2 (CCNA2) protein expression levels. Quantitative real-time PCR (qRT-PCR) and western blot analyses were conducted to detect transcriptomic and protein expression of CCNA2-targeting short interfering RNA (siRNA). The Cell Counting Kit-8 (CCK-8) assay, EdU staining, and scratch experiments were employed to observe the proliferation and migration of hepatoma cell lines subsequent to CCNA2 inhibition.

Results

Three HCC patterns were identified, among which pattern B exhibited the the most unfavorable survival outcomes. Five DRGs (STC2, PBK, CCNA2, SERPINE1, and SLC6A1) were involved to establish the 5-DRG prognostic signature. High-risk groups (HRGs) exhibited prolonged survival durations in comparison to low-risk groups (LRGs). Both bioinformatics analyses and experimental methodologies corroborated the association of CCNA2 with poor prognosis in HCC patients. Functional studies elucidated that interference with CCNA2 significantly inhibited proliferation and migration, while simultaneously promoting apoptosis in hepatoma cells and resulting in the downregulation of epithelial-mesenchymal transition (EMT)-related protein markers.

Conclusions

The 5-DRG prognostic signature is proficient in predicting clinical outcomes, informing therapeutic strategies, and elucidating the characteristics of the immune microenvironment in HCC patients. Furthermore, this study elucidates the potential of CCNA2 as an innovative biomarker for HCC.

Introduction

In 2020, liver cancer accounted for 410,000 new cases and 390,000 deaths in China, representing 45.6% and 47.1% of global totals, respectively—the highest in the world [1]. Epidemiological studies reveal that China’s annual incidence rate of hepatocellular carcinoma (HCC) is approximately 0.03%, markedly surpassing the rates observed in Japan, Europe, and the United States, where the incidence is 0.01% or lower [2]. HCC accounts for 75%-85% of primary liver cancer [3], while non-alcoholic steatohepatitis (NASH) is rapidly emerging as the leading contributor to the increasing incidence and mortality associated with liver cancer on a global scale [4]. The HCC patient demographic in China is highly heterogeneous, primarily attributable to hepatitis B virus infections and variations in immune function. Approximately 70% of these patients are diagnosed at intermediate to advanced stages, thereby complicating their management and treatment strategies [5]. The five-year survival rate for HCC in China stands at just 12.1%, even lower for those with advanced disease, underscoring the urgent need for more personalized treatment approaches [6].

Current medical perspectives classify malignant tumors as metabolic disorders, highlighting the role of aberrant energy metabolism in the diagnosis, treatment, and prognosis of cancers. Recent studies highlight that regulated cell death is crucial in managing cancer metabolism. Disulfidptosis, a rapid form of cell death distinct from other types and driven by metabolic processes, occurs due to disulfide stress from excessive intracellular cystine accumulation [7,8,9,10]. Liu et al. [11] found that high expression of SLC7A11 is associated with poor prognosis in adrenocortical carcinoma. Similarly, Huang et al. [12] discovered that a model based on disulfidptosis could effectively predict outcomes for patients with lung adenocarcinoma. However, the connection between disulfidptosis and HCC remains poorly understood. Thus, the development of a specific, personalized, and sensitive HCC signature based on disulfidptosis-related genes (DRGs) is clinically valuable. The advancement of genomics has facilitated precise disease typing and treatment. This study focuses on DRGs, seeking to understand their role in liver cancer and identify potential therapeutic targets for HCC using bioinformatics.

As shown in Fig. 1, machine learning was performed to construct the 5-DRG prognostic signature for HCC using The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) databases, clarify the correlation between the 5-DRG prognostic signature in prognosis, immune infiltration, immunotherapy and tumor microenvironment in HCC and provides experimental validation of CCNA2 as an innovative biomarker for HCC, thereby establishing a theoretical foundation for future research endeavors.

Fig. 1

Flow chart of the research

Materials and methods

Data sets

This study analyzed 1,041 HCC samples, comprising 371 from The Cancer Genome Atlas (TCGA), 439 from the Gene Expression Omnibus (GEO) accessed on the same date (including GSE76427 with 115 samples, GSE54236 with 81 samples, and GSE14520 with 243 samples), and 231 from the International Cancer Genome Consortium (ICGC). All samples included transcriptome expression and corresponding clinical data. Additionally, six GEO cohorts (GSE6764 with 45 samples, GSE102790 with 257 samples, GSE112790 with 198 samples, GSE121248 with 107 samples, GSE62232 with 136 samples, GSE89377 with 53 samples) comprising both normal and tumor samples were merged into a new database termed ‘six-GEO cohorts’ after removing batch effects. Copy number variation and somatic mutation data for tumor specimens were sourced from the UCSC Xena database. Chromosomal alterations caused by CNVs were visualized using the “Rcircos” package in R software. Transcriptome data have been converted to transcripts per million kilobases (TPM) and processed using log₂ (value + 1) for compatibility with GEO expression profiles and integrated with GSE76427 and GSE54236 to create a new data matrix, the TCGA-GEO cohort. Furthermore, transcriptional data from diverse sources were harmonized utilizing the “SVA” package in R software to eliminate batch effects.

Classification of different HCC subtypes under disulfidptosis-related genes (DRGs)

Table S1 listed 24 disulfidptosis-related genes (DRGs) collected from published literature [8]. Following the expression analysis of DRGs, HCC patients were classified into various subtypes through consistent unsupervised clustering analysis using the “ConsensusClusterPlus” package in R software. Gene Set Variation Analysis (GSVA) (c2.cp. kegg. v7.5.1) from the MsigDB was employed to explore the potential function of DRGs. Principal Component Analysis (PCA) was utilized to better delineate the distinctions among various HCC subtypes. The “Estimate” package in R software was used to calculate immune, stromal, and ESTIMATE scores for different HCC clusters.

Identification of the risking model and determination of hub genes

Initially, the “limma” package in R software was used to identify differentially expressed genes (DEGs) in three HCC subtypes, resulting in 221 overlapping DEGs with an adjusted p-value < 0.01 and |logFC| > 1. Subsequently, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses were conducted to elucidate the potential mechanisms. Following this, 80 DEGs were selected via one-way Cox regression analysis (P < 0.001) for constructing the risk model. The risk model was subsequently developed using the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm and multivariate Cox regression analysis. The risk score calculation equation is as follows: Risk Score = ∑Coef(i) * Exp(i).

Validation of the general applicability of the 5-DRG prognostic model

Three independent databases (ICGC, GSE14520, and GSE54236) served as external validation sets to evaluate the accuracy and generalizability of the 5-DRG prognostic model based on risk scores. Additionally, the precision of the 5-DRG prognostic model was assessed using a range of bioinformatics tools.

Identifying and demonstrating the Hub Gene in the HCC process

As previously mentioned, six GEO cohorts, comprising 190 non-tumor and 561 HCC patients, were combined into a new matrix. Initially, DEGs were identified as genes with an adjusted P < 0.05 and |logFC| > 1.5. Subsequently, the Random Forest (RF) method, a prevalent machine learning, was employed to identify hub genes with a variable importance score > 2.0. The gene identified by both methods was ultimately designated as the hub gene.

Cell culture and RNA interference

The human-derived hepatocyte cell line MIHA and four HCC cell lines (HepG2, SUN-449, HuH-7, Hep3B) were acquired from the Cell Bank of the Chinese Academy of Sciences, Shanghai, China. The MIHA, HepG2, HuH-7, and Hep3B cell lines were cultured in high fructose DMEM (Gibco), while SUN-449 was cultured in RPMI 1640 medium. All culture media were supplemented with 10% Fetal Bovine Serum (Gibco), 100 U/ml penicillin, and streptomycin (Gibco), and maintained at 37°C with 5% CO₂. Small interfering RNAs (siRNAs) targeting the CCNA2 gene and negative control siRNA were synthesized by Genepharma (Zhousu, China). The sequences of CCNA2 siRNAs were as follows: siRNA-1: F5’- GUAGCAGAGUUUGUGUACATT, F3’-UGUACACAAACUCUGCUACTT; siRNA-2: F5’-CAGCCAGACAUCACUAACATT, F3’- UGUUAGUGAUGUCUGGCUGTT; siRNA-3: F5’- GGGAGAAUUAAGUUUGAUATT, F3’- UAUCAAACUUAAUUCUCCCTT; siRNA-4: F5’-GCCAGUGAGUGUUAAUGAATT, F3’- UUCAUUAACACUCACUGGCT. Negative control (Si-NC): F5’-UUCUCCGAACGUGUCACGUTT, F3’- ACGUGACACGUUCGGAGAATT.

Cell proliferation assay

The transfected cells were inoculated into 96-well plates at 5000 cells per well, and cell proliferation was detected at 0, 24, 48, and 72 h using the CCK-8 kit (MCE, HY-K0301, USA) at 450 nm at dual wavelengths. The aforementioned test was conducted independently on three separate occasions.

Cell scratch assay

First, a horizontal line is drawn on the underside of the 6-well plate with a straightedge, and then logarithmically grown cells are inoculated into the 6-well plate. Then, cells with loose apposition were washed with PBS and then streaked again after 24 and 48 h. The aforementioned test was conducted independently on three separate occasions.

EdU fluorescent staining

Reagents for transfection of CCNA2 were added to a 6-well plate with 1 × 10^{^}5 cells per well. BeyoClick EdU Cell Proliferation Kit with Alexa Fluor 488 (Beyotime, C0071S, Jiangsu, China) was used to conduct experiments and results were presented by inverted fluorescence microscopy. The aforementioned test was conducted independently on three separate occasions.

Cell cycle analysis

Follow the cell cycle kit instructions (Beyotime, C1052, Jiangsu, China). First, cells were immobilized with previously cooled 70% ethanol at 4 °C for 30 min, centrifuged and washed several times by adding pre-cooled PBS, followed by the addition of 0.5 mL of propidium iodide staining solution of the corresponding system, and warmed in a warm bath protected from light for a half hour. The aforementioned test was conducted independently on three separate occasions.

Clinical sample collection and two HCC mouse models construction

HCC tumor samples with their paratumor tissues were collected and conducted with the permission of the ethics committee of Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine (No.2018-630-59-01). All the male C57BL/6 mice at 3–5 weeks of age were purchased from Shanghai Jiesijie Laboratory Animal Co. Ltd (Shanghai, China). The mice were randomized into the control and tumor groups (n = 5 per group). The two HCC models were constructed as described previously [13] and conducted with the permission of the ethics committee of Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine (PZSHUTCM220822019).

RT-qPCR

Total mRNA was extracted from the cell lines using TRIzol reagent (Bioteke Corporation, RP40002) and reverse-transcribed into cDNA using the Reverse Transcription Master Kit (Vazyme, R222-01). Quantitative real-time PCR (qRT-PCR) was conducted with ChamQ SYBR qPCR Master Mix (Vazyme, Q311-02). The mRNA levels were normalized to the expression in MIHA cells. Primer sequences used are listed as follows. CCNA2: Forward 5’-3’: CAGAAAACCATTGGTCCCTC, Reverse 5’-3’: CACTCACTGGCTTTTCATCTTC. β-actin: Forward5’-3’: TGTGTTTTCCTCCTTGCCTCTGAT, Reverse5’-3’: ATGCCACAGGATTCCATACC.

Western blot

Tissues and cells were lysed using RIPA lysis buffer (Beyotime, P0013B, Jiangsu, China) mixed with PMSF and phosphatase inhibitors (Beyotime, P1082) on ice for 30 min, followed by centrifugation. The supernatant was collected for BCA protein concentration quantification (Thermo Scientific, 23227, USA). The separated proteins were transferred to a PVDF membrane using 10% SDS-PAGE. Following blocking with 5% skimmed milk, the methanol-activated PVDF membrane was incubated with specific primary antibodies on a shaker at 4 °C overnight, followed by the addition of specific secondary antibodies at room temperature. Protein detection was carried out using equal volumes of ECL solutions A and B (UU-Bio technology, U10012, Zhousu, China). The specific antibodies used are detailed in Table S2.

Immunohistochemical (IHC) staining

Tissue sections were deparaffinized and subjected to antigen retrieval using a citric acid antigen repair solution (pH 6.0). This was followed by blocking endogenous peroxidase activity with 3% hydrogen peroxide. After three washes, the sections were incubated with the primary antibody overnight at 4 °C, then with the secondary antibody for 50 min at room temperature. The sections were then treated with DAB solution for visualization, observed under a microscope, and stained with Hematoxylin for nuclear counterstaining. Final dehydration and sealing were performed using neutral balsam.

Statistical analysis

Statistical analyses were conducted using R software (version 4.2.2) and its associated packages. The T-test was employed for comparisons between two groups, while one-way Analysis of Variance (ANOVA) was used for multiple-group comparisons. Unless otherwise stated, a p-value < 0.05 was considered statistically significant.

Results

The genetic landscape of 24 disulfidptosis-related genes (DRGs)

Figure S1A presents the transcriptional profiles of 24 DRGs in both normal and tumor samples, revealing a marked overexpression in HCC samples. The waterfall plot (Fig. S1B) shows that 40 (10.99%) of 364 HCC samples exhibited mutations in these DRGs, primarily missense mutations. Fig. S1C indicates widespread copy number amplification among the DRGs, with CAPZB showing the most significant copy number deletion. The chromosomal positions of 24 DRGs are shown in Fig. S1D. Additionally, PCA plots (Fig. S1E) effectively differentiate between normal and tumor samples based on the expression profiles of DRGs, while the correlation network plot (Figure S1F) elucidates the interactions among these genes. These observations underscore the pivotal role of the 24 DRGs in the pathogenesis and mutational landscape of HCC.

Identification of three HCC patterns and biological mechanisms

The network diagram (Fig. 2A) demonstrates the co-expression and prognostic values of 24 differential related genes (DRGs) in HCC patients, identifying most as risk factors in HCC progression, except for MYH10, which emerged as the sole favorable gene. The Principal Component Analysis (PCA) plot indicates that the batch effect of HCC samples was eliminated after processing (Fig. S2A-B). Cluster analysis shows three distinct HCC patterns differentiated based on the expression of these DRGs (Fig. 2B, Fig. S2 C-F). PCA depicted in Fig. 2C, highlights significant distinctions among the three patterns. Survival analysis, shown in Fig. 2D, indicates that pattern C is associated with the poorest prognosis compared to the other patterns. Figure 2E presents a heatmap that integrates the expression of the 24 DRGs with clinicopathological characteristics and survival status. Single-sample gene set enrichment analysis (ssGSEA) (Fig. 2F) examines the relationship between 22 immune cells and the three HCC patterns, revealing a notable upregulation of almost all immune cells in pattern B, underscoring the critical role of immune pathways. Lastly, the ESTIMATE algorithm calculates the highest estimate score (Fig. 2G), immune score (Fig. 2H), and stromal score (Fig. 2I) for pattern B, further emphasizing its unique immune and stromal characteristics.

Fig. 2

Creation of three HCC patterns. (A) A prognostic network diagram showing genetic risk, prognostic value, and genetic relationships. (B) Optimal fractal state at k = 3. (C) The PCA analysis of three HCC patterns. (D) Survival analysis of three HCC patterns. (E) A heatmap covering clinical traits, different HCC patterns, and gene expression. (F) Relationship between three HCC patterns and immune cells. (G-I) Relationship between three HCC patterns in (G) estimate score, (H)immune score, and (I) stromal score

Capturing of DRG cluster and analysis of latent differential molecular pathways

Gene Set Variation Analysis (GSVA) was conducted to assess functional differences across survival periods. Notably, pattern B exhibited significant enrichment in cancer pathways, including those associated with small-cell lung cancer and pancreatic cancer (Fig. S3A). In contrast, pattern A was enriched in olfactory transduction and neuroactive ligand-receptor interaction pathways (Fig. S3B), while pattern C correlated with metabolic processes like pyruvate and butanoate metabolism (Fig. S3C).

Using the “limma” package in R studio, 221 overlapping genes across the three HCC patterns were identified with an adjusted P-value < 0.01 and |logFC| > 0.585, illustrated in a Venn diagram (Fig. S3D). Gene Ontology (GO) analysis revealed that these genes primarily influence extracellular structures (Fig. 3A). Additionally, KEGG analysis underscored critical pathways including PI3-AKT and TNF signaling, focal adhesion, actin cytoskeleton regulation, and pathways related to proteoglycans in cancer and human papillomavirus infection in HCC (Fig. 3B).

Fig. 3

Identification of HCC clusters based on DEGs. (A) The GO pathway annotation and (B) the KEGG annotation in DEGs. (C) three DRG clusters were identified based on the optimal K value(K = 3). (D) PCA analysis of three DRG clusters. (E) Survival curves for three DRG clusters. (F) The heatmap containing clinical information, different HCC subtypes, and gene expression levels. (G) Box plot of differences in expression levels of 24 DRGs in three DRG clusters

Univariate Cox regression analysis pinpointed 79 of these genes as significantly correlated with HCC patient survival (P < 0.001) (Fig. S3E), forming the basis for defining three distinct DRG clusters with unique pathological mechanisms (Fig. 3C). PCA analysis confirmed these distinctions (Fig. 3D), with cluster B associated with the poorest prognosis compared to clusters A and C (Fig. 3E). A comprehensive heatmap, which includes variables such as gender, age, survival status, clinical stage, and the expression profiles of these 79 differentially expressed genes (DEGs) across HCC clusters, is illustrated in Fig. 3F.

Development and validation of the 5-DRG prognostic signature

The Least Absolute Shrinkage and Selection Operator(LASSO) regression identified 12 of the 79 DEGs as significant, as illustrated in Fig. 4A. Subsequent multivariate Cox regression analysis revealed five key genes—STC2, PBK, CCNA2, SERPINE1, and SLC6A1—as core components of the 5-DRG prognostic signature (Fig. 4B). The risk score formula is defined as: Risk-Score = exp (STC2) × 0.192025861858049 + exp (CCNA2) × 0.164522570871041 + exp (SERPINE1) × 0.0942345591963748 + exp (PBK) × 0.151623718743126 + exp (SLC6A1) × (-0.120172055823625). This algorithm was applied to score all HCC patients in the TCGA-GEO dataset, stratifying them into high-risk groups (HRGs) and low-risk groups (LRGs) based on the median score. Patients in HRGs showed a shorter life expectancy and poorer survival outcomes compared to those in LRGs (Fig. 4C-E). A heatmap (Fig. 4F) displayed the differential expression of these five genes, revealing that four of them, excluding SLC6A1, were overexpressed in HRGs. ROC curve analysis demonstrated a strong predictive capability with area under the curve (AUC) values > 0.65 across 1, 3, and 5 years (Fig. 4G). Comparisons of the risk score with clinical phenotypes suggested that the 5-DRG prognostic signature more accurately predicted outcomes than gender, age, and clinical stage (Fig. 4H). This finding was corroborated by the concordance index (C-index) (Fig. 4I). A nomogram incorporating age, gender, risk score, and clinical stage was developed to aid clinicians in predicting survival for HCC patients (Fig. 4J), with the accuracy confirmed by a calibration curve (Fig. 4K).

Fig. 4

Construction of the 5-DRG prognostic signature. (A) Coefficient display of 12 DRGs analyzed by LASSO analysis. (B) The partial likelihood deviance of LASSO analysis. (C) KM-analysis of two risk groups. (D) Risk values for two risk HCC patients. (E) Map of survival distribution points for two risk patients. (F) Expression levels of five DRGs in two risk HCC patients. (G) 1-year, 3-year, and 5-year survival period prediction. (H) The AUC of risk score and clinical characteristics. (I) The DCA for risk score and clinical traits. (J) Construction of nomogram to guide clinical practice. (K) Calibration curves for the nomogram

Multi-queue HCC data validation and advantages of clinical application

To assess the generalizability of the 5-DRG prognostic signature, median risk values from the TCGA-GEO cohort were applied to three independent HCC cohorts: ICGC, GSE14520, and GSE54236. As anticipated, patients classified as high-risk across these cohorts displayed unfavorable prognoses, thereby confirming the robustness of the model (Fig. S4A-C). This signature was designed for clinical use, leading to additional analyses to highlight its substantial benefits. Initial univariate and multivariate regression analyses aimed to verify the model’s performance as independent of other clinical factors. Univariate analysis identified stage (HR 1.377, 95% CI 1.240–1.529, P < 0.001) and risk score (HR 1.683, 95% CI 1.390–2.037, P < 0.001) as significant predictors. Further, multivariate analysis that included clinical stage and risk score confirmed the 5-DRG signature as an independent risk factor (Fig. 5A-B). Patients in DRG cluster B demonstrated the highest risk scores relative to other clusters, with notably higher DRG expression in HRGs compared to the LRG (Fig. 5C, G). Additionally, patients in gene cluster B showed the highest risk scores (Fig. 5D). The risk score correlated strongly with advanced clinical stages, with no marked differences in age and gender (Fig. 5E), confirmed by a heatmap (Fig. 5F). A Sankey diagram depicted the varying survival states across different HCC subtypes (Fig. 5H). Further, a higher tumor stemness index indicated increased tumor differentiation and malignancy. Spearman analysis showed a positive correlation between the risk score and RNAss (R = 0.13, p = 0.016) (Fig. 5I), supporting the utility of the risk score as a valuable clinical tool for diagnosing and predicting the survival of HCC patients.

Fig. 5

Clinical characteristics of the 5-DRG prognostic signature. (A) Univariate and (B) multivariate Cox analysis on the signature and clinical information such as age, gender, and stage. (C) Comparison of the risk values of three DRG clusters and (D) gene clusters. (E) Histogram of the percentage of clinical traits such as age, gender, and stage in LRGs and HRGs. (F) Heatmap of differences in survival status, age, gender, and stage in LRGs and HRGs. (G) Expression boxplots for LRG and HRG at 24 disulfidptosis genes. (H) A snakey on gene cluster, DRG cluster, and survival status. (I) Scatterplot of correlation between RNAss and risk score. (J) Expression boxplots for LRGs and HRGs at immune checkpoints

Correlation analysis of the risk score in tumor microenvironment and immunotherapy

The introduction of immunotherapy has brought both opportunities and challenges to cancer treatment. In the liver, immunosuppressive cells normally protect against autoimmunity and chronic inflammation. However, in cancerous conditions, these cells can facilitate immune escape and tumor progression, underscoring the need to evaluate immunotherapy efficacy. Initial analyses of common immune checkpoints in high-risk groups (HRGs) revealed significant upregulation (Fig. 5J), suggesting potential immune escape and a poor response to immunotherapy. Transcatheter arterial chemoembolization (TACE) and sorafenib represent prevalent non-surgical and molecularly targeted therapies, respectively, for advanced hepatic malignancies. It is vital to assess their effectiveness in treating hepatocellular carcinoma. This study involved 147 HCC patients treated with TACE (GSE104580 cohort) and 83 treated with sorafenib (GSE109211 cohort). Initial observations showed differential expression of the five genes in the prognostic model between responders and non-responders to both treatments (Fig. 6A, D). ROC curve analysis for predicting responses to TACE and sorafenib treatment yielded AUC values of 0.757 and 0.747, respectively (Fig. 6B, E). Furthermore, patients who responded to either treatment exhibited significantly lower risk scores (Fig. 6C, F). These results suggest that the prognostic signature may serve as a novel marker for assessing the efficacy of TACE and sorafenib treatments in hepatocellular carcinoma.

Fig. 6

Prognostic model response to TACE and sorafenib. (A) Expression of five DRGs in the TACE-treated GSE104580 cohort and (D) sorafenib-treated GSE109211. (B, E) Sensitivity of the 5-DRG prognostic signature to predict (B) TACE and (C) sorafenib response. (C, F) Risk values differ between (C) TACE-response and non-response groups and (F) sorafenib-response and non-response groups

Tumor mutation burden (TMB) and drug sensitivity

Tumor Mutation Burden (TMB) quantifies tumor gene mutations at the microscopic level. Analysis of gene mutations identified TP53 as the most frequently mutated gene, present in 38% of high-risk groups (HRGs) (Fig. S5A), while CTNNB1 was the most mutated in low-risk groups (LRGs), occurring in 29% of samples (Fig. S5B). This study supports existing research identifying TMB as a significant prognostic marker, showing that high TMB (H-TMB) patients have poorer survival outcomes compared to those with low TMB (L-TMB) (Fig. S5C). Additionally, patients with L-TMB in HRGs demonstrated the shortest survival times among the groups analyzed (Fig. S5D), highlighting the critical need for careful monitoring of these patients. However, no significant differences were found in TMB levels between HRGs and LRGs, likely due to the low mutation frequency in disulfidptosis-related genes (DRGs) in HCC (Fig. S5E). Moreover, an evaluation of over 100 drugs commonly used in clinical practice was performed, aiming to enhance the precision of HCC treatment (Fig. S5F). In summary, this risk model effectively predicts prognosis, clarifies aspects of the tumor microenvironment in liver cancer, guides immunotherapy decisions, and facilitates the development of precision treatment strategies.

Identification of the hub gene

Six GEO cohorts (GSE6764, GSE102790, GSE112790, GSE121248, GSE62232, GSE89377), comprising a comprehensive dataset, were utilized in finding the hub gene. Initially, six GEO cohorts were amalgamated into a novel dataset, following the removal of batch effects. Utilizing a line graph, we delineated the expression of the five genes for which the model was constructed and observed that, except SLC6A1, the expressions of the other four genes were elevated in tumor tissues relative to normal tissues (Fig. 7A). Subsequently, DEGs between normal and HCC samples were identified and filtered, applying criteria of adjusted P < 0.01 and |logFC| >1.5, as illustrated in the volcano plot (Fig. 7B). Candidate genes, characterized by relative importance greater than 2.0, were selected (Fig. 7C), and the corresponding decision tree is depicted in Fig. 7D. Following this, an intersection was performed between the genes identified by the random forest algorithm and those ascertained through LASSO regression analysis, revealing CCNA2 as the sole overlapping gene (Fig. 7E). The expression of CCNA2 was subsequently analyzed across 33 different cancer types in the TCGA database, demonstrating significant upregulation in all tumor instances (Fig. 7F). In conclusion, our findings suggest that CCNA2 could potentially serve as a novel prognostic marker in hepatocellular carcinoma.

Fig. 7

Identification of the hub gene. (A) Line graph of the expression of five DRGs in the six GEO cohorts. (B) Volcano map for six GEO cohorts. (C) Top 30 genes identified using random forest method. (D) Decision trees for random forests. (E) Random forest and LASSO regression intersection genes demonstrated by Wayne diagram. (F) CCNA2 expression in 33 tumors in the TCGA database

Diagnostic value of the CCNA2

Transcriptomic data from two GEO databases (GSE6764 and GSE89377), encompassing various stages of hepatocellular carcinoma, were analyzed to evaluate the diagnostic potential of CCNA2 in early-stage liver cancer. Comparative analysis revealed that the transcription levels of CCNA2 were markedly higher in HCC patients compared to individuals with dysplastic liver tissue (Fig. 8A-B), suggesting its potential as an early-stage liver cancer diagnostic marker, a notion supported by the significant AUC values (Fig. 8C-D). Moreover, the prognostic impact of CCNA2 expression on HCC was evaluated, revealing that patients with elevated CCNA2 levels experienced significantly shorter overall survival (OS) and disease-free survival (DFS) (Fig. 8E-F). This conclusion was further substantiated by analysis of two independent GEO cohorts, GSE116174 and GSE54236 (Fig. 8G-H). Lastly, the association between CCNA2 expression and clinical features was examined, revealing that HCC patients with higher CCNA2 levels tended to have more advanced clinical T stage, clinical stage, and G stage, with no significant difference in M stage (Fig. 8I-L). Collectively, this evidence underscores the pivotal role of CCNA2 in the diagnosis of hepatocellular carcinoma. Western Blot analysis revealed a significant upregulation of CCNA2 expression in HCC specimens compared to paracarcinoma tissue samples (Fig. 9A), a finding corroborated by Immunohistochemistry (IHC) results (Fig. 9B). Furthermore, consistent results were observed in two HCC mouse models via the Western Blot (Fig. 9C). These findings collectively demonstrate that CCNA2 is overexpressed in HCC.

Fig. 8

Clinical value of CCNA2. (A-B) Differential expression of CCNA2 in dysplastic liver tissue and HCC in GSE6764 and GSE89377. (C-D) Sensitivity of CCNA2 for differentiating dysplastic liver tissue and HCC in GSE6764 and GSE89377. (E-F) Differential survival of (E)OS and (F) DFS in high-expression and low-expression of CCNA2. (G-H) K-M survival curve of high-expression and low-expression of CCNA2 in GSE116174 and GSE54236. (I-L) Differential expression of CCNA2 in different clinical traits such as T stage, clinical stage, G stage, and M stage in TCGA-LIHC

Fig. 9

Relationship of CCNA2 to cell proliferation, migration, cell cycle, and apoptosis. (A) Protein levels of CCNA2 in HCC and paraneoplastic tissues. (B) Immunohistochemistry detected the protein expression level of CCNA2 in normal, paraneoplastic, and tumor tissues. (C) Protein expression levels of CCNA2 in normal and two HCC mouse models. (D) Protein expression and transcriptomic levels of CCNA2 in normal hepatocytes and tumor cell lines as demonstrated by the western blot and qRT-PCR. (E) Protein expression and mRNA levels of silenced CCNA2 are shown by the western blot and qRT-PCR. (F) Representative images on Edu staining and quantitative analysis. (G) Representative correlation image of the scratch experiment (H-I) quantitative analysis of the Edu staining, the scratch assay on 24 and 48 h. (J)Temporal fold plot of cell proliferative capacity at different periods as detected by CCK8 assay. (K) Plot of cell flow cycle changes and quantitative histograms for each cell cycle. (L) CDK1 protein expression levels after silencing of CCNA2 as detected by the western blot assay. (M) Relative protein degrees of apoptosis-related markers (Bax, Bcl2, and P53). (N-O) Relative protein levels of EMT-related markers (MMP7, MMP9, N-cadherin, E-cadherin, ZO-1, β-catenin, and vimentin). (P) Relative protein levels of PI3K/AKT pathways. *p < 0.05, **p < 0.01,*** p < 0.001, ns No Significance

CCNA2 promotes proliferation and migration in the HCC cell line

Protein levels of CCNA2 in various hepatocyte cell lines, including MIHA, HepG2, Hu-7, SNU-449, and Hep3B, were quantified using Western blot experiments. The results showed that transcriptional and translational levels of CCNA2 were significantly higher in HCC cell lines (HepG2, Hu-7, SNU-449, and Hep3B) compared to the normal human hepatocyte cell line MIHA, with the highest expression observed in SNU-449 (Fig. 9D). To investigate CCNA2’s role in liver cancer pathogenesis, CCNA2 expression was silenced in the SNU-449 cell line, with validation provided by qRT-PCR and Western blot analyses (Fig. 9E). Functional assays, including the EdU assay, wound healing assays (Fig. 9F-I), and CCK-8 assays (Fig. 9J), demonstrated that silencing CCNA2 significantly reduced the migration and proliferation of HCC cell lines.

CCNA2 induces cell cycle arrest and apoptosis

Given CCNA2’s role in the cell cycle, flow cytometry was used to examine its impact on cell cycle progression. The analysis indicated a notable increase in the proportion of cells in the G2 phase following CCNA2 silencing, suggesting potential cell cycle arrest at the G2/M transition (Fig. 9K). To support this finding, Western blot analysis was performed on proteins that regulate the G2/M phase, revealing a decrease in Cyclin-Dependent Kinase 1 (CDK1) levels (Fig. 9L). Additionally, the assay analyzed apoptosis-related proteins, including Bax, Bcl2, and P53 (Fig. 9M). Collectively, these results indicate that silencing CCNA2 may suppress tumor proliferation by inducing cell cycle arrest in the G2/M phase and enhancing apoptosis.

CCNA2 is involved in EMT activation in HCC lines

Given the critical role of epithelial-mesenchymal transition (EMT) in conferring invasive and migratory capabilities to tumor cells, the study examined the protein expression of EMT-related markers. Consistent with expectations, CCNA2 knockdown resulted in the attenuation of protein expression levels of N-cadherin, vimentin, MMP7, and MMP9, alongside an increase in E-cadherin expression (Fig. 9N-O). These findings suggest that CCNA2 plays a mediatory role in EMT processes in vitro.

CCNA2 activates the PI3K/AKT signal pathway in HCC cell lines

Given the established correlation between EMT and the PI3K/AKT pathway, and the previous findings indicating a close relationship between CCNA2 and the PI3K/AKT pathway (Fig. 3B), the study focused on examining the protein expression levels of genes within this pathway. The findings demonstrated a significant inhibition in the protein expression levels of phospho-PI3K and phospho-AKT, while the overall protein levels of PI3K and AKT remained unaltered following CCNA2 silencing (Fig. 9P). This suggests that CCNA2 might facilitate the activation of EMT in hepatocarcinogenesis through its mediation of the PI3K/AKT pathway.

Discussion

Primary liver cancer is the third leading cause of cancer-related deaths worldwide, with approximately one million new cases annually. Hepatocellular carcinoma (HCC) accounts for nearly 90% of these cases, often linked to the progression of various chronic liver diseases [14]. Recent significant advancements in liver cancer treatment, particularly systemic therapeutic agents, have shifted the paradigm in HCC management [15, 16]. Although nearly half of the global HCC cases are reported in China, incidence and mortality rates have been declining due to sustained government efforts at various levels. However, given China’s vast population and the low survival rate of 12.1%, liver cancer continues to pose a significant challenge [17]. Additionally, due to the insidious nature of HCC, over half of the patients are diagnosed at an advanced stage, at which point systemic therapy is the only treatment option that offers survival benefits [18]. Consequently, identifying biomarkers to predict treatment efficacy has become increasingly critical. The discovery of disulfidptosis, a significant form of cell death identified following the characterization of cuproptosis in 2022 by Xiaoguang Liu et al., may open new avenues in cancer therapy. Unfortunately, research into disulfidptosis in HCC remains scant. This study aims to explore the potential relationships and mechanisms linking disulfidptosis to HCC using bioinformatics, setting the stage for future molecular experiments.

By staging HCC, patients can be stratified for individualized treatment plans tailored according to the disease stage. This study initially identified three distinct HCC subtypes through the integration of DRGs from the GEO and TCGA datasets, subtypes that existing clinical criteria do not classify. The variation in survival cycles among these subtypes suggests divergent pathological mechanisms.

GSVA analysis revealed that pattern B exhibited enrichment in various cancer types. After this, a differential analysis of the three HCC patterns was conducted, yielding 221 DEGs. Following this, the 5-DRG prognostic model comprising PBK, SERPINE1, CCNA2, SLC6A1, and STC2 was developed using LASSO regression and multifactorial analysis on the 221 genes with prognostic significance. This 5-DRG prognostic model has demonstrated robust efficacy in predictive, immunological, and pharmacological aspects of treatment. Initially, HCC patients were stratified into Low-Risk Groups (LRGs) and High-Risk Groups (HRGs) based on risk scores, revealing that patients in LRGs exhibited significantly higher survival rates compared to those in HRGs. Crucially, these findings were validated in both the TCGA-GEO cohort and three independent HCC datasets (GSE14520, ICGC, and GSE54236). Concurrently, potential therapeutics for HCC were identified through drug sensitivity analysis, providing valuable insights for clinicians in treating this condition.

Transarterial chemoembolization (TACE) is the preferred treatment for patients with intermediate to advanced HCC and is acknowledged as a prevalent non-surgical intervention for HCC [19]. However, the tumor response rate to a single conventional TACE session is only 52% [20], and a progressive decline in tumor responsiveness is observed following repeated TACE procedures. Research indicates that sorafenib contributes to extending tumor progression-free survival (PFS) and overall survival (OS) in patients with TACE-resistant HCC [21]. The 5-DRG prognostic model effectively predicts the responsiveness to TACE and sorafenib, potentially reducing medical overuse and patient costs. Furthermore, the model exhibits substantial clinical utility. Univariate and multifactorial analyses have demonstrated that risk scores independently influence patient prognosis, beyond current risk factors like gender, age, and stage. Moreover, the clinical nomogram constructed for HCC patients assists physicians in predicting their survival rates. The advent of immunotherapy marks a new chapter in liver cancer treatment [22]. The study revealed that common immune checkpoints are significantly more upregulated in HRGs than in LRGs, indicating that patients in HRGs may derive greater benefit from immunotherapy. Additionally, Cyclin A2 (CCNA2) was identified through machine learning analysis, confirming its significant role in the tumor and immune microenvironment. Furthermore, it was observed that high CCNA2 expression is closely associated with a poorer prognosis and more advanced clinical stages. Moreover, the role of CCNA2 was further validated through silencing experiments, suggesting that CCNA2 could be a promising candidate for the treatment of hepatic malignancies.

Existing evidence suggests that five genes (PBK, SERPINE1, CCNA2, SLC6A1, and STC2) influence various tumor processes via diverse mechanisms. PBK overexpression has been linked to poor prognosis in lung [23], colorectal [24], and gastric cancers [25]. Yang et al. reported that PBK overexpression is associated with enhanced HCC proliferation, migration, and invasion, both in vitro and in vivo [26]. Furthermore, the prognostic significance of PBK has been well established through various bioinformatic analyses [27,28,29]. Elevated SERPINE1 expression has been demonstrated to correlate with poor prognosis in various tumors [30,31,32], potentially linked to its role in tumor microenvironment remodeling and immune cell infiltration. SLC6A1, notably the only gene among them with reduced expression in hepatocellular carcinoma tissue, has been researched in neurological disorders [33]. Its regulatory role in hepatocellular carcinoma remains to be elucidated, warranting further investigation. STC2 is well-documented to be upregulated in HCC tissues, and this upregulation is associated with shorter survival cycles [34,35,36]. In the era of big data, the importance of machine learning in extracting insights and making effective predictions from extensive datasets is paramount [37]. Random Forest, a supervised machine-learning algorithm based on classification trees, was developed by Breiman in 2001 [38]. The LASSO Cox analysis, initially proposed by Robert in 1996 [39], is a linear regression method employing L1 regularization and has become widely utilized in predictive models. The issue of covariance in LASSO regression has been addressed with the incorporation of Random Forest. Through the combined use of Random Forest and LASSO regression analysis, the crossover gene CCNA2 was identified. CCNA2, located in the Q27 region of human chromosome 4 and spanning 7489 bp, is expressed in nearly all human tissues and is classified as a cell cycle-related protein [40]. Studies have shown that CCNA2 is relevant to cytoskeletal dynamics and cell movements [41]. Our study found that CCNA2 was distinctly identified in the development and progression of HCC.

Initially, it was found that overexpression of CCNA2 is associated with advanced HCC stages and unfavorable outcomes, as evidenced by data from the TCGA and GEO databases. Subsequently, observations revealed that both the transcription and translation levels of CCNA2 are significantly higher in cancer cells and tumor tissues compared to normal tissues and cells. This observation was further corroborated by immunohistochemical analysis. The literature review indicated that reports on CCNA2 in HCC are scarce, predominantly comprising bioinformatic analyses with a lack of experimental validation. Consequently, the SNU-449 cell line, exhibiting the highest CCNA2 expression among HCC cell lines, was chosen for functional experiments to elucidate the biological role of CCNA2 in LIHC. Within the HCC tumor microenvironment, immune cells, including regulatory T cells and bone marrow-derived suppressor cells, along with certain factors, contribute to immunosuppression and tumor escape, thereby facilitating the progression of HCC. Additionally, the depletion of T prolif cells may further result in the diminished capacity of immune cells to clear malignant cells [42, 43]. Single-cell technology enables the delineation of cell community composition in diseased tissues and the localization of malignant and microenvironmental cells within tumors, thus enhancing clinical diagnosis and treatment strategies [44]. Therefore, we investigated the cellular subpopulations affected by CCNA2 at the single-cell level and discovered that CCNA2 predominantly accumulates in Tprolif and malignant cells, suggesting its potential role in promoting liver cancer progression through T-cell depletion.

The epithelial-mesenchymal transition (EMT) is intricately associated with tumor invasion and metastasis and is modulated by various signaling pathways, notably PI3K/AKT [45]. This study revealed that silencing CCNA2 in SNU-449 cells significantly increased the expression of E-cadherin, while reducing the expression of mesenchymal markers such as N-cadherin, Vimentin, and MMP9, thereby suggesting that CCNA2 silencing may inhibit the EMT process in these cells. Concomitantly, KEGG and GSVA analyses associated the PI3K/AKT signaling pathway with CCNA2 expression. Subsequent experimental investigations demonstrated that disruption of CCNA2 expression led to a significant decrease in PI3K/AKT pathway-related proteins, indicating that CCNA2 might influence hepatocellular carcinoma cell proliferation and the EMT process via this signaling pathway.

The strengths of our study include, firstly, the development of the 5-DRG prognostic model, which not only predicts immunological and clinical outcomes but also personalizes and refines patient treatment, and secondly, the validation of CCNA2’s adverse prognostic impact in HCC and the identification of the PI3K/AKT signaling pathway as a potential target of CCNA2. However, our study also presents certain limitations. First, the validation of this prognostic model on a larger scale requires a more extensive sample size. Second, there is a need for a larger sequencing database of HCC patients for further validation. Regarding CCNA2, a gap exists in systematic animal-level studies. Third, the precise relationship between CCNA2 and the PI3K pathway requires further verification. Future investigations will focus on exploring the molecular mechanisms by which CCNA2 mediates disulfidptosis, and to demonstrate that CCNA2 is capable of hindering liver cancer metastasis. Despite these limitations, the bioinformatics analysis and experimental validation of CCNA2 as a novel biomarker for hepatocellular carcinoma pave the way for further research into HCC mechanisms and drug development.

Conclusions

We developed a prognostic model based on five DRGs, demonstrating remarkable sensitivity in predicting the progression of HCC, and highlighted CCNA2 as a promising candidate for HCC diagnosis.