This study was conducted in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All animal experiments were approved by the Ethics Committee of the Third Affiliated Hospital of Naval Medical University (approval number: EHBHKY2014-03-006). Regarding human samples, the study was approved by the Institutional Review Board of Ethics Committee of the Third Affiliated Hospital of Naval Medical University (approval number: EHBHKY2023-KO39-P001). All procedures performed in studies involving human participants followed the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. Informed consent was obtained from all individual participants included in the study. All samples were anonymized to respect the privacy of participants.

The Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/, accessed on 23 August 2023) provided the GSE101685 microarray dataset, which comprises eight groups of normal liver tissues and 24 groups of HCC tissues. The probes were then summarized using Affymetrix annotation files using the “Affy” R package’s median polished probe set. DEGs were identified using the GEO2R tool, with a fold change (FC) >2 or <0.5 and p < 0.05 used to identify up-regulated DEGs and down-regulated DEGs.

The co-expression network of DEGs in the GSE101685 dataset was created using the Weighted gene co-expression network analysis (WGCNA) program in R. We first calculated the paired gene Pearson correlation matrix and then used the power function am = |cmn| to get the weighted adjacency matrix. Using the soft threshold power β, we identified weak and strong connections between genes. We also generated a topological overlap matrix (TOM) using the adjacency matrix. Genes with similar expression patterns were assigned to modules using ALM-based average linkage hierarchical clustering. The key module was chosen based on the strongest association between GSE101685 samples and gene modules.

RNAseq data was obtained from 371 individuals diagnosed with liver hepatocellular carcinoma (LIHC) through The Cancer Genome Atlas (TCGA; https://tcga-data.nci.nih.gov/tcga, accessed on 25 August 2023) database. Subsequently, the “forestplot” program was utilized to conduct a batch survival analysis for the genes in the key module, and the effect of each gene on disease-free survival (DFS) was visualized. Statistical significance was determined using p-values less than 0.05. The log-rank test and univariate Cox regression were used to calculate the hazards ratios (HR) and their 95% confidence intervals (CI).

The “glmnet” package in R software was utilized to investigate 56 genes in the turquoise module. The least absolute shrinkage and selection operator (LASSO) model was tuned using ten-fold cross-validation to determine the tuning parameters. The genes with the minimum criterion of tunning parameter (λ) were identified as the most predictive. The selected genes represent the most statistically significant predictors in the GSE101685 dataset and form the basis of the prognostic model. The LIHC cohort from the TCGA database was then separated into two categories according to the expression pattern of significant genes: high-risk and low-risk, respectively. Both cohorts underwent a risk assessment, and the DFS probability was calculated using a Kaplan-Meier (KM) analysis for each data set. In addition, the median survival time was calculated, and a log-rank test was employed to examine the statistical significance of the survival differences between the two cohorts, yielding a p-value. To provide additional context for the comparative risk, the HR for the high-risk group was also calculated. Lastly, the predictive ability of the risk model for patient 1-, 3-, and 5-year survival was assessed using the “timeROC” program, and the Area Under Curve (AUC) analysis was used to create the receiver operating characteristic (ROC) curve. It is worth noting that a higher AUC value suggests better prognosis-predicting ability.

After selecting nine genes with prognostic significance from our risk model, both univariate and multivariate Cox regression analyses were applied to these genes and clinical factors, including age, pT stage, pTNM stage, and tumor grade. Utilizing the “forestplot” package within R, these analyses facilitated the graphical representation of p-values, hazard ratios (HRs), and their 95% confidence intervals (CIs). For variables demonstrating significance (p < 0.05), a nomogram was developed to assess their influence on patient survival outcomes.

TGF-β is known to act as a tumor suppressor in the early stages of HCC [16] by inhibiting cell proliferation and promoting apoptosis, thus preventing the initiation and progression of this malignancy [17]. In this study, we evaluated CYB5D2 expression levels in both normal liver and TCGA-LIHC samples through the Wilcoxon test, followed by an analysis of the association between CYB5D2 and TGF-β. Additionally, TGF-β expression in TCGA-LIHC samples was assessed using the Wilcoxon test to explore its role in LIHC pathogenesis. Finally, the impact of TGF-β expression variations on patient OS was determined via KM survival analysis.

The HPA (https://www.proteinatlas.org/, accessed on 5 September 2023) database provides a vast collection of immunohistochemical images of human proteins in different tissues and cells, allowing for the exploration of protein distribution and expression patterns. In this database, immunohistochemical analysis was performed to assess the protein expression levels of CYB5D2 and TGF-β in both normal liver and HCC tissues.

Shanghai Institute of Cell Biology (Shanghai, China) provided the human HCC cell lines C3A (obtained from ATCC), HepG2, MHCC97H, and Hep3B, as well as the human normal liver cell line QSG7701. These cells were grown in RPMI 1640 media with 10% fetal bovine serum (FBS) added, and they were kept at 37°C in a 5% CO₂ humidified environment. The CYB5D2 overexpression plasmid was transfected into C3A and HepG2 cell lines for the transfection process, while the empty vector plasmid served as the control. Additionally, to investigate relevant cellular responses, 5 ng/mL of the active form of TGF-β was added to the medium of C3A and HepG2 cell lines. The transfection process utilized Lipofectamine 3000 (Invitrogen, Carlsbad, CA, USA), adhering strictly to the provided guidelines.

Total RNA was isolated from cells using TRIzol reagent (Invitrogen), following the accompanying protocol. RNA concentration and purity were assessed using a NanoDrop 2000 spectrophotometer. cDNA synthesis was conducted using the PrimeScript RT Reagent Kit (Takara Bio, Otsu, Japan), as per the manufacturer’s directions. The same kit was also used for reverse transcription. qRT-PCR was carried out using SYBR Green PCR Master Mix (Takara) and specific primers for CYB5D2, TGF-β, E-cadherin, N-cadherin, Snail, and Twist, detailed in Table S1. GAPDH served as the internal control, and gene expression levels were determined by the 2^−ΔΔCt method.

Protein lysates from cells were prepared using RIPA lysis buffer (Beyotime, Shanghai, China) supplemented with a protease inhibitor cocktail and a phosphatase inhibitor cocktail (both from Roche). Protein concentrations were quantified using the BCA protein assay kit from Thermo Fisher Scientific. Equal amounts of proteins were then subjected to separation via 10% SDS-PAGE gels, followed by their transfer onto PVDF membranes. After blocking for an hour with 5% nonfat milk in TBST, the membranes were incubated overnight at 4°C to be incubated with primary antibodies CYB5D2 (Invitrogen, #PA5-69862), TGF-β (Invitrogen, #MA5-15065), E-cadherin (Invitrogen, #13-1700), N-cadherin (Invitrogen, #33-3900), Snail (Invitrogen, #MA5-14801), and Twist (Invitrogen, #PA5-49688), they were diluted at a 1:1000 dilution. Horseradish peroxidase-conjugated secondary antibodies were then incubated on TBST-washed membranes for one hour at room temperature. Protein bands were visualized and quantified using the Millipore ECL Plus kit and Image J software.

To evaluate the expression level of CYB5D2, IHC analysis was performed on tissue samples obtained from the Third Affiliated Hospital of Naval Medical University, including HCC and surrounding normal liver tissue. The sections were rehydrated in a graded ethanol series after being deparaffinized in xylene. Antigens were extracted from citrate buffer (pH 6.0) using microwave heating. 3% hydrogen peroxide and 5% bovine serum albumin were used to block nonspecific binding and endogenous peroxidase activity, respectively. Sections were treated with a 1:1000 solution of a primary antibody specific to CYB5D2 for an entire night at 4°C. Hematoxylin staining was used to visualize the nuclei. Finally, stained slices were imaged to evaluate CYB5D2 expression between tumor and normal tissue.

Cellular proliferation was assessed using the CCK-8 test (Dojindo). Transfected cells were cultured in 96-well plates for 24, 48, 72, and 96 h each. After that, each well received a 10 μL aliquot of CCK-8 solution and incubated at 37°C. The absorbance was then measured at 450 nm using a microplate reader.

For the cell cycle assessment, cells were collected, rinsed with chilled phosphate-buffered saline (PBS), and preserved in 70% ethanol at −20°C overnight. Post-fixation, cells were washed with PBS and incubated with propidium iodide (Sigma-Aldrich) in darkness for 30 min to stain DNA. Subsequently, DNA content in stained cells was quantified using a BD Biosciences flow cytometer. Data acquired from the cytometer was processed and analyzed using FlowJo software to determine cell cycle distribution.

In transwell chambers with or without Matrigel (Corning Inc., Corning, NY, USA), cell migration and invasion were assessed. The trypsinized cells were resuspended in a medium devoid of serum after being cleaned. 3 × 10⁴ cells were injected into the upper chamber of each Transwell. The lower chamber was filled with 200 μL of serum-free media and 600 μL of chemoattractant-containing medium with 10% FBS. Following a 24-h incubation period, cells passing through the bottom wells of the membrane were fixed with 4% paraformaldehyde and stained with crystal violet. Five randomly chosen areas inside each well were counted for the quantity of migrating and invasive cells under the microscope.

The Ethics Committee of the Third Affiliated Hospital of Naval Medical University approved all animal treatments (approval number: EHBHKY2014-03-006). Six-week-old male BALB/c nude mice were obtained from SLAC Laboratory in Shanghai, China. In their lateral axillary region of the forelimb near the back, the mice received subcutaneous injections of either empty control C3A cells or C3A cells that overexpressed CYB5D2. After 2–3 weeks, the size and weight of the mouse tumors were measured and compared before and after injection to evaluate the impact of CYB5D2 overexpression on tumor development. Following the final measurement, the mice were euthanized 7 days after the last treatment using CO₂ inhalation at a flow rate of 20% chamber volume per minute until loss of consciousness, followed by cervical dislocation. All euthanasia procedures adhered to the guidelines of the Ethics Committee of the Third Affiliated Hospital of Naval Medical University and followed the American Veterinary Medical Association (AVMA) Guidelines for the Euthanasia of Animals.

Data processing was conducted utilizing R language packages, employing either Student’s t-test or one-way Analysis of Variance (ANOVA) for statistical analysis. A threshold of p < 0.05 was set for significance determination. The data was provided as mean ± SEM.

A total of 1033 DEGs from the GSE101685 dataset using the GEO2R tool. Of these, 363 were up-regulated DEGs, and 670 were down-regulated DEGs (Fig. 1A). The probes were then placed into modules using the WGCNA package on R through average linkage clustering (Fig. 1B–E). A power of β = 14 (R² = 0.85 without scale) was selected as a soft threshold, aiming to ensure a scale-free network, resulting in the identification of 2 modules (excluding the grey module). After comparing the correlation between the two modules and the two sets of samples in the GSE101685 dataset, it was discovered that the turquoise module (consisting of 452 genes) had the highest correlation with both sets of samples, with a correlation coefficient of 0.801.

After identifying the turquoise module as a key module, we conducted a DFS prognostic analysis on the genes within this module. As evident from the results in Table S2, a total of 56 genes showed significant prognosis. Subsequently, we performed LASSO Cox regression analysis on these 56 genes and identified 9 signature prognostic genes according to the optimal lambda value (lambda.min = 0.0688) (Fig. 2A,B) (The genes shown in (A) correspond to the genes in the provided Table S2.). The risk score for these genes was determined as follows: Riskscore = (−0.0212)*CFHR4 + (−0.0405)*SPP2 + (−0.0021)*FBLN5 + (−0.0157)*ANO1 + (−0.0261)*STEAP4 + (−0.0012)*GLUD1 + (−0.0336)*CYB5D2 + (−0.0321)*LCAT + (−0.2391)*IL18RAP. According to the risk model study, the high-risk group outperformed the low-risk group in terms of survival and death (Fig. 2C). KM survival analysis results, presented in Fig. 2D, elucidated that the median survival time for the high-risk group was notably shorter, at 1.2 years, compared to 3.9 years for the low-risk group, with an HR of 2.287 (>1), indicating a significant reduction in DFS probability for the high-risk group. The ROC curve analysis further accentuated the predictive accuracy of the risk model, especially pronounced at the 5-year mark (AUC = 0.729) (Fig. 2E). Our research concluded by highlighting the important prognostic potential of these nine genes.

In the analysis of nine genes and four clinical variables in the risk model, we identified five variables with statistical significance (p < 0.05), namely CFHR4, CYB5D2, IL18RAP, pT stage, and pTNM stage (Fig. 3A,B). The nomogram analysis results indicated that these variables had significant predictive power on the 1-, 3-, and 5-year survival rates of patients, which was also confirmed by the calibration curve results (Fig. 3C,D).

From the prognostic genes identified through nomogram analysis, CYB5D2, CFHR4, and IL18RAP have already been substantiated as participating in the pathogenesis of HCC. In this investigation, we chose CYB5D2 as the hub gene to examine its relationship to HCC. Fig. 4A illustrated a noticeable underexpression of CYB5D2 in LIHC, hinting towards a potential role in disease manifestation. These findings are further confirmed by immunohistochemistry results. Compared to normal liver tissue, the staining intensity of CYB5D2 in tumor tissue was weakened, indicating a lower expression level in tumor tissue (Fig. 4B). Correlation analysis revealed a significant negative association (r = −0.65) between CYB5D2 and TGF-β (Fig. 4C). To further substantiate our findings, we employed the Wilcoxon test, which confirmed TGF-β overexpression in LIHC (Fig. 4D,E). TGF-β overexpression was associated with a poorer OS prognosis, suggesting a potential prognostic role for TGF-β in LIHC.

By using the qRT-PCR and WB method in vitro experiments, we detected the CYB5D2 and TGF-β expression levels in HCC cells. In C3A and HepG2, CYB5D2 mRNA and protein levels were found to be much lower than in QSG7701 (Fig. 5A,B). In contrast, qRT-PCR and WB confirmed that C3A and HepG2 cells had elevated TGF-β expression levels (Fig. 5C,D). The HPA database also analyzed the levels of CYB5D2 and TGF-β in HCC tissues. Among them, no staining signal was detected for CYB5D2 in tumor tissue, and TGF-β showed high staining intensity in tumor tissue (Fig. 5E,F). This suggested that whilst TGF-β expression is markedly elevated in tumor tissues, CYB5D2 expression is low in malignant tissues. Therefore, we speculated that the downregulation of CYB5D2 and the upregulation of TGF-β may be related to the malignant progression of HCC.

In our study, overexpression experiments were conducted by transfecting the CYB5D2 plasmid into C3A and HepG2 cells. qRT-PCR analyses confirmed the upregulation of CYB5D2 mRNA level post-transfection (Fig. 6A). Subsequently, the impact of CYB5D2 overexpression on cell proliferation was assessed using the CCK-8 assay. Results demonstrated a significant reduction in the proliferation rates of both C3A and HepG2 cells following CYB5D2 overexpression (Fig. 6B,C). Further analysis, as depicted in Fig. 6D–G, revealed that CYB5D2 overexpression led to cell cycle arrest in the G1 phase, with an increase from 57.79% to 65.69% in C3A cells and from 55.67% to 62.21% in HepG2 cells. These findings collectively implied that CYB5D2 overexpression inhibits cell proliferation and induces cell cycle arrest, providing insight into the potential therapeutic implications of targeting CYB5D2 in HCC.

Transwell studies were then used to assay the effect of CYB5D2 overexpression on the migration and invasion of C3A and HepG2 cells. The group with CYB5D2 overexpression exhibited a significant decrease in the number of migrating and invading cells compared to the control group (Fig. 7A–D). It was demonstrated that overexpression of CYB5D2 prevented the migration and invasion capabilities of C3A as well as HepG2 cells.

The EMT plays a pivotal role in the development and metastasis of HCC [18,19], with key markers such as E-cadherin, and N-cadherin being closely linked to tumor progression [20]. To investigate the regulatory roles of CYB5D2 and TGF-β on these EMT-associated factors, we segregated C3A and HepG2 cells into four groups: control (over-NC), CYB5D2 overexpression (over-CYB5D2), TGF-β, and CYB5D2 overexpression with TGF-β (over-CYB5D2+TGF-β). qRT-PCR analysis revealed a significant increase in E-cadherin expression in the over-CYB5D2 group as compared to the control. However, with the addition of TGF-β, its expression level decreased. Conversely, the expression levels of N-cadherin, Snail, and Twist were downregulated by CYB5D2 overexpression, with their levels rising following TGF-β treatment, as evidenced by qRT-PCR and corroborated by WB analysis (Fig. 8A–C). Based on these findings, we conducted Transwell rescue experiments and discovered that overexpression of CYB5D2 significantly reduced the migratory and invasive behavior of C3A and HepG2 cells compared to the control group. The addition of TGF-β partially counteracted this inhibitory effect (Fig. 8D–G). Thus, our results suggested that CYB5D2 inhibited both the migration and invasion of HCC cells and modified the expression of EMT-related proteins. TGF-β might play a role in reversing the inhibitory effects of CYB5D2 on C3A and HepG2.

The study evaluated the in vivo effects of overexpressing CYB5D2 were assessed using a xenograft model of HCC in nude mice. The results showed a significant reduction in tumor size and weight after CYB5D2 overexpression in the nude mouse model (Fig. 9A,B). These findings underscored the impact of CYB5D2 modulation on tumorigenesis in HCC.