Development of a four-gene prognostic signature for hepatocellular carcinoma using TCGA and GEO datasets

Abstract

Background

Hepatocellular carcinoma (HCC) poses a major global health challenge due to its high mortality and recurrence, necessitating continuous research into biomarkers for better diagnosis and treatment.

Methods

Gene expression and clinical data from HCC patients were sourced from TCGA, with the GSE14520 dataset used for validation. Differentially expressed genes (DEGs) between HCC and normal tissues were identified. Patients were randomly divided into training and testing groups using a repeated sampling strategy, and various regression analyses were conducted to develop a prognostic gene signature. The four-gene signature’s prognostic ability was assessed using ROC curves, Kaplan–Meier analysis, multivariate Cox regression, and a nomogram, with further validation in the GSE76427 and ICGC cohorts.

Results

A four-gene signature (EFNA4, MMP1, NEIL3, TXNRD1) was developed for predicting HCC prognosis. The GSE14520 database confirmed these genes were overexpressed in HCC tissues (p < 0.001). Immunohistochemical (IHC) analysis revealed that EFNA4, MMP1, and TXNRD1 protein levels were significantly elevated in HCC compared to nearby normal tissues (p < 0.05), and the risk score successfully pinpointed high-risk patients with lower overall survival (p < 0.001). As an independent prognostic factor for HCC, the risk score was used alongside the stage to develop a nomogram, which surpassed the predictive ability of each factor individually for 1-, 3-, and 5-year survival rates (AUCs > 0.7). The nomogram’s predictions matched observed outcomes well in the entire cohort. Functional analysis linked cell cycle and pyrimidine metabolism pathways to HCC development.

Conclusion

Our study developed a new four-gene model combined with clinical parameters to effectively predict HCC prognosis, offering a reliable tool for clinicians and a foundation for future mechanistic investigation.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05244-1.

Introduction

Globally, hepatocellular carcinoma (HCC) is the most prevalent primary liver cancer, ranking sixth in terms of diagnosis and third in cancer-related mortality [1]. It poses challenges due to late detection and high recurrence. Therefore, the early detection and diagnosis of HCC through the use of specific and sensitive biomarkers, alongside the implementation of novel therapeutic approaches, are crucial for enhancing patient outcomes.

Screening for HCC is aimed at high-risk groups, like people with hepatitis virus, cirrhosis, or a family history of the disease [2]. For diagnosing HCC in patients with HBV infection and cirrhosis, the Liver Imaging Reporting and Data System (LI-RADS) has a sensitivity of 67% and a specificity of 93% for tumors larger than 1 cm [3, 4]. Biopsies are required for individuals without known risk factors or with atypical imaging results. While biomarkers like alpha-fetoprotein (AFP), des-gamma-carboxy prothrombin (DCP), and glypican-3 (GPC3) might improve outcomes [5], only about half of HCC cases are diagnosed early enough for curative treatment. Increasing attention is being given to using multiple markers for diagnosing HCC. Song et al. [6]. discovered that PIVKA-Ⅱ by itself is highly effective for diagnosing HCC, and its sensitivity increases when used alongside AFP. However, adding a third marker, GPC3, doesn’t improve diagnosis further. Wang et al. [7]. suggest that combining AFP, AFP-L3%, and DCP could be beneficial for HCC diagnosis and screening. The study by Xie et al. [8]. found that AFP, α-L-Fucosidase (AFU), and PIVKAII collectively improve the early detection accuracy of hepatitis B-related HCC, notably in patients suffering from chronic hepatitis B and cirrhosis. This indicates that uncovering new genes related to prognosis could result in prognostic models that are more accurate than those based on conventional clinical parameters.

In this research, we created a model incorporating several prognostic genes and clinical factors to forecast overall survival (OS) in patients with HCC. Using univariate Cox regression and the Least Absolute Shrinkage and Selection Operator (LASSO) method, we discovered new biomarkers and developed a four-gene signature using TCGA data. We validated the expression of the four prognostic genes using external mRNA datasets, public protein data, and in-house IHC, and computed risk scores with multivariate Cox coefficients. The model was validated with International Cancer Genome Consortium (ICGC) HCC and GEO’s GSE76427 data. We combined risk scores with clinical parameters to build a nomogram, evaluated by calibration plots and time-dependent receiver operating characteristic curve (tROC) analysis. Lastly, we conducted functional enrichment analysis to explore mechanisms linked to the four prognostic genes.

Materials and methods

Sources and processing of datasets

In August 2022, data from RNA sequencing and clinical details for 406 HCC samples and 58 nearby non-cancerous tissue samples were retrieved from the TCGA database (https://portal.gdc.cancer.gov/). The original raw count data were adjusted using the transcripts per million (TPM) approach and then transformed using log2. A total of 59,427 protein-coding genes were identified through a comprehensive annotation process. DEGs were determined using the ‘Limma’ package in R [9], characterized by an absolute log2 fold change above 2 and an FDR below 0.05.

Creation of the prognostic gene signature

HCC patients without complete clinical information were omitted from the study. Univariate Cox regression analysis was performed on statistically significant DEGs to determine their relationship with OS, identifying prognostic genes with a p-value under 0.01. Afterward, patients were randomly assigned to training and testing groups in a 1:1 ratio using the ‘caret’ package in R [10]. To ensure a robust and generalizable model, the random splitting process was repeated 1000 times. The final dataset partition used for model development was selected based on the optimal predictive performance (highest average AUC for 1-, 3-, and 5-year overall survival) of the resulting gene signature in the corresponding testing set. To further refine the selection of prognostic genes related to OS in HCC patients, LASSO penalized Cox regression was employed [11]. Prior to LASSO analysis, the mRNA expression levels of the candidate prognostic genes were standardized (Z-score normalization) to ensure equitable treatment of all features by the penalty term. LASSO penalized Cox regression with ten-fold cross-validation was employed to further refine the selection of prognostic genes. The optimal penalty parameter (lambda, λ) was determined as the value that yielded the minimum partial likelihood deviance (lambda.min) across the cross-validation folds. A prognostic gene signature was created by linearly combining LASSO Cox regression coefficients with mRNA expression levels, and the risk score was determined by summing the products of each gene’s expression and its corresponding coefficient. The R packages ‘survival’ and ‘survminer’ were utilized to find the optimal cut-off value through a two-sided log-rank test. Based on the median risk score, patients were divided into high-risk and low-risk categories. Using the ‘survivalROC’ R package, a time-dependent ROC curve was generated to assess the predictive value of the prognostic gene signature for OS [12]. To compare survival differences between high-risk and low-risk groups, Kaplan-Meier survival analysis and log-rank test were conducted using the ‘survival’ package in R. Following this, the predictive capability of the prognostic gene signature was evaluated in the testing cohort as well as the whole cohort.

External validation of the prognostic gene signature

The GSE76427 dataset, encompassing gene expression profiles and clinical data for HCC, was procured from the Gene Expression Omnibus (GEO) database, accessible via the National Center for Biotechnology Information (NCBI) at https://www.ncbi.nlm.nih.gov/geo/. Additionally, RNA sequencing data and clinical information for a separate cohort of 231 HCC samples were retrieved from the ICGC portal, found at http://dcc.icgc.org/releases/current/.

In the GSE76427 and ICGC datasets, the risk score for each HCC patient was assessed using a consistent prognostic gene-signature model. Afterward, the gene signature’s predictive efficacy was assessed through tROC analysis and Kaplan–Meier survival curves.

The prognostic gene signature demonstrates an independent prognostic role

Univariate and multivariate Cox regression analyses, using a forward stepwise approach, were conducted to determine if the prognostic gene signature functions independently of clinical factors like age, gender, tumor grade, and stage. A p-value of less than 0.05 was used to determine statistical significance.

Developing and validating a predictive nomogram

Based on the Cox regression analysis outcomes, a nomogram was created to estimate the 1-, 3-, and 5-year OS rates for HCC in the training cohort, combining the risk score with the disease stage. To evaluate the nomogram’s efficacy, calibration curves and tROC curves were analyzed using R software. This evaluation procedure was systematically applied to both the testing cohort and the whole cohort.

Gene set enrichment analysis

GSEA was performed to clarify possible biological pathways linked to prognosis-related genes included in the nomogram [13]. Samples from TCGA-HCC were categorized into two cohorts according to the median gene expression levels. The analysis was conducted using GSEA software (version 4.1.0) on JAVA 8.0, with significance criteria set at p < 0.05 and FDR q < 0.05.

Verification of the prognostic gene signature using immunohistochemistry analysis

Expression validation was performed at multiple levels. First, mRNA overexpression was confirmed using the GSE14520 dataset. For protein-level validation, we utilized both in-house immunohistochemistry (IHC) and publicly available protein expression data from The Human Protein Atlas (HPA) database.

Nineteen HCC tumor and adjacent normal tissue samples were collected from Jiangxi Cancer Hospital, China, between February and October 2023. Patients had confirmed HCC diagnoses and met criteria: HCC diagnosis, no pre-surgery chemotherapy/radiation, detailed hospitalization records, and residency in Nanchang for over 5 years. Exclusions included metabolic/infectious diseases, other tumor history, prior radiotherapy/chemotherapy, and incomplete clinical data. Samples were split: one part for paraffin-embedding and histopathology, the other frozen at − 80 °C. The tissue sections were prepared, dewaxed, hydrated, subjected to antigen retrieval, treated with hydrogen peroxide, washed, and blocked using BSA. Anti-EFNA4, anti-MMP1, and anti-TXNRD1 antibodies (1:100) were incubated overnight at 4 °C (Note: A validated antibody for NEIL3 suitable for IHC was not available in our laboratory; thus, NEIL3 protein expression was assessed using the HPA database). Detection involved the streptavidin peroxidase method with diaminobenzidine (DAB) staining and hematoxylin counterstaining.

The Visiopharm software from Denmark was utilized to determine tissue regions and compute the histochemistry score (H-Score) for IHC scan evaluation. Tissue staining is quantified using the H-Score method, which turns the number and intensity of positively stained cells into numerical values. The formula H-Score = ∑(pi × i) is used to calculate the H-Score, where ‘i’ corresponds to the staining intensity (0 for negative, 1 for light yellow, 2 for brown-yellow, 3 for tan) and ‘pi’ is the percentage of cells that are positive. The H-score, which can be between 0 and 300, measures the degree of positive intensity, with higher values showing more positivity. To calculate the positivity rate, divide the count of positive cells by the total cell count.

Protein expression data for EFNA4, TXNRD1, and NEIL3 in HCC and normal liver tissues were also retrieved from The Human Protein Atlas (HPA, www.proteinatlas.org) for supplementary validation.

Statistical analysis

R software (version 3.6.3) was used to conduct statistical analyses. Tumor tissue samples were allocated into two groups through random selection using the “sample” function. A heatmap, generated with the ‘pheatmap’ package and zero-mean normalization, was used to visualize DEGs. PCA was applied using the ‘ggfortify’ package to address batch effects and aid clustering [14]. Paired t-tests were used to evaluate quantitative variables for paired samples, while Wilcoxon rank-sum tests were applied to unpaired samples. Kaplan-Meier curves were used for survival analyses, and p-values and hazard ratios (HR) with 95% confidence intervals (CI) were calculated using log-rank tests and univariate Cox regression. All claims of significant differences are accompanied by specific p-values. Unless mentioned otherwise, a p-value below 0.05 was taken as a sign of statistical significance.

Results

Identification of DEGs

Initially, 59,427 protein-coding genes from TCGA-HCC RNA-sequencing data were analyzed. PCA demonstrated a significant divergence in transcript levels between cancerous and adjacent non-cancerous tissues, with distinct clustering observed between tumor and normal groups (Fig. 1A). Following this, 1,600 DEGs were identified (Supplementary Table S1), comprising 1,543 upregulated and 57 downregulated genes (Fig. 1B). A heatmap of the top 20 DEGs was presented (Fig. 1C). According to the screening criteria, 1,239 DEGs with HR greater than 1 and 361 DEGs with HR less than 1 were found in the training set (n = 186) (Fig. 2A). These intersected with the genes that were upregulated and downregulated, resulting in 1272 overlapping candidate genes (OCGs), of which 1227 had an HR greater than 1 and 45 had an HR less than 1 (Fig. 2A).

Fig. 1.

Processing and screening of differentially expressed genes (DEGs). A A principal component analysis (PCA) plot of the TCGA HCC dataset showing no batch effect, with red nodes for normal and blue for tumor clusters. B A volcano plot comparing DEGs in HCC to normal liver tissue, highlighting significantly up-regulated genes (red nodes, logFC > 2, FDR < 0.05) and down-regulated genes (green nodes, logFC < -2, FDR < 0.05). C A heatmap displaying the top 20 DEGs in HCC

Fig. 2.

Identification of four key prognostic genes and their expression levels in HCC. A Venn diagram shows 1600 DEGs, with 1239 upregulated (HR > 1) and 361 downregulated (HR < 1). B LASSO coefficients for 310 prognostic genes, with each line indicating changes in gene coefficients; axes represent log lambda and coefficient values, with numbers showing non-zero gene coefficients at different log lambda values. C Using LASSO regression with tenfold cross-validation, eight prognostic genes are identified at the minimum lambda value; the axes show log lambda and Partial Likelihood Deviance, with numbers representing non-zero protein coefficients. D Using multivariate Cox regression analysis, a gene signature is formed from the eight genes associated with prognosis. E-H The GSE14520 database was used to validate the mRNA expression of four prognostic genes in primary tumors and surrounding normal liver tissues. HR hazard ratio. LASSO least absolute shrinkage and selection operator. *p < 0.05, **p < 0.01

Screening, validating, and IHC of prognosis-related DEGs

Initially, 442 prognosis-related DEGs were identified from 1,272 OCGs using univariate Cox regression analysis, but only 310 DEGs with p < 0.01 were selected for further analysis. Afterward, LASSO regression using tenfold cross-validation was utilized to find the best lambda value, linked to 8 genes out of the 310 DEGs that showed a strong correlation with OS (Fig. 2B, C). Following this, multivariate Cox regression analysis identified an optimal prognostic model comprising four genes: TXNRD1 (Thioredoxin reductase 1, cytoplasmic), NEIL3 (Endonuclease 8-like 3), MMP1 (Interstitial collagenase), and EFNA4 (Ephrin-A4) (Fig. 2D, with HR and 95% CI displayed to three decimal places).

We validated the differential expression of TXNRD1, NEIL3, MMP1, and EFNA4 in HCC utilizing the GSE14520 dataset. As illustrated in Fig. 2E-H, the mRNA expression levels of these genes were markedly elevated in HCC tissues relative to adjacent normal tissues (p < 0.001). Protein levels of EFNA4, MMP1, and TXNRD1 were also assessed in 19 specimen pairs via IHC, showing increased expression in HCC tissues (Fig. 3A-I and Supplementary Fig. 1–8). Additionally, consistent with the mRNA findings, protein expression data from the Human Protein Atlas database indicated significant overexpression of EFNA4 and TXNRD1 in HCC tissues compared to normal liver (Fig. 3J-K). However, data for NEIL3 is not available in the Human Protein Atlas, and MMP1 does not exhibit differential expression in liver cancer. Kaplan-Meier survival analysis demonstrated that HCC patients with elevated expression levels of the four prognosis-related DEGs—TXNRD1, NEIL3, MMP1, and EFNA4—exhibited a significantly reduced survival rate compared to patients with lower expression levels of these DEGs in the GSE14520 dataset (Fig. 4A-D).

Fig. 3.

The expression levels of the prognosis-gene signature in HCC tissues were evaluated using immunohistochemistry (IHC). Panels (A, D, G) display representative IHC images demonstrating EFNA4 expression in lymph nodes (A), MMP1 in testicular tissue (D), and TXNRD1 in pancreatic tissue (G), which serve as positive controls. Panels (B, E, H) present IHC staining results for EFNA4 (B), MMP1 (E), and TXNRD1 (H) in HCC and adjacent normal tissues, with images captured at magnifications of 100 × and 400 ×. Panels (C, F, I) provide H-score statistics calculated for 19 paired specimens in panels C and I, and for 7 paired specimens in panel F. Panels (J-K) illustrate that EFNA4 (J) and TXNRD1 (K) expression levels are elevated in HCC compared to normal liver tissue, as indicated by data from The Human Protein Atlas database

Fig. 4.

Using data from the GSE14520 database, a Kaplan-Meier analysis was carried out to show the relationship between the expression levels of four prognosis-related DEGs and overall survival in HCC patients. A EFNA4. B MMP1. C NEIL3. D TXNRD1

Construction and appraisal of the four-gene prognostic signature

Based on the outcomes of the multivariate Cox regression analysis, a prognostic gene signature was developed. This signature is represented by a four-gene-based risk score, formulated according to their respective Cox coefficients: risk score = 0.0115*Exp_(TXNRD1) + 0.2031*Exp_(NEIL3) + 0.1147*Exp_(MMP1) + 0.0651*Exp_(EFNA4). Patients were categorized into high-risk (n = 89) and low-risk (n = 91) groups using the “survminer” R package to determine the median cut-off point. The high-risk group had significantly more deaths, as shown in Fig. 5B. Figure 5C displays a heatmap of the four genes, aligning with expression trends in Fig. 2E-H. Kaplan–Meier curves (Fig. 5D) demonstrated that patients with high-risk HCC had notably shorter survival durations than those in the low-risk category (p < 0.001). The four-gene signature exhibited greater AUC values in time-dependent ROC curves than single genes (Fig. 5E and Supplementary Fig. 9). This consistency was observed throughout the testing and entire datasets (Fig. 6A-D). The AUCs for 1-year, 3-year, and 5-year OS in all datasets [(0.70) (Figs. 5E and 6B, and 6D)] highlighted the robust predictive accuracy of the four-gene signature for OS.

Fig. 5.

Assessment of a four-gene signature’s prognostic potential in the TCGA HCC cohort training set. A Risk scores and their median, separating patients into groups of low and high risk. B Survival status, with higher risk scores linked to more deaths. C mRNA expression heatmap of the four-gene signature. D Kaplan-Meier survival curves for OS in both risk categories. E Time-dependent receiver operating characteristic (ROC) analysis for the four-gene signature

Fig. 6.

ICGC HCC and GSE76427 were used as external validation sets to confirm the prognostic and predictive value of the four-gene signature. Kaplan-Meier survival analysis was performed on the TCGA testing set (A), whole set (C), and external validation set (E and G). (B, D, F and H) The AUC of time-dependent ROC curves confirmed the risk score’s prognostic performance across these sets

Validation of the prognostic gene signature using external data

To validate the predictive significance of the four-gene signature, we determined risk scores for HCC patients in both the ICGC database and the GSE76427 dataset. Consistent with the TCGA cohort findings, high-risk patients demonstrated significantly worse OS than those at low risk (p < 0.05) (Fig. 6E and G). The 5-year OS AUCs were 0.732 and 0.868, respectively (Fig. 6F and H), suggesting that the four-gene signature is an effective predictor of OS in HCC.

Construction and validation of a gene-based prognostic nomogram

A cohort of 165 patients, with comprehensive data encompassing gender, age, tumor grade, and stage, was selected for further analysis. Using univariate and multivariate Cox regression analyses, the risk score based on the four-gene signature was identified as an independent predictor of OS in the training set (Fig. 7A, B). To estimate 1-year, 3-year, and 5-year OS in HCC patients, a nomogram incorporating this risk score and stage was created (Fig. 7C). In the TCGA training or testing sets, calibration plots showed that the four-gene prognostic nomogram might not accurately estimate mortality, potentially leading to underestimation or overestimation (Fig. 8A and E). However, the calibration plot for predicting patient survival showed a high level of agreement between the predicted and actual results in the TCGA whole dataset (Fig. 8I). The nomogram’s AUC values for predicting OS, which are greater than 0.7, exceeded those of the risk score and stage alone (Fig. 8B-D, F-H and J-L), indicating it is the most effective predictor.

Fig. 7.

Identifying independent prognostic factors and developing a gene-based prognostic model. A, B Forrest plot of univariate (A) and multivariate (B) Cox regression analysis in the TCGA HCC training set. C Nomogram predicting HCC patient OS, where points from two predictors (Stage and riskScore) are summed to estimate 1-, 3-, and 5-year OS probabilities

Fig. 8.

Evaluation of a gene-based nomogram for predicting survival probability, comparing its accuracy with that of a risk score and disease stage. Calibration plots evaluate the consistency between predicted and actual 1-, 3-, and 5-year OS results across the training (A), testing (E), and entire (I) datasets. For 1-, 3-, and 5-year OS predictions, time-dependent ROC curves compare the nomogram, risk score, and stage across the training set (B, C, D), testing set (F, G, H), and complete set (J, K, L)

Discovery of biological pathways linked to DEGs and four prognostic genes

To clarify the biological roles of 1,600 DEGs, GO and KEGG enrichment analyses were performed. In the GO biological analysis, the DEGs were predominantly enriched in categories such as chromosomal region, collagen-containing extracellular matrix, spindle, and condensed chromosome. The biological processes were primarily associated with organelle fission, nuclear division, chromosome segregation, and mitotic sister chromatid segregation, among others. The analysis of molecular functions revealed that these DEGs were primarily associated with structural components of the extracellular matrix, integrin binding, DNA helicase activity, and activity related to single-stranded DNA helicases (Fig. 9A). The KEGG pathway analysis revealed multiple pathways for DEGs, such as cell cycle, ECM-receptor interaction, and DNA replication (Fig. 9B). Furthermore, GSEA was conducted to investigate the possible biological processes linked to four prognostic genes. The findings indicated that samples with high levels of TXNRD1, NEIL3, MMP1, and EFNA4 were mainly enriched in pathways associated with cell growth, DNA replication and repair, and WNT signaling (Fig. 10; Tables 1, 2, 3 and 4).

Fig. 9.

Functional analysis of DEGs in the TCGA HCC cohort. A GO enrichment analysis. B KEGG enrichment analysis

Fig. 10.

Gene Set Enrichment Analysis (GSEA) was conducted in relation to the expression of four prognostic genes. The analysis revealed that the gene sets “WNT signaling pathway” (A), “DNA replication” (B), “Cell cycle” (C), and “Pyrimidine metabolism” (D) were significantly enriched in HCC samples exhibiting high expression levels of EFNA4, MMP1, NEIL3, and TXNRD1, respectively

Table 1.

GSEA gene enrichment results of EFNA4

Gene set name	NES	NOM p-val	FDR q-val
KEGG_PURINE_METABOLISM	2.00	0.000	0.011
KEGG_RNA_POLYMERASE	1.95	0.000	0.016
KEGG_PYRIMIDINE_METABOLISM	1.95	0.000	0.013
KEGG_NOTCH_SIGNALING_PATHWAY	1.91	0.000	0.013
KEGG_CELL_CYCLE	1.90	0.000	0.014
KEGG_WNT_SIGNALING_PATHWAY	1.85	0.002	0.022

NES normalized enrichment score, NOM nominal, FDR false discovery rate

Gene sets with NOM p-val < 0.05 and FDR q-val < 0.05 are considered as significant

Table 2.

GSEA gene enrichment results of MMP1

Gene set name	NES	NOM p-val	FDR q-val
KEGG_NUCLEOTIDE_EXCISION_REPAIR	2.07	0.000	0.003
KEGG_CELL_CYCLE	2.06	0.000	0.003
KEGG_MTOR_SIGNALING_PATHWAY	1.93	0.000	0.011
KEGG_DNA_REPLICATION	1.88	0.000	0.013
KEGG_WNT_SIGNALING_PATHWAY	1.88	0.000	0.012
KEGG_PYRIMIDINE_METABOLISM	1.87	0.002	0.012

NES normalized enrichment score, NOM nominal, FDR false discovery rate

Gene sets with NOM p-val < 0.05 and FDR q-val < 0.05 are considered as significant

Table 3.

GSEA gene enrichment results of NEIL3

Gene set name	NES	NOM p-val	FDR q-val
KEGG_CELL_CYCLE	2.23	0.000	0.000
KEGG_BASE_EXCISION_REPAIR	2.19	0.000	0.000
KEGG_PYRIMIDINE_METABOLISM	2.09	0.000	0.001
KEGG_PURINE_METABOLISM	2.04	0.000	0.003
KEGG_DNA_REPLICATION	1.98	0.000	0.005
KEGG_P53_SIGNALING_PATHWAY	1.87	0.000	0.013

NES normalized enrichment score, NOM nominal, FDR false discovery rate

Gene sets with NOM p-val < 0.05 and FDR q-val < 0.05 are considered as significant

Table 4.

GSEA gene enrichment results of TXNRD1

Gene set name	NES	NOM p-val	FDR q-val
KEGG_PENTOSE_AND_GLUCURONATE_INTERCONVERSIONS	2.22	0.000	0.000
KEGG_GLUTATHIONE_METABOLISM	2.11	0.000	0.001
KEGG_AMINO_SUGAR_AND_NUCLEOTIDE_SUGAR_METABOLISM	2.04	0.000	0.002
KEGG_STARCH_AND_SUCROSE_METABOLISM	2.00	0.000	0.003
KEGG_OTHER_GLYCAN_DEGRADATION	1.98	0.000	0.003
KEGG_PYRIMIDINE_METABOLISM	1.79	0.002	0.030

NES normalized enrichment score, NOM nominal, FDR false discovery rate.

Gene sets with NOM p-val < 0.05 and FDR q-val < 0.05 are considered as significant

Discussion

The global incidence of HCC is rising [15]. Detecting HCC early is vital for improving patient outcomes, as the prognosis is closely tied to the disease stage, with a 5-year survival rate of 36% for early stages and just 13% for advanced metastatic cases [16, 17]. As discussed above, relying on a single clinical parameter for prognosis is inadequate; therefore, combining multiple prognostic factors is recommended for more accurate predictions.

In our study, we found a unique four-gene signature (TXNRD1, NEIL3, MMP1, and EFNA4) that can predict the prognosis of HCC. These genes exhibited significantly higher mRNA expression in HCC tissues, which correlated with shorter patient survival in the external validation database GSE14520. Serving as an independent prognostic factor, the four-gene risk model effectively predicted prognosis in the TCGA, ICGC, and GSE76427 HCC cohorts. Moreover, high-risk patients had significantly poorer survival. The nomogram combining the four-gene signature and clinical stage (conventional clinical parameter) was the best predictor of OS. The calibration curves for the nomogram showed excellent agreement between predicted and observed outcomes in the entire TCGA cohort. Some deviation was observed in the training and testing subsets, which could be attributed to the smaller sample size in each subset leading to increased variability in the calibration estimates. The robust performance in the entire cohort and external validation sets supports the general stability of the model. The KEGG and GSEA analyses indicated that these genes are essential for cell cycle and DNA replication pathways, underscoring their involvement in HCC development and potential molecular mechanisms.

Thioredoxin reductase 1 (TXNRD1), a key redox regulator in mammalian cells, is implicated in cancer development. Huang et al. [18]. found that TXNRD1 is significantly elevated in HCC tumors and linked to poor patient survival. The mechanism involves USF2-driven TXNRD1 upregulation, which promotes HCC progression by activating Akt/mTOR signaling. Fu et al. [19]. discovered that TXNRD1 was overexpressed in HCC tissues and cells, with IHC analysis showing elevated levels in 47.5% (57/120) of clinical samples, increasing with clinical stage. Hua et al. [20]. reported that miR-125b-5p suppresses HCC cell proliferation, migration, and invasion by targeting TXNRD1. Lee et al. [21]. found that targeting TXNRD1 led to significant ROS accumulation, effectively suppressing HCC tumor growth. The Nei endonuclease VIII-like 3 (NEIL3) is part of the Fpg/Nei family and acts as a monofunctional glycosylase in the base excision repair (BER) pathway [22]. Wu et al. [23]. discovered that NEIL3 protein concentrations were much higher in cancerous tissues than in non-cancerous ones, pointing to an unfavorable prognosis for HCC. Zhang et al. [24]. found that SNHG3 regulates NEIL3 through E2F1, promoting HCC proliferation. Lai et al. [25]. showed that NEIL3 activates the BRAF/MEK/ERK/TWIST pathway, leading to epithelial-mesenchymal transition (EMT) and treatment resistance in HCC. MMP-1, which is also called collagenase-1, is responsible for breaking down interstitial collagens and is regulated by TIMP-1. The overproduction of MMP-1 and TIMP-1 boosts the invasiveness and migration of HCC cells by degrading the extracellular matrix (ECM) during EMT [26]. Xu et al. [27]. discovered that in the group of matrix metalloproteinases (MMPs), only MMP1, MMP3, MMP8, MMP9, MMP11, MMP12, MMP14, MMP15, MMP20, MMP21, and MMP24 show a significant increase in HCC compared to nearby tissues. However, survival analyses, along with univariate and multivariate Cox regression analyses, indicated that only MMP1 serves as a poor prognostic biomarker for HCC. Ephrin A4 (EFNA4), an EPH family ligand, aids blood vessel and epithelial development by controlling cell migration and rejection [28]. Ye et al. [29]. found EFNA4 overexpressed in HCC patients, acting as an oncogene and prognostic marker. Yuan et al. [30]. showed that disrupting EFNA4 reduces cell proliferation, invasion, and angiogenesis in HCC by downregulating pygopus-2 (PYGO2). Our findings align with previous results, showing that IHC analysis confirmed significant overexpression of TXNRD1, MMP1, and EFNA4 proteins in HCC tissues.

The GSEA results indicated that samples with high expression of EFNA4, MMP1, and NEIL3 were significantly enriched in the cell cycle signaling pathway. Furthermore, high expression of all four genes was linked to pyrimidine metabolism. These findings, combined with prior studies, suggest potential mechanistic roles: TXNRD1 may promote HCC via redox-mediated Akt/mTOR activation. NEIL3, a DNA repair enzyme, could facilitate rapid proliferation and genomic instability. MMP1 likely contributes to invasion and EMT through ECM degradation. EFNA4 may drive proliferation and angiogenesis via pathways such as Wnt/β-catenin or through interaction with its Eph receptor. The enrichment in pyrimidine metabolism underscores a potential role in sustaining nucleotide pools for unchecked DNA replication in cancer cells. These hypotheses, however, require direct experimental validation through future in vitro and in vivo studies to elucidate the precise molecular mechanisms.

Recent studies highlight the potential of gene-signatures based on aberrant mRNA for predicting HCC prognosis. Miao et al. [31]. identified seven oxidative stress-related genes (ENO1, NDRG1, NPM1, H2AX, IL33, MT3, TXNRD1) using the Cox regression model, which are significant for HCC prognosis. Hu et al. [32]. formulated a gene signature involving seven genes (PSRC1, CHORDC1, KPNA2, CDCA8, G6PD, NEIL3, EZH2) related to glycolysis and immunity to categorize HCC risk. As far as we know, this study introduces a four-gene (TXNRD1, NEIL3, MMP1, and EFNA4) signature prognostic model and nomogram that have not been reported before and could be a useful tool for diagnosing and predicting outcomes in HCC. Similarly, Wu et al. [33]. constructed a prognostic nine-gene signature (IQGAP3, BIRC5, PTTG1, STC2, CDKN3, PBK, EXO1, NEIL3, and HOXD9) based on DEGs in HCC, which can be utilized as a combined biomarker for the independent prediction of OS in HCC patients. Notably, the tROC analysis showed that the four-gene prognostic signature was more effective than individual genes for diagnosis. When combined with clinical parameters, such as stage, in the nomogram, it significantly improved survival predictions for HCC patients, indicating its potential as a strong survival predictor. Moreover, determining the risk score by assessing the expression levels of these four genes is both economical and fast compared to gene mutation and other genetic analyses.

To shed light on the molecular mechanisms that may contribute to the onset and advancement of liver cancer, we analyzed 1,600 DEGs identified in HCC through enrichment analysis. Our results are consistent with Guo et al [34], who found 152 DEGs in HCC tissues (165 samples) when compared to normal tissues (73 samples). The analysis of KEGG enrichment showed that these DEGs were primarily connected to the cell cycle. In a similar manner, Zhang et al [35]. identified 27 overlapping DEGs in HCC and normal tissues using data from both TCGA and the GSE60502 dataset, with ‘cell cycle’, ‘oocyte meiosis’, and ‘p53 signaling pathway’ as the top three enriched KEGG pathways. Additionally, Su et al [36]. reported 415 DEGs in HCC, which included 196 genes that were up-regulated and 219 that were down-regulated. According to KEGG pathway analysis, the up-regulated DEGs played a role in controlling the cell cycle and DNA replication pathways, while the down-regulated DEGs were associated with metabolic pathways. Overall, these results imply that the DEGs in HCC tissues might affect the progression of HCC by altering the proliferation of cancer cells. It is noteworthy that GSEA results highlighted significant enrichment in the cell cycle signaling pathway for samples with overexpression of EFNA4, MMP1, and NEIL3. Moreover, increased expression of EFNA4, MMP1, NEIL3, and TXNRD1 was linked to an enhancement in the pyrimidine metabolism signaling pathway. These findings imply that the four overexpressed DEGs are crucial in encouraging mutation and malignant growth in HCC cells.

In summary, our integrated bioinformatics analysis and validation suggest that the four-gene signature (EFNA4, MMP1, NEIL3, TXNRD1) is closely associated with HCC prognosis and may exert its effects through core cancer-related pathways such as cell cycle regulation and nucleotide metabolism. This provides a focused direction for future mechanistic exploration.

Limitations of the study

Nevertheless, the present study had several limitations. First, the sample size for IHC validation was relatively small (19 pairs). Second, we did not validate the mRNA expression levels of the four prognostic genes in our local cohort using RT-qPCR. Third, due to the unavailability of a validated anti-NEIL3 antibody suitable for IHC in our laboratory at the time of the study, we were unable to include it in our IHC validation. Fourth, the study lacked in vitro and animal model experiments to explore the functional roles of these genes in HCC development. Furthermore, the prognostic model was built and validated using retrospective public databases which lack detailed information on certain clinical confounders such as etiology, liver function status, specific treatment regimens, and patient comorbidities. Additional in-depth research, including future prospective studies incorporating comprehensive clinical variables, is required to confirm our results and improve the clinical use of the four-gene prognostic nomogram for assessing HCC prognosis.

Conclusions

We developed a distinctive four-gene signature and nomogram to predict OS in HCC, equipping clinicians with a reliable tool for prognosis assessment and treatment strategy formulation. The identified genes also represent promising candidates for future mechanistic and therapeutic investigations.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

W.G. Conceptualization, resources, data curation, funding acquisition, investigation, and writing-original draft. W.C. Conceptualization, resources, investigation, writing-original draft. Y.Z. Data curation, software, methodology. X.Y. Data curation, software. P.W. Conceptualization, supervision, validation, writing-review and editing. O.Z. Supervision, validation, visualization, project administration, writing-review and editing.

Funding

This research was supported by the Medical and Health Technology Projects in Zhejiang, China (Grant No. 2023KY405), the Science and Technology Projects of the Taizhou Science and Technology Bureau, China (Project No. 22ywb28), and the Zhejiang Provincial Natural Science Foundation of China (Grant No. LBY22H200001). The financial backers were not involved in designing the study, collecting and analyzing data, deciding to publish, or preparing the manuscript.

Data availability

The data generated from this study are accessible through the TCGA database (https://portal.gdc.cancer.gov/), the GEO database (https://www.ncbi.nlm.nih.gov/geo/), and the ICGC database (http://dcc.icgc.org/releases/current/). Supplementary Tables S1 (complete list of DEGs) is provided. The analysis code is available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This research received approval from the Ethics Committee of Taizhou Central Hospital (Taizhou University Hospital), under the ethical approval number TCH2025-093. All procedures were conducted in strict adherence to the principles outlined in the Declaration of Helsinki. Informed consent was secured from all individual participants involved in the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.