Integrated multi-dataset screening to predict prognosis and identify immunotherapy gene targets in hepatocellular carcinoma patients

Abstract

This study systematically predicts prognosis and key gene targets for immunotherapy in hepatocellular carcinoma (HCC) patients based on the joint screening of multiple datasets. Transcriptomic and clinical data from TCGA, GEO, and ICGC were integrated to construct a multi cohort analytical framework. Weighted Gene Co expression Network Analysis was applied to identify key functional modules within the GEO dataset. A comprehensive machine learning ensemble strategy—incorporating over one hundred combinations of classical feature selection and survival prediction algorithms—was employed to derive a robust multi gene prognostic signature. Model performance was evaluated through Kaplan–Meier survival analysis, time dependent ROC curves, and decision curve analysis across multiple independent validation cohorts. Additional analyses examined differential expression across clinical subgroups, immune cell infiltration patterns, immune checkpoint associations, and gene mutation profiles to further elucidate the biological and immunological relevance of the identified genes. Ten key genes were identified. TYMS was identified as a risk factor, while APOL3 and FBXO2 emerged as potential protective factors. Candidate genes were closely associated with features of the immune microenvironment, showing significant correlations with levels of immune cell infiltration and expression of immune checkpoint molecules such as PD-1 and CTLA-4. This study identified core HCC genes with prognostic and immunotherapeutic significance, providing novel targets and a theoretical basis for optimizing risk stratification and personalized treatment.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-38424-8.

Introduction

Hepatocellular carcinoma (HCC) stands as the sixth most common malignant neoplasm worldwide and is the third foremost cause of cancer-related death.^1,2 In 2023 global reports, over 900 000 new cases were recorded annually, with roughly 83 percent arising in Asia.³ Chronic hepatitis B infection, cirrhosis, and metabolic-associated fatty liver disease remain the primary drivers of HCC.⁴ Although interventions such as surgical resection, ablative therapies and targeted agents (for example, sorafenib or lenvatinib) have appreciably extended survival in early-stage patients, five-year survival for advanced HCC stagnates below 18 percent, and recurrence approaches 70 percent at five years.⁵ Most studies to date have concentrated on remodeling the tumour microenvironment, combining immune-checkpoint blockade with other modalities and probing epigenetic regulators.⁶ Nevertheless, a systematic delineation of the key molecular circuits underpinning HCC heterogeneity and therapeutic resistance has yet to be achieved.

In current clinical practice, reliance on TNM staging alongside the Child–Pugh score provides only limited precision in forecasting individual treatment responses, while targeted therapies demonstrate modest efficacy and often succumb to acquired resistance.⁷ Comprehensive genomic profiling has uncovered striking inter-tumour diversity: mutations in TP53, CTNNB1 and the TERT promoter correlate closely with tumour evolution, metastatic potential and drug sensitivity.^8,9 Notably, CTNNB1-mutant lesions typically respond poorly to immune-checkpoint inhibitors but may be vulnerable to Wnt/β-catenin pathway blockade. Thus, integrating multi-omics data to pinpoint prognostic hub genes promises both to refine risk stratification and to underpin bespoke therapeutic strategies—a crucial translational step toward overcoming current bottlenecks in precision HCC.^10–13.

In this study, we integrated transcriptomic and clinical data from multiple sources, including TCGA, GEO, and ICGC. By combining network co-expression analysis with differential expression screening for mutual validation, and employing batch effect correction strategies, we constructed a robust gene interaction network. On this basis, machine learning algorithms were used to identify core candidate genes, whose predictive performance was validated through multiple approaches such as time-dependent ROC, Kaplan–Meier survival analysis, and decision curve analysis. Finally, we further examined the expression differences of the candidate genes across different clinical subgroups, their associations with immune cell infiltration and immune checkpoint molecule expression, as well as gene mutation characteristics, thus providing a comprehensive and in-depth molecular foundation for prognostic evaluation and the discovery of immunotherapeutic targets in HCC.

Methods

Data acquisition and weighted gene co-expression network analysis (WGCNA)

This study comprised an integrative, multi‐database analysis. First, we characterized large‐scale HCC transcriptomes by retrieving the GSE244826 series from the Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/) and applying the WGCNA framework. To inform machine‐learning (ML) model development, we then incorporated GSE208036, GSE233421 and GSE273947.^14,15 Concurrently, we downloaded HCC RNA-sequencing expression profiles and corresponding clinical annotations from the International Cancer Genome Consortium (ICGC; https://dcc.icgc.org/) and The Cancer Genome Atlas (TCGA; https://portal.gdc.com/). Finally, Kaplan–Meier survival analyses for candidate gene signatures were performed independently in the ICGC and TCGA cohorts.

In our analysis, WGCNA proceeded through six meticulously designed stages. First, expression data were reduced to the top 50% most variably expressed genes and adjusted for batch effects using ComBat. Second, the optimal soft‐thresholding power was selected by examining the inflection point in the scale‐free topology fit index versus mean connectivity plot. Third, Pearson correlation coefficients defined gene–gene similarity, which was then transformed into a weighted topological overlap matrix (1 – TOM) to diminish indirect co‐expression biases. Fourth, co‐expression modules were identified via the dynamic tree‐cut algorithm, applying specified cutoffs for minimum module size, merge height and a deepSplit value of 2. Fifth, each module’s primary expression pattern was summarized as its eigengene, obtained from the first principal component. Finally, Spearman rank correlations and multivariable analyses evaluated links between module eigengenes and critical clinical parameters, including tumour stage and metastatic status.

Differential expression screening and functional enrichment analysis

In this study, differential expression analysis was conducted using the limma framework, wherein linear models were fitted to the expression matrix and empirical Bayes moderation was applied to refine variance estimations.¹⁶ To control for multiple hypothesis testing, the Benjamini–Hochberg procedure was employed, defining significance as P value below 0.05 coupled with an absolute log2 fold‐change exceeding 1. Results were depicted as volcano plots via ggplot2 (v3.4.2), wherein –log10 ( P value) values were plotted against log2 fold‐changes and non‐significant transcripts were shown in gray, with up‐ and downregulated genes distinguished in red and green, respectively.

Functional enrichment was performed through the DAVID platform: Gene Ontology categories (Biological Process, Molecular Function, Cellular Component) were assessed by hypergeometric testing (FDR < 0.05, enrichment fold > 2) and consolidated with REVIGO to eliminate redundant terms, while KEGG pathways were interrogated by Fisher’s exact test (adjusted p < 0.01), selecting only those pathways encompassing at least five differentially expressed genes and demonstrating an impact factor above 0.2.^17–19 Finally, an interaction network of enriched pathways was assembled in Cytoscape using the ClueGO plugin (kappa score = 0.4), illuminating key functional interrelationships.²⁰.

Construction of the VNN plot and batch effect correction across datasets

A panel of 93 candidate genes was defined by intersecting 2,012 WGCNA module members with 165 DEGs (|log₂ fold-change|> 1; P.Val < 0.05), employing dplyr’s inner_join to ensure exact concordance of gene symbols and Ensembl identifiers. These genes formed the basis of a Visual Nearest Neighbours (VNN) network, constructed to map their interrelationships.

We then harmonized four independent transcriptomic cohorts from GEO (GSE244826, GSE208036, GSE233421 and GSE273947). Raw matrices were converted to log₂ scale, followed by quantile normalization. Batch effects were subsequently mitigated with the ComBat algorithm (sva package; model =  ~ 1) to retain genuine biological variation.

To validate the integration, we applied three complementary strategies: UMAP embeddings (uwot) compared clustering patterns before versus after correction.Median expression per gene was visualized in grouped bar charts and tested by Kruskal–Wallis, with p > 0.05 taken as evidence of successful batch-effect removal. The coefficient of variation across batches was quantified, requiring a ≥ 30% reduction in inter-group CV disparity post-adjustment.

All figures were produced using ggplot2, encoding dataset origin by point shape and clinical subtype by color. Finally, density plots of expression profiles confirmed alignment of gene-level distributions across cohorts following correction.

Machine‐learning framework, stability assessment and predictive model development

We employed a multi-layered ensemble framework to derive both classification and survival-prediction models. For diagnostic classification, twelve established algorithms—including Lasso, Ridge, and Elastic Net regressions—were combined across feature-selection and modeling modules, resulting in 113 distinct workflows. Survival modeling integrated ten methods (e.g., random survival forests, CoxBoost, survival SVMs), yielding 101 prognostic frameworks.

Model training was performed using five-fold cross-validation, repeated ten times for robustness, with hyperparameter tuning via leave-one-out cross-validation. Performance was evaluated by area under the receiver-operating characteristic curve (AUC) for classification tasks and Harrell’s concordance index (C-index) for survival models. Stability was assessed across six independent validation cohorts, and a hierarchical clustering heatmap (tree height cutoff = 0.8) was used to identify clusters of superior model combinations.

To select the final gene signature, we applied a stepwise consensus-based feature prioritization strategy. First, genes were retained only if they demonstrated statistically significant coefficients (p < 0.05) in more than 50% of the top-performing models. Second, Shapley Additive Explanations (SHAP) were used to quantify feature contributions, and genes ranking within the top 10% of SHAP importance scores across models were prioritized.

Candidate genes were further cross-validated across the TCGA, ICGC, and GEO datasets to ensure consistent prognostic associations. Finally, functional annotation against the MSigDB Hallmark collection (FDR < 0.25) confirmed the involvement of selected genes in biologically relevant pathways, particularly those related to immunity and hepatocarcinogenesis.

This multi-tiered strategy led to the identification of ten consensus biomarkers, which formed the basis of our final risk prediction model. For clinical translation, we implemented a lightweight representation-learning predictor via a standardized API, capable of real-time risk estimation. In a prospective validation cohort, the system maintained an AUC > 0.85 and demonstrated added prognostic value over conventional clinical indices. Kaplan–Meier analysis was performed using ICGC and TCGA cohorts to evaluate survival differences between risk groups.

Time‐dependent ROC curves, Kaplan–Meier survival analyses and decision curve plotting

We validated prognostic performance in hepatocellular carcinoma (HCC) cohorts from the ICGC and TCGA repositories. Overall survival (OS) was defined as the interval from diagnosis to death from any cause, with censoring at last follow-up. Inter-study heterogeneity was quantified by meta-analysis, permitting the aggregation of subgroups exhibiting low variability.

Time-dependent areas under the receiver-operating characteristic curve (AUC) at 1, 3 and 5 years were estimated using the timeROC package in R. Predictor variables comprised clinicopathological parameters (TNM stage, differentiation grade) and the genomics-derived risk score; combined prognostic probabilities were obtained from a multivariable Cox proportional hazards model. AUC comparisons utilized DeLong’s test (two-sided p < 0.05). All plots were generated with ggplot2, employing TCGA/ICGC official color palettes.

Patients were stratified into high- and low-risk groups by the median risk score. Survival differences were evaluated with stratified log-rank tests (survival package v3.5), and hazard ratios (HRs) with 95% confidence intervals were calculated via Cox regression. Kaplan–Meier curves, accompanied by risk tables indicating numbers at risk at prespecified time points, were drawn using survminer (v0.4.9).

Differential analysis of candidate genes under different clinical characteristics

To systematically assess the association between candidate gene expression profiles and clinicopathological features, this study established a multidimensional stratified analysis framework. Patients were first standardized and grouped according to international clinical practice guidelines. Grouping was performed based on age (cutoff at 65 years), BMI (cutoff at 25), gender, tumor grade and stage, and whether the patient received radiotherapy or chemotherapy, followed by intergroup comparisons. Visualization was conducted using faceted boxplots (ggplot2 v3.4.2). Differential expression analysis employed mixed-effects models: independent samples t-tests were applied to normally distributed data, while Mann–Whitney U tests were used for non-normally distributed data. Multiple comparison correction was implemented through a two-stage strategy: Bonferroni correction was applied within-group comparisons, and false discovery rate control was used for between-group analyses.

Core endpoint indicators of candidate differential gene prediction model: COX regression and survival analysis

After the preliminary screening of candidate differential genes, this study employed the COX proportional hazards regression model and survival analysis to evaluate their association with clinical outcomes. The specific methods are as follows. Based on clinical follow-up data (including survival time and event information) from TCGA and ICGC, as well as the differential gene expression matrix, univariate COX regression analysis was performed for each candidate gene individually using the survival package in R language (v4.0.0). Genes with a p value < 0.05 were considered as candidate factors with potential prognostic value and were included in the subsequent multivariate model construction. To avoid multicollinearity, variance inflation factor (VIF) was applied to assess correlations among variables, and independent variables with VIF > 5 were excluded.

The selected candidate genes were then incorporated into a multivariate COX regression model. Model refinement was conducted using a forward stepwise method based on the Akaike Information Criterion (AIC), ultimately obtaining the optimal prognostic gene combination along with their corresponding regression coefficients (β). A risk score formula was constructed based on the multivariate regression coefficients: Risk Score = Σ (βi × expressioni). Subjects were divided into high-risk and low-risk groups according to the median risk score.

Survival curves for the two groups were plotted using the Kaplan–Meier method, and the log-rank test was applied to compare survival differences between groups. Visualization of curves was performed using the survminer package in R. The significance level was set at p < 0.05, and hazard ratios (HR) with 95% confidence intervals (CI) were reported.

Finally, Harrell’s C-index was used to assess the discriminatory ability of the model. Time-dependent receiver operating characteristic (ROC) curves at different follow-up time points were plotted using the timeROC package, and the corresponding area under the curve (AUC) values were calculated to evaluate the model’s predictive performance at various time points. To validate the robustness of the model, the above analyses were repeated in an independent validation cohort, with comparisons of C-index, AUC, and survival differences conducted to verify the model’s external validity and generalizability.

Immune cell infiltration and immunotherapy assessment

This study utilized transcriptome sequencing data of hepatocellular carcinoma (HCC) and corresponding clinical information obtained from the TCGA and ICGC databases. The CIBERSORT algorithm (LM22 gene signature matrix, with 1,000 permutations) was applied to quantitatively assess the relative abundance of 22 known immune cell subpopulations in each sample, and the stability of the results was validated through permutation testing. To further enhance the reliability of immune infiltration analysis, this study also performed cross-validation of the infiltration levels of major immune cell subpopulations using the TIMER (Tumor Immune Estimation Resource) database and the ssGSEA (single-sample Gene Set Enrichment Analysis) method based on the GSVA package. Finally, to predict patients’ potential responses to immune checkpoint inhibitor (ICI) therapy, the TIDE (Tumor Immune Dysfunction and Exclusion) algorithm was employed to estimate tumor immune dysfunction and exclusion status, and the scores were used to explore the sensitivity of different subgroups of patients to immune checkpoint blockade (ICB).

Analysis of gene mutation characteristics

Based on the previously screened key genes, this study further systematically evaluated the genomic mutation characteristics in samples from HCC patients. Mutation data were obtained from the TCGA data portal (https://portal.gdc.cancer.gov/), downloaded in MAF format, and preprocessed using the maftools package (v2.6.05) in the R environment (v4.0.3). First, patient samples were divided into high-expression and low-expression groups according to the median expression of each key gene. Subsequently, mutation data were imported using the read.maf function, and the plotmafSummary and oncoplot modules were used to generate overall tumor mutation burden (TMB) distribution maps, mutation type stacked bar charts, and gene hotspot mutation plots. To assess differences in mutation frequency of key genes between high- and low-expression groups, Fisher’s exact test was applied with Benjamini–Hochberg multiple testing correction on the p-values, setting a significance threshold at FDR < 0.05. Finally, combined with the mutation spectrum difference analysis results, this study explored the potential pathogenic mechanisms of key genes in the occurrence and development of HCC, providing a basis for subsequent functional validation and clinical translation.

Polymerase chain reaction (PCR) detection of key gene expression levels

To further investigate differences in the serum expression levels of key genes in patients with HCC, peripheral blood samples were collected from individuals with HCC and from healthy controls. Quantitative PCR analyses were performed in both groups to evaluate differences in the mRNA expression of TYMS, APOL3, and FBXO2. Serum was separated by centrifugation at 3000 rpm for 10 min, after which total RNA was isolated using a commercial serum RNA extraction kit. The purified RNA was subsequently reverse-transcribed into complementary DNA (cDNA) using a reverse transcription kit, and cDNA concentrations were normalized across all samples prior to downstream analyses. Gene-specific primers were obtained from Sangon Biotech (Shanghai, China), and the corresponding sequences are listed in Table 1. All primers were carefully designed to ensure optimal amplification efficiency. Quantitative real-time PCR was then conducted on an Applied Biosystems QuantStudio 6 system. Relative gene expression levels were calculated and statistically compared between groups to identify differential expression patterns.

Table 1 .

RT-qPCR primer sequences.

Genes	Forward	Sequences
ACTB	Forward primer	CCACCATGTACCCTGGCATT
	Reverse primer	GTCCTCGGCCACATTGTGAA
TYMS	Forward primer	TACAGCCTGAGAGATGAATTCCC
	Reverse primer	GGATTCCAAGCGCACATGAT
APOL3	Forward primer	CAGTGAGAACGGCATGTTGC
	Reverse primer	TCTCTGCAATCCCTTTGCGT
FBXO2	Forward primer	CCCACACCTTCACCGACTAC
	Reverse primer	GCTCAGGGTTCTACCCACAC

Statistical analysis

All data preprocessing and statistical analyses were performed using R (version 4.1.0) within the RStudio environment; the main R packages and their versions are detailed in the corresponding sections of the methodology. Raw expression matrices from the GEO, TCGA, and ICGC databases were subjected to background correction, quantile normalization, and batch effect correction, respectively. All statistical tests were two-sided, and p < 0.05 was considered statistically significant.

Results

Screening of key gene modules in the GEO dataset

For the WGCNA analysis results of GSE244826, the Scale Independence plot demonstrates the scale-free topology of the network under varying soft-thresholding power conditions. The results of this study show that when the soft-thresholding power is set to 8, the network reaches the desired scale-free state (Fig. 1A). In addition, the Mean Connectivity plot indicates that the average connectivity remains within a reasonable range at power = 8 in this study.

Fig. 1.

WGCNA analysis and identification of key gene modules in the GSE244826 dataset. (A) Scale independence and mean connectivity plots for selection of soft-thresholding power. (B) Gene dendrogram with assigned module colors. (C) Heatmap of module–trait relationships. (D) Volcano plot of differentially expressed genes. (E) KEGG enrichment analysis of selected gene modules. (F) GO enrichment analysis of selected gene modules.

The Gene Dendrogram and Module Colors plot provides an intuitive representation of the clustering relationships among all genes and divides genes into different modules based on co-expression patterns. The branches at the top of the dendrogram represent individual genes, while the colored bar at the bottom indicates module assignment (Fig. 1B). In the co-expression network constructed in this study, genes are classified into multiple modules, each corresponding to a unique co-expression characteristic, thus laying the foundation for subsequent correlation analyses between modules and clinical traits.

Finally, the Module–Trait Relationships heatmap displays the correlations and significance between different gene modules and key clinical tumor characteristics, as derived from the GSE244826 dataset. The plotted results show that the orange, darkorange, and lightyellow modules are significantly correlated with patient prognosis and tumor characteristics (p < 0.05). The core genes of these modules will be further screened as potential targets for precise therapy in hepatocellular carcinoma (Fig. 1C).

A volcano plot was generated to visualize gene expression levels in GSE244826, with a p-value (P.Val) cutoff of 0.05 and |logFC|= 1 as the threshold, resulting in a total of 2,012 differentially expressed genes. Upregulated genes are shown in red and downregulated genes in green (Fig. 1D). KEGG and GO enrichment analyses of the genes screened by WGCNA revealed that they are mainly enriched in drug metabolism, cytochrome P450–related pathways, and other drug metabolism pathways (Fig. 1E and F). Notably, the pathways of “Chemical Carcinogenesis” and “Mismatch Repair” are implicated in the pathogenesis of HCC, with alterations in these pathways potentially contributing to tumor initiation and progression by affecting DNA repair mechanisms and carcinogen metabolism.

VENN identifies key genes and integration of different GEO datasets

By intersecting WGCNA and DEGs, 93 key differential genes were obtained (Fig. 2A). KEGG and GO enrichment analyses indicated that these key differential genes are concentrated in drug metabolism and cytochrome P450-related pathways (Figs. 2B and C).

Fig. 2.

Integration and analysis of key genes across multiple GEO datasets. (A) Venn diagram showing the intersection of WGCNA modules and DEGs (n = 93). (B) KEGG enrichment analysis of key genes. (C) GO enrichment analysis of key genes. (D) Bar plots of gene expression levels before and after batch effect removal. (E) UMAP plots of sample distribution before and after batch effect correction.

Further, the datasets GSE208036, GSE233421, and GSE273947 were obtained. Due to significant batch effects among different datasets, rigorous batch effect removal was performed on the merged data. Systematic evaluation of expression distribution and sample clustering was conducted before and after data integration. From the bar expression plots, prior to batch effect removal, noticeable differences in gene expression levels were observed across batches, demonstrated by inconsistent bar heights of the same genes among different datasets. The fluctuations between batches exceeded the range of normal biological variation. After batch effect removal, the bar expression plots showed that expression levels of the same genes across datasets tended to be consistent, and the overall distribution became more uniform. Data commonalities were highlighted, and non-biological differences were significantly reduced (Fig. 2D).

In the UMAP dimensionality reduction visualization, prior to batch effect removal, samples clustered primarily according to dataset origin, with samples from the same batch grouping together, while samples from different batches were dispersed and independent. After batch effect correction, the sample distribution pattern in the UMAP plot changed markedly. Samples from different sources merged and clustered based on biological features rather than dataset origin. This indicates that batch effects were effectively eliminated (Fig. 2E).

Construction of ML and predictive models

Figure 3A illustrates the AUC distribution of various machine learning algorithm combinations (including dominant strategies such as Lasso regression, Ridge regression, Elastic Net, and their integrated derivative schemes) in the stratification and classification task of hepatocellular carcinoma patients. The x-axis represents the serial numbers of different algorithm combinations; values inside the boxes indicate AUC scores, with the average values highlighted in blue boxes. Among the various algorithmic combinations evaluated, the final model was constructed using LASSO regression for feature selection and a Cox proportional hazards model for survival prediction. This combination yielded favorable C-index values across multiple validation cohorts and was selected for its robustness and interpretability. Machine learning ultimately selected 10 genes: APOL3, FBXO2, GSTA4, LGALS4, PDZK1IP1, SDCBP2, SPINT2, SYT7, TYMS, and UGT1A3.To further investigate potential regulatory crosstalk and mechanistic connections among the ten key genes identified in our prognostic model, we performed a gene interaction network analysis using the GeneMANIA platform. The resulting network revealed extensive co-expression links and genetic interactions among core genes such as TYMS, GSTA4, SPINT2, and UGT1A3 (Figure S1). Functional enrichment indicated that these genes are predominantly involved in organic acid binding, glucuronate and uronic acid metabolism, and retinoid binding pathways. These findings suggest that the predictive gene signature may not only be statistically robust but also biologically coherent, particularly in pathways related to metabolic remodeling and immune modulation in HCC.

Fig. 3.

Construction and evaluation of machine learning-based predictive models for hepatocellular carcinoma. (A) AUC values for various machine learning algorithm combinations used in patient stratification. (B, C) Kaplan–Meier survival curves comparing high- and low-risk groups based on the ModelRiskScore in ICGC and TCGA cohorts. (D) Multivariate Cox regression analysis of the selected gene features. (E) Risk score analysis showing gene-specific contributions, with TYMS as a risk factor and APOL3 and FBXO2 as protective factors.

Based on the above genes, predictive models were constructed, and Cox regression analysis and ModelRiskScore were conducted. We performed Cox regression analysis based on the key genes mentioned above and found that the selected feature genes have significant predictive value for overall survival in patients. The multivariate Cox regression model demonstrated that high expression of some genes is closely associated with poor prognosis, indicating that these genes may serve as independent prognostic risk factors (Fig. 3D). Furthermore, we developed the ModelRiskScore risk scoring model using these genes, in which TYMS acts as a risk factor, while APOL3 and FBXO2 serve as potential protective factors (Fig. 3E).

Survival differences between high-risk and low-risk patient groups were analyzed using survival data from the hepatocellular carcinoma cohorts in the ICGC and TCGA databases, respectively. The results show that this score effectively distinguishes between high-risk and low-risk groups, with significant survival differences observed in survival analysis (Figs. 3B and C). The genes identified through combined screening have promising applications in prognosis evaluation and personalized immunotherapy.

Differential analysis of candidate genes under various clinical characteristics

In this study, we conducted a multi-dataset integrated screening of hepatocellular carcinoma patients to identify potential gene targets related to prognosis and immunotherapy. Patients were grouped and analyzed based on clinical indicators including age (cutoff: 65 years), BMI (cutoff: 25), gender, tumor grade and stage, and whether they received radiochemotherapy. The results showed that the expression of candidate genes differed significantly between groups stratified by T stage (p < 0.05) (Fig. 4E). However, no statistically significant differences in candidate gene expression were observed among groups stratified by age (Fig. 4A), BMI (Fig. 4B), gender (Fig. 4C), tumor grade, or radiochemotherapy status (Fig. 4D) (p > 0.05). These findings suggest that the identified genes are primarily associated with tumor T stage, while their correlation with other clinical characteristics is not significant, providing a reference for subsequent research targeting precise immunotherapy based on T stage.

Fig. 4.

Differential expression analysis of candidate genes across clinical subgroups in hepatocellular carcinoma. (A–D) Comparison of candidate gene expression among groups by age, BMI, gender, and radiochemotherapy status, showing no significant differences. (E) Differential expression of candidate genes among groups stratified by T stage, with significant differences observed (p < 0.05).

Analysis of survival differences and gene mutation profiling

Within the TCGA database, overall survival (OS), disease-free survival (DFS), disease-specific survival (DSS), and progression-free survival (PFS) were evaluated. The results indicated that the differences between the high-risk and low-risk groups were statistically significant for all endpoints (all p < 0.05) (Fig. 5A and B). In the ICGC database, survival analysis revealed a significant difference in overall survival (OS) between the high- and low-risk groups (P = 0.034), with the high-risk group having a poorer prognosis (Fig. 5A). These findings suggest that the risk model established through combined multi-dataset screening demonstrates robust prognostic stratification capability in hepatocellular carcinoma patients and provides important reference value for the identification of immunotherapeutic targets (Fig. 5C).

Fig. 5.

Survival analysis and mutation profiling of risk groups in hepatocellular carcinoma. (A, B) Kaplan–Meier survival curves for overall survival (OS), disease-free survival (DFS), disease-specific survival (DSS), and progression-free survival (PFS) between high- and low-risk groups in the TCGA cohort, all showing significant differences (p < 0.05). (C) Kaplan–Meier curve of overall survival differences between high- and low-risk groups in the ICGC cohort. (D) Mutation profile comparison between high- and low-risk groups, highlighting higher mutation frequencies and enrichment of key driver genes in the high-risk group.

To further investigate the molecular characteristics of different risk groups, we conducted a comparative analysis of gene alterations among patients. The results showed that the mutation frequency in the high-risk group was generally higher than that in the low-risk group, and certain key driver genes (such as TP53, TNN, and MUC) exhibited higher mutation proportions in the high-risk group (Fig. 5D). In addition, there were notable differences in the distribution of some types of gene mutations between the different risk groups.

Analysis of immune cell infiltration and immunotherapy response

We systematically evaluated the characteristics of immune cell infiltration and the potential response of patients to immunotherapy. To ensure the reliability of the estimation results, multiple mainstream immune cell infiltration prediction tools, including TIMER, EPIC, MCPcounter, Quantiseq, ESTIMATE, and xCell, were employed to cross-validate the abundance of major immune cell subsets. The results demonstrated a high degree of consistency. Among them, endothelial cells, CD4 + T cells, M2 macrophages, and B cells showed high consistency across the ICGC and TCGA cohorts (Fig. 6A).

Fig. 6.

Immune cell infiltration analysis and immunotherapy response in hepatocellular carcinoma. (A) Cross-validation of immune cell infiltration using TIMER, EPIC, MCPcounter, Quantiseq, ESTIMATE, and xCell algorithms demonstrates consistent abundance patterns of major immune cell subsets, especially endothelial cells, CD4 + T cells, M2 macrophages, and B cells, across ICGC and TCGA cohorts. (B, C) Assessment of immune checkpoint inhibitor response in the Riaz 2018 cohort and the Lauss 2017 cohort, showing significant differences between patient groups. (D) Progression-free survival (PFS) outcomes in the Lauss 2017 cohort stratified by risk group.

By assessing the efficacy of immune checkpoint inhibitors in several publicly available clinical cohorts, the results indicated heterogeneity in patient responses to immunotherapy between different cohorts. Specifically, the Riaz cohort 2018 and the Lauss cohort 2017 both exhibited significant differences in therapeutic response to immune checkpoint inhibitors (Fig. 6B, C, and D).

PCR detection of key gene expression levels

Quantitative PCR analysis of TYMS, APOL3, and FBXO2 revealed distinct expression patterns between HCC patients and healthy controls (Fig. 7). TYMS mRNA levels were significantly elevated in HCC patients, reflecting a gene expression profile associated with reduced sensitivity to chemotherapy. In contrast, APOL3 expression did not differ significantly between the two groups; as a protein involved in autophagy, endoplasmic reticulum stress, and antiviral responses, this finding underscores the complex molecular landscape of HCC. Notably, FBXO2 expression was markedly downregulated in HCC patients, consistent with its reported role in exerting antitumor effects through the degradation of specific glycoproteins.

Fig. 7.

Differential expression of senescence-associated genes in the serum of healthy individuals and HCC patients. qPCR analysis demonstrated a significant upregulation of TYMS and a marked downregulation of FBXO2 in HCC patients, whereas APOL3 expression did not differ significantly between groups. Data are presented as mean ± SD; ns, not significant, p < 0.05.

Discussion

This study systematically identified and validated key target genes related to the prognosis and immunotherapy of HCC by integrating large-scale transcriptomic data from multiple sources, including TCGA, GEO, and ICGC. Combining network co-expression analysis, differential gene expression screening, and batch effect correction, we constructed a robust biomarker screening and risk prediction model based on multi-omics evidence and machine learning methods, ultimately selecting ten key genes. Furthermore, the prognostic predictive ability of these genes across different clinical subgroups, as well as their associations with the immune microenvironment and relevant therapeutic targets, were comprehensively evaluated using time-dependent ROC curves, Kaplan–Meier survival analysis, decision curve analysis, and other approaches. The results indicated that TYMS is a significant risk factor, while APOL3 and FBXO2 serve as potential protective factors, providing a novel molecular foundation for clinical stratification and personalized treatment of HCC.

Compared with previous studies on hepatocellular carcinoma and other solid malignancies, this research demonstrates notable multidimensional innovation. On one hand, most prior work mainly relied on limited single-center clinical samples or a single database, making it difficult to exclude sample bias or data noise that could affect key gene selection.^21–23 Although some recent studies have attempted to integrate multi-omics data, they often lack effective batch effect correction and evidence-based bioinformatics, resulting in limited comparability and clinical applicability of their findings.^24,25 On the other hand, investigations into predictive biomarkers for HCC immunotherapy have predominantly focused on classic genes such as PD-1/PD-L1 signaling axis, TP53, and CTNNB1. In contrast, our study achieves cross-validation across multiple data sources and indicators at the molecular level and strengthens the theoretical basis for personalized prognosis through patient stratification and subgroup sensitivity analyses. Additionally, by integrating analyses of immune infiltration, gene mutations, various survival analyses, and correlations with immunotherapy targets, this work provides a systematic framework to further elucidate molecular mechanisms and translational value. Therefore, our methods and findings serve as both a model and a complementary contribution to the precise dissection of tumor heterogeneity and molecular subtyping research in complex cancers.

Among the key genes screened in this study, thymidylate synthase (TYMS), the core enzyme gene in the folate metabolism and DNA synthesis pathway, stands out prominently as a high-risk feature.^26–28 TYMS has been confirmed to be closely associated with tumor proliferation, metastasis, and drug resistance in various solid tumors such as colorectal cancer, lung cancer, and gastric cancer.^29–31 Its overexpression not only promotes the rate of DNA synthesis and enhances tumor cell proliferation but also mediates changes in the immune microenvironment, thereby affecting tumor immune evasion and the efficacy of immune checkpoint inhibitors. In HCC, relevant reports indicate that upregulation of TYMS is closely related to poor prognosis, increased risk of postoperative recurrence, and resistance to first-line chemotherapy agents such as fluoropyrimidines.³² This study further found that patients with high TYMS expression exhibited significantly poorer survival outcomes and a trend toward lower responsiveness to immunotherapy, suggesting that TYMS may serve as a molecular prognostic marker and a novel target for intervention in HCC. Additionally, gene mutations or dysregulation of TYMS may potentially reveal mechanisms underlying tumor drug resistance, providing molecular evidence for optimizing combination treatment strategies.^33,34 Furthermore, the association between TYMS expression and immune evasion highlights its potential as a predictive biomarker for the effectiveness of ICIs in HCC, suggesting that targeting TYMS could improve ICI-based treatment strategies.

Regarding protective factors, the expression of APOL3 and FBXO2 is negatively correlated, indicating their potential tumor suppressive functions.^35,36 APOL3, a member of the apolipoprotein L family, plays a critical role in regulating cellular autophagy, endoplasmic reticulum stress, and antiviral responses.^37,38 In HCC, the specific functions of APOL3 are still in early stages of exploration, but preliminary results suggest that its upregulation may enhance cellular homeostasis and impede tumor growth. FBXO2, a member of the F-box protein family involved in the ubiquitin–proteasome pathway, may promote the degradation of tumor-associated proteins and participate in regulating the cell cycle and apoptosis. Existing studies in prostate and breast cancers have shown that FBXO2 exerts antitumor effects by degrading specific glycoproteins. Our analysis indicates that high FBXO2 expression is closely associated with longer progression-free survival and favorable immune infiltration levels, providing strong evidence for its potential as a protective biomarker in HCC.

The establishment of a core gene predictive model for HCC is of significant importance for improving patient stratification, prognostic evaluation, and individualized treatment.^39,40 The ten-gene risk model established in this study demonstrates strong potential for clinical application, effectively distinguishing between high- and low-risk subgroups while accurately predicting overall survival and progression-free survival. Importantly, the model exhibits robust stability and generalizability across various clinical factors, including age, disease stage, and hepatitis B background, which are key components in the management of HCC as per established clinical protocols. Additionally, by integrating immune infiltration and key immune checkpoint analyses, the model offers valuable insights for identifying patient populations most likely to benefit from immunotherapy, as well as assessing treatment responses. As precision medicine continues to advance, multi-gene models such as the one proposed here are increasingly recognized as pivotal tools in tumor diagnosis and treatment. These models align with current trends in HCC management guidelines, contributing to the refinement of clinical decision-making frameworks and providing more tailored, effective management strategies for HCC patients.

This study also has several limitations. First, although the public database samples used are multi-sourced, they predominantly consist of populations from Europe, America, and East Asia, and lack large-scale prospective multicenter validation. The external reproducibility of the model requires further strengthening. Second, transcriptome data are limited in capturing dynamic regulation and functional execution at the protein level; expression changes of some genes may not fully align with their actual biological activities. Third, while machine learning algorithms have enhanced the robustness of the screening process, issues such as model overfitting and feature redundancy may still affect practical application. Future research should integrate real-world cohorts, single-cell sequencing, multidimensional validation, and mechanistic experiments to further elucidate the functions of core genes and expand the breadth and depth of the study.

Conclusion

In summary, this study systematically screened and validated HCC prognosis-related molecular markers based on the integration of multiple datasets and multi-omics strategies, and established a multi-gene predictive model. These findings not only deepen the understanding of molecular heterogeneity and mechanisms of drug resistance in hepatocellular carcinoma, but also provide theoretical support for individualized patient stratification and immunotherapy decision-making. The identified key genes lay a solid foundation for subsequent molecular mechanism studies and the development of novel therapeutic targets. In the future, efforts should be made to strengthen multi-center large-sample and translational research to promote the clinical application and precision treatment of these findings, ultimately improving the overall survival and quality of life of patients with HCC.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Lichen Zhou, Wenjie Zhang: project development, data analysis, manuscript writing Zhuoran Liu, Yaming Xie: project development, data analysis & collection Kangyi Jiang: project development, manuscript editing.

Data availability

The data that support the findings of this study are available on request from the corresponding author, upon reasonable request.

Declarations

Competing interests

The authors declare that none of the information in this study may have been affected by personal or financial links or interests that were known to be in conflict.