Decoding the cholesterol–apoptosis axis in HCC: a machine learning-based multi-omics integration and single-cell transcriptomic analysis

Abstract

Liver hepatocellular carcinoma (LIHC), a predominant form of primary hepatic malignancy, demonstrates a progressively escalating global incidence, imposing substantial health and economic burdens on patients and society. Early diagnosis remains challenging, often resulting in late-stage detection, which limits the efficacy of current therapeutic strategies. This study systematically examines the transcriptional signatures of apoptosis-associated and cholesterol metabolic pathways in LIHC, providing insights into its underlying mechanisms and identifying potential prognostic markers. We employed multi-omics and machine learning to evaluate gene expression variations and construct a prognostic risk scoring model. This study identified apoptosis- and cholesterol metabolism-related differentially expressed genes (ACMRDEGs). Importantly, LASSO regression analysis identified six hub genes (EPHX2, FABP5, SQLE, ADH4, HMGCS2, and CYP7A1) as critical prognostic biomarkers, demonstrating significant correlation with overall survival (OS). Furthermore, immune cell infiltration analysis indicated significant differences in 12 immune cell types within LIHC microenvironment, underscoring the immune system’s involvement in disease progression. cholesterol and alcohol metabolism pathways were significantly enriched among hub gene modules, as quantified by multiple gene enrichment analyses. Single-cell analysis identified six major cell types, providing a deeper understanding of the cellular heterogeneity within LIHC. In summarize, this study presents the first integrated apoptosis–cholesterol metabolic pathway-based six-gene prognostic model for LIHC, validated for robustness across multiple cohorts, which may facilitate personalized therapeutic strategies and refined risk assessment in clinical practice.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-04010-z.

Introduction

Liver hepatocellular carcinoma (LIHC), the predominant histological type dominating primary hepatic neoplasms, represents a critical global health challenge, accounting for approximately 90% of hepatic malignancies, while its mortality trajectory positions it as the third most lethal malignancy globally [1–3]. Chronic hepatic pathologies, notably viral hepatitis (HBV/HCV), alcohol-induced hepatotoxicity, and non-alcoholic fatty liver disease (NAFLD), constitute principal etiological drivers in hepatocarcinogenesis [4]. Despite advancements in therapeutic strategies—including surgical ablation, precision radiotherapy, cytotoxic regimens, kinase inhibitors, and immune checkpoint blockade—the 5-year survival probability for advanced LIHC persistently constrained to < 15% [5–10], largely due to delayed diagnosis, tumor heterogeneity, and acquired drug resistance [11–13]. This bleak outlook highlights the urgent necessity to clarify the molecular mechanisms driving LIHC progression and discover new biomarkers for early detection and personalized treatment.

Current research efforts have increasingly focused on the molecular subtyping, underlying pathogenesis, and therapeutic strategy of LIHC, particularly the roles of apoptosis evasion, dysregulated cholesterol metabolism, and immune microenvironment alteration [14–16]. Apoptosis, or programmed cell death, is critical in maintaining cellular homeostasis and its dysregulation has been implicated in multiple cancer progression and resistance to therapy [17]. Furthermore, alterations in cholesterol metabolism have been identified as significant contributors to tumorigenesis and cancer cell proliferation [18]. Notably, the interplay between these two biological pathways remains poorly characterized in LIHC, presenting a pivotal knowledge gap. Our study addresses this by systematically investigating apoptosis- and cholesterol metabolism-related differentially expressed genes (ACMRDEGs), which may serve as molecular linchpins connecting metabolic reprogramming to malignant phenotypes.

Through systematic analysis of LIHC samples and their association with apoptosis and cholesterol metabolism, this research aims to fill gap in current studies and provide new theoretical foundations and practical guidance for future clinical applications. We integrated multi-omics bioinformatics approaches and machine learning across four dimensions: (1) analyzing transcriptomic data from TCGA-LIHC and GEO cohorts to identify ACMRDEGs; (2) using Cox proportional hazards and LASSO regression to develop a prognostic risk model; (3) performing CIBERSORT-based immune infiltration analysis to characterize tumor microenvironment dynamics; and (4) utilizing single-cell sequencing to determine cellular components. This methodology enables a thorough investigation of ACMRDEGs expression change in LIHC samples, facilitating the development of a novel prognostic risk model.

While prior research has predominantly centered on isolated pathway models (such as cholesterol or fatty acid metabolism), few studies have addressed the interplay between distinct biological processes. This study provides mechanistic insights into LIHC pathobiology through an integrated apoptosis-cholesterol metabolism axis and proposes a clinically applicable tool for individualized prognosis assessment.

Materials and methods

Transcriptomic data acquisition

In this study, we utilized the R package TCGAbiolinks [19] (Version 2.30.4) to download the TCGA-LIHC dataset (https://portal.gdc.cancer.gov/). This dataset served as the primary dataset for subsequent analyses and validation. The LIHC dataset comprises 59 control samples (Group: Control) and 380 liver cancer samples (Group: LIHC). After excluding disease samples with missing clinical survival data, a total of 368 liver cancer samples were included for prognostic analysis. Clinical data corresponding to these samples were retrieved from the UCSC Xena database (https://xena.ucsc.edu/) [20], with specific details provided in Table 1.

Table 1.

LIHC data set information list

	TCGA-LIHC	GSE121248	GSE76427	GSE149614
Platform	/	GPL570	GPL10558	GPL24676
Species	Homo sapiens	Homo sapiens	Homo sapiens	Homo sapiens
Tissue	Liver	Liver	Liver	Liver
Samples in LIHC group	368	70	115	10
Samples in Control group	50	37	52	8
Reference	/	17,975,138	29,117,471	35,933,472

TCGA, The Cancer Genome Atlas; GEO, Gene Expression Omnibus; LIHC, Liver Hepatocellular Carcinoma

Table 2.

Clinical characteristics of LIHC patients in the TCGA-LIHC datasets

Characteristics	Control	LIHC	P value
n	50	368
Gender, n (%)			0.102
Female	22 (5.3%)	119 (28.5%)
Male	28 (6.7%)	249 (59.6%)
Age, n (%)			0.083
> 60	33 (7.9%)	195 (46.7%)
<=60	17 (4.1%)	173 (41.4%)
Stage, n (%)			0.765
Stage III	12 (3.1%)	83 (21.5%)
Stage I	18 (4.7%)	172 (44.6%)
Stage II	11 (2.8%)	85 (22%)
Stage IV	1 (0.3%)	4 (1%)
T, n (%)			0.548
T2	14 (3.4%)	92 (22.2%)
T3	13 (3.1%)	78 (18.8%)
T4	3 (0.7%)	13 (3.1%)
T1	20 (4.8%)	182 (43.9%)
N, n (%)			1.000
N0	31 (10.8%)	250 (87.4%)
N1	1 (0.3%)	4 (1.4%)
M, n (%)			0.937
M0	33 (10.9%)	265 (87.7%)
M1	1 (0.3%)	3 (1%)
OS.time, median (IQR)	686.5 (374.5, 1369.8)	601 (344.75, 1102.2)	0.235
OS, n (%)			< 0.001
Dead	35 (8.4%)	131 (31.3%)
Alive	15 (3.6%)	237 (56.7%)

TCGA, The Cancer Genome Atlas; LIHC, Liver Hepatocellular Carcinoma

Additionally, we acquired liver cancer datasets GSE121248 [21] and GSE76427 [22] from the GEO database (https://www.ncbi.nlm.nih.gov/geo/) [23] via the GEOquery R package (Version 2.70.0) [24]. Dataset GSE121248 includes 70 liver cancer samples (Group: LIHC) and 37 adjacent non-cancerous tissue samples (Group: Control), all derived from Homo sapiens and classified based on the GPL570 chip platform. Detailed information on the dataset is available in Table 1. Similarly, dataset GSE76427 consists of 115 liver cancer samples (Group: LIHC) and 52 adjacent non-cancerous tissue samples (Group: Control), derived from hepatocellular carcinoma tumor tissues, and annotated using the GPL10558 chip platform. All LIHC group samples from both GSE121248 and GSE76427 were incorporated into the model validation analysis. For data preprocessing, we applied the limma R package (Version 3.58.1) [25] to perform normalization, annotation of probes, and standardization of the data from GSE121248 and GSE76427.

Furthermore, apoptosis-related genes (ARGs) were retrieved from the GeneCards database (https://www.genecards.org/) [26]. Using “Apoptosis” as a search keyword and filtering for “Protein Coding” genes, a total of 13,297 ARGs were identified (refer to Table S1). Additionally, we collected a set of cholesterol metabolism-related genes (CMRGs) from the published literatures by searching “cholesterol metabolism” in PubMed (https://pubmed.ncbi.nlm.nih.gov/) [27], which resulted in the identification of 140 CMRGs (refer to Table S2).

Single-cell data acquisition

The GEO database provides an extensive repository of single-cell sequencing data, and in this study, we accessed the LIHC single-cell RNA sequencing dataset GSE149614 [28], which is based on the GPL24676 Illumina NovaSeq 6000 platform. The dataset includes 10 primary tumor samples and 8 adjacent non-tumor liver tissue samples, all derived from Homo sapiens, which were selected for inclusion in the single-cell analysis portion of this study.

Using the Seurat R package (Version 5.2.1) [29], we employed the “CreateSeuratObject” function to import the single-cell RNA-seq matrix for liver cancer and control groups from the GSE149614 dataset, creating a Seurat object with a minimum of 3 cells per gene and at least 200 genes per cell.

Subsequently, we normalized the sequencing depth of the dataset using the SCTransform function. Principal component analysis (PCA) was applied to identify significant principal components (PCs), and the “ElbowPlot” function was used to visualize the distribution of P-values. A total of 30 principal components were selected for further analysis using Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction. K-nearest neighbor (KNN) clustering was performed in the PCA space with default parameters using 30 PCs, and cell clusters were identified based on a resolution of 0.3, as determined by the FindClusters function and visualized with the clustree function. Finally, t-SNE (t-distributed stochastic neighbor embedding) was applied for further dimensionality reduction to facilitate data visualization and exploration.

Cell types within the single-cell dataset were annotated using cell-type-specific biomarkers (refer to Table S3) to identify and calculate the proportions of different cell types in the dataset. Genes expression across different cell types were calculated.

Differentially expressed genes (DEGs) with apoptosis and cholesterol metabolism in LIHC

The TCGA-LIHC dataset was categorized into two groups, namely the LIHC group and the control group, based on sample stratification. Differential gene expression analysis was performed using the DESeq2 R package (Version 1.42.0) [30] between the LIHC and control groups. A threshold of |logFold Change (FC)| >1 and p.value < 0.05 was set to identify DEGs. Specifically, genes with logFC >1 and p.value < 0.05 were classified as up-regulated genes, while those with logFC < -1 and p.value < 0.05 were considered down-regulated genes. The results of the differential expression analysis were visualized using a volcano plot generated by the ggplot2 R package (Version 3.4.4).

To identify the genes associated with apoptosis and cholesterol metabolism in LIHC, we performed an intersection analysis between the DEGs (|logFC| >1 and p.value < 0.05) obtained from the TCGA-LIHC dataset and ARGs and CMRGs. The overlapping genes were visualized using a Venn diagram to identify the ACMRDEGs in LIHC.

Finally, we employed univariate Cox regression analysis using the survival R package (Version 3.5-7) [31] to screen genes with p.value < 0.05, which were then used for the construction of a prognostic model based on the ACMRDEGs.

Construction and validation of the ACMRDEGs-related prognostic risk model for LIHC

In order to establish a prognostic risk model for LIHC, we conducted univariate and multivariate Cox regression analyses based on clinical information of the TCGA-LIHC dataset to assess the prognostic significance of the ACMRDEGs. These analyses helped determine whether these genes were independent prognostic factors for LIHC.

We began by performing univariate Cox regression analysis to select variables with a p.value < 0.05, which were subsequently subjected to Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis [32]. LASSO regression, implemented using the glmnet R package (Version 4.1-8) [33] with the parameter of “family = cox”, was conducted with 10 iterations. This method, which introduces a penalty term (lambda × absolute value of coefficients), reduces model overfitting while enhancing its generalizability. The results of the LASSO regression were visualized using prognostic risk model plots and coefficient trajectory plots. The final hub genes were identified through the LASSO regression, followed a multivariate Cox regression analysis to calculate the risk score for each patient. The formula is as follows:

The results of the multivariate Cox regression analysis, including the expression levels of hub genes, were visualized using a forest plot.

To assess the accuracy and resolution of the prognostic risk model, patients in the TCGA-LIHC dataset were stratified into high-risk and low-risk groups based on the median value of the Cox risks core. Kaplan-Meier (K-M) curve [34] was performed using the survival R package (Version 3.5-7) [31] to compare overall survival (OS) between these groups.

To further evaluate the performance of the prognostic model, we generated time-dependent Receiver Operating Characteristic (ROC) curves using the survivalROC R package (Version 1.0.3.1) [35]. The area under the ROC curve (AUC) was calculated to assess the predictive ability of the model for 1-year, 3-year, and 5-year survival. AUC values range from 0.5 to 1, with higher AUC values indicating better diagnostic accuracy. An AUC >0.5 suggests that the molecular expression trend promotes the event, with an AUC between 0.5 and 0.7 indicating lower accuracy, between 0.7 and 0.9 indicating moderate accuracy, and an AUC >0.9 indicating high accuracy.

Additionally, calibration curves were plotted to assess the goodness-of-fit between the observed and predicted probabilities of the model’s outcomes. Decision Curve Analysis (DCA) [36], performed using the ggDCA R package (Version 1.1), was employed to further evaluate the clinical utility of the prognostic risk model.

Finally, the hub genes identified from the TCGA-LIHC dataset were used to construct models in the GSE121248 and GSE76427 datasets. These datasets were similarly stratified into high-risk and low-risk groups based on the median risk score, and the expression differences of the hub genes between these groups were visualized through boxplots. The predictive ability of the hub genes for LIHC was further evaluated using ROC curves and AUC calculation in the GSE121248 and GSE76427 datasets by employing pROC R package (Version 1.18.5) [37].

Establishment and evaluation of prognostic risk model combining hub genes and clinical indicators for LIHC

Given the potential incremental predictive value of clinical variables, we performed univariate and multivariate analyses on clinical parameters within the TCGA-LIHC dataset-including pathological TNM stage, alpha-fetoprotein (AFP) level, fibrosis Ishak score, and Child-Pugh grade. Variables with p < 0.2 were incorporated into an updated predictive model combining both clinical factors and hub genes. The performance of the updated model was further evaluated using AUC, calibration curves, and DCA to assess its predictive accuracy and stability. Additionally, we employed the survIDINRI package to calculate the integrated discrimination improvement (IDI) index [38] in order to quantify the degree of model enhancement.

Functional annotation profiling

Gene Ontology (GO) annotation [39] was systematically conducted across three ontological domains: biological processes (BP), molecular functions (MF), and cellular components (CC), utilizing the clusterProfiler toolkit (v4.10.0) [40] for large-scale functional categorization. Concurrently, pathway-centric interrogation was performed through Kyoto Encyclopedia of Genes and Genomes (KEGG) [41] pathway mapping to elucidate disease-relevant molecular networks. Significance thresholds were defined as Benjamini-Hochberg (B-H) adjusted p-value (p.adj) < 0.05 coupled with a false discovery rate (FDR) < 0.05 to ensure stringent multiple testing correction.

Gene set enrichment framework

Phenotype-associated pathway dysregulation was assessed via Gene Set Enrichment Analysis (GSEA) following transcriptomic ranking by logFC differential expression magnitudes [42]. The analytical workflow incorporated:

1.Preprocessing: TCGA-LIHC cohort stratification into risk subgroups (high vs. low) based on prognostic signatures; 2.Gene Prioritization: Directional ranking of differentially expressed genes (DEGs) by phenotypic correlation; 3.Computational Execution: Implementation through clusterProfiler (v4.10.0) [40] with parameters: Permutation number: 1,000, Gene set size constraints: 10 ≤ n ≤ 500, Random seed initialization: 2022; 4.Reference Database: Curated gene sets from MSigDB [43] c2.cp.all.v2022.1.Hs (3,050 canonical pathways).

Significant enrichment required dual thresholds of p.adj < 0.05 and FDR < 0.05, maintaining analytical rigor.

Protein–protein interaction (PPI) network

PPI networks represent the interactions between individual proteins, which are essential for numerous biological processes, including signal transduction, gene expression regulation, energy metabolism, and cell cycle regulation. The comprehensive analysis of protein interactions within biological systems is pivotal for understanding the functional mechanisms of proteins, the regulatory mechanisms of biological signals and metabolic pathways, particularly in the context of disease and other physiological states [44]. Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) [44] is a widely-used database that provides information on known and predicted protein-protein interactions. It includes data from experimental results, literature mining of PubMed abstracts, and predictions based on bioinformatics approaches. STRING encompasses 14,094 species, containing 67.59 million proteins and over 200.5 million protein-protein interactions. In our study, a PPI network was constructed using the prognostic genes, with the interaction score set at a minimum threshold of 0.150 (medium confidence). The network was visualized using Cytoscape software [45]. In such networks, closely connected local regions may represent molecular complexes with specific biological functions. Genes with interactions in the PPI network were selected for further analysis.

Additionally, microRNAs (miRNAs) play crucial roles in regulating gene expression during biological development and evolution. A single target gene can be regulated by multiple miRNAs, and conversely, one miRNA can regulate various target genes. The miRDB database is a resource designed for the prediction of miRNA target genes and their functional annotations [46]. We utilized the miRDB database to predict miRNAs that interact with the hub genes. These interactions were subsequently visualized in an mRNA-miRNA regulatory network using Cytoscape.

Immune infiltration analysis

CIBERSORT is an analytical tool based on linear support vector regression, used for deconvoluting transcriptomic expression matrices to estimate the composition and abundance of immune cells in mixed cell populations [47]. Using the CIBERSORT algorithm in conjunction with the LM22 gene signature matrix, immune cell enrichment scores greater than zero were filtered and used to construct the immune cell infiltration matrix for the TCGA-LIHC dataset. Subsequently, the significant differences in the expression of LM22 immune cells across various sample groups were visualized using the R package ggplot2 (Version 3.4.4). Immune cell types that exhibited significant differences between the groups were selected for further analysis. The correlation between these immune cells was calculated using the Spearman correlation method, and the results were displayed as a heatmap using the R package heatmap (Version 1.0.12). Finally, the correlation between hub genes identified in the prognostic model and immune cells was analyzed, with results considered significant if p-value < 0.05. These correlations were visualized using a bubble chart created with ggplot2 (Version 3.4.4). Furthermore, the above analyses were validated using additional algorithms (e.g., ssGSEA, xCell) and the TIMER 3.0 database (an open-access resource for tumor immune analysis) [48].

Proteomics analysis and immunohistochemical (IHC) validation

To validate the expression of the six hub genes at the protein level, we performed proteomic analysis using The University of Alabama at Birmingham Cancer Data Analysis Portal (UALCAN) (https://ualcan.path.uab.edu/analysis-prot.html) [49], an open-access platform for cancer proteomics. Furthermore, immunohistochemical validation was carried out through the well-established Human Protein Atlas (HPA) database (https://www.proteinatlas.org/) [50], which provides publicly available protein expression data derived from tissue immunostaining.

Statistical analysis

Computational analyses were implemented in the R statistical computing environment (v4.3.0). Continuous measures were expressed as arithmetic means with standard deviation (SD) descriptors. Nonparametric comparative analyses were structured as follows: intergroup comparisons utilized the Wilcoxon rank-sum test, whereas multigroup assessments employed the Kruskal-Wallis rank variance framework. Categorical data distributions were evaluated through Pearson’s chi-square or Fisher’s exact probability tests, as appropriate for contingency table dimensions. Bivariate associations between molecular variables were quantified using Spearman’s ρ correlation coefficients, with statistical significance defined at α = 0.05 unless otherwise stipulated.

Results

The schematic framework for this study is demonstrated in Fig. 1.

Fig. 1.

Schematic workflow diagram. A Transcriptome-based prognostic analysis in Liver Hepatocellular Carcinoma (LIHC); B Single-cell sequencing-associated analysis in LIHC. TCGA The Cancer Genome Atlas, LIHC Liver Hepatocellular Carcinoma, ACMRDEGs Apoptosis and Cholesterol metabolism Related Differentially Expressed Genes, GO Gene Ontology, KEGG Kyoto Encyclopedia of Genes and Genomes, GSEA Gene Set Enrichment Analysis, ROC Receiver Operating Characteristic, KM Kaplan–Meier, DEGs Differentially Expressed Genes, PCA Principal component analysis

ACMRDEGs analysis

We conducted differential expression analysis on the TCGA-LIHC dataset, a total of 4,537 DEGs were identified between LIHC and control groups. Of these, 3,341 genes exhibited upregulated expression (logFC > 1 and p-value < 0.05), while 1,196 genes showed downregulated expression (logFC < -1 and p-value < 0.05) (Fig. 2A).

Fig. 2.

Differential gene expression analysis. A Volcano plot of differentially expressed genes (DEGs) in The Cancer Genome Atlas-Liver Hepatocellular carcinoma (TCGA-LIHC), comparing Liver Hepatocellular carcinoma (LIHC samples) (orange: disease group) and control tissues (blue: normal group); B Venn diagram illustrating overlaps among DEGs, apoptosis-related genes (ARGs), and cholesterol metabolism-related genes (CMRGs) in TCGA-LIHC; C Heatmap displaying expression differences of overlapping genes in TCGA-LIHC (red: high expression; purple: low expression); D Forest plot of univariate Cox regression analysis for prognostic genes in LIHC (TCGA-LIHC dataset). TCGA The Cancer Genome Atlas, LIHC Liver Hepatocellular Carcinoma, ARGs apoptosis-related genes, CMRGs cholesterol metabolism-related genes, DEGs differentially expressed genes

To focus on targeted genes related to LIHC, we intersected the set of all DEGs with ARGs and CMRGs, and visualized the results with a Venn diagram (Fig. 2B). This analysis yielded 28 ACMRDEGs. The expression differences of these 28 intersecting genes between normal and tumor groups were further visualized in a heatmap (Fig. 2C).

Building upon the intersection results, we performed univariate Cox regression analysis using clinical data from TCGA-LIHC samples (Table 2). Genes with a p-value < 0.05 in the univariate analysis were included in subsequent prognostic analyses. As the forest plot shown, EPHX2 (hazard ratio (HR) = 0.855), FABP5 (HR = 1.43), SQLE (HR = 1.19), ANXA2 (HR = 1.29), ADH4 (HR = 0.887), HMGCS2 (HR = 0.87), CYP7A1 (HR = 0.895), and ACADL (HR = 0.794) were considered as independent prognostic risk factors (Fig. 2D).

Establishment and validation of prognostic risk model

To further evaluate the prognostic value of the eight identified ACMRDEGs in LIHC, we included them in a LASSO regression analysis and constructed a corresponding regression model. The results of this analysis were visually represented through the LASSO regression model plot (Fig. 3A) and the LASSO variable trajectory plot (Fig. 3B). The LASSO regression model incorporated six genes—EPHX2, FABP5, SQLE, ADH4, HMGCS2, and CYP7A1—which were subsequently defined as hub genes. To explore the association between the expression levels of these six hub genes and their prognostic significance, we performed a multivariate Cox regression analysis (Fig. 3C).

Fig. 3.

Least absolute shrinkage and selection operator (LASSO) and Cox regression analysis. A, B Prognostic risk model (A) and coefficient trajectories (B) of the LASSO regression model; C Forest plot of the multivariable Cox regression analysis; D Kaplan–Meier (K-M) curve comparing overall survival (OS) between high- and low-risk groups in Liver Hepatocellular Carcinoma (LIHC) patients; E Time-dependent Receiver Operating Characteristic Curve (ROC) for LIHC prognosis in The Cancer Genome Atlas-Liver Hepatocellular Carcinoma (TCGA-LIHC) cohort; F) Differential expression of prognostic genes between high- and low-risk groups; G–I Calibration curves for 1-, 3-, and 5-years survival prediction accuracy; J–L Decision Curve Analysis (DCA) curves for 1-, 3-, and 5-years prognostic prediction accuracy. TCGA The Cancer Genome Atlas, LIHC Liver Hepatocellular Carcinoma, LASSO Least Absolute Shrinkage and Selection Operator, OS Overall Survival, K-M Kaplan–Meier, ROC Receiver Operating Characteristic Curve, AUC Area Under the Curve, DCA Decision Curve Analysis. * P < 0.05, ** P < 0.01, *** P < 0.001. Model performance: AUC > 0.5 indicates predictive capacity (closer to 1 = stronger diagnostic utility; 0.5–0.7 = limited accuracy). Calibration curves align with ideal predictions when near the diagonal

Subsequently, we conducted Kaplan-Meier (K-M) survival curve analysis to assess the prognostic value of the risk score of hub genes in conjunction with the overall survival (OS) data from the TCGA-LIHC cohort. The analysis, stratified by the median risk score, demonstrated a statistically significant difference in OS between the high-risk and low-risk groups (p < 0.001) (Fig. 3D). Furthermore, we plotted the time-dependent ROC curve for patients from TCGA-LIHC (Fig. 3E). The results indicated that the prognostic risk model exhibited satisfactory accuracy for 1-year survival (AUC > 0.7), although the accuracy decreased for 3-year and 5-year survival (0.7 > AUC > 0.5). A comparison of hub genes expression differences between the high-risk and low-risk groups revealed statistically significant findings (p < 0.001) (Fig. 3F). The risk score was calculated as follows (Table 3):

Table 3.

Results of univariate and multivariate Cox regression analyses

Univariate cox regression analysis
Gene	p.value	Mean	Lower	Upper	Coef.
EPHX2	0.0439	0.855	0.735	0.996	− 0.156
FABP5	0.000241	1.43	1.18	1.72	0.355
SQLE	0.0256	1.19	1.02	1.38	0.173
ANXA2	0.00412	1.29	1.09	1.54	0.258
ADH4	6.41e-05	0.887	0.836	0.941	− 0.12
HMGCS2	0.00109	0.87	0.8	0.946	− 0.139
CYP7A1	0.0085	0.895	0.824	0.972	− 0.111
ACADL	0.0143	0.794	0.66	0.955	− 0.231

Multivariate cox regression analysis
Gene	p.value	Mean	Lower	Upper	Coef.
EPHX2	0.163	1.16	0.941	1.44	0.151
FABP5	0.287	1.14	0.896	1.45	0.13
SQLE	0.0147	1.23	1.04	1.46	0.21
ADH4	0.0316	0.922	0.856	0.993	− 0.0814
HMGCS2	0.0529	0.891	0.793	1	− 0.115
CYP7A1	0.226	0.936	0.841	1.04	− 0.066

Additionally, we performed a calibration analysis for the 1-, 3-, and 5-year prognostic outcomes of the prognostic risk model(Fig. 3G-I). The results demonstrated that the model exhibited the best clinical predictive performance for the 1-year outcome (Fig. 3G). To further evaluate the clinical utility of the model, we performed decision curve analysis (DCA) for 1-, 3-, and 5-year survival outcomes (Fig. 3J-L). The results confirmed that the risk score from the multivariate Cox regression model provided the most accurate predictions for 1-year survival (Fig. 3J).

In addition, we calculated the risk scores through the same method in two independent datasets: GSE121248 and GSE76427. Based on the median risk score, LIHC samples in these datasets were classified into high-risk and low-risk groups. To investigate the expression differences of the hub genes in the GSE121248 liver cancer cohort, we performed differential expression analysis, showing significant differences in the expression levels of the six hub genes between the high-risk and low-risk groups (all p < 0.05) (Fig. 4A). We then generated ROC curves to evaluate the classification performance of the hub genes (Fig. 4B-C). The ROC analysis revealed that SQLE and FABP5 exhibited relatively low accuracy for classifying high-risk and low-risk groups (0.5 < AUC < 0.7), whereas the other genes demonstrated moderate accuracy (0.7 < AUC < 0.9).

Fig. 4.

Differential expression validation. A Comparative expression analysis of hub genes between high- and low-risk groups in hepatocellular carcinoma (HCC) samples from the GSE121248 dataset (red: high-risk group; blue: low-risk group); B, C Receiver Operating Characteristic Curve (ROC) and Area Under the Curve (AUC) values for hub genes in the GSE121248 HCC cohort; D Differential expression profiles of hub genes in high- vs. low-risk groups (GSE76427 HCC dataset); E, F ROC curves evaluating diagnostic performance of hub genes in the GSE76427 cohort. HCC hepatocellular carcinoma, ROC Receiver Operating Characteristic Curve, AUC Area Under the Curve, TPR True Positive Rate, FPR False Positive Rate. * P < 0.05, ** P < 0.01, *** P < 0.001. AUC > 0.5 indicates predictive capacity (0.5–0.7: limited accuracy; 0.7–0.9: moderate accuracy; closer to 1 = stronger diagnostic utility)

Similarly, to explore the expression differences of hub genes in the GSE76427 liver cancer cohort, we conducted differential expression analysis, which revealed that, except for CYP7A1, five hub genes showed statistically significant expression differences between the high-risk and low-risk groups (p < 0.05) (Fig. 4D). ROC curves generated for these genes (Fig. 4E, F) indicated that ADH4, HMGCS2, and SQLE exhibited moderate accuracy in classifying the high-risk and low-risk groups (0.7 < AUC < 0.9), while the other genes had lower classification accuracy (0.5 < AUC < 0.7).

Furthermore, three clinical parameters with significant prognostic value were identified from a set of indicators closely associated with HCC: Pathologic T stage, Pathologic M stage, and Child-Pugh grade (Table S4). These were integrated into the previously established hub gene-based prognostic risk model to construct a new combined clinical and genetic predictor. The updated model demonstrated improved predictive performance over the original model, with AUC values of 0.743, 0.745, and 0.793 for 1-, 3-, and 5-year survival, respectively (Fig. S1A), indicating enhanced discrimination ability over time. The model also exhibited good calibration across all three time points (Fig.S1B). Although the IDI analysis indicated a numerical enhancement in predictive performance with the new model, the difference did not reach statistical significance (p > 0.05) (Fig. S1C). Finally, DCA was performed to evaluate the clinical utility of the model at 1, 3, and 5 years. The results suggested favorable net benefit across all time points, with a trend toward increasing clinical value over longer follow-up periods (Fig. S1D-F).

Proteomics analysis

Following validation at the transcriptomic level, we further investigated the protein expression of these six hub genes. Compared with normal tissues, tumor tissues exhibited significantly up-regulated protein expression of FABP5 (p = 2.3E-68) and SQLE (p = 2.3E-12), while the expression of EPHX2 (p = 3.6E-70), ADH4 (p = 3.6E-101), and HMGCS2 (p = 2.2E-38) was markedly down-regulated. In contrast, no significant difference was observed in CYP7A1 expression (p = 1E-01). These findings are fully consistent with the previous results (Figure.S2). Furthermore, IHC results for the remaining five hub genes-excluding CYP7A1, for which data were unavailable-are presented in Figure S3, along with clinical and staining information, allowing direct observation of their expression patterns in hepatocellular carcinoma tissues.

Gene enrichment analysis

To further investigate the relationship between the six hub genes and LIHC in biological mechanism, we performed GO and KEGG enrichment analyses, focusing on BP, CC, MF, and relevant KEGG pathways. The detailed results are presented in Table 4. The bubble plot revealed that the six hub genes were predominantly enriched in BP such as alcohol metabolic process, cholesterol metabolic process, secondary alcohol metabolic process, sterol metabolic process, and sterol biosynthetic process. Additionally, they were significantly associated with MF including retinoid binding, isoprenoid binding, toxic substance binding, monooxygenase activity, and hydrolase activity, acting on ether bonds. Furthermore, these genes were enriched in several KEGG pathways, including PPAR signaling pathway, primary bile acid biosynthesis, steroid biosynthesis, terpenoid backbone biosynthesis, and butanoate metabolism (Fig. 5A). Additionally, we visualized the enrichment results by constructing network diagrams for BP, MF, and KEGG pathways. These diagrams illustrate the connections between the molecules and their corresponding annotations, where node size is proportional to the number of molecules involved in each term (Fig. 5B and D).

Table 4.

GO/KEGG enrichment analysis results

Ontology	ID	Description	GeneRatio	BgRatio	p.adjust	q.value
BP	GO:0006066	Alcohol metabolic process	5/6	361/18,800	1.5e-08	3.69e-06
BP	GO:0008203	Cholesterol metabolic process	4/6	139/18,800	4.24e-08	3.94e-06
BP	GO:1,902,652	Secondary alcohol metabolic process	4/6	149/18,800	5.61e-08	3.94e-06
BP	GO:0016125	Sterol metabolic process	4/6	154/18,800	6.41e-08	3.94e-06
BP	GO:0016126	Sterol biosynthetic process	3/6	65/18,800	7.83e-07	3.85e-05
MF	GO:0005501	Retinoid binding	2/6	38/18,410	6.19e-05	0.0011
MF	GO:0019840	Isoprenoid binding	2/6	38/18,410	6.19e-05	0.0011
MF	GO:0015643	Toxic substance binding	1/6	10/18,410	0.0033	0.0175
MF	GO:0004497	Monooxygenase activity	2/6	103/18,410	0.0005	0.0049
MF	GO:0016801	Hydrolase activity, acting on ether bonds	1/6	10/18,410	0.0033	0.0175
KEGG	hsa03320	PPAR signaling pathway	3/6	75/8164	1.46e-05	0.0003
KEGG	hsa00120	Primary bile acid biosynthesis	1/6	17/8164	0.0124	0.0725
KEGG	hsa00100	Steroid biosynthesis	1/6	20/8164	0.0146	0.0725
KEGG	hsa00900	Terpenoid backbone biosynthesis	1/6	23/8164	0.0168	0.0725
KEGG	hsa00650	Butanoate metabolism	1/6	27/8164	0.0197	0.0725

GO, Gene Ontology; BP, Biological Process; MF, Molecular Function; KEGG, Kyoto Encyclopedia of Genes and Genomes

Fig. 5.

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis. A Bubble plot displaying GO and KEGG enrichment results for six hub genes, categorized as Biological Process (BP), Molecular Function (MF), and KEGG pathways. X-axis: enriched GO/KEGG terms; B–D Network diagrams summarizing GO/KEGG enrichment relationships for hub genes: BP (B), MF (C), and KEGG pathways (D). Nodes represent enriched terms (red) and associated molecules (blue); edges indicate term-molecule relationships. GO Gene Ontology, BP Biological Process, MF Molecular Function, KEGG Kyoto Encyclopedia of Genes and Genomes

From another perspective, we employed GSEA to explore the associations between gene expression and BP, CC, and MF in LIHC patients from the TCGA-LIHC dataset (Fig. 6A). First, we stratified the TCGA-LIHC cohort into high- and low-risk groups based on the risk score. Differential expression analysis was performed using the DESeq2 R package, identifying genes with |logFC| >1 and p.adj < 0.05 as DEGs. The analysis revealed a total of 1907 DEGs, with 1377 genes showing upregulation (logFC > 1, p < 0.05) and 530 genes exhibiting downregulation (logFC < − 1, p < 0.05). Notably, the DEGs were significantly enriched in multiple biological functions and signaling pathways, including FATTY ACID METABOLISM (NES = − 2.768, p.adj < 0.001, FDR < 0.001)(Fig. 6B), DRUG METABOLISM CYTOCHROME P450 (NES = − 2.628, p.adj < 0.001, FDR < 0.001)(Fig. 6C), ACTIVATION OF ATR IN RESPONSE TO REPLICATION STRESS (NES = 1.798, p.adj < 0.01, FDR < 0.01)(Fig. 6D), SYNTHESIS OF IP2 IP AND INS IN THE CYTOSOL (NES = 1.791, p.adj < 0.01, FDR < 0.01)(Fig. 6E), and KINESINS (NES = 1.915, p.adj < 0.001, FDR < 0.001)(Fig. 6F). Detailed results were summarized in Table 5.

Fig. 6.

Differential gene expression analysis and gene set enrichment analysis (GSEA) for liver hepatocellular carcinoma (LIHC). A Ridge plot displaying five enriched biological functions from GSEA in hepatocellular carcinoma (HCC) samples; B–F GSEA results showing significant enrichment of all genes in HCC patients for the following pathways: KEGG_FATTY_ACID_METABOLISM (B); KEGG_DRUG_METABOLISM_CYTOCHROME_P450 (C); REACTOME_ACTIVATION_OF_ATR_IN_RESPONSE_TO_REPLICATION_STRESS (D); REACTOME_SYNTHESIS_OF_IP2_IP_AND_INS_IN_THE_CYTOSOL (E); REACTOME_KINESINS (F). TCGA The Cancer Genome Atlas, GSEA Gene Set Enrichment Analysis, HCC hepatocellular carcinoma, LIHC Liver Hepatocellular Carcinoma

Table 5.

GSEA enrichment analysis results

Description	setSize	NES	p.adjust	qvalue
KEGG_FATTY_ACID_METABOLISM	42	-2.76817	1.93E-08	1.63E-08
KEGG_DRUG_METABOLISM_CYTOCHROME_P450	70	-2.62842	1.93E-08	1.63E-08
REACTOME_ACTIVATION_OF_ATR_IN_RESPONSE_TO_REPLICATION_STRESS	37	1.79785	0.00818	0.00647
REACTOME_SYNTHESIS_OF_IP2_IP_AND_INS_IN_THE_CYTOSOL	14	1.79071	0.00767	0.00690
REACTOME_KINESINS	61	1.91536	0.00037	0.00031

GSEA, Gene Set Enrichment Analysis

Interaction network analysis

We constructed a PPI network for six prognostic hub genes using the STRING database. The results revealed a relatively tight interaction among these genes (Fig. 7A). Subsequently, we performed a subgroup comparison to further validate the expression differences of the six hub genes between the tumor and normal groups in the TCGA-LIHC dataset, which indicated statistically significant differences, with a p-value of less than 0.001 (Fig. 7B).

Fig. 7.

Correlation analysis of hub genes. A Protein–protein interaction (PPI) network of hub genes; B Comparative expression analysis of hub genes between tumor (orange: Liver Hepatocellular Carcinoma (LIHC)) and normal tissues (blue: control) in The Cancer Genome Atlas- Liver Hepatocellular Carcinoma (TCGA-LIHC) cohort; C Regulatory network of mRNA-miRNA interactions (red nodes: mRNAs; blue nodes: miRNAs). PPI Protein-protein interaction, LIHC Liver Hepatocellular Carcinoma, TCGA The Cancer Genome Atlas. * P < 0.05, ** P < 0.01, *** P < 0.001

The mRNA-miRNA interaction network was visualized using Cytoscape software (Fig. 7C). The resulting network comprises six hub genes—EPHX2, FABP5, SQLE, ADH4, HMGCS2, and CYP7A1—and 56 miRNA molecules, illustrating the complex mRNA-miRNA interactions.

Immune infiltration analysis

Immune cell landscape quantification was conducted via CIBERSORT deconvolution to assess the relative proportions of 22 lymphoid and myeloid subsets within the TCGA-LIHC cohort. Comparative immunoprofile visualization between neoplastic and adjacent non-tumorous specimens was achieved through a propotion bar plot delineating compartment-specific abundance distributions (Fig. 8A). Intergroup comparative analysis revealed marked compositional disparities, with 12 immunocyte subtypes—including Plasma cells, CD4 + memory-activated T cells, regulatory T cells (Tregs), gamma T cells, resting NK cells, Monocytes, M0/M2-polarized macrophages, resting dendritic cells, mast cells in activated/resting states, and Neutrophils—exhibiting statistically significant infiltration differences (all p < 0.05) (Fig. 8B).

Fig. 8.

Immune infiltration analysis by CIBERSORT Algorithm 1. A Bar plot showing immune cell composition in tumor (orange) vs. control (blue) tissues (The Cancer Genome Atlas- Liver Hepatocellular Carcinoma (TCGA-LIHC)); B Box plots comparing immune cell infiltration levels between tumor and control groups. TCGA The Cancer Genome Atlas; LIHC Liver Hepatocellular Carcinoma. * P < 0.05, ** P < 0.01, *** P < 0.001

Subsequently, we performed a correlation heatmap analysis to investigate the relationships between the abundance of the 12 immune cell types and LIHC samples (Fig. 9A). The results revealed a significant negative correlation between Mast cells resting and Mast cells activated (r-value = -0.41) in LIHC samples. This finding is consistent with the observed upregulation of Mast cells resting and downregulation of Mast cells activated in LIHC samples (Fig. 8B).

Fig. 9.

Immune infiltration analysis by CIBERSORT Algorithm 2. A Heatmap of correlation coefficients between immune cell subtypes in The Cancer Genome Atlas- Liver Hepatocellular Carcinoma (TCGA-LIHC); B Bubble plot displaying associations between hub genes and immune infiltration (red: positive; purple: negative; color intensity = |r|). TCGA The Cancer Genome Atlas, LIHC Liver Hepatocellular Carcinoma. * P < 0.05, ** P < 0.01, *** P < 0.001. Correlation strength: |r| < 0.3 = weak/none; 0.3–0.5 = moderate

We also employed a correlation bubble plot to illustrate the associations between the six hub genes and immunocyte infiltration levels (Fig. 9B). The bubble plot results showed a positive correlation between M0 macrophages and the FABP5 (r-value = 0.385, p < 0.001), while a negative association with the HMGCS2 (r-value = -0.362, p < 0.001) in LIHC samples.

Finally, the immune infiltration analysis was validated using the ssGSEA deconvolution algorithm. Both methods demonstrated strong concordance in the enrichment trends of the six hub genes and major immune cell subpopulations (Fig. S4A-B). Furthermore, we evaluated the correlation between FABP5, HMGCS2, and macrophage infiltration through multiple algorithms in the TIMER 3.0 database. Consistent results were obtained using CIBERSORT-ABS and XCell, whereas QUANTISEQ and TIDE did not yield statistically significant outcomes (Fig. 4C-D).

Single-cell sequencing analysis

To investigate the intricate tumor microenvironment of LIHC at the single-cell level, we conducted dimensionality reduction and clustering annotations using the single-cell sequencing dataset GSE149614. The study included 18 samples, with each sample exhibiting relatively uniform cell distribution, suggesting the absence of significant batch effects, which provides a solid foundation for subsequent analyses. After performing initial quality control and doublet removal (Fig. 10A, B), we successfully identified 58,380 cells from the single-cell data.

Fig. 10.

Annotation and visualization of the cellular microenvironment in liver hepatocellular carcinoma (LIHC). A, B Violin plots illustrating cell quality control metrics; C Uniform Manifold Approximation and Projection (UMAP) visualization of 22 cell clusters identified through unsupervised clustering; D UMAP projection annotated with cell types in LIHC; E, F Expression patterns of cell type-specific biomarkers (violin plot: (E); bubble plot: (F)). TCGA The Cancer Genome Atlas, LIHC Liver Hepatocellular Carcinoma, UMAP Uniform Manifold Approximation and Projection

Next, we applied the Uniform Manifold Approximation and Projection (UMAP) method to partition all cells into 22 distinct clusters (Fig. 10C). Cellular identities were resolved through integrative analysis of cluster-specific transcriptional signatures and lineage-defining molecular markers. In tumor tissues, we identified six major cell types, namely hepatocytes, T NK cells, myeloid cells, B cells, endothelial cells, and fibroblasts (Fig. 10D). The differential expression of specific biomarkers across these cell types was visualized using violin and bubble plots (Fig. 10E, F), with APCO2 showing the most pronounced differential expression in hepatocytes.

Subsequently, we constructed a heatmap to depict the differential expression of the top 10 most significantly DEGs between various cell types (Fig. 11A). We presented bar plots comparing the proportions of cell types between different sample groups, including single-cell LIHC and control groups (Fig. 11B, C). Notably, hepatocyte depletion emerged as a hallmark of LIHC lesions, potentially contributing to oncogenic niche formation. Finally, the expression distribution of the six identified hub genes across distinct cell types was visualized using violin plots (Fig.S5). The results demonstrate that HMGCS2 was most highly expressed in hepatocytes, while FABP5 showed marked expression across fibroblasts, endothelial cells, and myeloid cells.

Fig. 11.

Analysis of cellular components in liver hepatocellular carcinoma (LIHC). A Heatmap of the top 10 differentially expressed genes (DEGs) across cell types (color intensity represents correlation strength; red: positive; purple: negative); B Bar plot showing the proportion of cell types in the study cohort; C Comparative cell type abundance between LIHC (orange) and control (blue) groups. DEGs Differentially Expressed Genes, LIHC Liver Hepatocellular Carcinoma

Discussion

LIHC represents a significant global health threat because of its high incidence and mortality rates [2, 3]. LIHC, the most common type of primary liver cancer, is closely linked to chronic liver conditions such as HBV/HCV [51] and alcoholic/NAFLD [4], which promote tumorigenesis through complex oncogenic processes. The asymptomatic nature of early-stage HCC and the lack of effective diagnostic tools result in over 70% of cases being diagnosed at advanced stages, contributing to a five-year survival rate below 20% [5, 9]. This immense public health burden underscores the urgent need to develop novel biomarkers and targeted therapies to improve clinical outcomes and reduce mortality.

Recent research has indicated that cholesterol can modulate apoptosis through multiple mechanisms. In autoimmune myocarditis, cholesterol accumulation is linked to inducing both intrinsic and extrinsic apoptotic pathways [52]. In multiple myeloma cells, elevated levels of LRP8, a regulator of cholesterol metabolism, promote enhanced cholesterol uptake and utilization, thereby influencing mTORC1 signaling and apoptotic processes [53]. Within the context of LIHC, dysregulated cholesterol metabolism may contribute to oxidative stress, inflammatory responses, and activation of apoptotic signaling, which play pivotal roles in HCC progression and prognosis [27, 54]. However, the interactive mechanisms between apoptosis and cholesterol metabolism remain incompletely understood. By integrating multi-omics data with machine learning, our study systematically identifies key dual-phenotype genes linked to LIHC prognosis, offering novel insights into the underlying molecular mechanisms. The established model likely captures the integrated effects of these metabolic disturbances, providing a comprehensive perspective on the cholesterol metabolism–apoptosis axis in HCC, with the potential to improve prognostic accuracy and inform therapeutic strategies.

In this study, we identified six core genes-EPHX2, FABP5, SQLE, ADH4, HMGCS2, and CYP7A1-with significant prognostic value from 28 ACMRDEGs through univariate, multivariate, and LASSO regression analyses. We further interpreted their potential mechanisms from the perspective of a “cholesterol metabolism-apoptosis” interaction axis. Previous studies have indicated that EPHX2 is downregulated in multiple cancers and may regulate tumor progression by modulating epoxide metabolism and oxidative stress, thereby indirectly influencing apoptotic pathways such as PI3K/AKT and engaging in crosstalk with cholesterol metabolism [55–58]. FABP5, functioning as a lipid chaperone, participates in fatty acid transport and PPAR signaling, shaping a lipid-metabolic inflammatory microenvironment. Its significant correlation with M0 macrophage infiltration suggests an indirect role in suppressing tumor cell apoptosis by attenuating immunogenic cell death signals and facilitating immune evasion [59–63]. In contrast, SQLE exerts more direct control over apoptotic signaling. As a rate-limiting enzyme in cholesterol synthesis, it governs membrane cholesterol composition and lipid raft integrity. Cholesterol accumulation alters the membrane localization of death receptors (e.g., Fas, TNFR) and enhances survival pathways such as PI3K/Akt, thereby directly interfering with apoptotic execution and promoting tumor cell survival [64–66]. Notably, the pathways mediated by FABP5 and SQLE are functionally complementary, collectively forming a coordinated “lipid transport–cholesterol synthesis–apoptosis evasion” model that provides a novel theoretical framework for understanding how the cholesterol metabolism–apoptosis axis influences tumor progression. HMGCS2, another key cholesterol synthase, may affect tumor cell survival and targeted therapy response by modulating ketone body metabolism and energy supply [67–70]. ADH4, involved in ethanol metabolism, potentially promotes apoptotic signaling through alcohol-derived metabolites that induce oxidative stress and disrupt lipid homeostasis [71, 72]. Meanwhile, CYP7A1, the rate-limiting enzyme in the classical bile acid synthesis pathway, appears to suppress hepatocyte apoptosis when downregulated by reducing bile acid production. In LIHC, this may occur through pathways such as FXR/NF-κB, influencing disease progression and prognosis [73–75].

Analysis of gene expression across different clinical and pathological characteristics indicates that tumor microenvironments dynamically alter gene functions. This may provide new insights for the prognostic management of LIHC patients. Notably, for the first time, our study defines the dual mechanism of cholesterol metabolism and apoptosis in the six hub genes within LIHC, using them as a comprehensive marker to stratify LIHC patients into different risk groups. Assessing the correlation between these genes and patient survival helps address the limitations of current serum markers (such as AFP) in early LIHC detection, thus promoting the advancements in preclinical detection paradigms and precision therapeutic interventions for LIHC.

Additionally, we innovatively constructed a cholesterol-apoptosis-related risk score using six hub genes via LASSO regression, with prognostic validity confirmed by multivariate Cox analysis. This model improves upon the static nature of traditional TNM staging by enabling dynamic risk stratification, thus facilitating personalized treatment strategies. It demonstrated consistent predictive performance (AUC >0.7) and robust discriminative power across diverse populations-including European, American, and Asian cohorts-accounting for variations in race, gender, and age, as validated in multiple independent datasets (TCGA-LIHC, GSE121248, GSE76427) (Figs. 3 and 4). Recognizing the established value of integrating gene expression with clinical indicators for improved predictive accuracy-as demonstrated in studies such as those by Lu et al., Shi et al., and our colleagues [10, 32, 76]-we incorporated pathologic T stage, M stage, and Child-Pugh grade into the genetic prognostic model. This combined model showed improved predictive performance, particularly in long-term survival estimates (Fig.S1). Nevertheless, due to the relatively limited clinical annotations in public databases-particularly the lack of systematic etiological information such as HBV/HCV infection, alcohol-related liver disease, and metabolic dysfunction-associated steatotic liver disease—stratified analyses across different etiological subgroups could not be performed. This limitation may affect the applicability of the model across diverse etiological backgrounds. Additionally, while numeric improvement was observed in discrimination and clinical utility, not all metrics reached statistical significance. Future efforts will focus on validating and refining this model using real-world clinical cohorts with more complete prognostic variables, which will help better evaluate its translational applicability and robustness. In summary, these six genes collectively form an integrated “cholesterol–apoptosis regulatory axis” that operates through cholesterol synthesis, bile acid metabolism, fatty acid transport, alcohol metabolism, and oxidative stress. This system coordinately regulates metabolic reprogramming and apoptosis evasion in HCC pathogenesis. The proposed model not only offers a novel set of molecular biomarkers for prognostic stratification but also provides a theoretical foundation for developing new treatment strategies targeting the metabolic–apoptotic crosstalk.

Analysis of immune infiltration in LIHC samples indicated notable variations in immune cell levels, implying that the immune microenvironment may be shaped by diverse mechanisms impacting disease progression and patient prognosis. Application of the CIBERSORT and ssGSEA algorithm revealed a significant correlation between the infiltration levels of immune cells, including regulatory T cells (Tregs), mast cells, and macrophages, and clinical outcomes in LIHC. Tregs have been reported as a substantial determinant in shaping the immunosuppressive landscape in LIHC, contributing to an environment conducive to tumor progression [77]. Our findings (Fig. 8B) align with the observation that an increased proportion of Tregs correlates with advanced clinical stages and poor prognosis. Macrophages, especially the M0 unpolarized state and M2 polarized subtype, are crucial in LIHC development. Their dynamic regulation is closely linked to immune suppression, pro-cancer signaling, and therapy resistance in the tumor microenvironment (TME) [78]. M0 macrophages are typically more plastic, while M2 macrophages tend to promote tumor growth and immune suppression. In our study, we found that M0 macrophages were highly infiltrated, whereas M2 macrophages were less infiltrated in LIHC tissues. We hypothesize that this may reflect the early stage of tumor development, where M0 macrophages have not yet been extensively polarized to the M2 phenotype. Alternatively, differential expression of ACMRDEGs may regulate pathways such as oxidative stress and cholesterol metabolism, leading to a reversal of the M2 phenotype [79]. Furthermore, our analysis indicated a positive correlation between M0 macrophages and the FABP5 gene (r-value = 0.385, p < 0.001) and a negative correlation with the HMGCS2 gene (r-value = − 0.362, p < 0.001). Based on these findings, we hypothesize that these genes might potentially modulate the polarization of M0 macrophages through apoptosis-cholesterol metabolism pathways. However, this mechanism remains speculative, and further experimental validation such as immunohistochemistry, flow cytometry, or other immunological assays are needed to confirm such functional relationships.

Recent studies have suggested that mast cells may contribute to the progression and prognosis of liver diseases [80]. The complex relationship between mast cells, immune responses, and tumor behavior highlights the importance of understanding their role in liver pathophysiology. Retrospective studies have indicated that the presence of mast cells might correlate with better prognosis in HCC patients, with high-infiltration CD117 + mast cells in the TME offering protective effects against tumor progression [81]. Moreover, inter actions between mast cells-derived mediators and other immunocytes significantly affect patient prognosis [82]. However, previous research has not yet elucidated which mast cell subtypes are involved. This study is the first to demonstrate that resting mast cell infiltration is increased, while activated mast cells are decreased in LIHC tissue, showing a significant negative correlation (r-value = − 0.41). This suggests that mast cell activation may have a suppressive effect on LIHC, although the potential antagonistic regulatory mechanisms between mast cell subtypes require further investigation. These findings provide new targets for future immune-based therapeutic strategies and offer fresh perspectives on the complexity of the immune microenvironment.

We employed multiple gene enrichment analysis approaches and confirmed that the six hub genes are involved in BP such as alcohol and cholesterol metabolism, and are associated with well-established pathways including PPAR signaling axis. In normal tissues, cholesterol metabolism is essential for cellular membrane integrity [83], while alcohol metabolism facilitates detoxification and prevents hepatic injury [84]. Tumor cells reprogram metabolic pathways (e.g., the Warburg effect) to adapt to stressful microenvironmental conditions such as hypoxia [85], inwhich PPAR signaling may contribute to tumor progression by modulating lipid metabolism and inflammatory responses [86]. Through multi-omics integration—including PPI and miRNA regulatory networks—this study reaffirms the relevance of these pathways in LIHC and establishes a prognostic model with clinical potential. Recent developments in small-molecule inhibitors targeting metabolic pathways (e.g., PPARγ agonists and statins) have shown promise in cancer therapeutics [87, 88]. Our work further elucidates the upstream regulatory role of miRNAs, providing a more comprehensive understanding of regulatory networks in LIHC. Future researches should focus on key nodes within these pathways to develop more specific and effective targeted therapies, ultimately improving treatment outcomes for liver cancer patients.

This study reinforces the essential role of lipid metabolism in LIHC at the single-cell level. Through dimensionality reduction clustering, we identified six major cell types with significantly different proportions in LIHC samples. Notably, hepatocytes showed a marked decrease in proportion within LIHC, and this reduction was significantly associated with the expression of the APOC2 gene, which regulates triglyceride metabolism (Fig. 10E-F). Previous studies have demonstrated that, compared to cancer cell lines, primary hepatocytes exhibit higher lipoprotein synthesis [89]. A decrease in this synthesis is linked to lipid dysregulation during cancer progression, which is consistent with our results. On the other hand, changes in genes such as APOC2 may drive imbalances in nutrient utilization within the immune microenvironment, potentially impairing the tumoricidal function of immune cells [90]. In addition, our study observed high expression of HMGCS2 in hepatocytes, suggesting a potential close association with APOC2. Meanwhile, FABP5 was predominantly enriched in myeloid cells, which is consistent with our previous finding of a positive correlation between FABP5 and M0 macrophages. These observations imply that different hub genes may influence the tumor microenvironment of hepatocellular carcinoma by modulating distinct effector cells through the cholesterol-apoptosis axis. It should be noted, however, that these hypotheses require further experimental validation.

Understanding these metabolic alterations is essential for improving the efficacy of immunotherapy and enhancing patient prognosis. Integrating insights from metabolic regulation, immune response, and genetic susceptibility is critical for advancing LIHC management and therapeutic strategies, ultimately improving patient care and quality of life.

It is important to acknowledge several limitations of this study. Firstly, although comprehensive transcriptomic and proteomic data were obtained from public databases, the absence of experimental validation constrains the biological interpretation and clinical applicability of the findings. In particular, the interactions among genes and between genes and immune cells, as well as the underlying mechanisms, warrant further in-depth investigation. Secondly, the model’s robustness and generalizability may be limited by the relatively small sample size. Additionally, potential batch effects across different datasets may have introduced biases into the analytical outcomes. Therefore, future studies should incorporate a larger number of clinical samples and experimental validation-including in vitro functional assays, drug screening and clinical correlation analyses-to strengthen the biological interpretation and clinical translatability of our conclusions.

Conclusion

In conclusion, this study successfully identified transcriptomic biomarkers within the apoptotic-cholesterol regulatory axis in LIHC pathogenesis and established a LASSO-Cox-derived prognostic signature validated for stage-specific survival stratification. Despite its potential implications, these findings require further verification through independent cohorts and experimental studies to confirm their clinical relevance and biological mechanisms. We anticipate that continued research will help clarify the translational potential of these results, possibly contributing to improved diagnosis and treatment strategies for LIHC in the future.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conception and writing-original draft preparation, L.D; Charting and methodology, Y.L; data analysis and visualization, L.D; reference acquisition, J.S and A.Q.Z; comments and suggestions, J.H; manuscript revision and funding acquisition, L.D. All the authors approved the final manuscript.

Funding

This study was supported by Natural Science Foundation of Ningbo (2023J221), Medical and Health Research Project of Zhejiang Province (2024KY1480), Traditional Chinese Medicine Research Project of Zhejiang Province (2025ZL524) and Ningbo Public Welfare Science and Technology Project (2024S159).

Data availability

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author/s. All data and original files in our work are freely available under a ‘Creative Commons BY 4.0’ license. All methods were carried out in accordance with relevant guidelines and regulations.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.