Machine learning-based prognostic model of stemness and angiogenesis-related genes for predicting prognosis and immune infiltration in patients with HCC

Abstract

Hepatocellular carcinoma (HCC), the predominant form of primary liver cancer, ranks as the fourth leading cause of cancer-associated mortality globally. The heightened mortality associated with HCC is largely attributed to its propensity for recurrence and metastasis, which cannot be achieved without tumor stemness and angiogenesis. Here, we aimed to develop a novel signature of stemness and angiogenesis-related genes (SARGs) for the prediction of clinical prognosis and tumor microenvironment in HCC, with the overarching objective of uncovering novel therapeutic targets capable of concurrently disrupting these intertwined processes, thereby offering potential breakthroughs for more effective anti-HCC strategies. The differentially expressed SARGs were subjected to univariate Cox regression analysis to identify SARGs with prognostic significance. A nine-SARGs risk score model was constructed using Least Absolute Shrinkage Selection Operator (LASSO) Cox regression with 10-fold cross-validation. Furthermore, a nomogram incorporating the SARGs score and other clinicopathological features was developed for accurate prediction of survival rate in patients with HCC. Knockdown of ELOVL3 expression was performed, and its effects on tumor stemness and angiogenic potential were verified through in vitro and in vivo experiments. Patients with HCC were categorized into high- and low-risk groups based on the median risk score values, with higher risk scores indicating worse overall survival (log-rank P < 0.001). The nine-SARGs risk score, comprising DRD1, CDX2, ELOVL3, TKTL1, IGLON5, SHISA9, WNT1, CNTN6, and MMP3, demonstrated robust predictive performance (C-index: 0.72) for clinical prognosis, tumor microenvironment characteristics, and immunotherapy response in HCC. ELOVL3 knockdown reduced tumor stemness and angiogenic potential, leading to the inhibition of tumor growth and metastasis. This study established a direct molecular correlation between tumor stemness and angiogenesis, encompassing clinical features, tumor microenvironment, and immune response, thereby offering valuable insights for predicting clinical outcomes and immunotherapy responses in HCC. Our findings demonstrate that ELOVL3 correlates with cancer cell stemness and angiogenic potential, thereby identifying it as a promising therapeutic target. Further investigations are warranted to elucidate the downstream molecular pathways that mediate these functional effects.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-38379-w.

Introduction

Hepatocellular carcinoma (HCC) is the most common primary liver cancer and the fourth leading cause of cancer-related deaths worldwide¹.

Cancer stem cells (CSCs), a small subpopulation within tumors, possess unique characteristics similar to normal stem cells, including self-renewal, the ability to differentiate into various cell types within the tumor mass, and high tumor-initiating potential². These properties endow CSCs with the capacity to drive tumor growth, maintain tumor heterogeneity, and contribute to tumor recurrence after treatment³. For instance, in breast cancer, CSCs have been shown to be more resistant to chemotherapy and radiation, leading to treatment failure and disease relapse⁴. Understanding the molecular mechanisms underlying tumor stemness is crucial for developing more effective anti-cancer therapies that can specifically target this resilient subpopulation.​.

Angiogenesis, the formation of new blood vessels from pre-existing ones, is essential for tumors to grow beyond a few millimeters in size⁵. Tumors require a continuous supply of nutrients and oxygen, which is provided by the newly formed blood vessels⁶. In addition to facilitating tumor growth, angiogenesis also plays a critical role in tumor metastasis, as cancer cells can enter the bloodstream through the newly formed vessels and spread to distant organs^7,8. Anti-angiogenic therapies, such as bevacizumab, have been developed to target this process⁹. However, the development of resistance to these therapies has limited their long-term efficacy^10,11.

Remarkably, increasing evidence indicates a strong interplay between tumor stemness and angiogenesis^12,13. CSCs can secrete various pro-angiogenic factors, such as vascular endothelial growth factor (VEGF), to promote angiogenesis, creating a microenvironment conducive to their survival and expansion¹⁴. Conversely, the angiogenic niche can also regulate the self-renewal and stemness properties of CSCs, forming a positive feedback loop¹⁵. The surface markers of CSCs and neovessels have been extensively studied in the past decades¹⁶. Currently, a number of studies have focused on the common biomarkers shared by CSCs and tumor vascular systems, and these studies represent highly attractive and promising research endeavors in the field of cancer treatment. Kidney cancer, highly vascularized with immature vessels/loose endothelium, has abundant CSCs and leaky vessels worsening metastasis/prognosis. CD105 overexpresses on its neovascular ECs (endothelial cells) and CSCs¹⁷. Wang et al.’s CD105-targeted TDS inhibits angiogenesis (via nanofibers) and CSCs (via miR-19b/PTEN), reducing lung metastases in PDX models¹⁸. Moreover, in the treatment of glioma, targeting common biomarkers can not only directly achieve tumor vessels and CSCs dual inhibition but also improve the CSCs- specific drug delivery¹⁹.

Recent advances in machine learning have facilitated the development of robust gene signatures for cancer prognosis, with the integration of multi-omics data effectively enhancing predictive accuracy^20,21. Consistent with recent machine learning-based biological modeling studies that underscore the significance of multi-cohort validation and functional validation of key genes, our research constructed a 9-gene prognostic model for hepatocellular carcinoma (HCC) based on stemness and angiogenesis-related genes (SARGs). Among these genes, ELOVL3 was further subjected to in-depth analysis to explore its novel regulatory role in HCC cell proliferation, migration and invasion in vitro, as well as in tumor growth in vivo.

Collectively, the findings of this study are expected to offer more precise and individualized therapeutic options for HCC patients.

Materials and methods

Data resources

The training set patient data of liver hepatocellular carcinoma (LIHC) (including gene expression profiles and clinical data information: age, gender, tumor grade, TNM stage) was downloaded from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/), containing 50 normal samples and 374 cancer samples²². The validation set data of HCC, ICGC-LIRI-JP, was downloaded from the International Cancer Genome Consortium (ICGC) database (https://dcc.icgc.org/), including 232 cancer samples. Stemness-related and angiogenesis-related genes were obtained from GeneCards (https://www.genecards.org/)²³, a comprehensive human gene database that integrates transcriptomic, proteomic, genetic, and functional information. The GeneCards Relevance score is a composite metric that quantifies the strength of association between a gene and a specific keyword (e.g., “stemness” or “angiogenesis”) based on multiple lines of evidence, including direct annotations, literature co-occurrences, pathway memberships, and functional similarity. The score ranges from 0 to > 100, with higher values indicating stronger associations.

Stemness and angiogenesis-related genes (SARGs)

We retrieved 5,955 cell stemness-related genes and 2,711 angiogenesis-related genes using a Relevance score threshold of > 10. This threshold was selected based on the following rationale:

Stringency balance: Relevance score > 10 captures genes with moderate-to-strong evidence while excluding weakly associated genes (score < 10), reducing noise from spurious associations. Reproducibility: This threshold has been widely used in previous cancer genomics studies^23–25 and aligns with GeneCards recommendations for filtering functionally relevant genes. Biological coverage: Preliminary analysis showed that lowering the threshold (e.g., score > 5) increased gene count by 2-fold but added genes with limited functional annotations, whereas raising it (e.g., score > 20) excluded well-characterized stemness/angiogenesis regulators. The intersection of these two gene sets yielded 2,654 overlapping genes, which were designated as stemness and angiogenesis-related genes (SARGs) (Supplementary Table S1). Potential biases associated with this threshold include the exclusion of genes with context-specific roles in HCC stemness/angiogenesis that may have lower GeneCards relevance scores due to limited annotation. However, sensitivity analysis (Relevance score > 5, >15, > 20) confirmed that the core prognostic signature and key biological conclusions remained consistent across thresholds (Supplementary Figure S6 and Supplementary Table S4), mitigating concerns about threshold-dependent bias.

Clustering to identify cancer subtypes

Through univariate Cox regression analysis, we identified 1,147 genes significantly associated with prognosis (P < 0.05) from the 2,654 SARGs. Based on the expression profiles of these 1,147 prognostic SARGs, consensus clustering analysis was performed using the ConsensusClusterPlus R package²⁴, with 1,000 resampling iterations (pItem = 0.8, pFeature = 1.0) to ensure classification stability. TCGA samples (excluding those with survival information < 30 days and duplicate samples) were divided into 2 clusters: Cluster 1 (n = 185 and Cluster 2 (n = 159). The two clusters showed significant survival differences (log-rank test, P < 0.01).

Differential analysis of the two subtypes and the protein-protein interaction (PPI) network of DEGs

Differential expression analysis was conducted on Cluster 1 and Cluster 2 samples using the “edgeR” package (FDR < 0.05 and |log₂(FC)| >1)²⁴. Differentially expressed genes (DEGs) were selected and a PPI network was constructed using the STRING database (https://cn.string-db.org/) with a confidence score > 0.8²⁵. The resulting PPI network comprised 868 nodes (proteins) and 1,775 edges (interactions), with an average node degree of 4.09, network density of 0.0047, average local clustering coefficient of 0.347, and 81 connected components (largest component: 618 nodes). The PPI enrichment P-value (< 1.0e-16) indicated that observed interactions were significantly more frequent than expected by chance, suggesting non-random functional associations among DEGs. However, the low average degree and sparse network density indicate a loosely connected network architecture rather than a tightly integrated global functional module. To identify biologically meaningful functional modules within this sparse network, we applied the Molecular Complex Detection (MCODE) algorithm, which identifies densely connected subnetworks (clusters) representing potential protein complexes or functional modules. In summary, while the overall PPI network is sparse and does not suggest global coordinated pathway activity, specific densely connected MCODE submodules exist and likely drive distinct biological functions relevant to HCC molecular subtypes. This refined interpretation more accurately reflects the heterogeneous nature of differential gene expression between Cluster 1 and Cluster 2²⁶.

Construction and validation of the prognostic model

Genes from the PPI network were subjected to univariate Cox regression analysis (P < 0.05) to identify survival-associated genes. Subsequently, Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression with 10-fold cross-validation was applied to prevent overfitting and select the most informative predictors^27,28. The dataset was randomly partitioned into 10 equal-sized folds; in each iteration, 9 folds trained the model while 1 fold validated. This process repeated 10 times to ensure all samples were validated. The optimal regularization parameter (λ) was determined by maximizing the concordance index (C-index) across validation folds, yielding λ = 0.000422. Multivariate Cox proportional hazards regression was then performed on the LASSO-selected genes to construct the final prognostic model, ultimately identifying 9 genes with non-zero coefficients (Supplementary Table S2). The complete risk score formula is:

Enrichment analysis of DEGs between high- and low-risk groups

Differential expression analysis was performed on high- and low-risk groups using the edgeR package (FDR < 0.05, |log₂(FC)| > 1)²⁹. KEGG and GO enrichment analyses were conducted using the clusterProfiler package, with results visualized by the enrichplot R package³⁰.

Prediction of response to immunotherapy and immune infiltration analysis

Based on the single sample gene set enrichment analysis (ssGSEA) method, immune scores of high- and low-risk groups were obtained respectively³⁰. The Immunophenotype Score (IPS), a predictor of responses to anti-CTLA-4 and anti-PD-1 therapies, was analyzed to understand differences in immune microenvironment between low- and high-risk groups. IPS data for each patient were obtained from the TCIA database (https://tcia.at/home). ssGSEA was used to estimate the enrichment scores of immune cells and immune-related functional pathways³¹.

Tumor mutation analysis

For single nucleotide variant (SNV) mutation data of HCC, the “maftools” package was used to analyze and plot the mutation profiles, mutation types, SNV classes, and mutation rates between high- and low-risk groups³². Additionally, the top 20 genes with the highest mutation rates in both groups were screened to generate waterfall plots of their mutation profiles.

Cell lines and culture

The human HCC cell line HCCLM3 and PLC/PRF/5 was purchased from Procell Biotech (Wuhan, China). The cells were cultured in Dulbecco’s Modified Eagle Medium (DMEM) high-sugar medium (Procell, Wuhan, China) supplemented with 10% fetal bovine serum (Procell, Wuhan, China) at 37 °C in a 5% CO₂ atmosphere. Supplementary Materials are available at Supplementary information.

Data analysis

All statistical analyses were conducted using R software (version 4.5.2). Gene expression levels between tumor and adjacent normal tissues were compared using independent-sample t-tests. Specifically, we examined the differences in immune cell infiltration and activation of immune pathways between these two groups, and the statistical significance of the proportional differences was assessed using chi-squared tests. The coefficients of the prognostic characteristics were calculated using multivariable Cox regression analysis. Kaplan-Meier curves were used to generate survival curves for the high- and low-risk groups, and we used Pearson’s correlation test to analyze the correlation between variables, employing the log-rank test to assess the statistical significance of the differences. Univariate and multivariate Cox regression analyses were conducted to determine the independent prognostic factors for OS, with variables reaching statistical significance used in the multivariate Cox proportional hazards model. All statistical tests were two-sided, and statistical significance was set at P < 0.05 for all tests²².

Results

Identification and analysis of prognostic genes

To construct an HCC prognostic model based on stemness-angiogenesis-linked genes, we first acquired data and defined SARGs, then screened prognostic SARGs and performed subtype clustering. The TCGA-LIHC training cohort was retrieved from TCGA, including gene expression and clinical data. From GeneCards, we extracted 5,955 stemness-related and 2,711 angiogenesis-related genes (Relevance score > 10). Venn analysis identified 2,654 overlapping genes (Fig. 1A), defined as SARGs (Supplementary Table S1). Univariate Cox regression on 2,654 SARGs yielded 1,147 prognostic SARGs (P < 0.05). Adjusting the clustering parameter (k) within a range from 2 to 6, we identified the optimal clustering configuration at k = 2 (Supplementary Figure S1). Using the ConsensusClusterPlus R package, 342 qualified TCGA-LIHC samples were clustered into two subtypes based on these 1,147 genes: Cluster 1 (n = 178) and Cluster 2 (n = 164) (Fig. 1B, C). Clinical relevance of the two molecular subtypes was validated by comparing clinicopathological features between Cluster 1 and Cluster 2 (Supplementary Table S5, Figure S2). Significant differences were observed in: Pathologic Stage (P < 0.001): Cluster 1 enriched for Stage I (57.4% vs. 34.6%), Cluster 2 for Stage II (55.3% vs. 37.7%). T Stage (P < 0.001): Cluster 1 predominantly T1 (59.6% vs. 36.5%), Cluster 2 enriched for T3/T4 (35.2% vs. 17.5%). Vital Status (P < 0.001): Cluster 2 exhibited significantly higher mortality (47.2% vs. 26.2%). No significant differences were observed in age (P = 0.149), gender (P = 0.140), N stage (P = 0.146), or M stage (P = 0.344), indicating that molecular subtypes were driven by tumor biology rather than demographic factors. Kaplan–Meier analysis showed significant OS differences between clusters (log-rank test, P < 0.01), confirming effective prognostic stratification (Fig. 1D).

Fig. 1.

Identification of stemness and angiogenesis-related genes (SARGs) and consensus clustering-based subtype classification in HCC. (A) Venn diagram illustrating the overlap of stemness-related genes (n = 5,955, Relevance score > 10) and angiogenesis-related genes (n = 2,711, Relevance score > 10) from GeneCards, yielding 2,654 SARGs. (B) The cumulative distribution function (CDF) plots portray the variation in k values, spanning from 2 to 6. (C) Consensus clustering outcome at k = 2 showing clear separation of Cluster 1 (n = 183) and Cluster 2 (n = 159). (D) Kaplan-Meier survival analysis highlighting significant overall survival disparity between Cluster 1 and Cluster 2 (log-rank test, P < 0.01).

Construction of a PPI network

To explore the molecular differences underlying the distinct prognostic subtypes of HCC (Cluster 1 and Cluster 2) identified in Fig. 1, we first performed differential expression analysis on the two clusters. Using the edgeR R package, DEGs were screened with strict thresholds: false discovery rate (FDR) < 0.05 and absolute value of log₂ (fold change, log₂ FC) > 1. This analysis yielded a set of DEGs that effectively distinguished Cluster 1 from Cluster 2, laying the foundation for subsequent functional interaction analysis.

To dissect cooperative functional relationships among DEGs, we constructed a PPI network using STRING (confidence score > 0.8), comprising 868 nodes and 1,775 edges. While the overall network was sparse (average node degree = 4.09), MCODE analysis identified 51 densely connected submodules, including chemokine signaling, cytokine/immune pathways, and extracellular matrix organization (Supplementary Figure S3). These submodules likely drive distinct biological functions underlying HCC subtype differences. In summary, while the overall PPI network is sparse, specific densely connected submodules exist and likely drive distinct biological functions relevant to HCC molecular subtypes (Fig. 2).

Fig. 2.

Exploration of Differentially Expressed Genes (DEGs) through Protein–Protein Interaction (PPI) analysis. PPI network constructed using STRING database (confidence score > 0.8) for DEGs between Cluster 1 and Cluster 2. The network comprises 868 nodes and 1,775 edges, with average node degree of 4.09 and network density of 0.0047.

Construction of a prognostic model in the TCGA cohort

Genes obtained from the construction of the Protein-Protein Interaction (PPI) network were selected, and univariate Cox regression analysis (P < 0.05) was applied to identify genes associated with survival. Subsequently, the screened genes with P < 0.05 were used for subsequent LASSO regression analysis with 10-fold cross-validation (Fig. 3A, B, Figure S4) to narrow down the range of survival-related genes and select the optimal regularization parameter λ. Further, multivariate Cox regression confirmed 9 genes as critical prognostic genes, including DRD1, CDX2, ELOVL3, TKTL1, IGLON5, WNT1, CNTN6, MMP3, and SHISA9 (Fig. 3C). Then, we developed a prognostic model, and the risk score was computed using the risk score formula (see Methods section for complete formula with gene coefficients). In accordance with the median value that was used as the dividing line, the patients were classified as belonging to either a high-risk group or a low-risk group. Survival assays indicated that patients with high-risk scores exhibited a shorter OS (Fig. 3D). Using time-dependent ROC curves, the predictive ability of the risk score for OS was examined, and AUC achieved 0.755 at 1 year, 0.716 at 3 years, and 0.677 at 5 years (Fig. 3E). The model demonstrated robust performance with a concordance index (C-index) of 0.72 (95% CI: 0.68–0.76), indicating good discrimination between high- and low-risk patients (Supplementary Table S3). In the high-risk group, the risk-associated genes—including CDX2, ELOVL3, TKTL1, IGLON5, SHISA9, CNTN6, and MMP3—exhibited upregulated expression, whereas the other two genes (DRD1 and WNT1) showed downregulated expression (Fig. 3F). The risk score plot and survival status analysis revealed that the low-risk group exhibited favorable survival outcomes with an extended OS duration (Fig. 3G).

Fig. 3.

Construction and validation of the stemness-angiogenesis-related gene (SARG)-based prognostic risk model in HCC. (A) Partial likelihood deviance versus log (Lambda) for LASSO Cox regression analysis with 10-fold cross-validation (see Figure S4 for detailed cross-validation curve). (B) Coefficient profiles of the prognostic SARGs from LASSO Cox regression analysis. (C) Univariate Cox regression analysis of SARGs in HCC patients from the TCGA cohort, identifying 9 genes with significant prognostic value (P < 0.05). (D) Kaplan–Meier survival curves showcasing overall survival differences between high-risk (n = 169) and low-risk (n = 169) groups stratified by median risk score (log-rank test, P < 0.001). (E) Time-dependent ROC curves with AUC values: 0.755 (1-year), 0.716 (3-year), and 0.677 (5-year), demonstrating robust predictive accuracy. (F) Heatmap displaying characteristic gene expression patterns in TCGA cohorts. Risk genes (CDX2, ELOVL3, TKTL1, IGLON5, SHISA9, CNTN6, MMP3) show higher expression in high-risk group; protective genes (DRD1, WNT1) show higher expression in low-risk group. (G) The distribution and median value of the risk scores in the TCGA set (median = − 0.014, range: −41.18 to 5.71). Model performance: C-index = 0.72 (95% CI: 0.68–0.76).

Validation of the SARGs prognostic model in the ICGC database

The LIHC-JP cohort from the ICGC database was used as the validation group to evaluate the universality of the SARGs prognostic model in the training cohort. The same formula used for TCGA cohort was used to calculate the risk score for each patient in the validation cohorts. The results revealed that patients in the high-risk group had a significantly poorer survival status than those in the low-risk group (Supplementary Figure S5A). Furthermore, ROC analysis demonstrated that the SARGs prognostic model had AUC values of 0.66, 0.7, and 0.8 for 1-, 3-, and 5-year survival, respectively (Supplementary Figure S5B). The distribution plot depicting the risk score, survival status, and expression of the 9 genes indicated a correlation between an increased risk score and higher mortality rates (Supplementary Figure S5C-E).

Independent prognostic analyses of risk scores and construction of a predictive nomogram model

Univariate and multivariate Cox analyses were used to explore the independent prognostic factors for HCC. The results showed that the risk score and stage were significantly associated with OS, suggesting that the risk score of the SARG prognostic model constructed using the TCGA-LIHC cohort was an independent prognostic fac tor for patients with HCC (p < 0.01) (Fig. 4A, B). Therefore, we created a nomogram to assess patient prognosis more accurately based on independent predictive indicators of HCC (Fig. 4C). Notably, further analysis showed that the clinical efficiency of this combination was substantially superior to that of any single component involved (Fig. 4D). As shown in Fig. 4E, the nomogram had a better predictive ability when combined with clinical information, with an AUC of 0.80. Overall, using an SARGs-based prognostic model for patients with HCC integrated with clinical information, a nomogram with superior credibility and accuracy was successfully constructed. the calibration curves (1, 3, and 5 years) demonstrated favorable performance during the internal validation of the nomogram (Fig. 4F).

Fig. 4.

Independent prognostic validation of the SARG-based risk score and construction of a clinical nomogram for HCC. (A) Univariate Cox regression analysis of clinical characteristics and risk scores in the TCGA-LIHC cohort. (B) Multivariate Cox regression analysis confirming risk score and stage as independent prognostic factors for overall survival (both P < 0.01). (C) Nomogram integrating risk score, age, gender, and TNM stage for predicting 1-, 3-, and 5-year overall survival probability in HCC patients. (D) Decision curve analysis (DCA) plot demonstrating that the nomogram provides greater net benefit than individual predictors across a wide range of threshold probabilities, supporting its clinical utility. (E) Time-dependent ROC curves comparing the nomogram (AUC = 0.80) with individual predictors: risk score (AUC = 0.76), stage, age, and gender. The nomogram showed superior predictive performance. (F) Calibration curves for 1-, 3-, and 5-year overall survival predictions, showing good agreement between nomogram-predicted probabilities and observed outcomes (diagonal line represents perfect calibration).

TME characteristics in different SARGs subgroups

CIBERSORT analysis indicated substantial disparities in immune cell infiltration patterns across high-risk and low-risk groups. Our findings revealed a significant increase in the abundance of activated dendritic cells (aDCs), B cells, immature dendritic cells (iDCs), mast cells, macrophages, natural killer (NK) cells, neutrophils, T helper 1 (Th1) cells, and tumor-infiltrating lymphocytes (TIL) in the SARGs low-risk group. Immune function analysis demonstrated that cytolytic activity, T cell co-stimulation, Type I IFN Response, and Type II IFN Response were highly active in the low-risk group (Fig. 5A).

Fig. 5.

Differences in tumor microenvironment (TME) immune infiltration and computational prediction of immunotherapy response potential between high- and low-risk HCC patients. (A) Boxplots displaying a comprehensive analysis of immune cell infiltration profiles between high-risk and low-risk patients based on the ssGSEA algorithm. The low-risk group showed significantly higher infiltration of activated dendritic cells (aDCs), B cells, immature dendritic cells (iDCs), mast cells, macrophages, natural killer (NK) cells, neutrophils, Th1 cells, and tumor-infiltrating lymphocytes (TIL). Immune functional pathways (cytolytic activity, T cell co-stimulation, Type I/II IFN response) were also more active in the low-risk group. (B) Immunophenotype Score (IPS) analysis showing computational prediction of response to anti-PD-1 and anti-CTLA-4/PD-1 therapies. High-risk patients exhibited lower IPS values, suggesting potentially reduced responsiveness to immune checkpoint inhibitors. Note: IPS is a computational predictor and requires validation with actual immunotherapy response data. (C) Tumor Immune Dysfunction and Exclusion (TIDE) scores for high-risk and low-risk patients. P < 0.05, P < 0.01, P < 0.001.

Patients in the high-risk group had significantly lower Immunophenotype Score (IPS) values than the low-risk group, implying reduced responsiveness to anti-PD-1 monotherapy and anti-CTLA-4/anti-PD-1 combination therapy (Fig. 5B). As a gene expression-based computational predictor, IPS was not validated with immunotherapy response data in our cohort; these findings are thus hypothesis-generating and require prospective validation.

The high-risk group also showed higher Tumor Immune Dysfunction and Exclusion (TIDE) scores (Fig. 5C), indicating a predicted elevated immune evasion tendency. TIDE, a bioinformatic tool, does not directly measure immune cell function, tumor-immune interactions, or actual treatment responses, and its predictive accuracy varies by tumor type and regimen. Validation with real-world immunotherapy data was unavailable in our TCGA/ICGC cohorts.

Collectively, while TIDE provides mechanistic insights into immune evasion in high-risk HCC patients, these findings require rigorous validation in immunotherapy-treated prospective cohorts before clinical translation.

Functional enrichment analysis of DEGs and tumor mutation landscape comparison between high- and low-risk HCC groups

The biological functions of the differentially expressed SARGs were further determined through GO annotation and KEGG pathway enrichment analyses. GO annotation showed DEGs were enriched in trans-synaptic signaling, synaptic membrane, and passive transmembrane transporter activity. KEGG pathway analysis revealed significant enrichment in neuroactive ligand–receptor interaction, calcium signaling, and cAMP signaling—pathways known to regulate both tumor stemness and angiogenesis (Fig. 6A-B). Collectively, these findings indicate a substantial correlation between SARGs and neuro-cell communication-associated functions and pathways in HCC progression.

Fig. 6.

Functional enrichment analysis of differentially expressed genes (DEGs) and tumor mutation landscape comparison between high- and low-risk HCC groups. (A) Gene Ontology (GO) enrichment analysis for DEGs between high-risk and low-risk groups. Biological processes: regulation of trans-synaptic signaling, modulation of chemical synaptic transmission, regulation of membrane potential. Cellular components: synaptic membrane, transmembrane transporter complex, monoatomic ion channel complex. Molecular functions: passive transmembrane transporter activity, monoatomic ion channel activity. (B) Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis showing significant enrichment in neuroactive ligand–receptor interaction, calcium signaling pathway, hormone signaling, and cAMP signaling pathway. (C,D) Mutation waterfall plots showcasing distinct mutational landscapes between high-risk (C) and low-risk (D) groups.

Subsequently, we further investigated the distribution pattern of the top20 TCGA-based somatic mutations between the high-and low-risk groups. The differences in genetic mutation profiles were unveiled through the illustration of the mutation waterfall plot. TP53 was identified as the most frequently mutated gene in both subgroups, with the high-risk group exhibiting an 22% higher mutation frequency than the low-risk group (Fig. 6C, D). In addition, CTNNB1 significantly downregulated in the high-risk group and TTN was significantly downregulated in the low-risk group.

Functional validation of ELOVL3 as a regulator of HCC stemness, and angiogenesis

To validate ELOVL3’s biological function in HCC (a key gene in the SARGs signature), we first analyzed its prognostic value and then performed in vitro assays focusing on stemness and angiogenesis. Kaplan–Meier survival analysis of the TCGA-LIHC cohort (patients stratified by median ELOVL3 expression) showed that high ELOVL3 expression correlated with significantly shorter overall survival (Fig. 7A), confirming ELOVL3 as an independent prognostic risk factor for HCC—consistent with its role in the SARGs model. For ELOVL3 depletion studies, two si-ELOVL3 constructs (si-ELOVL3-1/2) or non-targeting si-NC were transfected into HCCLM3 and PLC/PRF/5 cells. qPCR revealed all si-ELOVL3s reduced ELOVL3 mRNA, with si-ELOVL3-2 showing the highest efficiency (Fig. 7B); Western blot further confirmed reduced ELOVL3 protein in si-ELOVL3-transfected cells (Fig. 7C).

Fig. 7.

Functional validation of ELOVL3 as a regulator of HCC stemness and angiogenesis. (A) Kaplan–Meier survival curves showing that high ELOVL3 expression (above median) correlates with significantly shorter overall survival in the TCGA-LIHC cohort (log-rank test, P < 0.01), confirming ELOVL3 as a prognostic risk factor. (B) qPCR analysis of ELOVL3 mRNA levels after transfection with si-NC (negative control), si-ELOVL3-1, or si-ELOVL3-2 in HCCLM3 and PLC/PRF/5 cells. si-ELOVL3-2 showed highest knockdown efficiency (n = 3 biological replicates). (C) Western blot confirming successful ELOVL3 protein knockdown in both cell lines. (D,E) Sphere formation assay demonstrating that ELOVL3 knockdown significantly reduced sphere-forming capacity. Representative images (D) and quantification (E) of spheroids after 7–10 days of serum-free culture (n = 3 independent experiments). (F) Western blot analysis showing that ELOVL3 knockdown significantly downregulated stemness markers CD133, SOX2, Nanog, CD90, and OCT4. (G,H) qPCR analysis revealing that si-ELOVL3 decreased mRNA levels of pro-angiogenic factors VEGFA (G) and FGF2 (H) in both cell lines (n = 3 biological replicates). (I) Western blot confirming reduced VEGFA and FGF2 protein levels upon ELOVL3 knockdown. (J,K) ELISA demonstrating that ELOVL3 knockdown inhibited extracellular secretion of VEGFA (J) and FGF2 (K) into cell culture supernatants (n = 3 independent experiments). Data are represented as mean ± SD. P* < 0.05, P** < 0.01, P*** < 0.001 (two-tailed Student’s t-test).

Tumor stemness was assessed via sphere formation and stemness marker detection. After 7–10 days of serum-free culture, si-ELOVL3 cells formed fewer/smaller spheroids (Fig. 7D, E). Western blot also showed si-ELOVL3 significantly downregulated CD133, SOX2, Nanog, CD90, and OCT4 protein levels (Fig. 7F), indicating ELOVL3 is associated with the maintenance of HCC cell stemness.

To explore ELOVL3’s role in angiogenesis, qPCR showed si-ELOVL3 decreased VEGFA (Fig. 7G) and FGF2 (Fig. 7H) mRNA; Western blot confirmed reduced protein levels (Fig. 7I). ELISA further demonstrated si-ELOVL3 inhibited extracellular secretion of VEGFA (Fig. 7J) and FGF2 (Fig. 7K), suggesting ELOVL3 may contribute to HCC angiogenesis by regulating these pro-angiogenic factors.

ELOVL3 knockdown inhibits HCC cell proliferation, migration, invasion in vitro and tumor growth in vivo

To further confirm ELOVL3’s role in HCC progression, we evaluated its impact on cell proliferation, migration, invasion (in vitro) and tumor growth (in vivo). CCK-8 assays showed si-ELOVL3-transfected HCCLM3 and PLC/PRF/5 cells had significantly lower viability at 3 Day vs. si-NC (Fig. 8A). Colony formation assays revealed fewer colonies (≥ 50 cells) in the si-ELOVL3 group (Figs. 8B–C). EdU staining further demonstrated a marked reduction in EdU-positive proliferating cells after ELOVL3 knockdown (Figs. 8D–E). Wound healing assays showed si-ELOVL3 cells had lower wound closure rates at 48 h vs. si-NC (Figs. 8F–G). Transwell invasion assays confirmed fewer si-ELOVL3 cells invaded through Matrigel (Figs. 8H–I). Subcutaneous xenografts were established by inoculating nude mice with HCCLM3 cells (lenti-shELOVL3 or lenti-Ctrl). Tumors in the lenti-shELOVL3 group were significantly lighter than controls (Figs. 8J–K). IHC staining of xenograft tissues confirmed ELOVL3 downregulation, along with upregulated E-cadherin and downregulated CD133/VEGFA (Fig. 8L)—consistent with in vitro findings of suppressed proliferation, stemness and angiogenesis.

Fig. 8.

ELOVL3 knockdown inhibits HCC cell proliferation, migration, invasion in vitro and tumor growth in vivo. (A) CCK-8 proliferation assay showing that ELOVL3 knockdown significantly reduced cell viability in HCCLM3 and PLC/PRF/5 cells at Day 3 compared to si-NC control (n = 5 replicates). (B,C) Colony formation assay: representative images (B) and quantification (C) demonstrating fewer colonies (≥ 50 cells) in si-ELOVL3 groups after 10–14 days (n = 3 independent experiments). (D,E) EdU incorporation assay: representative images (D) and quantification (E) showing marked reduction in EdU-positive proliferating cells upon ELOVL3 knockdown (n = 3 independent experiments). (F,G) Wound healing assay: representative images at 0 h and 48 h (F) and quantification of wound closure rate (G), demonstrating impaired migration in si-ELOVL3 cells (n = 3 independent experiments). (H,I) Transwell invasion assay: representative images (H) and quantification (I) showing significantly fewer invading cells (stained with crystal violet) in si-ELOVL3 groups after 24 h (n = 3 independent experiments). (J,K) Subcutaneous xenograft model: nude mice (n = 4 per group) were inoculated with HCCLM3 cells (1 × 10⁶) stably infected with lenti-Ctrl or lenti-shELOVL3. Representative images of excised tumors (J) and tumor weights (K) at Day 21 post-inoculation, showing significantly smaller tumors in the lenti-shELOVL3 group. (L) Immunohistochemistry (IHC) staining of xenograft tissues for H&E, ELOVL3, CD133, E-cadherin, and VEGFA. ELOVL3 knockdown resulted in reduced ELOVL3, CD133, and VEGFA expression, along with increased E-cadherin expression—consistent with reduced stemness, angiogenic potential, and epithelial phenotype restoration. Scale bar: 50 μm. Data are represented as mean ± SD. P* < 0.05, P** < 0.01, P*** < 0.001 (two-tailed Student’s t-test).

Discussion

Hepatocellular carcinoma (HCC) remains a global therapeutic challenge due to its high recurrence and metastasis rates³³, which are closely linked to tumor stemness and angiogenesis³⁴. This study innovatively integrated stemness and angiogenesis-related genes (SARGs) to construct a prognostic model, shedding light on the intertwined regulatory mechanisms of these two processes and identifying potential therapeutic targets for HCC.

The gene selection process was designed to balance stringency and biological coverage, using a GeneCards relevance score > 10 to capture moderately-to-strongly associated stemness and angiogenesis-related genes while minimizing noise. Sensitivity analysis across alternative thresholds confirmed the robustness of the derived signature, indicating that the core prognostic genes are consistently identified regardless of minor adjustments to the relevance score cutoff.

Then, a key strength of our study lies in the rigorous establishment of the 9-SARG prognostic model (comprising DRD1, CDX2, ELOVL3, TKTL1, IGLON5, SHISA9, WNT1, CNTN6, and MMP3). By leveraging TCGA and ICGC datasets, we validated that this model effectively stratifies HCC patients into high- and low-risk groups, with high-risk patients exhibiting significantly worse overall survival (OS). Time-dependent ROC analysis further confirmed its robust predictive performance, with 1-, 3-, and 5-year AUC values consistently exceeding 0.7 in the TCGA cohort. The model achieved a concordance index (C-index) of 0.72 (95% CI: 0.68–0.76), demonstrating good discrimination ability. When integrated with clinical factors (e.g., TNM stage) into a nomogram, the model showed enhanced prognostic accuracy (AUC = 0.80) and good calibration, as evidenced by calibration curves and decision curve analysis (DCA). This underscores the model’s potential to guide individualized clinical decision-making, addressing the unmet need for reliable prognostic tools in HCC management.

Mechanistically, functional enrichment analyses revealed that differentially expressed SARGs are primarily involved in neuroactive ligand–receptor interaction, calcium signaling pathway, hormone signaling, and cAMP signaling pathway. These pathways are well-recognized for their roles in regulating both tumor stemness and angiogenesis: Cell communication and signaling, like neuroactive ligand–receptor interaction, links to CSC self-renewal and angiogenesis, while cAMP signaling pathway activation drives the secretion of pro-angiogenic factors (e.g., VEGF) and reinforces CSC stemness.

Previous studies have shown that the Tumor Immune Dysfunction and Exclusion (TIDE) score, the expression levels of immune checkpoint molecules (such as PD-L1 and PD-1), and the degree of immune cell infiltration may serve as biomarkers for predicting immunotherapy response in HCC^35–38. The results of immune infiltration analysis uncovered distinct immune landscapes between high- and low-risk groups. High-risk patients displayed higher TIDE scores, suggesting a greater computational prediction of immune evasion potential.

Tumor mutation analysis further identified TP53 as the most frequently mutated gene in high-risk patients (with a 22% higher mutation rate than low-risk patients), suggesting a potential crosstalk between TP53 mutation, SARG dysregulation, and HCC progression.

Among the 9 SARGs, ELOVL3, elongation of very long chain fatty acids-3, emerged as a notable functional mediator. Previous studies have demonstrated that ELOVL3 is related to increased fatty acid oxidation in brown adipocytes^39,40. However, its role in HCC was previously unknown. Our experimental results showed that ELOVL3 knockdown significantly reduced the sphere-forming capacity of HCC cells and downregulated the expression of CSC markers (CD133, CD90, Sox2, and OCT4), indicating an association with tumor stemness properties. Concurrently, ELOVL3 depletion decreased the mRNA and protein levels of pro-angiogenic factors and inhibited the secretion of VEGFA and FGF2, suggesting an association with angiogenic processes. Also, downregulation of ELOVL3 inhibited cell proliferation, migration, and invasion. In vivo, ELOVL3 knockdown led to smaller tumor volumes and lower tumor weights in nude mouse xenograft models.

These findings position ELOVL3 as a promising candidate for further investigation as a potential therapeutic target—disrupting ELOVL3 could potentially influence both CSC maintenance and tumor vascularization, addressing two key drivers of HCC malignancy. However, comprehensive mechanistic studies and clinical validation are required before translating these findings into therapeutic strategies.

This study also has limitations. First, a major limitation of this study is the reliance on retrospective public datasets (TCGA and ICGC) for model validation. Prospective, multi-center clinical trials enrolling HCC patients with standardized treatment and follow-up are essential to verify the model’s performance in real-world clinical settings. Additionally, the model’s applicability to diverse ethnic groups, different treatment strategies (e.g., surgical resection, transarterial chemoembolization, immunotherapy), and early-stage HCC requires further investigation. Therefore, the current model should be considered a preliminary prognostic tool that needs rigorous clinical validation before routine clinical application.

Second, our experimental results demonstrate associations between ELOVL3 depletion and reduced stemness markers, pro-angiogenic factor secretion, and tumor progression phenotypes. However, definitive causal mechanisms linking ELOVL3 to these processes remain unclear. ELOVL3 catalyzes very-long-chain fatty acid elongation, which may influence membrane composition, signaling lipid production, or mitochondrial function—all of which are implicated in CSC maintenance and angiogenesis. Future studies should include rescue experiments (e.g., re-expression of ELOVL3 in knockdown cells), pathway inhibitor assays (e.g., targeting fatty acid synthesis pathways), and lipidomics profiling to elucidate the precise molecular mechanisms underlying ELOVL3’s role in HCC stemness and angiogenesis. Additionally, clinical validation in patient-derived samples and exploration of ELOVL3 inhibitors as potential therapeutic agents warrant investigation.

Third, a key constraint of the present model is its sole reliance on SARGs and conventional clinical factors. Incorporating a broader panel of emerging biomarkers, including circulating tumor DNA, exosomal RNAs, and radiomics features that reflect tumor heterogeneity and microenvironmental characteristics, may further optimize its prognostic performance and facilitate its translation into routine clinical practice.

Fourth, the predictions of immunotherapy response based on TIDE and IPS—two widely used bioinformatic algorithms—have not been validated using actual clinical response datasets, a critical gap that restricts their immediate translational value in clinical practice. We recognize that both TIDE and IPS possess inherent limitations as in silico prediction tools: neither is designed to directly quantify immune cell functional activity nor capture the dynamic, patient-specific nature of actual immunotherapy responses. Therefore, these findings should be interpreted as hypothesis-generating rather than definitive, and future validation in well-powered prospective cohorts with systematically documented immunotherapy outcomes is imperative to verify their clinical utility. Nevertheless, these results provide a strong rationale for exploring novel combination therapeutic strategies that concurrently target SARG pathways and utilize immune checkpoint inhibitors (ICIs), with the goal of optimizing treatment efficacy for patients.

Conclusion

In conclusion, this study constructs a reliable SARG-based prognostic model for HCC, which requires prospective, multi-center validation to confirm its translational potential, and identifies ELOVL3 as a potential therapeutic target. These findings not only provide insights into HCC pathogenesis but also offer a basis for developing novel combination therapies (e.g., ELOVL3 inhibitors plus ICIs) to improve the outcomes of HCC patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUC

Area under the curve

ECs

Endothelial cells

SARG

Stemness and angiogenesis-related gene

VEGF

Vascular endothelial growth factor

FGF2

Fibroblast growth factor 2

CSCs

Cancer stem cells

EMT

Epithelial-mesenchymal transition

GO

Gene Ontology

HCC

Hepatocellular carcinoma

KEGG

Kyoto Encyclopedia of Genes and Genomes

LASSO

Least absolute shrinkage and selection operator

OS

Overall survival

LIHC

Liver hepatocellular carcinoma

PFS

Progression-Free Survival

PPI

Protein–protein interaction

ROC

Receiver operating characteristics

TCGA

The Cancer Genome Atlas

TMB

Tumor mutation burden

TME

Tumor microenvironment

ICGC

International Cancer Genome Consortium

DCA

Decision curve analysis

ssGSEA

Single-sample gene set enrichment analysis

TIDE

Tumor immune dysfunction and exclusion

Author contributions

Wengang Chen and Xin Cheng conceived the idea; Feng Cheng, Yan Shi and Xia Gao drafted of the main text and performed the experiments. Jiayi Han, Bingjie Xu, Yajuan Chen and Chang Fan drafted the figure and table. All authors read and approved the final version of the work to be published.

Funding

This work was supported by the Anhui Provincial Quality Engineering Project of Institutions of Higher Education (2022jyxm1715 and 2022jyxm1717), the Horizontal Projects of Lunan Houpu Pharmaceutical (HXKT2022014), and the Anhui Provincial Health and Family Planning Scientific Research Project (AHWJ2022a028).

Data availability

The TCGA data presented in this study are openly available in the GDC data portal (https://portal.gdc.cancer.gov/). ICGC-LIRI-JP were obtained from the ICGC database (https://www.dcc.icgc.org/projects/LIRI-JP). 5955 cell stemness-related genes and 2711 angiogenesis-related genes were obtained from genecards (https://www.genecards.org/). Further information is available from the corresponding author upon request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

All animal experiments involved in this study were approved by the Welfare and Ethics Committee for Laboratory Animals of Wannan Medical College (Approval No. WNMC-AWE-2023360). All animal experiments were performed in compliance with the GB/T 43051 − 2023 “General Requirements for Biosafety in Animal Experiments of Laboratory Animals in China”, and all efforts were made to minimize animal suffering and ensure animal welfare.