Identification of novel circulating protein biomarkers for hepatocellular carcinoma superior to alpha-fetoprotein through a stemness index and secretome analysis

Abstract

Objective

This study aimed to identify novel circulating protein biomarkers for hepatocellular carcinoma (HCC) superior to alpha-fetoprotein (AFP) by integrating tumor stemness index and secreted protein screening, and to explore their roles in prognosis, immune microenvironment, and tumor mechanisms.

Methods

Stemness-associated differentially expressed genes were identified from TCGA (The Cancer Genome Atlas) and GEO (Gene Expression Omnibus) databases after applying ComBat batch effect correction and quantile normalization to harmonize the datasets. WGCNA (weighted gene co-expression network analysis), secreted protein databases, and plasma sequencing were used to select candidate genes. Functional and immune infiltration analyses were performed, alongside multivariable Cox regression modeling and inverse variance weighted (IVW) Mendelian Randomization (MR) for prognostic and causal validation.

Results

A total of 19 genes significantly associated with tumor stemness and detectable in circulation were identified. After filtering based on clinicopathological features and survival analysis, 10 key genes were selected. The receiver operating characteristic (ROC) curve analysis revealed that, aside from FCN3 (ficolin 3), the diagnostic sensitivity and specificity of CRHBP (corticotropin-releasing hormone binding protein), CLEC3B (C-type lectin domain family 3 member B), TEK (TEK receptor tyrosine kinase), SOGA1 (SOGA family member 1), IL33 (interleukin 33), CXCL12 (C-X-C motif chemokine ligand 12), NENF (neudesin neurotrophic factor), and ITM2B (integral membrane protein 2B) (AUCs [area under the curves] ranging from 0.67 to 0.99) surpassed those of AFP (AUC = 0.64). Cibersort analysis indicated significant associations of CRHBP, CLEC3B, TEK, IL33, NENF, CXCL12, and AFP with various immune cell infiltrates. The gene set enrichment analysis (GSEA) suggested that CRHBP, CLEC3B, TEK, NENF, and AFP were primarily enriched in cell cycle-related pathways, whereas IL33, CXCL12, SOGA1, and ITM2B were involved in MAPK, GNRH, and NOTCH signaling. A multivariable Cox regression model identified CLEC3B, TEK, and NENF for constructing a prognostic signature, which demonstrated strong predictive performance in the TCGA cohort, as evidenced by the concordance index (C-index), time-dependent ROC analysis (1-/3-/5-year AUCs: 0.74/0.72/0.70), and decision curve analysis. IVW MR analysis further confirmed a causal relationship between cis-expression quantitative trait loci (cis-eQTLs) of the NENF gene and HCC risk (P = 0.024), with no significant pleiotropy or heterogeneity detected. Elevated NENF protein expression, validated by proteomic and immunohistochemical analyses of public databases and a tissue microarray cohort, was significantly associated with HCC progression and reduced overall survival (hazard ratio [HR], 2.456; 95% CI, 1.271–4.960; P = 0.012).

Conclusion

This multi-omics study reveals a panel of eight circulating proteins with diagnostic performance superior to that of AFP and independent prognostic value in HCC, nominating NENF as a causal biomarker for disease progression and survival.

Introduction

Hepatocellular carcinoma (HCC), one of the most common malignant tumors worldwide, causes over 900,000 new cases and more than 800,000 deaths annually, and patients with metastatic or recurrent HCC face particularly poor prognoses [1]. Hepatitis B virus (HBV) infection remains a major risk factor for HCC development. Current treatment strategies include surgery, radiotherapy, and chemotherapy; however, the five-year survival rate drops to approximately 18% once metastasis occurs [2]. Alpha-fetoprotein (AFP), an oncofetal glycoprotein, has long served as a serum biomarker for HCC and is often used in combination with ultrasound and other imaging modalities [3]; nonetheless, its clinical utility is limited by suboptimal sensitivity and specificity, inconsistent elevation in some HCC patients, and variability across detection assays [4, 5]. Therefore, a deeper understanding of the molecular mechanisms driving HCC pathogenesis is essential to develop novel diagnostic strategies and improve patient outcomes.

Cancer stem cells (CSCs) represent a subpopulation of tumor cells with self-renewal capacity and the ability to generate heterogeneous tumor cell lineages [6]. They play a pivotal role in tumor initiation, progression, metastasis, and recurrence [7–9]. Accumulating evidence indicates that the growth, recurrence, and therapeutic resistance of HCC are also driven by CSCs [7, 8]. Highly tumorigenic CSCs maintain their self-renewal through activation of stemness-related signaling pathways such as WNT, Hedgehog (Hh), and NOTCH [10]. Moreover, CSCs can enter a dormant state (arrested in G0 phase), reducing proliferative activity and thereby developing resistance to chemotherapy, ultimately leading to tumor relapse [11]. Recent studies have further revealed complex interactions between CSCs and the tumor immune microenvironment (TIME) [12–14]. For instance, IL-17 secreted by Th17 cells enhances CSC self-renewal in multiple tumor models [15], and tumor-associated macrophages (TAMs) promote immune evasion by sustaining CSC stemness and suppressing immune cell infiltration [16]. Currently, stemness indices based on DNA methylation (mDNAsi) and mRNA expression (mRNAsi) have been developed to quantify the level of tumor stemness [17]. However, most studies remain at the pan-cancer level, and the prognostic value and biological functions of stemness-related genes in HCC have not yet been systematically explored.

Secreted proteins serve as crucial mediators of intercellular communication and play key roles in HCC proliferation, metastasis, and microenvironment regulation [18–21]. For example, high mobility group box 1 (HMGB1) promotes HCC invasion and metastasis by activating the RAGE/NF-κB pathway [18]; and exosome-derived 14-3-3ζ protein can induce Treg cell differentiation, thereby suppressing antitumor immune responses [19]. These findings highlight the potential of secreted proteins as diagnostic biomarkers for HCC. In this study, we integrated data from TCGA and GEO databases to identify differentially expressed genes associated with stemness indices in HCC. By further incorporating secreted protein profiles, we validated their expression in plasma dataset. Our ultimate goal is to identify novel circulating biomarkers that outperform AFP, providing new molecular tools for early diagnosis and prognostic assessment of HCC.

Methods

Data acquisition and differential gene expression analysis

RNA sequencing data and complete clinical records from patients with HCC were obtained from The Cancer Genome Atlas (TCGA) (https://portal.gdc.cancer.gov/). Clinical variables included age, histologic grade, and tumor stage. Differential expression analysis was performed using the R package “limma”. Genes with an absolute log2 fold change (|log2FC|) ≥ 0.3 and an adjusted P-value < 0.05 were considered differentially expressed genes (DEGs). Additionally, a gene set encoding proteins known to be secreted into the bloodstream was retrieved from the Human Protein Atlas (HPA) (https://www.immport.org).

Stemness index expression, clinicopathological and survival analysis

The mRNA expression-based stemness index (mRNAsi) was calculated to evaluate tumor stemness using the one-class logistic regression machine learning algorithm (OCLR), as described by Malta et al. [17]. The mRNAsi score ranges from 0 (no expression) to 1 (complete expression). Differences in mRNAsi expression between normal and HCC tissues, along with its association with clinicopathological features, were analyzed using GraphPad Prism 9. The correlation between mRNAsi and overall survival (OS) was assessed using the R package “survival”, with a P-value < 0.05 considered statistically significant.

Weighted co-expression network analysis and secreted gene screening

Weighted gene co-expression network analysis (WGCNA) was conducted using the R package “WGCNA” to identify co-expressed gene modules and explore their association with the stemness index, along with hub genes within the network. The top 50% of genes with the highest variance were retained for further analysis. A weighted adjacency matrix was converted into a topological overlap matrix (TOM) to assess intergenic connectivity, and a hierarchical clustering dendrogram was constructed to identify distinct gene modules, each represented by a unique branch and color. Genes from key WGCNA modules, HCC-related DEGs, and blood-secreted protein-coding genes were intersected, and the resulting gene list was validated using the manually annotated Swiss-Prot entries within the UniProt database.

Validation using GEO dataset and functional enrichment analysis

The gene expression dataset GSE142987 (platform GPL20795) was downloaded from the Gene Expression Omnibus (GEO) (http://www.ncbi.nlm.nih.gov/geo), comprising 35 HCC plasma samples and 30 normal plasma samples. To harmonize the datasets, ComBat batch effect correction and quantile normalization were applied. The R packages “stats” and “car” were used to evaluate differences in candidate gene expression between groups, and results were visualized using “ggplot2”. Genes showing consistent expression trends and significant differential expression (P < 0.05) in both tissue and blood were selected for subsequent analysis. Functional enrichment analysis of Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways and Gene Ontology (GO) terms was performed with the R package “clusterProfiler”, using a significance threshold of P < 0.05.

Association of gene expression with prognosis and clinicopathological features

The Kaplan-Meier method from the R package “survival” was used to analyze the relationship between gene expression and overall survival (OS) in HCC patients. The “ggplot2” package was employed to visualize differences in gene expression across clinicopathological variables. A P-value < 0.05 was considered statistically significant. Genes that were significantly associated with both prognosis and clinicopathological parameters were defined as core genes for further investigation.

ROC analysis and mechanistic exploration of core genes

Receiver operating characteristic (ROC) curves were generated and the area under the curve (AUC) was computed using the R package “pROC” to evaluate the diagnostic sensitivity and specificity of core genes in distinguishing HCC from normal plasma samples. The gene set enrichment analysis (GSEA) was conducted using the MsigDB database (reference set: c2.cp.kegg_legacy.v2023.2.Hs.symbols.gmt) from the GSEA platform to explore potential mechanisms and pathways involved in HCC pathogenesis. Samples were divided into high- and low-expression groups based on median gene expression values. Pathways with an adjusted P < 0.05, false discovery rate (FDR) < 0.25, and |NES| ≥ 1 were considered significantly enriched.

Single-cell RNA sequencing analysis and correlation with stemness and immune infiltration

The Tumor Immune Single-Cell Hub 2 (TISCH2) database (http://tisch.comp-genomics.org/home/) [22] was used to analyze the expression distribution and cell type specificity of core genes in HCC single-cell RNA sequencing datasets (e.g., GSE146409). Spearman correlation analysis was applied to assess the relationship between core genes and the stemness index, with results visualized using “ggplot2”; a P-value < 0.05 was considered significant. The R package “CIBERSORT” was utilized to analyze differences in immune cell infiltration between normal and HCC tissues. Correlations between gene expression and immune cell subsets were further evaluated, with statistical significance set at P < 0.05. Protein expression levels of core genes in HCC and normal liver tissues were examined using the UALCAN database (https://ualcan.path.uab.edu/analysis.html) [23].

Prognostic model construction and evaluation

Multivariable Cox regression analysis was performed using the R package “survival” to identify core genes independently associated with survival. A risk score model was constructed using the formula: RiskScore = Σ(βi × Si), where β represents the regression coefficient and S denotes the gene expression level. The optimal cutoff value was determined using the “survminer” package, and patients were stratified into high- and low-risk groups. Survival differences between groups were compared using Kaplan-Meier curves, and time-dependent ROC analysis was conducted with the “timeROC” package to evaluate predictive accuracy. The “ComplexHeatmap” package was used to visualize associations between gene expression and risk scores. A nomogram integrating clinicopathological parameters (stage, grade, TNM, and risk score) was developed with the “rms” package to predict 1-, 3-, and 5-year OS probabilities. Calibration curves were plotted to assess agreement between predicted and observed outcomes. The concordance index (C-index) was calculated using the “pec” and “survival” packages, and ROC analysis (“timeROC”) was used to compare the sensitivity of the risk score with conventional clinical parameters. Decision curve analysis (DCA) was performed with the “stdca” package to evaluate clinical net benefit, and results were visualized using “ggplot2”. To enhance the reliability of the prognostic model and mitigate overfitting, we employed least absolute shrinkage and selection operator (LASSO)-Cox regression for variable selection and regularization during the construction of the multivariable Cox proportional hazards model based on CLEC3B, TEK, and NENF. The optimal penalty parameter (λ) was determined through 10-fold cross-validation, ultimately yielding a parsimonious set of gene markers with the strongest prognostic predictive value, thereby ensuring model simplicity and robustness.

Mendelian Randomization (MR) analysis

Cis-expression quantitative trait loci (cis-eQTLs) of the genes of interest were obtained from the OpenGWAS database (https://gwas.mrcieu.ac.uk/) as instrumental variables. Strongly associated single-nucleotide polymorphisms (SNPs) were selected based on the following criteria: P < 5 × 10⁻⁸, linkage disequilibrium parameters r² < 0.1 within a 2000 kb window, and F-statistic > 10 using the “TwoSampleMR” R package. Outcome data were sourced from the HCC genome-wide association study (GWAS) dataset in OpenGWAS (ieu-b-4915), which included 372,366 samples (350 cases and 372,016 controls). Causal relationships between genes and HCC were evaluated using five MR methods: MR Egger, weighted median, inverse variance weighted, simple mode, and weighted mode.

Immunohistochemical (IHC) analysis

All tissue specimens were embedded in paraffin blocks. A high-throughput tissue microarray designed for (Cat. no. HLivH180Su14) containing 90 HCC tissue cores and 90 matched adjacent non-tumor tissue spots was utilized. IHC analysis was performed using the EliVision™ Plus two-step detection system (Product no. KIT-9903, Fuzhou Maixin) in accordance with the manufacturer’s protocol. A mouse monoclonal antibody against NENF (Cat. No. FNab05661) was applied at a dilution of 1:200. Phosphate-buffered saline (PBS) was used as the negative control. Two experienced pathologists, blinded to the clinical data, independently evaluated the immunostaining results according to both staining intensity and the proportion of positively stained cells; based on the NENF protein expression intensity scores, patients were classified into two groups: a high-expression group (scores ranging from + to +++, or expression scores of 3–9) and a low/negative-expression group (scores from negative to ±, or expression scores of 0–2) [24].

Statistical analysis

All statistical analyses were performed using R software (version 4.3.2) and GraphPad Prism 9. Group comparisons were conducted using the Student’s t-test (for normally distributed data with homogeneity of variance), Welch’s t-test (for normally distributed data with unequal variances), or the Wilcoxon rank-sum test (for non-normally distributed data). Survival differences were analyzed using the Kaplan-Meier method with the log-rank test. A two-sided P-value < 0.05 was considered statistically significant for all analyses.

Results

Identification of stemness-associated key gene modules via WGCNA and their expression, prognostic, and clinicopathological significance in HCC

To establish the theoretical foundation of tumor stemness within our analytical framework, we first evaluated the expression pattern and clinical relevance of the mRNA expression-based stemness index (mRNAsi) in HCC. mRNAsi expression was significantly elevated in HCC tissues compared with normal liver tissues (P < 0.01; Fig. 1A). Furthermore, mRNAsi levels were significantly higher in poorly differentiated (G3–G4) tumors than in well-differentiated (G1–G2) tumors (P < 0.01; Fig. 1B). Kaplan–Meier survival analysis revealed that patients with high mRNAsi had significantly shorter OS (P < 0.01; Fig. 1C). These results indicate that tumor stemness is closely associated with malignant progression and poor prognosis in HCC.

Fig. 1.

Identification of stemness index modules. A, B Expression analysis of the mRNA expression-based stemness index (mRNAsi) in hepatocellular carcinoma (HCC). C Kaplan-Meier survival analysis evaluating the association between mRNAsi and overall survival (OS) in patients with HCC. D Comparative analysis of mRNA expression profiles between HCC tissues (n = 374) and normal tissues (n = 50) identified 11,246 differentially expressed genes (DEGs). E Correlation analysis of 9,967 HCC genes and determination of the soft threshold β. F Dendrogram illustrating the correlation coefficients between each co-expression gene module and the stemness index trait. G Phenotypic correlation analysis of gene co-expression and stemness index in HCC. *** P < 0.001, ****P < 0.0001

To systematically identify core genes related to HCC stemness, we constructed a weighted gene co-expression network (WGCNA). Comparison of mRNA expression profiles between 374 HCC tissues and 50 normal tissues identified 11,246 DEGs, of which 1,175 were downregulated and 10,071 were upregulated (Fig. 1D). After preprocessing, 9,967 genes were included in the co-expression network analysis. A soft thresholding power β of 6 was selected to ensure a scale-free network topology (Fig. 1E), and a gene dendrogram was successfully constructed (Fig. 1F). Module–trait relationship analysis showed that the turquoise (r = − 0.66, P = 3 × 10⁻⁴⁸) and red (r = − 0.79, P = 8 × 10⁻⁸¹) modules were significantly negatively correlated with the stemness index, whereas the yellow (r = 0.62, P = 4 × 10⁻⁴⁰), blue (r = 0.46, P = 6 × 10⁻²¹), and magenta (r = 0.40, P = 6 × 10⁻¹⁶) modules were significantly positively correlated with the stemness index (Fig. 1G). Thus, genes from these five modules were included as stemness-related candidate genes for subsequent analyses.

Screening, independent validation, and functional enrichment of stemness-associated blood-candidate genes

Given the clinical potential of blood-based biomarkers, we focused on protein-coding genes that are secreted into the bloodstream. Intersecting genes from the key WGCNA modules, HCC DEGs, and genes encoding secreted proteins from the human secretome database yielded 67 overlapping genes, of which 40 were upregulated and 27 were downregulated in HCC (Fig. 2A and B). To validate the diagnostic potential of these 67 genes, we analyzed their expression in an independent plasma dataset (GSE142987), which included 35 HCC and 30 normal plasma samples. T-test analysis identified 19 genes that were significantly differentially expressed between HCC and normal plasma (P < 0.05). Among these, AFP, ANGPT1, CCL15, ITIH4, ITM2B, JAM3, NENF, SOGA1, VWF, and XYLT2 were upregulated in the HCC group, whereas CCL2, CCL21, CLEC3B, CRHBP, CXCL12, FCN2, FCN3, IL33, and TEK were downregulated (Fig. 2C). These 19 genes were selected as core candidates for further investigation.

Fig. 2.

Screening and comprehensive analysis of blood-secreted genes. A, B The intersection of genes from WGCNA, HCC DEGs, and genes encoding proteins secreted into the blood yielded 67 overlapping genes, including 40 upregulated and 27 downregulated. C T-test analysis identified 19 differentially expressed blood-secreted stemness-associated genes in the GSE142987 dataset. D, E KEGG and GO analyses were performed to characterize the functional pathways of these 19 selected differentially expressed blood-secreted genes. WGCNA: weighted gene co-expression network analysis, HCC: hepatocellular carcinoma, DEGs: differentially expressed genes, KEGG: Kyoto Encyclopedia of Genes and Genomes, GO: Gene Ontologyns: not significant, *P < 0.05, ** P < 0.01, *** P < 0.001

To explore the biological functions of these candidate genes, we performed GO and KEGG enrichment analyses. KEGG analysis indicated significant enrichment in pathways including rheumatoid arthritis, viral protein interaction with cytokines and cytokine receptors, cytokine–cytokine receptor interaction, chemokine signaling pathway, NF-κB signaling pathway, HIF-1 signaling pathway, leukocyte transendothelial migration, PI3K–Akt signaling pathway, and influenza A (Fig. 2D). GO biological process (BP) analysis revealed that these genes are primarily involved in monocyte chemotaxis, myeloid leukocyte migration, leukocyte chemotaxis, neutrophil chemotaxis, lymphocyte migration, cytokine-mediated signaling pathway, regulation of leukocyte chemotaxis, neutrophil migration, granulocyte migration, and T cell migration (Fig. 2E). These findings strongly suggest that the candidate genes are broadly involved in immune regulation and inflammatory responses, which are closely linked to the tumor microenvironment in HCC.

Prognostic value and diagnostic efficacy of core genes

We next evaluated the clinical relevance of the 19 genes. Kaplan–Meier survival analysis showed that the expression of FCN3, CRHBP, IL33, CLEC3B, TEK, NENF, and SOGA1 was significantly associated with overall survival in HCC patients (Fig. 3A and G). Furthermore, the expression of these genes was significantly correlated with various clinicopathological features. For example, in advanced-stage (T3–T4) tumors, CRHBP, CLEC3B, ITIH4, and TEK were downregulated, whereas SOGA1 was upregulated (Fig. 3H). In stage III–IV tumors, FCN3, CRHBP, IL33, CLEC3B, CCL15, ITIH4, and TEK were downregulated, and SOGA1 was upregulated (Fig. 3I). In poorly differentiated (G3–G4) tumors, FCN3, CRHBP, CXCL12, CLEC3B, VWF, TEK, and JAM3 were downregulated, whereas AFP, NENF, and SOGA1 were upregulated (Fig. 3J). In tumors with microvascular invasion, FCN3, CRHBP, CXCL12, CLEC3B, and TEK were downregulated (Fig. 3K). Integrating survival and clinicopathological analyses, we ultimately identified 10 molecules with significant clinical relevance: CRHBP, CLEC3B, TEK, SOGA1, FCN3, IL33, AFP, CXCL12, NENF, and ITM2B, which were defined as core genes for subsequent studies.

Fig. 3.

Analysis of feature genes in relation to hepatocellular carcinoma (HCC) pathological characteristics and diagnostic sensitivity. A to G Kaplan-Meier analysis was performed to evaluate the association between the expression of FCN3, CRHBP, IL33, CLEC3B, TEK, NENF, and SOGA1 and overall survival (OS) in patients with HCC. H Differential expression analysis of the 19 feature genes across various clinicopathological parameters in HCC. I Analysis of differential gene expression of the 19genes according to clinical stage. J Differential expression of the 19 genes based on pathological grade (G1 to G4). K Expression differences of the 19 genes stratified by microvascular invasion status. L Receiver operating characteristic (ROC) curve analysis assessing the diagnostic performance of the differentially expressed genes for HCC. ns: not significant, *P<0.05, ** P<0.01, *** P<0.001

We further evaluated the diagnostic efficacy of these 10 core genes in distinguishing HCC from normal tissues using ROC curve analysis. All genes except FCN3 had AUC values greater than 0.7, including CRHBP, CLEC3B, TEK, SOGA1, IL33, CXCL12, NENF, ITM2B, and AFP (Fig. 3L). Notably, CRHBP, CLEC3B, TEK, SOGA1, IL33, CXCL12, NENF, and ITM2B demonstrated diagnostic performance comparable to or better than that of the conventional biomarker AFP, highlighting their strong potential as novel diagnostic markers for HCC.

Elucidation of core gene mechanisms via GSEA and single-cell sequencing and expression profiling

To elucidate the potential biological functions and signaling pathways of the core genes, we conducted single-gene GSEA. The results revealed that the key genes do not exert broad effects on entire pathways but rather specifically regulate downstream core components. Specifically, CRHBP, CLEC3B, SOGA1, FCN3, and NENF were primarily enriched in pathways involving ECM–receptor interaction, cytokine receptor binding, chemokine activity, focal adhesion, cell cycle, and DNA replication (Fig. 4A). In contrast, TEK, IL33, CXCL12, and ITM2B were significantly enriched in NOTCH, MAPK, JAK-STAT, GNRH, Toll-like receptor, and TGF-β signaling pathways (Fig. 4B). Further analysis demonstrated pathway-specific regulatory roles: TEK and IL33 showed marked enrichment in gene sets related to cell differentiation and inflammatory responses within the JAK-STAT and NOTCH pathways, while CLEC3B and FCN3 predominantly influenced molecular functions associated with cell migration and adhesion in ECM-receptor interaction and focal adhesion pathways. Elucidation of this specific regulatory pattern not only strengthens the biological relevance of the GSEA findings to HCC but also provides a more precise mechanistic framework for understanding how these core genes promote HCC progression through modulation of the tumor microenvironment, immune responses, and key oncogenic pathways.

Fig. 4.

Functional mechanisms and single-cell expression distribution of feature genes in hepatocellular carcinoma (HCC). A, B The gene set enrichment analysis (GSEA) was conducted. C to N Expression distribution of core genes at single-cell resolution was analyzed in the HCC single-cell RNA sequencing dataset GSE146409 using TISCH2

Using single-cell RNA sequencing data from the TISCH2 database (GSE146409), we resolved the expression patterns of the core genes at single-cell resolution. The results showed that IL33, CRHBP, TEK, and FCN3 were predominantly expressed in endothelial cells; CXCL12 was mainly expressed in fibroblasts and mononuclear/macrophage cells; whereas NENF and ITM2B were widely expressed across malignant cells, endothelial cells, fibroblasts, and hepatocytes (Fig. 4C and N). These findings provide important spatial resolution regarding the cellular contexts in which these genes may operate within the HCC tumor microenvironment.

Core gene associations with stemness and immune infiltration and protein level validation

Spearman correlation analysis showed that CRHBP, CLEC3B, CXCL12, FCN3, IL33, ITM2B, SOGA1, and TEK were negatively correlated with the tumor stemness index (P < 0.05), whereas AFP and NENF were positively correlated with the stemness index (Fig. 5A). This indicates that most candidate markers are inversely associated with stemness features, while AFP and NENF may directly contribute to the maintenance or promotion of stemness.

Fig. 5.

Analysis of feature genes in relation to stemness index, immune cells, and protein-level expression. A Spearman correlation analysis between core genes and tumor stemness index. B Differential immune cell infiltration analysis between normal controls and patients with hepatocellular carcinoma (HCC) using the CIBERSORT method. C Correlation between core gene expression and immune cell infiltration. D to M Differential expression of core gene-encoded proteins in HCC based on the UALCAN dataset. ns: not significant, *P < 0.05, ** P < 0.01, *** P < 0.001

Analysis of immune cell infiltration using the CIBERSORT algorithm revealed that, compared with normal tissues, HCC tissues had significantly lower proportions of plasma cells, γδ T cells, monocytes, M2 macrophages, activated mast cells, and neutrophils, and significantly higher proportions of resting CD4⁺ memory T cells, regulatory T cells (Tregs), activated NK cells, M0 macrophages, resting dendritic cells, and resting mast cells (Fig. 5B). These results depict a characteristic immunosuppressive microenvironment in HCC.

Further correlation analysis revealed close relationships between core genes and specific immune cell subsets (Fig. 5C). For instance, CRHBP, CLEC3B, TEK, and IL33 showed positive correlations with anti-tumor immune cells (e.g., M1 macrophages, resting CD4 ⁺ memory T cells) and negative correlations with pro-tumor immune cells or M2-polarized macrophages (e.g., Tregs, M0 macrophages). In contrast, AFP and NENF were positively correlated with immunosuppressive cells such as Tregs. These associations strongly suggest that these candidate markers may not only have diagnostic value but could also influence HCC progression and response to immunotherapy by modulating the tumor immune microenvironment.

To confirm our findings at the protein level, we analyzed the protein expression of core genes using the UALCAN database. Consistent with mRNA trends, the protein levels of CRHBP, TEK, FCN3, IL33, and CXCL12 were significantly lower in HCC tissues than in normal liver tissues, whereas the protein levels of CLEC3B, AFP, NENF, and ITM2B were significantly elevated in HCC (Fig. 5D and M). These results further validate the reliability of these candidate markers at the translational level.

Construction of a core gene prognostic model and evaluation of its clinical utility

Based on the seven genes significantly associated with survival (FCN3, CRHBP, IL33, CLEC3B, TEK, NENF, SOGA1), we performed multivariable Cox regression analysis and ultimately identified three independent prognostic factors: CLEC3B, TEK, and NENF. A prognostic risk score model was constructed using these genes (Table 1). Using the median risk score as the cutoff, patients in the TCGA cohort were stratified into high-risk and low-risk groups. Kaplan–Meier curves showed that patients in the high-risk group had significantly shorter OS than those in the low-risk group (P < 0.001; Fig. 6A). Time-dependent ROC analysis demonstrated that the three-gene model performed well in predicting 1-, 3-, and 5-year survival, with AUC values of 0.738, 0.715, and 0.699, respectively (Fig. 6B). The distribution of risk scores, survival status, and gene expression visually illustrated the model’s discriminative power (Fig. 6C).

Table 1.

Parameters of the multivariable Cox model and the three identified biomarkers (CLEC3B, TEK, and NENF)

Gene	Coef	HR	95%CI low	95%CI high	P value
CLEC3B	-0.521779641	0.593463456	0.483836687	0.727929243	5.51E-07
TEK	0.211500612	1.235530722	0.928494296	1.644098593	0.146784740
NENF	0.203726361	1.225962636	0.946481338	1.587970437	0.122760136

HR hazard ratio

Fig. 6.

Construction and validation of a prognostic model in hepatocellular carcinoma (HCC). A, C Each patient with HCC was assigned a risk score based on coefficients derived from multivariate Cox regression analysis and stratified into high-risk or low-risk groups. The survival status and gene expression patterns were compared between the two groups. Panel C shows that patients in the high-risk group had lower survival rates than those in the low-risk group. B Performance of the prognostic model in predicting 1-, 3-, and 5-year overall survival (OS) of patients with HCC. D The nomogram incorporates predictive variables including pathological stage, pathological grade, T stage, N stage, M stage, and risk score. E to G Calibration curves demonstrated close agreement between the nomogram-predicted and observed probabilities of 1-, 3-, and 5-year survival, confirming the reliability of the nomogram. H Concordance index (C-index) analysis incorporating RiskScore, T stage, Stage, and Grade. I ROC curve analysis based on multiple clinical variables. J to L Decision curve analysis evaluating the net benefit of Model1 and Model2 for predicting 1-, 3-, and 5-year survival in patients with HCC. M External validation of the prognostic analysis for high- and low-risk groups stratified using the ICGC database. N External validation of a prognostic model based on the ICGC database for predicting 1-, 3-, and 5-year survival in patients with HCC

To facilitate the clinical application of this prognostic model, we constructed a nomogram that integrated the risk score with clinicopathological parameters (stage, grade, TNM stage) to predict the probability of 1-, 3-, and 5-year OS in HCC patients (Fig. 6D). Calibration curves showed excellent agreement between the nomogram-predicted survival probabilities and actual observations (Fig. 6E and G). Concordance index (C-index) analysis indicated that the risk score had higher predictive accuracy for 2- to 8-year survival than conventional T stage, overall stage, or grade (Fig. 6H). Multi-index ROC curve analysis showed that the comprehensive predictive ability of the risk score (AUC = 0.715) was superior to that of any single clinical parameter (Fig. 6I). Decision curve analysis (DCA) further demonstrated that the model including the risk score (Model 2) provided greater clinical net benefit than the model containing only traditional clinical parameters (Model 1) (Fig. 6J and L). Together, these validation steps indicate that the risk score based on CLEC3B, TEK, and NENF is a robust and clinically useful independent prognostic indicator.

To assess the model’s generalizability and mitigate overfitting, an external validation was performed using the ICGC cohort. The model achieved AUC values of 0.764, 0.676, and 0.604 for predicting 1-, 3-, and 5-year OS (Fig. 6M and N), respectively. These results confirm the model’s robust prognostic performance and generalizability.

MR analysis validates causal relationships at the genetic level

To investigate whether genetic causal relationships exist between these genes and HCC risk, we performed MR analysis. The results indicated that only the cis-eQTL of NENF showed a significant causal association with HCC risk (P = 0.024; Fig. 7A). Heterogeneity tests and pleiotropy tests (all with P > 0.05) indicated the validity of the instrumental variables, with no significant heterogeneity or horizontal pleiotropy observed (Table 2, and Table 3). The intercept from MR-Egger regression and the symmetry of the funnel plot supported these conclusions (Fig. 7C and D). Leave-one-out sensitivity analysis confirmed that no single SNP unduly influenced the overall results, indicating that the findings are robust and reliable (Fig. 7B). The MR analysis provides genetic evidence that elevated expression of NENF may be a potential risk factor for HCC.

Fig. 7.

Mendelian Randomization (MR) analysis. A Forest plot illustrating the causal association between genetically predicted CLEC3B, TEK, and NENF expression and hepatocellular carcinoma (HCC). B Leave-one-out sensitivity analysis evaluating the stability of the causal estimates. C Scatter plot depicting the relationship between instrumental variable effects and outcome associations. D Funnel plot assessing potential heterogeneity and directional pleiotropy in the MR estimates

Table 2.

Results of heterogeneity testing for the NENF gene

Id.exposure	Outcome	Exposure	Method	Q value	Q_df	Q_P value
eqtl-a-ENSG00000117691	Liver cancer	NENF	MR Egger	29.57426902	18	0.061796867
eqtl-a-ENSG00000117691	Liver cancer	NENF	Inverse variance weighted	29.71494396	19	0.055540164

Table 3.

Results of pleiotropy testing for the NENF gene

Id.exposure	Outcome	Exposure	Egger_intercept	Se	P value
eqtl-a-ENSG00000117691	Liver cancer	NENF	1.68E-05	5.73E-05	0.773167298

Expression of NENF in HCC and its prognostic significance

We conducted a systematic validation of the candidate biomarker NENF through a series of experiments. Analysis of the CPTAC cohort using the UALCAN proteomics database demonstrated that NENF protein expression was significantly elevated in HCC tissues compared with normal liver tissues (Fig. 8A and C). IHC data from the Human Protein Atlas (HPA) database indicated positive NENF expression in more than 80% of HCC samples (Fig. 8D). Survival analysis based on the HPA database revealed that high NENF expression was significantly associated with reduced OS among patients with HCC. In cohort 1, OS rates were 39% in the high-expression group versus 54% in the low-expression group; in cohort 2, the corresponding rates were 56% versus 83% (Fig. 8E and F), suggesting that high NENF expression may represent a potential indicator of poor prognosis in HCC.

Fig. 8.

External validation of NENF protein expression in hepatocellular carcinoma (HCC) tissues. A Pan-cancer analysis of NENF protein expression using the UALCAN database. B, C Analysis of NENF protein expression in patients with HCC from the UALCAN database. D Immunohistochemical (IHC) validation of NENF protein expression in HCC tissues based on the Human Protein Atlas (HPA) database. E, F Prognostic analysis of NENF expression in HCC using two independent cohorts from the HPA. G, H Representative images and scoring of NENF expression detected by tissue microarray in 90 HCC cases. I Prognostic value of high and low NENF expression levels based on tissue microarray data in patients with HCC

To obtain direct experimental validation, we performed IHC using a commercially available tissue microarray (Cat no. HLivH180Su14) comprising 90 HCC cases with paired adjacent non-tumor tissues. The results showed specific upregulation of NENF protein in the cytoplasm of HCC cells (Fig. 8G). Quantitative analysis based on IHC scores confirmed significantly higher NENF expression in tumor tissues compared with adjacent non-tumor tissues (Fig. 8H; P < 0.001). In addition, patients with high NENF expression had significantly shorter OS than those with low expression (hazard ratio [HR], 2.456; 95% CI, 1.271–4.960; P = 0.012; Fig.8I).

Discussion

Hepatocellular carcinoma (HCC) is characterized by its high malignancy and poor prognosis, primarily attributable to its rapid progression, aggressive invasiveness, and high metastatic potential [1, 25, 26]. Alpha-fetoprotein (AFP) is currently the most widely used serological biomarker for HCC in clinical practice [3, 4, 27]; however, its diagnostic utility is substantially limited by low sensitivity and a high false-negative rate in early-stage or small HCCs, as well as by false-positive elevations in certain benign liver diseases such as chronic hepatitis and cirrhosis [28, 29]. Therefore, the identification and validation of novel blood-based biomarkers with superior performance to AFP is of critical importance for improving early detection and patient outcomes.

Against this background, the present study innovatively integrated tumor stemness characteristics with secretome analysis to identify novel blood-based biomarkers with high sensitivity and specificity for HCC. Using a multi-omics data integration approach and systems biology methods, we first employed WGCNA to construct core gene modules closely associated with the HCC stemness index, then focused on protein-coding genes with the potential to be secreted into the bloodstream. Through a multi-step screening and validation process, we identified 10 candidate genes: CRHBP, CLEC3B, TEK, SOGA1, FCN3, IL33, AFP, CXCL12, NENF, and ITM2B. Subsequent ROC curve analysis highlighted eight of these (CRHBP, CLEC3B, TEK, SOGA1, IL33, CXCL12, NENF, and ITM2B) which exhibited diagnostic performance comparable to or better than AFP, indicating substantial potential for clinical translation.

Integrating multi-omics data analysis and independent validation using a GEO cohort, this study demonstrated that the expression of core biomarkers such as CLEC3B, TEK, and NENF showed significant variation across major HCC clinicopathological features, including T stage, histologic grade, clinical stage, and microvascular invasion, and was closely associated with tumor progression. For example, CLEC3B and TEK expression was markedly downregulated in advanced (Stage III–IV), poorly differentiated (G3–G4), and microvascular invasion–positive HCC tissues, whereas SOGA1 and NENF were upregulated during disease progression. These findings suggest that the expression of the identified biomarkers is not uniform across different HCC developmental stages or etiological backgrounds but exhibits dynamic changes that reflect malignant progression, indicating their potential utility as stage- or grade-specific diagnostic markers.

To further elucidate the biological underpinnings of these observations, we next discuss the functional roles of several key candidates based on existing literature and our enrichment analyses. Integral membrane protein 2B (ITM2B), located on chromosome 13q14.3 and existing in two transcript isoforms [30], is expressed in lymphoid, hematopoietic, and brain tissues [31, 32]. Previous research has shown that ITM2B promotes malignant progression in esophageal squamous cell carcinoma via interaction with miR-196a-5p [33]. The roles of ITM2B and SOGA1 in HCC remain poorly defined. Our GSEA linked elevated SOGA1 to activated cell cycle and DNA replication pathways, with upregulation in high-grade tumors implying a proliferative role. Single-cell sequencing showed ITM2B’s broad distribution across malignant and stromal cells. Both genes correlated with tumor stemness and specific infiltration patterns of immunosuppressive M0/M2 macrophages and regulatory T cells. These results propose that ITM2B and SOGA1 might influence HCC progression through stemness and immune microenvironment regulation, suggesting new research avenues and experimental targets. Nevertheless, these inferences require further validation through subsequent functional experiments and analyses of clinical samples.

Corticotropin-releasing hormone-binding protein (CRHBP) has been shown to inhibit HCC cell cycle progression from G2 to M phase, thereby suppressing proliferation and promoting apoptosis; it also exerts anti-tumor effects by antagonizing the pro-angiogenic activity of CRH, reducing tumor microvessel density [34]. These reports are consistent with our findings of downregulated CRHBP expression and its association with favorable prognosis in HCC. C-type lectin domain family 3 member B (CLEC3B) has been found to be downregulated in exosomes and to promote HCC metastasis and angiogenesis through the AMPK/VEGF signaling pathway [35]. Angiogenesis is a critical process in tumor growth and metastasis, particularly in highly vascularized tumors such as HCC [36]. Our results not only confirm the downregulation of CLEC3B in HCC but also highlight its clinical relevance through prognostic modeling. The TEK (TEK receptor tyrosine kinase) gene encodes a receptor tyrosine kinase of the Tie family, predominantly expressed in endothelial cells [37]. A recent study suggests that TEK activation may remodel the tumor immune microenvironment (TIME) by polarizing tumor-associated macrophages (TAMs) toward an M1 phenotype and reducing regulatory T cell (Treg) infiltration [38]. Furthermore, TEK can mediate mast cell activation and enhance their adhesion to VCAM-1, thereby promoting immune responses [39]. These insights provide a new immunological perspective on the role of TEK in HCC. Interleukin-33 (IL-33), located on chromosome 9, is widely expressed in epithelial and endothelial cells [40]. A recent study by Wang et al. [41] revealed that nuclear expression of IL-33 decreases during hepatocarcinogenesis and is higher in peri-tumoral tissue than in tumor centers, suggesting that nuclear IL-33 may possess transcriptional regulatory functions. However, the precise mechanisms underlying IL-33’s role in HCC pathogenesis warrant further investigation. C-X-C motif chemokine ligand 12 (CXCL12), located on chromosome 10, is a key inflammatory mediator [42] that contributes to tumor progression by recruiting hematopoietic progenitor cells, stem cells, endothelial cells, and leukocytes, and by promoting angiogenesis [42]. Neuron-derived neurotrophic factor (NENF) is hypomethylated and activated in HCC tissues and cell lines, where it facilitates HCC development through MAPK and AKT signaling pathways [43]. Moreover, NENF is overexpressed in several cancers, including breast cancer, and promotes cell invasion and metastasis [44]. Our MR analysis further substantiates NENF as a genetic risk factor for HCC, providing strong evidence for a causal relationship. Finally, suppressor of glucose by autophagy 1 (SOGA1), a downstream target of METTL16, has been implicated in regulating glycolysis and colorectal carcinogenesis [45]. Although SOGA1 has been proposed as a prognostic marker in breast cancer [45], its functional role in HCC remains unclear. Our finding that SOGA1 is upregulated in advanced HCC suggests it may contribute to HCC malignant progression.

The prognostic model based on CLEC3B, TEK, and NENF demonstrated promising diagnostic accuracy (AUC = 0.738) in the primary cohort, suggesting potential utility as an adjunct to AFP. This finding was supported by external validation, where the model achieved AUCs of 0.764, 0.676, and 0.604 for predicting 1-, 3-, and 5-year OS, respectively. The consistent performance across cohorts indicates robust model generalizability and mitigates concerns of overfitting. However, it should be noted that this model was derived from transcriptomic data, whereas AFP is measured at the protein level in clinical practice. Given the limited concordance between gene transcription and protein expression (approximately 30%–50%) [46, 47], the present results cannot be directly compared with the clinical performance of AFP, and the translational value of this model requires further validation at the protein level. Nevertheless, our subsequent validation with tissue microarrays confirmed that the protein expression of the final candidate biomarker, NENF, is elevated in HCC and is associated with poor patient prognosis, thereby verifying the reliability of our initial transcriptome analysis.

From the perspective of the tumor immune microenvironment, our correlation analysis revealed that AFP and NENF expression was positively correlated with immunosuppressive Tregs, whereas CRHBP, CLEC3B, TEK, and IL33 showed negative correlations with Tregs. Under physiological conditions, cytotoxic CD8⁺T cells, CD4⁺ T cells, and NK cells collectively maintain immune surveillance. However, the HCC microenvironment is enriched with myeloid-derived suppressor cells (MDSCs), Tregs, and TAMs, fostering an immunosuppressive niche that enables tumor cells to evade immune attack [48–53]. Tregs suppress anti-tumor immunity through multiple mechanisms: directly inhibiting the activation and differentiation of CD4 and CD8⁺ T cells [48]; impairing the cytotoxicity of CD8⁺ T cells by reducing the release of granzymes and perforin [52]; competing with effector T cells for IL-2 consumption [52]; inducing dendritic cell (DC) tolerance via CTLA-4 or LAG3 [49, 51]; and secreting immunosuppressive factors such as TGF-β, IL-10, and IL-35, which subsequently inhibit the functions of CTLs and NK cells and compromise the antigen-presenting capacity of DCs, ultimately leading to immune escape [50, 53]. The significant correlations between the biomarkers we identified and these key immune cell subsets suggest that they may not only serve as diagnostic tools but also as indicators of the immune microenvironment status in HCC, potentially offering new targets for immunotherapy.

Despite employing a systematic bioinformatics approach that identified promising novel diagnostic and prognostic biomarkers for HCC, this study has several limitations. First, our model was derived from transcriptomic data, while clinical AFP is measured at the protein level. Given the known discordance between mRNA and protein expression [46, 47], its translational relevance remains limited without further protein-level validation, such as our preliminary tissue microarray analysis for NENF. Second, the findings are primarily based on bioinformatic analyses and cross-database validation. Direct confirmation of candidate protein expression in blood and their diagnostic efficacy using local clinical samples or experimental assays such as ELISA (enzyme-linked immunosorbent assay) or Western blot is still lacking. Third, the biological functions and mechanistic roles of several candidate genes (e.g., NENF) in HCC pathogenesis remain unclear. Further functional studies, including in vitro and in vivo experiments involving gene knockout or overexpression, are needed to elucidate their effects on tumor stemness, the immune microenvironment, and relevant signaling pathways. Fourth, future studies should evaluate a multi-marker panel that combines these biomarkers with AFP and PIVKA-II to capture distinct pathways (e.g., TEK in angiogenesis, NENF in proliferation), which may improve sensitivity in early-stage or AFP-negative HCC. Fifth, underlying chronic liver diseases such as hepatitis or cirrhosis can alter circulating protein levels and confound the specificity of these biomarkers for HCC detection. Future studies must therefore validate their performance in cohorts that include the full spectrum of liver disease, from healthy controls to HCC. Finally, our model was developed on mixed-etiology data and lacks validation in key etiologic subgroups such as HBV-, HCV-, and non-viral HCC. Given the known molecular heterogeneity among these subtypes, future work must include stratified validation in prospective, etiology-defined cohorts to confirm generalizability.

Conclusions

This study establishes a novel circulating protein biomarker panel for HCC, derived from stemness and secretome analysis, that diagnostically rivals or surpasses AFP. The panel’s clinical utility is underscored by multi-omics validation and a robust prognostic model, with NENF specifically implicated in tumor progression and survival. These results offer a direct path toward developing combined diagnostic assays and targeted therapies, enhancing personalized management of HCC.

Authors’ contributions

Cai Gao, and Jingxian Han designed the study; Cai Gao, Xuebing Li, Xinwei Liu, Sha Yan, and Xiaodan Ran conducted the analysis of pooled data; Cai Gao wrote the manuscript; Xuebing Li helped to draft the manuscript; Jingxian Han proofread, revised and final approved the manuscript; all authors have approved the version to be published.

Funding

None.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.