Abstract
Background
Cancer stem cells (CSCs) drive drug resistance, recurrence, and heterogeneity in hepatocellular carcinoma (HCC). Although individual stemness-related genes (SRGs) have been explored, comprehensive SRG-based signatures for prognosis and immunotherapy prediction remain underdeveloped.
Materials and methods
Using an integrated machine learning framework comprising 101 algorithms, we screened SRGs by comparing high- versus low-stemness HCC samples from TCGA-LIHC cohort. The optimal algorithm, StepCox[backward]+Enet[alpha = 0.9], was selected based on the highest average C-index across four independent datasets. A risk stratification system was constructed and validated for survival prediction, immune microenvironment characterization, and immunotherapy response evaluation using TIDE, IPS, TMB analyses, and external ICB therapy cohorts. Functional experiments were performed to validate the role of NCAPG in HCC.
Results
A 10-gene SRG signature (CHAF1A, NCAPG, KIF22, KIF4A, PBK, ANLN, RFC4, MCM2, TRAF2, and NSMCE2) was established. Low-risk patients exhibited significantly favorable prognosis across all cohorts (P < 0.05). The risk score outperformed conventional clinical parameters (average C-index = 0.8015) and 31 published prognostic models. Low-risk tumors displayed an immune-“hot” phenotype with higher immune/stromal scores, elevated CD8 + T cell infiltration, and superior predicted ICB response. High-risk patients showed enrichment of oncogenic pathways and increased sensitivity to multiple chemotherapeutic agents. Knockdown of NCAPG suppressed HCC cell proliferation, migration, invasion, and sphere formation in vitro, and inhibited tumor growth in vivo, accompanied by reduced stemness marker expression.
Conclusions
This SRG-based signature provides robust prognostic stratification and immunotherapy response prediction for HCC. NCAPG represents a promising stemness-related therapeutic target.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1186/s12876-026-05195-5.
Keywords: Hepatocellular carcinoma, Cancer stemness, Machine learning, Non-SMC condensin I complex subunit G
Introduction
Currently, liver cancer remains a major type of cancer that severely endangers public health, with over 800,000 new cases, and over 750,000 cancer-associated deaths, reported annually [1]. HCC is the most common subtype of primary liver cancer (PLC), accounting for approximately 85% of cases of patients with PLC [2]. Although the incidence and mortality rates of HCC continue to rise, the epidemiological characteristics of HCC have changed in recent years. HCC arises in the context of chronic liver disease, including viral hepatitis (HBV/HCV), alcoholic liver disease, and metabolic dysfunction-associated steatotic liver disease (MASLD; previously termed non-alcoholic fatty liver disease, NAFLD). The incidence rates of hepatitis B virus (HBV)-induced HCC and aflatoxin B1-induced HCC are slowly decreasing in the areas traditionally associated with a high incidence of the disease, such as east Asia and south-eastern Asia, whereas the incidence of NAFLD-associated HCC and alcohol abuse-associated HCC are rapidly increasing globally, which poses challenges to the currently available HCC prevention strategies and associated health education policies [3, 4]. At present, the early diagnosis and treatment status of HCC are also giving widespread cause for concern. Several surveys have shown that approximately 30% of patients with HCC were diagnosed at an early stage, and less than half of these patients at the early stage had opportunities to undergo surgical treatment due to the high treatment costs and a lack of availability of professional equipment and technicists [5]. For patients with advanced disease, the option of multiple therapies may be available, including targeted therapy, immunotherapy and chemotherapy, which usually results in significantly prolonging their survival times. However, fewer than 30% of patients with advanced disease receive long-term planned anti-HCC treatment due to the associated high fees, severe adverse reactions and individual heterogeneity, which all contribute towards an unfavorable 5-year survival rate of < 25% in advanced patients [6, 7]. All the above epidemiological data demonstrate that, at present, the developments that have been made in the diagnosis, treatment and prognosis assessment of HCC do not meet the patients’ needs; therefore, further innovations in the associated fields of research are urgently required.
Cancer stem cells (CSCs) comprise a small subset of tumor cells that have the capability of self-renewal, with unlimited differentiation and tumor initiation potential [8]. Numerous studies have previously demonstrated that CSCs contribute to tumor growth, tumor recurrence, tumor metastasis, cell death resistance and drug resistance [9–11]. The presence of CSCs also leads to a reprogramming of the tumor immune microenvironment (TIME), cancer metabolism and cancer epigenetic regulation, thereby providing a suitable cell ecosystem for cancer cell survival [11, 12]. Therefore, the eradication of CSCs is considered to be a promising strategy for the treatment of various types of cancer, including HCC [13]. A series of classical stemness-related gene (SRG) markers, including cluster of differentiation 44 (CD44), cluster of differentiation 133 (CD133), cluster of differentiation 90 (CD90), epithelial cell adhesion molecule (EPCAM), octamer-binding transcription factor 4 (OCT4), sex determining region Y-box 2 (SOX2) and Nanog homeobox (NANOG), and stemness-related pathways, such as the Notch, transforming growth factor-beta (TGF-β), Hedgehog and Wnt/β-catenin pathways, have been identified and widely validated in both fundamental and clinical research studies [9, 14]. On the basis of the abovementioned findings, research groups have developed CSC-related marker or pathway inhibitors or have designed chimeric antigen receptor (CAR)-engineered immune cells to target CSC surface markers, such as CD133, CD44 and EPCAM, which operate through exerting anti-CSC activity to counter cancer occurrence and progression [15, 16]. However, the practical applications of these therapies remain limited currently. In patients with HCC, unsatisfactory progress in this regard was found to be closely associated both with the similarity of genetic markers comparing between CSCs and normal stem cells, and the heterogeneity of CSC markers among individuals [17]. Therefore, there is an urgent need to systematically identify novel SRGs, and to develop novel stemness-related biomarkers for individualized therapies for patients with HCC.
In recent years, the application of machine learning in cancer diagnosis, treatment guidance and prognosis assessment has garnered a lot of attention. To comprehensively identify novel disease markers and disease-associated molecular features based on machine learning strategies is a vital goal of this research topic, and this is considered to be potentially an effective approach for achieving personalized cancer care. For example, Zhou et al. [18] developed the “BAMBI” tool to select cancer RNA biomarkers in relatively small, yet high-dimensional transcriptomics datasets, which optimized the procedures of biologically informed statistical methods and feature selection to enhance the selection accuracy. In addition, Xia et al. [19] developed the “ieGENES” method for the identification of differentially expressed genes (DEGs) in cancer studies, and they also proposed a parsimonious kernel machine regularization (PKMR) model to improve cancer classification accuracy. Furthermore, Hu et al. [20] established an online machine learning-based platform for molecular diagnosis, which integrated multiple machine learning algorithms to analyze both bulk RNA sequencing (RNA-seq) data and single-cell RNA sequencing (scRNA-seq) data, enabling rapid, automatic and accurate diagnosis of the molecular subtypes of B-acute lymphoblastic leukemia. Considered together, the above studies have demonstrated researchers’ enthusiasm for utilizing machine learning to assist in the identification of cancer markers and associated molecular characteristics, thereby improving the diagnosis, treatment and prognosis assessment of cancer.
Gene signatures are unique sets of genes whose expression patterns are characteristic of a particular biological state, condition or phenotype, and the applications of gene signatures in cancer include disease diagnoses, prognostic evaluations, predictions of responses to adjuvant therapy, and the identification of therapeutic targets [21, 22]. A previous study by Bareche et al. [23] demonstrated that the construction of a gene signature based on a machine learning strategy could significantly increase the model interpretability and prediction accuracy. Focusing on SRGs, machine-learning based gene signatures have been constructed in several types of human cancer, including ovarian cancer, lung adenocarcinoma and breast cancer, and they exhibited relatively good predictive power with respect to prognostic outcomes and the immunotherapy response [24–26]. As far as HCC is concerned, however, relevant studies remain scarce. Moreover, the majority of the previous gene signatures were established according to pre-set algorithms, and these algorithms might not have been the optimal ones in terms of achieving the best predictive results. Therefore, there is an urgent need to establish an SRG signature with high predictive efficiency with respect to the prognosis of patients with HCC and their immunotherapy responses. The aim of the present study was to identify such an SRG signature, also offering a comprehensive analysis of the SRG signature for prognostic assessment and the immunotherapy response prediction of HCC.
Materials and methods
Data resources and collection
In the present study, The Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) dataset was enrolled from TCGA database (https://www.cancer.gov/ccg/research/genome-sequencing/tcga) for SRG selection, prognostic model construction, immune microenvironment analysis, immune checkpoint inhibitor (ICI) therapy response prediction, drug sensitivity analysis and gene set enrichment analysis (GSEA). The dataset of Liver Cancer-RIKEN, Japan Project from International Cancer Genome Consortium (ICGC-LIRI-JP), obtained from the ICGC database (https://docs.icgc-argo.org/docs/data-access/icgc-25k-data), and the GSE14520 and GSE76427 datasets from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) were enrolled to validate the predictive performance of our prognostic model. Moreover, the IMvigor210 dataset [27] and GSE91061 dataset were also collected to validate the predictive performance of the risk stratification system on immune checkpoint blockade (ICB) treatment responses. In these cohorts, samples that had survival data or transcriptomic data missing were excluded from the current study.
Stemness index calculation and SRG selection
Malta et al. [28] designed and established a machine learning tool to quantify the stemness based on the one-class logistic regression (OCLR) algorithm. The stemness index of each sample was calculated according to gene expression data and DNA methylation data. Briefly, two transcriptomic datasets from pluripotent stem cells were selected as the training groups to recognize the stemness features (based on a gene expression profile of 11,774 genes) and the DNA methylation features (based on 151 differentially methylated CpG sites) [29, 30]. The detailed workflow of Malta et al. [28] is described in the following website: https://bioinformaticsfmrp.github.io/PanCanStem_Web/. In the present study, the above method was employed to calculate the stemness index. The stemness index of each sample in TCGA-LIHC dataset was calculated according to the Spearman correlation (RNA expression data) or linear model (DNA methylation data) operators. The patients in TCGA-LIHC dataset were divided into a high-stemness group and a low-stemness group according to the median cutoff. The DEGs between the two groups were screened with |logFC| >4 and a P-value < 0.001. These DEGs were considered as the SRGs.
Prognostic model and nomogram construction
The DEGs comparing between HCC samples and normal liver samples in TCGA-LIHC cohort were identified by the “limma” package, with the cutoff of |logFC| >2 and P-value < 0.001. The abovementioned SRGs and DEGs between HCC samples and normal liver samples were compared, and the intersection was selected to perform further survival analysis. The SRGs with differential expression and prognostic differences were considered as the candidates for establishment of the prognostic model. An integrated machine learning selection framework containing 101 independent or combined algorithms was enrolled for optimal algorithm selection. In these algorithms, random survival forest (RSF), Enet, Lasso, Ridge, stepwise Cox, CoxBoost, plsRcox, SuperPC, gradient boosting machine (GBM) and survival-support vector machine (SVM) were considered as the basic algorithms. The statistical features concerning these algorithms have been described in detail in a previous study [31]. The combined algorithms contained at least one of the basic algorithms, including RSF, LASSO, CoxBoost or StepCox, owing to their feature selection capacity. First, the prognostic models were constructed from each of the respective 101 algorithms using TCGA-LIHC dataset. Subsequently, the 101 prognostic models were validated in the GSE76427, GSE14520 and ICGC-LIRI-JP datasets. The C-index of each algorithm in the corresponding cohort was then calculated. The algorithm with the highest average C-index among the four datasets was considered as the optimal algorithm for construction of the prognostic model. The risk score of each patient was then calculated according to expression of the genes and their coefficients. The patients were divided into high-risk and low-risk groups, as determined by the median value of the risk score, and the survival differences between the two groups were compared using the “survival” package. The “TimeROC” package was subsequently employed to perform time-dependent receiver operating characteristic (ROC) analysis. C-indices of clinical parameters were also calculated using the “rms” package. In addition, the “survival” package was used to identify the correlations between clinical parameters and the risk scores, as analyzed by univariate and multivariate Cox regression analyses.
The nomogram containing details of the age, sex, clinical stage and risk score of the patient, and corresponding calibration plots were generated using the “rms” package. In addition, C-index comparisons between our prognostic model and prognostic models from other studies were made using the “CompareC” package.
Immune microenvironment analysis
Estimation of stromal and immune cells in malignant tumor tissues using expression data (the ESTIMATE method) was utilized to assess both tumor purity and the overall infiltration levels of stromal and immune cells in cancer tissues [32]. The expression levels of human leukocyte antigen (HLA)-associated genes and common immune checkpoints were visualized using the “ggpubr” package. Based on the single sample GSEA (ssGSEA) method, the enrichment scores of classical immune functions and common immune cells were visualized using the “GSVA” package, and associations between the risk score and immune cell infiltration were subsequently analyzed using Pearson’s correlation analysis.
ICB therapy response prediction
Some of the most commonly used ICB treatment response predictive tools, including the tumor immune dysfunction and exclusion (TIDE) score, immunophenoscore (IPS) and the tumor mutation burden (TMB) score, were also enrolled [33–35]. Based on the transcriptomic data from cancer tissues, the TIDE score predicts the ICB therapy response of patients via modeling two fundamental mechanisms of tumor immune escape; namely, the inhibition of T cell infiltration, and the dysfunction of T cells. Moreover, the T cell dysfunction score, T cell exclusion score and immune escape score all belong to TIDE scoring system [33]. The IPS consists of four domains, containing major histocompatibility complex (MHC) molecules, effector cells, suppressor cells and immune checkpoints [34], and a higher IPS score is correlated with increased immunogenicity. Furthermore, the TMB score represents a quantitative biomarker that reflects the total number of mutations carried by tumor cells. Tumor cells with high TMB levels have a relatively high level of neoantigens, and are considered to assist the immune system in terms of recognizing tumor cells and stimulating the proliferation of anti-tumor T cells and anti-tumor responses [35]. Considered overall, a higher IPS or TMB score is associated with improved ICB therapy responses, whereas a higher TIDE score is associated with a poor ICB therapy response. Open access datasets containing survival data and transcriptomic data from patients with HCC receiving ICB therapy are lacking at present. The GSE91061 dataset [comprising clinical data and transcriptomic data from patients with advanced melanoma receiving anti-programmed cell death protein 1 (anti-PD-1) therapy] and the IMvigor210 dataset (comprising clinical data and transcriptomic data from patients with metastatic urothelial carcinoma receiving anti-PD-L1 therapy) were employed to externally validate the various ICB treatment responses between the high-risk group and the low-risk group. Study designs similar to this may be found in the articles published by Wei et al. [36] and Xiong et al. [37]. The treatment response of the patients in these two cohorts was assessed according to RECIST 1.1 criteria [38].
Drug sensitivity analysis and GSEA
The drug sensitivity data of individual patients in TCGA-LIHC cohort were downloaded from the Genomics of Drug Sensitivity in Cancer (GDSC) database (https://www.cancerrxgene.org/). The “oncoPredict” package was utilized for visualization of the overall distribution of IC50 values in the high-risk and low-risk groups. The enriched pathways and corresponding enrichment scores in both the high-risk and the low-risk groups were analyzed using the “GSVA” and the “GSEABase” package. Using these methods, several classical cancer-associated pathways were visualized in the present study.
Cell culture
A normal human liver epithelial cell line (THLE-2) and human HCC cell lines (HCCLM3, Huh-7, JHH-7, Mahlavu and MHCC97L) were purchased from Shanghai Zhongqiao Xinzhou Biotechnology Co., Ltd. THLE-2 cells were cultured in THLE-2 IMMOCELL® Special medium (purchased from Xiamen Yimo Biotechnology Co., Ltd.), whereas the HCCLM3, Huh-7, JHH-7, Mahlavu and MHCC97L cells were cultured in high-glucose Dulbecco’s Modified Eagle Medium (H-DMEM) (Wuhan Servicebio Technology Co., Ltd.) with 10% fetal bovine serum (cat. no. SA301.02.V; Cellmax) and 1% penicillin-streptomycin solution (Shanghai Yeasen Biotechnology Co., Ltd.). All the cell lines employed in the present study were incubated at 37˚C in an atmosphere containing 5% CO2.
RNA extraction and reverse transcription-quantitative polymerase chain reaction (RT-qPCR) assay
Total RNA from the cells was extracted using a Total RNA Extraction Kit (Beijing Sevenbio Biotechnology Co., Ltd.), following the manufacturer’s protocol. The purity and concentration of RNA samples were assessed using a UV spectrophotometer (Thermo Fisher Scientific, Inc.). RT Master Mix for qPCR Ⅱ (MedChemExpress) was purchased for complementary DNA (cDNA) synthesis and genomic DNA (gDNA) removal. RT-qPCR assay was subsequently performed using a CFX96 Touch Real-Time PCR Detection System (Bio-Rad Laboratories, Inc.) and SYBR Green qPCR Master Mix (MedChemExpress). The 2-ΔΔCq method was utilized for data quantification and data standardization. Glyceraldehyde − 3-phosphate dehydrogenase (GAPDH) was set as the housekeeping gene. The sequences of the primers used for RT-qPCR assay are shown listed in Supplementary Table 1.
Protein extraction and western blot analysis
Total protein for cells was extracted using RIPA lysis buffer (Beijing Solarbio Science & Technology Co., Ltd.), and the concentrations of the protein samples were evaluated using a BCA protein assay kit (Vazyme Biotech Co., Ltd.). SDS-PAGE (10% gels) was utilized for protein electrophoresis. The proteins in the gels were subsequently transferred to 0.22-µm polyvinylidene fluoride (PVDF) membranes (Shanghai Yeasen Biotechnology Co., Ltd.). The membranes were then blocked with 5% BSA blocking buffer (Beijing Solarbio Science & Technology Co., Ltd.) for 2 h, prior to an incubation with primary antibodies at 4˚C overnight. The membranes were subsequently washed 3 times with Tris-buffered saline/0.1% Tween-20 (TBST) solution, before being incubated with the secondary antibodies for 1 h at room temperature. The targeted proteins on the membranes were detected using an ECL detection kit (Shanghai Yeasen Biotechnology Co., Ltd.) and a ChemiDec MP platform (Bio-Rad Laboratories, Inc.). Western blot image analysis and quantification were performed using ImageLab software (Bio-Rad Laboratories, Inc.). The primary antibodies used for the western blotting experiments were as follows: Anti-NCAPG (1:8,000 dilution, cat. no. 68082-1-Ig; Proteintech Group, Inc.); anti-Beta-tubulin (1:20,000, cat. no. 66240-1-Ig; Proteintech Group, Inc.); anti-OCT4 (1:500, cat. no. R381335; Zen-Bioscience); anti-CD44 (1:1,000, cat. no. 200840; Zen-Bioscience) and anti-NANOG (1:500, cat. no. R381167; Zen-Bioscience).
Lentivirus infection
To obtain the stable NCAPG knockdown HCC cells, lentivirus vectors targeting NCAPG and negative control vector were designed and generated by Shanghai Focus Bioscience Co., Ltd. According to the multiplicity of infection (MOI) value, HCC cells were infected with lentivirus. The stable cell lines were subsequently selected using 2 µg/ml puromycin dihydrochloride (Wuhan Servicebio Technology Co., Ltd.). The efficiency of NCAPG knockdown was then tested using RT-qPCR and western blot assays. The target sequences of the shRNAs are shown in Supplementary Table 1. Lentiviral transduction was performed at a multiplicity of infection (MOI) of 5 for HCCLM3 cells and 8 for Huh-7 cells. These MOI values were optimized in pilot experiments to achieve > 90% infection efficiency while minimizing cytotoxicity. Polybrene (8 µg/mL) was added to enhance transduction efficiency. After 24 h of incubation, the medium was replaced with fresh medium containing puromycin (2 µg/mL) to select stably transduced cells. Knockdown efficiency was confirmed by qRT-PCR and western blotting after 7 days of selection.
CCK-8 assay
Huh-7 or HCCLM3 cells were seeded at a density of 3,000 cells/well into 96-well plates, and then the culture was allowed to proceed. At the time points of 0, 24, 48 and 72 h, the medium was changed to 100 µl complete medium containing 10 µL CCK-8 reagent (Abbkine Scientific Co., Ltd.). Subsequently, the plates were incubated at 37˚C for 60 min in the dark. Finally, the OD value of each well at 450 nm was measured using a microplate reader (Bio-Rad Laboratories, Inc.).
5-Ethynyl-2’-deoxyuridine (EdU) assay
EdU assays were performed using an EdU cell proliferation image kit (Abbkine Scientific Co., Ltd.). Briefly, Huh-7 or HCCLM3 cells were seeded at a density of 15,000 cells/well into 96-well plates, and then the culture was allowed to proceed. The cells were subsequently incubated with 10 µM EdU solution for 2 h. Following cell fixation and treatment with penetration enhancer, the cells were dyed using Click-iT reaction mix (AbFluor 488Azide; Abbkine Scientific Co., Ltd.) for 60 min. In addition, cell nuclei were also labeled by treatment with 4’,6-diamidino-2-phenylindole (DAPI) solution. Finally, the fluorescence intensities of the cell nuclei were observed and recorded under a fluorescence microscope (Nikon Corporation).
Transwell assay
To assess the migratory capability of the HCC cells, a total of 15,000 Huh-7 or HCCLM3 cells in 200 µl serum-free medium were cultured in the upper chambers of Transwell inserts, whereas 800 µl complete medium was added into the lower chambers of the Transwell inserts. After a 48 h incubation, the inserts were fixed with 4% paraformaldehyde solution (Biosharp Life Sciences) for 30 min, before being dyed with crystal violet solution (Beijing Solarbio Science & Technology Co., Ltd.) for 30 min. The remaining cells in the upper chambers were removed. The migrated cells were observed, and their numbers were recorded under an optical microscope (Leica Microsystems GmbH).
For the invasion assays, the upper chambers of the inserts were pre-laid with Matrigel™ (Corning, Inc.); otherwise, the other experimental processes were identical with those of the migration assay.
Sphere formation assay
A total of 5,000 Huh-7 or HCCLM3 cells were cultured with HyClone™ serum-free DMEM/F12 (Cytiva) containing 20 ng/ml epidermal growth factor (Novoprotein Scientific Inc.), 10 ng/ml fibroblast growth factor (Novoprotein Scientific Inc.), 2% B27 supplement (Beyotime Institute of Biotechnology) and 1% N2 supplement (Beyotime Institute of Biotechnology) in 6-well ultra-low-attachment plates (Beyotime Institute of Biotechnology). After cultivating the cells for 7 days, the numbers and sizes of tumor spheres were observed and recorded under an optical microscope (Olympus Corporation).
Animal experiments
Four-week-old male BALB/c nude mice (body weight: 14–16 g) were purchased from Hangzhou Hang Si Biotechnology Co., Ltd. A total of 10 mice were randomly allocated into the shNC group and the sh-NCAPG group (5 mice per group). They were housed in a specific pathogen-free room with a 12-h light-dark cycle, controllable light, temperature (22 ± 1℃) and humidity (55 ± 10%), and adequate water and food. And, no more than 5 mice lived in one mouse cage. A 7-day acclimation period was accepted for the mice prior to the subcutaneous injection of the cells. Following the acclimation period, 100 µl cold phosphate-buffered saline (PBS) containing 1 × 106 HCCLM3 cells were injected subcutaneously into the right flanks of the mice. The mice in the shNC group were injected subcutaneously with 1 × 106 negative control HCCLM3 cells, while the mice in the sh-NCAPG group were injected subcutaneously with 1 × 106 stable NCAPG knockdown HCCLM3 cells. The experimental end point of the current study was set as 21 days after the subcutaneous injection, and the overall condition, weight, tumor diameter and approximate tumor volume of each mouse were recorded twice one week (every Monday and Thursday). At the endpoint, the mice were euthanized by carbon dioxide (CO2 displacement rate, 50% cage volume/min). Mice were continuously exposed to CO2 for 6 minutes. After turning off the CO2 supply, continue observation for 1–2 minutes; mice were determined to be dead when respiratory and cardiac stop occurred along with the loss of corneal reflex. The subcutaneous tumors were entirely removed, and subsequently photographed. A vernier caliper was used to measure the length and width of the tumors, and the tumor volume was calculated using the following formula: Tumor volume (mm3) = 0.5 x length x (width)2. Moreover, the humane endpoints of the present study were also established: Weight loss, > 20%; tumor diameter, >2 cm; tumor volume, > 1500mm3; tumor with ulcer necrosis; and overall poor condition. In the present study, we confirmed that no mouse reached the humane endpoint.
Statistical analysis
R software (version 4.4.2) and GraphPad Prism software (version 10.1.2; Dotmatics) were utilized for statistical analyses and visualization of the data in the present study. The Cox regression test was employed for survival analysis. Correlation analyses were assessed using the Spearman correlation test. Student’s t-test or the Wilcoxon test was used for comparisons between two groups, whereas one-way ANOVA analysis was used for comparisons among more than two groups. A P-value < 0.05 was considered to indicate a statistically significant difference, and the statistics in the figures are shown by the following: NS, no statistical significance; *, P-value < 0.05; **, P-value < 0.01; and ***, P-value < 0.001.
Results
Identification of the candidates for the stemness-related prognostic model
According to the stemness index calculation framework proposed by Malta et al. [28], the stemness index of the individual samples in TCGA-LIHC dataset were calculated. The samples were divided into a high-stemness group and a low-stemness group based on the median stemness index. Between the two groups, a total of 1,050 genes were identified as SRGs. The gene list of the SRGs is shown in Supplementary Tables 2, and the heat map of partial SRGs is shown in (Fig. 1A). In comparing the survival differences between the high-stemness group and the low-stemness group, a significantly lower overall survival rate was observed in the high-stemness group (Fig. 1B). A total of 649 genes were identified as DEGs comparing between HCC tissues and the normal liver tissues in TCGA-LIHC cohort (Supplementary Table 3). Of these 649 DEGs, a total of 188 genes were also featured in the list of SRGs (Fig. 1C). The 188 differentially expressed SRGs were further investigated with regard to the survival analyses. These results demonstrated that 74 differentially expressed SRGs were associated with the prognosis of patients with HCC (Fig. 1D); these 74 SRGs were therefore selected as the candidate genes to build up the prognostic model.
Fig. 1.

Identification of the candidates for establishment of the stemness-related prognostic model. A A heatmap revealing the expression levels of partial SRGs identified by the OCLR algorithm in the high-stemness and low-stemness groups. B The survival differences between the low-stemness and high-stemness groups are shown. C The intersection between SRGs and DEGs is shown. D The SRGs with differential expression between normal tissues and HCC tissues, and the prognostic significance between the high- and low-expression groups in the intersection. SRG, stemness-related gene; OCLR, one-class logistic regression; DEG, differentially expressed gene
Stemness-related prognostic model construction and validation
Both the expression data and survival data of TCGA-LIHC dataset were enrolled as the training group, whereas the ICGC, GSE14520 and GSE76427 cohorts were enrolled as the validation groups. Based on the integrated machine learning algorithm selection tool containing 101 approaches, the average C-index of the corresponding algorithm among these four datasets was calculated. The results revealed that StepCox[backward]+Enet[alpha = 0.9] was the optimal algorithm for prognostic model construction, with the highest calculated C-index of 0.8015 (Fig. 2A).
Fig. 2.

Construction and validation of the stemness-related prognostic model based on the integrated machine learning algorithm selection framework. A The individual C-indexes and average C-indexes among the TCGA-LIHC, ICGC-LIRI-JP, GSE76427 and GSE14520 datasets corresponding to each algorithm are shown. B Survival analysis between the high-risk and low-risk patients in the TCGA-LIHC, ICGC-LIRI-JP, GSE76427 and GSE14520 datasets. C Time-dependent ROC analysis of 1-year, 3-year and 5-year survival rates was performed to evaluate the predictive efficacy of the present prognostic model. D C-index comparisons in the range from 1-year survival to 5-year survival among various clinical parameters of the patients, namely the risk score, age, sex and clinical stage. E Univariate and multivariate Cox regression analyses were performed to detect the independent risk factors affecting the prognosis of HCC patients among the risk score, age, gender and clinical stage. TCGA-LIHC, The Cancer Genome Atlas Liver Hepatocellular Carcinoma; ICGC-LIRI-JP, The Dataset of Liver Cancer-RIKEN, Japan Project from International Cancer Genome Consortium; ROC, receiver operating characteristic; HCC, hepatocellular carcinoma
To construct a robust prognostic signature from the 74 stemness-related genes (SRGs), we employed a two-step machine learning-based feature selection approach. Step 1-Stepwise Cox regression with backward selection (StepCox[backward]): To further refine the gene set, we applied stepwise Cox regression with backward elimination (StepCox[backward]) based on Akaike Information Criterion (AIC). The backward selection process iteratively removed genes with the least contribution to model fit, using AIC as the stopping criterion. This step reduced the gene list from 74 to 25 genes. Step 2-Elastic Net regularization (Enet[alpha = 0.9]): Finally, we applied Elastic Net (Enet) regularization with alpha = 0.9, which combines L1 (Lasso) and L2 (Ridge) penalties with a strong L1 weighting, to perform both gene selection and coefficient shrinkage simultaneously. Ten-fold cross-validation was used to select the optimal lambda (λ) value that minimized the partial likelihood deviance. This final step yielded the optimal 10-gene prognostic signature.
The risk score was calculated according to the expression levels of chromatin assembly factor 1 subunit A (CHAF1A), NCAPG, kinesin family member 22 (KIF22), kinesin family member 4 A (KIF4A), PDZ binding kinase (PBK), anillin (ANLN), replication factor C subunit 4 (RFC4), minichromosome maintenance complex component 2 (MCM2), TNF receptor associated factor 2 (TRAF2) and non-structural maintenance of chromosomes element 2 homolog (NSMCE2). The formula to calculate the risk score was as follows: Risk score = (0.038916629) x TRAF2 expression + (0.024665277) x NSMCE2 expression + (0.234752661) x MCM2 expression + (0.044600273) x RFC4 expression + (0.07676431) x PBK expression + (0.145310073) x ANLN expression + (0.126530395) x NCAPG expression + (0.010769802) x KIF4A expression + (0.048747861) x KIF22 expression + (-0.175383603) x CHAF1A expression. Using TCGA-LIHC cohort, the expression levels of the 10 SRGs in normal liver tissues and HCC tissues are visualized in (Supplementary Fig. 1A. According to the median risk score, the patients were divided into high-risk and low-risk groups. In the four datasets, the high-risk score group showed a significantly worse prognosis compared with that in the low-risk score group (Fig. 2B), which demonstrated a good distinguishing ability for survival risk of patients with HCC. Time-dependent ROC analysis was also performed, which revealed the 1-year, 3-year and 5-year area under the curve (AUC) values to be mostly > 0.75 in these four datasets (Fig. 2C). C-index comparisons between the risk score and other clinical parameters, including age, sex and clinical stage, were also made, and the risk score exhibited the highest C-index from the 1-year survival to the 5-year survival predictions in these four datasets (Fig. 2D). The results of the ROC analysis and C-index comparisons all revealed that our prognostic model had a comparatively good predictive performance for the prognosis of patients with HCC. Moreover, univariate and multivariate Cox regression analyses were also performed to identify the risk factors that affected the prognosis of the HCC patients, and these analyses revealed that the clinical stage and the risk score were the independent risk factors that influenced the prognosis of the HCC patients (Fig. 2E). Given that cancer stem cells are known to drive tumor recurrence, we next investigated whether our 10-gene stemness-related risk score was associated with recurrence patterns in HCC patients. We obtained recurrence-free survival (RFS) data from four independent HCC cohorts. Patients stratified as high-risk based on our 10-gene signature exhibited significantly shorter RFS compared with low-risk patients across all four cohorts (Supplementary Table 4).
Model comparison and nomogram construction
To further verify the predictive performance of our prognostic model, a total of 31 prognostic models established in previous publications were recruited in order to perform model comparison analysis. To ensure the objectivity of the model comparison, various themes of the enrolled models were taken into consideration, including immune cells, programmed cell death, epigenetics, cell metabolism, and the specific gene family (Supplementary Table 5). Moreover, the processes of target gene screening, risk score calculation and subsequent risk stratification in the other 31 prognostic models were found to be consistent with the original publications. Using the data from TCGA-LIHC cohort, the C-index of each model was calculated, and these are visualized in (Fig. 3A). Our prognostic model showed the highest C-index of 0.804 among these models, which also indicated that our prognostic model had a good predictive efficacy for the prognosis of patients with HCC. In order to facilitate clinical utilization, a nomogram incorporating the risk score, sex, age and clinical stage of the patients based on TCGA-LIHC cohort was also constructed (Fig. 3B). The 1-year, 3-year and 5-year survival rates of individual patients was assessed by the total scores, which were calculated by adding up the scores corresponding to each parameter. A higher total score was correlated with a lower survival rate at the 1-year, 3-year and 5-year stages. A good fit comparing between the current model and the ideal model was also determined through the calibration plots corresponding to the 1-year, 3-year and 5-year survival rates (Fig. 3C).
Fig. 3.

External model comparisons and stemness-related nomogram establishment. A C-index comparisons between the prognostic model in the present study and 31 other prognostic models with various themes are shown. B a stemness-related nomogram was constructed according to the risk score, age, clinical stage and sex. C The calibration plots of the nomogram corresponding to 1-year, 3-year and 5-year survival rates of the HCC patients are shown. HCC, hepatocellular carcinoma
Immune microenvironment landscape and ICB treatment response prediction between the high-risk and the low-risk cohorts
At present, ICB treatment is commonly employed as a strategy to reduce postoperative recurrence and to alleviate cancer progression in advanced patients. However, the therapeutic effects of ICB therapy are known to vary greatly among individuals. Recent studies have reported that the differences in ICB therapy response among individuals are closely associated with the individual TIME and corresponding immune function status [39, 40]. In order to expand the potential applications of our risk stratification system such that it could accommodate both immune microenvironment distinction and ICI therapy response prediction, the ESTIMATE tool was first employed to assess the overall distribution of stromal cells and immune cells. As shown in Fig. 4A, the ESTIMATE score, immune score and stromal score of high-risk patients were found to be significantly lower compared with those in low-risk patients. HLA class I genes and HLA class Ⅱ genes are known to be important regulators in anti-tumor immunity activation [41]. The expression levels of classical HLA class I genes, HLA class Ⅱ genes and common immune checkpoints were found to be significantly increased in low-risk patients (Fig. 4B and C). The enrichment scores of B cells, CD8+ T cells, dendritic cells (DCs), macrophages, mast cells, natural killer (NK) cells, Th2 cells and tumor infiltrating lymphocytes (TILs) were also shown to be elevated in low-risk patients (Fig. 4D). In terms of the common immune functions, the enrichment scores of antigen-presenting cell (APC) co-stimulation, cytolytic activity, HLA, pro-inflammation, para-inflammation, T cell co-stimulation and type Ⅰ interferon (IFN) response in the low-risk group were found to be significantly higher compared with those in the high-risk group (Fig. 4E). In addition, the associations between the risk score and immune cell infiltration were also analyzed with multiple algorithms, including the XCELL, TIMER, EPIC, CIBERSORT, CIBERSORT-ABS, MCPCOUNTER and QUANTISEQ algorithms. The correlation analyses with P-values < 0.05 are shown in (Fig. 4F). Significantly, in multiple algorithms, the risk score was found to be negatively correlated with the infiltration levels of the classical anti-tumor immune effector cells, including CD8+ T cells and DCs, whereas the risk score was positively correlated with the infiltration level of M2 macrophages; namely, a classical immunosuppressive cell type (Fig. 4G-I). Considered altogether, the evidence presented above suggest that the TIME in the low-risk group resembled the TIME of “hot” tumors as opposed to “cold” tumors, based on the “hot” and “cold” theory of tumors [42].
Fig. 4.

Immune microenvironment analysis comparing between high-risk patients and low-risk patients, and the association between immune cell infiltration and the risk score. A The ESTIMATE method was used to reveal the overall infiltration status of immune cells and stromal cells. B The mRNA levels of HLA-related markers in low-risk and high-risk patients are shown. C The mRNA levels of common immune checkpoints in low-risk and high-risk patients are shown. D The enrichment scores of immune cell types based on the ssGSEA method in low-risk and high-risk patients are shown. E The enrichment scores of common immune functions based on the ssGSEA method in low-risk and high-risk patients are shown. F The correlations between the risk score and the infiltration levels of the immune cell types from multiple algorithms are shown. G-I The correlations between the risk score and the infiltration levels of CD8+ T cells, dendritic cells and M2 macrophages in various algorithms are shown. *P < 0.05; **P < 0.01; ***P < 0.001. HLA, human leukocyte antigen; ssGSEA, single sample gene set enrichment analysis
Several digital predictive tools for the ICB response were also enrolled. Low-risk patients were found to have higher TMB scores, PD-1 + cytotoxic T-lymphocyte associated protein 4 (CTLA4)-IPS scores, CTLA4-IPS scores and PD1-IPS scores (Fig. 5A and B), but lower TIDE scores, T cell dysfunction scores, T cell exclusion scores and immune escape scores (Fig. 5C), which suggested that low-risk patients might be more likely to derive benefits from receiving ICB therapy. To externally validate the results through using the GSE91061 and IMvigor210 cohorts, the risk scores in responders [complete response (CR) / partial response (PR)] were found to be significantly lower compared with those in non-responders [stable disease (SD) / progressive disease (PD)]. The percentage of responders in the low-risk group was also significantly higher compared with that of the non-responders. Regarding the survival analysis, high-risk patients also showed a worse prognosis compared with low-risk patients (Fig. 5D and E).
Fig. 5.

The risk score as an indicator for ICB treatment response prediction in patients with HCC. A the TMB scores of high-risk and low-risk patients are shown. B The PD-1༆CTLA4-IPS, CTLA4-IPS and PD1-IPS scores in high-risk and low-risk patients are shown. C The TIDE score and its branch score, including T cell dysfunction score, T cell exclusion score and Immune escape score, in high-risk and low-risk patients. D and E The risk distribution of responders and non-responders, and the differences of responders’ percentages and the survival differences between high-risk and low-risk patients in the GSE91061 and IMvigor210 cohorts. *P < 0.05; ***P < 0.001. ICB, immune checkpoint blockade; TMB, tumor mutation burden; PD-1, programmed cell death protein 1; CTLA4, cytotoxic T-lymphocyte associated protein 4; IPS, immunophenoscore; TIDE, tumor immune dysfunction and exclusion
Drug sensitivity analysis and GSEA comparing between the high-risk and low-risk cohorts Chemotherapy and targeted therapy continue to be the treatment methods of choice that are utilized to prevent HCC recurrence and to prolong the survival time of patients with advanced disease. The IC50 data of HCC samples in TCGA-LIHC dataset was downloaded from the GDSC database. Due to the poor prognosis and poor outcomes of ICB treatment in high-risk patients, we were interested in which chemotherapy drugs and targeted drugs may have been more sensitive to high-risk patients. As shown in Fig. 6A, high-risk patients were more sensitive to lapatinib, nilotinib, oxaliplatin, sapitinib, erlotinib, docetaxel, crizotinib, afatinib and 5-fluorouracil, which indicated that, as far as the high-risk patients were concerned, these drugs might be more suitable for chemotherapy or targeted therapy. GSEA with a view towards investigating the classical cancer hallmarks was also performed in the two groups. Higher enrichment levels of IL6-JAK-STAT3 signaling, mTORC1 signaling, Notch signaling, PI3K-AKT-mTOR signaling, DNA repair, E2F targets, epithelial-mesenchymal transition (EMT) signaling, glycolysis, hypoxia and IL2-STAT5 signaling were detected in high-risk patients (Fig. 6B), which might partly account for why the high-risk patients were associated with both poor prognosis and low therapy benefits resulting from ICB treatment.
Fig. 6.

Drug sensitivity analysis and gene set enrichment analysis comparing between the high-risk and low-risk patients. A Comparisons of the IC50 values of lapatinib, nilotinib, oxaliplatin, sapitinib, erlotinib, docetaxel, crizotinib, afatinib and 5-fluorouracil between high-risk and low-risk patients were made. B The enrichment levels of cancer-associated hallmarks are shown, including IL6-JAK-STAT3 signaling, mTORC1 signaling, Notch signaling, PI3K-AKT-mTOR signaling, DNA repair, E2F targets, EMT signaling, glycolysis, hypoxia and IL2-STAT5 signaling, in high-risk patients and low-risk patients. EMT, epithelial-mesenchymal transition
Knockdown of NCAPG inhibits the proliferation, migration, invasion and stemness of HCC cells
Among the 10 genes comprising the prognostic signature, we selected NCAPG for detailed functional validation based on four predefined criteria. First, NCAPG exhibited the highest hazard ratio in multivariate Cox regression analysis, indicating the strongest independent prognostic value. Second, NCAPG showed the strongest correlation with the OCLR-derived stemness index, suggesting a close biological link to cancer stemness. Third, NCAPG demonstrated the most consistent differential expression across all eight independent validation cohorts, with an average log2 fold-change of 3.8. NCAPG is a component of non-SMC condensin I complex, which is closely associated with the maintenance of chromosome stability [43]. In the above analyses, NCAPG was identified as an SRG in HCC; however, no previous experimental evidence has demonstrated the effects of NCAPG on HCC stemness maintenance. Therefore, we performed a series of experiments to validate our hypothesis. First, the expression levels of NCAPG in a normal liver cell line and multiple HCC cell lines were compared, which indicated that NCAPG expression was markedly upregulated in HCC cells (Fig. 7A). The results of the EdU and CCK-8 assays revealed that NCAPG knockdown suppressed HCC cell proliferation (Fig. 7B-D). Furthermore, the results of the migration and invasion assays based on the Transwell system revealed that NCAPG knockdown could suppress HCC cell migration and invasion (Fig. 7E and F). Subsequently, the tumor sphere formation assay also illustrated that silencing NCAPG reduced the stemness of the HCC cells (Fig. 7G). NCAPG knockdown efficiency was confirmed via RT-qPCR and western blot analyses (Supplementary Fig. 1B-1E and Fig. 7H). To further detect the expression of classical stemness markers, NCAPG knockdown led to decreased protein expression levels of OCT4, CD44 and NANOG (Supplementary Fig. 1D, 1E and 7H). Finally, in vivo experiments showed that NCAPG knockdown could inhibit HCC growth in a subcutaneous xenograft mouse model (Fig. 7I).
Fig. 7.

NCAPG knockdown suppresses the proliferation, migration, invasion and stemness of HCC cells. A The mRNA and protein expression levels of NCAPG in a normal liver cell line and several HCC cell lines (see the Materials and methods section for further details) are shown. B and C EdU assay was performed to assess the proliferative capability of the cells in the control and NCAPG knockdown groups (scale bar = 50 μm). D CCK-8 assay was performed to assess the proliferative capability of the cells in the control and NCAPG knockdown groups. E and F Transwell assays were performed to detect the migratory and invasive capabilities of the cells in the control and NCAPG knockdown groups (scale bar = 100 μm). G Tumor sphere formation assay was performed to reflect the ability of stemness maintenance in the control and NCAPG knockdown groups (scale bar = 100 μm). H The protein expression levels of NCAPG, CD44, OCT4 and NANOG in the control and NCAPG knockdown groups are shown. I A subcutaneous xenograft mouse model was constructed to detect HCC growth in the control and NCAPG knockdown groups. *P < 0.05; **P < 0.01; ***P < 0.001. NCAPG, non-SMC condensin I complex subunit G; HCC, hepatocellular carcinoma; CD44, cluster of differentiation 44; OCT4, octamer-binding transcription factor 4; NANOG, Nanog homeobox
Discussion
In the present study, stemness-related gene markers according to the OCLR algorithm were identified, and a novel 10-gene stemness-related prognostic model with a good predictive performance was subsequently constructed to assess the survival risk of patients with HCC. A corresponding risk stratification system was also established, and the risk score was identified as an independent risk factor for the prognosis of HCC patients. Although our prognostic model appeared to serve as a predictive tool with high predictive accuracy and wide applicability, large-scale prospective studies or retrospective studies are still required to validate and improve the current model. The 10 SRGs comprised the genes CHAF1A, NCAPG, KIF22, KIF4A, PBK, ANLN, RFC4, MCM2, TRAF2 and NSMCE2. In previous studies, several of the members in this collection of 10 genes have been identified as biomarkers or therapeutic targets in HCC. Hu et al. [44] reported that KIF4A was overexpressed in HCC tissues, and that KIF4A knockdown could inhibit the proliferation of HCC cells. Furthermore, Yang et al. [45] reported that the upregulation of PBK was correlated with an unfavorable prognosis in HCC, and PBK could also promote HCC invasion and migration via the ETV4-uPAR signaling pathway. Lian et al. [46] found that ANLN is an essential gene for the growth of HBV-associated HCC. MCM2 was shown to promote the stemness and sorafenib resistance of HCC through regulating Hippo signaling [47]. A study by Liang et al. [48] demonstrated that TRAF2 depletion could suppress HCC proliferation through inhibiting mTORC1 activity. As for NCAPG, through employing genome-wide CRISPR knockout technology, it was found that NCAPG is an essential oncogene for HCC growth [49]. Ai et al. [50] reported that NCAPG promoted the proliferation, invasion and migration of HCC cells. Xiao et al. [51] showed that NCAPG could promote HCC progression through inhibiting pyroptosis. The previously reported promoting effects of NCAPG on HCC proliferation, invasion, migration and stemness maintenance have been corroborated by the findings of the current study. We were also interested in elucidating the underlying molecular mechanisms governing how NCAPG maintains HCC stemness, an aspect that will be explored and discussed further in future studies by our research group.
While our data demonstrate that NCAPG knockdown reduces the expression of stemness markers including OCT4, NANOG, and CD44, the precise molecular mechanism by which NCAPG—a condensin complex subunit—regulates the transcription of these stemness factors remains to be fully elucidated. We acknowledge that the current study does not provide direct mechanistic evidence, and we consider this an important limitation that warrants discussion. NCAPG is a core component of the condensin I complex, which is primarily known for its essential role in mitotic chromosome condensation and segregation. However, emerging evidence has revealed that condensin complexes also participate in interphase chromatin organization and transcriptional regulation. Specifically, condensin I has been shown to influence higher-order chromatin structure and modulate chromatin accessibility at specific genomic loci, thereby affecting gene expression programs. Based on these established functions, we hypothesize that NCAPG may regulate stemness factor expression through its role in chromatin organization. One plausible model is that NCAPG, as part of the condensin I complex, modulates chromatin accessibility at the promoters of key stemness regulatory genes, thereby influencing their transcriptional output. Alternatively, NCAPG may affect the expression or activity of upstream signaling pathways that converge on stemness transcription factors.
Subsequently, we found that tumors in high-risk groups were prone to the “cold” tumor model, and that low-risk patients were more likely to derive benefits from ICB therapy. ICB therapy is a common strategy that is employed in the treatment of postoperative patients and advanced patients. A major concern of ICB therapy, however, is that its therapeutic effectiveness varies greatly among individuals, especially in cancer types with high heterogeneity, including HCC and cholangiocarcinoma [52, 53]. The current risk stratification system shows a relatively good predictive performance with respect to ICB therapy responses, and this has been validated by several widely used digital predictive tools and real-world clinical trial cohorts. Our hope was that the predictive efficacy of our risk stratification system with respect to the ICB therapy response could be validated by future researchers and additional clinical trials. A critical limitation of our immunotherapy response analysis is that the external validation of the risk score’s predictive value relied on melanoma (GSE91061) and urothelial carcinoma (IMvigor210) datasets, as open-access HCC cohorts with detailed immunotherapy response and survival data were not available at the time of our analysis. We fully acknowledge that this cross-cancer extrapolation represents a major limitation that must be carefully considered when interpreting our findings. Moreover, patients in these retrospective datasets may have received heterogeneous treatments, which could affect survival outcomes independently of the risk score. While the consistent association between the risk score and survival across multiple cohorts with different treatment backgrounds strengthens our confidence in the robustness of the signature. Additionally, the lack of prospective clinical validation restricts translation into practice. An additional limitation is that our in vivo experiments were performed in immunodeficient BALB/c nude mice, which lack functional T cells. While this model is standard for human HCC xenograft studies and allowed us to evaluate the tumor-intrinsic effects of NCAPG knockdown, it is not suitable for validating our immune-hot versus immune-cold classification or our immunotherapy response predictions. The tumor-immune interactions we characterized through in silico analyses (CIBERSORT and ESTIMATE) could not be recapitulated in this model. Future studies using immunocompetent syngeneic models are necessary to validate these immune-related findings and to test the utility of our risk score for guiding immunotherapy.
For practical translation, the risk score is calculated as a linear combination of the 10 gene expression levels weighted by their Cox regression coefficients, with patients stratified into high- and low-risk groups using the median score from the training cohort. While RNA-seq or microarray platforms are suitable for research applications, a targeted RT-qPCR panel offers a more cost-effective and clinically feasible alternative for routine implementation. The risk score provides complementary prognostic information to the BCLC staging system and could potentially guide adjuvant therapy decisions—for instance, high-risk patients may be candidates for more intensive surveillance or adjuvant chemotherapy given their elevated recurrence risk and predicted chemotherapy sensitivity. We propose that the risk score could be integrated into the BCLC framework to refine risk stratification: for example, BCLC 0-A patients with high-risk scores might warrant intensified follow-up, while BCLC B-C patients with high-risk scores could be considered for more aggressive treatment or clinical trial enrollment. However, we emphasize that this conceptual framework requires prospective validation before clinical adoption. Future studies should evaluate whether risk-score-guided management improves patient outcomes in a stepwise manner, starting with retrospective validation, followed by prospective observational studies, and ultimately randomized controlled trials if warranted.”
CSCs in HCC have been shown to be strongly associated with heterogeneity constitution, tumorigenesis, tumor relapse, metastasis and resistance to anti-cancer therapies, partly contributing to the poor prognosis of HCC patients [54]. Interestingly, the results of the GSEA in the present study also revealed that the enrichment scores of IL6-JAK-STAT3 signaling, mTORC1 signaling, Notch signaling, PI3K-AKT-mTOR signaling, DNA repair, E2F targets, EMT signaling, glycolysis, hypoxia and IL2-STAT5 signaling were all increased in the high-risk group. Existing evidence has implied a relationship among the poor prognosis associated with the high-risk group, HCC stemness and the enrichment of these cancer hallmarks. Previous studies had reported that activation of PI3K-AKT-mTOR signaling could promote HCC progression and stemness [55, 56]. The activation of aerobic glycolysis in HCC cells was also shown to be both a cancer-promoting factor and a stemness-maintenance factor [57, 58]. Furthermore, the Notch signaling pathway has been widely acknowledged as one of the classical pathways to maintain cancer stemness, and to be a putative target for the development of anti-cancer drugs [59]. Hypoxia is also an important condition that has been shown to promote cancer stemness and drug resistance [60].
Certain limitations of the present study, however, should be acknowledged. Our reliance on bulk RNA-seq data for the calculation of the stemness index and the development of the prognostic signature. Bulk RNA-seq averages gene expression across all cells in a tumor sample, including malignant cells, stromal cells, and immune cells. Consequently, the OCLR-based stemness index reflects a population-averaged measure rather than the actual percentage or functional behavior of physical cancer stem cells. Secondly, we acknowledge that our chemotherapy sensitivity predictions have not been experimentally validated. In vitro cell viability assays with the predicted chemotherapeutic agents are necessary to confirm that the risk score truly identifies patients with differential drug sensitivity. Thirdly, the GSE91061 and IMvigor210 cohorts were recruited for external validation of our risk stratification system in ICB treatment response prediction due to the lack of open access datasets concerning ICB therapy for patients with HCC. Although the similar study designs could be found in previous publications [36, 37], ICB therapy cohorts toward patients with HCC were still the optimal to validate the predictive efficacy of our signature, and therefore we hoped the future researchers could provide or share their related data to validate our findings. A limitation of this study is that functional validation was primarily focused on NCAPG, while the remaining nine genes in the signature were not experimentally characterized. This was a pragmatic decision based on the selection criteria described above and the recognition that six of the remaining genes have been independently validated in the literature. Nevertheless, we acknowledge that the absence of experimental evidence for all signature genes represents an important limitation. Future studies should investigate the functional roles of CDK1, BUB1B, TOP2A, CCNB1, AURKA, and KIF20A specifically in the context of HCC stemness maintenance, and should explore whether combinatorial targeting of these genes yields additive or synergistic effects on CSC properties. Additionally, the relative contribution of each gene to the overall prognostic performance of the signature warrants further investigation through systematic perturbation experiments. These studies, while beyond the scope of the current work, would provide a more comprehensive understanding of the molecular architecture underlying this stemness-related prognostic model. Additionally, while we demonstrate that NCAPG knockdown impairs sphere formation and reduces CSC marker expression, the gold-standard in vivo limiting dilution assay to definitively confirm a reduction in tumor-initiating capacity was not performed. The substantial animal requirements and extended experimental timeline precluded this analysis in the current study. We acknowledge that these in vivo self-renewal experiments would provide the most rigorous evidence for NCAPG’s role in modulating tumor-initiating capacity and represent an important direction for future investigation. A significant limitation of this study is the absence of experiments testing small-molecule inhibitors or other NCAPG-targeting strategies. While we have reviewed the literature on available condensin complex inhibitors, including Q15 and Michellamine A, and discussed the therapeutic potential and challenges of targeting NCAPG, we have not performed any experimental validation of these agents. In vitro dose-response studies and in vivo efficacy studies using NCAPG-targeting compounds would be required to demonstrate the therapeutic potential of NCAPG inhibition. We acknowledge that this limitation restricts the direct clinical translation of our findings. Future studies should evaluate available condensin inhibitors or develop novel NCAPG-targeting strategies to confirm whether pharmacological inhibition of NCAPG recapitulates the tumor-suppressive effects of genetic knockdown observed in our study. Moreover, Future in vivo studies should incorporate additional HCC cell lines, such as Huh-7 and Hep3B, to confirm the reproducibility of NCAPG’s tumor-promoting function across diverse HCC backgrounds. Beyond the mechanistic studies discussed above, a key future direction will be the integration of single-cell RNA-seq to characterize the cellular heterogeneity underlying our 10-gene signature.
Conclusions
In conclusion, in the present study a novel stemness-related gene signature has been developed which has served as an indicator for survival risk evaluation and ICB treatment response prediction for patients with HCC. Moreover, the effects of NCAPG on the proliferation, invasion, migration and stemness maintenance of HCC were also verified, suggesting that NCAPG is a potential biomarker and therapeutic target for HCC.
Supplementary Information
Supplementary Material 2: Supplementary Figure 1. The expression level of the 10 SRGs used for prognostic model construction in TCGA‑LIHC dataset, and the knockdown efficiency of NCAPG in Huh‑7 and HCCLM3 cells. (A) The mRNA levels of the 10 SRGs used for construction of the prognostic model in TCGA‑LIHC dataset. (B and C) RT‑qPCR assay was performed to detect the knockdown efficiency in stable NCAPG knockdown Huh‑7 cells or HCCLM3 cells. (D and E) Western blotting assay to assess the protein expression levels of NCAPG, CD44, OCT4 and NANOG in the control and NCAPG knockdown groups. *P<0.05; **P<0.01; ***P<0.001. SRG, stemness‑related gene; TCGA‑LIHC, The Cancer Genome Atlas Liver Hepatocellular Carcinoma; NCAPG, non‑SMC condensin I complex subunit G. Supplementary Table 1. Primer sequences for qRT-PCR and target sequences of shRNAs utilized in the current study. Supplementary Table 2. The 1050 SRGs identified by the OCLR algorithm. Supplementary Table 3. The 649 DEGs between normal liver samples and HCC samples in TCGA-LIHC cohort. Supplementary Table 4. Association of the 10-gene risk score with recurrence-free survival in independent HCC cohorts. Supplementary Table 5. The related information of the 31 prognostic models for C-index comparison.
Acknowledgements
Not applicable.
AI Disclosure
None of AI tools was used in our study.
Authors’ contributions
Xiongpei Huang and Zecheng Qiu designed and performed the bioinformatics analyses, the experiments in the current study. Liusheng Li and Jinglin Gong drafted and revised the manuscript and participated in data visualization. All authors approved the submitted version of the manuscript.
Funding
None.
Data availability
The data for bioinformatic analyses were obtained from TCGA database (https://www.cancer.gov/ccg/research/genome-sequencing/tcga), ICGC database (https://docs.icgc-argo.org/docs/data-access/icgc-25k-data) and GEO database (https://www.ncbi.nlm.nih.gov/geo/).
Declarations
Ethics approval and consent to participate
Animal experiments were approved by the Institutional Animal Care and Use Committee of Nanchang Royo Biotech Co., Ltd. (RYE2024060701) and were performed in accordance with the ARRIVE guidelines.
Consent for publication
Not Applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Koh JH, Wang M, Suzuki H, Muthiah M, Ng CH, Huang DQ. NAFLD and NAFLD-related HCC in Asia: Burden and Surveillance. J Clin Exp Hepatol. 2024;14:101213. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Sia D, Villanueva A, Friedman SL, Llovet JM. Liver cancer cell of origin, molecular class, and effects on patient prognosis. Gastroenterology. 2017;152:745–61. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Massarweh NN, El-Serag HB. Epidemiology of hepatocellular carcinoma and intrahepatic cholangiocarcinoma. Cancer Control. 2017;24:1073274817729245. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Ha NB, Yao F. Alcohol and hepatocellular carcinoma. Clin Liver Dis. 2024;28:633–46. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Gofton C, Agar M, George J. Early implementation of palliative and supportive care in hepatocellular carcinoma. Semin Liver Dis. 2022;42:514–30. [DOI] [PubMed] [Google Scholar]
- 6.Bicer F, Kure C, Ozluk AA, El-Rayes BF, Akce M. Advances in immunotherapy for hepatocellular carcinoma (HCC). Curr Oncol. 2023;30:9789–812. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Wu TK, Hui RW, Mak LY, Fung J, Seto WK, Yuen MF. Hepatocellular carcinoma: advances in systemic therapies. F1000Res. 2024;13:104. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Sun JH, Luo Q, Liu LL, Song GB. Liver cancer stem cell markers: progression and therapeutic implications. World J Gastroenterol. 2016;22:3547–57. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Xia P, Liu DH. Cancer stem cell markers for liver cancer and pancreatic cancer. Stem Cell Res. 2022;60:102701. [DOI] [PubMed] [Google Scholar]
- 10.Ahmad G, Amiji MM. Cancer stem cell-targeted therapeutics and delivery strategies. Expert Opin Drug Deliv. 2017;14:997–1008. [DOI] [PubMed] [Google Scholar]
- 11.Sarabia-Sánchez MA, Tinajero-Rodríguez JM, Ortiz-Sánchez E, Alvarado-Ortiz E. Cancer stem cell markers: symphonic masters of chemoresistance and immune evasion. Life Sci. 2024;355:123015. [DOI] [PubMed] [Google Scholar]
- 12.Niharika, Garg M. Understanding the autophagic functions in cancer stem cell maintenance and therapy resistance. Expert Rev Mol Med. 2024;26:e23. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Sloan AR, Silver DJ, Kint S, Gallo M, Lathia JD. Cancer stem cell hypothesis 2.0 in glioblastoma: where are we now and where are we going? Neuro Oncol. 2024;26:785–95. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Wang Z, Li R, Yang G, Wang Y. Cancer stem cell biomarkers and related signalling pathways. J Drug Target. 2024;32:33–44. [DOI] [PubMed] [Google Scholar]
- 15.Liu Y, Wang H. Biomarkers and targeted therapy for cancer stem cells. Trends Pharmacol Sci. 2024;45:56–66. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Masoumi J, Jafarzadeh A, Abdolalizadeh J, Khan H, Philippe J, Mirzaei H, Mirzaei HR. Cancer stem cell-targeted chimeric antigen receptor (CAR)-T cell therapy: challenges and prospects. Acta Pharm Sin B. 2021;11:1721–39. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Lee TK, Guan XY, Ma S. Cancer stem cells in hepatocellular carcinoma - from origin to clinical implications. Nat Rev Gastroenterol Hepatol. 2022;19:26–44. [DOI] [PubMed] [Google Scholar]
- 18.Zhou P, Li Z, Liu F, Kwon E, Hsieh TC, Ye S, Vasudevan S, Lee JA, Tran KV, Zhou C. BAMBI integrates biostatistical and artificial intelligence methods to improve RNA biomarker discovery. Brief Bioinform. 2025;26:bbaf073. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Xia XL, Zhou SM, Liu Y, Lin N, Overton IM. ieGENES: A machine learning method for selecting differentially expressed genes in cancer studies. J Biomed Inf. 2025;164:104803. [DOI] [PubMed] [Google Scholar]
- 20.Hu Z, Jia Z, Liu J, Mao A, Han H, Gu Z. MD-ALL: an integrative platform for molecular diagnosis of B-acute lymphoblastic leukemia. Haematologica. 2024;109:1741–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Latha NR, Rajan A, Nadhan R, Achyutuni S, Sengodan SK, Hemalatha SK, Varghese GR, Thankappan R, Krishnan N, Patra D, et al. Gene expression signatures: a tool for analysis of breast cancer prognosis and therapy. Crit Rev Oncol Hematol. 2020;151:102964. [DOI] [PubMed] [Google Scholar]
- 22.Pourbagheri-Sigaroodi A, Fallah F, Bashash D, Karimi A. Unleashing the potential of gene signatures as prognostic and predictive tools: a step closer to personalized medicine in hepatocellular carcinoma (HCC). Cell Biochem Funct. 2024;42:e3913. [DOI] [PubMed] [Google Scholar]
- 23.Bareche Y, Kelly D, Abbas-Aghababazadeh F, Nakano M, Esfahani PN, Tkachuk D, Mohammad H, Samstein R, Lee CH, Morris LGT, et al. Leveraging big data of immune checkpoint blockade response identifies novel potential targets. Ann Oncol. 2022;33:1304–17. [DOI] [PubMed] [Google Scholar]
- 24.Zhou K, Ma X, Yan T, Zheng H, Xie S, Guo L, Lu R. Stemness-relevant gene signature for chemotherapeutic response and prognosis prediction in ovarian cancer. Stem Cells Int. 2025;2025:2505812. [DOI] [PMC free article] [PubMed]
- 25.Zheng W, Zhou C, Xue Z, Qiao L, Wang J, Lu F. Integrative analysis of a novel signature incorporating metabolism and stemness-related genes for risk stratification and assessing clinical outcomes and therapeutic responses in lung adenocarcinoma. BMC Cancer. 2025;25:591. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Yang X, Yang X, Tang H, Chen X, Wang J, Zhao H. Characterization of stem cell landscape and identification of stemness-relevant prognostic gene signature to aid immunotherapy in breast cancer. Discov Oncol. 2025;16:9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Rosenberg JE, Galsky MD, Powles T, Petrylak DP, Bellmunt J, Loriot Y, Necchi A, Hoffman-Censits J, Perez-Gracia JL, van der Heijden MS, et al. Atezolizumab monotherapy for metastatic urothelial carcinoma: final analysis from the phase II IMvigor210 trial. ESMO Open. 2024;9:103972. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Malta TM, Sokolov A, Gentles AJ, Burzykowski T, Poisson L, Weinstein JN, Kamińska B, Huelsken J, Omberg L, Gevaert O, et al. Machine learning identifies stemness features associated with oncogenic dedifferentiation. Cell. 2018;173:338–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Daily K, Ho Sui SJ, Schriml LM, Dexheimer PJ, Salomonis N, Schroll R, Bush S, Keddache M, Mayhew C, Lotia S, et al. Molecular, phenotypic, and sample-associated data to describe pluripotent stem cell lines and derivatives. Sci Data. 2017;4:170030. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Salomonis N, Dexheimer PJ, Omberg L, Schroll R, Bush S, Huo J, Schriml L, Ho Sui S, Keddache M, Mayhew C, et al. Integrated genomic analysis of diverse induced pluripotent stem cells from the progenitor cell biology consortium. Stem Cell Rep. 2016;7:110–25. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Xu H, Liu Z, Weng S, Dang Q, Ge X, Zhang Y, Ren Y, Xing Z, Chen S, Zhou Y, et al. Artificial intelligence-driven consensus gene signatures for improving bladder cancer clinical outcomes identified by multi-center integration analysis. Mol Oncol. 2022;16:4023–42. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Yoshihara K, Shahmoradgoli M, Martínez E, Vegesna R, Kim H, Torres-Garcia W, Treviño V, Shen H, Laird PW, Levine DA, et al. Inferring tumour purity and stromal and immune cell admixture from expression data. Nat Commun. 2013;4:2612. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Jiang P, Gu S, Pan D, Fu J, Sahu A, Hu X, Li Z, Traugh N, Bu X, Li B, et al. Signatures of T cell dysfunction and exclusion predict cancer immunotherapy response. Nat Med. 2018;24:1550–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Charoentong P, Finotello F, Angelova M, Mayer C, Efremova M, Rieder D, Hackl H, Trajanoski Z. Pan-cancer Immunogenomic analyses reveal genotype-immunophenotype relationships and predictors of response to checkpoint blockade. Cell Rep. 2017;18:248–62. [DOI] [PubMed] [Google Scholar]
- 35.Chalmers ZR, Connelly CF, Fabrizio D, Gay L, Ali SM, Ennis R, Schrock A, Campbell B, Shlien A, Chmielecki J, et al. Analysis of 100,000 human cancer genomes reveals the landscape of tumor mutational burden. Genome Med. 2017;9:34. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Wei C, Chen M, Deng W, Bie L, Ma Y, Zhang C, Liu K, Shen W, Wang S, Yang C, et al. Characterization of gastric cancer stem-like molecular features, immune and pharmacogenomic landscapes. Brief Bioinform. 2022;23:bbab386. [DOI] [PubMed] [Google Scholar]
- 37.Xiong D, Wang Y, You M. A gene expression signature of TREM2hi macrophages and γδ T cells predicts immunotherapy response. Nat Commun. 2020;11:5084. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Schwartz LH, Seymour L, Litière S, Ford R, Gwyther S, Mandrekar S, Shankar L, Bogaerts J, Chen A, Dancey J, et al. RECIST 1.1 - Standardisation and disease-specific adaptations: perspectives from the RECIST Working Group. Eur J Cancer. 2016;62:138–45. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Kawashima S, Togashi Y. Resistance to immune checkpoint inhibitors and the tumor microenvironment. Exp Dermatol. 2023;32:240–9. [DOI] [PubMed] [Google Scholar]
- 40.Wang Z, Wang Y, Gao P, Ding J. Immune checkpoint inhibitor resistance in hepatocellular carcinoma. Cancer Lett. 2023;555:216038. [DOI] [PubMed] [Google Scholar]
- 41.Ivanova M, Shivarov V. HLA genotyping meets response to immune checkpoint inhibitors prediction: a story just started. Int J Immunogenet. 2021;48:193–200. [DOI] [PubMed] [Google Scholar]
- 42.Wu B, Zhang B, Li B, Wu H, Jiang M. Cold and hot tumors: from molecular mechanisms to targeted therapy. Signal Transduct Target Ther. 2024;9:274. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43.Li R, Wang D, Yang H, Pu L, Li X, Yang F, Zhu R. Important role and underlying mechanism of non–SMC condensin I complex subunit G in tumours (Review). Oncol Rep. 2024;51:77. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Hu G, Yan Z, Zhang C, Cheng M, Yan Y, Wang Y, Deng L, Lu Q, Luo S. FOXM1 promotes hepatocellular carcinoma progression by regulating KIF4A expression. J Exp Clin Cancer Res. 2019;38:188. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Yang QX, Zhong S, He L, Jia XJ, Tang H, Cheng ST, Ren JH, Yu HB, Zhou L, Zhou HZ, et al. PBK overexpression promotes metastasis of hepatocellular carcinoma via activating ETV4-uPAR signaling pathway. Cancer Lett. 2019;452:90–102. [DOI] [PubMed] [Google Scholar]
- 46.Lian YF, Huang YL, Wang JL, Deng MH, Xia TL, Zeng MS, Chen MS, Wang HB, Huang YH. Anillin is required for tumor growth and regulated by miR-15a/miR-16-1 in HBV-related hepatocellular carcinoma. Aging. 2018;10:1884–901. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 47.Zhou X, Luo J, Xie H, Wei Z, Li T, Liu J, Liao X, Zhu G, Peng T. MCM2 promotes the stemness and sorafenib resistance of hepatocellular carcinoma cells via hippo signaling. Cell Death Discov. 2022;8:418. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 48.Liang X, Yao J, Cui D, Zheng W, Liu Y, Lou G, Ye B, Shui L, Sun Y, Zhao Y, et al. The TRAF2-p62 axis promotes proliferation and survival of liver cancer by activating mTORC1 pathway. Cell Death Differ. 2023;30:1550–62. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 49.Wang Y, Gao B, Tan PY, Handoko YA, Sekar K, Deivasigamani A, Seshachalam VP, OuYang HY, Shi M, Xie C, et al. Genome-wide CRISPR knockout screens identify NCAPG as an essential oncogene for hepatocellular carcinoma tumor growth. FASEB J. 2019;33:8759–70. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 50.Ai J, Gong C, Wu J, Gao J, Liu W, Liao W, Wu L. MicroRNA–181c suppresses growth and metastasis of hepatocellular carcinoma by modulating NCAPG. Cancer Manag Res. 2019;11:3455–67. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 51.Xiao C, Gong J, Jie Y, Liang W, Tai Y, Qin W, Lu T, Chong Y, Hei Z, Hu B, et al. E2F1-mediated Up-regulation of NCAPG promotes hepatocellular carcinoma development by inhibiting pyroptosis. J Clin Transl Hepatol. 2024;12:25–35. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 52.Peeters F, Cappuyns S, Piqué-Gili M, Phillips G, Verslype C, Lambrechts D, Dekervel J. Applications of single-cell multi-omics in liver cancer. JHEP Rep. 2024;6:101094. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 53.Zhang M, Huang K, Yin Q, Wu X, Zhu M, Li M. Spat heterogeneity hepatocellular carcinoma microenvironment determines efficacy immunotherapy. 2025;16:15. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 54.Fang X, Yan Q, Liu S, Guan XY. Cancer stem cells in hepatocellular carcinoma: intrinsic and extrinsic molecular mechanisms in stemness regulation. Int J Mol Sci. 2022;23:12327. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 55.Li Q, Chen Q, Wang W, Xie R, Li Z, Chen D. KGF secreted from HSCs activates PAK4/BMI1, promotes HCC stemness through PI3K/AKT pathway. IUBMB Life. 2025;77:e2929. [DOI] [PubMed] [Google Scholar]
- 56.Yang X, Jiang W, Kong X, Zhou X, Zhu D, Kong L. Genistein restricts the epithelial mesenchymal transformation (EMT) and stemness of hepatocellular carcinoma via upregulating miR-1275 to Inhibit the EIF5A2/PI3K/Akt Pathway. Biology (Basel). 2022;11:1383. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 57.Wu T, Liao L, Wu T, Chen S, Yi Q, Xu M. IGF2BP2 promotes glycolysis and hepatocellular carcinoma stemness by stabilizing CDC45 mRNA via m6A modification. Cell Cycle. 2023;22:2245–63. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 58.Ji W, Bai J, Ke Y. Exosomal ZFPM2-AS1 contributes to tumorigenesis, metastasis, stemness, macrophage polarization, and infiltration in hepatocellular carcinoma through PKM mediated glycolysis. Environ Toxicol. 2023;38:1332–46. [DOI] [PubMed] [Google Scholar]
- 59.Iluta S, Nistor M, Buruiana S, Dima D. Notch and hedgehog signaling unveiled: crosstalk, roles, and breakthroughs in cancer stem cell research. Life (Basel). 2025;15:228. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 60.Shang T, Jia Z, Li J, Cao H, Xu H, Cong L, Ma D, Wang X, Liu J. Unraveling the triad of hypoxia, cancer cell stemness, and drug resistance. J Hematol Oncol. 2025;18:32. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Supplementary Material 2: Supplementary Figure 1. The expression level of the 10 SRGs used for prognostic model construction in TCGA‑LIHC dataset, and the knockdown efficiency of NCAPG in Huh‑7 and HCCLM3 cells. (A) The mRNA levels of the 10 SRGs used for construction of the prognostic model in TCGA‑LIHC dataset. (B and C) RT‑qPCR assay was performed to detect the knockdown efficiency in stable NCAPG knockdown Huh‑7 cells or HCCLM3 cells. (D and E) Western blotting assay to assess the protein expression levels of NCAPG, CD44, OCT4 and NANOG in the control and NCAPG knockdown groups. *P<0.05; **P<0.01; ***P<0.001. SRG, stemness‑related gene; TCGA‑LIHC, The Cancer Genome Atlas Liver Hepatocellular Carcinoma; NCAPG, non‑SMC condensin I complex subunit G. Supplementary Table 1. Primer sequences for qRT-PCR and target sequences of shRNAs utilized in the current study. Supplementary Table 2. The 1050 SRGs identified by the OCLR algorithm. Supplementary Table 3. The 649 DEGs between normal liver samples and HCC samples in TCGA-LIHC cohort. Supplementary Table 4. Association of the 10-gene risk score with recurrence-free survival in independent HCC cohorts. Supplementary Table 5. The related information of the 31 prognostic models for C-index comparison.
Data Availability Statement
The data for bioinformatic analyses were obtained from TCGA database (https://www.cancer.gov/ccg/research/genome-sequencing/tcga), ICGC database (https://docs.icgc-argo.org/docs/data-access/icgc-25k-data) and GEO database (https://www.ncbi.nlm.nih.gov/geo/).
