Comprehensive multi-omics analysis identifies cancer subtypes and prognostic signatures of hepatitis B virus-associated hepatocellular carcinoma Authors: Rui Wang, Bo Ren, Xijie Zhang, Bo Liu, Wence Zhou Published: 2026 ABSTRACT Hepatitis B virus-associated hepatocellular carcinoma (HBV-HCC) is a heterogeneous malignancy with poor prognosis, necessitating refined classification and novel therapeutics. Leveraging multi-omics data, we utilized the MOVICS package to identify two robust cancer subtypes (CSs) from the TCGA-LIHC cohort. CS1 was characterized by a favorable prognosis and a “hot” immune microenvironment, whereas CS2 exhibited aggressive clinicopathological features and a “cold” immune landscape. To predict prognosis, we constructed a machine learning-derived prognostic signature (MLPS) using 92 algorithm combinations. The optimal model, identified as RSF+Enet[alpha = 0.2], demonstrated superior predictive accuracy and outperformed ten existing HBV-HCC models. Furthermore, we implemented a drug repurposing strategy for high-risk patients and identified KX2-391, a dual inhibitor from the CTRP and PRISM databases, as a potent therapeutic agent. In vitro assays confirmed that KX2-391 dose-dependently induces apoptosis in high-MLPS HBV-HCC cell lines, and in vivo xenograft models demonstrated its superior efficacy in inhibiting tumor growth compared to lenvatinib, without significant organ toxicity. Mechanistically, we discovered that F9 acts as a tumor suppressor by interacting with SERPINC1 to inhibit the Wnt/β-catenin signaling pathway. Collectively, our study provides a novel HBV-HCC classification system, a high-performance prognostic tool, and identifies both KX2-391 and F9 as promising avenues for precision therapy. Introduction Hepatitis B virus (HBV) infection is one of the most common viral infections worldwide, and long-term infection can lead to the development of liver cancer1. HBV-associated hepatocellular carcinoma (HBV-HCC) is one of the leading causes of liver cancer cases globally, contributing to a large number of deaths and burdens, especially in Asia and Africa2,3. Despite significant progress in the prevention and treatment of hepatitis B in recent years, the treatment of HBV-HCC remains challenging4. Current treatment options for HCC include surgical resection, local ablation, chemotherapy, targeted therapy, and immunotherapy; however, the overall efficacy is still unsatisfactory5. The development of HBV-HCC is a multifactorial, multistep, and complex process. After HBV infection enters the hepatocyte, the viral genome integrates into the genome of the host cell, leading to genetic variation and altered gene expression in the host cell6. These changes may promote malignant transformation of cells and tumor growth. In addition, HBV infection may promote hepatocarcinogenesis by activating the inflammatory response of hepatocytes and interfering with apoptosis7. Previous studies have found that patients with HBV-HCC have a better prognosis and are more susceptible to targeted therapies and immunotherapies than HCC patients caused by other factors8–10. However, heterogeneity still exists among HBV-HCC patients, and not all HBV-HCC patients show the same response profile to the same treatment. HBV-HCC is fundamentally distinct from other forms of HCC—including hepatitis C virus (HCV)-related HCC, alcohol-associated HCC, and non-viral metabolic HCC—yet current classification frameworks for HCC largely ignore this etiological divide, pooling HBV-HCC with non-HBV subtypes and obscuring its unique biology. The core distinction lies in etiological drivers and molecular mechanisms. HBV-HCC is driven by direct virus-host genomic and functional interactions: HBV integrates into the host genome via covalently closed circular DNA, disrupting tumor suppressors or activating oncogenes11. The HBV X protein (HBx) further hijacks signaling pathways (Wnt/β-catenin, NF-κB), degrades p53, and suppresses DNA repair, creating a “virus-engineered” oncogenic landscape12,13. In contrast, non-HBV HCC arises from chronic injury-repair cycles: HCV-HCC relies on persistent viral replication, inducing oxidative stress14,15; alcohol-HCC is mainly attributed to the direct toxic effects of alcohol metabolites: alcohol is metabolized into acetaldehyde in the liver, which directly damages hepatocyte DNA, leading to DNA mutations and impaired repair16; and metabolic dysfunction-associated steatotic liver disease-HCC is fueled by lipotoxicity and metabolic gene variants (PNPLA3, TM6SF2, MBOAT7)17. Despite advances in HCC subtyping, existing frameworks predominantly focus on pan-HCC (aggregating mixed etiologies), which dilutes HBV-specific oncogenic signals and limits the development of etiology-tailored therapies. The Cancer Genome Atlas (TCGA) classification18 and Hoshida et al.’s seminal “S1–S3” subtypes19 grouped HBV-HCC with non-viral HCC, obscuring HBV-driven pathways such as Wnt/β-catenin activation via HBx6. This pan-HCC approach fails to capture the multifactorial heterogeneity of HBV-HCC, where epigenetic alterations and immune microenvironment dynamics collectively drive tumor evolution. Critical limitations of current classifications underscore the need for HBV-HCC-specific frameworks. First, most studies lack etiological specificity: early HBV-HCC subtyping efforts20,21relied on small, single-center cohorts with homogeneous ethnicities, limiting generalizability to diverse global HBV-HCC populations (e.g., Chinese, African, or European cohorts with distinct viral genotypes and host genetics). Second, single-omics bias prevails: prior work often focused on isolated layers (e.g., transcriptome-only22; genomics-only23), ignoring the interplay between immune cell infiltration and metabolic reprogramming unique to HBV-HCC. In this study, we propose to systematically identify consensus cancer subtypes (CSs) of HBV-HCC patients and elucidate their molecular and clinical features differences by integrating multi-omics data of HBV-HCC patients to provide a more detailed and comprehensive understanding of HBV-HCC and to use multi-omics data to help overcome the limitations of single-omics. Results Identification of multi-omics HBV-HCC cancer subtypes Based on the results of the CPI and the Gap-statistics, the number of CS with the highest average statistical value for TCGA-LIHC cohort was found to be two (Fig. 1a). Consensus integration for robustness was performed for the respective CS obtained by the ten algorithms (Fig. 1b). Consensus matrix for two CSs based on the ten algorithms (Fig. 1c). Silhouette analysis demonstrated sample similarity (0.35 and 0.68), which were named CS1 (n = 72) and CS2 (n = 27), respectively (Fig. 1d). Significant differences in molecular features were demonstrated in CS1 and CS2 (Fig. 1e). There were significant differences in the clinical characteristics between CS1 and CS2; i.e., patients with HBV-HCC in the CS2 group had a worse prognosis, higher pathological stage, higher TNM stage, higher grade, a higher proportion of vascular invasion, and a higher proportion of severe type of adjacent hepatic tissue inflammation extent type (Fig. 1e and Supplementary Table 1).Fig. 1Identification of cancer subtypes based on HBV-HCC multi-omics data from the TCGA-LIHC cohort.a The number of cancer subtypes and their corresponding CPI and Gap-statistics. b Clustering of HBV-HCC patients by ten clustering methods. c Consensus clustering matrix for two cancer subtypes based on ten algorithms. d Silhouette analysis assessment of the quality of consensus clustering results. e Comprehensive heatmap of consensus cancer subtypes of HBV-HCC, including mRNA, lncRNA, miRNA, DNA CpG methylation sites, and mutated genes, as well as the distribution of clinical features among different cancer subtypes. f KM curves revealing differences in OS between different cancer subtypes. g ROC curves reveal the ability of CSs to predict OS in HBV-HCC. h, i Univariate and multivariate Cox regression analyses revealing the effect of CSs on HBV-HCC patients’ OS. Identification of cancer subtypes based on HBV-HCC multi-omics data from the TCGA-LIHC cohort. a The number of cancer subtypes and their corresponding CPI and Gap-statistics. b Clustering of HBV-HCC patients by ten clustering methods. c Consensus clustering matrix for two cancer subtypes based on ten algorithms. d Silhouette analysis assessment of the quality of consensus clustering results. e Comprehensive heatmap of consensus cancer subtypes of HBV-HCC, including mRNA, lncRNA, miRNA, DNA CpG methylation sites, and mutated genes, as well as the distribution of clinical features among different cancer subtypes. f KM curves revealing differences in OS between different cancer subtypes. g ROC curves reveal the ability of CSs to predict OS in HBV-HCC. h, i Univariate and multivariate Cox regression analyses revealing the effect of CSs on HBV-HCC patients’ OS. The KM curves showed that the OS, DSS, DFI, and PFI of HBV-HCC patients in the CS1 group were significantly better than those of patients in the CS2 group (Fig. 1f and Supplementary Fig. 1a). ROC curves showed that different CSs could accurately predict the OS, DSS, DFI, and PFI of HBV-HCC patients (Fig. 1g and Supplementary Fig. 1b), and Cox regression analysis showed that CS2 was a risk factor for OS, DSS, DFI, and PFI in HBV-HCC patients (Fig. 1h, i and Supplementary Fig. 1c, d). Differences in genomic variation, biological function, and immune landscape between different HBV-HCC cancer subtypes A study of SNV from the TCGA HBV-HCC cohort found a higher proportion of mutations in CTNNB1, OBSCN, APOB, ZFHX4, COL4A4, ZNF638, KIF4B, GRID1, CHRM3, KIAA1217, PLCL2, KCNB2, NLRP2, and ZFYVE26 in the CS2 patients compared to the CS1 patients, but there was no difference in TMB between patients in the CS1 and CS2 groups (Fig. 2a). In terms of CNV, significant differences in CNV and gistic score were demonstrated between CS1 and CS2 patients (Fig. 2a and Supplementary Fig. 2), especially CS2 patients had a higher proportion of amplifications at loci 5p15.33, 5p13.2, 17q11.2, 17q21.33, 17q23.1, and 17q25.3, and a higher proportion of deletions at loci 14q23.3, 14q32.33, and 18q21.2. Meanwhile, TOP2A and STAT3 had a higher frequency of CNVs in the CS2 group of patients.Fig. 2Differences in genomic variation, biological function and immune landscape between different HBV-HCC CSs.a ComplexHeatmap revealing genomic variation between different CSs. b, c Biological functional differences between different HBV-HCC CSs based on hallmark and KEGG terms. d–g Differences in the immune landscape between different HBV-HCC CSs. Differences in genomic variation, biological function and immune landscape between different HBV-HCC CSs. a ComplexHeatmap revealing genomic variation between different CSs. b, c Biological functional differences between different HBV-HCC CSs based on hallmark and KEGG terms. d–g Differences in the immune landscape between different HBV-HCC CSs. Based on KEGG and hallmark terms showed that CS1 group patients were significantly enriched for multiple coagulation-related and metabolism-related pathways, including complement and coagulation cascades, coagulation, glycine serine and threonine metabolism, alanine aspartate and glutamate metabolism, and linoleic acid metabolism, whereas CS2 group patients were remarkably concentrated on glycosphingolipid biosynthesis, DNA repair, reactive oxygen species pathway, and Myc targets (Fig. 2b, c). With the IOBR package, we found that CS1 group HBV-HCC patients showed a distinct “hot tumor” type, while CS2 showed a “cold tumor” type (Fig. 2d–g). A variety of anti-tumor cells (CD4/CD8 + T cells, B cells, naïve CD4 + T cells, etc.) had high infiltration levels in the CS1 group, while pro-tumor cells such as MDSC and M2 type macrophages had high levels in the CS2 group patients. In addition, CS1 group HBV-HCC patients had higher levels of Pan_F_TBRs, CAF_Peng_et_al, EMT1, and CD_8_T_effector, whereas APM was at a higher level in CS2 group HBV-HCC patients. Validation of four external cohorts To explore the robustness of our findings, four external cohorts (GSE14520, GSE144269, ICGC-LIRI-JP, and CHCC-HBV) were used for validation. The CS to which each patient in each cohort belonged was predicted by the CS-specific up-regulated biomarkers obtained earlier according to the NTP method (Fig. 3a). Interestingly, a consistently poorer prognosis of patients in the CS2 group compared to patients in the CS1 group was observed in four external validation cohorts (Fig. 3b).Fig. 3Validation of four external cohorts.a Using the NTP method to validate CSs in four external validation cohorts. b KM curves revealing OS differences between CS1 and CS2 in the four external validation cohorts. Validation of four external cohorts. a Using the NTP method to validate CSs in four external validation cohorts. b KM curves revealing OS differences between CS1 and CS2 in the four external validation cohorts. Construction of machine learning-derived prognostic signature (MLPS) A total of 31 intersected genes were obtained (Fig.4a). Supplementary Fig. 3 shows the differences before and after batch effect correction, with results indicating that the batch effect was effectively removed post-correction. HBV-HCC from the TCGA cohort was used as the training set to construct the MLPSs using 92 machine learning models or their combinations, and the other four cohorts were used as the testing set. The results showed that the modeling method of RSF+Enet[alpha = 0.2] had the highest average C-index and was considered the best modeling method (Fig.4b). The algorithm combination of RSF+Enet[alpha=0.2] included 10 genes, of which REG3A, SPP1, and TKT were highly expressed in tumor tissues, whereas AFM, ALDOB, F9, FETUB, HRG, IGFALS, and SPP2 were lowly expressed in tumor tissues (Fig.4c) and that these 10 genes could accurately discriminate the diagnosis of HBV-HCC tumor and non-tumor tissues, especially for F9 (AUC = 0.929, 95% CI: 0.886–0.967), IGFALS (AUC = 0.929, 95%CI: 0.885–0.965), and TKT (AUC = 0.925, 95%CI: 0.877–0.964) (Fig.4d). The KM curves showed that high expression of AFM, F9, FETUB, HRG, IGFALS, and SPP2 in HBV-HCC had a better OS, whereas high expression of REG3A, SPP1, and TKT in HBV-HCC had a poor OS (Supplementary Fig. 4).Fig. 4Construction of MLPS.a Take the intersection of DEGs between HBV-HCC tumor and non-tumor patients, CS-specific up-regulated biomarkers, and shared genes in five cohorts. b C-index of models derived from machine learning algorithms of 92 methods in five HBV-HCC cohorts. c Expression differences between HBV-HCC tumors and non-tumors for the ten genes included in the algorithmic combination of RSF + Enet[alpha = 0.2]. d The ability of the ten genes included in the algorithmic combination of RSF+Enet[alpha = 0.2] in differentiating HBV-HCC tumor from non-tumor patients. e Distribution of patients with high and low MLPS scores among different CSs of HBV-HCC patients in the TCGA-LIHC cohort. f KM curves revealing OS differences between patients with high and low MLPS scores in five HBV-HCC cohorts. Construction of MLPS. a Take the intersection of DEGs between HBV-HCC tumor and non-tumor patients, CS-specific up-regulated biomarkers, and shared genes in five cohorts. b C-index of models derived from machine learning algorithms of 92 methods in five HBV-HCC cohorts. c Expression differences between HBV-HCC tumors and non-tumors for the ten genes included in the algorithmic combination of RSF + Enet[alpha = 0.2]. d The ability of the ten genes included in the algorithmic combination of RSF+Enet[alpha = 0.2] in differentiating HBV-HCC tumor from non-tumor patients. e Distribution of patients with high and low MLPS scores among different CSs of HBV-HCC patients in the TCGA-LIHC cohort. f KM curves revealing OS differences between patients with high and low MLPS scores in five HBV-HCC cohorts. In addition, we found that patients with low MLPS scores were almost exclusively classified as CS1 (48/49), whereas the vast majority of patients in the CS2 group were patients with high MLPS scores (26/27) (Fig.4e). The KM curves showed significantly better prognosis for HBV-HCC patients with low MLPS scores than for HBV-HCC patients with high MLPS scores in all five cohorts (Fig.4f), and MLPS scores accurately predicted patients’ OS (Supplementary Fig. 5). Cox regression analyses showed MLPS to be an independent risk factor for OS across multiple cohorts (Supplementary Fig. 6). For DSS, PFI, and DFI, HBV-HCC patients with high MLPS scores had poor DSS, PFI, and DFI (Supplementary Fig. 7a), and MLPS accurately predicted DSS, PFI, and DFI in HBV-HCC patients in the TCGA-LIHC cohort (Supplementary Fig. 7b). Cox regression analyses revealed MLPS as an independent risk factor for HBV-HCC DSS, PFI, and DFI in the TCGA-LIHC cohort (Supplementary Fig. 7c, d). The superiority and robustness of MLPS A comparison of ten previously published models on the prognosis of HBV-HCC patients (the details of the ten models were shown in Supplementary Table 2) revealed that, except in the GSE14520 cohort (where the C-index of MLPS was lower than that of Jianming_Dong’s model) and CHCC-HBV cohort (where the C-index of MLPS was lower than that of Zhenfu_Gao’s and Xiwen_Liao’s models), the rest of the cohort MLPS exhibits the highest C-index. In addition, Jianming_Dong, Zhenfu_Gao, and Xiwen_Liao’s models in TCGA-LIHC, ICGC-LIRI-JP, and GSE144269 suffered from a lack of robustness with a C-index of less than 0.6, which demonstrated that the MLPS we developed had a good superiority and robustness (Supplementary Fig. 8a). Subsequently, we compared the C-index of MLPS with multiple clinical features in predicting the prognosis of HBV-HCC patients and showed that MLPS demonstrated the highest C-index in predicting OS, DSS, PFI, and DFI of HBV-HCC patients in the TCGA-LIHC cohort (Supplementary Fig. 8b). MLPS also demonstrated good superiority in predicting OS in HBV-HCC patients in multiple external cohorts (Supplementary Fig. 8c). Differences in genome variation, biological function, and immune landscape between high-MLPS and low-MLPS scores HBV-HCC patients ComplexHeatmap reveals differences in genomic variants between patients in different MLPS groups (Supplementary Fig. 9a). Patients in the high-MLPS group had high proportions of mutations in CTNNB1, OBSCN, APOB, and EVC2, and patients in the high-MLPS group had higher levels of TMB. For CNV, significant differences were demonstrated between patients in the high and low MLPS groups, mainly in that patient in the high MLPS group had a higher proportion of amplifications at 5p15.33, 5p13.2, 5q35.3, 17q11.2, 7q21.33, 17q23.1, and 17q25.3, and deletions at 14q23.3, and a lower proportion of 11q13.3 amplification (Supplementary Fig. 9a). In addition, there were significant differences in CNV frequencies and gistic scores at multiple loci and chromosomes between patients in the high and low MLPS groups (Supplementary Fig. 9a–c). GSVA showed that oxidative phosphorylation, pyrimidine metabolism, glutathione metabolism, the pentose phosphate pathway, Myc targets, the reactive oxygen species pathway, and mTORC1 signaling were significantly enriched in patients with high-MLPS scores, whereas the low-MLPS scores group was concentrated in the Notch signaling pathway, PPAR signaling pathway, TGF-β signaling pathway, Hedgehog signaling pathway, and complement and coagulation cascades, etc. (Supplementary Fig. 10a). In terms of immune landscape, patients with low-MLPS scores had higher CD8+ naive T−cells_xCell, CD4_T cells_EPIC, B_cells_Rooney_et_al, Macrophages_EPIC, CAFs_EPIC, Pan_F_TBRs, Fibroblasts_MCPcounter, Fibroblasts_xCell, Monocytes_quantiseq, Neutrophils_MCPcounter, Dendritic cells resting_CIBERS, NK cells activated_CIBERSORT, NK cells_Bindea et al, Pan_F_TBRs, Fibroblasts_MCPcounter, Tregs_quantiseq, CAF_Peng_et_al, CAFs_EPIC, EMT2, TGFb Family Member_Li et al, TMEscoreB_CIR, and Pan_F_TBRs, while patients in the high-MLPS scores group had higher levels of CD4+ memory T−cells_xCell, pro_B−cells_xCell, Macrophages M0_CIBERSORT, Macrophages_xCell, MDSC_Peng_et_al., TMEscore_CIR, IPS_IPS and APM (Supplementary Fig. 10b). Identification and validation of compounds targeting patients in the high-MLPS scores group The flowchart of this part of the analysis is shown in Fig. 5a, and five CTPR-derived compounds (isoevodiamine, cabozantinib, KX2-391, lovastatin, and nakiterpiosin) and nine PRISM-derived compounds (ingenol-mebutate, PF-03814735, epothilone-b, trimebutine, KX2-391, docetaxel, gemcitabine, volasertib, and LGX818) were obtained according to log2 FC (log2FC > 0.02 for CTRP and log2FC > 0.05 for PRISM) and correlation coefficient (cor <-0.3) as screening criteria (Fig. 5b–e), and KX2-391, a compound shared by CTRP and PRISM, was considered to have great potential for the treatment of high MLPS scores HBV-HCC. From the GSE97098 cohort, 18 HBV‑HCC cell lines were screened, and their corresponding MLPS scores were obtained. The results showed that Huh1 and SNU878 exhibited the highest MLPS scores (Fig. 5f). We then investigated the effects of KX2‑391 on Huh1 and SNU878. The data revealed a dose‑ and time‑dependent response to KX2‑391 treatment (IC₅₀ values: Huh1—18.15 μM at 24 h, 12.21 μM at 48 h, and 7.178 μM at 72 h; SNU878—14.93 μM at 24 h, 7.362 μM at 48 h, and 5.714 μM at 72 h) (Fig. 5g, h). Flow cytometry analysis demonstrated that KX2‑391 could promote apoptosis in both Huh1 and SNU878 cells (Fig. 5i).Fig. 5Screening and validation of small-molecule compounds targeting HBV‑HCC patients with high MLPS scores.a Workflow of compound screening. b–e Correlation and differential analysis of sensitivity of CTRP-derived and PRISM-derived compounds. f MLPS scores of 18 HBV‑HCC cell lines. g, h CCK‑8 assay showing the proliferation of Huh1 and SNU878 cells after KX2‑391 treatment. i Flow cytometry analysis demonstrating apoptosis of Huh1 and SNU878 cells following KX2‑391 treatment. Screening and validation of small-molecule compounds targeting HBV‑HCC patients with high MLPS scores. a Workflow of compound screening. b–e Correlation and differential analysis of sensitivity of CTRP-derived and PRISM-derived compounds. f MLPS scores of 18 HBV‑HCC cell lines. g, h CCK‑8 assay showing the proliferation of Huh1 and SNU878 cells after KX2‑391 treatment. i Flow cytometry analysis demonstrating apoptosis of Huh1 and SNU878 cells following KX2‑391 treatment. In vivo experiments reveal the effect of KX2-391 on high-MLPS score HBV-HCC We established a subcutaneous tumor model of HCC in nude mice using Huh1 cells and treated them with different concentrations of KX2-391 (2.5, 5, and 10 mg/kg) and lenvatinib (10 mg/kg). The results showed that KX2-391 significantly reduced the weight of established tumors, inhibited tumor growth rate, and decreased tumor volume (Fig. 6a–c). Furthermore, the high-dose KX2-391 group (10 mg/kg) exhibited superior inhibitory effects on subcutaneous tumor formation compared to lenvatinib (10 mg/kg) (Fig. 6a–c). No significant differences in body weight were observed among different groups (Fig. 6d). Staining of heart, liver, lung, and kidney tissues from nude mice revealed that neither KX2-391 nor lenvatinib caused significant organ damage (Fig. 6e). Additionally, tumors in the high-dose KX2-391 group had the highest TUNEL-positive rate and the lowest Ki67-positive rate (Fig. 6e). These findings indicate that KX2-391 exhibits potent antitumor activity against HCC in vivo, with the high-dose regimen being more effective than lenvatinib while maintaining good organ safety.Fig. 6In vivo investigation of the effect of KX2-391 on high-MLPS score HBV-HCC.a Representative images of tumor masses. b Tumor weight. c Tumor growth curve. d Mouse body weight. e HE staining of heart, liver, lung, and kidney tissues, and immunofluorescence staining for Ki67 and TUNEL. In vivo investigation of the effect of KX2-391 on high-MLPS score HBV-HCC. a Representative images of tumor masses. b Tumor weight. c Tumor growth curve. d Mouse body weight. e HE staining of heart, liver, lung, and kidney tissues, and immunofluorescence staining for Ki67 and TUNEL. Effects of F9 on HBV-HCC in vivo and in vitro and its mechanism To investigate the effect of F9 on the biological behavior of HBV-HCC cell lines, we overexpressed F9 in Hep3B and MHCC97H, and validated it using qPCR and WB (Fig. 7a, b and Supplementary Fig. 11). CCK-8 assay results showed that both F9 gene overexpression and rF9 treatment significantly inhibited the proliferation ability of Hep3B and MHCC97H cells (Fig. 7c). Flow cytometry analysis revealed that both F9 gene overexpression and rF9 treatment led to an increase in apoptosis levels in Hep3B and MHCC97H cells (Fig. 7d). WB analysis demonstrated that both F9 gene overexpression and rF9 treatment in Hep3B and MHCC97H cells significantly promoted the expression of apoptosis-related proteins and suppressed the expression of anti-apoptotic proteins (Fig. 7e and Supplementary Fig. 11).Fig. 7Effects of F9 on the biological behavior of HBV-HCC and underlying mechanisms.a, b Overexpression of F9 in Hep3B and MHCC97H cells was verified using qPCR and WB. c, d Cell proliferation and apoptosis were assessed using the CCK-8 assay and flow cytometry after F9 overexpression or rF9 treatment in Hep3B and MHCC97H cells. e WB analysis showing the effect of F9 overexpression and rF9 treatment on apoptosis-related protein expression. f GSEA showing differences in enrichment of the Wnt/β-catenin signaling pathway in hallmark terms between high and low F9 expression patients in the TCGA-LIHC cohort. g WB analysis demonstrating the effect of F9 overexpression and rF9 treatment on the Wnt/β-catenin signaling pathway. h Images of tumors isolated from the xenograft model. i, j Tumor weight and volume monitoring over time. k Immunofluorescence images of Ki-67 and TUNEL staining in tumor samples from each group. Effects of F9 on the biological behavior of HBV-HCC and underlying mechanisms. a, b Overexpression of F9 in Hep3B and MHCC97H cells was verified using qPCR and WB. c, d Cell proliferation and apoptosis were assessed using the CCK-8 assay and flow cytometry after F9 overexpression or rF9 treatment in Hep3B and MHCC97H cells. e WB analysis showing the effect of F9 overexpression and rF9 treatment on apoptosis-related protein expression. f GSEA showing differences in enrichment of the Wnt/β-catenin signaling pathway in hallmark terms between high and low F9 expression patients in the TCGA-LIHC cohort. g WB analysis demonstrating the effect of F9 overexpression and rF9 treatment on the Wnt/β-catenin signaling pathway. h Images of tumors isolated from the xenograft model. i, j Tumor weight and volume monitoring over time. k Immunofluorescence images of Ki-67 and TUNEL staining in tumor samples from each group. To further investigate the molecular biological function of F9, the pathway enrichment differences between high-F9 and low-F9 were analyzed by GSEA in the TCGA-LIHC cohort, and the results showed that hallmark_Wnt/β-catenin signaling was significantly enriched in the low-F9 expression patients (Fig. 7f), suggesting that F9 may exert its effects by modulating the Wnt/β-catenin signaling pathway. Subsequently, we validated this finding via WB. The results showed that F9 overexpression and rF9 treatment in Hep3B and MHCC97H significantly increased the level of phosphorylated β-catenin (Ser33/Ser37/Thr41), decreased the levels of β-catenin and phosphorylated GSK3β (Ser9), while leaving the total GSK3β level unchanged, thereby suppressing the Wnt/β-catenin signaling pathway (Fig. 7g and Supplementary Fig. 11). Furthermore, we found that tumor growth was markedly inhibited following F9 overexpression or rF9 treatment (Fig. 7h–j). Immunofluorescence assays demonstrated that F9 overexpression and rF9 treatment significantly suppressed tumor cell proliferation while promoting apoptosis (Fig. 7k). F9 modulates the Wnt/β-catenin signaling pathway in HBV-HCC via SERPINC1 To investigate how F9 regulates the Wnt/β-catenin signaling pathway in HBV-HCC cells, we searched four protein–protein interaction databases—the Database of Interacting Proteins, Integrated Interactions Database, IntAct, and STRING—for proteins potentially interacting with F9. Intersection analysis identified SERPINC1 as the only common interactor (Fig. 8a). We next examined the correlation between F9 and SERPINC1 mRNA and protein levels across multiple cohorts, which revealed a significant positive correlation at both the transcriptional and translational levels (Fig. 8b). Co-immunoprecipitation assays confirmed a direct physical interaction between F9 and SERPINC1 in Hep3B and MHCC97H cells (Fig. 8c and Supplementary Fig. 11). Immunofluorescence staining further demonstrated co-localization of F9 (red) and SERPINC1 (green) in these cell lines (Fig. 8d). WB analysis showed that overexpression of SERPINC1 did not alter F9 expression (Fig. 8e and Supplementary Fig. 11), whereas overexpression of F9 markedly upregulated SERPINC1 (Fig. 8f and Supplementary Figure 11). To explore the impact of F9 and SERPINC1 on the Wnt/β-catenin pathway, we knocked down SERPINC1 in Hep3B and MHCC97H cells (Fig. 8g and Supplementary Fig. 11). This knockdown significantly increased the level of phospho-β-catenin (Ser33/Ser37/Thr41), decreased the levels of β-catenin and phospho-GSK3β (Ser9), and left total GSK3β levels unchanged (Fig. 8h and Supplementary Fig. 11). Finally, simultaneous overexpression of F9 and knockdown of SERPINC1 reversed the Wnt/β-catenin pathway alterations induced by F9 overexpression alone (Fig. 8i and Supplementary Fig. 11). These results demonstrate that F9 interacts with SERPINC1 and modulates its protein expression to regulate the Wnt/β-catenin signaling pathway, thereby playing an important role in HBV-HCC.Fig. 8F9 modulates the Wnt/β-catenin signaling pathway in HBV-HCC through interaction with SERPINC1.a Venn diagram showing SERPINC1 as the only common interactor of F9 identified from four databases. b Correlation analysis of F9 and SERPINC1 mRNA and protein levels across multiple cohorts. c Co-immunoprecipitation assay confirming direct physical interaction between F9 and SERPINC1 in Hep3B and MHCC97H cells. d Immunofluorescence staining showing co-localization of F9 (red) and SERPINC1 (green) in Hep3B and MHCC97H cells. e, f WB analysis of F9 and SERPINC1 expression after SERPINC1 overexpression or F9 overexpression. g Knockdown efficiency of SERPINC1 in Hep3B and MHCC97H cells. h WB analysis showing the effects of SERPINC1 knockdown on the Wnt/β-catenin pathway. i The impact of simultaneous F9 overexpression and SERPINC1 knockdown on the Wnt/β-catenin signaling pathway. F9 modulates the Wnt/β-catenin signaling pathway in HBV-HCC through interaction with SERPINC1. a Venn diagram showing SERPINC1 as the only common interactor of F9 identified from four databases. b Correlation analysis of F9 and SERPINC1 mRNA and protein levels across multiple cohorts. c Co-immunoprecipitation assay confirming direct physical interaction between F9 and SERPINC1 in Hep3B and MHCC97H cells. d Immunofluorescence staining showing co-localization of F9 (red) and SERPINC1 (green) in Hep3B and MHCC97H cells. e, f WB analysis of F9 and SERPINC1 expression after SERPINC1 overexpression or F9 overexpression. g Knockdown efficiency of SERPINC1 in Hep3B and MHCC97H cells. h WB analysis showing the effects of SERPINC1 knockdown on the Wnt/β-catenin pathway. i The impact of simultaneous F9 overexpression and SERPINC1 knockdown on the Wnt/β-catenin signaling pathway. Discussion Primary liver cancer is a disease with high incidence, high mortality, and poor prognosis24,25. HBV-HCC constitutes a major proportion of primary liver cancer and leads to a serious burden worldwide. Despite the emergence of targeted therapies and immunotherapy, the prognosis of HBV-HCC patients remains poor and heterogeneous. In this study, multi-omics data analysis identified two distinct HBV-HCC cancer subtypes with significant differences in genetic and clinical features. A machine learning prognostic signature (MLPS) was subsequently constructed to accurately predict the prognosis of HBV-HCC. We further discovered that F9 acts as a therapeutic target in HBV-HCC, regulating the Wnt/β-catenin signaling pathway through interaction with SERPINC1, thereby exerting its function in liver cancer. Finally, in vitro and in vivo experiments confirmed that KX2-391 is a promising antitumor agent for HBV-HCC patients with high MLPS scores. Zhang et al. classified HBV‑HCC into two subtypes by non‑negative matrix factorization based on transcriptomic immune scores; the subtypes differed markedly in clinical features, pathway enrichment, and genomic variation, but single‑omics classification may not fully capture HBV‑HCC heterogeneity26. To avoid the limitations of a single omics, we used ten clustering algorithms based on multi-omics data to investigate the internal heterogeneity of HBV-HCC, which resulted in the acquisition of two CSs with significant molecular and clinical features. Several genes have high-frequency SNVs in CS2, including CTNNB1, OBSCN, APOB, etc. The CTNNB1 gene encodes β-catenin, a major oncogene, which is mutated in a variety of cancers, including melanoma and HCC27. β-catenin-mutated HCC lacks an immune infiltrate and is considered a “cold tumor” that is resistant to immunotherapy. Mutated β-catenin in HCC suppressed the expression of SDC4 and RAB27A and inhibited exosome release, and thus inhibited the infiltration of immune cells27. OBSCN encodes the giant multifunctional protein OBSCURIN, which is mainly expressed in the heart and skeletal muscle. OBSCURIN can participate in the coordination of tumor cells as a tumor suppressor, preventing the cells from entering the process of epithelial-mesenchymal transition, thus inhibiting tumor cell metastasis and carcinogenesis28. We constructed an MLPS based on a 10-gene panel, including F9, IGFALS, TKT, etc., that can predict the prognosis of HBV-HCC. F9 is located on the X chromosome and encodes the vitamin K-dependent coagulation factor IX, and mutations in this gene lead to hemophilia B29. However, with further research, it has been found that F9 is associated with tumorigenesis, progression, and drug resistance. The first epidermal growth factor motif of coagulation factor IX attenuates the adhesion of oral squamous cell carcinoma cells to the extracellular matrix and activates apoptosis in a caspase-3-dependent manner30. Moreover, F9 is a key gene responsible for CDK4/6 inhibitor-induced cell cycle arrest and senescence phenotypes, whereas knockdown of F9 prevents the induction of senescence31. Studies have shown that a high level of F9 expression is a good marker for survival prognosis of BRCA32, and a low level of F9 was found to be associated with liver metastasis in colorectal cancer33. However, the role of F9 in HBV-HCC has not been reported. In this study, we found that overexpression of F9 could significantly inhibit the malignant phenotype of HCC cell lines, and the overexpression of F9 could inhibit the activation of the Wnt/β-catenin signaling pathway. The Wnt/β-catenin signaling pathway is critical for liver function, including hepatobiliary development and repair processes. Although the levels of β-catenin are low in normal hepatocytes from healthy livers and β-catenin is predominantly located in the membrane, during the development of HCC, its levels in the cytoplasm and nucleus are elevated, suggesting that typical Wnt/β-catenin signaling is reactivated34. Subsequently, we confirmed the interaction between F9 and SERPINC1 through co-immunoprecipitation and immunofluorescence assays, and demonstrated that F9 regulates the Wnt/β-catenin signaling pathway via SERPINC1. SERPINC1 encodes antithrombin, a member of the serine protease inhibitor family, and plays a central role in hemostatic regulation35. With further research, SERPINC1 has also been implicated in tumorigenesis and development. In nasopharyngeal carcinoma, SERPINC1 was found to be significantly upregulated in tumor tissues, and knockdown of SERPINC1 promoted apoptosis of nasopharyngeal carcinoma cells and inhibited the PI3K/Akt/mTOR signaling pathway36. Studies on colorectal cancer revealed that SERPINC1 was significantly overexpressed in cases with aggressive clinicopathological features (higher T/N stage, vascular invasion, elevated CEA and CA19-9 levels, and therapeutic resistance), and high SERPINC1 expression was correlated with poor patient prognosis. Moreover, knockdown of SERPINC1 markedly inhibited the growth and promoted apoptosis of colorectal cancer cells, potentially through TGF-β-mediated epithelial-mesenchymal transition37. In lung cancer, overexpression of SERPINC1 promoted proliferation, migration, and invasion of lung cancer cells in vitro while suppressing apoptosis, possibly via regulation of related factors through the PI3K/AKT pathway38. However, the precise mechanistic link between SERPINC1 and the Wnt/β-catenin pathway in HCC had not been clearly defined prior to our work. Our study provides new insights by demonstrating that SERPINC1 interacts with F9, and together they regulate the activation status of the Wnt/β-catenin signaling axis, ultimately impacting the malignant phenotype of HBV-HCC cells. This finding expands the functional repertoire of SERPINC1 beyond its classical role and positions it as a potential modulator of key oncogenic pathways in liver carcinogenesis. The identification of therapeutic drugs specifically designed for molecularly defined subsets of HCC is crucial for advancing the field of precision oncology. In this study, we employed a dual connectivity mapping strategy, utilizing the CTRP and PRISM databases, to screen for compounds with potential efficacy against HBV-HCC patients exhibiting high MLPS scores—a subgroup characterized by aggressive behavior and poor prognosis. Among the candidate compounds, KX2-391 emerged as the only molecule common to both datasets, identified based on stringent screening criteria. Both in vitro and in vivo experiments demonstrated that KX2-391 exerts significant antitumor effects against high-MLPS HBV-HCC in a dose-dependent manner. Notably, in the subcutaneous xenograft mouse model, the high-dose group (10 mg/kg) exhibited superior inhibition of tumor formation compared to the clinically approved multi-kinase inhibitor lenvatinib (10 mg/kg). Importantly, KX2-391 treatment did not induce systemic toxicity, as evidenced by stable body weight and the absence of pathological abnormalities in the heart, liver, lung, and kidney tissues. KX2-391 primarily targets cytoplasmic tyrosine kinases, and emerging evidence indicates that Src inhibition can modulate the Wnt/β-catenin signaling pathway—a key driver of HCC pathogenesis—and suppress tumor growth in HCC models39,40. The clinical significance of our findings lies in the potential repurposing of KX2-391 for the treatment of high-MLPS HBV-HCC patients. Stratifying patients according to MLPS ensures that KX2-391 is administered to those most likely to benefit, addressing the unmet need for personalized therapy in HCC. Moreover, the superior efficacy of KX2-391 over lenvatinib in our preclinical model highlights its promise as an alternative or complementary treatment option for patients who are refractory to current standard therapies. In conclusion, KX2-391 demonstrates potent antitumor activity against high-MLPS HBV-HCC both in vitro and in vivo, with a favorable safety profile. Its identification through comprehensive drug screening, combined with robust pharmacological validation, supports its prioritization for further translational development. Future studies should focus on exploring biomarkers to guide patient selection, optimizing dosing regimens, and evaluating combination strategies to maximize clinical outcomes. The TME in HCC is a complex and dynamic ecosystem that is critical for understanding disease progression and response to various treatments41. Our study was focused on HBV-HCC and revealed the complex interactions of the TME within it, with a special emphasis on patients with high-MLPS score (and CS1) who were classified as “hot tumors” and patients with low-MLPS score (and CS2) who were identified as “cold tumors.” This not only provides insight into tumor behavior but also has important implications for HBV-HCC outcomes and treatment strategies. In addition, different TMEs may be one of the main reasons for the prognostic differences between patients with different phenotypes. Multiple immune cells in the TME work together to maintain homeostasis. But a variety of immune cells play immunosuppressive roles, including myeloid-derived suppressor cells, tumor-associated macrophages, tumor-associated neutrophils, tolerogenic dendritic cells, regulatory T cells, etc., which not only promote tumorigenesis but also mediate tumor progression, metastasis, and even treatment resistance42. Immune cells do not only interact with each other, but also tumor cells, and immune cells can interact with each other, affecting the function of immune cells, such as inhibiting their killing activity, inducing immune tolerance, or promoting inflammatory response, which contributes to the survival and colonization of tumor cells43–45. Several limitations should be acknowledged. Our molecular subtyping was based on the TCGA‑LIHC cohort, utilizing transcriptome, DNA methylation, SNV, and CNV data. However, this public resource lacks proteomic and metabolomic measurements; thus, our analysis was confined to a limited set of omics layers. This constraint may preclude capturing critical aspects of HBV‑HCC biology, such as post‑translational modifications of oncoproteins and metabolic reprogramming at the enzymatic level, potentially leading to an incomplete representation of disease heterogeneity. Another limitation is the lack of validation in real‑world HBV‑HCC cohorts. Although our model showed robust performance across five cohorts, real‑world populations often exhibit greater variability in comorbidities, treatment histories, and specimen handling, which may affect generalizability. To address this gap, we plan to establish collaborations with international medical centers to access real‑world HBV‑HCC cohorts with linked multi‑omics and longitudinal clinical data. Prospective studies with standardized omics profiling across diverse ethnicities and regions are also warranted. Furthermore, we aim to construct an open‑access HBV‑HCC repository to facilitate continuous external validation and iterative model refinement. These efforts will enhance the clinical applicability of our findings. From a clinical perspective, although MLPS shows strong prognostic performance, its implementation feasibility needs consideration. The model requires only the mRNA expression of 10 genes, which can be measured in routinely available FFPE or fresh tumor samples using multiplex qRT‑PCR or small‑panel RNA‑seq. Sample accessibility is high, and the cost is low. The scoring algorithm can be integrated into existing pathology workflows with simple calculators or web tools, without requiring specialized bioinformatics support, facilitating risk stratification in multidisciplinary meetings. However, inter‑platform variability should be minimized by standardized protocols, and prospective real‑world validation and cost‑effectiveness analysis are needed for full clinical adoption. In summary, comprehensive multi-omics analysis of HBV-HCC revealed two distinct cancer subtypes that differ markedly in genomic and immunological profiles, as well as clinical characteristics. Based on these findings, we developed the MLPS to accurately predict prognosis in HBV-HCC patients, demonstrating broad applicability and superior performance. Through a series of basic experimental studies, we identified and validated F9 as a potential therapeutic target in HBV-HCC. Mechanistically, F9 functions in HBV-HCC by interacting with SERPINC1 to modulate the activation of the Wnt/β-catenin signaling pathway. Finally, we identified and validated KX2-391 as a promising therapeutic agent for HBV-HCC, particularly for patients with high MLPS scores. Methods Data download and processing Multi-omics data from the TCGA liver hepatocellular carcinoma (LIHC) cohort, including transcriptomic profiles, DNA methylation CpG sites, copy number variation (CNV), single nucleotide variation (SNV), and clinical information, were obtained from the TCGA-GDC portal (https://portal.gdc.cancer.gov/). A total of 99 HBV-HCC patients from the TCGA-LIHC cohort with all-omics data and prognostic information were screened, and transcriptomic profiles of non-tumor tissues were also obtained. RNA transcriptomic and clinical information were obtained from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) for the GSE14520 and GSE144269 cohorts, and the number of HBV-HCC patients screened with prognostic information was 212 and 49, respectively. International Cancer Genome Consortium (ICGC) (https://dcc.icgc.org/) is a database dedicated to an in-depth understanding of the genetic variation and molecular mechanisms of cancer. We included 57 HBV-HCC patients with prognostic information from the Liver Cancer-RIKEN, Japan Project of International Cancer Genome Consortium (ICGC-LIRI-JP) cohort46. In addition, 159 Chinese hepatocellular carcinoma patients with Hepatitis B virus infection (CHCC-HBV) from Zhongshan Hospital of Fudan University were got from NODE (https://www.biosino.org/node)47. Utilize the ComBat function in the sva package to perform correction, removing batch differences caused by sequencing platforms/batches48. The TCGA-LIHC cohort was used to internally identify HBV-HCC cancer subtypes (CSs), and the GSE14520, GSE144269, ICGC-LIRI-JP, and CHCC-HBV cohorts were used for external validation. In addition, we downloaded the RNA‑seq dataset GSE97098 from the GEO, which contains 81 liver cancer cell lines, and selected 18 HBV‑HCC cell lines for subsequent analyses49. Identification of consensus CSs of HBV-HCC in TCGA-LIHC cohort based on multi-omics data Consensus CSs were determined using multi-omics data from 99 HBV-HCC patients in the TCGA-LIHC cohort, according to the guidelines of the R-project “MOVICS” package50. First, for mRNA, lncRNA, miRNA, and DNA methylation CpG sites, the “method” parameter of the “getElites” function was set to “sd” for the HBV-HCC patients of the TCGA-LIHC cohort. The parameter “method” of the “getElites” function was set to “sd” to filter the top 20% of genes with the highest degree of variation. Subsequently, univariate Cox regression was used to assess the OS-related factors, and p-value < 0.05 was used as the threshold for screening. For mutation data, the “method” parameter was set to “freq” to identify genes with a mutation frequency higher than 10%. In order to find the appropriate number of CSs, the clustering prediction index (CPI) and Gap-statistics were evaluated based on multi-omics data. Subsequently, multiple clustering CS results were obtained using ten clustering algorithms, including iClusterBayes, SNF, PINSPlus, NEMO, COCA, LRAcluster, ConsensusClustering, IntNMF, CIMLR, and MoCluster, respectively. The getConsensusMOIC function was used to integrate the results of the ten clustering algorithms and ultimately generate a robust CS, and the similarity of samples in the CSs was quantified using silhouette scores. Nearest template prediction validation Using the R-project “limma” package, 50 specific biomarkers for each consensus CS were identified. Subsequently, based on the specific biomarkers, nearest template prediction (NTP) was used to evaluate CS in external cohorts. Characterization of genomic variation Tumor mutation burden (TMB) was calculated for each sample from SNV data and visualized using the R-project “maftools” and the “ComplexHeatmap” package. For CNV data, significantly amplified and deleted genomic regions were identified using the Genomic Identification of Significant Targets in Cancer_2.0 (GISTIC_2.0) module of GenePattern (https://cloud.genepattern.org/gp/pages/index.jsf). Construction of a machine learning driver prognostic signature Differentially expressed genes (DEGs) between non-tumor tissues and HBV-HCC tumor tissues in the TCGA-LIHC cohort were screened according to log2|Foldchange (FC) | ≥ 1 and adjusted p-value < 0.05. Subsequently, DEGs, the CS-specific up-regulated biomarkers obtained above, and the shared genes of the five cohorts were taken for intersection to obtain intersecting genes. Then, ten machine learning algorithms (including CoxBoost, elastic network (Enet), generalized boosted regression modeling (GBM), least absolute shrinkage and selection operator (Lasso), partial least squares regression for Cox (plsRcox), Ridge, random survival forest (RSF), StepCox, supervised principal components (SuperPC) and survival support vector machine (survival-SVM)) and/or their combinations algorithms were used to construct machine learning-driver prognostic signatures (MLPSs) for predicting OS in HBV-HCC’ patients using the TCGA-LIHC cohort as the training set, and the GSE14520, GSE144269, ICGC-LIRI-JP, and CHCC-HBV cohorts as the testing set, according to the previous method51. The algorithm/algorithm combination corresponding to the highest mean concordance index (C-index) among the five cohorts was considered the optimal method. Predictive capability and robustness of MLPS The optimal algorithms/algorithm combination used to construct the MLPS was previously screened to obtain MLPS scores for each HBV-HCC patient of the five cohorts, who were subsequently divided into high-MLPS and low-MLPS groups based on median values. Kaplan-Meier (KM) curves were used to assess prognostic differences in patients with different phenotypes. Cox regression analyses were used to investigate the prognostic impact of MLPS on prognosis. Prognostic indicators included overall survival (OS), disease-specific survival (DSS), progression-free interval (PFI), and disease-free interval (DFI). Then, we systematically retrieved ten prognostic models related to HBV-HCC (Supplement Table 2) and compared their predictive power between them and the MLPS we developed using the C-index52–61. Biological and immunological characteristics analysis Hallmark and Kyoto Encyclopedia of Genes and Genomes (KEGG) terms were downloaded from MSigDB (https://www.gsea-msigdb.org/gsea/index.jsp), and gene set enrichment scores were calculated for each sample using gene set variation analysis (GSVA). A variety of previously published features associated with tumor microenvironment (TME) cell types, immunotherapy responses, immune suppression, immune exclusion, and immunotherapy biomarkers were collected and quantified using the Immuno-Oncology Biological Research (IOBR) package62. Compound screening for targeted therapy in HBV-HCC patients with high MLPS scores Expression data on human tumor cell lines were obtained from the Broad Institute Cancer Cell Line Encyclopedia (CCLE) (https://sites.broadinstitute.org/ccle/). Cancer Therapeutics Response Portal (CTRP) (http://portals.broadinstitute.org/ctrp.v2.1/) and Profiling Relative Inhibition Simultaneously in Mixtures (PRISM) (https://depmap.org/portal/prism/) were used to predict the sensitivity of each sample to treatment with each compound, and AUC values were used to evaluate sensitivity63. Cell lines and reagents The human hepatocellular carcinoma HBV-associated cell lines Hep3B, MHCC97H, Huh1, and SNU878 were cultured in Minimum Essential Medium/Dulbecco’s Modified Eagle Medium/Roswell Park Memorial Institute Medium 1640 (Gibco, USA) containing 1% penicillin-streptomycin (Servicebio, China) and 10% fetal bovine serum (ExCell Bio, China) at 37°C in a 5% CO2 incubator. Cell lines were identified by short tandem repeat. KX2-391, Lenvatinib, and recombinant F9 protein (rF9) were used according to the instructions provided by their respective manufacturers. Lentivirus and plasmid The coding sequences of F9 and SERPINC1 were retrieved from the NCBI database and separately cloned into lentiviral expression vectors to achieve their respective overexpression, with an empty vector serving as the negative control. To knock down SERPINC1, small interfering RNA targeting SERPINC1 (si-SERPINC1) was employed (sequence: 5'-AUCACAUUGGAAUACAUGGCC-3'), along with a negative control siRNA (sequence: 5'-CAUGUGGUCUGUCGCAUAAUA-3'). All transfection procedures were performed following the manufacturer’s protocols. Cell counting kit-8 assay Cells were inoculated in 96-well plates (5*103 cells/well), and then cell viability was assessed using the cell counting kit-8 (CCK-8) kit (APExBIO, USA) (10 μl/well). Wound healing assay Cells were inoculated in 6-well plates (2*105 cells/well), scratches were made, and the healing of scratches was photographed after 48 h. Cell apoptosis analysis Cells were inoculated into 6-well plates (2*105 cells/well) and subsequently resuspended in 500 μL of 1x binding buffer (Bestbio, China), followed by the addition of 5 μL of Annexin V-FITC and 5 μL of propidium iodide, and incubated for 20 min at room temperature and finally detected using a flow cytometer. EdU assay Cells (2*104 cells/well) were added to the EdU working solution (APExBIO, USA) and incubated for 2 h. 1 mL of 4% paraformaldehyde was added and fixed for 15 min. Then 1 mL of PBS containing 0.3% Triton X-100 was added and incubated for 15 min. 500 μL of reaction solution was immediately added and incubated away from light for 30 min. 1 mL of 1× Hoechst 33342 was added and incubated away from light for 15 min. Finally, photographs were taken and analyzed using fluorescence microscopy. Western blot Cells were lysed in lysis buffer (Biosharp, China), and proteins were extracted and separated on SDS/PAGE gels and then transferred to PVDF membranes (Millipore, USA). Subsequently, they were sequentially incubated with appropriate dilutions of primary and secondary antibodies and visualized. Antibodies and dilutions used are listed below: GAPDH (Proteintech, 1:10,000), β-actin (Proteintech, 1:10,000), β-tubulin (Proteintech, 1:10,000), F9 (Abcam, 1:1000), SERPINC1 (Abcam, 1:1000), BAX (Proteintech, 1:5000), BCL-2 (Proteintech, 1:5000), β-catenin (Proteintech, 1:5000), phospho-β-catenin (Ser33/37/Thr41) (Cell Signaling Technology, 1:5000), GSK3β (Proteintech, 1:5000), phospho-GSK3β (Ser9) (Proteintech, 1:5000), caspase-3 (Cell Signaling Technology, 1:5000), PARP (Ser9) (Cell Signaling Technology, 1:5000). Quantitative real-time polymerase chain reaction Total cellular RNA was extracted using Trizol (Servicebio, China) according to the manufacturer’s instructions and then reverse transcribed to cDNA using SweScript RT I First Strand cDNA Synthesis Kit (Servicebio, China). Finally, quantitative real-time polymerase chain reaction (qPCR) was performed using HotStart™ Universal 2× FAST SYBR Green qPCR Master Mix (APExBIO, USA, Cat#K1172). The mRNA expression levels of the target genes were determined using GAPDH as an internal reference and evaluated by 2−ΔΔCt. The primer sequences were F9-F: 5'-GATGCATTCTGTGGAGGCTCTAT-3', F9-R: 5'-AGAAGGGCAATGTCATGGTTGTA-3', GAPDH-F: 5'-GGAGTCCACTGGCGTCTTCA-3', GAPDH-R: 5'-GTCATGAGTCCTTCCACGATACC-3'. Prediction of potential interacting proteins for F9 Predict potential interacting proteins that may associate with F9 through STRING (https://cn.string-db.org/), the Database of Interacting Proteins (http://dip.doe-mbi.ucla.edu/), IntAct (https://www.ebi.ac.uk/intact/), and the Integrated Interactions Database (https://iid.ophid.utoronto.ca/search_by_proteins/). Immunoprecipitation (IP) Lyse the cells and quantify the lysate, then incubate the antibody with ProteinA/G magnetic beads to allow binding, followed by incubation with the protein sample so that the immune complexes bind to the beads. Wash and elute the beads, then proceed with subsequent analysis. Xenograft models All institutional and national guidelines for the care and use of laboratory animals were followed. Animal experiments were performed with ethical approval for research involving animals of The Second Hospital of Lanzhou University (D2024-856). The xenograft model used 4–5-week-old immunodeficient BALB/c nude mice produced by Hefei Qingyuan Biotechnology Company. Mice were housed in a specific pathogen-free environment, with room temperature maintained between 20 °C and 26 °C, relative humidity controlled at 40%–70%, and under conditions of adequate ventilation and appropriate lighting. Animals were allocated to groups using computer-generated random numbers. Nude mice were injected with the cell line (5 × 10⁶ cells per mouse) subcutaneously in the upper back and euthanized after 3 weeks. Tumor size was measured with a vernier caliper. Tumor volume was calculated using the following formula: volume = 0.5 * length * width2. Euthanasia was performed via intraperitoneal injection of sodium pentobarbital at a dosage of 100 mg/kg body weight. The depth of anesthesia was assessed by monitoring the loss of hindlimb withdrawal reflex. Following euthanasia, tumor tissues were collected for subsequent experimental analyses. Immunohistochemistry and immunofluorescence Tumor tissues from xenograft models were collected and fixed in 4% paraformaldehyde (Servicebio, G1101), then paraffin-embedded and cut into 5μm thick sections. Immunohistochemical staining was performed with horseradish peroxidase/DAB (Servicebio, G1212) according to the instructions. Sequentially, dewaxing, antigen repair, blocking endogenous peroxidase, and serum closure were performed. This was followed by incubation with anti-Ki-67 (Proteintech, 1:5000) primary and secondary antibodies. Finally, the results were observed under a microscope, and images were acquired. The TMR (Red) Tunel cell apoptosis detection kit (Servicebio, G1502) was used for immunofluorescence experiments according to the instructions. After deparaffinization, proteinase K repair, and membrane rupture, the reaction solution was added, and the nuclei were restained with DAPI. Finally, the results were observed under a microscope, and images were acquired. In short, the steps for immunofluorescence double-labeling are as follows: cells were evenly seeded in a cell culture plate, and then the permeabilization working solution, primary antibodies, secondary antibodies, and DAPI staining solution were added sequentially. After the slide was gently blotted dry, it was mounted with an anti-fade mounting medium. Finally, images were captured. Statistical analysis R-project (version: 4.3.1) and GraphPad Prism (version: 9.5.0) were used for data analysis and visualization. Correlations between variables were analyzed using Spearman’s correlation analysis. p < 0.05 represents statistical significance, and “ns”, “*”, “**”, “***”, “****” represent p > 0.05, p < 0.05, p < 0.01, p < 0.001, and p < 0.0001, respectively. Ethics approval The experiments were performed with ethical approval for research involving animals from The Second Hospital of Lanzhou University (D2024-856). Supplementary information Supplementary information Supplementary information