Identification and prognostic analysis of genes related to CTNNB1 mutations in hepatocellular carcinoma Authors: Ningqi Zhu, Guo Zhang, Xiaotian Shen, Zule Chen, Ying Li, Diyu Chen, Huliang Jia Published: 2026 ABSTRACT BackgroundMutations in CTNNB1 are recognized oncogenic drivers of hepatocellular carcinoma (HCC); however, the downstream effector molecules and their prognostic significance remain incompletely defined. In this study, we developed a mutant CTNNB1 gene signature (MCGS) to predict CTNNB1 mutation status and to evaluate the prognostic relevance of the constituent genes.MethodsHydrodynamic tail vein injection mouse models of HCC were generated by co-expressing Sleeping Beauty transposon/transposase and mutant CTNNB1 with either AKT-c-Myc or AKT-NRASG12V, followed by RNA sequencing of tumor tissues. Cross-species differential expression analyses were performed by integrating data from The Cancer Genome Atlas (TCGA) and the International Cancer Genome Consortium (ICGC). Machine-learning models were trained to identify MCGS. The prognostic value of signature genes was validated using public transcriptomic datasets and immunohistochemistry on an institutional tissue microarray (TMA) cohort.ResultsMutant CTNNB1 markedly accelerated hepatocarcinogenesis in mice, leading to earlier tumor development and reduced survival. Transcriptomic profiling revealed enrichment of the Hippo and MAPK signaling pathways in CTNNB1-mutant HCCs. We established a 15-gene MCGS using a Categorical Boosting algorithm, which accurately predicted CTNNB1 mutation status (AUC = 0.957 in TCGA; 0.893 in ICGC). Among these genes, metallothionein-3 (MT3) emerged as a prognostic marker, with high expression independently associated with poorer overall survival in public datasets. Consistently, elevated MT3 protein expression on the TMA was also independently associated with worse overall and recurrence-free survival in our institutional cohort.ConclusionWe identified a robust mutant CTNNB1 gene signature capable of accurately predicting CTNNB1 mutation status, and uncovered MT3 as a novel, clinically relevant prognostic biomarker in HCC.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13062-026-00873-6. Introduction Liver cancer is currently the sixth most common malignancy and the third leading cause of cancer-related mortality worldwide [1]. Among liver cancers, hepatocellular carcinoma (HCC) accounts for approximately 90% of cases [2]. Although immune checkpoint blockade (ICB) therapy has achieved remarkable success across various tumor types, only about 15% of HCC patients exhibit sensitivity to ICB [3, 4]. HCC displays considerable genetic heterogeneity, with mutations in the CTNNB1 gene found in approximately 26–37% of cases [5]. These mutations often result in constitutive activation of the Wnt signaling pathway. They are primarily located in exon 3, affecting serine/threonine residues or adjacent sites, thereby conferring a gain-of-function phenotype. The mutant β-catenin protein lacks key phosphorylation and ubiquitination sites required for degradation, leading to its accumulation in the cytoplasm and subsequent translocation into the nucleus. Once inside the nucleus, it forms complexes with TCF/LEF transcription factors, driving aberrant transcriptional activation or repression of downstream target genes [6–8]. Multiple studies have demonstrated that mutant β-catenin significantly influences tumor initiation, differentiation, growth, proliferation, metabolism, and the immune microenvironment [9]. CTNNB1 gain-of-function mutations are strongly associated with the development of an immunosuppressive tumor microenvironment, which may explain the poor responsiveness of this HCC subtype to immunotherapy [9, 10]. In recent years, inhibition of the Wnt/β-catenin pathway has been widely investigated as a potential therapeutic approach for HCC [10]. However, its clinical efficacy remains limited, primarily due to substantial on-target toxicity when directly inhibiting the β-catenin transcriptional machinery in normal cells. Consequently, direct targeting of core pathway components has proven to be a suboptimal clinical strategy [11]. This highlights the urgent need to discover downstream or pathway-associated genes that are functionally relevant to tumor prognosis and therapeutic resistance, and to evaluate their potential as druggable targets. In the present study, we established a CTNNB1 gain-of-function (CTNNB1GOF) mutant mouse model of HCC and a corresponding control using the hydrodynamic tail vein injection (HTVi) method. RNA sequencing (RNA-seq) was then performed to identify differentially expressed genes between the mutant and control groups. By integrating these results with RNA-seq datasets from The Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC), the International Cancer Genome Consortium Liver Cancer - RIKEN, JP (ICGC-LIRI-JP), National Omics Data Encyclopedia (NODE), and the Gene Expression Omnibus (GEO) public databases, we identified a distinct gene signature associated with CTNNB1 mutation and systematically evaluated its diagnostic and prognostic potential. Materials and methods Patients and specimens 147 adult patients with HCC were enrolled. HCC and matched tumor-adjacent tissue samples were collected from Huashan Hospital, Fudan University between January 2020 and December 2021. All patients underwent surgical resection and were pathologically diagnosed of HCC. Clinical characteristics of the patients, including age, sex, tumor size, tumor number, degree of tumor differentiation, TMN stage, BCLC stage, laboratory data, and follow-up data were collected. All patients were followed up until the end of August 2025. The primary endpoints of the study were OS and RFS. OS was calculated as the duration between the date of surgery and the date of death or the date of last follow-up, while RFS was calculated as the duration between date of surgery and the date of first confirmed tumor recurrence or the date of last follow-up. This study was approved by the Ethics Committee of Huashan Hospital, Fudan University (Approval No. KY2024-1113). Verbal informed consent was obtained from all individual participants included in the study. Considering the observational and retrospective nature of this study, the ethics committee approved the use of telephone-based verbal informed consent. Animal studies All animal studies were approved by The Animal Welfare and Ethical Committee of Shanghai Rat&Mouse Biotech Co.,Ltd (Approval No. 20250304(12)). Specific pathogen-free (SPF) animals were obtained from GemPharmatech (Jiangsu, China) and maintained under environmentally controlled conditions with a 12/12-hour light‒dark cycle at 22 ± 2 °C and a relative humidity of 50 ± 10%. Prior to experiment, the mice were randomly assigned via computer-generated randomization method. 6-week-old male C57BL/6J mice (n = 10 / group) were used for for hydrodynamic tail vein injection (HTVi) of HCC models. For each mouse, pT3-N90-β-catenin (15 µg) / pT3-vector (15 µg), pT3-EF1aH-myr-AKT-MYC (25 µg) / pT3-EF1aH-myr-AKT-NRASG12V(25 µg) and pCMV(CAT)T7-SB100 (10ug) were dissolved in 2 mL sterile 0.9% saline and injected via mouse tail vein in 5–8 s. Mice were manually restrained during injection and monitored carefully for signs of distress immediately after injection and daily thereafter. Mice were monitored daily and euthanized when they developed large abdominal masses and become moribund. Mice in another repeat of the experiment were euthanized at the 18-day after injection time point. Body and liver weights were measured at time of tissue harvesting. The experiment was repeated 3 time, and the total number of mice was 120 (In 2 experiments, mice were euthanized when they became moribund. In a separate experiment, mice were euthanized at the 18-day after injection time point). Animal euthanasia was performed by exsanguination under anesthesia with 500ul Avertin (1.25%) intraperitoneal injection. Avertin was purchased from Tigergene Technology LLC (Jiangsu, China). Plasmids and constructions c-Myc coding (1317 bp) sequence and NRASG12V coding sequence (600 bp) was constructed into pT3-myr-AKT (addgene: #179909), with P2A sequence between them. pT3-myr-AKT (addgene: #179909), pT3-N90-β-catenin (addgene: #31785), pCMV(CAT)T7-SB100 (addgene: #34879), pT3-vector (addgene: #180149) was purchased from Miaoling Biotech Corp.,Ltd, and Plasmid construction service was provided by OBiO Technology (Shanghai, China) Corp.,Ltd. Library preparation and RNA-seq analysis Total RNA was extracted from tissues using the RNA Easy Fast Tissue/Cell kit (#DP451, TIANGEN, China). A NanoDrop 2000 spectrophotometer (Thermo Scientific, USA) was applied for RNA purity measurement and quantification. Sequencing libraries were generated using the TruSeq Stranded mRNA LT Sample Prep Kit (Illumina, USA) following the manufacturer’s recommendations. Prepared libraries were sequenced on an Illumina Novaseq platform. Sequence alignment for each sample was performed against the reference genome of the sequenced species using STAR (version 2.5.3a) with the two-pass mapping mode. The quality of the alignment output was assessed using the RSeQC (version v2.6.4) software package. The abundances of transcripts (including mRNAs, pseudogenes, noncoding RNAs, and other predicted RNAs) were calculated and normalized in fragments per kilobase of transcript per million mapped reads. Immunohistochemistry and analysis The paraffin-embedded tissues were cut to 4 μm thickness sections. The paraffin sections then went through steps including deparaffinization, rehydration, hydrogen peroxide treatment, antigen retrieval, and blocking. Then tissue slides were reserved with specific dilutions of the antibodies against β-catenin (F0487, Selleck, USA), HA-tag (51064-2-AP, Proteintech, China), MT3 (12179-1-AP, Proteintech, China) for 8 h at 4 ◦C. The slides were then incubated with horseradish peroxidase-linked secondary antibody. The IHC staining score was calculated according to the staining intensity and the percentage of positive cells. The rules of the staining intensity scoring were as follows: 0 point (Negative); 1 point (Low, light brown); 2 points (Medium, brown); 3 points (High, dark brown). The extent score was rated according to the extent of positive staining: 0, no positive cells; 1, positive cells ≤ 25%; 2, positive cells 26%-50%; 3, positive cells 51%-75%; and 4, positive cells ≥ 76%. The final immunohistochemical score is the product of staining intensity and extent score (score range: 0–12). If the final score of samples was less than 7, the expression of MT3 was considered low expression, while the score of 7 or higher was considered high expression. Bioinformatics analysis RNA expression profiles, gene mutation profiles [.maf (Mutation Annotation Format) files] and clinicopathological characteristics from TCGA database and ICGC database were retrieved from The Cancer Genome Atlas (http://gdac.broadinstitute.org/), and the International Cancer Genome Consortium (https://icgc.org/). GSE14520, GSE76427 were sourced from the GEO database. OEP000321 was sourced from the NODE database. The R package “clusterProfile” was used for GSEA and KEGG analysis. The R package “DEseq2” was used for differential gene expression analysis. The R package “Maftools” was used for tumor somatic mutation profiles analysis. The R package “caret” was used for construction and validation of machine learning models including: Random Forest, Gradient Boosting, Support Vector Machine (SVM), Logistic Regression, k-Nearest Neighbors Method, Partial Least Squares (PLS) Model, Boosting Method, Neural Networks, Naive Bayes, Discriminant Analysis, least absolute shrinkage and selection operator (Lasso) Regression, Adaptive Boosting, and R package “catboost” and “lightgbm” for Categorical Boosting (CATboost) and Light Gradient Boosting Machine (LightGBM). SHAP analysis based on the CATboost model was performed with R package “shapviz” and “kernelshap”. RNA-seq expression profiles [log2(FPKM + 1)] from TCGA-LIHC were used as the training cohort and those from ICGC-LIRI as the independent validation cohort. 14 machine-learning algorithms mentioned above were implemented. Model tuning employed nested five-fold cross-validation. Evaluation metrics included AUC, sensitivity, specificity, accuracy, positive predictive value, negative predictive value, F1-score, and Youden’s index. The final optimal model was retrained on the full TCGA-LIHC RNA-seq dataset and validated on ICGC-LIRI. Model robustness was evaluated through bootstrap confidence intervals of AUC and decision-curve, calibration, and SHAP value analyses to confirm interpretability and stability. Survival analyses were conducted in R. Univariate Cox proportional hazards models were fitted using the “survival” package to identify prognostic genes (p < 0.05). Significant genes and clinical covariates were included in multivariate Cox regression to assess independent prognostic value. The optimal cutoff for high and low transcriptional gene expression was determined with the “maxstat” package (smethod = “LogRank”, pmethod = “exactGauss”). Kaplan-Meier survival curves were visualized with “survminer” package. Statistical analysis R 4.4.1 and Prism 10.5.0 (GraphPad Software, La Jolla, CA, USA) software were used to analyze the data. Unpaired Student’s t-test was used to determine statistical significance for regularly distributed variables for comparisons between two groups, and the Mann-Whitney U test was used to analyze non-normally distributed data. The log-rank (Mantel-Cox) test was used to identify statistically significant differences, and the Kaplan-Meier method was employed to produce survival curves for the subgroups in each data set. The univariate Cox proportional hazards regression model was used to generate the hazard ratios for single-variate studies. To identify independent prognostic markers, the multivariate Cox regression model was employed. In our work, p ≤ 0.05 was deemed to be statistically significant. Animal experiment was conducted 3 times. Results CTNNB1GOF gene could accelerate HCC progression in mouse HTVi model Previous studies have shown that AKT-c-Myc and AKT-NRASG12V can induce HCC in HTVi-induced murine hepatocarcinogenesis models [12–15]. We sought to investigate the potential cooperative effects of Wnt/β-catenin signaling activation with those known oncogenic drivers. In this study, we introduced a gain-of-function mutant CTNNB1 (ΔN90-β-catenin) into these models to assess its impact on tumor development and survival. We forced expression of ΔN90-β-catenin with or without AKT-c-Myc/AKT-NRASG12V plasmids in 6-week-old C57BL/6J mice through HTVi with sleeping beauty transposon/transposase(Fig. 1A). Mice injected with ΔN90-β-catenin + AKT-c-Myc/AKT-NRASG12V displayed signs of early morbidity and mortality post-HTVi compared with models without ΔN90-β-catenin (Fig. 1B and C). At the 18-day timepoint, we observed livers with notable gross HCC and significantly increased liver weight (LW)/body weight (BW) ratio in ΔN90-β-catenin expressed mice (p < 0.0001, p = 0.0001) compared with LW/BW in non-ΔN90-β-catenin expressed mice (Fig. 1D and E). Fig. 1Establishment of murine liver cancer models with or without mutant gain-of-function CTNNB1. (A) Schematic showing the timeline of hydrodynamic tail vein injection of AKT-c-Myc/AKT-NRASG12V + transposon/transposase with or without ΔN90-β-catenin in 6-week-old C57BL6/J mice. (B) Kaplan-Meier curve showing decreased survival of AKT-c-Myc + ΔN90-β-catenin compared with AKT-c-Myc + Vector mice. Log-rank test p < 0.0001 for comparison. (C) Kaplan-Meier curve showing decreased survival of AKT-NRASG12V + ΔN90-β-catenin compared with AKT-NRASG12V + Vector mice. Log-rank test p < 0.0001 for comparison. (D) Bar graph showing significant increase in liver weight / body weight ratio (%) in AKT-c-Myc + ΔN90-β-catenin mice compared with AKT-c-Myc + Vector mice liver at same timepoint of euthanasia (Student’s t-test ****p < 0.0001). (E) Bar graph showing significant increase in liver weight / body weight ratio (%) in AKT-NRASG12V + ΔN90-β-catenin mice compared with AKT- NRASG12V + Vector mice liver at same timepoint of euthanasia (Student’s t-test ***p = 0.0001). (F) Representative liver gross image at moribund time, Hematoxylin-eosin staining and immunohistochemical staining of mouse liver tumors Establishment of murine liver cancer models with or without mutant gain-of-function CTNNB1. (A) Schematic showing the timeline of hydrodynamic tail vein injection of AKT-c-Myc/AKT-NRASG12V + transposon/transposase with or without ΔN90-β-catenin in 6-week-old C57BL6/J mice. (B) Kaplan-Meier curve showing decreased survival of AKT-c-Myc + ΔN90-β-catenin compared with AKT-c-Myc + Vector mice. Log-rank test p < 0.0001 for comparison. (C) Kaplan-Meier curve showing decreased survival of AKT-NRASG12V + ΔN90-β-catenin compared with AKT-NRASG12V + Vector mice. Log-rank test p < 0.0001 for comparison. (D) Bar graph showing significant increase in liver weight / body weight ratio (%) in AKT-c-Myc + ΔN90-β-catenin mice compared with AKT-c-Myc + Vector mice liver at same timepoint of euthanasia (Student’s t-test ****p < 0.0001). (E) Bar graph showing significant increase in liver weight / body weight ratio (%) in AKT-NRASG12V + ΔN90-β-catenin mice compared with AKT- NRASG12V + Vector mice liver at same timepoint of euthanasia (Student’s t-test ***p = 0.0001). (F) Representative liver gross image at moribund time, Hematoxylin-eosin staining and immunohistochemical staining of mouse liver tumors Based on immunohistochemistry, we observed all HCC were simultaneously positive for HA-tag (representing AKT-c-Myc and AKT + NRASG12V plasmids). β-catenin protein was highly expressed in ΔN90-β-catenin + AKT-c-Myc/AKT-NRASG12V HCC (Fig. 1F). Several specific pathways were enriched in CTNNB1GOF HCC Next, we perform RNA-seq on those two pairs of models (AKT-c-Myc + Vector, n = 3 vs. AKT-c-Myc + ΔN90-β-catenin, n = 3, AKT-NRASG12V + Vector, n = 3 vs. AKT-NRASG12V + ΔN90-β-catenin, n = 3). Differentially expressed genes (DEGs) were identified based on DEseq2 algorithm. To elucidate the molecular mechanisms underlying hepatocarcinogenesis driven by the gain-of-function mutant CTNNB1, we performed Gene Set Enrichment Analysis (GSEA) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis on tumors from the different experimental cohorts. Based on GSEA results, “Hippo_signaling_pathway_Signal_transduction”, and “Neuroactive_ligand_receptor_interaction_Signaling_molecules_and_interaction” were enriched in CTNNB1 GOF HCC in both two pairs of models (Fig. 2A and B). Based on KEGG enrichment analysis results, “Hippo_signaling_pathway_Signal_transduction” and “MAPK_signaling_pathway_Signal_transduction” were enriched in CTNNB1 GOF HCC in both two pairs of models (Fig. 2E and F). Fig. 2Transcriptomic enrichment analysis comparing mutant CTNNB1 to non-mutant CTNNB1 HCC. (A) Top 20 pathways based on NES from GSEA of differentially expressed genes comparing AKT-c-Myc + ΔN90-β-catenin to AKT-c-Myc + Vector. (B) Top 20 pathways based on NES from GSEA of differentially expressed genes comparing AKT-NRASG12V + ΔN90-β-catenin to AKT-NRASG12V + Vector. (C-D) GSEA plots of “Hippo_signaling_pathway_Signal_transduction”. (E) Top 20 pathways based on p value from KEGG pathway analysis of up-regulated genes comparing AKT-c-Myc + ΔN90-β-catenin to AKT-c-Myc + Vector. (F) Top 20 pathways based on p value from KEGG pathway analysis of up-regulated genes comparing AKT-NRASG12V + ΔN90-β-catenin to AKT- NRASG12V + Vector. HCC: hepatocellular carcinoma; NES: normalized enrichment score; GSEA: gene set enrichment analysis; KEGG: Kyoto Encyclopedia of Genes and Genomes Transcriptomic enrichment analysis comparing mutant CTNNB1 to non-mutant CTNNB1 HCC. (A) Top 20 pathways based on NES from GSEA of differentially expressed genes comparing AKT-c-Myc + ΔN90-β-catenin to AKT-c-Myc + Vector. (B) Top 20 pathways based on NES from GSEA of differentially expressed genes comparing AKT-NRASG12V + ΔN90-β-catenin to AKT-NRASG12V + Vector. (C-D) GSEA plots of “Hippo_signaling_pathway_Signal_transduction”. (E) Top 20 pathways based on p value from KEGG pathway analysis of up-regulated genes comparing AKT-c-Myc + ΔN90-β-catenin to AKT-c-Myc + Vector. (F) Top 20 pathways based on p value from KEGG pathway analysis of up-regulated genes comparing AKT-NRASG12V + ΔN90-β-catenin to AKT- NRASG12V + Vector. HCC: hepatocellular carcinoma; NES: normalized enrichment score; GSEA: gene set enrichment analysis; KEGG: Kyoto Encyclopedia of Genes and Genomes Transcriptionally comparison of CTNNB-mutated HCC and CTNNB1-wildtype HCC identifies a mutant CTNNB1 gene signature To define a robust transcriptional signature driven by oncogenic mutant CTNNB1, we integrated data from our controlled murine models with human HCC cohorts. We first identified DEGs from our two paired mouse models. 105 up-regulated and 23 down-regulated common genes were overlapped (Fig. 3A and B). Then we analyzed two independent human HCC databases: TCGA-LIHC, ICGC-LIRI). We identified DEGs in RNA-seq datasets of these two databases between CTNNB1-mutated and CTNNB1-wildtype HCC. 215 up-regulated genes and 1044 down-regulated gene were overlapped (Fig. 3A and B). Then we overlapped those up-regulated and down-regulated genes of human and mice, and finally identified 14 up-regulated genes and 4 downregulated genes including LGR5, NOTUM, GREB1, RNF43, ADCY8, AXIN2, TCF7, SLC13A3, ZNRF3, TNFRSF19, ATP1A2, MT3, SP5, NKD1 (up-regulated) and COL8A2, PAK3, BNC1, PRND (down-regulated) (Fig. 3A and B). Fig. 3Mutant CTNNB1 gene signature identification and construction of machine learning model to identify mutant CTNNB1 HCC. (A) Venn diagram illustrating the overlap of up-regulated genes in the two groups of mutant CTNNB1 mouse HCC models, TCGA-LIHC and ICGC-LIRI-JP mutant HCC. (B) Venn diagram illustrating the overlap of down-regulated genes in the two groups of mutant CTNNB1 mouse HCC models, TCGA-LIHC and ICGC-LIRI-JP mutant HCC. (C, E) ROC curve of fourteen machine learning models on training (TCGA-LIHC) and validation (ICGC-LIRI-JP) sets. (D, F) Performance line plots on training and validation sets. HCC: hepatocellular carcinoma; TCGA-LIHC: The Cancer Genome Atlas Liver Hepatocellular Carcinoma; ICGC-LIRI-JP: International Cancer Genome Consortium Liver Cancer - RIKEN, JP; ROC: Receiver Operating Characteristic; NPV: negative predictive value. PPV: positive predictive value Mutant CTNNB1 gene signature identification and construction of machine learning model to identify mutant CTNNB1 HCC. (A) Venn diagram illustrating the overlap of up-regulated genes in the two groups of mutant CTNNB1 mouse HCC models, TCGA-LIHC and ICGC-LIRI-JP mutant HCC. (B) Venn diagram illustrating the overlap of down-regulated genes in the two groups of mutant CTNNB1 mouse HCC models, TCGA-LIHC and ICGC-LIRI-JP mutant HCC. (C, E) ROC curve of fourteen machine learning models on training (TCGA-LIHC) and validation (ICGC-LIRI-JP) sets. (D, F) Performance line plots on training and validation sets. HCC: hepatocellular carcinoma; TCGA-LIHC: The Cancer Genome Atlas Liver Hepatocellular Carcinoma; ICGC-LIRI-JP: International Cancer Genome Consortium Liver Cancer - RIKEN, JP; ROC: Receiver Operating Characteristic; NPV: negative predictive value. PPV: positive predictive value Construction of machine learning models based on transcriptome level to identify patients with CTNNB1 mutations Leveraging the differentially expressed genes identified from the integrated analysis, we sought to develop a robust classifier to predict the CTNNB1 mutation status in HCC patients. The expression levels of those genes were used as the feature set for model construction. We utilized the TCGA-LIHC cohort (n = 371) as the training set and the independent ICGC-LIRI-JP cohort (n = 226) as the external validation set. To identify the optimal algorithm for this prediction task, we implemented and rigorously compared 14 distinct machine learning algorithms: Random Forest, Gradient Boosting, Support Vector Machine (SVM), Logistic Regression, k-Nearest Neighbors Method, Partial Least Squares (PLS) Model, Boosting Method, Neural Networks, Naive Bayes, Discriminant Analysis, least absolute shrinkage and selection operator (Lasso) Regression, Adaptive Boosting, Categorical Boosting (CATboost), and Light Gradient Boosting Machine (LightGBM). Each model was evaluated on the training and validation cohorts, based on its Receiver Operating Characteristic (ROC) curve (Fig. 3C and E) and a comprehensive set of performance metrics, including accuracy, F1-score, Negative Predictive Value (NPV), Positive Predictive Value (PPV), sensitivity, specificity, and Youden’s index (Fig. 3D and F). Notably, the CATBoost algorithm demonstrated superior and consistent performance across both datasets. On the TCGA-LIHC training set, the CATBoost model achieved an area under the ROC curve (AUC) of 0.957, indicating favorable discriminative ability (Fig. 3C). The high performance was maintained on the external ICGC-LIRI-JP validation set, where the model attained an AUC of 0.893 (Fig. 3E). The confusion metrics for the CATboost model on training and test datasets are shown by Fig. 4A and B. On training set, the model exhibited high accuracy of 88.4%, so as on test set with an accuracy of 86.2%, confirming its ability to identify CTNNB1-mutated HCC. Fig. 4Evaluation and interpretation of CATboost model. (A-B) The confusion metrics for the CATboost model on training and test datasets. (C) SHAP bar plot of genes included in CATboost model. (D) SHAP beeswarm plot of genes included in CATboost model. (E-H) SHAP feature dependence plot of top 4 important genes. CATboost: Categorical Boosting, SHAP: SHapley Additive explanation Evaluation and interpretation of CATboost model. (A-B) The confusion metrics for the CATboost model on training and test datasets. (C) SHAP bar plot of genes included in CATboost model. (D) SHAP beeswarm plot of genes included in CATboost model. (E-H) SHAP feature dependence plot of top 4 important genes. CATboost: Categorical Boosting, SHAP: SHapley Additive explanation To evaluate contribution of those signature genes to the model, we used SHapley Addictive exPlanations (SHAP) to illustrate the importance degree. Figure 4C shows the average absolute SHAP value of genes. ADCY8, COL8A2 and ZNRF3 are excluded because their SHAP value are 0, which means they have negligible contribution to the prediction model. The remained 15 genes comprise the mutant CTNNB1 gene signature (MCGS). Figure 4D is a SHAP beeswarm plot. Each dot represents a single patient from training cohort. The horizontal position shows whether a gene’s expression value for a given patient pushes the model prediction towards CTNNB1-mutated (positive SHAP value) or wild-type (negative SHAP value). The color indicates the gene’s expression value, from low (purple) to high (yellow). We also show SHAP feature dependence of top 4 important gene in Fig. 4E-H. The x-axis is gene expression, and the y-axis is SHAP value. Survival analysis of CTNNB1 mutant gene signature identified metallothionein-3 (MT3) as a prognostic gene Having established a conserved gene signature for CTNNB1 mutant HCC, we next sought to investigate its clinical prognostic relevance. We aimed to determine if individual genes within the signature harbor independent prognostic power. To identify genes within the MCGS that are independently associated with patient survival, we performed survival analysis in the TCGA-LIHC cohort. The expression level of each gene in the MCGS was subjected to both univariate and multivariate Cox proportional hazards regression analyses, with overall survival as the endpoint. Univariate cox analysis demonstrated that the expression of MT3 (HR = 1.08, P = 0.000155) is significantly associated with overall survival (OS) (Fig. 5A). We then included it into multivariate cox analysis with clinicopathological characteristics including Age, gender, pathologic stage and tumor grade. We found that MT3 expression (HR = 1.2, P = 0.043) is a prognostic adverse factor for OS (Fig. 5B). Fig. 5High MT3 mRNA expression is associated with OS in HCC. (A) Univariate Cox regression analysis demonstrating MT3 is OS associated. (B) Multivariate Cox regression analysis demonstrating MT3 is independent prognostic factor. (C-H) Kaplan-Meier curve for OS on several public HCC RNA-seq datasets including TCGA-LIHC, ICGC-LIRI-JP, TCGA + ICGC integration, GSE14520, OEP000321 and GSE76427. MT3: Metallothionein-3. OS: overall survival; HCC: hepatocellular carcinoma; TCGA-LIHC: The Cancer Genome Atlas Liver Hepatocellular Carcinoma; ICGC-LIRI-JP: International Cancer Genome Consortium Liver Cancer - RIKEN, JP High MT3 mRNA expression is associated with OS in HCC. (A) Univariate Cox regression analysis demonstrating MT3 is OS associated. (B) Multivariate Cox regression analysis demonstrating MT3 is independent prognostic factor. (C-H) Kaplan-Meier curve for OS on several public HCC RNA-seq datasets including TCGA-LIHC, ICGC-LIRI-JP, TCGA + ICGC integration, GSE14520, OEP000321 and GSE76427. MT3: Metallothionein-3. OS: overall survival; HCC: hepatocellular carcinoma; TCGA-LIHC: The Cancer Genome Atlas Liver Hepatocellular Carcinoma; ICGC-LIRI-JP: International Cancer Genome Consortium Liver Cancer - RIKEN, JP Based on the Cox model results, we further validated the prognostic value of MT3 by stratifying patients in several HCC cohorts (TCGA-LIHC, ICGC-LIRI-JP, TCGA-ICGC integration, GSE14520, OEP000321, GSE76427) into high and low MT3 expression groups using the optimal cut-off value determined by maximally selected rank statistics test for Kaplan-Meier survival analysis. The results demonstrated that patients with high tumor expression of MT3 had a significantly shorter OS compared to those with low MT3 expression (Fig. 5C-H). Validation of prognostic significance of MT3 based on HCC microarray cohort To clinically validate the prognostic significance of MT3 at the protein level, we constructed a tissue microarray (TMA) from a well-annotated, institutional cohort of 147 patients with HCC. MT3 protein expression was evaluated by immunohistochemistry (IHC). The samples of the four classifications of immunohistochemical staining intensity are shown in Fig. 6A. And IHC score integrating staining intensity and extent of positive staining was used to stratify patients into high and low expression groups. The prognostic value of MT3 was assessed for both OS and recurrence-free survival (RFS) through univariate and multivariate Cox proportional hazards analyses, adjusting for key clinical-pathological variables including age, gender, liver cirrhosis, tumor number, tumor size, tumor grade, macro/micro vascular invasion, AFP level, and BCLC stage. The clinicopathological characteristics of the 147 patients with HCC are shown in Table S1. Fig. 6High MT3 protein expression is associated with OS and RFS in HCC. (A) four classifications of MT3 immunohistochemical staining intensity. (B-C) Kaplan-Meier curve for OS and RFS demonstrating bad prognosis in MT3 protein highly expressed patients. MT3: Metallothionein-3; OS: overall survival; RFS: recurrence free survival; HCC: hepatocellular carcinoma, IHC: immunohistochemistry High MT3 protein expression is associated with OS and RFS in HCC. (A) four classifications of MT3 immunohistochemical staining intensity. (B-C) Kaplan-Meier curve for OS and RFS demonstrating bad prognosis in MT3 protein highly expressed patients. MT3: Metallothionein-3; OS: overall survival; RFS: recurrence free survival; HCC: hepatocellular carcinoma, IHC: immunohistochemistry Univariate Cox analysis for OS identified several factors significantly associated with OS, including Age ≥ 60 years old [hazard ratio (HR) = 1.81, P = 0.041], male gender (HR = 0.47, P = 0.018), presence of liver cirrhosis (HR = 2.42, P = 0.006), tumor numbers > 2 (HR = 2.35, P = 0.008), and high MT3 expression (HR = 2.00, P = 0.016) (Table S2). Subsequent multivariate analysis, which included all significant variables from the univariate model, confirmed that Age ≥ 60 years old (HR = 1.86, P = 0.037), male gender (HR = 0.49, P = 0.031), presence of liver cirrhosis (HR = 2.38, P = 0.008), tumor numbers > 2 (HR = 2.42, P = 0.008), and high MT3 expression (HR = 1.97, P = 0.023) each retained their status as independent factors for OS (Table S3). We next investigated the association between characteristics and RFS. In univariate analysis, Age ≥ 60 years old (HR = 1.74, P = 0.042), male gender (HR = 0.49, P = 0.021), presence of liver cirrhosis (HR = 2.14, P = 0.009), tumor numbers > 2 (HR = 2.34, P = 0.006), G3/G4 tumor grade (HR = 1.68, P = 0.048) and high MT3 expression (HR = 1.89, P = 0.019) was significantly associated with RFS (Table S4). Multivariate Cox regression model for RFS demonstrated that Age ≥ 60 years old (HR = 2.19, P = 0.008), presence of liver cirrhosis (HR = 2.20, P = 0.009), tumor numbers > 2 (HR = 2.37, P = 0.007), G3/G4 tumor grade (HR = 1.87, P = 0.020) and high MT3 expression (HR = 1.76, P = 0.048) were independent factor for RFS (Table S5). In summary, high MT3 expression was independent risk factor for both OS and RFS. We also divided patients into two groups according to IHC score of MT3 expression. As shown in Fig. 6B-C, there were significant differences in OS and RFS between high MT3 expression group and low MT3 expression group. Discussion HCC exhibits substantial molecular heterogeneity, which is closely associated with diverse mutational patterns. HCCs harboring CTNNB1 mutations represent a distinct molecular subtype characterized by constitutive activation of the Wnt signaling pathway. This subtype typically displays an immunosuppressive tumor microenvironment and shows resistance to immunotherapy [9, 10]. In this study, we employed HTVi mouse model of HCC and integrated transcriptomic data from the TCGA-LIHC and ICGC-LIRI-JP cohorts to identify a CTNNB1- mutation-associated gene expression signature and to develop a predictive model for CTNNB1 mutational status based on transcriptomic profiles. Through differential gene expression analysis, we identified 18 genes that were consistently dysregulated in CTNNB1-mutant HCC. Among them, NOTUM, AXIN2, RNF43, ZNRF3, and NKD1 function as negative feedback regulators of Wnt pathway activation. These genes are transcriptionally induced upon pathway activation and serve to attenuate signaling intensity [16–19]. NOTUM is a well-established Wnt/β-catenin target gene upregulated in HCC with activated Wnt/β-catenin signaling. It suppresses Wnt activity by deacylating Wnt proteins [20]. A 2021 study demonstrated that in colorectal cancer, APC-mutant tumor cells secrete NOTUM to inhibit the differentiation of adjacent wild-type crypt cells, thereby gaining a growth advantage [21]. AXIN2, another classical Wnt/β-catenin target, encodes a critical component of the β-catenin destruction complex. RNF43 and ZNRF3 are transmembrane E3 ubiquitin ligases that ubiquitinate Frizzled receptors, driving their endocytosis and lysosomal degradation. Deficiency of RNF43/ZNRF3 predisposes to hepatocarcinogenesis by impairing liver regeneration and disturbing hepatic lipid metabolism [22]. NKD1, also a canonical Wnt target, interacts directly with Dishevelled (Dvl), a key adaptor protein in Wnt signaling. By binding to Dvl, NKD1 disrupts its polymerization and activation, thereby inhibiting downstream signal propagation [23]. In contrast, LGR5, TCF7, TNFRSF19, and SP5 function as positive feedback amplifiers of the Wnt pathway. LGR5 forms a complex with its natural ligands, the R-spondin family proteins. The LGR5-R-spondin complex promotes internalization and degradation of RNF43 and ZNRF3, increasing cell-surface Wnt receptor abundance and enhancing cellular sensitivity to Wnt ligands [24]. Intracellularly, β-catenin associates with TCF7 to form a transcriptional activation complex that induce the expression of Wnt/β-catenin target genes. TCF7 itself is a direct transcriptional target of the β-catenin-TCF complex, thereby forming a strong auto-amplification loop [25]. TNFRSF19, a member of the TNF receptor superfamily, acts as a co-receptor by interacting with LRP5/6 and recruiting AXIN1 to the plasma membrane. This sequestration disintegrates the cytoplasmic β-catenin destruction complex and prevents β-catenin phosphorylation and degradation [26, 27]. SP5 can bind to the promoters of Wnt/β-catenin target genes and cooperate with TCF/LEF transcription factors to enhance transcriptional activation [28]. The remaining eight differentially expressed genes—upregulated SLC13A3, ADCY8, MT3, and ATP1A2, and downregulated COL8A2, PAK3, BNC1, and PRND—are not established components of the canonical Wnt feedback circuitry. SLC13A3 encodes a transporter for glutathione (GSH); its silencing depletes intracellular GSH, induces autophagic ferroptosis, and downregulates the leucine transporter SLC7A5, leading to leucine deprivation and mTOR pathway inhibition. SLC13A3 contributes to liver cancer pathogenesis as an effector downstream of activated β-catenin [29]. ADCY8 is a calcium/calmodulin-stimulated adenylate cyclase that catalyzes the conversion of ATP to cAMP, sharply elevating intracellular cAMP levels in response to increased calcium [30]. ADCY8 is overexpressed in HCC patients and is associated with poorer prognosis [31]. ATP1A2 encodes the α2 catalytic subunit of the Na⁺/K⁺-ATPase. Predominantly expressed in astrocytes and muscle tissue, ATP1A2 exhibits higher affinity for extracellular K⁺, facilitating rapid potassium clearance [32]. Although little is known about ATP1A2 in cancer, its high expression may reflect a notable metabolic investment by tumor cells to maintain specific ion gradients. MT3, rich in cysteine residues, binds zinc with high affinity and participates in regulating intracellular zinc homeostasis [33]. It effectively scavenges reactive oxygen species, conferring resistance to oxidative stress, and has been implicated in the development of sorafenib resistance in HCC cells [34]. The downregulated genes COL8A2, PAK3, BNC1, and PRND may also play diverse roles in tumor progression. COL8A2 encodes the collagen type VIII alpha 2 chain, a basement membrane component that inhibits cell migration [35]. Its downregulation may lead to basement membrane disorganization and increased tumor invasiveness. PAK3 regulates actin cytoskeleton dynamics and promotes cell protrusion formation [36]. Although PAK3 can facilitate HCC metastasis [37], the strong proliferative drive conferred by CTNNB1 mutation may suppress PAK3 expression as a resource-allocation strategy—prioritizing proliferation over energetically costly migratory programs. BNC1 is a transcription factor that acts as a potent inhibitor of epithelial–mesenchymal transition (EMT), promoting epithelial marker expression (e.g., E-cadherin) and suppressing mesenchymal markers. Downregulation of BNC1 triggers EMT, leading to loss of epithelial identity, increased motility, and enhanced invasiveness [38]. PRND encodes Doppel, a prion-like protein with incompletely understood functions; it may be involved in apoptosis regulation and oxidative stress response [39]. Among all candidates, MT3 emerged as a significant and independent prognostic gene consistently validated across both large-scale public RNA-seq datasets and our institutional HCC TMA cohort. This concordance—from transcriptomic to proteomic levels and across independent cohorts—strengthens the reliability of our findings and underscores the pivotal role of MT3 in HCC progression. While MT3 is traditionally known for its functions in metal ion homeostasis and detoxification, our results position MT3 as an active driver of tumor aggressiveness rather than merely a passive stress-response molecule. High MT3 expression may confer survival advantages by mitigating oxidative stress and resisting apoptosis induced by the tumor microenvironment or therapeutic interventions. Moreover, through regulation of intracellular zinc dynamics, MT3 may modulate the activity of multiple transcription factors and signaling pathways critical for tumor proliferation and invasion. As for the relationship with Wnt/β-catenin, MT3 may affect β‑catenin stability and transcriptional activity by regulating cellular redox balance and zinc homeostasis. Redox‑sensitive modulation of Wnt signaling through the nucleoredoxin-Dishevelled axis and zinc‑regulated kinases such as AKT and GSK3β provide possible mechanistic links [40, 41]. Moreover, MT3 may interact with PI3K/AKT or MAPK cascades that crosstalk with Wnt/β‑catenin signaling, potentially amplifying CTNNB1‑driven oncogenic pathways [42]. Further functional in vitro and in vivo studies, including manipulation of MT3 in CTNNB1-mutant HCC models and assessment of β-catenin nuclear activity, are warranted to clarify its role and therapeutic potential. Using the identified gene set, our CatBoost model demonstrated robust predictive performance for CTNNB1 mutational status. SHAP analysis indicated minimal contributions from ADCY8, COL8A2, and ZNRF3. Although these genes exhibited limited predictive influence in the computational model, this should not be interpreted as biological insignificance. In conclusion, we identified a CTNNB1-mutation–associated gene signature in HCC and constructed an effective machine-learning model to predict CTNNB1 mutation status. Furthermore, we revealed MT3 as a key prognostic driver within this signature, highlighting its substantial research and clinical relevance. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1