Radiogenomic MRI biomarkers for noninvasive prediction of GPC3 expression and tumor microenvironment in hepatocellular carcinoma.

Background
Hepatocellular carcinoma (HCC) poses a significant global health burden, causing approximately 830,000 annual deaths with advanced-stage 5-year survival rates below 18%, largely due to delayed diagnosis and therapeutic resistance [1, 2]. The insidious onset of early HCC frequently leads to late detection, necessitating systemic therapies such as immunotherapy whose efficacy is often compromised by tumor microenvironment (TME) heterogeneity [3, 4]. This underscores an urgent need for robust biomarkers to refine risk stratification and therapeutic strategies [5]. Glypican-3 (GPC3), overexpressed in 70–90% of HCC cases, has emerged as a pivotal biomarker and therapeutic target by enhancing Wnt/β-catenin signaling, promoting tumor proliferation and suppressing cytotoxic T-cell infiltration [6, 7]. Under hypoxic conditions, GPC3 interacts with HIF-1, preventing its proteasomal degradation and amplifying angiogenesis via VEGF upregulation [8].
Noninvasive imaging, particularly dynamic contrast-enhanced MRI (DCE-MRI), offers superior soft-tissue resolution for characterizing HCC biology. Multiparametric sequences (contrast-enhanced T1-weighted imaging [CE-T1WI] and diffusion-weighted imaging [DWI]) enable precise interrogation of intratumoral hemodynamics and TME remodeling [9, 10].
The integration of radiomics—high-throughput extraction of quantitative imaging features augmented by artificial intelligence—has further revolutionized TME biomarker discovery [11].
Despite advances in radiomic models in predicting GPC3 expression, their clinical applicability remains limited due to lack of external validation, reliance on a single machine learning approach, and the absence of mechanistic insights into how GPC3-associated imaging phenotypes influence the TME [12–14]. Recent evidence highlights the peritumoral niche,especially a metabolically active 500-µm zone at the tumor-liver interface (TLI), as crucial for invasion and immunosuppression [15, 16], and models based on 3-mm peritumoral region frequently manifest the predominant performance [17–19]. Yet, radiomic profiling of this strategic region remains unexplored.
To address these gaps, we propose a weighted gene co-expression network analysis (WGCNA)-enhanced radiogenomic framework that transcends conventional univariate correlations. By identifying co-expression gene modules holistically adapted to imaging phenotypes, this approach mitigates false positives inherent in single-gene analyses [20, 21] and alleviate limitations of tissue scarcity, high costs with conventional approaches based on bulk RNA sequencing or immunohistochemistry (IHC) [22].
Our study uniquely investigates HIF-1/PPAR pathway activation as the mechanistic bridge connecting GPC3-driven radiomic signatures to immunosuppressive TME remodeling, advancing precision immunotherapy for HCC.

Materials and methods

Data sources and histopathology
This multicenter retrospective study integrated institutional clinical imaging data (three tertiary centers) and public genomic repositories. Consecutive HCC patients with pretreatment DCE-MRI and immunohistochemically confirmed GPC3 status were enrolled from Center I (July 2013-July 2023, n = 196), Center II (January 2021-June 2023, n = 82), and Center III (January 2022-January 2023, n = 80). Institutional cases (n = 274) comprised the imaging cohort used for model training and internal validation; Center III (n = 80) served as an independent institutional external test set. Inclusion criteria: (1) Pathologically confirmed primary HCC; (2) Treatment-naïve status; (3) Preoperative MRI within 30 days; (4) Absence of extrahepatic malignancies or severe hepatic/renal dysfunction. In addition, 34 cases with available MRI and bulk RNA-seq data (FPKM) data from the Cancer Imaging Archive (TCIA, https://www.cancerimagingarchive.net/) database were used as an extra radiogenomic validation cohort. 363 cases with bulk RNA-seq data (FPKM) and clinical annotations from the Cancer Genome Atlas - Liver Hepatocellular Carcinoma cohort (TCGA-LIHC, https://portal.gdc.cancer.gov/, (n = 363) were enrolled for transcriptomic validation and pathway analysis. ComBat harmonized batch effects.
Histology and GPC3 IHC: All institutional tumor specimens were formalin-fixed paraffin-embedded (FFPE). GPC3 expression was assessed by immunohistochemistry on FFPE sections according to institutional pathology protocols. Positivity was defined as the proportion of positively stained cells ≥ 25%. For multi-nodule patients, the largest index lesion (by maximum diameter on imaging) was used for radiomic analysis unless otherwise specifie

Radiomic processing

Imaging acquisition and segmentation
Due to CE-T1WI and DWI were prioritized based on prior literature demonstrating their predictive performance for molecular markers in HCC and due to availability across centers, arterial-phase CE-T1WI and DWI sequences (1.5T/3.0T) were retrieved from PACS in Digital Imaging and Communications in Medicine (DICOM) format (Supplementary Figure 1).
Tumor segmentation was performed on CE-T1WI using 3D Slicer (v5.2.0) by two board-certified radiologists ( > 5 years experience) who were blinded to histopathology. Based on prior reports demonstrating optimal performance with a 3 mm three-dimensional volume, for each lesion we delineated: (1) tumor core (whole tumor excluding visible capsule/vessels), (2) 3-mm peritumoral ring extending into adjacent liver parenchyma from the tumor-liver interface (TLI), and (3) a bidirectional 3-mm zone across the TLI as described in prior studies. When lesion borders were ill-defined, segmentation followed a conservative approach: regions of unequivocal tumor signal were delineated and ambiguous margins were excluded; cases with grossly non-segementable lesions were recorded and excluded from imaging feature analysis. Arteriovenous shunts and adjacent large vessels were excluded from ROIs. DWI ROIs were co-registered to CE-T1WI and manually adjusted on ADC maps to ensure coverage of the same tumor volume; when DWI suffered severe artifacts the sequence was excluded for that case and the model relying only on CE-T1WI was applied. DSC > 0.80 confirmed interobserver concordance (20% random subset).

Feature extraction and selection
Images were resampled to isotropic voxels (3 × 3×3 mm3) prior to feature extraction. Radiomic features were computed from CE-T1WI and DWI using PyRadiomics (https://pyradiomics.readthedocs.io/). A total of 1,688 features were initially extracted from tumor and peritumoral ROIs, including first-order statistics, shape, and texture (GLCM, GLRLM, GLSZM, NGTDM, GLDM) and wavelet variants. Prior to feature selection, features were standardized (z-score) and extreme outliers were trimmed by the 1.5 IQR rule.
To reduce redundancy, Spearman correlation filtering (|ρ| > 0.75, p < 0.05) was applied and one representative feature from each highly correlated pair was retained. Least absolute shrinkage and selection operator (LASSO) regression optimized feature selection (λ chosen by minimal binomial deviance via 10-fold CV). Features retained in > 90% of bootstrap iterations were considered robust and included in final modeling.

Model development
Patients from Centers I/II were partitioned into cross-validation (n = 136) and internal validation (n = 58) sets (7:3). Center III (n = 80) and TCIA (n = 34) comprised external cohorts.
Machine learning models, including logistic regression (LR), random forest (RF), and support vector machine (SVM), K-Nearest Neighbor (KNN), and decision tree (DT) were trained. Feature selection and model training underwent a 10-fold cross-validation with Synthetic Minority Over-sampling Technique (SMOTE) balancing. Features selected > 90% of iterations were retained. Hyperparameter tuning details are provided in Supplementary File 1. Model performance was validated using both internal and external validation sets. The optimal model was chosen by comparing the area under the curve (AUC), sensitivity, specificity with 95% confidence interval (CI), and validated via calibration curves, Decision curve analysis (DCA), and the SHapley Additive exPlanations (SHAP) analysis.
A radscore was computed to generate an interpretable quantitative score as the weighted feature coefficients of the linear logistic regression model, and predict_proba output of the remained non-linear models. The correlation between radscore and GPC3 expression was confirmed using spearman correlation analysis.

Radiogenomic workflow
With TCIA cases, Radscores derived from imaging models was computed and identified radscore-related genes using Spearman analysis (|R| > 0.3 and p < 0.05, FDR adjusted).
TCGA-LIHC (n = 363) cohort with bulk RNA-seq data and clinical annotation were projected radscore-related genes and subjected to WGCNA analysis to identify signatures associated with the TME. Hierarchical clustering removed outlier genes. A soft thresholding power (β = 4) was determined by achieving scale-free topology fit (R2 > 0.9, mean connectivity < 100). Modules correlated with ESTIMATE-derived TME scores, including immune score, tumor purity, stromal score, and ESTIMATE score, were identified using dynamic tree cutting (minModuleSize = 30, MEDissThres = 0.25). The turquoise module of 108 genes showed strongest association.
The Search Tool for the Retrieval of Interacting Genes/Proteins (STRING, https://string-db.org/) selected those with an interaction score > 0.15. Cytoscape (http://www.cytoscape.org/) with MCODE score > 3 and CytoHubba top 10 identified hub genes.

Risk model development
A multi-Cox regression model was developed using the formula:
where β represents regression coefficients, i denotes hub genes, and Exp is gene expression level. Patients were stratified into high- and low-risk groups based on median risk scores. R package ConsensusClusterPlus confirmed the rationality of binary classification (maxK = 6, reps = 50, pItem = 0.8, pFeature = 0.8).

Immune-related responder and enrichment analyses
Predicted immunotherapy response was estimated by the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm using TCGA expression profiles. The CIBERSORT algorithm compute the abundances of 22 tumor-infiltrating immune cells in tumor tissues of patients with HCC between high- and low-risk group.
Gene-ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses revealed the potential mechanism and pathway of hub genes.

Statistical analysis
Kaplan–Meier survival curves were analyzed using the log-rank test. Cox proportional hazards models assessed risk factors. A nomogram was developed and validated using Harrell’s C-index and 1,000 bootstrap resamples. Analyses were conducted in R v4.3.3) and Python v3.7/3.11(p < 0.05 significant).

Results

Study workflow and patient characteristics
Figure 1 depicts the overall study workflow. The combined cohort comprised 274 institutional, 363 TCGA, and 34 TCIA HCC patients. Demographic and clinical characteristics were consistent across centers (Table 1). Median age ranged from 57 to 61 years, with female patients comprising 22–33% of each cohort. Median overall survival (OS) was 810 days in TCGA and 874 days in TCIA (institutional OS data were unavailable). GPC3 was positive in 34% of institutional tumors. TCGA and TCIA data were obtained from public repositories.

Radiomics model performance
From multi-ROIs of each sequence, 1,688 radiomic features were initially extracted and underwent preprocessing. A total of 20 features were extracted in the CE-T1WI sequence, 15 features in the DWI sequence, and 25 features in the integrated CE-T1WI+DWI sequence. Detailed lists and definitions of all extracted features are provided in Supplementary Figure 1.
Table 2 presents the performance metrics of 10-fold cross-validation. Results indicated CE-T1WI and combined CE-T1WI+DWI sequences achieved the highest performance. The RF and SVM models on CE-T1WI yielded demonstrated high sensitivity ( > 96%) and stable AUC values, with RF achieving a training AUC of 0.965 (95% CI 0.945–0.985) and a validation AUC of 0.935 (95% CI 0.863–0.995), while SVM exhibited near-perfect sensitivity in training but slightly lower validation AUC. The integrated CE-T1WI+DWI sequence produced the best overall performance, the RF model attained the highest validation AUC of 0.979 (95% CI 0.956–0.995) with balanced sensitivity and specificity, and the KNN model achieved the highest specificity (0.867, 95 CI% 0.787–0.932) (Fig. 2A).
Calibration curves (Fig. 2B) showed near-perfect calibration for the CE-T1WI RF model across high probability ranges. The combined RF model’s predictions closely matched the ideal line. These results confirm that multi-sequence integration substantially improves model stability.
DCA (Fig. 2C) showed the CE-T1WI RF model had the highest net benefit across threshold probabilities of 0.2–0.8, outperforming both “treat all” and “treat none” strategies. The combined RF model further increased net benefit over all thresholds, indicating broader clinical applicability. These findings underscore the complementary value of CE-T1WI and DWI. Considering CE-T1WI’s robust performance and the absence of DWI data in TCIA, subsequent analyses focused on the CE-T1WI sequence.
Based on selected features and training model in the CE-T1WI sequence, SHAP analysis (Fig. 3A) was used to interrogate how individual radiomic features influence the model’s prediction of GPC3 positivity. The SHAP beeswarm plot ranks features by their overall contribution that illustrated via mean absolute SHAP and shows the direction and magnitude of each feature’s effect on the model output, in which points to the right increase the predicted probability of GPC3 positivity, and points to the left decrease it. Overall, the top-ranked features are dominated by texture and wavelet-transformed metrics. Specifically, the highest-impact feature, lbp-3D-k_firstorder_Entropy, shows that higher entropy values are associated with positive SHAP values, namely higher entropy increases the predicted probability of GPC3 positivity. This indicates that tumors with greater intensity disorder/heterogeneity are more likely to be predicted as GPC3-positive by our model. Other high-ranking features such as lbp-3D-k_glszm_SmallAreaHighGrayLevelEmphasis and wavelet-HHL_glrlm_ShortRunHighGrayLevelEmphasis similarly show that larger values (indicating more or stronger small-area/high-intensity foci and short high-intensity runs) push the prediction toward GPC3 positivity. Model performance metrics in the external validation set in the CE-T1WI sequence, implicating the SVM (AUC = 0.756) possessed the best effect (Fig. 3B, Table 3); the strong correlation with GPC3-positive expression and distribution in the training and internal-validation sets of GPC3 was confirmed (R = 0.78, p < 0.05) and manifested (Figs. 3C and D).

Radiogenomic analyses
A radscore was calculated for TCIA patients using the CE-T1WI model. Representative cases (Fig. 4A) show that high-radscore tumors exhibit uniform density, clear boundaries, and absence of vascular invasion, whereas low-radscore tumors are irregular with necrosis and prominent vessels. High-radscore patients tended to be younger and have lower tumor stage, consistent with the favorable prognosis of GPC3-positive tumors.
WGCNA with a soft threshold of β = 4identified four gene modules from 409 radscore-correlated genes (Fig. 4B). The 108-gene turquoise module showed the strongest association with TME scores (Fig. 4C–E).
STRING network analysis and MCODE/CytoHubba identified 10 hub genes (Fig. 5A): glyceraldehyde-3-phosphate dehydrogenase (GAPDH), Arginase-1 (ARG1), Cytochrome b5 (CYB5A), Aldo-keto reductase family 1 member B1 (AKR1B1), Secreted Phosphoprotein 1 (SPP1), ATP-binding cassette subfamily C member 1 (ABCC1), phosphofructokinase (PFKP), Acyl-CoA oxidase 2 (ACOX2), 7-dehydrocholesterol reductase (DHCR7), fatty acids binding protein 4 (FABP4) (Fig. 5B), in which ACOX2, DHCR7, FABP4, CYB5A correlated positively with radscore, whereas PKFP, GAPDH, AKR1B1, ABCC1, ARG1 and SPP1 correlated negatively (Fig. 5D). Immunohistochemistry (Supplementary Figure 2) confirmed their differential expression in tumor versus normal tissues.
A Cox regression model based on these hub genes stratified patients into high- and low-risk groups (optimal k = 2, Fig. 5C). High-risk patients had significantly poorer overall survival than low-risk patients (p < 0.05, Fig. 5E). In multivariate analysis (C-index = 0.720), both the gene-based risk score (HR = 4.07, p < 0.0001) and tumor T stage (HR = 1.55, p = 0.0007) were independent predictors of mortality (Fig. 5F).

Immune-related responder and enrichment analyses
Overall immunotherapy response was low (8.8%). High-risk patients had a significantly better response than low-risk patients (p = 0.015, Fig. 6A), suggesting that radscore may stratify treatment benefit. (Fig. 6A). CIBERSORT analysis (Fig. 6B) showed differential immune correlations for the hub genes. GAPDH correlated negatively with resting and T cells gamma delt, and PFKP negatively with monocytes, suggesting an immunosuppressive microenvironment. Conversely, SPP1 correlated positively with M2 macrophages and monocytes, and FABP4 with activated dendritic cells, indicating potential roles in immunosuppression.
High-risk patients exhibited significantly higher ImmuneScore and ESTIMATEScore than low-risk patients (Fig. 6C), indicating greater immune/stromal infiltration. StromalScore was similar between groups, while tumor purity was higher in low-risk tumors, supporting the validity of the WGCNA-derived modules.
GO analysis (Fig. 6D) indicated hub genes were enriched in steroid and hexose metabolic processes and oxidoreductase activity. KEGG pathways (Fig. 6E) included glycolysis/gluconeogenesis, HIF-1 signaling, PPAR signaling, and steroid biosynthesis, reflecting metabolic reprogramming and hypoxic adaptation in HCC.

Discussion
We developed a noninvasive MRI radiomics model that accurately predicts GPC3 expression in HCC. It demonstrated robust performance in both internal cross-validation and independent external cohorts, effectively stratifying GPC3-positive tumors. Radscore correlated strongly with GPC3 status, implying that the image-derived signature reflects underlying GPC3-driven biology. This aligns with prior studies showing that MRI radiomics can noninvasively identify GPC3-positive HCC [23]. By linking Radscore with gene co-expression modules, our framework positions Radscore as a surrogate marker for GPC3-associated hypoxic and lipid metabolic alterations.
Radscore-associated gene modules revealed a metabolic shift in GPC3-positive tumors. Radscore-high profiles upregulated glycolytic and proliferative genes (e.g. PFKP, GAPDH, AKR1B1, ABCC1, ARG1), whereas Radscore-low tumors expressed genes involved in fatty acid and cholesterol metabolism (e.g. ACOX2, FABP4, DHCR7, CYB5A). This suggests that GPC3+ HCCs undergo enhanced glycolysis with altered lipid handling. Pathway analysis confirmed significant enrichment of HIF-1 and PPAR signaling among Radscore-associated genes [24, 25]. HIF-1α stabilization under hypoxia drives glycolysis and lactic acid production, fostering adaptation to low-oxygen conditions and promoting immunosuppression [8]. Conversely, PPAR activation enhances fatty acid oxidation and polarizes macrophages toward an M2-like immunosuppressive phenotype via RIPK3-ROS signaling [26]. These intertwined metabolic–immune pathways likely underlie the imaging phenotypes observed: high intratumoral entropy may reflect intermingled HIF-1-driven glycolytic and PPAR-driven lipogenic regions, whereas uniform texture patterns correspond to homogeneous lipid deposition by PPAR-regulated processes.
High entropy features of lbp-3D-k_firstorder_Entropy in the imaging of the tumor suggests a high degree of disorder in the pixel intensity distribution, which may be related to the spatial interlacing of HIF-1 pathologically driven glycolytic regions (high metabolic activity) and PPAR-regulated lipid deposition regions (low metabolic activity). Similarly, low values of logarithm_glszm_ZoneEntropy and wavelet-HLL_glcm_JointEnergy may reflect local metabolic homogenization, corresponding to ACOX2/DHCR7-mediated bile acid-cholesterol homeostasis [27].
High values of wavelet-HHL_glcm_SumAverage and wavelet-LHH_glcm_Correlation reflect texture uniformity and directional consistency and may correspond to uniform lipid deposition regulated by FABP4/CYB5A, while the peak signal of wavelet-LHH_firstorder_Maximum may indicate the accumulation of lipid droplets in the PPAR-activated region [28, 29].
lbp-3D-k_glszm_SmallAreaHighGrayLevelEmphasis and wavelet-HLH_glszm_SmallAreaHighGrayLevelEmphasis features emphasizes high signal intensity within the scope of small area, it may map glycolytic active foci with high expression of PFKP/GAPDH, and the heterogeneity of spatial distribution may due to the intratumoral gradient activation of HIF-1 [30, 31].
Wavelet-HHL_glrlm_ShortRunHighGrayLevelEmphasis and wavelet-LHH_gldm_LargeDependenceHighGrayLevelEmphasis respectively capture the short-range high signal area and long-range high signal dependent on area, it may induces sustained reinforcement in the delayed period and correspond to the immunosuppressive microenvironment of SPP1+ TAMs aggregation [27, 32] and the fibrotic interstitial mediated by ARG1. Differences in the spatial distribution of these features may explain the coexistence of pro-inflammatory dominated via HIF-1 and anti-inflammatory dominated via PPAR phenotypes in GPC3-positive tumors [33, 34].
Wavelet-HLL_firstorder_Range and squareroot_firstorder_RobustMeanAbsoluteDeviation prompt pixel intensity of expanding the range of high value, It may correspond to a difference in metabolic gradient between the ARG1-mediated arginine depletion area and surrounding normal liver tissue, as a “halo sign” or edge blurring in enhanced scans [33].
Heterogeneous features of lbp-3d-k_glszm_GrayLevelNonUniformityNormalized and wavelet-HLH_glszm_SmallAreaEmphasis may reflect oxidation sterol accumulation caused by the dysfunction of DHCR7/ACOX2 [27, 32], the activation of PPAR-induced lipophagy may alter the pattern of fat infiltration at tumor margin.
The peak distribution of lbp-3D-k_firstorder_Kurtosis may reflect AKR1B1-mediated oxidative stress regions (such as necrotic lesions), and ROS accumulation through the Warburg effect may lead to high and low mixed signals in imaging [35].
Based on spatial distribution of the wavelet-HHL_glcm_SumAverage uniformity and wavelet-HLH_glszm_SmallAreaHighGrayLevelEmphasis glycolysis, HCC can be noninvasively divided into two metabolic subtypes: Glycolytic-oriented which is appropriate for HIF-1 inhibitors combined with anti-PD-1 therapy to inhibit lactate-mediated immune escape, and lipometabolic-oriented which is appropriate for PPAR inhibitors to block fatty acid uptake. Besides, before and after the treatment, the declining of lbp-3d-k_glszm_SmallAreaHighGrayLevelEmphasis could prompt PFKP inhibition, and the decrease of squareroot_firstorder_MeanAbsoluteDeviation may reflect metabolic homogenization because of sensitive to treatment.
Therefore, we can assume that the medical imaging of GPC3+ patients in HCC shows that GPC3 directly acts on HIF-1/PPAR pathway or indirectly affects HIF-1/PPAR pathway activity, thereby causing tumor hypoxia and metabolic reprogramming, that’s perhaps why high-risk group has a highly correlation to poor prognosis and COX proportional risk model has good forecasting ability (C-index = 0.720) even hub genes without cox regression screening.
Several limitations warrant consideration. First, this research was designed retrospectively, future prospective validation is underway to assess real-world performance and generalizability. Second, although the approach based once-T1WI is both reasonable and reliable-given the robust performance metrics observed, it may limit the capture of diffusion-specific tumor characteristics lacking the radiogenomic analyses with DWI sequence. Moreover, our analysis did not perform a single-target region analysis due to sample size constraints, which might further refine prognostic insights. Future studies with larger cohorts and multi-target analyses may help overcome these limitations. Second, although multi-database analysis mitigates this, future studies should validate HIF-1/PPAR protein expression in GPC3+ tumors and correlate with radiomic features. Third, the biological basis linking specific radiomic features to metabolic reprogramming remains hypothetical, further spatial transcriptomics or multi-region radiogenomic mapping is ideal but was beyond the scope of this retrospective analysis. Such independent sample validation of IHC or prospective matched MRI-tissue cohort is planned as the next step. Forth, the predicting effect of the external test set seemed inferior to that of the internal validation set, which was considered to be the influence of mild overfitting or batch effect. We hypothesis that a major contributor to lower external AUC is prevalence mismatch that our internal cohorts had a 3:1 positive:negative ratio, whereas the independent external cohort had a different case mix. While we applied SMOTE within cross-validation, prevalence differences across cohorts can still degrade external performance; future prospective cohorts with balanced sampling or reweighting approaches are needed. Meanwhile, although ComBat was applied appropriately, residual inter-center differences in radiomic distributions remain and likely contributed to the AUC drop, suggesting that applying ComBat directly to radiomic features might further mitigate site effects in future analyses. To further minimize potential bias and enhance model generalizability, future strategies including class-weighted learning and probability recalibration, feature stability assessment with ICC analysis, nested cross-validation donduction, and so forth.
Unlike prior radiogenomic studies focusing on Wnt/β-catenin pathway-driven imaging traits [36], our work uncovers HIF-1 and PPAR as novel regulators shaping HCC radiophenotypes via metabolic reprogramming. This aligns with emerging evidence that GPC3 not only promotes Wnt signaling but also modulates hypoxia adaptation through HIF-1 stabilization, suggesting a dual oncogenic role targetable by metabolic inhibitors. Tumors with signatures indicating PPAR/lipid-metabolism activation may have distinct immunometabolic microenvironments and might be considered for metabolic modulation strategies or combined approaches with immunotherapy. For GPC3, because it is a surface tumor antigen, MRI-predicted GPC3 positivity may help preselect patients for GPC3-targeted therapies (antibodies, CAR-T, ADCs). We cite relevant literature and emphasize that our radiogenomic model would need prospective validation before clinical decision-making.

Conclusion
In a word, our study pioneers a radiomic framework to non-invasively decode GPC3-associated metabolic vulnerabilities in HCC. By linking radscore to HIF-1/PPAR-driven TME remodeling, we provide a rationale for image-guided combination therapies and lay the groundwork for longitudinal monitoring of metabolic adaptation during treatment. These findings personalized risk stratification and inform clinical decision-making in HCC management, paving the way for more precise and effective therapeutic strategies.

Electronic supplementary material
Below is the link to the electronic supplementary material.