Redox-senescence function of PON1 in hepatocellular carcinoma and its non-invasive assessment using super-resolution radiomics: a multi-center study

Abstract

Background

Risk stratification in hepatocellular carcinoma (HCC) is limited by the lack of robust biomarkers reflecting tumor biology. The antioxidant enzyme Paraoxonase-1 (PON1) shows prognostic potential, yet its role in tumor tissues and the feasibility of non-invasive assessment remain unclear.

Methods

Senescence-related pathways and prognostic candidates were screened using transcriptomic data from TCGA-LIHC and GTEx. PON1 expression was validated in multicenter cohorts through qPCR, immunohistochemistry, and Western blotting. Functional assays in PON1 knockdown and overexpression models evaluated oxidative stress, glutathione balance, mitochondrial dysfunction, and senescence markers. We developed a radiomics model based on contrast-enhanced CT scans with super-resolution reconstruction to predict tumoral PON1 expression. This radiomics signature was then integrated with clinical variables to build a combined model, which was evaluated across training, validation, test, and external cohorts.

Results

PON1 was identified as a downregulated senescence-related prognostic gene. Low PON1 expression was associated with poorer overall and progression-free survival across independent clinical cohorts. PON1 depletion increased intracellular and mitochondrial ROS, lowered the GSH/GSSG ratio, impaired mitochondrial membrane potential, and induced senescence phenotypes, while its restoration mitigated these effects. SR-enhanced radiomics improved prediction of tumoral PON1 expression across all cohorts. Integration of radiomics signatures with clinical variables further improved discrimination, achieving the highest accuracy and net clinical benefit.

Conclusion

PON1 downregulation contributes to oxidative stress–driven senescence and unfavorable clinical outcomes in HCC. SR-enhanced radiomics provides an accurate, non-invasive method for estimating tumoral PON1 expression, demonstrating potential value for radiogenomic profiling and preoperative risk stratification.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-026-08519-x.

Introduction

Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality worldwide. Despite advances in early detection and therapeutic strategies, long-term outcomes for most patients remain poor [1, 2]. One major challenge is the substantial biological heterogeneity of HCC, which limits accurate risk stratification and impedes personalized therapy decisions [3–5]. Consequently, there is an urgent need to identify molecular regulators that drive tumor behavior and to develop reliable, non-invasive tools capable of characterizing the underlying tumor biology.

Cellular senescence is increasingly recognized as a key tumor-intrinsic process in HCC. Rather than acting solely as a tumor-suppressive barrier, senescent tumor cells can develop a senescence-associated secretory phenotype (SASP) that promotes inflammation, extracellular-matrix remodeling, and immune dysregulation within the tumor microenvironment [6]. Senescence-associated pathways are commonly upregulated in HCC and have been implicated in promoting tumor progression, particularly under oxidative stress and metabolic disturbance [7–9]. Although several senescence-related transcriptional signatures have been proposed, the molecular regulators that mechanistically link senescence biology to aggressive tumor behavior remain largely unknown [10]. Identifying such regulators is essential both for elucidating disease biology and for developing clinically meaningful biomarkers for risk stratification.

Paraoxonase-1 (PON1) is a liver-derived antioxidant enzyme that detoxifies lipid peroxides and maintains redox homeostasis [11]. Although PON1 has been widely investigated in cardiovascular and metabolic disorders, emerging evidence indicates its potential relevance in liver diseases characterized by oxidative stress and inflammation [12, 13]. In HCC, PON1 has mainly been explored in multigene bioinformatic signatures or as a serum biomarker for microvascular invasion and AFP-negative tumors, yet its functional role in tumor tissues remains largely unexplored [14–17]. Given that oxidative stress is a key upstream trigger of hepatocyte senescence and that PON1 is a major enzymatic regulator of lipid peroxide clearance, PON1 may in turn influence redox- and senescence-related tumor biology. Nevertheless, whether PON1 directly contributes to tumor progression by modulating redox imbalance or senescence phenotypes remains unknown, representing a critical gap in current knowledge.

With increasing interest in radiogenomics, non-invasive imaging-based estimation of tumor molecular features has emerged as a promising complement to tissue-based assessment in HCC [18, 19]. Radiomics can extract quantitative descriptors of intratumoral heterogeneity, yet its performance is inherently limited by the spatial resolution of conventional CT imaging [20]. Low-resolution images may obscure fine-scale texture patterns associated with tumor biology, thereby reducing model robustness and multi-center generalizability. Deep learning–based super-resolution (SR) reconstruction has recently been proposed to enhance anatomical detail and recover high-frequency information beyond the capacity of conventional interpolation [21–24]. Although SR has shown encouraging results in various oncologic imaging tasks, its potential to improve radiomics-based prediction of molecular expression in HCC remains largely unexplored. SR-enhanced radiogenomic modeling may therefore provide a more accurate and clinically practical approach for preoperative molecular characterization.

In this study, we investigated the clinical and biological relevance of PON1 in hepatocellular carcinoma by integrating public transcriptomic datasets, multicenter patient cohorts, and functional experiments. We assessed whether PON1 downregulation is linked to oxidative stress–related senescence features and poor outcomes. To address the limitations of tissue sampling and the potential clinical value of non-invasive molecular assessment, we further developed CT-based radiomics models to predict intratumoral PON1 expression and evaluated whether SR reconstruction could enhance predictive performance. This integrated approach aimed to improve understanding of redox- and senescence-related mechanisms in HCC and to establish a practical imaging strategy for preoperative molecular characterization.

Methods

Patient cohorts

This retrospective study included a total of 238 patients from Zhengzhou University People’s Hospital and 82 patients from Henan Cancer Hospital between January 2020 and September 2025. The inclusion criteria were as follows: [1] histopathologically confirmed HCC with available complete tissue sections; [2] availability of preoperative contrast-enhanced liver CT imaging. Patients were excluded if: [1] histological sections or imaging data were incomplete; [2] the CT images were of insufficient quality, based on predefined criteria listed in Table S1. Patients from Zhengzhou University People’s Hospital were randomly split into the training and validation cohorts in a 7:3 ratio using a computer-generated randomization sequence. The overall patient selection workflow is illustrated in Fig. S1.

Ethic approval

This study was conducted in accordance with the ethical principles of the Declaration of Helsinki and was approved by the Ethics Committee of Zhengzhou University People’s Hospital (Approval No. (2024) Ethical Review 176) and the Ethics Committee of Henan Cancer Hospital (Approval No. 2022-455-001). Given the retrospective nature of the study, the requirement for informed consent was waived.

Public database information

A total of 416 HCC cases with available RNA-seq data and clinical annotations were assessed from the TCGA-LIHC cohort. For comparison with non-tumor liver tissues, 104 normal liver samples were downloaded from the GTEx database. Additionally, contrast-enhanced CT images corresponding to 49 TCGA-LIHC patients were obtained from the TCIA repository and used as an external cohort for radiomics validation. Only samples with complete clinical information and adequate image or sequencing quality were included in the final analyses.

Data collection

Preoperative clinical variables, conventional imaging features, and survival outcomes (overall survival (OS) and progression-free survival (PFS)) were collected. OS was defined as the interval from surgery to death or last follow-up. PFS was defined as the interval from surgery to radiologically confirmed tumor recurrence. Detailed variable lists and follow-up procedures are provided in the Supplementary Material.

Cell culture

The human HCC cell lines Huh7 (Cat. No. CL-0120) and HEK293T (Cat. No. CL-0005) were obtained from Procell Life Science & Technology (Wuhan, China), and MHCC97H cells (Cat. No. iCell-h143) were purchased from iCell Bioscience Inc. (Shanghai, China). All cell lines were authenticated by short tandem repeat (STR) profiling prior to experimentation and were verified to be mycoplasma-free using the ATCC Universal Mycoplasma Detection Kit. Cells were maintained in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10% fetal bovine serum (FBS), 100 U/mL penicillin, and 100 µg/mL streptomycin, and incubated at 37 °C in a humidified atmosphere containing 5% CO₂.

Cellular and functional experiments

The human HCC cell lines Huh7 and MHCC97H were used to establish PON1 knockdown and overexpression models, respectively, via lentiviral transduction (Fig. S2). The impact of PON1 on cellular phenotypes was assessed through functional assays measuring migration and invasion. To mechanistically evaluate PON1, we assessed the following: oxidative stress, encompassing both intracellular and mitochondrial reactive oxygen species; glutathione redox balance; mitochondrial membrane potential; and cellular senescence. Detailed protocols for cell culture, lentiviral construction, and all biochemical assays are provided in the Supplementary Material.

DCFH-DA assay

Intracellular reactive oxygen species (ROS) levels were measured using a DCFH-DA fluorescent probe kit (Beyotime, Cat. No. S1105S). According to the manufacturer’s instructions, DCFH-DA was diluted in serum-free medium to prepare a 10 µM working solution. After removing the culture medium, cells were incubated with the DCFH-DA working solution at 37 °C in 5% CO₂ for 20–30 min. During incubation, plates were gently agitated every 5–10 min to facilitate probe uptake. At the end of incubation, cells were washed three times with serum-free medium to remove residual extracellular probe. Fluorescence intensity was then immediately measured using a fluorescence microscope or a microplate reader.

Measurement of mitochondrial reactive oxygen species

Mitochondrial ROS (mtROS) levels were measured using MitoSOX™ Red (Thermo Fisher Scientific, Cat. No. M36008). Huh7 and MHCC97H cells were incubated with 2.5 µM MitoSOX™ working solution for 10 min at 37 °C, followed by washing with PBS to remove residual dye. Mean fluorescence intensity (MFI) was quantified by flow cytometry, and data were analyzed using FlowJo software (version 10).

Measurement of mitochondrial membrane potential

Mitochondrial membrane potential (ΔΨm) was assessed using the JC-1 Mitochondrial Membrane Potential Assay Kit (Beyotime, Cat. No. C2006). After experimental treatments, Huh7 and MHCC97H cells were incubated with JC-1 working solution (5 µg/mL) for 20 min at 37 °C in the dark. Cells were then washed twice with JC-1 staining buffer and resuspended in fresh buffer for flow cytometric analysis. JC-1 aggregates (red fluorescence, FL2) and JC-1 monomers (green fluorescence, FL1) were detected by flow cytometry. Mitochondrial membrane potential was calculated as the FL2/FL1 fluorescence ratio, and data were processed using FlowJo (version 10).

Measurement of lipid peroxidation

Lipid peroxidation levels were assessed using the Lipid Peroxidation Assay Kit with BDPY 581/591 C11 (Beyotime, Cat. No. S0043S). Following experimental treatments, Huh7 and MHCC97H cells were incubated with 2 µM BDPY 581/591 C11 working solution for 20 min at 37 °C in the dark. After incubation, the cells were washed with PBS to remove the residual probe. The oxidation of BDPY 581/591 C11 by lipid peroxides results in a fluorescence shift from red to green, which was detected by flow cytometry. The extent of lipid peroxidation was quantified by calculating the ratio of green to red mean fluorescence intensity, and the data were analyzed using FlowJo software (version 10).

Apoptosis assay

Apoptosis was assessed using the Annexin V-FITC/Propidium Iodide (PI) Apoptosis Detection Kit (Beyotime, Cat. No. C1062L). After treatment, cells were harvested, washed twice with cold PBS, and resuspended in binding buffer. Each sample was incubated with 5 µL Annexin V-FITC and 10 µL PI for 15 min at room temperature in the dark. Apoptotic cells were quantified by flow cytometry. Cells positive for Annexin V-FITC alone (Annexin V⁺/PI⁻) or double-positive for Annexin V-FITC and PI (Annexin V⁺/PI⁺) were classified as apoptotic. Data were analyzed using FlowJo (version 10).

SA-β-gal staining

Senescence-associated β-galactosidase (SA-β-gal) staining was performed using a commercial kit (Beyotime, Cat. No. C0602). Cells were fixed with β-galactosidase fixative solution for 15 min at room temperature, rinsed with PBS, and then incubated with freshly prepared X-gal staining working solution at 37 °C in a CO₂-free incubator overnight. Stained cells were observed under a bright-field microscope, and the percentage of SA-β-gal–positive cells was quantified based on randomly selected fields (n = 5 per sample).

Image acquisition and super-resolution reconstruction

Contrast-enhanced CT images from two centers (n = 320) were acquired and anonymized. To enhance spatial resolution before radiomics analysis, images were reconstructed using the deep learning–based 3D SR framework. Both the original non-reconstructed (NR) images and the reconstructed SR images were retained for comparative radiomics modeling. Preprocessing procedures are detailed in Supplementary Material.

Tumor segmentation and ROI delineation

Tumor regions of interest (ROIs) were manually segmented using ITK-SNAP (version 3.8). Initial delineation was performed by a hepatobiliary surgeon with 5 years of experience based on visible tumor boundaries. Interobserver agreement was assessed by a senior surgeon with 10 years of experience who repeated segmentation in a randomly selected subset of 50 cases. Intraclass correlation coefficients (ICCs) for radiomic features extracted from the two segmentations were computed, and features with ICC > 0.75 were considered robust and retained for further analyses.

Radiomic feature extraction

Radiomic features were extracted from both NR and SR images using PyRadiomics. A total of 1,538 features were computed for each ROI, including:

(1) Shape (n = 14); (2) First-order statistics (n = 302); (3) GLCM (n = 369); (4) GLDM (n = 234); (5) GLRLM (n = 268); (6) GLSZM (n = 268); (7) NGTDM (n = 83). Wavelet-transformed images were included to capture multi-scale patterns. Voxel intensities were discretized into 25 Gy-level bins. All features were standardized using Z-score normalization prior to modeling.

Feature selection and dimensionality reduction

Dimensionality reduction was performed in the training cohort using LASSO logistic regression with 5-fold cross-validation. Features with non-zero coefficients at the optimal penalty parameter were retained, yielding 13 radiomic features for model construction.

Radiomics model development

Five supervised machine-learning classifiers were trained using NR- and SR-derived features: (1) Logistic Regression (LR); (2) Support Vector Machine (SVM); (3) k-Nearest Neighbors (KNN); (4) Random Forest (RF); (5) ExtraTrees.

Models were trained in the training cohort and evaluated in internal validation, internal test, and external TCIA cohorts. A radiomics score (Radscore) was generated for each patient from the corresponding classifier output. Both NR-based and SR-based radiomics pipelines followed identical workflows to ensure methodological comparability.

Clinical variable analysis and clinical–radiomics integration

Clinical variables were collected for all participants. Group comparisons were performed to assess baseline differences. Univariate and multivariate logistic regression analyses were performed in the training cohort to identify independent predictors of PON1 expression. Three predictive models were constructed: (1) clinical model; (2) radiomics model; and (3) integrated clinical–radiomics model. A nomogram was generated to visualize the integrated model. Model performance was assessed via: (1) ROC curve analysis; (2) calibration curves; and (3) decision curve analysis.

Statistical analysis

Statistical analyses were conducted using Python (3.12.0), SPSS (27.0), and R (4.5.2). Categorical variables were reported as frequencies and percentages. Continuous variables were expressed as mean ± SD or median (IQR), depending on distribution. Comparisons between groups were performed using the χ² test or Fisher’s exact test for categorical variables and the Mann–Whitney U test for continuous variables. OS and PFS were analyzed using Kaplan–Meier curves and compared using the log-rank test. Time-dependent ROC curves were utilized to assess the prognostic performance at 1, 2, and 3 years. For multiple hypothesis testing, P values were adjusted using the Benjamini-Hochberg false discovery rate (FDR) method. A two-sided P < 0.05 was considered statistically significant. Statistical significance in figures and tables was annotated as: NS, not significant; * P < 0.05; ** P < 0.01; *** P < 0.001.

Results

Transcriptomic senescence signatures and prognostic gene identification in HCC

GSEA of TCGA-LIHC and GTEx data revealed significant enrichment of senescence-associated pathways, particularly cell-cycle checkpoint signaling and regulation of cell-cycle phase transition in HCC tissues, with core regulators (TP53, P21, P16, RB1) at the leading edge (Fig. S3A-D). To identify senescence-associated genes with prognostic relevance, we next performed multivariate Cox regression in the TCGA-LIHC cohort, which yielded 27 genes significantly associated with overall survival (Fig. S4A-D). These candidates were further refined using five machine-learning algorithms. The consensus across these models identified a panel of 15 robust prognostic genes (Fig. S4E-L). All 15 genes showed consistent and significant differential expression between HCC and normal liver tissues in GTEx (Fig. S4M-N).

PON1 identified as a senescence-associated prognostic gene implicated in redox dysregulation

Cross-referencing the 15 prognostic genes with senescence-associated gene sets from Aging Atlas and GeneCards identified PON1 and TNFRSF11B as potential senescence-related regulators (Fig. 1A–B). Among these candidates, PON1 showed consistently reduced expression in tumor tissues across TCGA-LIHC and GSE14520 (Fig. 1C–E, H), whereas TNFRSF11B displayed variable patterns (Fig. 1G, I). We therefore selected PON1 as the most biologically plausible senescence-associated gene. High PON1 expression was associated with prolonged overall survival in both cohorts and improved progression-free survival in GSE14520 (Fig. 1K–M). To control for age-related differences between datasets (PON1’s association with age > 65 years in TCGA-LIHC (Fig. S5A) but only a non-significant trend in GSE14520 (Fig. S5F)), age-stratified analyses were performed. Low PON1 remained strongly associated with poorer prognosis in both age ≥ 65 and < 65 subgroups (Fig. S5K–P), confirming its independent prognostic value.

Fig. 1.

Identification of PON1 as a senescence-associated prognostic gene and its associated transcriptomic features. (A) Venn diagram of the 15 prognosis-related genes identified by multi-algorithm machine learning. (B) Overlap between these genes and senescence gene sets from Aging Atlas and GeneCards. (C) Volcano plot of differentially expressed genes between TCGA-LIHC and GTEx. (D-E) Boxplot comparing PON1 (D) and TNFRSF11B (E) expression in TCGA-LIHC and GTEx. (F) Clinical feature heatmap of PON1 expression in TCGA-LIHC. (G) Volcano plot of differentially expressed genes in GSE14520. (H-I) Boxplot validating PON1 (H) and TNFRSF11B (I) expression in tumor vs. adjacent tissues (GSE14520). (J) Clinical heatmap of PON1 expression in GSE14520. (K-M) Kaplan–Meier curves of overall survival in TCGA (K) and GSE14520 (L) and progression-free survival in GSE14520 (M). (N-P) GO and KEGG enrichment analyses comparing Q1 vs. Q4 of PON1 expression

Comparative enrichment analyses between the lowest (Q1) and highest (Q4) PON1 expression quartiles revealed that PON1-low tumors were enriched in flavin-dependent oxidoreductase activity and cytochrome P450-related metabolic pathways, suggesting increased oxidative burden and altered redox metabolism (Fig. 1N–P).

Clinical validation of PON1 downregulation and its prognostic impact

To validate the transcriptomic findings in clinical samples, we examined PON1 expression across multiple independent patient cohorts. Quantitative PCR in 86 paired tumor-adjacent tissues demonstrated a marked reduction of PON1 expression in HCC tumors compared with adjacent liver (Fig. 2A-B). This pattern was further confirmed at the protein level, as immunohistochemistry in 320 patients consistently showed decreased PON1 expression in tumor (Fig. 2C-E), and Western blotting in 10 representative pairs yielded concordant results (Fig. 2F). After stratifying patients into training, validation, and test sets, survival analyses demonstrated that patients with low tumoral PON1 expression were associated with significantly shorter OS and PFS across all cohorts (Fig. 2G-H), supporting the clinical relevance and prognostic robustness of PON1 downregulation in HCC.

Fig. 2.

Multicenter validation of tumor-specific downregulation of PON1 and its prognostic significance in HCC. (A) Relative PON1 mRNA expression in 86 paired primary HCC and adjacent non-tumor tissues measured by qPCR. (B) Quantile–quantile (Q–Q) plot comparing the distribution of relative PON1 mRNA expression between primary HCC tissues and paired adjacent non-tumor tissues. (C) Representative IHC staining of PON1 in adjacent non-tumor liver and primary HCC. (D-E) Quantification of PON1 IHC scores of primary HCC and adjacent non-tumor tissues in Cohort 1 (n = 238) (D) and Cohort 2 (n = 82) (E). (F) Western blot analysis of PON1 protein expression in paired tumor (T) and adjacent non-tumor (N) tissues from 10 patients. (G) Kaplan–Meier OS curves for the training (n = 168), validation (n = 70), and external test (n = 82) cohorts, stratified by PON1 expression. (H) Kaplan–Meier PFS curves for the same three cohorts

Cellular models demonstrate the tumor-suppressive role of PON1

To determine the role of PON1 in HCC, we evaluated its expression across multiple HCC cell lines (Fig. 3A). Same-gel Western blotting confirmed high PON1 levels in Huh7 cells and low levels in MHCC97H cells (Fig. 3B), establishing these lines as models for loss- and gain-of-function experiments. We silenced PON1 in Huh7 cells using shRNAs, prioritizing shPON1#1 for further assays based on its knockdown efficiency (Fig. S6A). We also established PON1-overexpressing MHCC97H cells via lentiviral transduction (Fig. 3B). PON1 silencing increased Huh7 cell migration and invasion, whereas PON1 overexpression suppressed these phenotypes in MHCC97H cells (Fig. 3C-H). Overexpressing PON1 in Huh7 cells or knocking it down in MHCC97H cells produced no further phenotypic changes (Fig. 3D-E, G-H), consistent with phenotypic saturation at their respective baselines. Given that these reverse models exhibited phenotypic saturation, we primarily utilized the Huh7-knockdown and MHCC97H-overexpression models for subsequent mechanistic investigations. To exclude cell-line-specific effects, we evaluated a third HCC cell line, PLC/PRF/5. Similar to the Huh7 model, PON1 knockdown in PLC/PRF/5 cells (Fig. 3B) promoted cell migration and invasion (Fig. S6B). These results indicate that PON1 suppresses HCC cell motility and invasiveness.

Fig. 3.

PON1 suppresses migration and invasion in hepatocellular carcinoma cells. (A) Western blot analysis of PON1 protein expression levels in five HCC cell lines. (B) Validation of PON1 knockdown and overexpression efficiencies in Huh7, MHCC97H, and knockdown in PLC/PRF/5 cells. (C) Representative images of Transwell migration and invasion assays for Huh7 cells. (D-E) Quantification of migrated (D) and invaded (E) Huh7 cells. (F) Representative Transwell migration and invasion images of MHCC97H cells with PON1 overexpression and knockdown. (G-H) Quantification of migrated (G) and invaded (H) MHCC97H cells

PON1 preserves mitochondrial redox homeostasis and prevents mitochondrial dysfunction

We next examined its impact on intracellular oxidative stress and mitochondrial function. DCFH-DA fluorescence analysis showed that PON1 knockdown markedly increased intracellular ROS levels to an extent comparable to exogenous H₂O₂ stimulation, whereas PON1 overexpression significantly reduced ROS accumulation (Fig. 4A-B). Consistently, the reduced GSH/GSSG ratio in PON1-deficient cells further indicated a shift toward a pro-oxidant state intracellular environment, supporting the notion that PON1 helps maintain cellular redox homeostasis (Fig. 4C-D). Following PON1 silencing, MitoSOX staining revealed a pronounced increase in mitochondrial ROS. This was accompanied by a significant decline in the mitochondrial membrane potential, as measured by the JC-1 ratio (Fig. 4E-L), collectively indicating impaired mitochondrial function. Furthermore, as a downstream consequence of GSH depletion and severe oxidative stress, C11-BODIPY flow cytometry confirmed that PON1 knockdown drove a significant accumulation of lipid peroxides within the cells (Fig. 4M-P).

Fig. 4.

PON1 maintains mitochondrial redox homeostasis, stabilizes mitochondrial membrane potential, and suppresses lipid peroxidation in HCC cells. (A-B) Representative fluorescence images of intracellular ROS levels measured by DCFH-DA staining in PON1-knockdown Huh7 cells (A) and PON1-overexpression MHCC97H cells (B), under oxidative stress modulation (H₂O₂ or NAC). (C-D) Quantification of intracellular GSH/GSSG ratios in Huh7 (C) and MHCC97H (D) cells. (E-H) MitoSOX-based flow cytometric analysis of mitochondrial ROS levels and corresponding quantification in Huh7 (E-F) and MHCC97H (G-H) cells. (I-L) JC-1 flow cytometry analysis of mitochondrial membrane potential in Huh7 (I-J) and MHCC97H (K-L) cells. (M-P) Flow cytometric analysis and quantification of lipid peroxidation using the C11-BODIPY probe in Huh7 (M-N) and MHCC97H (O-P) cells

PON1 restrains cellular senescence and the senescence-associated secretory phenotype

Despite increased oxidative stress and mitochondrial dysfunction, Annexin V/PI staining showed no significant difference in apoptosis, suggesting that PON1 primarily modulates cellular function rather than inducing cell death (Fig. 5A–D). SA-β-gal staining demonstrated a substantial increase in senescent cells upon PON1 knockdown, accompanied by upregulation of p53 and p21 and downregulation of pRb, PCNA, and CCNA2 (Fig. 5E-J). As senescent tumor cells actively remodel the microenvironment through the SASP [25], we subsequently quantified key SASP factors. ELISA revealed that PON1 depletion significantly elevated the secretion of IL-6, IL-1β, and TNF-α in vitro, a hyper-secretory pattern consistently mirrored in the serum of HCC patients with low tumoral PON1 expression (Fig. 5K-L). Together, these results indicate that PON1 restrains cellular senescence by preserving mitochondrial redox balance, thereby inhibiting the acquisition of pro-invasive and pro-inflammatory phenotypes.

Fig. 5.

PON1 restrains cellular senescence and the senescence-associated secretory phenotype (SASP) without inducing cell death in HCC. (A–D) Annexin V/PI staining and quantification of apoptotic cells in Huh7 (A–B) and MHCC97H (C–D) cells. (E–H) Representative SA-β-gal staining images (E, G) and quantification of senescent cells (F, H). (I–J) Western blot analysis of senescence-related proteins (p53, p21, pRb, PCNA, and CCNA2) in Huh7 (I) and MHCC97H (J) cells under oxidative stress–modulating treatments. (K) ELISA quantification of SASP factors (IL-6, IL-1β, TNF-α, and IFN-γ) in the culture supernatants of Huh7 and MHCC97H cells. (L) ELISA quantification of circulating SASP factors (IL-6, TNF-α, IL-1β, and IFN-γ) in the serum of clinical HCC patients, stratified by tumoral PON1 expression

Consistent with the phenotypic saturation observed in our earlier migration and invasion assays, reverse genetic interventions (PON1 overexpression in Huh7 or knockdown in MHCC97H) failed to further alter oxidative stress, mitochondrial function, or senescence markers. In contrast, PON1 knockdown in the intermediate-high PLC/PRF/5 cell line fully reproduced the severe oxidative stress, mitochondrial dysfunction, and SASP phenotypes observed in the Huh7 model (Fig. S6C-G). These data confirm that the PON1-mediated regulation of mitochondrial redox balance and cellular senescence is highly dependent on endogenous baseline levels and represents a generalized mechanism across HCC cell lines.

To rule out potential shRNA off-target effects, we performed genetic rescue experiments. Introducing an shRNA-resistant exogenous PON1 construct into PON1-knockdown Huh7 cells normalized migration and invasion, attenuated ROS accumulation, and reversed the SASP-related senescence phenotypes (Fig. S7).

Super-resolution CT radiomics enables non-invasive prediction of tumoral PON1 expression

To enable preoperative assessment of tumoral PON1 expression, we established CT-based radiomics pipelines using both original NR images and SR–enhanced images (Fig. 6). A total of 1,538 radiomic features were extracted (Fig. 7A-B). LASSO regression using 5-fold cross-validation identified 13 non-zero-coefficient features as the final radiomic signature (Table S2), which was used to train five supervised classifiers (Fig. 7C-E; Table S3). Comparative evaluation across the training, validation, test, and TCIA cohorts revealed that SR-based logistic regression achieved stable discrimination, with AUCs of 0.818, 0.767, 0.710, and 0.749, respectively. In contrast, the best-performing NR model (ExtraTrees) exhibited reduced generalizability, particularly in the external TCIA cohort (AUC 0.513), indicating poor cross-center generalizability (Fig. 7F-G). These findings demonstrate that SR reconstruction enhances spatial fidelity and texture robustness.

Fig. 6.

Workflow of super-resolution reconstruction and generation of high-quality CT images. A GAN-based super-resolution model generated high-resolution CT images from low-resolution inputs using a generator–discriminator architecture. Original enhanced CT scans were processed by the trained generator to produce super-resolved images with higher spatial clarity and improved tumor depiction, enabling more reliable radiomics feature extraction

Fig. 7.

Radiomics, clinical, and combined models for predicting tumoral PON1 expression. (A–B) Distribution of 1,538 extracted radiomics features across seven categories and corresponding value distributions. (C–D) LASSO regression for feature selection: coefficient profiles (C) and 5-fold cross-validation for optimal λ identification (D). (E) Thirteen radiomics features with non-zero coefficients retained after LASSO selection. (F) ROC curves of the SR-radiomics model built using logistic regression across the training, validation, test, and TCIA cohorts. (G) ROC curves of the optimal NR-radiomics model built using ExtraTrees across the same four cohorts. (H) Univariate and multivariate logistic regression identifying clinical predictors associated with high tumoral PON1 expression. (I) Nomogram integrating Radscore, age, peritumoral star sign, and AFP to construct the combined model for individualized prediction. (J) Comparative ROC curves of combined, clinical, and SR-radiomics models across the three internal cohorts. (K) Decision curve analysis (DCA) of the three models in the training, validation, and test cohorts. (L) Calibration curves illustrating agreement between predicted and observed probabilities for the three models across the three cohorts

Integrated clinical–radiomics modeling improves predictive performance

Then we integrated significant clinical predictors with the SR-derived radiomic signature (Table S4). In the training cohort, multivariate logistic regression identified age > 65 years as an independent predictor of high PON1, whereas peritumoral star sign and AFP > 400 ng/mL were independently associated with low PON1 (Table S5). Based on these variables, three models were constructed—a clinical model, an SR-radiomics model, and an integrated clinical-radiomics model—and presented as nomograms for individualized estimation (Fig. 7H-I). The integrated model consistently outperformed both individual models, achieving AUCs of 0.852, 0.878, and 0.805 in the training, validation, and test cohorts, respectively, compared with 0.818, 0.767, 0.710 for the radiomics model and 0.805, 0.784, 0.733 for the clinical model (Table S6; Fig. 7J). Decision curve analysis demonstrated that the integrated model provided the greatest net clinical benefit across a wide range of threshold probabilities, and calibration curves confirmed excellent agreement between predicted and observed outcomes (Fig. 7K-L). These results indicate that combining imaging-derived heterogeneity with clinical markers yields a more accurate and clinically practical tool for the non-invasive preoperative assessment of tumoral PON1 expression, potentially guiding personalized management.

PON1 and radiomics models exhibit superior long-term prognostic performance

To explicitly demonstrate the clinical utility of our approach and validate the rationale for predicting an intermediate biomarker (PON1) rather than directly modeling survival from radiomic features, we performed time-dependent ROC analysis for 1-, 2-, and 3-year overall survival. We systematically compared our intermediate PON1 model against a traditional clinical factor-based model (Table S7) and a direct radiomics-to-survival model (Table S8). While the clinical model exhibited adequate short-term predictive value for 1-year OS, its discriminative ability rapidly deteriorated over time across all cohorts. In stark contrast, both the PON1 model and the direct radiomics model demonstrated superior robustness and stability for long-term prognosis at 2 and 3 years (Fig. 8A-C; Table S9). Most importantly, predicting the intermediate biomarker achieved long-term prognostic accuracy highly comparable to the pure radiomics survival model, while crucially preserving mechanistic transparency by linking macroscopic imaging phenotypes directly to the microscopic redox-senescence tumor biology.

Fig. 8.

Time-dependent ROC curves comparing the predictive performance of the Clinical, PON1, and Radiomics models for 1-, 2-, and 3-year overall survival. (A) Time-dependent ROC curves for 1-year, 2-year, and 3-year overall survival in the training cohort. (B) Time-dependent ROC curves for 1-year, 2-year, and 3-year overall survival in the validation cohort. (C) Time-dependent ROC curves for 1-year, 2-year, and 3-year overall survival in the test cohort

Discussion

Our study establishes PON1 as a senescence-associated regulator that suppresses invasive phenotypes in HCC by maintaining redox homeostasis, with its reduced expression associated with unfavorable prognosis across different age groups. We further translated this finding into a non-invasive imaging tool by developing a CT radiomics model, where super-resolution reconstruction enhanced spatial fidelity, improving its performance for estimating intratumoral PON1 expression.

Senescent tumor cells remodel the microenvironment mainly through SASP factor release, which promotes angiogenesis and epithelial–mesenchymal transition, thereby enhancing tumor invasiveness [26, 27]. In this context, our analyses identified PON1 as the only prognostic gene overlapping with curated senescence gene sets, underscoring its potential relevance in senescence regulation in HCC.

In multicenter cohorts, low PON1 expression robustly correlated with worse OS and PFS. Although higher PON1 levels were more frequently observed in older patients, age-stratified analyses confirmed that the prognostic impact of PON1 was independent of age, effectively ruling out age as a confounding factor. The concordant downregulation of PON1 across qPCR, IHC, and Western blotting analyses not only validates our findings but suggests its potential as a reliable tissue-based biomarker in HCC. These results are consistent with and extend previous studies linking intratumoral PON1 expression to HCC prognosis [28, 29].

Mechanistically, our in vitro data indicate that PON1 modulates mitochondrial redox homeostasis and senescence-associated cellular states. PON1 depletion induced a profound redox imbalance, characterized by marked increases in intracellular and mitochondrial ROS, a decreased GSH/GSSG ratio, and loss of mitochondrial membrane potential. Given that PON1 functions as a potent lipid antioxidant, its depletion and the subsequent reduction in the GSH/GSSG ratio naturally drove a significant accumulation of lipid peroxides. Notably, however, this oxidative lipid damage did not cross the threshold required to trigger execution-phase ferroptosis or apoptosis, as evidenced by the preservation of plasma membrane integrity [30]. This indicates that PON1 loss primarily induces a sub-lethal state of metabolic stress, actively diverting the cell fate towards survival through senescence rather than cell death [31]. Consistent with this, PON1 depletion drove a senescent phenotype, as evidenced by elevated SA-β-gal activity and coordinated alterations in key senescence and cell-cycle regulators, linking oxidative stress to the acquisition of invasive properties. Importantly, our findings explicitly demonstrate that this senescent state translates into a robust SASP. The elevated secretion of pro-inflammatory cytokines, including IL-6, IL-1β, TNF-α, was observed not only in PON1-depleted HCC cells but also in the systemic circulation of HCC patients with low tumoral PON1 expression. This systemic SASP burden actively orchestrates a pro-inflammatory and tumor-permissive microenvironment, providing a clear translational link between PON1-mediated intracellular redox imbalance and the observed adverse clinical outcomes [32, 33]. These findings are consistent with previous reports linking mitochondrial dysfunction, excessive ROS accumulation, and senescence-driven invasiveness in cancer progression [34, 35].

In the radiomics analysis, SR reconstruction improved tumor boundary delineation and enhanced the visibility of fine textural details. These improvements translated into more robust predictive performance across all cohorts. The SR-based model also showed better cross-center generalizability than the model built from non-reconstructed images. This implies that the limited spatial resolution of conventional CT may constrain the capture of biologically relevant heterogeneity. Although SR techniques have been applied in other oncologic imaging studies, their use in radiogenomic prediction remains unexplored [21–24]. Our results indicate that SR-enhanced radiomics may provide additional value for non-invasive molecular characterization in HCC.

Building on these findings, we constructed an integrated model that combined SR-derived radiomics features with key clinical variables to improve predictive performance and clinical applicability. In multivariate analysis, AFP > 400 ng/mL and the peritumoral star sign were independently associated with low PON1 expression, whereas age > 65 years predicted higher expression — findings that align with prior studies on the clinical determinants of PON1 status [11, 16]. The integrated model achieved better discrimination, calibration, and clinical utility than either the clinical or radiomics model alone. Thus, the integration of imaging-derived heterogeneity with routine clinical markers provides a more robust framework for the preoperative assessment of molecular features.

A critical conceptual advancement of our study is the rationale to predict an intermediate biological effector instead of directly predicting survival outcomes from raw radiomic features. Direct radiomics-to-survival models often function as analytical “black boxes” and are highly susceptible to confounding by post-surgical interventions or underlying liver disease progression [36]. Indeed, our time-dependent ROC analyses revealed that while traditional clinical features possessed adequate short-term prognostic value, their predictive accuracy rapidly deteriorated over time. In stark contrast, modeling tumoral PON1 expression yielded highly stable and robust long-term prognostic stratification, rivaling the performance of direct high-dimensional radiomics survival models. By non-invasively capturing this specific redox-senescence hallmark, our SR-radiomics approach not only provides long-term prognostic value but also offers mechanistic transparency, potentially identifying patients with highly oxidative, SASP-driven tumors who might benefit from tailored neoadjuvant strategies.

Several limitations should be considered. Mechanistic conclusions are based on in vitro experiments, and in vivo studies are needed to further validate the biological role of PON1. Additionally, although the use of multicenter cohorts enhances robustness, larger prospective studies are required to assess the generalizability of the radiomics and integrated models. Despite these limitations, our work establishes a link between PON1 and senescence-associated redox regulation in HCC, and demonstrates the potential of SR-enhanced radiomics for the non-invasive estimation of tissue-level molecular features. These findings offer a foundation for future studies exploring senescence-related pathways and radiogenomic approaches in hepatocellular carcinoma.

Conclusion

In summary, PON1 downregulation in HCC was associated with oxidative stress–related senescence features and adverse clinical outcomes. The SR-enhanced radiomics model achieved non-invasive prediction of intratumoral PON1 expression, supporting the potential clinical translation of radiogenomic methods for preoperative molecular evaluation.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AFP

Alpha-fetoprotein

ATCC

American Type Culture Collection

AUC

Area under the curve

CT

Computed tomography

DCA

Decision curve analysis

DCFH-DA

2′,7′-dichlorodihydrofluorescein diacetate

GLCM

Gray-level co-occurrence matrix

GLDM

Gray-level dependence matrix

GLRLM

Gray-level run length matrix

GLSZM

Gray-level size zone matrix

GSEA

Gene set enrichment analysis

GSH/GSSG

Reduced/oxidized glutathione ratio

GTEx

Genotype-Tissue Expression

H₂O₂

Hydrogen peroxide

HCC

Hepatocellular carcinoma

ICC

Intraclass correlation coefficient

ITK-SNAP

ITK-SNAP software

JC-1

JC-1 mitochondrial membrane potential dye

KNN

k-nearest neighbors

LASSO

Least absolute shrinkage and selection operator

LR

Logistic regression

MFI

Mean fluorescence intensity

mtROS

Mitochondrial reactive oxygen species

MitoSOX

Mitochondrial superoxide indicator

NGTDM

Neighboring gray tone difference matrix

NR

Non-reconstructed images

OS

Overall survival

PBS

Phosphate-buffered saline

PFS

Progression-free survival

PON1

Paraoxonase-1

PyRadiomics

Python Radiomics package

qPCR

Quantitative polymerase chain reaction

Radscore

Radiomics score

RF

Random forest

ROC

Receiver operating characteristic

ROI

Region of interest

ROS

Reactive oxygen species

SA-β-gal

Senescence-associated β-galactosidase

SASP

Senescence-associated secretory phenotype

SR

Super-resolution

STR

Short tandem repeat

SVM

Support vector machine

TCGA-LIHC

The Cancer Genome Atlas–Liver Hepatocellular Carcinoma

TCIA

The Cancer Imaging Archive

Author contributions

All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Chiyu Cai, Yuqi Hao and Yushu Xue. The first draft of the manuscript was written by Chiyu Cai and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (Grant Nos. 82403462 and 82470653), the Henan Clinical Medical Scientist Program (Grant Nos. HNCMS202403 and HNCMS202504), the Henan Provincial Natural Science Foundation Youth Science Fund (Category A, Grant No. 262300421036), and the Scientific and Technological Project of Henan Province (Grant No. 252102310251).

Data availability

The datasets used and analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethical approval

This study was conducted in accordance with the ethical principles of the Declaration of Helsinki and was approved by the Ethics Committee of Zhengzhou University People’s Hospital (Approval No. (2024) Ethical Review 176) and the Ethics Committee of Henan Cancer Hospital (Approval No. 2022-455-001). Given the retrospective nature of the study, the requirement for informed consent was waived.

Consent for publication

All authors have read and approved the final manuscript and consent to its publication.

Competing interests

The authors have no relevant financial interests to disclose.