Integrated Multi-Omics and Pathomics Analysis for Prognostic Modeling of Hepatocellular Carcinoma Reveals Oxidative Stress-Driven Immunometabolic Regulatory Mechanisms

Integrated Multi-Omics and Pathomics Analysis for Prognostic Modeling of Hepatocellular Carcinoma Reveals Oxidative Stress-Driven Immunometabolic Regulatory Mechanisms | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Integrated Multi-Omics and Pathomics Analysis for Prognostic Modeling of Hepatocellular Carcinoma Reveals Oxidative Stress-Driven Immunometabolic Regulatory Mechanisms Lunwei Yang, Yuanliang Li, Xiaoying Zhong This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7489138/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Hepatocellular carcinoma (HCC), a major cause of cancer mortality, exhibits strong ties to oxidative stress (OS), though integrated multi-omics studies linking OS mechanisms to clinically predictive models remain scarce. To address this, we integrated European-descent GWAS data (2,852 HCC cases vs. 447,587 controls) with 1,065 OS-related genes, identifying 176 potential HCC-associated genes (P < 0.05) via TWAS (UTMOST/GBJ tests), including 12 key OS drivers. Pathomic features extracted from 379 TCGA HCC histopathological images (ResNet-50/CellProfiler) informed prognostic modeling, with histopathology-gene correlations mapped via Spearman analysis. Single-cell transcriptomics (GSE125449) uncovered CXCL1⁺ malignant cell interactions with NCF4⁺ macrophages through the CCL20-CCR6 axis. An elastic net-selected 70-feature gradient boosting machine (GBM) model demonstrated robust prognostic performance (training: 1/3/5-year AUC = 0.834/0.888/0.918; validation: AUC = 0.747/0.814/0.826), with the risk score serving as an independent prognostic factor (HR = 25.402, P < 0.001). TCGA analyses further linked risk scores to altered immune microenvironments, somatic mutations (e.g., TP53), and activated energy/metabolic pathways. This study elucidates OS-driven immunometabolic regulatory mechanisms in HCC and delivers an integrated histology-genomic prognostic model with implications for immunotherapy strategies. Health sciences/Biomarkers Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Health sciences/Oncology Hepatocellular carcinoma (HCC) Oxidative stress Multi-omics integration Prognostic model Tumor microenvironment Pathomics Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Hepatocellular carcinoma (HCC) is one of the leading causes of cancer-related mortality worldwide. Despite advances in diagnosis and treatment in recent years, patient outcomes remain unsatisfactory, and significant challenges persist in clinical decision-making [ 1 , 2 ].Currently, the traditional prognostic evaluation methods based on TNM staging have certain limitations [ 3 ]. Emerging technologies in the fields of genomics and pathology offer potential opportunities for more accurate prognostic prediction [ 4 ]. However, how to effectively integrate multi-omics data and extract clinically meaningful biomarkers remains a critical challenge that needs to be addressed. Emerging evidence suggests that the oxidative stress (OS) pathway may play a crucial role in the initiation and progression of hepatocellular carcinoma (HCC)[ 5 ]. However, its specific regulatory mechanisms have not yet been fully elucidated. Recent studies have found that OS can induce mitochondrial metabolic reprogramming through the NRF2/KEAP1 axis, thereby promoting tumor immune escape by inhibiting ferroptosis [ 6 ]. However, there remains a lack of integrated evidence regarding population-level genetic susceptibility for the aforementioned mechanisms. Meanwhile, although genome-wide association studies (GWAS) have identified multiple genetic loci associated with hepatocellular carcinoma (HCC) incidence [ 7 ], their contribution to prognostic prediction warrants further investigation. Furthermore, the integration of digital pathology and deep learning techniques has provided new insights for tumor heterogeneity analysis; however, current approaches still have room for optimization in terms of feature interpretability and clinical applicability [ 8 , 9 ]. This study aims to explore prognostic prediction methods for hepatocellular carcinoma (HCC) from a multi-omics perspective, with a focus on the key roles of oxidative stress (OS)-related genes in tumor initiation and progression. First, transcriptome-wide association analysis (TWAS) was conducted in combination with a known gene set related to oxidative stress, aiming to systematically identify susceptibility genes significantly associated with HCC prognosis and to preliminarily investigate their potential mechanisms in gene expression regulation and clinical outcomes. Second, an automated image analysis algorithm was employed to select representative regions from histopathological slides and to extract quantitative pathomic features, thereby capturing the heterogeneity of tumor microstructures. Subsequently, machine learning approaches were applied to integrate genetic variants, transcriptomic data, and pathological image features to construct a prognostic model with predictive performance. Furthermore, we analyzed the characteristics of immune cell infiltration in the tumor microenvironment and their correlations with risk scores, shedding light on potential immune regulatory mechanisms mediated by oxidative stress. This study intends to provide novel biomarkers and theoretical foundations for precise prognostic evaluation and personalized treatment strategies in HCC. Method GWAS data The GWAS data for hepatocellular carcinoma were obtained from the dataset by Verma A et al [ 10 ], which includes 2,852 cases of European ancestry and 447,587 control samples. Oxidative Stress Genes Genes associated with oxidative stress were retrieved and downloaded from the GeneCards database ( https://www.genecards.org/ ), using a relevance score greater than 7 as the screening criterion [ 11 ].A total of 1065 oxidative stress-related genes were ultimately identified. Trans-organizational TWAS analysis We employed the UTMOST analysis ( https://github.com/Joker-Jerome/UTMOST?tab=readme-ov-file ) to quantify overall gene–trait associations across tissues. This method enables the identification of a greater number of genes in tissues with substantial trait heritability and improves imputation accuracy[ 12 ]. Subsequently, we applied the generalized Berk-Jones (GBJ) test to integrate gene-trait associations by leveraging the covariance of single-tissue summary statistics [ 13 ]. A significance level of P < 0.05 was considered statistically significant. Histopathological slide data This study obtained hematoxylin and eosin (H&E)-stained whole-slide images (40× magnification, 0.25 µm/pixel resolution) from 379 patients with hepatocellular carcinoma (HCC) through the TCGA database ( https://gdc.cancer.gov/ ). After loading the images using the Slide class, the most representative 512×512 pixel regions (n = 10 per slide) were selected using the ScoreTiler algorithm, ensuring a tissue coverage of no less than 75%. The image preprocessing pipeline included color deconvolution to eliminate staining artifacts, conversion from RGB to grayscale, Otsu’s thresholding, and morphological operations. Regions with areas smaller than 500 pixels² (holes) or 1500 pixels² (tissue fragments) were removed during processing. All subsequent analyses were conducted on image data resampled to a uniform magnification of 20× or 40×. Deep Learning Features This study employs the ResNet-50 architecture for deep feature extraction from pathological images [ 14 ]. All images were resized to 224×224 pixels and preprocessed using ImageNet normalization parameters. The 2048-dimensional features from the output of the fourth residual block were specifically extracted and compressed into compact vectors through global max pooling. This approach effectively retains critical pathological information while significantly reducing computational complexity. Visualization of Prognostic-Related Features This study employs the Grad-CAM algorithm to achieve visual interpretation of the decision-making process of the deep learning model [ 15 ]. This method captures feature maps and gradient information by registering hook functions on the target network layers. The global average of the gradients corresponding to the target class is calculated as the channel weights, which are then used to generate weighted feature maps followed by ReLU activation. Subsequently, the activated feature maps are upsampled to the input image size via bilinear interpolation and normalized. Finally, the normalized maps are superimposed onto the original images with a transparency factor of 0.4 using the Jet colormap to produce heatmaps that visually highlight the critical visual regions relied upon by the model for decision-making [ 16 ]. Pathological Image Features This study conducted digital pathological image analysis using the CellProfiler platform [ 17 , 18 ]. Initially, hematoxylin- and eosin-stained channels were separated through image preprocessing. A multi-scale analytical strategy was employed to extract features from both global and local perspectives: at the global level, features including image quality, staining co-localization, granularity distribution, and multi-scale texture characteristics were extracted; at the local level, a hierarchical object recognition algorithm was applied to quantify staining intensity, morphology, spatial distribution, and local texture features. The mean, median, and standard deviation of each feature were calculated to construct a quantitative pathomic feature set [ 19 ]. Construction of Machine Learning Model A two-stage modeling strategy was employed in this study. First, univariate Cox regression analysis (P < 0.05) was conducted for preliminary feature selection. Subsequently, a predictive model was developed using the GBM (Gradient Boosting Machine) algorithm implemented in the Mime1 package (V 0.12)[ 20 ]. Model performance was evaluated using the C-index, time-dependent ROC curves (at 1-, 3-, and 5-year time points), and risk stratification based on Kaplan-Meier analysis. The GBM algorithm minimizes the loss function through iterative optimization of ensembles of decision trees, combining the feature selection advantages of StepCox [forward] with the nonlinear modeling capabilities of machine learning techniques. Feature importance was extracted from the trained GBM model, and features with non-zero importance scores were selected as the final feature set. Survival Analysis This study conducted survival analysis based on the clinical and risk score data from TCGA-LIHC. Key clinical features (pathological stage, TNM stage, age, and gender) were extracted. Univariate Cox regression analysis was performed to evaluate the impact of clinical factors (age, gender, T stage, N stage, M stage, and pathological stage) as well as the risk score on patient prognosis. Subsequently, multivariate Cox regression analysis was carried out on the variables that showed significance in the univariate analysis (T stage, pathological stage, and risk score). Gene Set Enrichment Analysis (GSEA) A differential expression analysis was performed between the high-risk and low-risk groups based on a significance threshold of FDR < 0.05, and a ranked gene list was generated according to the logFC values. GSEA was subsequently conducted using the clusterProfiler package (v 4.2.2) against the Hallmark gene sets from the MSigDB database (c2.cp.kegg.v2023.1.Hs.entrez.gmt). The distribution of each gene set within the ranked list was assessed by calculating the enrichment score, thereby identifying potential differences in biological pathways between the two groups. Mutation Analysis Somatic mutation data were obtained from the TCGA database and analyzed using the Maftools package (v 2.21.1) [ 21 ]. MAF-formatted mutation annotation files were generated for each group based on the high- and low-risk stratification. Comparative analysis of mutation features between the two groups focused on overall mutation burden, distribution of mutation types, and intergroup differences in frequently mutated genes. Immune-Related Analysis To investigate the role of immune checkpoints in patients stratified into high- and low-risk groups and to analyze their antigen-presenting capacity, we assessed the differential expression of multiple immune checkpoint genes and human leukocyte antigen (HLA) genes. Using the IOBR package (V 0.99.8), immune cell infiltration analysis was performed on gene expression data across risk groups employing the MCPcounter algorithm, allowing evaluation of the abundance of various immune cell types between the two groups[ 22 ]. Additionally, the TIDE algorithm[ 23 ]( http://tide.dfci.harvard.edu/ ) was utilized to predict the response status of samples to immune checkpoint blockade therapy.Differentially expressed genes between the high- and low-risk groups were analyzed in relation to immune response status. Upregulated and downregulated genes were separately subjected to Pathway-Level Metabolism and Proteome Analysis (ProteoMaps)[ 24 ]( https://www.proteomaps.net/ ). The association between risk scores and immunotherapy response was further explored by comparing the similarity of upregulated and downregulated genes between the two groups. Moreover, a random subset of 22 samples from the high- and low-risk groups was selected to generate a matrix for TIP analysis ( http://biocc.hrbmu.edu.cn/TIP/ ), which was used to track and analyze the anti-cancer immune status and the proportion of tumor-infiltrating immune cells [ 25 ]. Single-Cell Analysis This study analyzed single-cell transcriptomic data from GSE125449[ 26 ] using the Seurat package (v 4.3.0). Raw UMI count data were log-normalized (scale factor = 10,000) and the top 3,000 most variable genes were selected for further normalization and scaling. PCA was performed for dimensionality reduction, followed by batch effect correction across samples using the Harmony package (v 0.1.1). Cell clustering was conducted based on the first 10 principal components, and results were visualized using UMAP. Malignant cells were classified into CXCL1 + and CXCL1 − subpopulations according to their gene expression profiles, while tumor-associated macrophages were categorized into NCF4 + and NCF4 − subgroups. Intercellular Communication Analysis We employed the CommPath package (v1.0.0) to analyze the intercellular communication network [ 26 ]. Pathway enrichment analysis was performed based on the KEGG database, and pathway activity was quantified using the GSVA method (minimum gene set size = 10). Significantly enriched pathways with a positive score and p < 0.05 were selected, and the communication strength between cell subpopulations as well as the activity patterns of key signaling pathways were visualized. Transcription Factor Activity Analysis Transcription factor activity analysis was performed using the DoRothEA package (v1.6.0), which assesses transcription factor activity based on known transcription factor–target gene regulatory networks. High-confidence regulatory interactions (confidence levels A, B, and C) were selected from the DoRothEA database. The VIPER algorithm was then applied to calculate transcription factor activity scores from the single-cell RNA sequencing data. Finally, the top 40 transcription factors with the highest variability were selected for further analysis. Statistical Analysis A stratified random sampling strategy (stratified by clinical stage) was employed to divide the total sample into a training cohort (n = 303) and an independent validation cohort (n = 76) at a ratio of 8:2. Spearman’s rank correlation test (two-sided) was used to evaluate the associations between radiomic features and oxidative stress gene scores, as well as between digital pathological features and key gene expression levels. The strength of correlations was quantified using the Spearman correlation coefficient (ρ). For feature selection, elastic net regression (α = 0.5) was applied, with the regularization parameter λ determined via 10-fold cross-validation. This hybrid penalized regression approach combines the advantages of L1 and L2 regularization, allowing for feature sparsity control while mitigating multicollinearity. Survival curves were constructed using the Kaplan-Meier method, and log-rank tests were used to compare survival differences between risk subgroups, with statistical significance defined as P < 0.05. Univariate Cox regression analysis was first performed to screen potential prognostic factors (inclusion criterion: univariate P < 0.1), followed by multivariate analysis to assess their independent prognostic value. Hazard ratios (HRs) are presented with 95% confidence intervals. Model performance was evaluated through time-dependent receiver operating characteristic (ROC) analysis to assess 1-, 3-, and 5-year survival prediction accuracy, along with calculation and comparison of C-indices. The log-rank test was also used to evaluate the model’s ability to stratify patients into distinct risk groups. All computational implementations were conducted using Python 3.8.3 with the PyTorch 1.10 framework for deep learning tasks, and OpenCV 4.5 was used for image processing. Statistical analyses were carried out in R 4.1.3, primarily using the survival package (version 3.3.1) for survival modeling, the limma package (version 3.50) for differential analysis, and the pROC package (version 1.18) for ROC analysis. Multiple hypothesis testing was adjusted using the Benjamini–Hochberg procedure, and a false discovery rate (FDR) < 0.05 was considered statistically significant. All tests were two-sided. Results In the cross-tissue transcriptome-wide association analysis (TWAS), we identified 176 genes (P < 0.05) (Table S1 ). After intersecting with oxidative stress-related genes (Table S2 ), 12 significantly associated genes were ultimately selected. This study performed tissue region extraction and cell density scoring analysis on whole-slide images (WSIs) of hepatocellular carcinoma (HCC) from The Cancer Genome Atlas (TCGA). Samples were divided into 20× (n = 10) and 40× (n = 369) magnification groups. A tile size of 512×512 pixels was used, and a multi-step image processing pipeline was applied, including hematoxylin artifact removal, grayscale conversion, Otsu thresholding, morphological operations (erosion and dilation), and small region filtering. A tissue content threshold of 75% was set, and 10 representative tissue tiles were selected per group. This method successfully achieved standardized processing and analysis across WSIs of different magnifications. A ResNet-50 pretrained model was employed for deep feature extraction from 3,783 pathological image tiles. The layer4 output of ResNet-50 was selected as the feature extraction layer. Each image was preprocessed to a standard size of 224×224 pixels and normalized across RGB channels. A 2048-dimensional feature vector was extracted per image (corresponding to the output channel number of layer4). Maximum spatial pooling was applied across the spatial dimensions to obtain the most representative feature value for each channel, resulting in a 2048-dimensional deep feature representation for each pathological image. Pathological features were extracted using CellProfiler for subsequent analysis. Granularity, intensity, and texture features were extracted, along with mean, median, and standard deviation features for each object. All pathological features were grouped by sample and averaged. Similarly, deep features were also averaged per sample. Finally, pathological and deep features were integrated into a single feature matrix for further analysis (Table S3 ). Single-sample gene set enrichment analysis (SSGSEA) was used to compute enrichment scores for the 12 associated genes. Spearman correlation analysis revealed that 378 features were significantly correlated with the gene set scores in the training cohort (n = 287) (P < 0.05) (Table S4 ). Subsequently, elastic net regression (α = 0.2) was applied to select 70 most predictive features, including 63 ResNet-derived features and 7 CellProfiler-derived features. Network analysis demonstrated significant associations between these features and gene set scores (correlation coefficient range: -0.229 to 0.274) (Supplementary Fig. 1, Table S5 ). Univariate Cox regression analysis was performed on the 70 selected features in the training cohort to identify survival-associated features (p < 0.05) (Table S6 ). Grad-CAM was applied to visualize the deep learning features on pathological images. The results indicated that the model primarily focused on specific regions within the images (Fig. 1 ). Heatmap visualization illustrated the spatial distribution of key features that contributed to model decision-making, which may correspond to diagnostically relevant pathological features. A prognostic signature constructed using the "StepCox [forward] + GBM" model demonstrated good predictive performance in training and validation cohorts (Fig. 2 A). Survival analysis showed that the model effectively stratified patients into high-risk and low-risk groups (log-rank p < 0.05) (Fig. 3 A). Time-dependent ROC curve analysis revealed AUC values of 0.834, 0.888, and 0.918 at 1, 3, and 5 years in the training cohort, and 0.747, 0.814, and 0.826 in the validation cohort, respectively, indicating stable predictive capability (Fig. 3 B). The C-index reached 0.75 in the validation cohort, suggesting strong discriminative performance (Fig. 2 B). Univariate and multivariate Cox regression analyses using TCGA-LIHC data evaluated the prognostic impact of RiskScore and clinical parameters in HCC patients (Figs. 3 C, D; Tables S7, S8). Univariate analysis identified significant associations for T stage (HR = 1.802, 95%CI = 1.399–2.322, p < 0.001), tumor stage (HR = 1.855, 95%CI = 1.41–2.441, p < 0.001), and RiskScore (HR = 26.801, 95%CI = 11.859–60.571, p < 0.001). Multivariate analysis confirmed RiskScore (HR = 25.402, 95%CI = 10.687–60.381, p < 0.001) as an independent prognostic factor. HCC patients were stratified into high- and low-risk groups by median RiskScore. KEGG pathway GSEA revealed significantly enriched pathways (p.adjust < 0.05) with |NES|≥2, including fatty acid metabolism (NES=-2.83), PPAR signaling (NES=-2.48), and steroid hormone biosynthesis (NES=-2.50), associated with HCC prognosis (Fig. 4 A, Table S9 ). Somatic mutation profiles were compared between risk groups, revealing significant differences in the top 20 mutated genes. TP53 (31%) and TNN (27%) exhibited higher mutation frequencies in high-risk patients, with TP53 mutations differing between groups. Mutation waterfall plots showed distinct signatures, indicating association between RiskScore and molecular subtypes. Clinical annotation confirmed mutation distribution consistency with risk stratification, validating reliability (Supplementary Figs. 2, 3). Tumor immune phenotypes (TIP) were analyzed in 22 samples per risk group. Significant differences in immune cell infiltration were observed between groups. A heatmap revealed distinct immune cell composition profiles, associating RiskScore with tumor immune status (Supplementary Fig. 4). Differential expression analysis of immune checkpoint and HLA genes in TCGA-LIHC high- vs. low-risk groups revealed significant differences (P < 0.05) for several immune checkpoints and HLA-B/C (Figs. 4 B, C), suggesting prognostic links. MCPcounter assessed immune infiltration, showing higher T cell and monocyte levels in high-risk patients (Fig. 4 D). Proteomaps visualization indicated significant enrichment of lipid/steroid and amino acid metabolism pathways in both high-risk and immunotherapy response groups, demonstrating similarity. These findings provide insights into tumor risk stratification and immunotherapy response mechanisms (Figs. 5 E, F). Integration of pathological image features and gene expression data revealed significant Spearman correlations (p < 0.05) between histopathological features and multiple genes (Fig. 5 A). Network analysis illustrated interactions between pathological features (ResNet, CellProfiler) and genes, identifying NCF4 and CXCL1 with prognostic relevance (Fig. 5 B). Survival analysis showed better overall survival in low-expression groups of NCF4 and CXCL1 versus high-expression groups (p < 0.05), suggesting prognostic value (Figs. 5 C, D). Analysis of GSE125449 single-cell RNA sequencing data characterized HCC at the cellular level. Clustering identified major cell types like malignant cells and TAMs. Differential expression of NCF4 and CXCL1 across subpopulations and samples revealed key gene expression features at single-cell resolution (Figs. 6 A, Supplementary Fig. 5). Transcription factor activity analysis in NCF4/CXCL1 subpopulations identified the top 40 variable TFs. SPI1 and AR showed differential activity in NCF4 + TAMs, while NFKB1 and RELA did so in CXCL1 + tumor cells, indicating cell-type-specific regulation(Figs. 6 B, Supplementary Fig. 6). Cell-cell communication analysis revealed pathway activation with distinct specificity in tumor subpopulations: Leukocyte Transendothelial Migration was enriched in NCF4 + TAMs, while the Rap1 Signaling Pathway was enriched in CXCL1 + tumor cells (Figs. 6 C, Supplementary Fig. 7). Network analysis showed intercellular communication strength, suggesting functional synergy among specific subpopulations (Figs. 6 D, Supplementary Fig. 8). Discussion This study constructed an interpretable HCC prognostic model centered on oxidative stress by integrating genomic, transcriptomic, pathomic, and single-cell sequencing data. We elucidated the biological basis of risk stratification across molecular mechanisms, microenvironment heterogeneity, and immune regulation. Cross-tissue TWAS identified 12 oxidative stress pathway genes significantly associated with HCC, potentially playing critical roles in its progression. NCF4, an essential NADPH oxidase subunit, may cause dysregulated ROS levels, promoting DNA damage and tumor evolution [ 27 ].CXCL1 recruits MDSCs, inducing immunosuppression and facilitating immune evasion [ 28 ].Notably, at single-cell resolution, we observed for the first time the cell type-specific expression patterns of these two genes: CXCL1 is predominantly enriched in malignant epithelial cells, whereas NCF4 is highly expressed in tumor-associated macrophages. This finding not only reveals the potential for intercellular functional coordination within the hepatocellular carcinoma microenvironment, but also provides a foundation for future investigations into targeted interventions directed at specific cell types. Moreover, it underscores the significant role of oxidative stress in modulating the tumor immune microenvironment. Our digital pathology analysis developed a multimodal feature extraction pipeline integrating deep learning with traditional histopathological features, identifying multiple quantitative prognostic biomarkers. Over 90% of the 70 predictive features derived from deep ResNet-50 layer 4, indicating convolutional neural networks capture subtle microstructural information challenging to quantify manually [ 29 , 30 ].Using Grad-CAM, we revealed model attention aligned with pathological features like tumor heterogeneity and mitotic figures, aiding model interpretability [ 31 ].This aligns with AI-assisted pathology trends, highlighting deep learning's potential in digital pathology [ 32 ]. We developed a clinically applicable prognostic model based on oxidative stress features using a "StepCox [forward] + GBM" approach. It demonstrated strong discriminative performance in both training (n = 287) and validation (n = 96) cohorts, with time-dependent AUCs of 0.834/0.888/0.918 (1/3/5-year, training) and 0.747/0.814/0.826 (validation). The risk score was an independent prognostic factor regardless of clinical stage (HR = 25.402, p < 0.001) and maintained stable long-term performance (< 5% ROC-AUC decrease over time). This predictive stability holds significant clinical relevance given HCC's prolonged course [ 33 ]. Pathway analysis showed distinct metabolic alterations between high- and low-risk groups. GSEA revealed suppression of fatty acid metabolism, PPAR signaling, and steroid hormone biosynthesis in the high-risk group (NES < − 2, FDR < 0.05), potentially promoting tumor proliferation by disrupting energy supply and membrane synthesis [ 34 – 36 ]. ProteoMaps indicated similar lipid/amino acid metabolism patterns in the high-risk and immunotherapy-responsive groups, suggesting these pathways could predict immune checkpoint blockade efficacy and support integrating metabolic modulation into immunotherapy strategies [ 37 , 38 ]. These results further validate our proposed "oxidative stress–metabolism–immune dysregulation" hepatocarcinogenesis triad. In terms of the tumor immune microenvironment, TIP analysis revealed significant differences in the infiltration profiles of immune cells between the high- and low-risk groups. Integrating single-cell RNA sequencing data, we found that NCF4 and CXCL1 exhibited distinct expression patterns in tumor-associated macrophages and malignant cells, respectively. These findings not only provide cellular-level evidence supporting the biological plausibility of the risk stratification, but also suggest that specific intercellular interactions may play a pivotal role in driving disease progression[ 39 ]. Cell communication analysis substantiated this, identifying multiple ligand–receptor pairs mediating high-intensity interactions[ 40 ], collectively highlighting the multidimensional role of oxidative stress-related genes in shaping the tumor microenvironment. Mutation signature analysis revealed an association between the risk score and HCC molecular subtypes. The high-risk group showed increased TP53 mutation frequency and more complex mutation patterns, suggesting greater genomic instability. As a classical tumor suppressor central to DNA repair and cell cycle regulation [ 41 ], TP53 mutations are closely linked to hepatocarcinogenesis and are an established biomarker for poor prognosis [ 42 ]. The high-risk group also had a significantly higher TNN (Titin) mutation rate. Though encoding a large skeletal muscle protein with incompletely understood roles in liver cancer[ 43 ], evidence suggests its mutations may reflect tumor mutational burden and associate with worse outcomes. In the high-risk HCC subgroup, we observed significant upregulation of key immune checkpoint molecules PDCD1 (PD-1), CTLA4, and LAG3, suggesting severe T-cell exhaustion that may impair anti-tumor immunity[ 44 ]. The PD-1/PD-L1 pathway is an established mechanism of immune evasion in HCC and a major target for immunotherapy[ 45 ]. However, concurrent elevation of CTLA-4 and LAG3 suggests activation of multiple immunosuppressive pathways, which could explain immunotherapy resistance and variable treatment responses. We also observed aberrant upregulation of HLA class I molecules (HLA-B, HLA-C) in the high-risk group. While HLA class I downregulation is a recognized immune evasion mechanism in immunogenic tumors[ 46 ], studies in HCC suggest abnormal upregulation may represent a compensatory immune activation or adaptive response to chronic inflammation [ 47 ]. This "elevated yet dysfunctional" state warrants further functional validation. Immune infiltration analysis revealed higher levels of T cells and monocytes in the high-risk group, suggesting a more active immune response[ 48 , 49 ]. More intriguingly, comparing protein functional networks revealed high concordance between the high-risk group and immunotherapy-responsive group in lipid metabolism. This suggests lipid metabolism may critically link disease progression and immunotherapy response in the HCC microenvironment [ 50 ]. Lipid metabolism pathways are frequently dysregulated in HCC, promoting tumor proliferation, invasion, and immune evasion [ 51 , 52 ]. Crucially, lipid metabolism modulates T cell function: FAO enhances memory T cells, while cholesterol accumulation can cause T cell exhaustion [ 53 ].This may underlie the immune features observed in our high-risk patients. Although these patients have poorer outcomes, their preserved lipid metabolism suggests retained potential for antitumor immune responses under specific conditions. Thus, even conventionally high-risk HCC patients may benefit from checkpoint inhibitors, especially combined with lipid metabolism-targeting therapies[ 54 ]. Transcription factor activity analysis in NCF4/CXCL1-associated subpopulations revealed distinct regulatory patterns within the TME. SPI1 and AR activity significantly differed in TAMs with differential NCF4 expression, suggesting involvement in regulating polarization or functional states [ 55 ].In contrast, tumor cells with differential CXCL1 expression showed significant NFKB1/RELA activity differences, indicating NF-κB pathway activation [ 56 ].Given NF-κB's role in inflammation, survival, and plasticity, this suggests CXCL1-high tumor cells exhibit enhanced inflammatory responsiveness and may actively recruit/remodel immune cells via chemokine secretion. These findings support a potential positive feedback loop between oxidative stress and NF-κB signaling in HCC progression. Intercellular communication analysis further supports this transcriptional regulation. Pathways for Leukocyte Transendothelial Migration and Rap1 Signaling were significantly enriched in NCF4 + TAMs and CXCL1 + tumor cells [ 57 ]. These pathways mediate cell adhesion and migration—key processes in immune infiltration and metastasis—indicating distinct cell populations drive coordinated yet functionally divergent programs. Differential oxidative stress-related gene expression likely drives this functional heterogeneity. This study's innovation includes: developing the first oxidative stress-centered multimodal prognostic framework, characterizing single-cell oxidative stress-related gene expression profiles, and elucidating an oxidative stress-mediated metabolic-immune-genomic network. Despite limitations in external validation and functional experiments, it provides a foundation for oxidative stress-targeted precision therapy in hepatocellular carcinoma. This study has limitations. Model validation was primarily limited to the TCGA-LIHC cohort; external validation in independent cohorts is lacking and requires future prospective studies. Furthermore, while mechanistic hypotheses were inferred from multi-omics analysis, direct functional evidence is absent, necessitating validation in vitro and in vivo. Conclusion In summary, this study developed an oxidative stress-centered prognostic model for hepatocellular carcinoma through multi-omics analysis. The model demonstrates high discriminative ability and favorable generalization. Underlying biological mechanisms – spanning molecular, metabolic, immune, and cellular interactions – were systematically elucidated. Notably, the identified oxidative stress-related biomarkers linked to immune escape and metabolic regulation provide foundations and targets for novel diagnostics and therapeutics. Declarations Disclosure The authors report no conflicts of interest in this work. Acknowledgements Not applicable. Funding None. Ethics approval and consent to participate Not applicable. Author contributions statement Lunwei Yang (conceptualization, methodology, formal analysis, investigation, writing—original draft), Yuanliang Li (conceptualization, methodology, formal analysis, writing—original draft), Xiaoying Zhong (writing—review and editing, supervision, project administration). All authors read and approved the final manuscript. All authors have read and approved the final manuscript. Data availability statement The GWAS data for hepatocellular carcinoma were obtained from the dataset by Verma A et al.The TCGA dataset is publicly available at the TCGA portal (https://portal.gdc.cancer.gov). Patient consent for publication Not applicable. References Hwang, S. Y. et al. Hepatocellular carcinoma: updates on epidemiology, surveillance, diagnosis and treatment. Clin. Mol. Hepatol. 31 (Suppl), S228–S54 (2025). Rumgay, H. et al. Global burden of primary liver cancer in 2020 and predictions to 2040. J. Hepatol. 77 (6), 1598–1606 (2022). Trevisani, F. et al. Merits and boundaries of the BCLC staging and treatment algorithm: Learning from the past to improve the future with a novel proposal. J. Hepatol. 80 (4), 661–669 (2024). Chaudhary, K., Poirion, O. B., Lu, L. & Garmire, L. X. Deep Learning-Based Multi-Omics Integration Robustly Predicts Survival in Liver Cancer. Clin. Cancer Res. 24 (6), 1248–1259 (2018). Huang, Z. et al. Oxidative Stress Promotes Liver Cancer Metastasis via RNF25-Mediated E-Cadherin Protein Degradation. Adv. Sci. (Weinh) . 11 (13), e2306929 (2024). Sun, X. et al. Activation of the p62-Keap1-NRF2 pathway protects against ferroptosis in hepatocellular carcinoma cells. Hepatology 63 (1), 173–184 (2016). Hassan, M. M. et al. Genome-wide association study identifies high-impact susceptibility loci for HCC in North America. Hepatology 80 (1), 87–101 (2024). Nakatsuka, T. et al. Deep learning and digital pathology powers prediction of HCC development in steatotic liver disease. Hepatology 81 (3), 976–989 (2025). Nam, D., Chapiro, J., Paradis, V., Seraphin, T. P. & Kather, J. N. Artificial intelligence in liver diseases: Improving diagnostics, prognostics and response prediction. JHEP Rep. 4 (4), 100443 (2022). Verma, A. et al. Diversity and scale: Genetic architecture of 2068 traits in the VA Million Veteran Program. Science 385 (6706), eadj1182 (2024). Xu, M. et al. Identification and validation of immune and oxidative stress-related diagnostic markers for diabetic nephropathy by WGCNA and machine learning. Front. Immunol. 14 , 1084531 (2023). Hu, Y. et al. A statistical framework for cross-tissue transcriptome-wide association analysis. Nat. Genet. 51 (3), 568–576 (2019). Sun, R., Hui, S., Bader, G. D., Lin, X. & Kraft, P. Powerful gene set analysis in GWAS with the Generalized Berk-Jones statistic. PLoS Genet. 15 (3), e1007530 (2019). Marme, F. et al. Deep learning to predict breast cancer sentinel lymph node status on INSEMA histological images. Eur. J. Cancer . 195 , 113390 (2023). Selvaraju, R. R., Das, M. C. A., Vedantam, R., Parikh, D. & Batra, D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. IEEE International Conference on Computer Vision (ICCV). 2017;Venice:618 – 26. (2017). Ramprasaath, R., Selvaraju, A. D., Vedantam, R., Cogswell, M. & Parikh, D. Dhruv Batra. Grad-CAM: why did you say that? arXiv. ;arXiv:1611.07450 (2016). Carpenter, A. E. et al. CellProfiler: image analysis software for identifying and quantifying cell phenotypes. Genome Biol. 7 (10), R100 (2006). Jones, T. R. et al. CellProfiler Analyst: data exploration and analysis software for complex image-based screens. BMC Bioinform. 9 , 482 (2008). Jones, T. R. et al. Scoring diverse cellular morphologies in image-based screens with iterative feedback and machine learning. Proc. Natl. Acad. Sci. U S A . 106 (6), 1826–1831 (2009). Liu, H. et al. Mime: A flexible machine-learning framework to construct and visualize models for clinical characteristics prediction and feature selection. Comput. Struct. Biotechnol. J. 23 , 2798–2810 (2024). Mayakonda, A., Lin, D. C., Assenov, Y., Plass, C. & Koeffler, H. P. Maftools: efficient and comprehensive analysis of somatic variants in cancer. Genome Res. 28 (11), 1747–1756 (2018). Zeng, D. et al. IOBR: Multi-Omics Immuno-Oncology Biological Research to Decode Tumor Microenvironment and Signatures. Front. Immunol. 12 , 687975 (2021). Fu, J. et al. Large-scale public data reuse to model immunotherapy response and resistance. Genome Med. 12 (1), 21 (2020). Liebermeister, W. et al. Visual account of protein investment in cellular functions. Proc. Natl. Acad. Sci. U S A . 111 (23), 8488–8493 (2014). Xu, L. et al. TIP: A Web Server for Resolving Tumor Immunophenotype Profiling. Cancer Res. 78 (23), 6575–6580 (2018). Ma, L. et al. Tumor Cell Biodiversity Drives Microenvironmental Reprogramming in Liver Cancer. Cancer Cell. 36 (4), 418–430 (2019). e6. Li, L. et al. NCF4 attenuates colorectal cancer progression by modulating inflammasome activation and immune surveillance. Nat. Commun. 15 (1), 5170 (2024). Korbecki, J. et al. The Effect of Hypoxia on the Expression of CXC Chemokines and CXC Chemokine Receptors-A Review of Literature. Int. J. Mol. Sci. ; 22 (2). (2021). Wang, X. et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature 634 (8035), 970–978 (2024). Acs, B., Rantalainen, M. & Hartman, J. Artificial intelligence as the next step towards precision pathology. J. Intern. Med. 288 (1), 62–81 (2020). Sauter, D. et al. Validating Automatic Concept-Based Explanations for AI-Based Digital Histopathology. Sens. (Basel) ; 22 (14). (2022). Baxi, V., Edwards, R., Montalto, M. & Saha, S. Digital pathology and artificial intelligence in translational medicine and clinical practice. Mod. Pathol. 35 (1), 23–32 (2022). Bialecki, E. S. & Di Bisceglie, A. M. Clinical presentation and natural course of hepatocellular carcinoma. Eur. J. Gastroenterol. Hepatol. 17 (5), 485–489 (2005). Qin, L. et al. ELOVL1 is upregulated and promotes tumor growth in hepatocellular carcinoma through regulating PI3K-AKT-mTOR signaling. Heliyon 10 (15), e34961 (2024). Zhao, X. et al. NR4A3 inhibits the tumor progression of hepatocellular carcinoma by inducing cell cycle G0/G1 phase arrest and upregulation of CDKN2AIP expression. Int. J. Biol. Sci. 20 (15), 5850–5867 (2024). Wei, Q. et al. Estrogen suppresses hepatocellular carcinoma cells through ERbeta-mediated upregulation of the NLRP3 inflammasome. Lab. Invest. 95 (7), 804–816 (2015). Xia, L. et al. The cancer metabolic reprogramming and immune response. Mol. Cancer . 20 (1), 28 (2021). Zheng, J. et al. Hepatocellular carcinoma: signaling pathways and therapeutic advances. Signal. Transduct. Target. Ther. 10 (1), 35 (2025). Mao, X. et al. Crosstalk between cancer-associated fibroblasts and immune cells in the tumor microenvironment: new findings and future perspectives. Mol. Cancer . 20 (1), 131 (2021). Liu, Y. et al. Identification of a tumour immune barrier in the HCC microenvironment that determines the efficacy of immunotherapy. J. Hepatol. 78 (4), 770–782 (2023). Voskarides, K. & Giannopoulou, N. The Role of TP53 in Adaptation and Evolution. Cells ; 12 (3). (2023). Ortiz-Cuaran, S. et al. Association between TP53 R249S mutation and polymorphisms in TP53 intron 1 in hepatocellular carcinoma. Genes Chromosomes Cancer . 52 (10), 912–919 (2013). Han, X., Chen, J., Wang, J., Xu, J. & Liu, Y. TTN mutations predict a poor prognosis in patients with thyroid cancer. Biosci. Rep. 42 , 7 (2022). Meng, Y. et al. Immunosuppressive CD10(+)ALPL(+) neutrophils promote resistance to anti-PD-1 therapy in HCC by mediating irreversible exhaustion of T cells. J. Hepatol. 79 (6), 1435–1449 (2023). Zhou, B. et al. Hepatoma cell-intrinsic TLR9 activation induces immune escape through PD-L1 upregulation in hepatocellular carcinoma. Theranostics 10 (14), 6530–6543 (2020). Bauer, J. et al. The oncogenic fusion protein DNAJB1-PRKACA can be specifically targeted by peptide-based immunotherapy in fibrolamellar hepatocellular carcinoma. Nat. Commun. 13 (1), 6401 (2022). Stolarz, W. et al. [The HLA-I and HLA-II expression evaluation in liver biopsy specimens of patients with chronic viral hepatitis]. Przegl Epidemiol. 53 (3–4), 385–393 (1999). Tacke, F. Targeting hepatic macrophages to treat liver diseases. J. Hepatol. 66 (6), 1300–1312 (2017). Nicole, L. et al. Necroptosis-driving genes RIPK1, RIPK3 and MLKL-p are associated with intratumoral CD3(+) and CD8(+) T cell density and predict prognosis in hepatocellular carcinoma. J. Immunother Cancer ; 10 (3). (2022). Bian, X. et al. Lipid metabolism and cancer. J. Exp. Med. ; 218 (1). (2021). Su, F. & Koeberle, A. Regulation and targeting of SREBP-1 in hepatocellular carcinoma. Cancer Metastasis Rev. 43 (2), 673–708 (2024). Liu, Z. et al. Deficiency of SIAH1 promotes the formation of filopodia by increasing the accumulation of FASN in liver cancer. Cell. Death Dis. 15 (7), 537 (2024). Loschinski, R. et al. IL-21 modulates memory and exhaustion phenotype of T-cells in a fatty acid oxidation-dependent manner. Oncotarget 9 (17), 13125–13138 (2018). Ramapriyan, R. et al. Altered cancer metabolism in mechanisms of immunotherapy resistance. Pharmacol. Ther. 195 , 162–171 (2019). Qi, S. et al. SPI1-mediated macrophage polarization aggravates age-related macular degeneration. Front. Immunol. 15 , 1421012 (2024). Taniguchi, K. & Karin, M. NF-kappaB, inflammation, immunity and cancer: coming of age. Nat. Rev. Immunol. 18 (5), 309–324 (2018). Lu, Z. H. et al. Early administration of Wumei Wan inhibit myeloid-derived suppressor cells via PI3K/Akt pathway and amino acids metabolism to prevent colitis-associated colorectal cancer. J. Ethnopharmacol. 333 , 118260 (2024). Additional Declarations No competing interests reported. Supplementary Files SupplementaryFigure1.pdf Supplementary Figure 1. Elastic Net Analysis of Pathological Features and Genomic Scores. SupplementaryFigure2.pdf Supplementary Figure 2. Mutation analysis of low risk group. SupplementaryFigure3.pdf Supplementary Figure 3. Mutation analysis of the high risk group. SupplementaryFigure4.pdf Supplementary Figure 4. Analysis of Tumor Immune Phenotype (TIP) between high risk group and low risk group. SupplementaryFigure5.pdf Supplementary Figure 5. Differential expression of NCF4 and CXCL1 genes among subpopulations in single-cell analysis. SupplementaryFigure6.pdf Supplementary Figure 6. Analysis of transcription factor activity in the NCF4/CXCL1 subpopulation. SupplementaryFigure7.pdf Supplementary Figure 7. Intercellular communication analysis of the NCF4/CXCL1 subpopulation. SupplementaryFigure8.pdf Supplementary Figure 8. Communication network analysis of the NCF4/CXCL1 subpopulation. SupplementaryTable.xlsx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7489138","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":536351413,"identity":"ac051ae7-95a4-4f50-8962-463fe5d65991","order_by":0,"name":"Lunwei Yang","email":"","orcid":"","institution":"Guangzhou University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Lunwei","middleName":"","lastName":"Yang","suffix":""},{"id":536351414,"identity":"f61a8d61-2078-4bcd-8451-9f16cc91f492","order_by":1,"name":"Yuanliang Li","email":"","orcid":"","institution":"The First Affiliated Hospital of Guangzhou University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yuanliang","middleName":"","lastName":"Li","suffix":""},{"id":536351415,"identity":"e0014eb9-abf3-493f-80d6-554c844d180a","order_by":2,"name":"Xiaoying Zhong","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1UlEQVRIiWNgGAWjYDACCQjFw8befODAhx8kaJHh4zmWeHBmDwlabOQkcowPc7ARoUN+dvOxh1/bDvOwSeR8OMzAwyDPL3YAvxbGOcfSjWXOALXwvN1wuMCCwXDm7AT8WpglcsykJSqAWthzNxyewcOQYHCbgBY2sBYDoBaGnAcgkrAWHqAWyQ8gWzhyGIjTIiGRlibNcCYd6JdjBsBAliDsF/kZycckf7ZZ28u3Nz/+8OGHjTy/NAEtIMDMw9AMt5WwchBg/MFQR5zKUTAKRsEoGJkAANe/QCRufNf9AAAAAElFTkSuQmCC","orcid":"","institution":"Chongqing Nanan District Traditional Chinese and Western Medicine Hospital","correspondingAuthor":true,"prefix":"","firstName":"Xiaoying","middleName":"","lastName":"Zhong","suffix":""}],"badges":[],"createdAt":"2025-08-29 13:38:37","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7489138/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7489138/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":94874484,"identity":"347e3bb2-a055-4cc7-902f-4e92f597d1a3","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":154503,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/8090e7a0eec950324b6c190b.docx"},{"id":94987030,"identity":"aa175a72-6acf-4c47-b039-f242dcc29541","added_by":"auto","created_at":"2025-11-03 07:01:07","extension":"json","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5061,"visible":true,"origin":"","legend":"","description":"","filename":"5e87e849c9e34581bf0ff8957a28546c.json","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/8e29d110795c8c7ed6d4fa7c.json"},{"id":94874495,"identity":"8018a870-ddae-4f6f-8f13-4a0f54b98fc6","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":68284,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/8563b2398a17b88709a8671c.pdf"},{"id":94986462,"identity":"37cbb3e0-e9e3-4bde-b2bb-9f616e7ebcbd","added_by":"auto","created_at":"2025-11-03 07:00:19","extension":"pdf","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":69929,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/752b7f82f04262d0bd9f9a87.pdf"},{"id":94985834,"identity":"a1a5ea52-d9bd-4f7e-ab58-f9ad592cb394","added_by":"auto","created_at":"2025-11-03 06:59:04","extension":"pdf","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":69085,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/90b472e9343222c31ca3aefd.pdf"},{"id":94874500,"identity":"b0683f17-7243-4aae-ba57-5d150f3b0072","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":44260,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/a32cd260ce3f920bb2a4b198.pdf"},{"id":94986497,"identity":"a6756d12-1142-4744-9c65-9228357769d5","added_by":"auto","created_at":"2025-11-03 07:00:22","extension":"pdf","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1833246,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure5.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/b02e4e07917d26b551e7ac56.pdf"},{"id":94874493,"identity":"07573e0a-6de9-4886-9358-6d6cae063b50","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20335,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure6.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/7cf40ecdfee8122d894ab381.pdf"},{"id":94986281,"identity":"38b798c8-82d4-4198-918e-6fc37dc13a64","added_by":"auto","created_at":"2025-11-03 07:00:09","extension":"pdf","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":32405,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure7.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/6c8e132d4bcb5e120fcc127a.pdf"},{"id":94874501,"identity":"d0ee45a4-b10c-4703-88a4-bc3143436389","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":150625,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure8.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/8a18d416729bf7c6cceb0870.pdf"},{"id":94874512,"identity":"7e32f63c-409c-4ea6-9731-d1a566104715","added_by":"auto","created_at":"2025-10-31 15:31:49","extension":"xlsx","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11501798,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTable.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/f81bc3599f69a606e3ef253b.xlsx"},{"id":94986555,"identity":"42910a04-fdf9-41db-bd62-784fdc83db59","added_by":"auto","created_at":"2025-11-03 07:00:26","extension":"xml","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":117169,"visible":true,"origin":"","legend":"","description":"","filename":"5e87e849c9e34581bf0ff8957a28546c1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/64e6d313b4401ec76af34a75.xml"},{"id":94874515,"identity":"190f4ba7-844f-4e6d-bb71-d84413e2fe91","added_by":"auto","created_at":"2025-10-31 15:31:49","extension":"pdf","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26642867,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/e2ae434e9c6ce44a297c0dbe.pdf"},{"id":94874507,"identity":"bd0143dd-0358-4559-a4a9-6c39fc497375","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":721228,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/1fb7bf8f2c742ca2737a70f3.pdf"},{"id":94986219,"identity":"c663347e-d752-44cb-8a39-47d34e4ffb34","added_by":"auto","created_at":"2025-11-03 07:00:04","extension":"pdf","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":581711,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/9d6c6724bd60b177aee5db6c.pdf"},{"id":94986827,"identity":"6711855b-21bd-478b-a8ef-46653dc9b440","added_by":"auto","created_at":"2025-11-03 07:00:50","extension":"pdf","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2328355,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/b4b9a93d37bee7cad4798798.pdf"},{"id":94874506,"identity":"6ca557b1-05c3-478e-96ae-69ea56b140d9","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1030843,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/f5596fb22fadba4c51f74e68.pdf"},{"id":94874513,"identity":"28ebe9e1-cba6-4c33-9d07-b68c332eda17","added_by":"auto","created_at":"2025-10-31 15:31:49","extension":"pdf","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15639939,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/cd44209a8133076991a5352b.pdf"},{"id":94874503,"identity":"3d42cec1-0e56-4e70-8dce-3d7a366ca692","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"xml","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":116621,"visible":true,"origin":"","legend":"","description":"","filename":"5e87e849c9e34581bf0ff8957a28546c1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/6d7357edba98a7148d886470.xml"},{"id":94986663,"identity":"18325b77-e3f3-4f46-92b3-8d5ef4a1bae4","added_by":"auto","created_at":"2025-11-03 07:00:32","extension":"html","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":131113,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/ec829b19a13d0481ce860cfb.html"},{"id":94986685,"identity":"aaae73a4-95b7-4a31-99bd-4b990445ee39","added_by":"auto","created_at":"2025-11-03 07:00:37","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":4989948,"visible":true,"origin":"","legend":"\u003cp\u003eGrad-CAM heatmap visualizes the regions of interest in the pathological image as interpreted by the model. Redder areas indicate higher attention weights from the model, while bluer areas indicate lower attention.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/f2851895b3d9f8ab1eb0bbe5.png"},{"id":94874482,"identity":"a7c2bcdb-942c-4e8e-9691-0a77ca25542a","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":166197,"visible":true,"origin":"","legend":"\u003cp\u003e(A) The C-index distribution plot illustrates the predictive performance of the prognostic model in both the training and validation cohorts. The values depicted in the plot reflect the model's discriminatory capacity across different datasets, with a higher C-index indicating greater predictive accuracy.(B) Distribution of the C-index based on the model combination \"StepCox [forward] + GBM\".\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/733623e3f09e4d5f2fe0c4b1.png"},{"id":94874479,"identity":"99a602bc-6cfc-4afb-9b28-11faca830a28","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":80247,"visible":true,"origin":"","legend":"\u003cp\u003e(A) The survival curves (Kaplan–Meier curves) for high-risk and low-risk groups based on the \"StepCox [forward] + GBM\" model in both the training and validation cohorts, with a cutoff value of 0.5;(B) Receiver operating characteristic (ROC) curves for the \"StepCox[forward] + GBM\" model in the training and validation cohorts at 1, 3, and 5 years. The area under the curve (AUC) at each time point reflects the model's predictive performance for the probability of event occurrence at corresponding follow-up time points;(C) Univariate Cox regression analysis of variables including age, sex, T stage, N stage, M stage, clinical stage, and risk score in survival analysis;(D) Multivariate analysis based on the Cox proportional hazards model included the following variables: T stage, clinical stage, and risk score.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/551e6000aecf806a6514231e.png"},{"id":94874492,"identity":"56d1b7eb-250a-4353-bcef-477ce429996e","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":171083,"visible":true,"origin":"","legend":"\u003cp\u003e(A) The GSEA enrichment analysis results of differentially expressed genes between the high- and low-risk groups were based on the KEGG pathway database. The enrichment score (NES) reflects the relative activation level of each pathway in either the high-risk or low-risk group, with |NES| ≥ 2 considered to indicate statistically significant enrichment;(B-D) Expression differences in immune checkpoint genes, HLA family genes , and immune cell infiltration scores between the high-risk (High) and low-risk (Low) groups; (E-F) Functional landscape of upregulated differentially expressed genes in the high-risk (High) and low-risk (Low) groups, as well as in the immune therapy responder and non-responder groups. (*p\u0026lt;0.05, **p\u0026lt;0.01, ***p\u0026lt;0.001, ****p\u0026lt;0.0001).\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/4537a26dec02400963f80a42.png"},{"id":94986207,"identity":"c97ebc9f-e7a5-4a9d-9033-6916bd49e6ba","added_by":"auto","created_at":"2025-11-03 07:00:03","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":163126,"visible":true,"origin":"","legend":"\u003cp\u003e(A) Spearman correlation heatmap between clinical features and gene expression;(B) Network of significant Spearman correlations between clinical features (including resnet and cellprofiler features) and genes;(C-D) Patients were divided into low-expression and high-expression groups based on NCF4 and CXCL1 expression levels, and Kaplan-Meier curves for overall survival (OS) were plotted.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/f8cb594298d21ac9ce08b761.png"},{"id":94874491,"identity":"a232381f-ced5-4350-b1e3-45b1de680153","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":186922,"visible":true,"origin":"","legend":"\u003cp\u003e(A) A circlize plot visualizing the spatial distribution of different cell clusters from the single-cell transcriptomic data; (B) Heatmap of highly variable transcription factor (TF) activities across different cell types;(C) The Circos plot illustrates the strength of ligand-receptor interactions among different cell types;(D) Heatmap of pathway activity differences based on GSVA scores. The top 5 significantly differentially activated signaling pathways (KEGG) across different cell types are shown.\u003c/p\u003e","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/e450c44170d460193beeec47.png"},{"id":95655729,"identity":"e074ee22-7731-4a44-a2de-d77ece3b6dcd","added_by":"auto","created_at":"2025-11-11 16:16:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7498422,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/1f76b88a-c16d-47e8-a073-06a545e5b61c.pdf"},{"id":94874480,"identity":"9a7a5503-312d-4f53-9d05-cd9f031bc8fb","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":68284,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 1. Elastic Net Analysis of Pathological Features and Genomic Scores.\u003c/p\u003e","description":"","filename":"SupplementaryFigure1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/ffd359842cdd54b6bd684ebd.pdf"},{"id":94874481,"identity":"2a2d1d39-5589-47df-90f1-1a0938fe20a2","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":69929,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 2. Mutation analysis of low risk group.\u003c/p\u003e","description":"","filename":"SupplementaryFigure2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/aeef6bda506d37b25612df8b.pdf"},{"id":94986331,"identity":"ce58ae5c-f3a7-4bb0-ba62-5752d13d7792","added_by":"auto","created_at":"2025-11-03 07:00:12","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":69085,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 3. Mutation analysis of the high risk group.\u003c/p\u003e","description":"","filename":"SupplementaryFigure3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/47dec880895f7102f45dd3ea.pdf"},{"id":94986757,"identity":"051ce643-c60d-47d8-841c-51b19eb6a96c","added_by":"auto","created_at":"2025-11-03 07:00:44","extension":"pdf","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":44260,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 4. Analysis of Tumor Immune Phenotype (TIP) between high risk group and low risk group.\u003c/p\u003e","description":"","filename":"SupplementaryFigure4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/edec33e97917b78678e5521b.pdf"},{"id":94874497,"identity":"80b41c79-0302-4277-82dd-a047f6e1a7bd","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":1833246,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 5. Differential expression of NCF4 and CXCL1 genes among subpopulations in single-cell analysis.\u003c/p\u003e","description":"","filename":"SupplementaryFigure5.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/9bf783f3c8588ccf7bb92add.pdf"},{"id":94874488,"identity":"36ef0308-988c-4ba8-8304-f8392e5cb749","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":20335,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 6. Analysis of transcription factor activity in the NCF4/CXCL1 subpopulation.\u003c/p\u003e","description":"","filename":"SupplementaryFigure6.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/e1b29f2fd8eef360c9786373.pdf"},{"id":94874490,"identity":"eec74b3e-123d-45ad-9c20-5bc1aec6ca24","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":32405,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 7. Intercellular communication analysis of the NCF4/CXCL1 subpopulation.\u003c/p\u003e","description":"","filename":"SupplementaryFigure7.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/85d6641c6e328c5d1decb352.pdf"},{"id":94874504,"identity":"2ed8c3ab-9c54-495a-bf71-94735574bc53","added_by":"auto","created_at":"2025-10-31 15:31:48","extension":"pdf","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":150625,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 8. Communication network analysis of the NCF4/CXCL1 subpopulation.\u003c/p\u003e","description":"","filename":"SupplementaryFigure8.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/5cbf9404d2136d5e5eeecbef.pdf"},{"id":94985675,"identity":"3326474e-a4aa-470f-a28d-b432980a5326","added_by":"auto","created_at":"2025-11-03 06:58:39","extension":"xlsx","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":11501798,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTable.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7489138/v1/6126e390eb4f4b95fa9e25cc.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Integrated Multi-Omics and Pathomics Analysis for Prognostic Modeling of Hepatocellular Carcinoma Reveals Oxidative Stress-Driven Immunometabolic Regulatory Mechanisms","fulltext":[{"header":"Introduction","content":"\u003cp\u003eHepatocellular carcinoma (HCC) is one of the leading causes of cancer-related mortality worldwide. Despite advances in diagnosis and treatment in recent years, patient outcomes remain unsatisfactory, and significant challenges persist in clinical decision-making [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].Currently, the traditional prognostic evaluation methods based on TNM staging have certain limitations [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Emerging technologies in the fields of genomics and pathology offer potential opportunities for more accurate prognostic prediction [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. However, how to effectively integrate multi-omics data and extract clinically meaningful biomarkers remains a critical challenge that needs to be addressed.\u003c/p\u003e\u003cp\u003eEmerging evidence suggests that the oxidative stress (OS) pathway may play a crucial role in the initiation and progression of hepatocellular carcinoma (HCC)[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. However, its specific regulatory mechanisms have not yet been fully elucidated. Recent studies have found that OS can induce mitochondrial metabolic reprogramming through the NRF2/KEAP1 axis, thereby promoting tumor immune escape by inhibiting ferroptosis [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. However, there remains a lack of integrated evidence regarding population-level genetic susceptibility for the aforementioned mechanisms. Meanwhile, although genome-wide association studies (GWAS) have identified multiple genetic loci associated with hepatocellular carcinoma (HCC) incidence [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], their contribution to prognostic prediction warrants further investigation.\u003c/p\u003e\u003cp\u003eFurthermore, the integration of digital pathology and deep learning techniques has provided new insights for tumor heterogeneity analysis; however, current approaches still have room for optimization in terms of feature interpretability and clinical applicability [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThis study aims to explore prognostic prediction methods for hepatocellular carcinoma (HCC) from a multi-omics perspective, with a focus on the key roles of oxidative stress (OS)-related genes in tumor initiation and progression. First, transcriptome-wide association analysis (TWAS) was conducted in combination with a known gene set related to oxidative stress, aiming to systematically identify susceptibility genes significantly associated with HCC prognosis and to preliminarily investigate their potential mechanisms in gene expression regulation and clinical outcomes. Second, an automated image analysis algorithm was employed to select representative regions from histopathological slides and to extract quantitative pathomic features, thereby capturing the heterogeneity of tumor microstructures. Subsequently, machine learning approaches were applied to integrate genetic variants, transcriptomic data, and pathological image features to construct a prognostic model with predictive performance. Furthermore, we analyzed the characteristics of immune cell infiltration in the tumor microenvironment and their correlations with risk scores, shedding light on potential immune regulatory mechanisms mediated by oxidative stress. This study intends to provide novel biomarkers and theoretical foundations for precise prognostic evaluation and personalized treatment strategies in HCC.\u003c/p\u003e"},{"header":"Method","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eGWAS data\u003c/h2\u003e\u003cp\u003eThe GWAS data for hepatocellular carcinoma were obtained from the dataset by Verma A et al [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], which includes 2,852 cases of European ancestry and 447,587 control samples.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eOxidative Stress Genes\u003c/h3\u003e\n\u003cp\u003eGenes associated with oxidative stress were retrieved and downloaded from the GeneCards database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.genecards.org/\u003c/span\u003e\u003cspan address=\"https://www.genecards.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), using a relevance score greater than 7 as the screening criterion [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].A total of 1065 oxidative stress-related genes were ultimately identified.\u003c/p\u003e\n\u003ch3\u003eTrans-organizational TWAS analysis\u003c/h3\u003e\n\u003cp\u003eWe employed the UTMOST analysis (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/Joker-Jerome/UTMOST?tab=readme-ov-file\u003c/span\u003e\u003cspan address=\"https://github.com/Joker-Jerome/UTMOST?tab=readme-ov-file\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) to quantify overall gene\u0026ndash;trait associations across tissues. This method enables the identification of a greater number of genes in tissues with substantial trait heritability and improves imputation accuracy[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Subsequently, we applied the generalized Berk-Jones (GBJ) test to integrate gene-trait associations by leveraging the covariance of single-tissue summary statistics [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. A significance level of P\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e\n\u003ch3\u003eHistopathological slide data\u003c/h3\u003e\n\u003cp\u003eThis study obtained hematoxylin and eosin (H\u0026amp;E)-stained whole-slide images (40\u0026times; magnification, 0.25 \u0026micro;m/pixel resolution) from 379 patients with hepatocellular carcinoma (HCC) through the TCGA database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://gdc.cancer.gov/\u003c/span\u003e\u003cspan address=\"https://gdc.cancer.gov/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). After loading the images using the Slide class, the most representative 512\u0026times;512 pixel regions (n\u0026thinsp;=\u0026thinsp;10 per slide) were selected using the ScoreTiler algorithm, ensuring a tissue coverage of no less than 75%. The image preprocessing pipeline included color deconvolution to eliminate staining artifacts, conversion from RGB to grayscale, Otsu\u0026rsquo;s thresholding, and morphological operations. Regions with areas smaller than 500 pixels\u0026sup2; (holes) or 1500 pixels\u0026sup2; (tissue fragments) were removed during processing. All subsequent analyses were conducted on image data resampled to a uniform magnification of 20\u0026times; or 40\u0026times;.\u003c/p\u003e\n\u003ch3\u003eDeep Learning Features\u003c/h3\u003e\n\u003cp\u003eThis study employs the ResNet-50 architecture for deep feature extraction from pathological images [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. All images were resized to 224\u0026times;224 pixels and preprocessed using ImageNet normalization parameters. The 2048-dimensional features from the output of the fourth residual block were specifically extracted and compressed into compact vectors through global max pooling. This approach effectively retains critical pathological information while significantly reducing computational complexity.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eVisualization of Prognostic-Related Features\u003c/h2\u003e\u003cp\u003eThis study employs the Grad-CAM algorithm to achieve visual interpretation of the decision-making process of the deep learning model [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. This method captures feature maps and gradient information by registering hook functions on the target network layers. The global average of the gradients corresponding to the target class is calculated as the channel weights, which are then used to generate weighted feature maps followed by ReLU activation. Subsequently, the activated feature maps are upsampled to the input image size via bilinear interpolation and normalized. Finally, the normalized maps are superimposed onto the original images with a transparency factor of 0.4 using the Jet colormap to produce heatmaps that visually highlight the critical visual regions relied upon by the model for decision-making [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003ePathological Image Features\u003c/h3\u003e\n\u003cp\u003eThis study conducted digital pathological image analysis using the CellProfiler platform [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Initially, hematoxylin- and eosin-stained channels were separated through image preprocessing. A multi-scale analytical strategy was employed to extract features from both global and local perspectives: at the global level, features including image quality, staining co-localization, granularity distribution, and multi-scale texture characteristics were extracted; at the local level, a hierarchical object recognition algorithm was applied to quantify staining intensity, morphology, spatial distribution, and local texture features. The mean, median, and standard deviation of each feature were calculated to construct a quantitative pathomic feature set [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eConstruction of Machine Learning Model\u003c/h3\u003e\n\u003cp\u003eA two-stage modeling strategy was employed in this study. First, univariate Cox regression analysis (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05) was conducted for preliminary feature selection. Subsequently, a predictive model was developed using the GBM (Gradient Boosting Machine) algorithm implemented in the Mime1 package (V 0.12)[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Model performance was evaluated using the C-index, time-dependent ROC curves (at 1-, 3-, and 5-year time points), and risk stratification based on Kaplan-Meier analysis. The GBM algorithm minimizes the loss function through iterative optimization of ensembles of decision trees, combining the feature selection advantages of StepCox [forward] with the nonlinear modeling capabilities of machine learning techniques. Feature importance was extracted from the trained GBM model, and features with non-zero importance scores were selected as the final feature set.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eSurvival Analysis\u003c/h2\u003e\u003cp\u003eThis study conducted survival analysis based on the clinical and risk score data from TCGA-LIHC. Key clinical features (pathological stage, TNM stage, age, and gender) were extracted. Univariate Cox regression analysis was performed to evaluate the impact of clinical factors (age, gender, T stage, N stage, M stage, and pathological stage) as well as the risk score on patient prognosis. Subsequently, multivariate Cox regression analysis was carried out on the variables that showed significance in the univariate analysis (T stage, pathological stage, and risk score).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eGene Set Enrichment Analysis (GSEA)\u003c/h2\u003e\u003cp\u003eA differential expression analysis was performed between the high-risk and low-risk groups based on a significance threshold of FDR\u0026thinsp;\u0026lt;\u0026thinsp;0.05, and a ranked gene list was generated according to the logFC values. GSEA was subsequently conducted using the clusterProfiler package (v 4.2.2) against the Hallmark gene sets from the MSigDB database (c2.cp.kegg.v2023.1.Hs.entrez.gmt). The distribution of each gene set within the ranked list was assessed by calculating the enrichment score, thereby identifying potential differences in biological pathways between the two groups.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eMutation Analysis\u003c/h2\u003e\u003cp\u003eSomatic mutation data were obtained from the TCGA database and analyzed using the Maftools package (v 2.21.1) [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. MAF-formatted mutation annotation files were generated for each group based on the high- and low-risk stratification. Comparative analysis of mutation features between the two groups focused on overall mutation burden, distribution of mutation types, and intergroup differences in frequently mutated genes.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003eImmune-Related Analysis\u003c/h2\u003e\u003cp\u003eTo investigate the role of immune checkpoints in patients stratified into high- and low-risk groups and to analyze their antigen-presenting capacity, we assessed the differential expression of multiple immune checkpoint genes and human leukocyte antigen (HLA) genes. Using the IOBR package (V 0.99.8), immune cell infiltration analysis was performed on gene expression data across risk groups employing the MCPcounter algorithm, allowing evaluation of the abundance of various immune cell types between the two groups[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Additionally, the TIDE algorithm[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e](\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://tide.dfci.harvard.edu/\u003c/span\u003e\u003cspan address=\"http://tide.dfci.harvard.edu/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) was utilized to predict the response status of samples to immune checkpoint blockade therapy.Differentially expressed genes between the high- and low-risk groups were analyzed in relation to immune response status. Upregulated and downregulated genes were separately subjected to Pathway-Level Metabolism and Proteome Analysis (ProteoMaps)[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e](\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.proteomaps.net/\u003c/span\u003e\u003cspan address=\"https://www.proteomaps.net/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The association between risk scores and immunotherapy response was further explored by comparing the similarity of upregulated and downregulated genes between the two groups. Moreover, a random subset of 22 samples from the high- and low-risk groups was selected to generate a matrix for TIP analysis (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://biocc.hrbmu.edu.cn/TIP/\u003c/span\u003e\u003cspan address=\"http://biocc.hrbmu.edu.cn/TIP/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), which was used to track and analyze the anti-cancer immune status and the proportion of tumor-infiltrating immune cells [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eSingle-Cell Analysis\u003c/h2\u003e\u003cp\u003eThis study analyzed single-cell transcriptomic data from GSE125449[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] using the Seurat package (v 4.3.0). Raw UMI count data were log-normalized (scale factor\u0026thinsp;=\u0026thinsp;10,000) and the top 3,000 most variable genes were selected for further normalization and scaling. PCA was performed for dimensionality reduction, followed by batch effect correction across samples using the Harmony package (v 0.1.1). Cell clustering was conducted based on the first 10 principal components, and results were visualized using UMAP. Malignant cells were classified into CXCL1\u0026thinsp;+\u0026thinsp;and CXCL1\u0026thinsp;\u0026minus;\u0026thinsp;subpopulations according to their gene expression profiles, while tumor-associated macrophages were categorized into NCF4\u0026thinsp;+\u0026thinsp;and NCF4\u0026thinsp;\u0026minus;\u0026thinsp;subgroups.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003eIntercellular Communication Analysis\u003c/h2\u003e\u003cp\u003eWe employed the CommPath package (v1.0.0) to analyze the intercellular communication network [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Pathway enrichment analysis was performed based on the KEGG database, and pathway activity was quantified using the GSVA method (minimum gene set size\u0026thinsp;=\u0026thinsp;10). Significantly enriched pathways with a positive score and p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 were selected, and the communication strength between cell subpopulations as well as the activity patterns of key signaling pathways were visualized.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003eTranscription Factor Activity Analysis\u003c/h2\u003e\u003cp\u003eTranscription factor activity analysis was performed using the DoRothEA package (v1.6.0), which assesses transcription factor activity based on known transcription factor\u0026ndash;target gene regulatory networks. High-confidence regulatory interactions (confidence levels A, B, and C) were selected from the DoRothEA database. The VIPER algorithm was then applied to calculate transcription factor activity scores from the single-cell RNA sequencing data. Finally, the top 40 transcription factors with the highest variability were selected for further analysis.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eA stratified random sampling strategy (stratified by clinical stage) was employed to divide the total sample into a training cohort (n\u0026thinsp;=\u0026thinsp;303) and an independent validation cohort (n\u0026thinsp;=\u0026thinsp;76) at a ratio of 8:2. Spearman\u0026rsquo;s rank correlation test (two-sided) was used to evaluate the associations between radiomic features and oxidative stress gene scores, as well as between digital pathological features and key gene expression levels. The strength of correlations was quantified using the Spearman correlation coefficient (ρ).\u003c/p\u003e\u003cp\u003eFor feature selection, elastic net regression (α\u0026thinsp;=\u0026thinsp;0.5) was applied, with the regularization parameter λ determined via 10-fold cross-validation. This hybrid penalized regression approach combines the advantages of L1 and L2 regularization, allowing for feature sparsity control while mitigating multicollinearity.\u003c/p\u003e\u003cp\u003eSurvival curves were constructed using the Kaplan-Meier method, and log-rank tests were used to compare survival differences between risk subgroups, with statistical significance defined as P\u0026thinsp;\u0026lt;\u0026thinsp;0.05. Univariate Cox regression analysis was first performed to screen potential prognostic factors (inclusion criterion: univariate P\u0026thinsp;\u0026lt;\u0026thinsp;0.1), followed by multivariate analysis to assess their independent prognostic value. Hazard ratios (HRs) are presented with 95% confidence intervals.\u003c/p\u003e\u003cp\u003eModel performance was evaluated through time-dependent receiver operating characteristic (ROC) analysis to assess 1-, 3-, and 5-year survival prediction accuracy, along with calculation and comparison of C-indices. The log-rank test was also used to evaluate the model\u0026rsquo;s ability to stratify patients into distinct risk groups.\u003c/p\u003e\u003cp\u003eAll computational implementations were conducted using Python 3.8.3 with the PyTorch 1.10 framework for deep learning tasks, and OpenCV 4.5 was used for image processing. Statistical analyses were carried out in R 4.1.3, primarily using the survival package (version 3.3.1) for survival modeling, the limma package (version 3.50) for differential analysis, and the pROC package (version 1.18) for ROC analysis. Multiple hypothesis testing was adjusted using the Benjamini\u0026ndash;Hochberg procedure, and a false discovery rate (FDR)\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant. All tests were two-sided.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eIn the cross-tissue transcriptome-wide association analysis (TWAS), we identified 176 genes (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). After intersecting with oxidative stress-related genes (Table \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e), 12 significantly associated genes were ultimately selected.\u003c/p\u003e\u003cp\u003eThis study performed tissue region extraction and cell density scoring analysis on whole-slide images (WSIs) of hepatocellular carcinoma (HCC) from The Cancer Genome Atlas (TCGA). Samples were divided into 20\u0026times; (n\u0026thinsp;=\u0026thinsp;10) and 40\u0026times; (n\u0026thinsp;=\u0026thinsp;369) magnification groups. A tile size of 512\u0026times;512 pixels was used, and a multi-step image processing pipeline was applied, including hematoxylin artifact removal, grayscale conversion, Otsu thresholding, morphological operations (erosion and dilation), and small region filtering. A tissue content threshold of 75% was set, and 10 representative tissue tiles were selected per group. This method successfully achieved standardized processing and analysis across WSIs of different magnifications.\u003c/p\u003e\u003cp\u003eA ResNet-50 pretrained model was employed for deep feature extraction from 3,783 pathological image tiles. The layer4 output of ResNet-50 was selected as the feature extraction layer. Each image was preprocessed to a standard size of 224\u0026times;224 pixels and normalized across RGB channels. A 2048-dimensional feature vector was extracted per image (corresponding to the output channel number of layer4). Maximum spatial pooling was applied across the spatial dimensions to obtain the most representative feature value for each channel, resulting in a 2048-dimensional deep feature representation for each pathological image.\u003c/p\u003e\u003cp\u003ePathological features were extracted using CellProfiler for subsequent analysis. Granularity, intensity, and texture features were extracted, along with mean, median, and standard deviation features for each object. All pathological features were grouped by sample and averaged. Similarly, deep features were also averaged per sample. Finally, pathological and deep features were integrated into a single feature matrix for further analysis (Table \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eSingle-sample gene set enrichment analysis (SSGSEA) was used to compute enrichment scores for the 12 associated genes. Spearman correlation analysis revealed that 378 features were significantly correlated with the gene set scores in the training cohort (n\u0026thinsp;=\u0026thinsp;287) (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Table \u003cspan refid=\"MOESM4\" class=\"InternalRef\"\u003eS4\u003c/span\u003e). Subsequently, elastic net regression (α\u0026thinsp;=\u0026thinsp;0.2) was applied to select 70 most predictive features, including 63 ResNet-derived features and 7 CellProfiler-derived features. Network analysis demonstrated significant associations between these features and gene set scores (correlation coefficient range: -0.229 to 0.274) (Supplementary Fig.\u0026nbsp;1, Table \u003cspan refid=\"MOESM5\" class=\"InternalRef\"\u003eS5\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eUnivariate Cox regression analysis was performed on the 70 selected features in the training cohort to identify survival-associated features (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Table \u003cspan refid=\"MOESM6\" class=\"InternalRef\"\u003eS6\u003c/span\u003e). Grad-CAM was applied to visualize the deep learning features on pathological images. The results indicated that the model primarily focused on specific regions within the images (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Heatmap visualization illustrated the spatial distribution of key features that contributed to model decision-making, which may correspond to diagnostically relevant pathological features.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eA prognostic signature constructed using the \"StepCox [forward]\u0026thinsp;+\u0026thinsp;GBM\" model demonstrated good predictive performance in training and validation cohorts (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). Survival analysis showed that the model effectively stratified patients into high-risk and low-risk groups (log-rank p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA). Time-dependent ROC curve analysis revealed AUC values of 0.834, 0.888, and 0.918 at 1, 3, and 5 years in the training cohort, and 0.747, 0.814, and 0.826 in the validation cohort, respectively, indicating stable predictive capability (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). The C-index reached 0.75 in the validation cohort, suggesting strong discriminative performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eUnivariate and multivariate Cox regression analyses using TCGA-LIHC data evaluated the prognostic impact of RiskScore and clinical parameters in HCC patients (Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC, D; Tables S7, S8). Univariate analysis identified significant associations for T stage (HR\u0026thinsp;=\u0026thinsp;1.802, 95%CI\u0026thinsp;=\u0026thinsp;1.399\u0026ndash;2.322, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), tumor stage (HR\u0026thinsp;=\u0026thinsp;1.855, 95%CI\u0026thinsp;=\u0026thinsp;1.41\u0026ndash;2.441, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), and RiskScore (HR\u0026thinsp;=\u0026thinsp;26.801, 95%CI\u0026thinsp;=\u0026thinsp;11.859\u0026ndash;60.571, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Multivariate analysis confirmed RiskScore (HR\u0026thinsp;=\u0026thinsp;25.402, 95%CI\u0026thinsp;=\u0026thinsp;10.687\u0026ndash;60.381, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) as an independent prognostic factor.\u003c/p\u003e\u003cp\u003eHCC patients were stratified into high- and low-risk groups by median RiskScore. KEGG pathway GSEA revealed significantly enriched pathways (p.adjust\u0026thinsp;\u0026lt;\u0026thinsp;0.05) with |NES|\u0026ge;2, including fatty acid metabolism (NES=-2.83), PPAR signaling (NES=-2.48), and steroid hormone biosynthesis (NES=-2.50), associated with HCC prognosis (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA, Table \u003cspan refid=\"MOESM9\" class=\"InternalRef\"\u003eS9\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSomatic mutation profiles were compared between risk groups, revealing significant differences in the top 20 mutated genes. TP53 (31%) and TNN (27%) exhibited higher mutation frequencies in high-risk patients, with TP53 mutations differing between groups. Mutation waterfall plots showed distinct signatures, indicating association between RiskScore and molecular subtypes. Clinical annotation confirmed mutation distribution consistency with risk stratification, validating reliability (Supplementary Figs.\u0026nbsp;2, 3).\u003c/p\u003e\u003cp\u003eTumor immune phenotypes (TIP) were analyzed in 22 samples per risk group. Significant differences in immune cell infiltration were observed between groups. A heatmap revealed distinct immune cell composition profiles, associating RiskScore with tumor immune status (Supplementary Fig.\u0026nbsp;4).\u003c/p\u003e\u003cp\u003eDifferential expression analysis of immune checkpoint and HLA genes in TCGA-LIHC high- vs. low-risk groups revealed significant differences (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05) for several immune checkpoints and HLA-B/C (Figs.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB, C), suggesting prognostic links. MCPcounter assessed immune infiltration, showing higher T cell and monocyte levels in high-risk patients (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eD). Proteomaps visualization indicated significant enrichment of lipid/steroid and amino acid metabolism pathways in both high-risk and immunotherapy response groups, demonstrating similarity. These findings provide insights into tumor risk stratification and immunotherapy response mechanisms (Figs.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eE, F).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIntegration of pathological image features and gene expression data revealed significant Spearman correlations (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) between histopathological features and multiple genes (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA). Network analysis illustrated interactions between pathological features (ResNet, CellProfiler) and genes, identifying NCF4 and CXCL1 with prognostic relevance (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eB). Survival analysis showed better overall survival in low-expression groups of NCF4 and CXCL1 versus high-expression groups (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05), suggesting prognostic value (Figs.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC, D).\u003c/p\u003e\u003cp\u003eAnalysis of GSE125449 single-cell RNA sequencing data characterized HCC at the cellular level. Clustering identified major cell types like malignant cells and TAMs. Differential expression of NCF4 and CXCL1 across subpopulations and samples revealed key gene expression features at single-cell resolution (Figs.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA, Supplementary Fig.\u0026nbsp;5).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTranscription factor activity analysis in NCF4/CXCL1 subpopulations identified the top 40 variable TFs. SPI1 and AR showed differential activity in NCF4\u0026thinsp;+\u0026thinsp;TAMs, while NFKB1 and RELA did so in CXCL1\u0026thinsp;+\u0026thinsp;tumor cells, indicating cell-type-specific regulation(Figs.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eB, Supplementary Fig.\u0026nbsp;6).\u003c/p\u003e\u003cp\u003eCell-cell communication analysis revealed pathway activation with distinct specificity in tumor subpopulations: Leukocyte Transendothelial Migration was enriched in NCF4\u0026thinsp;+\u0026thinsp;TAMs, while the Rap1 Signaling Pathway was enriched in CXCL1\u0026thinsp;+\u0026thinsp;tumor cells (Figs.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eC, Supplementary Fig.\u0026nbsp;7). Network analysis showed intercellular communication strength, suggesting functional synergy among specific subpopulations (Figs.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eD, Supplementary Fig.\u0026nbsp;8).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study constructed an interpretable HCC prognostic model centered on oxidative stress by integrating genomic, transcriptomic, pathomic, and single-cell sequencing data. We elucidated the biological basis of risk stratification across molecular mechanisms, microenvironment heterogeneity, and immune regulation.\u003c/p\u003e\u003cp\u003eCross-tissue TWAS identified 12 oxidative stress pathway genes significantly associated with HCC, potentially playing critical roles in its progression. NCF4, an essential NADPH oxidase subunit, may cause dysregulated ROS levels, promoting DNA damage and tumor evolution [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].CXCL1 recruits MDSCs, inducing immunosuppression and facilitating immune evasion [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].Notably, at single-cell resolution, we observed for the first time the cell type-specific expression patterns of these two genes: CXCL1 is predominantly enriched in malignant epithelial cells, whereas NCF4 is highly expressed in tumor-associated macrophages. This finding not only reveals the potential for intercellular functional coordination within the hepatocellular carcinoma microenvironment, but also provides a foundation for future investigations into targeted interventions directed at specific cell types. Moreover, it underscores the significant role of oxidative stress in modulating the tumor immune microenvironment.\u003c/p\u003e\u003cp\u003eOur digital pathology analysis developed a multimodal feature extraction pipeline integrating deep learning with traditional histopathological features, identifying multiple quantitative prognostic biomarkers. Over 90% of the 70 predictive features derived from deep ResNet-50 layer 4, indicating convolutional neural networks capture subtle microstructural information challenging to quantify manually [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e].Using Grad-CAM, we revealed model attention aligned with pathological features like tumor heterogeneity and mitotic figures, aiding model interpretability [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].This aligns with AI-assisted pathology trends, highlighting deep learning's potential in digital pathology [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eWe developed a clinically applicable prognostic model based on oxidative stress features using a \"StepCox [forward]\u0026thinsp;+\u0026thinsp;GBM\" approach. It demonstrated strong discriminative performance in both training (n\u0026thinsp;=\u0026thinsp;287) and validation (n\u0026thinsp;=\u0026thinsp;96) cohorts, with time-dependent AUCs of 0.834/0.888/0.918 (1/3/5-year, training) and 0.747/0.814/0.826 (validation). The risk score was an independent prognostic factor regardless of clinical stage (HR\u0026thinsp;=\u0026thinsp;25.402, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and maintained stable long-term performance (\u0026lt;\u0026thinsp;5% ROC-AUC decrease over time). This predictive stability holds significant clinical relevance given HCC's prolonged course [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e].\u003c/p\u003e\u003cp\u003ePathway analysis showed distinct metabolic alterations between high- and low-risk groups. GSEA revealed suppression of fatty acid metabolism, PPAR signaling, and steroid hormone biosynthesis in the high-risk group (NES \u0026lt; \u0026minus;\u0026thinsp;2, FDR\u0026thinsp;\u0026lt;\u0026thinsp;0.05), potentially promoting tumor proliferation by disrupting energy supply and membrane synthesis [\u003cspan additionalcitationids=\"CR35\" citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. ProteoMaps indicated similar lipid/amino acid metabolism patterns in the high-risk and immunotherapy-responsive groups, suggesting these pathways could predict immune checkpoint blockade efficacy and support integrating metabolic modulation into immunotherapy strategies [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. These results further validate our proposed \"oxidative stress\u0026ndash;metabolism\u0026ndash;immune dysregulation\" hepatocarcinogenesis triad.\u003c/p\u003e\u003cp\u003eIn terms of the tumor immune microenvironment, TIP analysis revealed significant differences in the infiltration profiles of immune cells between the high- and low-risk groups. Integrating single-cell RNA sequencing data, we found that NCF4 and CXCL1 exhibited distinct expression patterns in tumor-associated macrophages and malignant cells, respectively. These findings not only provide cellular-level evidence supporting the biological plausibility of the risk stratification, but also suggest that specific intercellular interactions may play a pivotal role in driving disease progression[\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. Cell communication analysis substantiated this, identifying multiple ligand\u0026ndash;receptor pairs mediating high-intensity interactions[\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e], collectively highlighting the multidimensional role of oxidative stress-related genes in shaping the tumor microenvironment.\u003c/p\u003e\u003cp\u003eMutation signature analysis revealed an association between the risk score and HCC molecular subtypes. The high-risk group showed increased TP53 mutation frequency and more complex mutation patterns, suggesting greater genomic instability. As a classical tumor suppressor central to DNA repair and cell cycle regulation [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], TP53 mutations are closely linked to hepatocarcinogenesis and are an established biomarker for poor prognosis [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. The high-risk group also had a significantly higher TNN (Titin) mutation rate. Though encoding a large skeletal muscle protein with incompletely understood roles in liver cancer[\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e], evidence suggests its mutations may reflect tumor mutational burden and associate with worse outcomes.\u003c/p\u003e\u003cp\u003eIn the high-risk HCC subgroup, we observed significant upregulation of key immune checkpoint molecules PDCD1 (PD-1), CTLA4, and LAG3, suggesting severe T-cell exhaustion that may impair anti-tumor immunity[\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. The PD-1/PD-L1 pathway is an established mechanism of immune evasion in HCC and a major target for immunotherapy[\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. However, concurrent elevation of CTLA-4 and LAG3 suggests activation of multiple immunosuppressive pathways, which could explain immunotherapy resistance and variable treatment responses.\u003c/p\u003e\u003cp\u003eWe also observed aberrant upregulation of HLA class I molecules (HLA-B, HLA-C) in the high-risk group. While HLA class I downregulation is a recognized immune evasion mechanism in immunogenic tumors[\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e], studies in HCC suggest abnormal upregulation may represent a compensatory immune activation or adaptive response to chronic inflammation [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. This \"elevated yet dysfunctional\" state warrants further functional validation. Immune infiltration analysis revealed higher levels of T cells and monocytes in the high-risk group, suggesting a more active immune response[\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eMore intriguingly, comparing protein functional networks revealed high concordance between the high-risk group and immunotherapy-responsive group in lipid metabolism. This suggests lipid metabolism may critically link disease progression and immunotherapy response in the HCC microenvironment [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eLipid metabolism pathways are frequently dysregulated in HCC, promoting tumor proliferation, invasion, and immune evasion [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e]. Crucially, lipid metabolism modulates T cell function: FAO enhances memory T cells, while cholesterol accumulation can cause T cell exhaustion [\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e].This may underlie the immune features observed in our high-risk patients. Although these patients have poorer outcomes, their preserved lipid metabolism suggests retained potential for antitumor immune responses under specific conditions. Thus, even conventionally high-risk HCC patients may benefit from checkpoint inhibitors, especially combined with lipid metabolism-targeting therapies[\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTranscription factor activity analysis in NCF4/CXCL1-associated subpopulations revealed distinct regulatory patterns within the TME. SPI1 and AR activity significantly differed in TAMs with differential NCF4 expression, suggesting involvement in regulating polarization or functional states [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e].In contrast, tumor cells with differential CXCL1 expression showed significant NFKB1/RELA activity differences, indicating NF-κB pathway activation [\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e].Given NF-κB's role in inflammation, survival, and plasticity, this suggests CXCL1-high tumor cells exhibit enhanced inflammatory responsiveness and may actively recruit/remodel immune cells via chemokine secretion. These findings support a potential positive feedback loop between oxidative stress and NF-κB signaling in HCC progression.\u003c/p\u003e\u003cp\u003eIntercellular communication analysis further supports this transcriptional regulation. Pathways for Leukocyte Transendothelial Migration and Rap1 Signaling were significantly enriched in NCF4\u0026thinsp;+\u0026thinsp;TAMs and CXCL1\u0026thinsp;+\u0026thinsp;tumor cells [\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e]. These pathways mediate cell adhesion and migration\u0026mdash;key processes in immune infiltration and metastasis\u0026mdash;indicating distinct cell populations drive coordinated yet functionally divergent programs. Differential oxidative stress-related gene expression likely drives this functional heterogeneity.\u003c/p\u003e\u003cp\u003eThis study's innovation includes: developing the first oxidative stress-centered multimodal prognostic framework, characterizing single-cell oxidative stress-related gene expression profiles, and elucidating an oxidative stress-mediated metabolic-immune-genomic network. Despite limitations in external validation and functional experiments, it provides a foundation for oxidative stress-targeted precision therapy in hepatocellular carcinoma.\u003c/p\u003e\u003cp\u003eThis study has limitations. Model validation was primarily limited to the TCGA-LIHC cohort; external validation in independent cohorts is lacking and requires future prospective studies. Furthermore, while mechanistic hypotheses were inferred from multi-omics analysis, direct functional evidence is absent, necessitating validation in vitro and in vivo.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn summary, this study developed an oxidative stress-centered prognostic model for hepatocellular carcinoma through multi-omics analysis. The model demonstrates high discriminative ability and favorable generalization. Underlying biological mechanisms \u0026ndash; spanning molecular, metabolic, immune, and cellular interactions \u0026ndash; were systematically elucidated. Notably, the identified oxidative stress-related biomarkers linked to immune escape and metabolic regulation provide foundations and targets for novel diagnostics and therapeutics.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eDisclosure\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors report no conflicts of interest in this work.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eNone.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLunwei Yang (conceptualization, methodology, formal analysis, investigation, writing—original draft), Yuanliang Li (conceptualization, methodology, formal analysis, writing—original draft), Xiaoying Zhong (writing—review and editing, supervision, project administration). All authors read and approved the final manuscript. All authors have read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe GWAS data for hepatocellular carcinoma were obtained from the dataset by Verma A et al.The TCGA dataset is publicly available at the TCGA portal (https://portal.gdc.cancer.gov).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePatient consent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eHwang, S. Y. et al. Hepatocellular carcinoma: updates on epidemiology, surveillance, diagnosis and treatment. \u003cem\u003eClin. Mol. Hepatol.\u003c/em\u003e \u003cb\u003e31\u003c/b\u003e (Suppl), S228\u0026ndash;S54 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRumgay, H. et al. Global burden of primary liver cancer in 2020 and predictions to 2040. \u003cem\u003eJ. Hepatol.\u003c/em\u003e \u003cb\u003e77\u003c/b\u003e (6), 1598\u0026ndash;1606 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTrevisani, F. et al. Merits and boundaries of the BCLC staging and treatment algorithm: Learning from the past to improve the future with a novel proposal. \u003cem\u003eJ. Hepatol.\u003c/em\u003e \u003cb\u003e80\u003c/b\u003e (4), 661\u0026ndash;669 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChaudhary, K., Poirion, O. B., Lu, L. \u0026amp; Garmire, L. X. Deep Learning-Based Multi-Omics Integration Robustly Predicts Survival in Liver Cancer. \u003cem\u003eClin. Cancer Res.\u003c/em\u003e \u003cb\u003e24\u003c/b\u003e (6), 1248\u0026ndash;1259 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang, Z. et al. Oxidative Stress Promotes Liver Cancer Metastasis via RNF25-Mediated E-Cadherin Protein Degradation. \u003cem\u003eAdv. Sci. (Weinh)\u003c/em\u003e. \u003cb\u003e11\u003c/b\u003e (13), e2306929 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSun, X. et al. Activation of the p62-Keap1-NRF2 pathway protects against ferroptosis in hepatocellular carcinoma cells. \u003cem\u003eHepatology\u003c/em\u003e \u003cb\u003e63\u003c/b\u003e (1), 173\u0026ndash;184 (2016).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHassan, M. M. et al. Genome-wide association study identifies high-impact susceptibility loci for HCC in North America. \u003cem\u003eHepatology\u003c/em\u003e \u003cb\u003e80\u003c/b\u003e (1), 87\u0026ndash;101 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNakatsuka, T. et al. Deep learning and digital pathology powers prediction of HCC development in steatotic liver disease. \u003cem\u003eHepatology\u003c/em\u003e \u003cb\u003e81\u003c/b\u003e (3), 976\u0026ndash;989 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNam, D., Chapiro, J., Paradis, V., Seraphin, T. P. \u0026amp; Kather, J. N. Artificial intelligence in liver diseases: Improving diagnostics, prognostics and response prediction. \u003cem\u003eJHEP Rep.\u003c/em\u003e \u003cb\u003e4\u003c/b\u003e (4), 100443 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVerma, A. et al. Diversity and scale: Genetic architecture of 2068 traits in the VA Million Veteran Program. \u003cem\u003eScience\u003c/em\u003e \u003cb\u003e385\u003c/b\u003e (6706), eadj1182 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu, M. et al. Identification and validation of immune and oxidative stress-related diagnostic markers for diabetic nephropathy by WGCNA and machine learning. \u003cem\u003eFront. Immunol.\u003c/em\u003e \u003cb\u003e14\u003c/b\u003e, 1084531 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHu, Y. et al. A statistical framework for cross-tissue transcriptome-wide association analysis. \u003cem\u003eNat. Genet.\u003c/em\u003e \u003cb\u003e51\u003c/b\u003e (3), 568\u0026ndash;576 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSun, R., Hui, S., Bader, G. D., Lin, X. \u0026amp; Kraft, P. Powerful gene set analysis in GWAS with the Generalized Berk-Jones statistic. \u003cem\u003ePLoS Genet.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (3), e1007530 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMarme, F. et al. Deep learning to predict breast cancer sentinel lymph node status on INSEMA histological images. \u003cem\u003eEur. J. Cancer\u003c/em\u003e. \u003cb\u003e195\u003c/b\u003e, 113390 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSelvaraju, R. R., Das, M. C. A., Vedantam, R., Parikh, D. \u0026amp; Batra, D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. IEEE International Conference on Computer Vision (ICCV). 2017;Venice:618\u0026thinsp;\u0026ndash;\u0026thinsp;26. (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRamprasaath, R., Selvaraju, A. D., Vedantam, R., Cogswell, M. \u0026amp; Parikh, D. Dhruv Batra. Grad-CAM: why did you say that? arXiv. ;arXiv:1611.07450 (2016).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCarpenter, A. E. et al. CellProfiler: image analysis software for identifying and quantifying cell phenotypes. \u003cem\u003eGenome Biol.\u003c/em\u003e \u003cb\u003e7\u003c/b\u003e (10), R100 (2006).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJones, T. R. et al. CellProfiler Analyst: data exploration and analysis software for complex image-based screens. \u003cem\u003eBMC Bioinform.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 482 (2008).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJones, T. R. et al. Scoring diverse cellular morphologies in image-based screens with iterative feedback and machine learning. \u003cem\u003eProc. Natl. Acad. Sci. U S A\u003c/em\u003e. \u003cb\u003e106\u003c/b\u003e (6), 1826\u0026ndash;1831 (2009).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, H. et al. Mime: A flexible machine-learning framework to construct and visualize models for clinical characteristics prediction and feature selection. \u003cem\u003eComput. Struct. Biotechnol. J.\u003c/em\u003e \u003cb\u003e23\u003c/b\u003e, 2798\u0026ndash;2810 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMayakonda, A., Lin, D. C., Assenov, Y., Plass, C. \u0026amp; Koeffler, H. P. Maftools: efficient and comprehensive analysis of somatic variants in cancer. \u003cem\u003eGenome Res.\u003c/em\u003e \u003cb\u003e28\u003c/b\u003e (11), 1747\u0026ndash;1756 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZeng, D. et al. IOBR: Multi-Omics Immuno-Oncology Biological Research to Decode Tumor Microenvironment and Signatures. \u003cem\u003eFront. Immunol.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e, 687975 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFu, J. et al. Large-scale public data reuse to model immunotherapy response and resistance. \u003cem\u003eGenome Med.\u003c/em\u003e \u003cb\u003e12\u003c/b\u003e (1), 21 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiebermeister, W. et al. Visual account of protein investment in cellular functions. \u003cem\u003eProc. Natl. Acad. Sci. U S A\u003c/em\u003e. \u003cb\u003e111\u003c/b\u003e (23), 8488\u0026ndash;8493 (2014).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu, L. et al. TIP: A Web Server for Resolving Tumor Immunophenotype Profiling. \u003cem\u003eCancer Res.\u003c/em\u003e \u003cb\u003e78\u003c/b\u003e (23), 6575\u0026ndash;6580 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMa, L. et al. Tumor Cell Biodiversity Drives Microenvironmental Reprogramming in Liver Cancer. \u003cem\u003eCancer Cell.\u003c/em\u003e \u003cb\u003e36\u003c/b\u003e (4), 418\u0026ndash;430 (2019). e6.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi, L. et al. NCF4 attenuates colorectal cancer progression by modulating inflammasome activation and immune surveillance. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (1), 5170 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKorbecki, J. et al. The Effect of Hypoxia on the Expression of CXC Chemokines and CXC Chemokine Receptors-A Review of Literature. \u003cem\u003eInt. J. Mol. Sci.\u003c/em\u003e ;\u003cb\u003e22\u003c/b\u003e(2). (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang, X. et al. A pathology foundation model for cancer diagnosis and prognosis prediction. \u003cem\u003eNature\u003c/em\u003e \u003cb\u003e634\u003c/b\u003e (8035), 970\u0026ndash;978 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAcs, B., Rantalainen, M. \u0026amp; Hartman, J. Artificial intelligence as the next step towards precision pathology. \u003cem\u003eJ. Intern. Med.\u003c/em\u003e \u003cb\u003e288\u003c/b\u003e (1), 62\u0026ndash;81 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSauter, D. et al. Validating Automatic Concept-Based Explanations for AI-Based Digital Histopathology. \u003cem\u003eSens. (Basel)\u003c/em\u003e ;\u003cb\u003e22\u003c/b\u003e(14). (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBaxi, V., Edwards, R., Montalto, M. \u0026amp; Saha, S. Digital pathology and artificial intelligence in translational medicine and clinical practice. \u003cem\u003eMod. Pathol.\u003c/em\u003e \u003cb\u003e35\u003c/b\u003e (1), 23\u0026ndash;32 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBialecki, E. S. \u0026amp; Di Bisceglie, A. M. Clinical presentation and natural course of hepatocellular carcinoma. \u003cem\u003eEur. J. Gastroenterol. Hepatol.\u003c/em\u003e \u003cb\u003e17\u003c/b\u003e (5), 485\u0026ndash;489 (2005).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQin, L. et al. ELOVL1 is upregulated and promotes tumor growth in hepatocellular carcinoma through regulating PI3K-AKT-mTOR signaling. \u003cem\u003eHeliyon\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (15), e34961 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao, X. et al. NR4A3 inhibits the tumor progression of hepatocellular carcinoma by inducing cell cycle G0/G1 phase arrest and upregulation of CDKN2AIP expression. \u003cem\u003eInt. J. Biol. Sci.\u003c/em\u003e \u003cb\u003e20\u003c/b\u003e (15), 5850\u0026ndash;5867 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWei, Q. et al. Estrogen suppresses hepatocellular carcinoma cells through ERbeta-mediated upregulation of the NLRP3 inflammasome. \u003cem\u003eLab. Invest.\u003c/em\u003e \u003cb\u003e95\u003c/b\u003e (7), 804\u0026ndash;816 (2015).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXia, L. et al. The cancer metabolic reprogramming and immune response. \u003cem\u003eMol. Cancer\u003c/em\u003e. \u003cb\u003e20\u003c/b\u003e (1), 28 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZheng, J. et al. Hepatocellular carcinoma: signaling pathways and therapeutic advances. \u003cem\u003eSignal. Transduct. Target. Ther.\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (1), 35 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMao, X. et al. Crosstalk between cancer-associated fibroblasts and immune cells in the tumor microenvironment: new findings and future perspectives. \u003cem\u003eMol. Cancer\u003c/em\u003e. \u003cb\u003e20\u003c/b\u003e (1), 131 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, Y. et al. Identification of a tumour immune barrier in the HCC microenvironment that determines the efficacy of immunotherapy. \u003cem\u003eJ. Hepatol.\u003c/em\u003e \u003cb\u003e78\u003c/b\u003e (4), 770\u0026ndash;782 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVoskarides, K. \u0026amp; Giannopoulou, N. The Role of TP53 in Adaptation and Evolution. \u003cem\u003eCells\u003c/em\u003e ;\u003cb\u003e12\u003c/b\u003e(3). (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOrtiz-Cuaran, S. et al. Association between TP53 R249S mutation and polymorphisms in TP53 intron 1 in hepatocellular carcinoma. \u003cem\u003eGenes Chromosomes Cancer\u003c/em\u003e. \u003cb\u003e52\u003c/b\u003e (10), 912\u0026ndash;919 (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHan, X., Chen, J., Wang, J., Xu, J. \u0026amp; Liu, Y. TTN mutations predict a poor prognosis in patients with thyroid cancer. \u003cem\u003eBiosci. Rep.\u003c/em\u003e \u003cb\u003e42\u003c/b\u003e, 7 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMeng, Y. et al. Immunosuppressive CD10(+)ALPL(+) neutrophils promote resistance to anti-PD-1 therapy in HCC by mediating irreversible exhaustion of T cells. \u003cem\u003eJ. Hepatol.\u003c/em\u003e \u003cb\u003e79\u003c/b\u003e (6), 1435\u0026ndash;1449 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhou, B. et al. Hepatoma cell-intrinsic TLR9 activation induces immune escape through PD-L1 upregulation in hepatocellular carcinoma. \u003cem\u003eTheranostics\u003c/em\u003e \u003cb\u003e10\u003c/b\u003e (14), 6530\u0026ndash;6543 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBauer, J. et al. The oncogenic fusion protein DNAJB1-PRKACA can be specifically targeted by peptide-based immunotherapy in fibrolamellar hepatocellular carcinoma. \u003cem\u003eNat. Commun.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e (1), 6401 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eStolarz, W. et al. [The HLA-I and HLA-II expression evaluation in liver biopsy specimens of patients with chronic viral hepatitis]. \u003cem\u003ePrzegl Epidemiol.\u003c/em\u003e \u003cb\u003e53\u003c/b\u003e (3\u0026ndash;4), 385\u0026ndash;393 (1999).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTacke, F. Targeting hepatic macrophages to treat liver diseases. \u003cem\u003eJ. Hepatol.\u003c/em\u003e \u003cb\u003e66\u003c/b\u003e (6), 1300\u0026ndash;1312 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNicole, L. et al. Necroptosis-driving genes RIPK1, RIPK3 and MLKL-p are associated with intratumoral CD3(+) and CD8(+) T cell density and predict prognosis in hepatocellular carcinoma. \u003cem\u003eJ. Immunother Cancer\u003c/em\u003e ;\u003cb\u003e10\u003c/b\u003e(3). (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBian, X. et al. Lipid metabolism and cancer. \u003cem\u003eJ. Exp. Med.\u003c/em\u003e ;\u003cb\u003e218\u003c/b\u003e(1). (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSu, F. \u0026amp; Koeberle, A. Regulation and targeting of SREBP-1 in hepatocellular carcinoma. \u003cem\u003eCancer Metastasis Rev.\u003c/em\u003e \u003cb\u003e43\u003c/b\u003e (2), 673\u0026ndash;708 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu, Z. et al. Deficiency of SIAH1 promotes the formation of filopodia by increasing the accumulation of FASN in liver cancer. \u003cem\u003eCell. Death Dis.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (7), 537 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLoschinski, R. et al. IL-21 modulates memory and exhaustion phenotype of T-cells in a fatty acid oxidation-dependent manner. \u003cem\u003eOncotarget\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e (17), 13125\u0026ndash;13138 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRamapriyan, R. et al. Altered cancer metabolism in mechanisms of immunotherapy resistance. \u003cem\u003ePharmacol. Ther.\u003c/em\u003e \u003cb\u003e195\u003c/b\u003e, 162\u0026ndash;171 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQi, S. et al. SPI1-mediated macrophage polarization aggravates age-related macular degeneration. \u003cem\u003eFront. Immunol.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e, 1421012 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTaniguchi, K. \u0026amp; Karin, M. NF-kappaB, inflammation, immunity and cancer: coming of age. \u003cem\u003eNat. Rev. Immunol.\u003c/em\u003e \u003cb\u003e18\u003c/b\u003e (5), 309\u0026ndash;324 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLu, Z. H. et al. Early administration of Wumei Wan inhibit myeloid-derived suppressor cells via PI3K/Akt pathway and amino acids metabolism to prevent colitis-associated colorectal cancer. \u003cem\u003eJ. Ethnopharmacol.\u003c/em\u003e \u003cb\u003e333\u003c/b\u003e, 118260 (2024).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Hepatocellular carcinoma (HCC), Oxidative stress, Multi-omics integration, Prognostic model, Tumor microenvironment, Pathomics","lastPublishedDoi":"10.21203/rs.3.rs-7489138/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7489138/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eHepatocellular carcinoma (HCC), a major cause of cancer mortality, exhibits strong ties to oxidative stress (OS), though integrated multi-omics studies linking OS mechanisms to clinically predictive models remain scarce. To address this, we integrated European-descent GWAS data (2,852 HCC cases vs. 447,587 controls) with 1,065 OS-related genes, identifying 176 potential HCC-associated genes (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05) via TWAS (UTMOST/GBJ tests), including 12 key OS drivers. Pathomic features extracted from 379 TCGA HCC histopathological images (ResNet-50/CellProfiler) informed prognostic modeling, with histopathology-gene correlations mapped via Spearman analysis. Single-cell transcriptomics (GSE125449) uncovered CXCL1⁺ malignant cell interactions with NCF4⁺ macrophages through the CCL20-CCR6 axis. An elastic net-selected 70-feature gradient boosting machine (GBM) model demonstrated robust prognostic performance (training: 1/3/5-year AUC\u0026thinsp;=\u0026thinsp;0.834/0.888/0.918; validation: AUC\u0026thinsp;=\u0026thinsp;0.747/0.814/0.826), with the risk score serving as an independent prognostic factor (HR\u0026thinsp;=\u0026thinsp;25.402, P\u0026thinsp;\u0026lt;\u0026thinsp;0.001). TCGA analyses further linked risk scores to altered immune microenvironments, somatic mutations (e.g., TP53), and activated energy/metabolic pathways. This study elucidates OS-driven immunometabolic regulatory mechanisms in HCC and delivers an integrated histology-genomic prognostic model with implications for immunotherapy strategies.\u003c/p\u003e","manuscriptTitle":"Integrated Multi-Omics and Pathomics Analysis for Prognostic Modeling of Hepatocellular Carcinoma Reveals Oxidative Stress-Driven Immunometabolic Regulatory Mechanisms","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-31 15:31:43","doi":"10.21203/rs.3.rs-7489138/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"56bfc663-cff1-40c8-bbf4-233eeb28af08","owner":[],"postedDate":"October 31st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":57022189,"name":"Health sciences/Biomarkers"},{"id":57022190,"name":"Biological sciences/Cancer"},{"id":57022191,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":57022192,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2025-11-11T02:08:34+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-31 15:31:43","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7489138","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7489138","identity":"rs-7489138","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}