Ferroptosis-related gene signature predicts prognosis and immunotherapy response in hepatocellular carcinoma: A multi-cohort retrospective study

Abstract

Hepatocellular carcinoma (HCC), characterized by elevated incidence and mortality rates, has a considerable economic impact worldwide. The function of ferroptosis within the tumor microenvironment of HCC is crucial, and the specific contributions of ferroptosis-related genes (FRGs) are yet to be fully explored. FRG expression levels and relevant clinical data were sourced from The Cancer Genome Atlas. Two distinct ferroptosis-related subtypes were identified in liver cancer and their interrelationships were comprehensively examined. Using Cox regression and least absolute shrinkage and selection operator regression analysis, we created a predictive model based on FRGs to forecast overall survival and assess the potential benefits of immunotherapy in patients with HCC. Quantitative reverse transcription-polymerase chain reaction and IHC assays were conducted on clinical HCC specimens to validate the key FRGs. Significant differences in gene mutations, immune reactions, and prognostic outcomes were observed between the 2 distinct ferroptosis-related subtypes. An eight-gene signature consisting of SLC1A5, KIF20A, SLC7A11, CARS1, MYCN, PRDX6, GPX4, and KLF2 was established as a predictive model for liver cancer and was validated against data from GSE76427 and the International Cancer Genome Consortium cohorts. According to the FRG model, individuals classified as low-risk exhibited more favorable survival prospects compared to their high-risk counterparts (P < .05). Overall, our investigation highlights the promise of FRGs as prognostic biomarkers and immunotherapy response in HCC, providing a novel approach for personalized patient management.

1. Introduction

Hepatocellular carcinoma (HCC) is the predominant type of primary liver cancer and ranks third in terms of cancer-related mortality globally, claiming the lives of 7.6 million people in 2022.^[1] Projections suggest that by the year 2040, the incidence and mortality of liver cancer will exceed a 55% increase.^[2] Curative interventions for HCC, such as surgical resection and liver transplantation, are primarily applicable to patients diagnosed at early disease stages, while traditional systemic therapies (chemotherapy and radiotherapy) demonstrate limited efficacy and poor patient tolerance. Patients with unresectable HCC confront a grim prognosis, with median survival duration of approximately 13.6 months for those treated with lenvatinib and 12.3 months for sorafenib.^[3] This prognosis stems from profound tumor heterogeneity and complex pathogenesis driven by diverse molecular alterations. Research has shown that approximately a quarter of patients with HCC exhibit a significant mutational landscape, yet most cancer driver genes are either not effectively targeted or are undruggable.^[4,5] These challenges highlight the critical need to elucidate HCC’s molecular mechanisms and develop innovative therapeutic strategies.

Ferroptosis, an iron-dependent non-apoptotic form of cell death, is characterized by the lethal accumulation of lipid hydroperoxides.^[6,7] It emerges from a complex network comprising various elements of cellular metabolism such as iron, lipids, and redox reactions.^[8] Accumulating evidence has linked ferroptosis to the pathology of various cancers. Research by Dongbao Li et al^[9] reported that CST1 improves GPX4 stability and diminishes intracellular reactive oxygen species (ROS) by recruiting OTUB1 to inhibit ferroptosis and promoting metastatic behavior in gastric cancer. Zhang et al^[10] revealed that agents such as erastin and RSL3 enhanced the expression of NEAT1 by facilitating the interaction of p53 with the NEAT1 promoter, which in turn facilitated ferroptosis in HCC. Notably, PGAM1 has been shown to suppress tumor growth by augmenting ferroptosis and enhancing CD8+ T cell infiltration, exhibiting potential synergy with anti-programmed cell death protein 1 (PD-1) immunotherapy.^[11] These collective insights indicate that targeting cell ferroptosis pathways may become a promising therapeutic approach in oncology.

Immunotherapy has become a cornerstone in the treatment of numerous cancers. Inflammation typically accompanies ferroptotic stress. Recent studies have revealed that upon stimulation by immunotherapy, CD8+ T cells amplify lipid peroxidation, which subsequently induces ferroptosis of malignant cells.^[12] Moreover, Ping Yu et al^[13] highlighted the importance of PD-1 signaling in regulating phospholipid metabolism within CD8+ T cells, which bears therapeutic significance for immunotherapy. These findings collectively imply that ferroptosis can modulate tumor immunotherapy in a certain way. However, a detailed understanding of how ferroptosis interacts with immune responses, particularly in HCC, remains elusive.

Here, we initially identified ferroptosis-related molecular subtypes in The Cancer Genome Atlas (TCGA) cohort via unsupervised clustering and analyzed their enriched pathways, immune infiltration, and immunotherapy response. Then, we built a prognostic model for HCC that incorporated eight prognostic gene signatures identified through Cox and Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis. Simultaneously, we estimated the survival impact of immune checkpoint blockade (ICB) and chemotherapy using various datasets. A nomogram was also established, combining risk scores with various clinical parameters to enhance survival assessment. Finally, quantitative reverse transcription-polymerase chain reaction (qRT-PCR) and immunohistochemistry (IHC) assays were performed to validate key ferroptosis-related genes (FRGs) in the HCC specimens.

2. Materials and methods

2.1. Data acquisition

The liver hepatocellular carcinoma (LIHC) dataset, including Million mapped reads tumor-promoting immune macrophages expression profiles (374 tumor and 50 control samples) and clinical data (Table S1, Supplemental Digital Content, https://links.lww.com/MD/R183), was obtained from TCGA (https://portal.gdc.cancer.gov/) via the TCGAbiolinks package.^[14] We focused on 362 tumor samples (“01A”) with valid overall survival (OS) data, excluding para-cancerous tissues. Additionally, 50 control samples (11A) were included. Protein-coding genes were retained from the transcripts per million (TPM) expression data after removing low-expression genes, resulting in 16,587 genes.

Somatic mutation data for 371 LIHC cases (“Masked Somatic Mutation”) were analyzed using maftools, while copy number variation was assessed across 374 cases (“Masked Copy Number Segment”) with ggplot2. The GSE76427 dataset from GEO (https://www.ncbi.nlm.nih.gov/geo/) (115 tumor and 52 para-cancer samples) and the LIRI-JP dataset from International Cancer Genome Consortium (ICGC) (https://dcc.icgc.org/), consisting of 233 tumor samples, were also analyzed. Due to the lack of LIHC immunotherapy data, we referenced immunotherapy datasets from other cancers: clear cell renal cell carcinoma data (281 samples; 44 responders, 237 nonresponders) from PMID: 32472114,^[15] and additional immunotherapy data from PMID: 35444656.^[16]

Finally, 316 FRGs expressed in TCGA–LIHC were selected from the FerrDb database (see Table S2, Supplemental Digital Content, https://links.lww.com/MD/R183), excluding non-expressed genes.

2.2. Molecular subtype construction based on FRGs

We employed the Consensus ClusterPlus package of R (Bioconductor Core Team) to conduct consensus clustering on TCGA–LIHC expression profile data with FRGs to distinguish ferroptosis-related subtypes of HCC. This analysis involved 1000 iterations, where 80% of the total samples were randomly selected, and we explored cluster counts ranging from 2 to 8.

2.3. Gene-set enrichment analysis and gene set variation analysis (GSVA)

Gene-set enrichment analysis and GSVA were performed on the TCGA–LIHC expression dataset using “h.all.v7.5.1.symbols” gene set from the MSigDB database to identify biological process differences and molecular subtype variations related to ferroptosis, with significance set at adjp <.05.

2.4. Immune infiltration analysis

Immune-related genes were sourced from PMID: 28052254,^[17] comprising 782 genes and representing 28 immune cell subtypes (e.g., activated CD8+ T cells, macrophages, natural killer T cells, and regulatory T cells). Using the GSVA package of R,^[18] we applied single-sample gene set enrichment analysis to quantify immune cell infiltration levels in TCGA–LIHC transcriptomic data. Differential immune infiltration patterns across FRG clusters were visualized via ggplot2, with statistical significance assessed using Kruskal–Wallis tests (false discovery rate [FDR]-adjusted P < .05).

2.5. Analysis of differential genes related to ferroptosis

Differential analysis among FRGs molecular subtypes was performed using the Linear Models for Microarray Data (limma) package in R,^[19] based on TCGA grouping data. Ferroptosis-related differentially expressed genes (FRDEGs) were identified using a threshold of |log₂FC| >1 and FDR-adjusted P-value <.01 (Benjamini–Hochberg method). Genes with logFC >1 were classified as up-regulated and those with logFC <‐1 as down-regulated FRDEGs.

2.6. Function enrichment analysis

Gene ontology and Kyoto encyclopedia of genes and genomes pathway enrichment analyses were conducted on FRDEGs in TCGA–LIHC dataset using the R package clusterProfiler.^[20] Terms with a FDR <.05 were considered statistically significant. The screening criteria for enriched terms were an adjusted P-value <.05 and a q-value <.05, with P-value adjustment performed using the Benjamini–Hochberg method.

2.7. Immunotherapy analysis

Tumor immune dysfunction and exclusion (TIDE) website (http://tide.dfci.harvard.edu) was used to normalize TCGA–LIHC expression profile data to calculate TIDE scores and evaluate the immunotherapy response. Visualization of the results was facilitated through the ggplot2 package in R. A total of 20 cancer immunogenicity scores or immunophenoscores (IPS), which could predict CTLA-4 and PD-1 responsiveness, were retrieved from The Cancer Immunome Atlas (https://tcia.at/home) for liver cancer and analyzed using ggplot2 in R. To explore drug treatment differences among FRGs molecular subtypes, we used the pRRophetic package in R to retrieve gene mutation data for cell lines and IC50 values of different anticancer drugs from the Cancer Genome Project plan. Correlations between FRGs molecular subtypes and sensitivity to anticancer drugs were then analyzed.

2.8. Construction of protein–protein interaction (PPI) network and identification of hub genes

The STRING database (https://string-db.org/) encompasses 14,094 organisms, containing 67.59 million proteins and 20.05239 billion PPI. The database integrates evidence from experimental data, text mining of PubMed abstracts, consolidated results from other databases, and bioinformatic predictions. This STRING database was used to build a PPI network for FRGs, with a coefficient of 0.4 as the setting parameter. We visualized them with Cytoscape and the CytoHubba plug-in to assess hub genes within PPI network.

The gene set cancer analysis (GSCA) (https://guolab.wchscu.cn/GSCA/#/)^[21] is a platform designed for cancer analysis at genomic, pharmacogenomic, and immunogenomic levels. It integrates expression, mutation, drug sensitivity, and clinical data across 33 cancer types from 4 public data sources. We used this platform to analyze pan-cancer expression of hub genes and the ggplot function (R software) to draw bubble diagrams.

2.9. Risk model construction based on FRGs

We established a prognostic model for liver cancer patients using FRGs derived from the TCGA dataset. Key FRGs associated with OS were initially screened via univariate Cox regression (P < .05), followed by LASSO regression for variable selection. Multivariate Cox regression and stepwise regression were applied to identify optimal gene signatures and construct a risk score formula.

${\displaystyle \mathrm{r}\mathrm{i}\mathrm{s}\mathrm{k}\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\sum _i\mathrm{C}\mathrm{o}\mathrm{e}\mathrm{f}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{e}\mathrm{n}\mathrm{t}\left(\mathrm{g}\mathrm{e}\mathrm{n}{\mathrm{e}}_{\mathrm{i}}\right)*\mathrm{m}\mathrm{R}\mathrm{N}\mathrm{A}\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\left(\mathrm{g}\mathrm{e}\mathrm{n}{\mathrm{e}}_{\mathrm{i}}\right)}$

The optimal risk score cutoff was determined using the R surv_ cutoff function, categorizing TCGA–LIHC patients into high- and low-risk groups. Kaplan–Meier analysis demonstrated significant differences in survival outcomes between these risk groups, while time-dependent receiver operating characteristic (ROC) curves quantified predictive accuracy (AUC values) using the timeROC package. External validation using the GSE76427 and ICGC datasets confirmed model robustness, with risk distribution plots (ggplot2) and AUC analysis further verifying its stability. Drug information related to the model gene was retrieved using DGIdb (https://www.dgidb.org/), and the 3D structure of the potential therapeutic drug was visualized using PubChem (https://pubchem.ncbi.nlm.nih.gov/).

2.10. Construction of clinical prediction model based on FRGs model

This study developed a prognostic model by integrating risk scores with clinicopathological features (P < .05) through multivariate regression. A nomogram was constructed using the R rms package to predict HCC survival probabilities based on cumulative risk scores. The prognostic value of the combined risk stratification and clinical parameters was validated through Cox regression analysis of the TCGA–LIHC dataset. Model performance was evaluated using calibration curves and decision curve analysis, demonstrating alignment between predicted and observed survival outcomes.

2.11. Quantitative reverse transcription-polymerase chain reaction

Between July 2021 and November 2022, 30 HCC tissues and paired normal liver specimens were obtained from patients with HCC at The First Affiliated Hospital of Guangxi Medical University, with ethical approval granted by the hospital’s Ethics Committee. The baseline clinicopathological characteristics of this cohort are detailed in Table 1. Total RNA extraction was performed utilizing TRIzol reagent (Invitrogen), followed by cDNA synthesis using PrimeScript RT Master Mix (Takara) in accordance with the manufacturer’s instructions. qRT-PCR was carried out using Green Mix SYBR (Promega) to detect the mRNA expression levels of key FRGs, using GAPDH as the internal control. The primers utilized were detailed in Table S3, Supplemental Digital Content, https://links.lww.com/MD/R183 and relative quantification was computed by the 2^‐ΔΔCt method. Expression levels were derived from triplicate technical replicates (mean ± SD; n = 30 samples).

Table 1.

Baseline clinical and pathological characteristics of the HCC patient cohort (n = 30).

Characteristic	Number of patients (n = 30)	Percentage (%)
Age (yr), mean ± SD (range)	57.5 ± 8.5 (36–74)	–
Gender
Male	26	86.7
Female	4	13.3
Tumor size (cm)
≤5	17	56.7
>5	13	43.3
AJCC stage
I	14	46.7
II	11	36.7
III	5	16.6
IV	0	0.0
Histologic grade
G1	0	0.0
G2	8	26.7
G3	21	70.0
G4	1	3.3
Cirrhosis background
Yes	20	66.7
No	10	33.3
HBsAg (+)
Yes	18	60.0
No	12	40.0
Vascular invasion
Yes	13	43.3
No	17	56.7

AJCC = American Joint Committee on Cancer, HCC = hepatocellular carcinoma, SD = standard deviation.

2.12. Immunohistochemistry

The tissue samples were deparaffinized with xylene and hydrated with alcohol. The slides were then placed in sodium citrate buffer at 80 °C for antigen retrieval and allowed to cool to 24 °C. After rinsing thrice with PBS, 3% H₂O₂ was used to block endogenous peroxidase for 15 minutes at 24 °C. The sections were incubated overnight at 4 °C with primary antibodies: SLC1A5 (1:500, 20350-1-AP, Proteintech), KIF20A (1:200, 15911-1-AP, Proteintech), SLC7A11 (1:200, 26864-1-AP, Proteintech), CARS1 (1:500, 15296-1-AP, Proteintech), MYCN (1:250, 10159-2-AP, Proteintech), PRDX6 (1:100, 13585-1-AP, Proteintech), GPX4 (1:2000, 67763-1-Ig, Proteintech), and KLF2 (1:200, bs-2772R, Bioss), and then with the secondary antibody for 1 hour at 24 °C. Protein expression was quantified via ImageJ (integrated optical density; 5 fields/section) across n = 30 samples (mean ± SD).

2.13. Statistical analysis

All statistical analyses were conducted using R software (v4.2.1). Continuous variables were assessed via Student t test (normal distribution) or Mann–Whitney U test (non-normal), while categorical variables were analyzed using Chi-square/Fisher exact tests. Survival outcomes were evaluated with Kaplan–Meier curves (survival package) and log-rank tests. Univariate/multivariate Cox regression and LASSO analysis (glmnet package) identified prognostic factors. Two-tailed P-values < .05 were considered statistically significant.

3. Results

3.1. Project overall analysis flow chart

A flow diagram of the study is presented in Figure S1, Supplemental Digital Content, https://links.lww.com/MD/R182.

3.2. Effect of FRGs on liver cancer in TCGA–LIHC dataset

We systematically characterized FRG alterations in LIHC through multi-omics analyses. Single-nucleotide mutation analysis identified the top 20 mutated FRGs (Fig. 1A), revealing an overall mutation rate of 84.91% (351 samples), with TP53 showing the highest frequency (28%). Chromosomal mapping via RCircos (Fig. 1B) localized these genes’ genomic positions. Copy number variation analysis demonstrated frequent amplifications over deletions (Fig. 1C), particularly in tumor suppressors like ANXA13, ATAD2, and CFH (Fig. 1D). Transcriptomic validation revealed significant differential expression (P < .05) in 80% of FRGs between tumor and normal tissues (Fig. 1E and F): YY1AP1, PRDX6, and CYB5R1 were markedly upregulated (P < .001), whereas ATF3 and CTSB showed tumor-specific downgulation (P < .05).

Figure 1.

Effect of FRGs on liver cancer in TCGA–LIHC dataset. (A) Mutation profile of FRGs in TCGA–LIHC patients; (B) chromosomal location map of FRGs; (C) copy number change frequency of FRGs in TCGA–LIHC (orange: amplification; blue: deletion); (D) copy number change frequency of the tumor suppressor gene in TCGA–LIHC (blue: altered; orange: unaltered); (E) box plot of difference expression of FRGs in normal and LIHC samples (normal: blue; LIHC: orange); (F) heat map of FRGs expression in LIHC versus normal group (orange: high; blue: low). *P < .05; **P < .01; ***P < .001; ****P < .0001; ns = no statistical significance; LIHC = liver hepatocellular carcinoma, TCGA = The Cancer Genome Atlas.

3.3. Molecular subtype construction of FRGs in TCGA–LIHC dataset

We stratified LIHC patients through unsupervised consensus clustering of 316 FRGs. Optimal clustering (k = 2) divided samples into 2 distinct subtypes: cluster 1 (n = 146) and cluster 2 (n = 216) (Fig. 2A). PCA confirmed clear intergroup separation (Fig. 2B). Survival analysis demonstrated significantly superior prognosis in cluster 2 (log-rank P = .00027, Fig. 2C), with Sankey diagrams (Fig. 2D) revealing predominant stage I to II distribution in this cluster. Differential expression analysis identified 30 FRGs showing marked inter-subtype variation (heatmap, Fig. 2E; full gene list in Table S4, Supplemental Digital Content, https://links.lww.com/MD/R183). These molecular subtypes exhibit strong correlations with clinical outcomes and disease progression.

Figure 2.

Molecular subtype construction of FRGs in TCGA–LIHC dataset. (A) Consensus clustering (k = 2): color intensity indicates co-clustering probability (0–1); (B) principal component analysis (blue circles: cluster 1; pink triangles: cluster 2); (C) survival curve between subtypes (blue: cluster 1; pink: cluster 2); (D) Sankey diagram of survival status, FRGs molecular subtype and stage; (E) FRGs expression heat map between subtypes; *P < .05. LIHC = liver hepatocellular carcinoma, TCGA = The Cancer Genome Atlas.

3.4. Immune analysis of molecular subtypes of FRGs in TCGA–LIHC dataset

We systematically delineated the functional divergence between FRG subtypes in LIHC through multidimensional analyses. Pathway enrichment (Table S5, Supplemental Digital Content, https://links.lww.com/MD/R183) revealed cluster 1 exhibited significant activation of adverse pathways: glycolysis (Normalized enrichment score (NES) = 1.553, P < .001), angiogenesis (NES = 1.967, P < .001), and apoptosis (NES = 1.412, P < .001) (Fig. 3A–C). Hallmark analysis (GSVA) (Table S6, Supplemental Digital Content, https://links.lww.com/MD/R183) further identified cluster 1 enrichment in mitotic spindle and TGF-β signaling, while cluster 2 showed oxidative phosphorylation and adipogenesis activation (Fig. 3D).

Figure 3.

Immunological analysis of molecular subtypes of FRGs in the TCGA–LIHC dataset. (A–C) Pathway analysis of FRGs molecular subtypes. These panels illustrate 3 key pathways related to FRGs molecular subtypes: glycolysis (A), angiogenesis (B), and apoptosis (C); (D) heat map of hallmark gene set enrichment results across different subtypes; (E) immune cell infiltration boxplot (blue: cluster 1; pink: cluster 2); (F–G) molecular feature variation: MHC (F), T cell stimulators (G); cluster 1 (blue), cluster 2 (pink). *P < .05; **P < .01; ***P < .001; ****P < .0001. ns = no statistical significance, TCGA = The Cancer Genome Atlas.

In order to evaluate the influence of different subtypes of FRGs on the different levels of immune cell infiltration in LIHC patients, the single-sample gene set enrichment analysis single-sample gene set enrichment analysis algorithm was used to quantify the relative abundance of each immune cell infiltration, marking each infiltrating immune cell type (Table S7, Supplemental Digital Content, https://links.lww.com/MD/R183), the difference box plot was drawn by the ggplot function. Immune profiling demonstrated subtype-specific infiltration patterns: >60% of immune cell types showed significant abundance differences (P < .05), with cluster 1 displaying elevated MHC expression (Fig. 3E and F) and T-cell stimulators compared to cluster 2 (Fig. 3G).

3.5. Functional enrichment analysis of FRDEGs in TCGA–LIHC dataset

To examine the biological variances between the FRGs subtypes in LIHC, we carried out differential expression analysis utilizing the limma package in R. Through this analysis, we identified 1531 significant ferroptosis related differentially expressed genes (FRDEGs, Table S8, Supplemental Digital Content, https://links.lww.com/MD/R183). Following this, we performed gene ontology analysis on the 1531 FRDEGs (Fig. S2A–C, Supplemental Digital Content, https://links.lww.com/MD/R182; Table S9, Supplemental Digital Content, https://links.lww.com/MD/R183) and Kyoto encyclopedia of genes and genomes functional enrichment analysis (Fig S2D, Supplemental Digital Content, https://links.lww.com/MD/R182; Table S10, Supplemental Digital Content, https://links.lww.com/MD/R183).

3.6. Immunotherapy analysis of molecular subtypes of FRGs in TCGA–LIHC dataset

To assess the clinical relevance of FRGs subtypes in HCC, we evaluated their association with immunotherapy response using the TIDE algorithm (Table S11, Supplemental Digital Content, https://links.lww.com/MD/R184). Cluster 2 (favorable prognosis) exhibited significantly lower TIDE scores than cluster 1 (poor prognosis) (P < .0001, Fig. 4A), indicating superior ICB response potential. Stratification by TIDE threshold (score = 0) revealed cluster 2 predominance in low-TIDE groups (79.9%), whereas cluster 1 dominated high-TIDE groups (52.2%, Fig. 4B). Violin plots visualizing immune checkpoint gene expression across FRG molecular subtypes revealed significant differential expression (Fig. 4C–G). Multiple checkpoints showed pronounced upregulation in cluster 1, including BTN2A1 and CD276 (both P < .001). Using the expression matrix and drug treatment information in the Cancer Genome Project plan built into the pRRophetic package of R, the response of TCGA–LIHC patients to chemotherapeutic drugs and small molecule anticancer drugs was predicted according to the cell line response IC50 values of different drugs (Table S12, Supplemental Digital Content, https://links.lww.com/MD/R184). The results showed substantial differences between FRG molecular subtypes for 251 anticancer agents. Cluster 2 exhibited significantly lower IC50 values than cluster 1 (P < .001; Fig. 4H–L), indicating enhanced sensitivity to targeted therapies including erlotinib and trametinib.

Figure 4.

Immunotherapy analysis of molecular subtypes of FRGs in the TCGA–LIHC dataset. (A) Boxplot of TIDE score differences between FRG subtypes (blue: cluster 1; pink: cluster 2); (B) bar graph showing proportions of FRG subtypes within TIDE groups; (C–G) violin plots for BTN2A1 (C), CD276 (D), CD80 (E), PDCD1 (F), and CD27 (G) depicting the variations in immune checkpoint gene expression across different FRG molecular subtypes; (H–L) erlotinib (H), trametinib (I), cetuximab (J), talazoparib (K), and TGX221 (L) show the violin plots of the IC50 values of chemotherapeutic agents across various FRGs molecular subtypes. LIHC = liver hepatocellular carcinoma, TCGA = The Cancer Genome Atlas, TIDE = tumor immune dysfunction and exclusion.

3.7. Construction of PPI network and identification of hub genes

We constructed the FRG interaction network using STRING, revealing 297 nodes and 3002 edges (Table S13, Supplemental Digital Content, https://links.lww.com/MD/R184). In the generated network graph (Fig. S3A, Supplemental Digital Content, https://links.lww.com/MD/R182), the node size was proportional to the connectivity within the network. Cytoscape visualization identified 30 hub genes via CytoHubba (Fig. S3B, Supplemental Digital Content, https://links.lww.com/MD/R182; Table S14, Supplemental Digital Content, https://links.lww.com/MD/R184). Intersection with TCGA–LIHC differential expression data (1531 FRDEGs: 1352 up/179 down; Fig. S3C, Supplemental Digital Content, https://links.lww.com/MD/R182) yielded 4 key HCC-associated FRGs (HIF1A, SRC, EZH2, androgen receptor [AR]; Fig. S3D, Supplemental Digital Content, https://links.lww.com/MD/R182). Pan-cancer analysis across 14 malignancies showed SRC/EZH2 tumor upregulation and AR/HIF1A downregulation (Fig. S3E, Supplemental Digital Content, https://links.lww.com/MD/R182; Table S15, Supplemental Digital Content, https://links.lww.com/MD/R184).

3.8. Expression analysis and verification of hub genes

Hub gene expression analysis revealed significant upregulation in LIHC versus normal tissues across both TCGA and GSE76427 datasets (P < .05; Fig. 5A and B). HIF1A and EZH2 exhibited the most pronounced elevation, prompting their selection with AR and SRC for further investigation. Subsequent evaluation across FRG molecular subtypes demonstrated marked differential expression (P < .0001; Fig. 5C), where SRC and EZH2 showed significantly higher levels in poor-prognosis cluster 1. To assess the prognostic impact of hub genes in LIHC, we conducted univariate Cox regression analysis. Results visualized via forest plot (Fig. 5D) identified HIF1A, SRC, and EZH2 as significant risk factors and AR as a protective factor for LIHC survival. Copy number alteration analysis revealed predominant EZH2 amplification versus other hub genes, with SRC exclusively showing amplification patterns (Fig. 5E–H). Survival stratification by median expression demonstrated superior outcomes for HIF1A and EZH2 low-expression groups (HIF1A: P = .01, Fig. 5I; EZH2: P = .0002, Fig. 5J). Multivariate Cox regression confirmed EZH2 as an independent prognostic factor (P < .001; Fig. 5K).

Figure 5.

Expression analysis and verification of hub genes. (A and B) Boxplots showing hub gene differences between LIHC and normal groups in TCGA (A) and GSE121248 (B) datasets; (C) boxplot of hub gene differences among FRG subtypes in TCGA dataset; (D) univariate Cox forest plot: genes with log2HR > 0 are risk factors; <0 are protective; (E–H) Pie charts of copy number alteration proportions for HIF1A (E), SRC (F), EZH2 (G), and AR (H) in TCGA–LIHC (orange: amplification; blue: deletion; gray: no change); (I and J) survival curves for HIF1A and EZH2 by expression level (yellow: low-risk; purple: high-risk); (K) multivariate Cox forest plot following the same log2HR interpretation as (D). *P < .05; **P < .01; ***P < .001; ****P < .0001. AR = androgen receptor, LIHC = liver hepatocellular carcinoma, TCGA = The Cancer Genome Atlas.

3.9. Construction and verification of FRGs model in TCGA–LIHC dataset

We developed a prognostic risk model for LIHC patients using 316 FRGs. Univariate Cox regression of TCGA–LIHC data identified 135 significant FRGs (P < .05), with top associations shown in Figure 6A, with details in Table S16, Supplemental Digital Content, https://links.lww.com/MD/R184. Subsequent LASSO regression refined these to 13 prognostic candidates (Fig. 6B). Multivariate Cox analysis with stepwise selection established an 8-gene signature: SLC1A5, KIF20A, SLC7A11, CARS1, MYCN, PRDX6, GPX4, and KLF2 (Fig. 6C). Model parameter visualization revealed positive coefficient correlations for all genes except KLF2 (Fig. 6D). The risk score formula was derived as:

Figure 6.

FRGs model construction in TCGA–LIHC dataset. (A–C) Construction of the Cox-LASSO-multivariate Cox model: (A) univariate Cox forest plot; (B) LASSO regression cross-validation for lambda selection; (C) multivariate Cox forest plot; (D) model gene parameter lollipop diagram. Lollipop’s size correlates with the absolute value of the model coefficient; (E) C-index bar chart of model and model gene; (F) Sankey plot of FRGs molecular subtype, risk model and survival status; (G–I) The survival curves of TCGA (G), GSE76427 (H) and ICGC (I) datasets (yellow: low-risk; purple: high-risk). ICGC = International Cancer Genome Consortium, LASSO = least absolute shrinkage and selection operator, LIHC = liver hepatocellular carcinoma.

${\displaystyle \begin{array}{cc}@l@l@l@l@l@l@l@l@l@l@l@l@l@l@l@l@l@l@l@l@& {\displaystyle }\mathrm{r}\mathrm{i}\mathrm{s}\mathrm{k}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\left(0.1*\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{o}\mathrm{f}\mathrm{S}\mathrm{L}\mathrm{C}1\mathrm{A}5\right)+\left(0.22*\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{o}\mathrm{f}\mathrm{K}\mathrm{I}\mathrm{F}20\mathrm{A}\right)\mathrm{\backslash vspace}\mathrm{0.5mm}\\ & {\displaystyle }+\left(0.19*\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{o}\mathrm{f}\mathrm{S}\mathrm{L}\mathrm{C}7\mathrm{A}11\right)+\left(0.26*\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{o}\mathrm{f}\mathrm{C}\mathrm{A}\mathrm{R}\mathrm{S}1\right)\mathrm{\backslash vspace}\mathrm{0.5mm}\\ & {\displaystyle }+\left(0.18*\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{o}\mathrm{f}\mathrm{M}\mathrm{Y}\mathrm{C}\mathrm{N}\right)+\left(0.38*\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{o}\mathrm{f}\mathrm{P}\mathrm{R}\mathrm{D}\mathrm{X}6\right)\mathrm{\backslash vspace}\mathrm{0.5mm}\\ & {\displaystyle }+\left(0.27*\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{o}\mathrm{f}\mathrm{G}\mathrm{P}\mathrm{X}4\right)+(-0.32*\mathrm{E}\mathrm{x}\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{o}\mathrm{f}\mathrm{K}\mathrm{L}\mathrm{F}2)\mathrm{\backslash vspace}\mathrm{0.5mm}\end{array}}$

Our model showed superior predictive performance with a higher Cox regression C-index than single-gene models (Fig. 6E). A Sankey diagram (Fig. 6F) illustrated that most deaths occurred in the high-risk group, while low-risk cases were mainly linked to better prognosis clusters and survival. Poor prognosis cluster 1 mostly comprised high-risk samples.

To evaluate prognostic discrimination, patients in TCGA–LIHC were split by a risk score cutoff of 7.460741 into high- and low-risk groups, with the low-risk group showing improved survival (P < .0001, Fig. 6G). This method was also applied to 2 independent cohorts, GSE76427 and ICGC, using cutoffs of 11.41843 and 6.271458, respectively. Both validation sets showed consistent results, with low-risk patients exhibiting better survival rates compared to their high-risk counterparts (GSE76427: P = .029, Fig. 6H; ICGC: P = .0023, Fig. 6I).

3.10. FRGs model verification in TCGA–LIHC dataset

Analysis of risk score distribution and survival in TCGA (Fig. 7A), GSE76427 (Fig. 7B), and ICGC (Fig. 7C) showed higher mortality in high-risk groups. Time-dependent ROC curves (Fig. 7D–F) demonstrated the risk score’s strong predictive power for OS across all datasets. TCGA AUCs at 1, 2, and 3 years were 0.814 (95% CI: 0.754–0.873), 0.788 (95% CI: 0.727–0.849), and 0.755 (95% CI: 0.708–0.842); GSE76427 values were 0.666 (95% CI: 0.502–0.833), 0.729 (95% CI: 0.592–0.867), and 0.788 (95% CI: 0.656–0.921); and ICGC values were 0.635 (95% CI: 0.504–0.766), 0.635 (95% CI: 0.531–0.739), and 0.664 (95% CI: 0.553–0.775), respectively.

Figure 7.

FRGs model validation in TCGA–LIHC dataset. (A–C) Visualization of risk score distribution, patient survival, and gene expression in TCGA (A), GSE76427 (B), and ICGC (C) datasets. Orange: low expression; purple: high expression. (D–F) Validation of the prognostic accuracy of the FRGs signature across independent cohorts. ROC curves illustrate the performance of the signature in predicting overall survival at 1, 2, and 3 years in the TCGA (D), GSE76427 (E), and ICGC (F) cohorts. The orange, blue, and red lines represent predictions for 1-, 2-, and 3-year survival, respectively. AUC = area under curve, ICGC = International Cancer Genome Consortium, LIHC = liver hepatocellular carcinoma, ROC = receiver operating characteristic, TCGA = The Cancer Genome Atlas.

3.11. Immunotherapy analysis of FRGs model in TCGA–LIHC dataset

To evaluate the risk score model’s predictive capacity for immunotherapy, we compared immune checkpoint gene expression between high-risk and low-risk groups using boxplots (Fig. 8A). Most genes, such as PDCD1 (P < .001) and CTLA4 (P < .0001), exhibited significant upregulation within the high-risk cohort. We then analyzed LIHC-related IPS from The Cancer Immunome Atlas (Table S17, Supplemental Digital Content, https://links.lww.com/MD/R184) and found all 4 IPS categories were considerably elevated in the low-risk group (P < .01, Fig. 8B). Microsatellite instability (MSI) was compared between risk groups using a boxplot, showing higher MSI in the high-risk group (P = .015, Fig. 8C). Using an MSI cutoff of 0.3295, patients were stratified, with the low-MSI group showing better survival (P = .0002, Fig. 8D). The TIDE algorithm was subsequently employed to evaluate immunotherapy response in TCGA-LIHC (Table S11, Supplemental Digital Content, https://links.lww.com/MD/R184), revealing significantly elevated TIDE scores in the high-risk group (P < .0001, Fig. 8E). Tumor mutation burden (TMB) was computed using maftools R package, showing lower TMB in the low-risk cohort (P < .001, Fig. 8F). Using TMB cutoff of 1.918386, patients were stratified, with the low-TMB group having better survival (P = .00073, Fig. 8G). Finally, applying a risk score cutoff of 6.473342 to an immunotherapy cohort (PMID_32472114) also distinguished groups with significant survival differences favoring the low-risk group (P = .0012, Fig. 8H).

Figure 8.

Immunotherapy analysis of FRGs model in TCGA–LIHC dataset. (A and B) Boxplots of immune checkpoint gene expression (A) and IPS scores (B) by risk group (purple: high risk; yellow: low risk); (C–F) boxplots of MSI (C), TIDE (E), and TMB (F) scores between risk groups (yellow: low risk; purple: high risk); (D and G) survival curves based on MSI (D) and TMB (G) groups (pink: high; blue: low); (H) Survival curve for PMID_32472114 data (pink: high risk; blue: low risk). *P < .05; **P < .01; ***P < .001; ****P < .0001. IPS = immunophenoscore; LIHC = liver hepatocellular carcinoma; MSI = microsatellite instability; TCGA = The Cancer Genome Atlas; TIDE = tumor immune dysfunction and exclusion; TMB = tumor mutation burden.

3.12. Discrimination and drug treatment of FRGs molecular subtypes by FRGs model in TCGA–LIHC dataset

To assess the risk score’s ability to differentiate liver cancer FRG subtypes, boxplots showed a significant differentiation among clusters (P < .001), with cluster 2 (favorable prognosis) having a lower risk score than cluster 1 (poor prognosis) (Fig. S4A, Supplemental Digital Content, https://links.lww.com/MD/R182). Distribution analysis revealed cluster 2 was predominant in the low-risk cohort (75.4% vs 24.6%), while cluster 1 was more common in the high-risk population (78.3% vs 24.6%) (Fig. S4B, Supplemental Digital Content, https://links.lww.com/MD/R182). The ROC curve for risk score distinguishing FRG subtypes showed an AUC of 0.790 (Fig. S4C, Supplemental Digital Content, https://links.lww.com/MD/R182).

We assessed risk score’s impact on drug responsiveness using pRRophetic R package to estimate IC50 values in TCGA-LIHC. Significant differential responses were observed between risk groups (Fig S4D–K, Supplemental Digital Content, https://links.lww.com/MD/R182), with the high-risk group showing substantially elevated IC50 values for saracatinib (Fig S4F, Supplemental Digital Content, https://links.lww.com/MD/R182, P < .05) and trametinib (Fig S4G, Supplemental Digital Content, https://links.lww.com/MD/R182, P < .001). Drug–gene interaction screening via DGIdb (Table S18, Supplemental Digital Content, https://links.lww.com/MD/R184) yielded targeted small-molecule agents for SLC1A5, SLC7A11, and MYCN: glutamine, riluzole, tretinoin, dinutuximab, panobinostat, vincristine, cyclophosphamide, birabresib, cisplatin, etoposide, and sonidegib. PubChem-derived 3D structures of these compounds are shown in Fig. S4L–N, Supplemental Digital Content, https://links.lww.com/MD/R182.

3.13. Construction of clinical prediction model based on risk score in TCGA–LIHC dataset

Univariate Cox analysis showed that risk score (P < .001), T stage (P < .001), S stage (P < .001), and M stage (P = .0193) exhibited significant correlations with OS (Fig S5A, Supplemental Digital Content, https://links.lww.com/MD/R182, Table 2). Multivariate analysis confirmed both risk score and T stage functioned as independent prognostic indicators (P < .001, Fig. S5B, Supplemental Digital Content, https://links.lww.com/MD/R182, Table 3). A nomogram was devised to integrate risk score, age, and T stage for the purpose of predicting OS (Fig. S5C, Supplemental Digital Content, https://links.lww.com/MD/R182). Decision curve analysis demonstrated its clinical utility (Fig. S5D, Supplemental Digital Content, https://links.lww.com/MD/R182), and calibration curves showed a strong concordance between the predicted and actual 1-, 3-, and 5-year OS (Fig. S5E–G, Supplemental Digital Content, https://links.lww.com/MD/R182). ROC curves further confirmed the nomogram’s superior prognostic accuracy compared to individual factors (Fig. S5H–J, Supplemental Digital Content, https://links.lww.com/MD/R182).

Table 2.

Univariate Cox regression results of ferroptosis related scores on clinical prognosis.

	Hazard ratio (HR)	Lower 95% CI	Upper 95% CI	P-value
riskScore	2.74	2.19	3.41	4.21E‐19
TNM_T	1.43	0.903	2.27	1.74E‐07
Stage	1.42	0.867	2.31	1.02E‐05
TNM_M	3.98	1.25	12.7	.0193
Age	1.01	0.998	1.03	.0821
Gender	0.834	0.583	1.19	.322
TNM_N	1.98	0.485	8.08	.341

TNM_M = tumor node metastasis_metastasis, TNM_N = tumor node metastasis_ node, TNM_T = tumor node metastasis_tumor.

Table 3.

Multivariate Cox regression results of ferroptosis related score on clinical prognosis

	Hazard ratio (HR)	Lower 95%CI	Upper 95% CI	P-value
riskScore	2.71	2.15	3.41	<.001
TNM_T T2	1.04	0.65	1.67	.856
TNM_T T3	2.11	1.38	3.23	.001
TNM_T T4	4.23	2.1	8.51	<.001

TNM_T = tumor node metastasis_tumor.

3.14. Verification of key genes

To confirm the expression levels of key FRGs, qRT-PCR was performed on 30 HCC clinical samples. Compared to the para-tumor samples, significantly upregulated expression in HCC were observed for SLC7A11 (Fig. 9A) and KIF20A (Fig. 9D) (P < .05), whereas CARS1 (Fig. 9C), KLF2 (Fig. 9E), and PRDX6 (Fig. 9G) demonstrated significantly reduced expression in HCC samples (P < .05). For GPX4 (Fig. 9B), MYCN (Fig. 9F), and SLC1A5 (Fig. 9H), no statistically significant differences were noted between HCC and para-tumor samples (P > .05).

Figure 9.

Clinical validation of FRGs mRNA expression. (A–L) The mRNA expression levels of key genes HIF1A (A), SRC (B), EZH2 (C), AR (D), SLC1A5 (E), KIF20A (F), SLC7A11 (G), CARS1 (H), MYCN (I), PRDX6 (J), GPX4 (K), and KLF2 (L) were further validated using 30 clinical HCC samples *P < .05; **P < .01; ***P < .001; ****P < .0001. AR = androgen receptor; ns: no statistical significance.

For protein levels, upregulated expression in HCC were observed for SLC1A5 (Fig. 10A, I), CARS1 (Fig. 10D, L), MYCN (Fig. 10E, M), and KLF2 (Fig. 10H, P) (P < .05), whereas SLC7A11 (Fig. 10C, K) displayed a significant reduction in expression in HCC samples (P < .05). No significant differences in protein expression were identified between the HCC and para-tumor samples for KIF20A (Fig. 10B, J), PRDX6 (Fig. 10F, N), and GPX4 (Fig. 10G, O) (P > .05).

Figure 10.

Clinical validations of FRGs protein expression. (A–H) Representative images of immunohistochemical staining for SLC1A5 (A), KIF20A (B), SLC7A11 (C), CARS1 (D), MYCN (E), PRDX6 (F), GPX4 (G), and KLF2 (H) in tumor and para-tumor tissues. Scale bars: 50 μm (main), 25 μm (insets). (I–P) Mean optical density (MOD) of FRGs proteins in tumor and para-tumor tissues, including SLC1A5 (I), KIF20A (J), SLC7A11 (K), CARS1 (L), MYCN (M), PRDX6 (N), GPX4 (O), and KLF2 (P). *P < .05; **P < .01; ****P < .0001. ns = no statistical significance.

4. Discussion

HCC exhibits profound molecular and microenvironmental heterogeneity that complicates therapeutic interventions. Mounting evidence indicates that modulating ferroptosis status during immune infiltration could substantially amplify the benefits of immunotherapy for malignancies, including glioma,^[22] breast cancer,^[23] and HCC.^[24,25] Conche et al^[26] revealed that ferroptosis activates an adaptive immune infiltration reaction of myeloid-derived suppressor cells rather than serving as a direct tumor suppressor in HCC. Hao et al^[27] reported that APOC1 inhibition could facilitate the conversion of M2 macrophages to the M1 phenotype via the activation of the ferroptosis pathway. Tang et al^[28] further explored the impact of ferroptosis on tumor-associated macrophages, which are pivotal in regulating the tumor microenvironment (TME). These insights collectively underscore the intricate interactions between ferroptosis pathways and immune responses within HCC. Despite these promising findings, its 20% to 30% response rate in HCC^[29] highlights the unmet need for biomarkers predicting ferroptosis-modulated immune responses.

Initial FRG-based stratification of HCC patients revealed 2 distinct subtypes with significant prognostic differences. Patients in cluster 2 exhibited superior survival outcomes, a finding further emphasized by Sankey diagram visualization showing their predominance in early-stage (I/II) disease. Genomic profiling supported the clinical relevance of these subtypes, demonstrating frequent mutations, copy number alterations, and differential expression of FRGs in tumors compared to normal tissues, collectively underscoring their prognostic impact in HCC. These findings align with recent analyses of ferroptosis patterns, where Zhang et al^[16] similarly identified 2 ferroptosis clusters (A/B) in colon cancer with distinct immune microenvironments and prognostic outcomes, further validating the universal role of ferroptosis subtypes in shaping tumor biology. Yoo et al^[30] demonstrated that the overexpression of the SLC1A5 variant promotes the production of ATP and the synthesis of glutathione triggered by glutamine, resulting in resistance to gemcitabine in pancreatic cancer cells. Research on genetically engineered mice revealed that the deletion of the xC-subunit of SLC7A11 triggered tumor-selective ferroptosis and effectively suppressed growth of pancreatic cancer cells.^[31] Above analysis underscores the fact that mutations or altered expression of FRGs could influence susceptibility to ferroptosis, thereby impacting cancer progression and response to therapy.

Tumor immune microenvironment analysis provides insights into prognostic outcomes. Cluster 1 also demonstrated significantly lower MHC expression, T cell stimulators, and higher tumor immune dysfunction scores than cluster 2. These findings implicate variations in immune infiltration that affect HCC prognosis, with cluster 1 being potentially less responsive to immunotherapy. Analyses of immune checkpoint genes and drug sensitivity values support this finding. Most immune checkpoint genes (BTN2A1, CD276, CD80, PDCD1, and CD27) and IC50 values for anticancer drugs (erlotinib, trametinib, cetuximab, talazoparib, and TGX221) differed substantially between the clusters, with cluster 2 demonstrating improved outcomes. These results suggest that immune checkpoint activation protects tumors and is inversely correlated with HCC prognosis. Currently, ICB therapies show potent efficacy in a subgroup of patients with diverse malignancies, including HCC. A meta-analysis by Liu et al^[32] revealed that the combination of atezolizumab (PD-L1) and bevacizumab potentially offers enhanced efficacy and improved safety compared to lenvatinib in treating unresectable HCC. Overall, FRGs profiling identified distinct HCC subtypes with implications for prognosis, tumor genetics, and immunotherapy responses through alterations in the immune microenvironment.

We identified 4 key FRGs (SRC, HIF1A, EZH2, and AR) within the PPI network, showing significant expression differences across HCC subtypes. Survival analysis indicated that HIF1A, SRC, and EZH2 act as risk factors, whereas AR serves a protective role. Clinical qRT-PCR confirmed decreased AR expression in HCC, consistent with its reported inhibitory effect on tumor invasion.^[33] Notably, HIF1A is implicated in immune evasion via modulation of immune checkpoints,^[34] and EZH2 expression correlates with poor prognosis,^[35,36] highlighting its potential as a target for therapeutic intervention. These findings underscore critical role of FRGs in HCC progression and prognosis.

We developed a prognostic model utilizing 8 key FRGs to evaluate their impact on HCC outcomes. This model effectively stratified patients into high- and low-risk categories, with the low-risk group and cluster 2 showing significantly improved survival. Validation across TCGA, GSE76427, and ICGC cohorts confirmed the model’s robust predictive power, as high-risk patients exhibited shorter median OS and greater mortality risk. Extending beyond prognosis, the FRGs model was closely associated with critical immunological markers (including immune checkpoints, IPS, MSI, TIDE, and TMB) highlighting its potential to guide personalized immunotherapy strategies. Notably, patients with lower MSI, TIDE scores, and TMB had better survival outcomes. Supporting these findings, previous studies report that ferroptosis modulation enhances cancer immunotherapy efficacy.^[37] Similarly, SLC7A11-deficient tumors show increased immunotherapy sensitivity in vivo,^[38] collectively suggesting an important role for SLC7A11 in modulating the response to immunotherapy. Furthermore, our experimental results show that the mRNA and protein expression levels of SLC7A11 are inconsistent. This discordance suggests that posttranscriptional regulation^[39] plays a significant role in shaping the ferroptosis–immune axis in HCC, adding another layer of biological complexity to our prognostic model. Beyond SLC7A11, KLF2 is recognized as a significant tumor suppressor due to its frequently low expression across various cancer types. Research has demonstrated that KLF2 can induce ferroptosis by inhibiting the PI3K/AKT signaling pathway,^[40] suggesting a potentially crucial role for KLF2 in the context of HCC. Notably, we observed elevated KLF2 protein expression coupled with reduced mRNA levels in HCC tissues, implying important posttranslational regulation that potentiates its tumor-suppressive activity. In contrast to the tumor-suppressive role of KLF2, GPX4 represents another critical ferroptosis regulator that often exerts oncogenic functions by protecting cancer cells from lipid peroxidation-driven death. Emerging evidence shows GPX4-driven ferroptosis resistance facilitates HCC metastasis via the GRHL3/PTEN/PI3K/AKT axis^[41] and GPX4 activates innate immunity via the cGAS–STING pathway.^[42] For SLC1A5, evidence from a pan-cancer analysis positions SLC1A5 as a key regulator that influences both ferroptosis susceptibility and the composition of the TME,^[43] highlighting its important role in cancer biology. KIF20A, another oncogene in our signature, exhibits transcriptional upregulation with stable protein levels in HCC, consistent with its reported role as a predictive biomarker for PD-1 inhibitor response.^[44] This discordant mRNA–protein pattern suggests posttranslational regulation may sustain KIF20A’s oncogenic function, proposing its targeting as a strategy to augment immunotherapy efficacy. Similarly, PRDX6 shows reduced mRNA but maintained protein expression in our signature. Ito J et al^[45] reported that PRDX6 governs ferroptosis sensitivity by regulating selenium allocation to GPX4 and other selenoproteins, with its loss exacerbating lipid peroxidation and cell death. Our 8-FRG signature unifies ferroptosis susceptibility and immune evasion in HCC by integrating multigene regulatory layers into a prognostic and therapeutic framework.

Several recent studies have also explored ferroptosis-related signatures in HCC. For example, a study demonstrated that a prognostic model incorporating 4 ferroptosis-related genes (G6PD, SLC7A11, MYCN, and KIF20A) (constructed and validated using TCGA and ICGC datasets) effectively stratified HCC patients into high- and low-risk groups, with the risk score serving as an independent prognostic factor.^[46] Hu et al^[47] developed a prognostic model using ten ferroptosis-related genes (CAPG, FLT3, G6PD, HAVCR1, HMOX1, IL33, MT3, SLC7A11, SRXN1, and STMN1), which effectively stratified HCC patients into high- and low-risk groups with distinct clinical outcomes. Similarly, Wu et al^[48] constructed a novel 5-gene FRGs that effectively predicts prognosis and immunotherapy response in HCC, with high-risk patients showing stronger immune suppression and poorer survival. While these studies established the prognostic value of ferroptosis-related signatures, our study extends these efforts by: employing multi-cohort validation (TCGA, ICGC, and GEO) to enhance generalizability; providing experimental validation of key FRGs at mRNA and protein levels in clinical samples, revealing clinically relevant posttranscriptional disparities; identifying significant associations between the FRG signature and immunotherapy biomarkers (TIDE, IPS, TMB).

To assess the clinical applicability of our results, we analyzed the correlation between the risk scores derived from FRGs model and sensitivity of HCC to various chemotherapeutic agents. Our analysis indicated significant disparities in the IC50 values of the targeted drugs between patient subtypes, with lower risk scores correlating with a better drug response. This insight prompted the creation of a nomogram prognostic model that integrates the FRG-derived risk score with some clinical indicators, exhibiting superior discriminative ability in evaluating the OS status of patients with LIHC compared with conventional single-factor approaches.

However, there are several limitations that warrant discussion. First, cohort-specific risk stratification cutoffs reflect inherent heterogeneity across sample sources, sequencing platforms, and patient demographics, necessitating unified thresholds derived from large-scale prospective HCC studies. Second, while TIDE/IPS analyses and cross-cancer datasets provided mechanistic hypotheses about immune evasion, disease-specific biological differences limit their direct applicability to HCC, underscoring the need for dedicated HCC immunotherapy cohorts with annotated outcomes. Third, the direct application of colon/ccRCC immunotherapy data to HCC is limited by fundamental differences in TME and treatment response patterns. Most importantly, our 30-sample validation cohort, though informative for initial correlations, constrains statistical power, and the absence of functional validation experiments (in vitro and in vivo) precludes definitive mechanistic conclusions about the roles of identified FRGs.

5. Conclusions

Overall, our investigation highlights the promise of FRGs as prognostic biomarkers and immunotherapy response in HCC, thereby providing a novel approach for personalized patient management. While our findings suggest potential applications in immunotherapy response prediction, further validation in dedicated immunotherapy cohorts is needed to establish their clinical utility in this context.

Acknowledgments

We thank the Laboratory of Guangxi Zhuang Autonomous Region Engineering Research Center for Artificial Intelligence Analysis of Multimodal Tumor Images, Key Laboratory of Ultrasonic Molecular Imaging and Artificial Intelligence, Guangxi Key Laboratory of Early Prevention and Treatment for Regional High Frequency Tumor/Key Laboratory of Early Prevention and Treatment for Regional High Frequency Tumor (Guangxi Medical University), Ministry of Education for their support of this study.

Author contributions

Conceptualization: Hong Yang.

Data curation: Lipeng Li, Jinshu Pang, Yunjing Pan.

Formal analysis: Jinshu Pang, Jiamin Chen, Yunjing Pan.

Funding acquisition: Yun He, Hong Yang.

Investigation: Lipeng Li, Jinshu Pang, Hong Yang.

Methodology: Lipeng Li.

Project administration: Jinshu Pang, Yun He.

Resources: Jiamin Chen.

Software: Jiamin Chen.

Supervision: Yun He, Hong Yang.

Validation: Li Liang, Xiumei Bai.

Visualization: Jiamin Chen, Yunjing Pan.

Writing – original draft: Lipeng Li.

Writing – review & editing: Li Liang, Xiumei Bai, Hong Yang.