The genes related to ferroptosis were downloaded from a previously published public data center (www.zhounan.org/ferrdb), and noncoding genes were removed from the dataset. The website also divided ferroptosis-related genes into three groups as follows: suppressors, drivers, and markers. We further analyzed the genes closely related to this research.

We obtained the mRNA-sequencing data and clinical information of HCC patients from the liver hepatocellular carcinoma (LIHC) cohort in The Cancer Genome Atlas (TCGA) database (https://www.cancer.gov/), which included 19,676 annotated mRNA sequences and other clinical data from 370 tumor tissue samples and 50 normal tissue samples. After matching mRNA sequences with ferroptosis-related genes, the differentially expressed genes (DEGs) were selected by limma, an R package, with numerical conditions log2-fold change (FC)  >  1 and an adjusted p-value < 0.05.

We used the ConsensusClusterPlus package in R software for consistent clustering. The Euclidean squared distance metric and the K-means clustering algorithm were used to classify the TCGA-LIHC cohort into k clusters, with k = 2 to k = 9. The optimal number of clusters was determined by the consistent cumulative distribution function (CDF) graph and the delta region graph (Wilkerson and Hayes, 2010).

To predict the 1-, 3-, and 5-year survival probability of HCC patients, we combined all independent prognostic factors to construct a nomogram. Calibration curves were generated to assess the consistency between predicted survival rates and actual survival rates.

The Kaplan-Meier (K-M) curve was a visualized tool for comparing the overall survival (OS) and progression-free survival (PFS) in different ferroptosis molecular subgroups, with log-rank tests to compare the curves. The receiver operating characteristic (ROC) curve, which was built by the R package “survivalROC,” was used to evaluate model prognostic performance by calculating the area under the ROC curve (AUC).

Human hepatocellular carcinoma cell lines, including SK-HEP1 and HCC-LM3, were obtained from the American Type Culture Collection (ATCC) (Manassas, VA, United States). The cells were cultured in DMEM (containing 10% fetal bovine serum and 100 U/ml penicillin–streptomycin) and placed in a 37°C, 5% CO₂ incubator.

We transfected FTL shRNAs into SK-HEP1 and HCC-LM3 cells through Lipofectamine 2000 (Invitrogen, CA, United States), which was synthesized by GeneChem (Shanghai, China). After cultivation in basic DMEM, the cells were cultured in DMEM supplemented with FBS and penicillin–streptomycin.

The total protein of the two liver cancer cell lines after FTL shRNA transfection was extracted using RIPA lysis buffer (Invitrogen) containing PMSF (Bio–Rad, Shanghai, China), and the protein concentration was determined and quantified. The total protein was treated with 10% sodium dodecyl sulfate–polyacrylamide gel electrophoresis (SDS–PAGE) and transferred to a polyvinylidene fluoride membrane (PVDF) (Invitrogen, Carlsbad, United States). After transfer, the membranes were blocked at room temperature for 2 h, then the primary antibody was added and incubated at 4°C overnight, and the secondary antibody was added and incubated at room temperature for 2 h. Finally, the iBright FL1500 intelligent imaging system (Invitrogen, Carlsbad, United States) was used to analyze the absorbance of the protein bands and calculate the relative protein expression level. The antibodies used in the study are listed in the Supplementary Material.

SK-HEP1 and HCC-LM3 cells were digested and counted, seeded in 96-well plates (3,000 cells/plate in 200 µl DMEM), and cultured in a 37°C, 5% CO₂ incubator. After 0, 24, 48, and 72 h, we washed the culture medium off the cells to be tested, added CCK8 solution according to the instructions, and continued to culture the cells for 2 h. The absorbance value (OD) of each group was measured at 450 nm on a microplate reader and recorded for statistical analysis. We also used the 5-ethynyl-2′-deoxyuridine (EdU) reagent (Ruibo, Guangzhou, China) and a viability/cytotoxicity kit (Invitrogen, Carlsbad, United States) according to the manufacturer’s protocol.

SK-HEP1 and HCC-LM3 cells were trypsinized for 24 h, rinsed with PBS three times (3 min/time), fixed with 4% paraformaldehyde for 15 min at room temperature, and then stabilized in 0.2% Triton for 10 min to rupture the cell membrane. Nonspecific antigen-binding sites were blocked with 2% BSA for 30 min, and then the cells were incubated with anti-PCNA overnight at 4°C. After washing, the cells were incubated with the anti-rabbit antibody for 60 min, and the nuclei were stained with DAPI for 2 min and then washed with PBS. Finally, a fluorescence microscope was used to observe and photograph the cells, and the expression levels of PCNA were detected.

The t-test was used for the measurement data, the χ² test was used for the enumeration data, and the Kaplan-Meier method and log-rank test were used for the survival analysis. For all statistical calculations, the final results were determined to be statistically significant at p < 0.05.

To identify prognostic ferroptosis regulators in HCC, we first conducted differential gene expression analysis between tumor tissues and normal tissues in the TCGA-LIHC cohort, and the expression characteristics of the DEGs are shown in Figures 1A,B and Supplementary Table S1. Then, we performed Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis and Gene Ontology (GO) analysis to explore the signaling pathways, biological processes, cellular components, and molecular functions of the DEGs. We confirmed “ferroptosis” as a mainly enriched signaling pathway (Figures 1C,D) and obtained a total of 239 ferroptosis-related DEGs that were enriched in the ferroptosis pathway. Subsequently, we drew a Venn diagram to identify the role of ferroptosis-related DEGs as markers, suppressors, and drivers in ferroptosis (Figure 1E). The gene function, interaction, and prognostic role among the ferroptosis-related DEGs are shown in Figure 1F. Using univariate Cox regression, we finally identified 41 prognostic ferroptosis regulators from markers, suppressors, and drivers of ferroptosis (Figure 1G).

Unsupervised consensus clustering based on the expression characteristics of the 41 prognostic ferroptosis regulators was performed to establish ferroptosis-related subtypes in HCC. The results revealed that the optimal number of clusters was two (k value = 2) (Figures 2A–C). Hence, the patients in the TCGA-LIHC cohort were divided into two clusters, namely, fescluster A and fescluster B. The survival analysis indicated that compared with patients in fescluster B, patients in fescluster A showed a significant survival advantage (p < 0.001) (Figure 2D). We further explored the differences in clinical features between the two clusters, and the results revealed that patients in fescluster B had a higher proportion of death, age <60 years old, vascular invasion occurrence, and pathological grades G3/G4 than patients in fescluster A, and the patients in both fescluster A and fescluster B were predominantly male (Figure 2E). The results indicated that the clinical features of patients in fescluster B were worse than those in fescluster A. In recent years, tumor mutation burden (TMB), which represents the characteristics of the total number of somatic coding mutations in tumors, has been increasingly shown to have predictive value in tumor prognosis as a potential biomarker for non-small-cell lung cancer, liver cancer, and other cancer types (Rizvi et al., 2015; Tang et al., 2021). Based on this, we then confirmed the characteristics of TMB in fescluster A and fescluster B (Figures 2F,G) and explored the correlation between TMB and fesclusters (Figure 2H). The survival analysis shows that there was no statistically significant difference in survival probability between the high TMB group and the low TMB group (p > 0.05) (Figure 2I). We further divided the patients into fescluster A with low TMB, fescluster A with high TMB, fescluster B with low TMB, and fescluster B with high TMB, and survival analysis indicated that patients in fescluster A with low TMB showed a significant survival advantage compared with the other groups, while patients in fescluster B with high TMB had the worst prognosis (Figure 2J). These results revealed that the established ferroptosis-related subtypes in HCC showed effective prognostic predictive performance in HCC and were closely associated with the clinical characteristics of patients.

Next, random forest analysis was used to analyze critical ferroptosis-related genes in HCC, and the results indicated that MT3, NRAS, STMN1, FTL, and SLC1A5 were the five most important ferroptosis regulators in HCC associated with prognosis (Figure 3A). According to previous publications, there have been a lot of research on the function of three genes NRAS (Dietrich et al., 2019; Ding et al., 2020), STMN1 (Zhang et al., 2020), and SLC1A5 (Zhao et al., 2021) in HCC, and in this study, our pre-experiment indicated the expression of MT3 didn’t affect the proliferation of HCC, so we focused on the role of FTL in HCC. With the FTL expression level at 75% by quartile as the criterion, we divided the high FTL group and the low FTL group in the TCGA-LIHC cohort. The K-M survival curve and ROC curve showed that higher FTL expression predicted poor survival with superior reliability in HCC patients (Figures 3B,C). Figures 3D–H indicates that the expression level of FTL is closely associated with clinical features (p < 0.05), including sex, history grade, and TNM stage, but not age, in HCC. Interestingly, the predictive value of FTL was also closely related to clinical features, including sex, history grade, and TNM stage (p < 0.05), but not age in HCC. High expression levels of FTL showed poor survival time in the HCC patients with male sex, TNM stage I-II, and G3-G4 grade (p < 0.05), however, in HCC patients with female sex, TNM stage III-IV, and G1-G2 history grade, there was no sign of FTL expression in the prognosis of HCC (Figure 3I-L).

To verify the expression level of FTL in normal tissue and tumor tissue in HCC patients, we obtained three different HCC cohorts, including the TCGA-LIHC cohort, ICGC HCC cohort, and GSE14520 cohort. The results indicated that FTL was obviously higher in HCC tissue than in normal tissue (Figures 4A,E,I). Then, calibration curves, ROC curves, and decision curve analysis (DCA) were used to validate the diagnostic ability of FTL in HCC patients. The AUC values of the calibration curve and ROC curve were 0.716, 0.769, and 0.694, respectively, in these three cohorts (Figures 4B,C, F–G, J–K), indicating good predictive performance of this diagnostic model. The DCA curve of our diagnostic model showed some net benefit for prediction (Figures 4D,H,L). Overall, these results indicated that FTL was more highly expressed in HCC tissue and that the diagnostic model showed good predictive performance for distinguishing between HCC samples and normal samples.

To further clarify the immune microenvironment associated with the FTL expression level in HCC, a single sample gene set enrichment analysis (ssGSEA) was used to identify immune cell infiltration and immune activity. The results showed that higher expression indicated more immune cell infiltration, including activated CD8⁺ T cells and Gamma delta T cells (Figure 5A). In addition, the enrichment scores of cytolytic activity, CCR, HLA, and immune checkpoint were significantly increased in high FTL HCC patients (Figure 5B). Figure 5C shows the relationship between FTL expression and immune checkpoints, including PDCD1, CTLA4, TIGIT, and CD83; these checkpoints were expressed at higher levels in the high FTL group than in the low FTL group (Figures 5D–G). GSEA indicated that high FTL expression was also positively associated with immune activation-related signaling pathways, including the IL2-STAT5 signaling pathway, IL6-JAK-STAT3 signaling pathway, and interferon-gamma response signaling pathway (Figures 5H–J). This evidence suggests that FTL expression is positively associated with the immune-activation microenvironment in HCC.

To predict the chemotherapeutic and targeted therapeutic responses, half the maximum inhibitory concentration (IC50) of 266 anticancer drugs was obtained from the Genomics of Drug Sensitivity in Cancer (GDSC) website. The results showed that traditional chemotherapeutic drugs, including cisplatin, paclitaxel, and vinorelbine, had lower IC50 values in the high FTL group than in the low FTL group in HCC (p < 0.05) (Figures 6A–C). In addition, HCC patients with higher FTL expression showed lower IC50 values of targeted drugs such as sorafenib, dasatinib, imatinib, roscovitine, sunitinib, thapsigargin, tipifarnib, temsirolimus, and rapamycin (p < 0.05) (Figure 6D–L). These results indicate that the expression level of FTL is related to the sensitivity of some chemotherapeutic drugs and targeted drug treatments.

To evaluate the independent performance of the FTL level in predicting prognosis compared with other clinical features, including age, AFP, weight, vascular invasion, sex, history grade, and TNM stage, univariate and multivariate Cox regression were used. The results showed that age, TNM stage, and FTL level were independent predictive factors in HCC patients (Figure 7A). Then, a nomogram predictive model was constructed based on these three independent predictive factors to quantify the survival probability of HCC patients at 1, 3, and 5 years (Figure 7B). The calibration curve of the nomogram for predicting the overall survival probability of HCC patients at 1, 3, and 5 years was close to the 45° line (Figures 7C–E). DCA was used to evaluate the guiding significance of these independent predictive factors for predicting overall survival time at 1, 3, and 5 years, and the results show that the nomogram has superior predictive value in clinical practice (Figures 7F–H). These results showed that FTL is an independent prognostic factor and that our nomogram has valuable predictive performance.

To confirm the role of FTL in the tumor progression of HCC, HCC cells including SK-HEP1 and HCC-LM3 were transfected with two different FTL shRNA and scramble shRNA. Western blot indicated the FTL shRNA effectively inhibited FTL expression in SK-HEP1 and HCC-LM3 cells (Figures 8A,B). Then CCK8 kit and Edu assay were used to assess the role of FTL on HCC-cell proliferation. The result showed FTL inhibition obviously suppressed the proliferation of SK-HEP1 and HCC-LM3 cells (Figures 8C–F). To further explore the role of FTL in HCC-cell proliferation, immunofluorescence was used to measure the expression of PCNA in SK-HEP1 and HCC-LM3 cells. The result indicated FTL inhibition effectively decreased PCNA expression levels in HCC cells (Figures 8G,H). Moreover, silencing FTL expression induced lipid peroxidation levels of SK-HEP1 and HCC-LM3 cells (Figure 8I-L) and the concentration of iron inSK-HEP1 and HCC-LM3 cells was significantly increased with FTL inhibition (Figure 8M–N). These results indicated FTL inhibition effectively suppressed HCC-cell proliferation and triggered ferroptosis in HCC cells.