Ferroptosis and radiotherapy-related genes in head and neck squamous cell carcinoma: diagnostic and prognostic significance

Abstract

Background

Head and Neck Squamous Cell Carcinoma (HNSCC) is an aggressive malignancy with poor prognosis and low five-year survival rates. Ferroptosis, an iron-dependent form of cell death, has emerged as a potential factor in cancer therapy. This study aimed to identify differentially expressed genes (DEGs) associated with HNSCC, ferroptosis, and radiotherapy, and to develop a diagnostic model to establish novel biomarkers, early diagnosis and individualized therapy.

Methods

We analyzed TCGA-HNSCC, GSE6631, and GSE107591 datasets to identify Ferroptosis- and Radiotherapy-Related Differentially Expressed Genes (F&RRDEGs).Ferroptosis & radiotherapy scores were calculated, followed by immune infiltration and immunotherapy analyses. A Cox prognostic model was evaluated using nomograms, calibration curves, and decision curve analysis. Real-time quantitative PCR (RT-qPCR) and Western blotting (WB) were used for experimental validation.

Results

11 F&RRDEGs were identified classifying two subtypes related to HNSCC.Significant differences were found in immune cell infiltration and immunophenoscore in CTLA4 and PD1/PD-L1/PD-L2 + CTLA4 between high and low ferroptosis & radiotherapy scores(F&Rs) groups. The 1-year calibration curve of the prognostic risk model showed better and more accurate performance and performance for the five-year prediction. RT-qPCR identified significant differences in the genes ALOX12, EMP1, PIM1, and CA9 between the control and HNSCC groups, with WB confirming radiotherapy-induced upregulation of VDAC1/TFRC (pro-ferroptosis) and downregulation of GPX4/SLC7A11/FTH1 (anti-ferroptosis) in irradiated HNSCC cells.

Conclusion

ALOX12, EMP1, PIM1, and CA9 exhibited promising potential as diagnostic biomarkers. This study applies immune infiltration analysis to HNSCC research, offering fresh perspectives on patient-specific treatment plans, and accelerating the creation of focused therapies.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-03713-7.

Introduction

Sixth among frequent cancers, head and neck cancer ranks sixth among the most common malignancies, with a 5-year survival rate of less than 70%, making it a major public health concern [1, 2]. Current treatments for HNSCC include, several treatments, such as immunotherapy, targeted therapy, chemotherapy, radiotherapy, and surgical resection [3]. Despite advances in HNSCC treatment the 5-year survival rate has improved minimally in the last few decades. This lack of treatment improvement is largely attributed to location-specific recurrence [4]. Therefore, to support individualized therapy and enhance clinical outcomes, it is imperative to identify high risk HNSCC patients and create cutting-edge treatments.

Iron is an essential nutrient, crucial in various biological processes. It catalyzes reactions in numerous proteins, facilitating processes such as hemoglobin and DNA synthesis, oxygen transport, and immune defense, among others [5]. However, excess unbound iron generates free radicals via oxygen interactions, causing lipid peroxidation, reactive oxygen species (ROS) elevation, oxidative stress, and DNA damage [6, 7]. Ferroptosis—distinct from apoptosis or necrosis—results from iron-dependent peroxidation of polyunsaturated fatty acid-rich phospholipids [8, 9]. Its pivotal role in cancer therapy is increasingly recognized [10, 11].

While iron metabolism dysregulation in Oral squamous cell carcinoma (OSCC) is increasingly recognized, elevated transferrin receptor and suppressed ferroportin expression promote iron accumulation, enabling ferroptosis susceptibility [12, 13]. This metabolic rewiring creates a therapeutically exploitable vulnerability. Yet, integrating immunotherapy with conventional modalities, like radiotherapy, in a reasonable and efficient manner is challenging in oncology [14]. Emerging evidence confirms radiotherapy induces ferroptosis in HNSCC through ROS-dependent lipid peroxidation [15]. Radiation-triggered mitochondrial damage synergizes with FIN56 to amplify ferroptotic cell death in vivo. Jordon et al. found that compared to radiotherapy alone, the combination of the radio enhancer NBTXR3 significantly increased lipid peroxidation and cancer cell killing capacity, which might be achieved by promoting ferroptosis [16]. Ferroptosis mediates the tumor-suppressive effects of interferon-gamma (IFN-γ)—produced by CD8⁺ T cells during immune checkpoint blockade—suggesting its role as an immune mechanism against tumor development [17, 18]. By causing ferroptosis, drugs such as erastin suppresse the expression of Glutathione Peroxidase 4 (GPX4) and increase the radiosensitivity of breast, cervical, and Non-Small Cell Lung Cancer to X-rays [19]. Nevertheless, the prognostic utility of radiotherapy-ferroptosis crosstalk genes in HNSCC remains unexplored, hindering clinical translation of this synergistic mechanism.

Based on this premise, we analyzed datasets related to HNSCC from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) databases. Differential and enrichment analyses were performed to pinpoint genes and pathways linked to both ferroptosis and radiotherapy. Using a consensus clustering method based on F&RRDEGs, we identified two molecular subtypes of HNSCC with distinct gene expression profiles and clinical characteristics. We conducted an immune infiltration evaluation to investigate the differences in immune cell characteristics and associated genes between the HNSCC and control groups and between the high and low F&Rs groups (the high F&Rs group and the low F&Rs group refer to the division of all samples into high-score group and low-score group based on the median of the F&Rs calculated by single-sample gene set enrichment analysis (ssGSEA)). In addition, we conducted a more thorough evaluation of the prognostic risk model’s capacity to infer the outcomes of HNSCC patients.

Materials and methods

Data acquisition and preprocessing

The TCGA-HNSCC dataset was acquired using the TCGAbiolinks software. This dataset served as the trial set for our analysis. After excluding samples lacking clinical prognostic data, a total of 434 samples with both count and FPKM sequencing data were retained, comprising 414 HNSCC tumor samples and 20 normal controls. For validation we also obtained microarray data from the GEO database using the R tool GEOquery. Specifically the GSE6631 dataset included 22HNSCC and 22 normal samples [20, 21].

The dataset was produced on the Affymetrix Human Genome U95 Version 2 Array, specifically designated as GPL8300. Furthermore, we utilized the GSE107591 dataset which encompasses 24 HNSCC samples and 23 control samples. For comprehensive details, please see Table 1.

Table 1.

HNSCC data set information list

	TCGA-HNSCC	GSE6631	GSE107591
Platform	/	GPL8300	GPL6244
Species	Homo sapiens	Homo sapiens	Homo sapiens
Tissue	human head and neck	human head and neck	human head and neck
Samples in the HNSCC group	414	22	24
Samples in the Control group	20	22	23
Reference	/	Selection and validation of differentially expressed genes in head and neck cancer.	The oncogenic role of circPVT1 in head and neck squamous cell carcinoma is mediated through the mutant p53/YAP/TEAD transcription-competent complex.

We collected data on human genes from the GeneCards database by searching for “Ferroptosis” and “Radiotherapy” to identify Ferroptosis-Related Genes (FRGs) and Radiotherapy-Related Genes (RRGs). Next, we applied the selection criterion of “Protein Coding” to filter the findings, resulting in 738 FRGs and 2291 RRGs.

Screen ferroptosis & radiotherapy-related differential genes

In order to discover F&RRDEGs, we originally employed the R package limma to perform differential analysis on the Counts data obtained from the TCGA-HNSCC dataset, as well as the GSE6631 and GSE107591 datasets. We identified Differentially Expressed Genes (DEGs) for our investigation by selecting genes that met the criterion of absolute log fold change (|logFC|) greater than 0.5 and adjusted p-value (adj.P.Val ) less than 0.05. Genes with a logFC more than 0.5 and an adj.P.Val less than 0.05 were categorized as up-regulated DEGs, while genes with a logFC less than − 0.5 and an adj.P.Val less than 0.05 were categorized as down-regulated DEGs. We then created a volcano map to graphically represent the differential expression study’s findings. To further show the overlap of DEGs and FRGs & RRGs in the three datasets, we also made a Venn diagram. This allowed us to identify the genes that overlap, which we refer to as F&RRDEGs. Furthermore, we generated an expression heatmap for F&RRDEGs.

Expression difference analysis of F&RRDEGs

Mann-Whitney U test was performed to assess the expression disparities of F&RRDEGs between HNSCC and control samples across the TCGA-HNSCC, GSE6631, and GSE107591 datasets. We utilized the R tool ggplot2 to generate group comparison plots.

Subsequently, diagnostic performance was analyzed using the receiver operating characteristic (ROC) curve on F&RRDEGs with consistent expression levels across the three datasets. Using the R proc package, we assessed the relationship between the survival rate of HNSCC patients and the diagnostic performance of F&RRDEGs.

Mutation analysis and copy number variation

We employed the VarScan software for preprocessing the mutation data linked to F&RRDEGs observed in HNSCC patients, from the TCGA-HNSCC dataset. To visualize the mutation data, we employed the maftools package. Copy number variation (CNV) data which refers to genomic alterations that result in abnormal copy numbers of one or more genes, was obtained using the TCGAbiolinks package in R, (“Masked Copy Number Segment”) from the TCGA database. This data was acquired for the purpose of investigating the CNV of F&RRDEGs in HNSCC patients within the TCGA-HNSCC dataset. We also used the R ggplot2 tool to provide a visual examination of the CNV.

Gene set enrichment analysis of differentially expressed genes

Gene Set Enrichment Analysis (GSEA) measures the distribution trend of a preset gene set in a gene table arranged according to the gene’s association with the phenotype, allowing one to ascertain the gene’s contribution to the phenotype. Based on the logFC determined from cancer and normal groups, all differentially examined genes in the TCGA-HNSCC dataset were first sorted in this analysis. Subsequently, all genes underwent enrichment analysis with the cluster Profiler software. The parameters are as follows:

• Seed = 2022.

• Calculation number = 1000.

• Minimum gene count per gene set = 10.

• Maximum gene count per gene set = 500.

• Correcting p-values with the Benjamini-Hochberg method.

We obtained the “c2.cp.all.v2022.1.Hs.symbols.gmt [All Canonical Pathways] (3050)” gene set, which belonged to the Homo sapiens species. An adj.P.Val of less than 0.05 and a false discovery rate( value of less than 0.05 were used to identify significant enrichment.

Functional enrichment analysis using gene ontology

During the investigation of F&RRDEGs in the TCGA-HNSCC dataset, we performed gene ontology (GO) enrichment analysis using the same inclusion screening criteria and statistical significance limits as indicated earlier. Researchers employed the Benjamin-Hochberg method to correct for p-values.

Consistency cluster analysis

Consensus clustering is a widely used method for discovering novel disease subtypes or comparing multiple subtypes. We utilized the R package ConsensusClusterPlus for this work to do consensus clustering analysis. The analysis focused on the F&RRDEGs extracted from the TCGA-HNSCC dataset, aiming to classify samples within the HNSCC group into distinct subgroups, facilitating the construction of cancer group subtypes. The parameters were set as follows: 100 repetitions (reps = 100), 80% resampled samples (pItem = 0.8), 100% resampled features (pFeature = 1), and the Partitioning Around Medoids clustering algorithm. In addition, we used principal component analysis (PCA) to figure out how different the discovered subtypes were. Subsequently, a grouping comparison plot was generated to visualize the expression variations of the F&RRDEGs across the different subtypes.

Gene set variation analysis for pathway enrichment assessment

Gene Set Variation Analysis (GSVA) transforms the gene expression matrix from one sample to another into a quantitative matrix representing the expression of gene sets across samples. This analytic method is non-parametric and unsupervised. It is used to assess gene enrichment outcomes in the microarray data’s nuclear transcriptome. This technique evaluates if distinct pathways are enriched in various samples. In this inquiry, we utilized the “h.all.v7.4.symbols.gmt” reference gene set from the Molecular Signatures Database for GSVA on the gene expression matrix of the HNSCC group in the TCGA-HNSCC dataset. We analyzed the enrichment pathway results from the GSVA using the limma package only among the consistent clustering groups (cluster1/cluster2) of the HNSCC group in the TCGA-HNSCC dataset. Pathways with adjusted p. <0.05 were selected for heat map and group comparison chart visualization.

Assessment of ferroptosis and radiotherapy scores via SsGSEA

Single-sample GSEA (ssGSEA) technique determines the relative abundance of each gene in a dataset sample, was used to calculate F&Rs scores per sample using the R package GSVA and the ssGSEA algorithm, depending on the F&RRDEGs expression matrix. Samples were then grouped according to the F&Rs median (High, Low). Clinically relevant stacked histograms were created to visualize the high as well as low F&Rs groups combined with TCGA-HNSCC and clinical data. Additionally, we created a group comparison chart to illustrate the expression distinctions of F&RRDEGs between these two groups within the TCGA-HNSCC dataset.

Immune infiltration analysis using SsGSEA

The relative abundance of each immune cell infiltration, such as activated CD8 T cells, active dendritic cells, macrophages, natural killer T cells, regulatory T cells, and other human immune cell subtypes, was measured using the ssGSEA algorithm.

In addition, we calculated the enrichment score to evaluate the degree of different immune cell infiltration in each specimen. We utilized the F&Rs’ high and low groups obtained in previous analyses for subsequent analysis. Box plots compared the disparity in infiltration density among 28 distinct categories of immune-related cells of HNSCC samples. The correlation between the High as well as Low groups was calculated using Pearson’s correlation coefficient and visualized using ggplot2. Moreover, we assessed the association between immune cells and F&RRDEGs across different groups by integrating the gene expression matrix of TCGA HNSCC patients. Subsequently, we visualized these results using ggplot2 to create a correlation dot plot.

Analysis of immunogenicity and immunotherapy response

Immunogenicity denotes the capacity of an antigen or its epitope to trigger T cells and B cells, eliciting humoral and cell-mediated immune responses. This process is influenced by a multitude of genes related to effector cells, Major Histocompatibility Complex molecules, immunomodulators, and immunosuppressive cells. Machine learning techniques could be utilized to estimate and quantify immunogenicity. The Cancer Immunome Atlas (TCIA) database provides the immunophenoscore (IPS) for 20 different cancers, which has been shown to effectively forecast the reactivity of CTLA-4 and PD-1. In this study, we downloaded the IPS data for HNSCC from the TCIA database and visualized it using the ggplot2 package in R.

Clinical prognosis model of HNSCC using F&Rs and Cox regression

We integrated F&Rs, F&RRDEGs, and clinical factors (such as age, gender, and stage) from the TCGA-HNSCC dataset and used a single-factor Cox regression analysis to assess how well F&Rs could predict HNSCC. To build a prognostic model, variables with P-values less than 0.1 were chosen for a subsequent multi-factor analysis. Diagnostic plots were generated by multi-factor analysis, utilizing the R package “rms”.

To assess the multi-factor Cox prognostic model’s precision and accuracy, calibration curves were created. Calibration graphs assess the model’s predictive performance by comparing actual probabilities with predicted probabilities in different scenarios.

We also performed decision curve analysis (DCA) to assess the model’s accuracy and resolution. DCA is employed to evaluate molecular markers, diagnostic assays, and clinical prediction models.

To show DCA curves and assess the multi-factor Cox prognostic model’s prediction ability, we employed the R package ‘ggDCA’. These studies aimed to determine the clinical prognostic value of F&Rs in HNSCC and offer a thorough assessment of the multi-factor Cox prognostic model’s prediction accuracy and resolution.

Validation of gene and protein expression in HOK and SCC25 cell lines

Human oral keratinocytes (HOK) and the HNSCC cell line SCC25 were generously provided by Professor Yongcong Cai (University of Electronic Science and Technology of China), and exempt from ethical review. They were maintained in high-glucose Dulbecco’s modified Eagle’s medium (DMEM, GIBCO, USA) supplemented with 10% fetal bovine serum (FBS, GIBCO, USA). All cells were incubated at 37 °C in a humidified atmosphere containing 5% CO2. SCC25 cells were irradiated with the CellRad (Faxitron Bioptics, US) device at a dose rate of 2 Gy/min and a total of 4 Gy [22], and cells were collected 3 days after irradiation (RT-SCC25).

The HOK, SCC25, and RT-SCC25 cells were collected and treated with radio-immunoprecipitation assay (RIPA) lysis buffer containing protease inhibitors. Cell samples were quantified by a bicinchoninic acid (BCA) kit (Beyotime). Then, the total proteins were separated by a sodium dodecyl sulfate (SDS)-polyacrylamide gel and transferred to polyvinylidene difluoride (PVDF) membranes (Millipore, Billerica, MA, USA). After sealing with 5% non-fat milk in Tris-buffered saline Tween (TBST), the membranes were incubated with specific primary antibodies overnight at 4 ℃. The primary antibodies used were as follows: VDAC1 (R26067, Zenbio, CN), TFRC (381603, Zenbio), GPX4 (R24461, Zenbio), SLC7A11 (382036, Zenbio), FTH1 (R23306, Zenbio), PIM1 (ET1609-57, Huabio, CN), ALOX12 (R389216, Zenbio), EMP1 (AP17400a, Abcepta, CN), CA9 (ET1701-51, Huabio), NF-κB p65 (ET1603-12, Huabio), p-NFκBp65 (YP0191, Immunoway, CN), IKBα (ET1603-6, Huabio), p-IKBα(HA722770, Huabio). On the next day, the horseradish peroxidase-conjugated secondary antibody anti-rabbit IgG (HA1001, Huabio) was incubated for 2 h at room temperature. Finally, the targeted strips were visualized using an electrochemiluminescence (ECL) solution, and β-actin was used as an internal control.

Total RNA was extracted from cells using TRIzol reagent (Invitrogen). First-strand cDNA was synthesized with the RevertAid First Strand cDNA Synthesis Kit (Thermo Fisher Scientific) under the following conditions: 42 °C for 60 min and 70 °C for 5 min. The relative mRNA expression of hub genes was quantified by RT-qPCR on a QuantStudio 3 Real-Time PCR System (Applied Biosystems) using SYBR Green chemistry. All primers, listed in Supplementary Table S1, were synthesized by Beijing Biotechnology Corporation. Each 20 µL reaction mixture contained 10 µL of 2× SYBR Green Master Mix, 0.5 µL of each forward and reverse primer (10 µM), and 2 µL of cDNA template. The thermal cycling profile consisted of an initial denaturation at 95 °C for 10 min, followed by 40 cycles at 95 °C for 15 s and 60 °C for 60 s. Gene expression was normalized to the endogenous control GAPDH, and all reactions were performed in triplicate.

Statistical analysis

The data processing and analysis in this study were done using R software (Version 4.2.2). Mann-Whitney U tests examined the association between the fluctuations in non-normally distributed variables while comparing two sets of continuous data. We used Spearman correlation analysis to ascertain the correlation coefficient between several variables. We conducted survival analysis using the R survival program. The variation in survival rates was depicted using Kaplan-Meier survival curves, and the log-rank test was utilized to evaluate the significance of the variation in survival times among patient groups. All statistical P values are two-sided unless otherwise indicated, and P less than 0.05 is regarded as statistically significant.

Results

Technical roadmap

The data collection and analysis workflow is depicted in Figure S1. Each step will be elaborated upon in the subsequent section.

Data preprocessing and batch effect elimination

We normalized the GSE6631 and GSE107591 HNSCC datasets using the R package limma. Figure S2 showsthat after standardization, batch effects among samples in the GSE6631 and GSE107591 datasets have been significantly reduced, and the samples’ expression patterns have become more consistent.

Radiotherapy-related differentially-expressed genes

We obtained the intersection of the DEGs, FRGs, and RRGs from the three datasets, resulting in 11 overlapping genes (FSCN1, AURKA, EMP1, CA9, ENO2, ALOX12, IGF2BP3, GJA1, FAT1, PIM1, MUC1), which were referred to as F&RRDEGs (Fig. 1A-D). A depiction of the F&RRDEGs expression heatmap in the TCGA-HNSCC dataset is displayed in Fig. 1E. PCA computations were also performed on the expression matrix of F&RRDEGs. The PCA clustering findings revealed significant disparities between the two groups (Fig. 1F). According to the Wilcoxon rank sum tests’ outcomes, there were statistically significant expression differences for 10 F&RRDEGs in the TCGA-HNSCC dataset (Fig. 1G). Furthermore, as illustrated in Figs. 1H and 1I, all F&RRDEGs in the GSE6631 and GSE107591 datasets had statistical significance.

Fig. 1.

Screen F&RRDEGs A. Analysis of the TCGA-HNSCC dataset’s expression profile differences. B. Analysis of the GSE6631 dataset’s expression profile differences. C. Analysis of the GSE107591 dataset’s expression profile differences. D. Wayne plots of differential genes with Ferroptosis and radiotherapy. E. Expression of F&RRDEGs in TCGA-HNSCC dataset. F. PCA visualization of the TCGA-HNSCC dataset’s F&RRDEG expression profile. G–I. Group comparison chart of F&RRDEGs in the TCGA-HNSCC dataset, GSE6631 dataset, and GSE107591 dataset. (no significance, P ≥ 0.05; *P < 0.05; **P < 0.01; ***P < 0.001)

Analysis of expression differences of F&RRDEGs and ROC analysis

In the TCGA-HNSCC, GSE6631, and GSE107591 datasets, we compared the gene expression discrepancies of F&RRDEGs across the Control and HNSCC groups and created ROC curves for ten of the genes (CA9, ENO2, ALOX12, IGF2BP3, GJA1, FAT1, PIM1). The ROC curves demonstrated that ALOX12, EMP1, FAT1, GJA1, and IGF2BP3 exhibit significant diagnostic potential in the TCGA-HNSCC dataset. Consistently, AURKA, CA9, ENO2, and FSCN1 displayed an even higher diagnostic value in this dataset, while PIM1 exhibited a comparatively lower diagnostic value (Fig. 2A-E). Additionally, ALOX12, CA9, ENO2, FAT1, GJA1, IGF2BP3, and PIM1 manifested diagnostic significance in the GSE6631 dataset, while AURKA, EMP1, and FSCN1 demonstrated increased diagnostic capability in this particular dataset (Fig. 2F-J). Furthermore, the genes ALOX12, AURKA, CA9, EMP1, ENO2, FAT1, FSCN1, GJA1, IGF2BP3, and PIM1 exhibited diagnostic value in the GSE107591 dataset (Fig. 2K-O). The 95% confidence interval information for the AUC of the ROC curve is presented in Table S2.

Fig. 2.

Diagnostic Potential of Selected Genes Based on ROC Curve Analysis. A–E. F&RRDEGs ALOX12 (A blue line), CA9 (B blue line), EMP1 (B red line), and PIM1 (E red line) in the TCGA-HNSCC data set ROC curve results. F–J. F&RRDEGs ALOX12 (F blue line), CA9 (G blue line), EMP1 (G red line), and PIM1 (J red line) in the GSE6631 data set ROC curve results. K–O F&RRDEGs ALOX12 (K blue line, CA9 (L blue line), EMP1 (L red line), PIM1 (O red line) in the GSE107591 data set ROC curve results. FPR(False Positive Rate), TPR(True Positive Rate)

Overall, ALOX12, CA9, ENO2, FAT1, GJA1, and IGF2BP3 consistently demonstrated high diagnostic value across multiple datasets, indicating their potential utility as reliable biomarkers for HNSCC diagnosis. This highlights the importance of these genes in the clinical assessment of HNSCC and supports their further investigation in future studies.

Mutation and copy number variation analysis of F&RRDEGs

We applied the maftools software to present the mutational landscape of F&RRDEGs to assess the mutational status of 11 F&RRDEGs in patients with HNSCC (Figure S3A). The predominant mutation types are nonsense mutations, missense mutations, frameshift deletions, frameshift insertions, and in-frame deletions of the transcription start site, with nonsense mutations constituting a significant proportion. Moreover, single-nucleotide polymorphisms were the most common form of mutation among F&RRDEGs in HNSCC patients, with C > T and C > G being the most common variants.

In examining the Copy number variations (CNVs) of F&RRDEGs in HNSCC (showing the top 20 genes where copy number changes occur most frequently) (Figure S3B), it was observed that a significant proportion of these genes exhibited CNVs, with a greater prevalence in the gene CA9. Moreover, the majority of copy number changes were identified as significant amplifications (Figure S3C). For instance, the gene FSCN1 exhibited a high frequency of copy number amplification (above 20%), while deep copy number deletions were rarely observed.

Pathway and functional enrichment analysis of F&RRDEG

Our analysis of the TCGA-HNSCC dataset revealed significant enrichment of genes in several key pathways, including Negative Regulation of Notch4 Signaling, PDT-induced NFκB Survival Signaling, TP53 Regulates Transcription of Cell Cycle Genes, Oxidative Stress-Induced Senescence, Cellular Senescence, and DNA Damage Telomere Stress-Induced Senescence (Fig. 3) (Table S3). Furthermore, F&RRDEGs were found to be prevalent in various biological processes, including cell polarity, apical/basal cell polarity, monopolar cell polarity, and the maintenance of cell polarity in HNSCC samples. These genes were associated with cellular components such as actin-based cell projections, apical plasma membranes, microvilli, the apical part of the cell, and filopodia. In terms of enzymatic functions, F&RRDEGs were linked to hydrolases, carbon-oxygen cleavage enzymes, laccases, hydrolases acting on ether bonds, and oligopeptide transmembrane transporters (Table S4).

Fig. 3.

GSEA. A. The six primary biological features are represented as a mountain map. B–G. Six pathways exhibited significant enrichment among the genes within the TCGA-HNSC data

The results of enrichment analysis are visualized through histograms (Fig. 4A-C), showing significant enrichment of F&RRDEGs in biological processes (BP), cellular components (CC), and molecular functions (MF). Additionally, circular network diagrams illustrate the outcomes of the Gene Ontology (GO) functional enrichment analysis, highlighting the interconnections(Fig. 4D-F).

Fig. 4.

GO enrichment analysis of F&RRDEGs. A–C. GO enrichment analysis results. D–F. Circular network diagram of BP, CC, and MF in the GO analysis results. Blue dots within the network diagrams denote individual genes, while red dots represent distinct pathways

Correlation analysis of F&RRDEGs and subtype construction

We analyzed the expression of novel and alternative mRNA splicing-related genes in HNSCC samples from the TCGA-HNSCC dataset using “ConsensusClusterPlus”. This analysis revealed two distinct HNSCC subtypes: cluster 1 and cluster 2, as shown in Fig. 5A. HNSCC subtype 1 (cluster 1) consisted of 260 samples, while HNSCC subtype 2 (cluster 2) included 154 samples.

Fig. 5.

Construction of related subtypes of HNSCC. A. The outcome of consensus clustering (k = 2). B. PCA analysis of the two HNSCC subtypes. C. CDF charts for various consensus clustering numbers of clusters. D the Delta area under the CDF curve. E. Expression profiles of F&RRDEGs in different HNSCC subtypes. F. Grouped comparison of F&RRDEG expression in two subtypes in the TCGA-HNSCC dataset. (no significance, p ≥ 0.05; *p < 0.05; **p < 0.01, ***p < 0.001)

Notable distinctions between the two HNSCC subtypes were evident in the PCA clustering results (Fig. 5B). Cumulative Distribution Function (CDF) plots were generated for various cluster numbers in the consensus clustering results (Fig. 5C), accompanied by the Delta area under the CDF curve (Fig. 5D). The optimal consensus clustering outcome for the TCGA-HNSCC dataset was observed at k = 2.

Figure 5E illustrates the expression levels of the 11 F&RRDEGs across the two clusters, revealing significant negative correlations in their expression levels. From Fig. 5F, we identified that the 10 hub genes exhibited significantly different expression changes (P < 0.05). Specifically, EMP1, ALOX12, and MUC1 were up-regulated in cluster 2, while the remaining genes were up-regulated in cluster 1.

GSVA of hallmark gene sets in HNSCC subtypes

For the purpose of examining the distinct hallmark gene sets among the emerging HNSCC subtypes, we employed GSVA to analyze the gene expression data. Selecting hallmark gene sets with notable enrichment changes between clusters 1 and 2 was done using the limma program (adjusted p value < 0.05) (Table S5). A heatmap and a grouped comparison chart were generated for the 13 screened hallmark gene sets (Figure S4).

The 13 selected hallmark gene sets are Hypoxia, Glycolysis, DNA Repair, Notch Signaling, Kras Signaling Dn, Hedgehog Signaling, Pancreas Beta Cells, Epithelial-Mesenchymal Transition, Mtorc1 Signaling, Wnt Beta Catenin Signaling, Bile Acid Metabolism, Myc Targets V1, and Angiogenesis.

To assess the differences in genetic enrichment between the two subtypes, we employed a Mann-Whitney U test. The findings revealed that eleven routes significantly differed (P < 0.05), indicating distinct biological pathways are associated with the different HNSCC subtypes.

F&Rs construction and clinical correlation analysis

The ssGSEA method was employed to find the F&Rs for each sample. On the basis of the F&Rs’ median, we divided the HNSCC samples into groups with low and high scores. We constructed stacked histograms to examine the distribution of prognostic data, such as age, OS, gender, and stage, between the Low and High F&Rs groups (Fig. 6A-D). Additionally, We produced a grouped comparison chart by comparing the expression degrees of F&RRDEGs. The results of the comparison chart showed that nine F&RRDEGs (ALOX12, AURKA, CA9, EMP1, ENO2, FAT1, FSCN1, IGF2BP3, MUC1)showed strong expression in the high score group, suggesting their relevance in prognosis and potential clinical applications (Fig. 6E).

Fig. 6.

F&Rs construction and clinical correlation analysis of F&RRDEGs. A–D. Stacked bar charts depicting the distribution of age, OS, gender, and stage in the high as well as low groups of F&Rs. E. Grouped comparison chart showing the expression differences of F&RRDEGs between the high as well as low groups of F&Rs. (no significance, p ≥ 0.05; *p < 0.05; **p < 0.01, ***p < 0.001)

Immune infiltration and immunotherapy analysis

The TCGA-HNSCC dataset’s F&Rs Low, as well as High groups and the expression profiling data of 28 immune cell groups, were correlated using the ssGSEA algorithm. We made comparison charts according to the variations in immune infiltration levels across the Low and High groups(Fig. 7A). The findings revealed notable variations in infiltration (P < 0.05) among 19 immune cells. Correlations between specific F&RRDEGs and immune cell infiltration levels were established, highlighting the relationship between gene expression and the immune microenvironment.

Fig. 7.

Immune infiltration analysis of F&Rs grouping based on F&RRDEGs. A. Infiltration abundance of 28 immune cells. B–C. The immune infiltration abundance in the F&Rs low (B) along with high (C) groups of the TCGA-HNSCC dataset. D–E. Relationship among immune cells and F&RRDEGs in the F&Rs low (D) as well as high (E) groups. (no significance, p ≥ 0.05; *p < 0.05; **p < 0.01, ***p < 0.001)

Next, we performed additional calculations to determine the level of 19 immune cell infiltration. The findings demonstrated a positive correlation between immune cells and genes in the F&R groups classified as Low (Fig. 7B) and High (Fig. 7C).

We investigated the correlation between 11 F&RRDEG expression levels and 19 immune cell infiltration levels in TCGA-HNSCC. Figure 7D shows that FAT1, EMP1, and PIM1 were strongly linked to immune cells in the F&Rs Low group. It was seen in Fig. 7E that ENO2 and ALOX12 had a strong negative relationship with immune cells in the High group. We also retrieved the IPS related to HNSCC from the TCIA database in order to further explore the prognostic effect of F&RRDEGs in immune treatment. Next, we used R’s ggplot function to generate box plots illustrating the differences in the different IPS scores in TCGA-HNSCC patients (Figure S5). Analysis of Immune Phenotype Scores (IPS) indicated significant differences in CTLA4 and PD1/PD-L1/PD-L2 expression levels between low- and high-risk groups, suggesting that F&RRDEGs could serve as predictive markers for immunotherapy outcomes.

Construction of the clinical prognosis Cox model

Based on the clinicopathological parameters (age, gender, and stage) of HNSCC patients as well as the F&Rs of 11 F&RRDEGs, we ran a single-factor Cox regression analysis. In the multivariate analysis, all factors with P < 0.1 were included (Tables 2 and 3). Figure 8A illustrates the findings of the one-factor Cox regression. To evaluate the predictive ability of the model, a nomogram analysis was then performed (Fig. 8B). In the multivariate Cox prognostic model, tumor stage has a greater influence than the other factors. Furthermore, we created calibration curves for the multifactorial Cox prognostic models and carried out prognostic calibration studies for one, three, and five years (Fig. 8C-E). The closer the different colored lines are to the grey ideal situation, the better the line indicates a better prediction now.

Table 2.

Patient characteristics of HNSC patients in the TCGA datasets

Characteristics	Deceased	Alive	P value
n	173	241
Age, n (%)			0.221
> 50	146 (35.3%)	192 (46.4%)
<=50	27 (6.5%)	49 (11.8%)
Gender, n (%)			0.042
Male	118 (28.5%)	186 (44.9%)
Female	55 (13.3%)	55 (13.3%)
Stage, n (%)			0.001
Stage IV	123 (29.7%)	130 (31.4%)
Stage III	26 (6.3%)	49 (11.8%)
Stage II	21 (5.1%)	42 (10.1%)
Stage I	3 (0.7%)	20 (4.8%)
F&Rs,median (IQR)	1.6641 (1.5808, 1.7622)	1.6245 (1.5192, 1.7127)	0.002

Table 3.

COX regression analysis for F&Rs with OS in TCGA-HNSC

Characteristics	Total(N)	Univariate analysis			Multivariate analysis
		Hazard ratio (95% CI)	P value		Hazard ratio (95% CI)	P value
age	414		0.384
> 50	338	Reference
<=50	76	0.836 (0.553–1.262)	0.394
Gender	414		0.059
Male	304	Reference			Reference
Female	110	1.372 (0.995–1.890)	0.053		1.463 (1.058–2.024)	0.021
Stage	414		< 0.001
Stage IV	253	Reference			Reference
Stage III	75	0.627 (0.406–0.967)	0.035		0.631 (0.408–0.974)	0.038
Stage II	63	0.537 (0.337–0.857)	0.009		0.493 (0.306–0.793)	0.004
Stage I	23	0.256 (0.081–0.805)	0.020		0.240 (0.076–0.758)	0.015
F&Rs	414	2.731 (0.903–8.255)	0.075		1.380 (0.433–4.394)	0.586
ALOX12	414	0.898 (0.774–1.043)	0.158
AURKA	414	1.243 (1.010–1.528)	0.040		1.231 (0.997–1.521)	0.054
CA9	414	1.049 (0.963–1.142)	0.275
EMP1	414	1.021 (0.875–1.191)	0.794
ENO2	414	1.084 (0.959–1.226)	0.195
FAT1	414	1.068 (0.935–1.220)	0.333
FSCN1	414	1.183 (0.972–1.440)	0.093		1.176 (0.963–1.437)	0.113
GJA1	414	1.023 (0.910–1.150)	0.703
IGF2BP3	414	1.096 (0.942–1.274)	0.234
MUC1	414	1.014 (0.910–1.130)	0.801
PIM1	414	1.013 (0.834–1.232)	0.893

Fig. 8.

Construction of the clinical prognosis Cox model. A.Univariate Cox regression model. B. Multivariate Cox regression model. C–E. Calibration curves. F–H. DCA plot. (no significance, p ≥ 0.05; *p < 0.05; **p < 0.01, ***p < 0.001)

Next, we employed DCA to assess the multivariate Cox regression model’s effects on clinical utility at one, three, and five years (Fig. 8F-H, respectively). The clinical prediction impact of our multi-factor model ranked the periods as follows: 5 years had a higher effect than 3 years, which in turn had a higher effect than 1 year.

Univariate and multivariate Cox regression analyses, focusing on clinical parameters and F&Rs, established a prognostic model, with stage being the most influential factor. This model’s predictive ability was further validated through nomogram analyses and decision curve analyses for various time frames, demonstrating its clinical utility.

Gene and protein expression verification

WB verified the expression of ferroptosis-related proteins in SCC25 cells and RT-SCC25 cells. It was confirmed that after HNSCC cells were treated with radiotherapy, the expressions of VDAC1 and TFRC were increased, GPX4, SLC7A11, and FTH1 were down-regulated (Fig. 9A). It was confirmed that HNSCC cell radiotherapy was closely related to the ferroptosis pathway.

Fig. 9.

Expression and analysis of ferroptosis-related proteins and genes. A. WB of ferroptosis-related proteins in SCC25 and RT-SCC25 cells. B. RT-qPCR of ferroptosis-related gene mRNA levels in HOK and SCC25 cells. C. WB of ferroptosis-related proteins in HOK and SCC25 cells. D. WB of NF-κB signaling pathway-related proteins in HOK and SCC25 cells

Verification of gene and protein expression in HOK and SCC25 cells by RT-qPCR and WB confirmed that ALOX12, PIM1, and EMP1 were down-regulated in HNSCC compared with the control, while CA9 was significantly overexpressed, suggesting its potential role in HNSCC pathology (P < 0.05) (Fig. 9B-C). Meanwhile, WB verified the expression of proteins related to the NF-κB signaling pathway in HOK and SCC25 cells and confirmed that the expressions of p-NFκBp65 and p-IKBα were significantly increased in HNSCC, suggesting the potential role of the NF-κB signaling pathway in HNSCC (Fig. 9D).

Discussion

HNSCC is a prevalent malignancy worldwide, accounting for 900,000 cases and half a million deaths annually [23]. The poor prognosis in advanced stages, where 10–30% of patients develop metastases with a median survival of 6–15 months [24, 25], underscores the urgent need for biomarkers to guide therapy [26]. In this study we identified and validated F&RRDEGs with potential diagnostic and prognostic biomarkers in HNSCC. We identified ALOX12, EMP1, PIM1, and CA9 as key diagnostic markers (Fig. 2). Furthermore, our results demonstrate that radiotherapy induces ferroptosis in this context, mediated by VDAC1/TFRC upregulation and GPX4/SLC7A11 downregulation (Fig. 9A), providing evidence for this mechanism in HNSCC. By integrating radiotherapy response, ferroptosis regulation, and immune microenvironment analysis, a prognostic model with clinical application value was constructed [27, 28].

Ferroptosis is a unique form of cell death driven by iron buildup and toxic lipid peroxides [29, 30]. Cancer cells are especially vulnerable because their altered iron metabolism (like high iron uptake) promotes this process. Its pivotal role in cancer therapeutics stems from the inherent susceptibility of malignant cells to oxidative stress, where dysregulated iron metabolism (e.g., elevated transferrin receptor and suppressed ferroportin) promotes intracellular iron accumulation, thereby enabling ferroptosis. Notably, radiotherapy synergizes with ferroptosis induction by repressing SLC7A11-mediated cystine uptake, depleting glutathione (GSH), and inactivating GPX4—a master regulator of lipid peroxidation—ultimately amplifying tumor cell death [31, 32]. This makes targeting ferroptosis a promising way to overcome radiation resistance in head and neck cancer (HNSCC).

Building on the identification of 11 radiotherapy-ferroptosis DEGs and their pathway alterations (GO/GSEA/GSVA), our study establishes a novel molecular subtype classification for HNSCC. Crucially, unlike previous studies, which primarily suggested a potential role for ferroptosis in cancer therapy enhancement [6] or explored its relevance in other cancers like colorectal adenocarcinoma [33]. Meanwhile, few studies have explored the association between the prognosis of HNSCC patients and genes related to ferroptosis and radiotherapy. Some studies have demonstrated the feasibility of developing AI-based diagnostic models through the analysis of extensive clinical, histological, and genetic datasets. These patterns and data can be leveraged to evaluate tumors, thereby enhancing diagnostic accuracy [34]. Our work provides direct experimental validation for radiotherapy-induced ferroptosis specifically in HNSCC. Through Western blot analysis, we demonstrated that irradiated HNSCC cells exhibit a distinct ferroptosis-promoting molecular signature: upregulation of VDAC1 (mitochondrial lipid peroxidation) and TFRC (iron uptake), coupled with downregulation of GPX4 (antioxidant defense), SLC7A11 (cystine/glutathione axis), and FTH1 (iron storage). Consequently, our findings advocate for the incorporation of these validated ferroptosis biomarkers (VDAC1, TFRC, GPX4, SLC7A11, FTH1) into prognostic frameworks and the development of personalized therapeutic strategies to optimize HNSCC management.

11 DEF&RGs were identified based on the TCGA-HNSCC, GSE6631, and GSE107591 data sets. Among the 11 DEGs, only MUC1 showed no statistically significant difference in expression in the TCGA-HNSCC dataset. Moreover, through ROC curve analysis, MUC1 and FAT1 expression levels were insignificant for the survival diagnosis in patients with HNSCC. Besides MUC1, the expression of the remaining ten genes from the GSE6631 and GSE107591 datasets contributed to predicting survival outcomes in HNSCC patients. In both datasets, EMP1 has the highest diagnostic value (AUC = 0.988/0.971, respectively). The mutation analysis revealed that the most common mutations in the F&RRDEGs within the TCGA-HNSCC dataset were nonsense mutations, missense mutations, and frameshift deletions, with nonsense mutations predominant. Additionally, most genes in HNSCC patients showed CNVs, particularly the CA9, which exhibited a higher frequency of copy number alterations.

Our analysis identified CA9 and EMP1 as particularly noteworthy among the 11 radiotherapy-ferroptosis DEGs due to their diagnostic and prognostic significance. Crucially, CA9 demonstrated exceptionally high diagnostic value for HNSCC (AUC = 0.964), and exhibited a high frequency of copy number variations within the TCGA-HNSCC cohort. While previous studies have established CA9 as a hypoxia-responsive enzyme linked to poor treatment response and prognosis in various cancers, and implicated its upregulation in blocking ferroptosis and promoting malignancy [35, 36], our findings specifically highlight its strong diagnostic potential (high AUC) and prevalent genomic instability (frequent CNVs) in the context of HNSCC radiotherapy-ferroptosis DEGs [37, 38]. This positions CA9 as a valuable biomarker candidate for HNSCC diagnosis and prognostication.

Conversely, EMP1 expression was significantly downregulated in HNSCC. Database analysis, RT-qPCR, and Western blot consistently confirmed lower EMP1 levels in SCC25 compared to HOK (P < 0.05) (Fig. 9B-C). Although EMP1’s role in cancer has been contentious [39], our multi-faceted validation strongly supports its tumor-suppressive function in HNSCC, aligning with reports linking its downregulation to lymph node metastasis and suggesting roles in inhibiting invasion/migration and promoting apoptosis [40, 41]. Importantly, EMP1 also showed the highest diagnostic value for survival prediction in both GSE6631 and GSE107591 datasets (AUC = 0.988/0.971), a novel finding specific to the HNSCC radiotherapy-ferroptosis context not previously emphasized in the literature. The distinct expression patterns of CA9 (upregulated) and EMP1 (downregulated) in our patient cohort (P < 0.05) further validate their relevance in HNSCC pathogenesis and potential as biomarkers.

Functional enrichment analyses (GSEA/GO) of the 11 radiotherapy-ferroptosis DEGs revealed their significant association with key biological processes, including DNA damage repair, oxidative stress, and tumor microenvironment remodeling – pathways critically linked to radiotherapy response and ferroptosis. Notably, our Western blot analysis provided novel experimental evidence specific to HNSCC: expression of phosphorylated NF-κB p65 (p-NFκBp65) and phosphorylated IκBα (p-IκBα) was significantly elevated in SCC25 compared to HOK (Fig. 9, P < 0.05). While previous studies have established that DNA damage can activate both pro-ferroptotic (e.g., p53 network [42, 43]) and anti-ferroptotic pathways (e.g., NF-κB-mediated survival signals [36]), and that oxidative stress is a core driver of ferroptosis [37], our findings uniquely demonstrate the constitutive activation of the NF-κB pathway in HNSCC cells at the protein level. This suggests that in HNSCC, the anti-apoptotic/anti-ferroptotic arm of the DNA damage response (NF-κB) may be aberrantly active, potentially enabling tumor cell survival and resistance under radiotherapy-induced stress - a mechanism not previously experimentally documented in this specific context. Consequently, the observed NF-κB pathway activation, coupled with its known role in mitigating oxidative stress and ferroptosis [44, 45], positions it as a promising therapeutic target. We propose that combining NF-κB pathway inhibitors with ferroptosis inducers could synergistically block pro-survival signaling and enhance radiotherapy sensitivity in HNSCC, particularly for tumors exhibiting NF-κB overactivation and iron metabolism dysregulation.

Leveraging the expression profiles of the 11 F&RRDEGs within the TCGA-HNSCC cohort, we established a novel molecular subtype classification of HNSCC using consensus clustering [46, 47]. This analysis robustly identified two distinct subtypes (Cluster 1 and Cluster 2). Crucially, significant differential expression of 10 key hub genes defined these subtypes (Wilcoxon rank-sum test, P < 0.05): Cluster 1 exhibited upregulation of FSCN1, AURKA, CA9, ENO2, IGF2BP3, GJA1, and FAT1, while Cluster 2 showed upregulation of EMP1, ALOX12, and MUC1.

While large-scale omics analyses have revolutionized cancer subtyping and personalized treatment approaches in general [48], and specific molecular classifications (e.g., intrinsic subtypes in breast cancer [49]) have proven highly valuable for prognosis and therapy prediction; our work uniquely focuses on integrating radiotherapy and ferroptosis biology to define HNSCC subtypes. The distinct expression patterns of genes like CA9 (hypoxia/ferroptosis), EMP1 (putative tumor suppressor), and ALOX12 (lipid peroxidation) across subtypes suggest potential differences in radiotherapy sensitivity and ferroptosis susceptibility. This classification system provides a preliminary framework for exploring subtype-specific therapeutic responses; however, its clinical utility requires validation in prospective cohorts and functional studies.

GSVA analysis [50] revealed significant enrichment differences across 13 hallmark pathways between our two novel HNSCC subtypes (Cluster 1 vs. Cluster 2), with 11 pathways showing statistically significant differences (P < 0.05). Notably, Cluster 1 exhibited significant enrichment in pathways linked to treatment resistance and poor prognosis, including Hypoxia, Glycolysis, and DNA Repair [51, 52]. This pathway signature aligns with the Cluster 1 gene expression profile (e.g., upregulation of CA9, a known hypoxia marker).

Immune infiltration analysis identified significant differences (P < 0.05) in 19 immune cell types between the subtypes. Subtype-specific correlation analysis further delineated the relationships between F&RRDEGs and immune cells: in Cluster 1 (High-risk signature), key genes like ENO2 and ALOX12 showed significant negative correlations with immune infiltration. Conversely, in Cluster 2 (characterized by EMP1/FAT1/PIM1 upregulation), genes such as FAT1, EMP1 and PIM1 demonstrated significant positive correlations with immune cells.

Experimental validation of ALOX12 and P1M1 provided additional critical insights into the two key correlating genes. Although ALOX12 is established as a crucial executor in p53-mediated ferroptosis, independent of GPX4 [53], Our RT-qPCR and WB results revealed significantly low expression levels of ALOX12 in HNSCC patient samples (P < 0.05). This finding, coupled with its negative correlation with immune cells in the aggressive Cluster 1, suggests a potential suppression of its pro-ferroptotic role in HNSCC, warranting investigation into its role in radiotherapy response and immune evasion. Despite its well-documented oncogenic role in promoting tumor growth and therapy resistance in cancers like prostate cancer [54, 55], our RT-qPCR and WB analysis similarly confirmed significantly low expression of PIM1 in HNSCC patients (P < 0.05). This unexpected downregulation, observed alongside its positive correlation with immune cells in Cluster 2, suggests a context-dependent role in HNSCC distinct from its typical oncogenic function and merits focused study.

Together, these distinct pathway activities, immune microenvironments, and the validated dysregulation (particularly the low expression) of ALOX12 and PIM1 between subtypes underscore the biological and potential clinical relevance of our F&RRDEG-based classification. The contrasting immune correlations of these genes further suggest subtype-specific immune interaction patterns.

Significant variations in the presence of immune cells were observed between high- and low-risk groups, strongly suggesting that our F&R signature holds predictive potential for immunotherapy response in HNSCC patients.

Building upon known prognostic checkpoint regulators in HNSCC, such as PD-1 and CTLA-4 [56–58], our study uniquely integrates these immunological contexts with a novel F&R-based risk stratification. To further explore the clinical prognostic value of the F&R signature, we constructed a multifactorial Cox prognostic model incorporating both the risk score and key clinical indicators. Tumor stage demonstrated the highest predictive utility within this integrated model compared to other variables. Validation analyses confirmed the robustness of our model: Calibration curves at 1, 3, and 5 years indicated optimal predictive accuracy at the 1-year mark. Furthermore, Decision Curve Analysis (DCA) demonstrated increasing net benefit and larger AUC values over time (1-, 3-, 5-year), collectively affirming the model’s strong and progressively reliable prognostic predictive power for HNSCC. These findings underscore the practical clinical utility of our F&R-based multifactorial model as a novel prognostic tool.

While this study identified F&R-linked genes relevant to HNSCC, future validation requires integrating comprehensive clinical data and conducting prospective clinical studies to definitively establish the prognostic model’s therapeutic utility and mitigate potential selection bias inherent in retrospective, open-source datasets.

Beyond model validation, our findings highlight critical mechanistic questions: The role of NF-κB pathway activation (evidenced by radiotherapy-induced p-NFκBp65/p-IKBα upregulation) in suppressing ferroptosis warrants investigation, as its pro-survival signaling may counteract lipid peroxidation, contributing to therapeutic resistance. This potential link between radio-induced NF-κB and ferroptosis evasion represents a novel mechanistic axis suggested by our data. The functional impact of key identified genes (e.g., ALOX12, EMP1, PIM1, CA9) on ferroptosis sensitivity in HNSCC needs elucidation to inform targeted therapy development.

Furthermore, acknowledging HNSCC heterogeneity, longitudinal studies tracking F&RRDEG profile evolution during treatment are essential to understand dynamic biomarker-response relationships. Ultimately, validating the F&R signature’s ability to stratify patients could pave the way for personalized treatment optimization in HNSCC.

This study identified differentially expressed genes associated with ferroptosis and radiotherapy in HNSCC and established a prognostic model based on multi-omics statistical analysis. However, this study still has certain limitations. First, the retrospective nature of public datasets (TCGA/GEO) may introduce selection bias. Second, although we validated key genes (ALOX12, EMP1, PIM1, CA9) and proteins in cell lines, their functional roles in ferroptosis and radioresistance require further mechanistic investigation. Third, the prognostic model was constructed from transcriptomic data without external clinical validation. Future studies should include functional assays, larger prospective cohorts, and multi-omics integration to refine clinical applicability.

Conclusion

In summary, our study thoroughly investigated the link between ferroptosis and radiotherapy-related genes and their impact on the prognosis of HNSCC. We successfully developed and validated a novel prognostic model using 11 F&RRDEGs for HNSCC patients.This model demonstrates potential prognostic value in our retrospective cohort, but further validation in independent datasets and prospective studies is needed to assess its clinical utility for immunotherapy prediction. Additionally, we observed lower expression of ALOX12, PIM1, and EMP1 genes and higher expression of CA9 in HNSCC patients (P < 0.05). Importantly, the results of this research indicate that ferroptosis constitutes a core mechanism of tumor cell death induced by radiotherapy, thereby emphasizing the profound significance of ferroptosis-related biomarkers in the context of personalized therapy and prognostication for HNSCC. Consequently, this work furnishes a new theoretical basis and delineates future practical directions for investigating the molecular crosstalk between ferroptosis and radiotherapy, developing targeted interventions, and enhancing patient care in HNSCC.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

HNSCC

Head and neck squamous cell carcinoma

DEGs

Differentially expressed genes

RRGs

Radiotherapy-related genes

FRGs

Ferroptosis-related genes

F&RRDEGs

Ferroptosis and radiotherapy-related differentially expressed genes

F&Rs

Ferroptosis & radiotherapy scores

HOK

Human oral keratinocytes

SCC25

Squamous cell carcinoma 25

RT-qPCR

Real-time quantitative PCR

WB

Western blotting

DCA

Decision curve analysis

ROS

Reactive oxygen species

GPX4

Glutathione peroxidase 4

TCGA

The Cancer Genome Atlas

GEO

Gene Expression Omnibus

IPS

Immunophenoscore

OS

Overall survival

adj.P.Val

Adjusted p-value

ROC

Receiver operating characteristic

AUC

Area under the curve

CNV

Copy number variation

GSEA

Gene Set enrichment analysis

BP

Biological process

MF

Molecular function

CC

Cellular component

GO

Gene ontology

PCA

Principal component analysis

GSVA

Gene set variation analysis

SsGSEA

Single-sample gene-set enrichment analysis

TCIA

The Cancer Immunome Atlas

cDNA

Complementary DNA

EMP1

Epithelial membrane protein 1

CA9

Carbonic anhydrase 9

ALOX12

Arachidonate 12-lipoxygenase

PIM1

Moloney murine leukemia virus 1

Author contributions

Conceptualization and design of the study were conducted by LF and YM. Data analysis and interpretation were performed by LF, JW and ZW. LF and LS wrote manuscript. YM and LF revised the manuscript.

Funding

This work was supported by the Chinese National Natural Science Foundation (Grant Number: 82470953). The Chengdu Scinece and Technology Project (Grant Number:2024-YF09-00026-SN) .

Data availability

The datasets analyzed during this study are publicly available in the following repositories: TCGA-HNSCC: Accessed via The Cancer Genome Atlas (TCGA) portal (https://portal.gdc.cancer.gov/). GSE6631 and GSE107591: Available through the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE6631, https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc= GSE107591).Ferroptosis-related genes (FRGs) and Radiotherapy-related genes (RRGs): Curated from the GeneCards database (https://www.genecards.org/).

Declarations

Ethics approval and consent to participate

All experimental procedures involving human cell lines (HOK and SCC25) were conducted in accordance with relevant guidelines and regulations. The HOK and SCC25 cell lines used in this study were kindly donated by Prof. Yongcong Cai (University of Electronic Science and Technology of China). Ethical approval for the use of these cell lines was waived as they are commercially available and anonymized, with no direct involvement of human participants or personal data.

Consent for publication

All authors approved the final manuscript and the submission to this journal.

Competing interests

The authors declare no competing interests.