Immunogenomic signatures defining the molecular subtypes of head and neck cancer

Abstract

The immune system is crucial in the development and advancement of cancerous tumors, particularly in head and neck squamous cell carcinoma (HNSC). This study aimed to identify immune-related gene signatures (IRGs) for predicting the prognosis of HNSC. Transcriptome data from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO), in addition to immune gene data from ImmPort, were examined. Using Cox-LASSO screening, nine IRGs were identified, and patients were classified into high- and low-risk cohorts based on risk scores. Differential expression, survival analysis, Gene Set Enrichment Analysis (GSEA), Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), single-sample GSEA (ssGSEA), CIBERSORT, and drug sensitivity analyses were performed between the cohorts. The high-risk cohorts exhibited lower immune scores and survival rates, while the low-risk cohorts exhibited higher immune scores and better outcomes. Cox regression identified CD19, CD79A, CTLA4, ICOS, and LAT as protective genes and CHGB, DKK1, PDGFA, and PTX3 as risk genes. Based on these nine genes, we established a nomogram prediction model to further assess patient prognosis. These findings highlight the prognostic value of IRGs in HNSC, offering the potential for personalized treatment strategies based on immune risk profiles.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04518-y.

Introduction

Head and neck cancers encompass malignancies arising in the oral cavity, pharynx, larynx, nasopharynx, thyroid, and salivary glands, contributing significantly to global morbidity and mortality. According to 2022 global cancer statistics, thyroid cancer ranks eighth worldwide in incidence, with a mortality rate of 0.5%. Oral cancer ranks 16th with a mortality rate of 1.9%, while laryngeal and nasopharyngeal cancers rank 19th and 23rd, with mortality rates of 1.1% and 0.8%, respectively [1]. The 5-year survival rate for patients with head and neck squamous cell carcinoma (HNSC) is approximately 40%~50% [2], primarily due to local recurrence or distant metastasis observed in 80%~90% of patients with advanced-stage disease. Conventional treatments, including surgery, radiotherapy, and chemotherapy, often face limitations related to tumor heterogeneity and drug resistance.

Common risk factors of HNSC include smoking, alcohol abuse, Epstein-Barr virus, and human papillomavirus infections [3–6]. Tumorigenesis of HNSC is driven by complex mechanisms, including oncogene overexpression, tumor suppressor gene inactivation, dysregulated cellular signaling, microenvironmental changes, and immune evasion [7]. These factors either independently or synergistically promote tumor development, progression, and metastasis [8]. Recent progress in immunotherapy and immune checkpoint inhibitors, such as PD-1/PD-L1 blockers, has demonstrated potential in cancer treatment [9, 10]. However, the effectiveness of immunotherapy significantly depends on the patient’s immunophenotype, emphasizing the need to understand distinct immunophenotypes of HNSC and their biological characteristics to develop personalized treatment strategies and improve prognostic assessments.

The development of immune-related gene signatures (IRGs) for HNSC is anticipated to provide crucial insights into the prognostic relevance of immunogenomic profiles and the influence of infiltrating immune cells. This approach is strongly supported by recent studies. Research has demonstrated that individual IRGs (SLC2A3) can serve as key regulators of the tumor immune microenvironment, where their expression levels are negatively correlated with CD8⁺ T cell infiltration and strongly linked to immune escape and poor prognosis [11]. Furthermore, the therapeutic potential of targeting immune-related pathways in HNSC, such as the Stat3 and Bcl-2/Bax axis, highlights the critical importance of deepening our understanding of the immunogenomic landscape [12]. Collectively, these findings highlight the necessity of exploring integrated IRG-based prognostic models to more accurately stratify patients and identify novel therapeutic targets. In this study, we categorized patients with HNSC into high- and low-risk cohorts and compared their immune microenvironments, prognoses, and drug sensitivities. Using data from Transcript data from The Cancer Genome Atlas (TCGA), Gene Expression Omnibus (GEO), and ImmPort data sources, we conducted a comprehensive analysis, including differential gene expression analysis, Cox regression analysis, least absolute shrinkage and selection operator regression analysis (LASSO), single-sample gene set enrichment analysis (ssGSEA), CIBERSORT, joint immune checkpoint analysis, and drug sensitivity assessments.

The aim of this study was to identify and evaluate the prognostic immune-related genes in HNSC and to develop a predictive model for patient survival based on these genes.

Materials and methods

Data set

Gene expression matrices (raw counts from the “STAR - Counts” dataset) and corresponding clinical data for the Head and Neck Squamous Cell Carcinoma (TCGA-HNSC) cohorts were retrieved from the UCSC Xena platform (https://xenabrowser.net/). A total of 546 samples were included, comprising 494 primary tumor samples (01 A) and 44 adjacent normal samples (11 A). Overall survival information was available for all 494 tumor patients. These raw count data were used for all subsequent differential expression analyses.

To enhance the robustness and generalizability of our findings, we incorporated two independent validation cohorts from the Gene Expression Omnibus (GEO): GSE37791, which contains 40 oral squamous cell carcinoma (OSCC) samples and 40 matched normal tissues; and GSE41613, comprising 97 OSCC samples with available clinical outcome data [13]. The integration of these multi-center cohorts strengthens the statistical power and clinical relevance of our immunogenomic subtyping approach.

A comprehensive list of 2,483 IRGs was obtained from the ImmPort database (https://www.immport.org/home) [14]. This curated gene set served as the foundation for identifying prognostic immune features and constructing the subsequent molecular subtypes.

Differential expression analysis

Differentially expressed genes (DEGs) exhibit significant variations in RNA expression levels under specific conditions, such as environmental stress, temporal changes, or spatial differences. Differential expression analysis was performed to identify DEGs between distinct sample cohorts. The DESeq2 package in R was used to compare gene expression levels between disease (tumor) and control (normal) cohorts [15]. Genes were considered significantly differentially expressed if they met the |log₂FC| ≥ 1 criterion and had an adjusted p-value < 0.05.

Cox regression

The Cox regression model, commonly referred to as the proportional hazards model (Cox model), operates under the assumption that the hazard ratio (HR) remains constant over time, which is known as the proportional hazards assumption. This model uses survival outcome and time as response variables, enabling the simultaneous evaluation of the effects of multiple factors on survival. The prognostic significance of genes was evaluated using univariate Cox proportional hazards regression via the survival package in R. This model accommodates truncated survival data and does not assume a specific survival distribution [16].

LASSO regression

LASSO regression analysis, a technique used for compression estimation, refines models by constructing a penalty function that compresses specific regression coefficients. This process forces the sum of the absolute values of the coefficients to remain below a specified threshold [15]. This process can set some coefficients to zero, effectively performing the variable selection. LASSO regression retains the benefits of subset shrinkage and is a biased, valuable estimator for data with covariates. Genes identified through Cox analysis were further screened via LASSO regression via the glmnet package in R.

Quantitative real-time PCR (qPCR)

Total RNA was extracted from 10 pairs of snap-frozen HNSC tumor and matched normal tissues. cDNA was synthesized from 1 µg of total RNA. qPCR was performed with SYBR Green chemistry. Reactions were set up in triplicate with gene-specific primers (sequences listed in Supplementary Table 1). The PCR protocol comprised an initial denaturation at 95 °C for 30 s, followed by 40 cycles of 95 °C for 5 s and 60 °C for 30 s, with a final melting curve analysis. Relative gene expression was calculated using the 2^–ΔΔCt method normalized to GAPDH, and statistical significance was assessed by paired Student’s t-test (p < 0.05).

Immune stroma score calculation

The tumor microenvironment (TME) encompasses the surroundings of tumor cells, including the extracellular matrix, soluble molecules, and stromal cells. Immune and stromal cells within the TME are key non-tumor components that play an essential role in tumor diagnosis and prognostic evaluation [17]. The immune and stromal scores for the samples were calculated using the Estimate package in R. The Estimate method relies on gene expression data to infer the abundance of immune and stromal cells by analyzing gene expression features associated with these components. This approach uses a weighted scoring system and predefined gene sets to estimate the immune score and stromal score for each sample.

GO and pathway enrichment analysis

The Gene Ontology (GO) database categorizes gene functions into three primary cohorts: cellular components, which denote specific locations within a cell where functions take place; molecular functions, which outline the roles of genes at the molecular level; and biological processes, which encompass broader processes involved in biological reactions. The Kyoto Encyclopedia of Genes and Genomes (KEGG) is a comprehensive database that integrates genomic, chemical, and functional information. It links gene catalogs from fully sequenced genomes with higher-level system functions at cellular, species, and ecosystem levels. KEGG analysis is valuable for understanding the complex functions and utilities of biological systems [18, 19]. Functional enrichment of the GO and KEGG pathways was performed using the clusterProfiler package in R (p.value < 0.05), with results visualized via the ggplot2 package.

Immune infiltration

ssGSEA is effective for studying complex TME and immune responses. It starts by selecting relevant gene sets, such as those related to immune cell types, and ranks genes in each sample based on expression levels. The enrichment score (ES) reflects the concentration of genes from the set in the ranked list, with higher scores indicating greater enrichment. This method allows precise assessment of immune cell infiltration and other biological features in individual samples [20]. In this study, ssGSEA was used to calculate ES for 28 immune cell types, using immune-related genes from the literature and the e1071 package for parallel processing and abundance estimation.

CIBERSORT

CIBERSORT is a method used to estimate immune cell infiltration by deconvolving gene expression matrices with linear support vector regression. It utilizes the LM22 matrix, which contains gene signatures for 22 immune cell types, to analyze the gene expression data from samples. By minimizing the error between the sample data and reference gene sets, CIBERSORT estimates the relative abundance of each immune cell type. Additionally, it provides p-values to assess the reliability of the estimated abundances.

Survival analysis

Kaplan-Meier (K-M) curves are commonly used in survival analysis to estimate survival rates and assess the effect of a single factor on outcomes. The K-M method calculates survival probabilities at each time point, adjusting for endpoint events (such as death). When an event occurs, the survival probability decreases, with the magnitude of the drop reflecting the proportion of events relative to patients still under observation [21]. This results in a smooth survival curve that shows the relationshipbetween survival time and survival rate. In this study, survival analysis was conducted using the Survival package, and K-M curves for high- and low-risk cohorts were generated with the Survminer package. The survfit function was used to fit the data, and the ggsurvplot function was employed to compare survival differences between cohorts.

Nomogram construction

Using the survival and RMS packages, risk scores were calculated based on multi-factor Cox regression analysis, incorporating significant clinical features and gene data. The formula for calculating the risk score is: Risk Score=∑ (Coefficient of each variable×Value of the variable). Based on the Cox analysis, N stage and Age were selected as significant factors and included in the nomogram for intuitive prognostic prediction.

Gene set enrichment analysis

GSEA (Gene Set Enrichment Analysis) assesses the enrichment of predefined gene sets in samples to identify significant differences in biological processes, pathways, or functions. It calculates an ES based on the overall ranking of genes within a set, rather than focusing on individual gene significance [22]. The analysis compares gene set enrichment across sample cohorts to reveal functional variations. To examine the functional differences between the two clusters, GSEA was performed in this study.

Mutation analysis

Tumor mutation burden (TMB) reflects the total number of somatic mutations present in the tumor genome [23]. It quantifies the number of somatic non-synonymous mutations, such as gene coding errors, base substitutions, and gene insertion or deletion errors, per megabase pair of the tumor genome. TMB serves as a quantitative biomarker that reflects the overall mutational load of tumor cells and indirectly indicates the potential for neoantigen production. Tumors with high TMB are thought to generate more neoantigens, which may enhance immune system recognition of cancer cells, stimulate the proliferation of antitumor T cells, and activate subsequent immune responses. Therefore, a higher TMB is theoretically associated with increased neoantigen recognition by the immune system and potentially greater efficacy of immunotherapy. The maftools package in R was used to analyze somatic mutations between the high- and low-risk cohorts.

Immune checkpoints

Immune checkpoint molecules are proteins that regulate immune responses and maintain autoimmune functions within normal limits [24]. Notable examples include PD-1 [25], CTLA-4 [26], LAG-3 [27], LGALS9, HAVCR2, PDCD1, PDCD1LG2, and TIGIT. Immune checkpoint molecules are crucial for preventing excessive immune activation, thus avoiding conditions such as immune escape, immunosuppression, and autoimmune diseases. The abnormal expression of these molecules can disrupt normal immune functions and contribute to immune-related disorders. Differential expression of immune checkpoint genes between high- and low-risk cohorts was evaluated.

Transcription factors

Messenger RNAs can be translated into various functional proteins, including a prominent class of regulatory proteins known as transcription factors (TFs). TFs play crucial roles in regulating the initiation of transcription, making them central to transcriptional regulation studies [28]. TFs related to prognostic genes were predicted using miRNet on the basis of the JASPAR database.

Drug sensitivity

Drug sensitivity was predicted using the oncoPredict R package [29, 30]. The model was trained on the GDSC2 database, which contains gene expression profiles and half-maximal inhibitory concentration (IC50) data for 198 drugs across 805 cancer cell lines. For the HNSC samples, RNA-seq expression data (FPKM) were log2-transformed after adding a pseudo-count of 1 and corrected for batch effects using the ComBat empirical Bayes method (batchCorrect = ‘eb’) to align with the GDSC2 training set. The calcPhenotype function was used to predict IC50 values by automatically matching each HNSC sample to the most similar GDSC2 cell line based on genome-wide expression. To ensure robustness, we applied the following parameters: removal of the 20% lowest-varying genes (remove Low Varying Genes = 0.2), power transformation of IC50 values (powerTransformPhenotype = TRUE), and a minimum of 10 samples per drug (minNumSamples = 10). The resulting IC50 matrix was used to evaluate drug sensitivity correlations across risk cohorts.

Genemania

This study aimed to provide insights into protein-gene interactions, pathways, co-expression, co-localization, and protein structural domain similarities of the submitted genes. GeneMANIA was used to analyze gene interactions, pathways, co-expression, co-localization, and structural domain similarities.

Results

Differential expression analysis

A total of 4,785 differentially expressed genes (DEGs) were identified between normal and tumor samples in TCGA-HNSC, including 2,413 upregulated and 2,372 downregulated genes in tumors. The volcano plot of these DEGs is shown in Fig. 1A. These DEGs were then intersected with 2,027 immune-related genes obtained from the ImmPort Portal, resulting in 456 candidate genes (Fig. 1B). Differential expression analysis was performed using the DESeq2 package in R, with thresholds of |log2 fold change| ≥ 1 (i.e., ≥ 2-fold change) and P < 0.05.

Fig. 1.

A Volcano plot showing differentially expressed genes. B Intersection of differentially expressed genes in HNSC and immune-related genes. C COX regression analysis identified relevant genes, with the top10 genes (p < 0.05) displayed. D LASSO Cox analysis selected nine immune-related genes. E Patients were divided into high-risk and low-risk cohorts based on their risk scores, with prognostic survival rates showing statistically significant differences. F Concordance index for the LASSO-Cox model reached 0.72

Modeling and evaluation

Cox regression analysis was performed on the 456 candidate genes, with a p-value cutoff of 0.05. In total, 96 genes were identified, and the top ten genes with the strongest expression and p-values are shown in Fig. 1C. Genes with a higher HR were classified as risk genes, while those with an HR value less than 1 were categorized as protective genes. Specifically, CD19, CD40LG, CD79A, CTLA4, ICOS, and LAT were classified as protective genes, while CHGB, DKK1, PDGFA, and PTX3 were classified as risk genes. Notably, aberrant expression of CHGB has been reported in various cancers, including pancreatic cancer [31] and low-grade gliomas [32]. Its low expression is often associated with poor prognosis and is considered a protective factor, while its upregulation is often associated with cancer metastasis [33].

LASSO regression analysis

LASSO regression analysis was then applied to the 96 significantly associated genes, reducing the set to nine essential genes: CD19, CD79A, CTLA4, ICOS, LAT, CHGB, DKK1, PDGFA, and PTX3 (Fig. 1D). Patients were categorized into high- and low-risk cohorts based on their risk scores derived from the nine key genes identified through LASSO regression analysis. Survival rates differed significantly between these cohorts, with a p-value of 2e-09 (Fig. 1E). The low-risk cohorts exhibited a markedly higher survival rate compared to the high-risk cohorts. The concordance index (C-index) for the model reached 0.72, indicating good predictive accuracy (Fig. 1F).

Analysis of variances

Analysis of the differences in IRG expression between the tumor and normal cohorts revealed that all nine genes were statistically significant (Fig. 2A). Specifically, CD19, CD79A, CTLA4, ICOS, LAT, DKK1, and PDGFA were significantly upregulated in tumor tissues. This upregulation may be due to the ability of the tumor to attract a substantial influx of immune cells to eliminate cancer cells, leading to the marked upregulation of various immune-related genes. Additionally, tumor cells may alter the expression of surface antigens or inhibit the immune response through the secretion of inhibitory cytokines, leading to the upregulation of immune escape-related genes that are not significantly expressed in normal tissues. In contrast, CHGB and PTX3 were highly expressed in normal tissues to maintain immune homeostasis, but their expression was reduced in tumor tissues, likely due to suppression by TME. To experimentally validate the expression patterns of the nine key immune-related genes identified through our bioinformatic pipeline, we performed quantitative real-time PCR (qPCR) on a set of clinical HNSC specimens and matched adjacent normal tissues. The qPCR results robustly confirmed the differential expression patterns observed in the TCGA cohort analysis. As shown in Supplementary Fig. 1, the mRNA expression levels of the protective genes CD19, CD79A, CTLA4, ICOS, and LAT were significantly upregulated in tumor tissues compared to normal controls (p < 0.05). Conversely, the risk genes CHGB and PTX3 were significantly downregulated in tumor tissues (p < 0.05), consistent with our initial findings that their loss of expression is associated with poor prognosis. Based on the previous COX analysis, DKK1, PDGFA, CHGB, and PTX3 are considered risk genes, and their elevated expression might be associated with immune evasion, promoting tumorigenesis and progression.

Fig. 2.

A IRGs between the tumor and normal cohorts. B, C A heatmapof the top10 functional enrichment terms, each bubble plot represents a cluster. D Comparison of stromal, immune, and ESTIMATE scores between the high-risk and low-risk cohorts. E, F. Violin plots illustrating the distribution of immune cell scores between high- and low-risk cohorts of patients with HNSC using the TCGA database: E Distribution of scores for 28 immune cell types calculated using the ssGSEA method; F Distribution of scores for 22 immune cell types calculated using the CIBERSORT method. *p < 0.05, **p < 0.01, and ***p < 0.001

Enrichment analysis

GO analysis identified various gene functions, including roles in myoarthrosis, modulation of immune processes, cytokine activity, leukocyte-mediated cytotoxicity, immune receptors, natural killer cell-mediated immunity, cytokine receptor activity, and immune synapses (Fig. 2B). Several related pathways were also significantly enriched.

KEGG analysis demonstrated significant enrichment in several related pathways, including those associated with the cytoskeleton of myocytes, cell adhesion molecules, natural killer cell-mediated cytotoxicity, chemokine signaling pathways, interactions between viral proteins and cytokines and their receptors, T-cell receptor signaling pathways, primary immune deficiencies, and allograft rejection (Fig. 2C).

Estimate score

Using the ESTIMATE algorithm to assess stromal and immune cell infiltration in malignant tumor tissues based on expression data, we analyzed stromal and immune cell content, as well as tumor purity, in both the high- and low-risk cohorts (Fig. 2D). Comparisons of the immune, stromal, and ESTIMATE scores (the composite of the first two) revealed significant differences between the cohorts. Notably, both immune and ESTIMATE scores were significantly lower in the high-risk cohorts than in the low-risk cohorts (p < 0.05), indicating that the reduced relative abundance of immune cells in samples was associated with a higher risk of HNSC.

SsGSEA and CIBERSORT analysis

The Wilcoxon test revealed 20 immune cells with significantly different ES between the high- and low-risk cohorts (p < 0.05). Macrophage infiltration was notably lower in the high-risk cohorts, suggesting its role in tumor progression or immune escape. Activated B cells, CD4⁺ T cells, and CD8⁺ T cells were higher in the low-risk cohorts, indicating a stronger antitumor immune response. Significant differences were observed among Tfh, Th1, and Th17 cells. CIBERSORT analysis further identified 15 immune cell types with varying abundances, underscoring the differences in the TME between the risk cohorts. Immune infiltration was compared using both ssGSEA and CIBERSORT (Fig. 2E, F).

Tumor mutation analysis

Somatic gene mutations in tumor samples were analyzed using the maftools package in R [23]. The top20 mutated genes were visualized in both the high-risk (Fig. 3A) and low-risk cohorts (Fig. 3B). The analysis indicated that the proportion of samples with TP53 mutations was significantly higher in the high-risk cohorts than in the low-risk cohorts. Notably, TP53 exhibited the highest mutation frequency among HNSC samples [34].

Fig. 3.

A, B. Evaluation of somatic mutation: A Mutation profile of the top20 mutation genes in high-risk patients; B Mutation profile of the top20 mutation genes in low-risk patients. C GSEA analysis identified key pathways significantly enriched for differential genes between tumor and normal samples. D Survival analysis. K-M survival curves comparing the overall survival between high- and low-risk cohorts

Gene set enrichment analysis (GSEA)

All DEGs (tumor vs. normal) were included in the analysis. This approach revealed several critical immune-related pathways from the MSigDB collection (msigdb_v7.0_GMTs), including adaptive immune response, cellular activation, lymphocyte activation, B-cell activation, positive regulation of the immune response, and immune process regulation (Fig. 3C).

Survival analysis

K-M curves for the high- and low-risk cohorts were plotted separately using the survival package in R with optimal threshold differentiation applied (Fig. 3D). There was a significant difference in survival between high- and low-risk patients (p < 0.05), with the high-risk cohorts exhibiting a poorer prognosis than the low-risk cohorts.

Nomogram

A Cox proportional hazards model was developed to identify independent prognostic factors in patients with HNSC. The analysis confirmed that N stage, age, and risk score (mRNA) were significant independent risk factors, with higher N stage, advanced age, and elevated risk scores correlating with worse patient prognosis. Based on these findings, a nomogram was constructed by integrating N stage, age, and risk score to effectively predict the prognosis of patients with HNSC (Fig. 4A, B).

Fig. 4.

A Cox multivariate regression analysis assessing prognostic factors associated with the survival of patients with HNSC using data from the TCGA database. B A nomogram quantitatively predicting the 1-, 3-, and 5-year survival of patient with HNSC based on the prognostic signature and clinicopathological characteristics of the training set. C Differential analysis of immune checkpoint expression between high- and low-risk cohorts showing significant differences in the expression levels of several immune checkpoints. *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. (D) TF-mRNA regulatory network prediction of 32 TFs associated with nine prognostic genes

Checkpoints

Differential analysis of immune checkpoints was conducted by comparing the high- and low-risk cohorts (Fig. 4C). In the low-risk cohorts, elevated levels of immune checkpoint expression suggested that T cells and other immune cells were continuously activated, with tumor-infiltrating immune cells exhibiting heightened activity. This indicates that the tumor remained under some degree of immune surveillance and control. Low-risk samples typically harbor a greater abundance of tumor-infiltrating immune cells, and the high immune checkpoint expression reflects a natural response to prolonged immune activation [17].

TFs regulatory network

Using miRNet (https://www.mirnet.ca/) and the JASPAR database [35, 36], we predicted 32 TFs associated with nine prognostic genes, each with a degree ≥ 1. The resulting TF-mRNA regulatory network is illustrated (Fig. 4D) and comprises 41 nodes and 55 edges (interactions). CD79A and PDGFA interact with multiple key genes, suggesting that they are core immune regulators of HNSC carcinogenesis.

Drug sensitivity analysis

We examined the sensitivity of different risk cohorts to common anticancer drugs. Differences between the high- and low-risk cohorts, as well as correlations between IC50 values and risk scores, were analyzed. A total of 18 drugs were identified based on the following criteria: (1) significant absolute correlation of IC50 value of a drug with a risk score of ≥ 0.35, and (2) significant differences between the high- and low-risk cohorts.

The results for BI-2536 indicated a positive correlation between CD19 expression levels and the IC50 value (sensitivity). A lower IC50 value indicates higher sensitivity to BI-2536. The IC50 for the high-risk cohorts was significantly lower than that of the low-risk cohorts, implying that BI-2536 may be more beneficial in treating high-risk patients (Fig. 5A, B). One drug with an IC50 < 10 and the highest correlation was selected for presentation; the remaining results are provided in the Supplementary Information. The most effective drugs were BI-2536, SB505124, Gefitinib, Erlotinib, Lapatinib, Acetalax, IRAK4_4710, OSI-027, JAK1_8709, AGI-6780, Mitoxantrone, Wnt-C59, AZD1208, Entinostat, GSK343, AZD5991, LGK974 and AGI-5198.

Fig. 5.

A, B. Drug sensitivity analysis: A Differences in drug sensitivity between high-risk and low-risk cohorts; (B) Positive correlation between CD19 expression levels and the IC50 (sensitivity) of BI-2536. C GeneMANIA analysis of the correlations between the nine immune genes and their neighboring genes. D Validation of immune-related gene expression using the GSE37991 dataset. E Validation of the predictive ability of the nomogram prognostic model using the GSE41613 dataset, showing a concordance index (C-index) of 0.73

Genemania analysis

We used GeneMANIA to analyze the correlations between the nine immune genes and their neighboring genes. CD19, CD79A, ICOS, CTLA4, DKK1, PDGFA, CHGB, LAT, and PTX3 were closely associated with ICOSLG, ASAH1, FOXP3, CD79B, RSPO1, CD22, FGF18, FGF22, CD86, DKK2, CD80, LRP5, BLNK, FGF20, FGF16, SPHK2, PDAP1, LAT2, MED12, and KREMEN2. We identified the topseven pathways associated with cell signaling and functional regulation of the immune system, including the regulation of protein kinase B signaling, antigen receptor-mediated signaling pathways, lymphocyte proliferation, mono-nuclear cell proliferation, leukocyte proliferation, and cell surface interactions. These pathways play crucial roles in immune response, cellular proliferation, and cancer development (Fig. 5C).

GEO validation set

We used the GSE37991 dataset from the GEO database to validate the differential expression of IRGs identified in the TCGA head and neck cancer cohorts. Furthermore, we assessed the predictive performance of the nomogram prognostic model by calculating the C-index. Validation using the GSE41613 dataset confirmed the robustness of the selected IRGs, and the C-index evaluation demonstrated the predictive reliability of the model (C-index = 0.73), supporting its potential utility in the prognosis of patients with head and neck cancer (Fig. 5D, E).

Discussion

Emerging evidence highlights the pivotal role of immune modulation in cancer initiation and progression. The advent of immunotherapy has markedly transformed the landscape of cancer treatment. By targeting immune cells within or outside the TME, immunotherapy enhances the recognition and elimination of cancer cells, theoretically offering higher specificity and lower side effects [37, 38]. ICIs alleviate immune suppression by targeting immune checkpoints, thereby enhancing T-cell-mediated anti-tumor responses. Specifically, PD-1 and CTLA-4 antibodies have shown promise in improving patient prognosis. Significant advancements have been achieved in the treatment of HNSC using PD-1 and CTLA-4 inhibitors [39]. The CheckMate trials demonstrated that combining nivolumab and ipilimumab significantly improves survival and response rates in patients with advanced HNSC [40]. However, heterogeneity in patient responses to immunotherapy poses a significant barrier. Not all patients derive equal benefits, and some experience immune-related side effects such as rashes and colitis [24]. This underscores the growing importance of personalized treatment to optimize patient selection for immunotherapy.

In this study, we integrated immune-related gene data from TCGA, GEO, and ImmPort to construct a prognostic model based on nine key IRGs. Patients were classified into high- and low-risk cohorts based on their risk scores. Comprehensive analyses, including differential expression, GSEA, GO, KEGG, ssGSEA, CIBERSORT, and drug sensitivity evaluations, were conducted to deepen our understanding of the TME in HNSC. Critical genes such as DKK1 and PDGFA were identified as potential immune escape biomarkers, suggesting promising targets for novel therapeutic strategies. Our nine-gene model demonstrated superior predictive capability compared to existing frameworks, achieving an AUC of 0.72 for distinguishing risk cohorts. This enhanced accuracy supports its utility in personalizing HNSC treatment approaches. The prognostic power of our signature may be rooted in its ability to capture key aspects of an anti-tumor immune microenvironment. In this context, our findings align with recent research highlighting Tertiary Lymphoid Structures (TLS) as a critical determinant of favorable prognosis and enhanced immune infiltration in HNSC [41]. It is plausible that our high-risk group, characterized by a poor prognosis, may represent a TLS-low phenotype with diminished effective anti-tumor immunity, whereas the low-risk group may reflect a TLS-high, immune-active state. This connection strengthens the rationale for investigating our nine-gene signature not only as a prognostic tool but also as a potential predictor for immunotherapy response, mirroring the emerging role of TLS in guiding immunotherapeutic strategies. Future investigations should explore these identified genes as therapeutic targets, particularly in combination regimens with PD-1 or CTLA-4 inhibitors to potentiate treatment efficacy.

To elucidate potential interactions among the nine immunoregulatory genes—including protective genes (CD19, CD79A, ICOS, CTLA4) and risk genes (PDGFA, DKK1)—within the head and neck squamous cell carcinoma (HNSC) immune microenvironment, we integrated transcription factor binding predictions (via JASPAR and miRNet) with protein–protein interaction networks (using GeneMANIA). These analyses revealed a plausible mechanistic model in which these genes not only individually modulate key processes, such as lymphocyte activation, B-cell receptor (BCR) signaling, PI3K/AKT pathway, and Wnt signaling, but may also interact through a coordinated BCR–Wnt–Growth Factor axis, collectively shaping immune cell function and promoting tumor immune evasion. Specifically, BCR-related genes likely support antigen recognition and immune memory maintenance; ICOS–ICOSLG and CTLA4–CD80/CD86 interactions fine-tune lymphocyte activation and tolerance; PDGFA and related growth factors modulate tumor proliferation and immune context via PI3K/AKT signaling; while Wnt pathway activity may influence immune cell differentiation and facilitate immune escape [42–44].

Regarding the association between TP53 mutation frequency and immune checkpoint expression in the high-risk group, our analysis indicated that although TP53 mutation was significantly enriched in this group, it did not correlate significantly with key immune checkpoint molecules such as PD-L1. This may be attributable to the predominant role of TP53 in cell cycle and apoptosis regulation, exerting only indirect effects on immune gene expression. Moreover, immune checkpoint levels are influenced by a multitude of factors, including overall tumor mutational burden, inflammatory signals, and immune cell infiltration, which may overshadow the contribution of TP53 status alone. Additionally, bulk RNA-seq captures tissue-level expression and might obscure cell-type-specific regulatory relationships. Thus, the absence of a direct TP53–PD-L1 correlation is biologically consistent and does not undermine the overall immunogenomic characterization of the high-risk group [45].

The complex interplay of immunoregulatory genes within the TME is ultimately reflected in the composition and functional state of infiltrating immune cells. In our analysis of immune cell infiltration, we employed two complementary approaches, ssGSEA and CIBERSORT, to capture this complexity. Notably, the observed discrepancies in macrophage estimates between the methods are not contradictory but rather reflect their distinct algorithmic principles. While ssGSEA, as a gene set enrichment method, assesses the overall transcriptional activity of macrophages as a bulk population, CIBERSORT deconvolutes them into specific subsets (M0, M1, and M2). The opposing trends for macrophages in high- versus low-risk groups identified by CIBERSORT likely underscore a critical shift in macrophage polarization states, a nuance captured by its subset resolution but averaged out in the ssGSEA analysis. This interpretation aligns with the well-documented opposing roles of M1 and M2 macrophages in tumor progression. Therefore, the integration of both methods provides a more holistic view, suggesting not only an alteration in global macrophage presence but also a potential re-polarization within the HNSC immune landscape, a finding that warrants further investigation using spatial or single-cell technologies [46].

There are several limitations to this study. First, validation in larger, independent cohorts — especially within diverse clinical settings—is essential to establish the broad applicability of our immune gene signature. Second, although we employed GEO datasets for external validation, these cohorts were limited to OSCC. This design was consistent with our supplementary experimental validation, which also focused on OSCC; however, additional validation across other HNSC subtypes (laryngeal and nasopharyngeal carcinoma) will be necessary to confirm the generalizability of the model. Third, while DKK1 and PDGFA were implicated in immune evasion, their precise mechanistic roles in HNSC pathogenesis require further elucidation. Finally, the proposed integration of immunotherapy with conventional treatments (chemotherapy, radiotherapy, or targeted therapy) remains theoretical and necessitates empirical validation. Finally, the prognostic analysis of clinical stages was limited by the imbalanced distribution of our cohorts, which predominantly consisted of advanced-stage (III/IV) patients. Future studies with larger, prospectively enrolled cohorts encompassing more early-stage cases are needed to validate the generalizability of our molecular subtypes across the entire disease spectrum.

Conclusion

This study developed a novel prognostic model based on IRGs that effectively predicts the risk and protective factors in patients with HNSC. The model exhibited robust predictive performance, demonstrating significant correlations between immune infiltration levels and overall survival. These findings offer valuable insights into personalized treatment strategies, particularly for optimizing immunotherapy approaches. Despite its reliance on public databases, which poses a notable limitation, further validation through in vitro, in vivo, and clinical studies is essential to confirm its clinical applicability. Moreover, the identification of immune-related biomarkers, such as DKK1 and PDGFA, presents promising avenues for combination therapies aiming at enhancing the efficacy of immunotherapy. The integration of multi-omics data provides a deeper understanding of the TME, potentially guiding the development of more targeted treatments. Future research should explore the mechanisms of these genes in immune escape and validate the applicability of the model in a broader and diverse patient population.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

HNSC

Head and neck squamous cell carcinoma

IRGs

Immune-related gene signatures

TCGA

Transcriptome data from The Cancer Genome Atlas

GEO

Gene Expression Omnibus

GSEA

Gene Set Enrichment Analysis

ssGSEA

Single-sample GSEA

GO

Gene Ontology

KEGG

Kyoto Encyclopedia of Genes and Genomes

OSCC

Oral squamous cell carcinoma

DEGs

Differentially expressed genes

HR

Hazard ratio

TME

Tumor microenvironment

K-M

Kaplan-Meier

ES

Enrichment score

TMB

Tumor mutation burden

TFs

Transcription factors

Author contributions

Z.L.L . and W.A. designed the project, performed data extraction and analysis, and contributed to article editing & reviewing. L.B.G. and B. W. performed data extraction and analysis, data Curation, methodology and original draft. J.M.W., Y. C. and N. L. performed the quality assessment and methodology. All authors read and approved the final manuscript.

Funding

This work was supported by Natural Science Foundation of Shanxi Province (202103021224374, 202403021221256) and the scientific research project of Health Commission of Shanxi Province (2024100).

Data availability

This study analyzed publicly available datasets. These datasets are available from the TCGAand GEO databases. The gene expression datasets analyzed during this study are available in the TCGA databases (https://xenabrowser.net/datapages/) and GEO databases (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE37791、https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE41613).

Declarations

Ethics approval and consent to participate

The analysis of public datasets was conducted using data from The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO), which are de-identified and publicly available. Therefore, this part of the study did not require separate ethical approval. The use of human head and neck squamous cell carcinoma tissue samples for quantitative PCR validation in this study was reviewed and approved by the Medical Research Ethics Committee of Shanxi Provincial Cancer Hospital(Approval Number: 2021005). All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Consent for publication

Not applicable for the public database analysis. All participants (or their legal guardians) who provided tissue samples have consented to the publication of the anonymized research results derived from their samples.

Informed consent

Not applicable for the public database analysis. Informed consent was obtained from all individual participants (or their legal guardians) who provided tissue samples for the experimental validation part of this study.

Competing interests

The authors declare no competing interests.