Integrative single-cell and bulk transcriptomic analysis identifies lesion-associated gene signatures for prognostic stratification and therapeutic guidance in head and neck squamous cell carcinoma Authors: Meihua Xiong, Yong Luo, Shiping Hu Published: 2026 ABSTRACT BackgroundHead and neck squamous cell carcinoma (HNSCC) is a highly heterogeneous malignancy with poor prognosis and variable response to therapy. Effective biomarkers for prognosis and treatment guidance remain limited.MethodsWe performed integrative analysis of single-cell and bulk RNA-seq data to dissect the tumor microenvironment and identify lesion-associated transcriptional changes. Gene modules related to lesion progression were identified using weighted gene co-expression network analysis (WGCNA) and integrated with differential expression and genomic data. Prognostic modeling was conducted using multiple machine learning algorithms, and the best-performing model was validated across independent cohorts. A nomogram combining the risk score and clinical features was developed. Multi-omics analyses were applied to characterize the genomic landscape, immune infiltration, and treatment response associated with the risk groups.ResultsThe final prognostic model effectively stratified patients by survival and was validated in an external cohort. The nomogram showed high predictive accuracy and clinical utility. High-risk tumors were enriched in oncogenic pathways such as hypoxia, glycolysis, and epithelial–mesenchymal transition, and displayed elevated genomic instability. Immune profiling revealed suppressed antitumor activity and increased infiltration of immunosuppressive cells in high-risk patients. Risk score was negatively associated with immunotherapy response and correlated with differential sensitivity to chemotherapy and targeted agents.ConclusionThis study presents a robust prognostic model based on lesion-associated gene expression signatures in HNSCC. It offers valuable tools for individualized prognosis and therapeutic decision-making, with implications for advancing precision oncology in this challenging cancer type. Introduction Head and neck squamous cell carcinoma (HNSCC) represents one of the most common malignancies globally, characterized by high morbidity, mortality, and marked clinical heterogeneity [[1], [2], [3], [4], [5]]. Despite advancements in surgical techniques, radiotherapy, chemotherapy, and immune checkpoint inhibitors, the prognosis of advanced HNSCC remains unsatisfactory, with considerable variation in treatment response and survival outcomes among patients [[6], [7], [8], [9]]. This clinical variability underscores the need for reliable prognostic biomarkers and personalized therapeutic strategies. The complexity of HNSCC arises largely from its extensive molecular and cellular heterogeneity, which poses significant challenges for accurate prognosis and effective therapeutic decision-making [[10], [11], [12]]. Recent advances in single-cell RNA sequencing (scRNA-seq) have enabled unprecedented insights into tumor microenvironment (TME) heterogeneity, revealing distinct cellular compositions and functional states within tumors [[13], [14], [15]]. However, comprehensive analyses integrating single-cell profiles, molecular pathways, and clinical outcomes are scarce in HNSCC. Specifically, the molecular alterations associated with lesion-associated changes (LAC)—an intermediate state critical to malignant transformation—and their prognostic implications remain poorly characterized. Weighted gene co-expression network analysis (WGCNA) has emerged as a powerful method to identify clinically relevant gene modules and hub genes from high-dimensional transcriptomic data [16,17]. Previous studies applying WGCNA have uncovered gene networks associated with cancer progression and immune modulation across various tumor types [[18], [19], [20]]. Nevertheless, the integration of WGCNA-derived modules with survival modeling and clinical stratification has not been thoroughly investigated in the context of HNSCC, particularly regarding their value in predicting immunotherapy response and drug sensitivity. In this study, we systematically addressed these gaps by integrating single-cell and bulk transcriptomic data to dissect the cellular landscape and transcriptional dynamics of HNSCC. Initially, single-cell RNA-seq was employed to characterize cellular heterogeneity and define immune-related gene signatures significantly activated within malignant tissues. Next, we identified LAC-associated transcriptional modules via WGCNA and integrated these findings with differential expression and genomic alteration analyses to prioritize candidate genes. Using advanced machine learning algorithms, we developed and validated a robust prognostic risk model across multiple independent cohorts. Subsequently, we evaluated the clinical utility of the risk signature through comprehensive subgroup analyses and constructed an individualized prognostic nomogram, demonstrating superior accuracy over traditional clinical parameters. Further functional exploration of the identified gene signature revealed enrichment of hallmark pathways related to hypoxia, glycolysis, epithelial–mesenchymal transition (EMT), and angiogenesis, implicating a biologically aggressive and immunosuppressive phenotype in high-risk patients. We additionally delineated the genomic landscape associated with distinct risk groups and explored associations with immune checkpoint blockade response and drug sensitivity, thereby providing therapeutic insights that warrant further validation. Collectively, our integrative analysis enhances understanding of the molecular mechanisms underlying HNSCC progression and provides a candidate framework for individualized prognostic assessment and treatment stratification in HNSCC. Methods Single-cell RNA-seq data processing Single-cell RNA sequencing (scRNA-seq) data from HNSCC samples (GEO accession: GSE181919) were analyzed using the Scanpy python package. Quality control involved filtering out cells with high mitochondrial content or extremely low/high gene counts. Batch effects were corrected using the Harmony algorithm. Dimensionality reduction was performed using UMAP, and clustering was conducted using the Leiden algorithm. Cell type annotation was conducted based on canonical lineage markers. Bulk RNA-seq data acquisition and preprocessing Bulk RNA-seq data were retrieved from the TCGA-HNSC cohort (used as the training set) and the GSE41613 dataset (used as the validation set). Gene expression data were normalized, and differential expression analysis was conducted using the limma package. Genes with adjusted p-values <0.05 and |logFC| > 0.5 were selected as DEGs. Co-expression network construction and module identification Weighted gene co-expression network analysis (WGCNA) was applied to the TCGA dataset to identify gene modules correlated with lesion-associated changes (LAC). A soft-thresholding power was selected, and adjacency matrices were constructed to define topological overlap. Modules were identified using dynamic tree cutting and correlated with clinical traits. Identification and functional annotation of LAC-associated genes Genes in LAC-associated modules were intersected with DEGs to identify candidate LAC-related genes. These genes were further filtered based on fold-change and module membership. Functional enrichment was performed using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis. Construction and validation of the prognostic model Univariate Cox regression analysis was performed to identify prognosis-associated genes. Multiple machine learning algorithms, including Lasso, Ridge, stepwise Cox regression, CoxBoost, random survival forest (RSF), GBM, and SuperPC, were used to build prognostic models. These models were compared within a unified framework to reduce dependence on a single modeling strategy, and the final model was selected based on cross-cohort concordance index (C-index) performance together with interpretability. Risk scores were calculated from model coefficients, and survival analysis was conducted in both training and validation cohorts. Clinical correlation and subgroup analysis Associations between risk scores and clinical parameters (e.g., age, gender, T stage, N stage, AJCC stage) were evaluated. Subgroup survival analyses were performed across different clinical strata using Kaplan–Meier methods. Nomogram development and performance evaluation Multivariate Cox regression was applied to assess the independent prognostic value of the risk score. A nomogram was constructed integrating the risk score with clinical variables to predict 1-, 3-, and 5-year survival. Predictive accuracy was evaluated using ROC curves, calibration plots, decision curve analysis (DCA), and time-dependent C-index. Functional enrichment analysis Gene Set Variation Analysis (GSVA) and Gene Set Enrichment Analysis (GSEA) were performed using hallmark gene sets to identify differentially enriched pathways between risk groups. Survival analysis was used to evaluate the prognostic impact of key pathways. Genomic alteration analysis Tumor mutational burden (TMB), mutant-allele tumor heterogeneity (MATH), and somatic mutations were analyzed to assess genomic instability. Copy number variations (CNVs) were evaluated and compared between risk groups. Co-occurrence and mutual exclusivity analyses were performed. Immune landscape characterization Immune infiltration scores were estimated using ESTIMATE, ssGSEA, and CIBERSORT algorithms. Differences in immune cell composition and immune pathway activity between risk groups were analyzed. Correlations between immune features and the risk model were assessed. Immunotherapy response and drug sensitivity prediction Immune checkpoint gene expression was analyzed to estimate immune responsiveness. Immunotherapy response predictions were validated using the IMvigor210 dataset. Drug sensitivity prediction was performed using pharmacogenomic data, with correlation analysis conducted between risk scores and drug IC50 values. Statistical analysis All statistical analyses were performed using R and Python software. Comparisons between groups were assessed using Wilcoxon rank-sum tests, chi-square tests, or Fisher's exact tests where appropriate. Survival differences were evaluated using Kaplan–Meier curves and the log-rank test. Cox proportional hazards regression was used for univariate and multivariate analyses. For analyses involving multiple comparisons, p values were adjusted using the Benjamini–Hochberg method where appropriate, and both nominal p values and adjusted values were interpreted cautiously. Unless otherwise specified, p values < 0.05 were considered statistically significant. Results Single-cell transcriptomic profiling reveals cellular heterogeneity and functional activation landscape To dissect the cellular composition and functional characteristics of the HNSCC microenvironment, we performed integrated analysis of single-cell RNA sequencing data comprising a total of 50,733 cells. UMAP-based dimensionality reduction revealed that cells were well mixed across different samples (Fig. 1A) and patient sources (Fig. 1B), suggesting successful batch correction and data integration. Clustering using the Leiden algorithm identified 24 distinct cellular clusters (Fig. 1C), which were subsequently annotated into major cell types based on canonical markers, including epithelial cells, plasma cells, endothelial cells, fibroblasts, macrophages, T cells, B cells, mast cells, dendritic cells, and myocytes (Fig. 1D). Marker gene analysis confirmed the accuracy of cell type annotation: EPCAM was highly expressed in epithelial cells, CD3D in T cells, MS4A1 in B cells, CD68 and CD86 in macrophages, and PECAM1 in endothelial cells (Fig. 1E). These classical lineage markers validated the biological identity of each cluster. We next assessed the pathway activity of a predefined gene signature using AUCell scoring. Violin plots revealed significant differences in AUCell scores across cell types, with macrophages, dendritic cells, and T cells exhibiting the highest activation levels (Fig. 1F, p < 0.0001). Mapping the AUCell scores back onto the UMAP embedding showed that pathway activity was spatially concentrated in specific cellular populations (Fig. 1G). Finally, we compared AUCell scores across different tissue types, including normal (NL), low-grade preinvasive (LP), cancerous (CA), and lymph node metastases (LN). Several immune-related cell types, particularly macrophages and T cells, displayed elevated AUCell scores in the cancer (CA) group, indicating enhanced pathway activation under malignant conditions (Fig. 1H). Together, these results demonstrate the successful integration and annotation of scRNA-seq data, and reveal that the defined gene signature is preferentially enriched in immune-related populations, particularly within cancerous tissues—highlighting the relevance of this signature in tumor immune modulation.Fig. 1Single-cell transcriptomic profiling of HNSCC reveals immune-related activation. (A–B) UMAP plots showing effective data integration across samples and patients.(C) Identification of 24 cell clusters by the Leiden algorithm.(D) Annotation of major cell types using canonical markers.(E) Dot plot of representative genes validating cell type identity.(F) Violin plots showing higher AUCell scores in macrophages, dendritic cells, and T cells.(G) Spatial distribution of AUCell scores on UMAP highlights immune activation regions.(H) AUCell scores across tissue types, with macrophages and T cells showing increased activation in cancer tissues (****p < 0.0001). Single-cell transcriptomic profiling of HNSCC reveals immune-related activation. (A–B) UMAP plots showing effective data integration across samples and patients.(C) Identification of 24 cell clusters by the Leiden algorithm.(D) Annotation of major cell types using canonical markers.(E) Dot plot of representative genes validating cell type identity.(F) Violin plots showing higher AUCell scores in macrophages, dendritic cells, and T cells.(G) Spatial distribution of AUCell scores on UMAP highlights immune activation regions.(H) AUCell scores across tissue types, with macrophages and T cells showing increased activation in cancer tissues (****p < 0.0001). Identification of LAC-associated modules and key risk genes through integrative transcriptomic analysis To elucidate the transcriptional landscape underlying lesion-associated changes (LAC) in HNSCC, we first performed weighted gene co-expression network analysis (WGCNA). Hierarchical clustering of samples and trait heatmaps revealed clear segregation between LAC and non-LAC phenotypes (Fig. 2A). Gene modules were identified using dynamic tree cutting, resulting in multiple color-coded modules (Fig. 2B). Module–trait correlation analysis indicated that the red module was most strongly associated with LAC (cor = 0.68, p = 2e–69), followed by the yellow module (cor = 0.62, p = 6e–56) (Fig. 2C). These modules were prioritized for further investigation due to their robust correlation with the disease phenotype. Differential expression analysis identified 10,298 DEGs between LAC and non-LAC samples (Fig. 2D), including significantly upregulated genes such as MMP1, MMP9, COL1A1, and LAMC2, all of which are known to participate in extracellular matrix degradation and tumor invasion. The intersection of DEGs and WGCNA module genes yielded 927 overlapping candidates (Fig. 2E), reflecting transcriptional reprogramming within co-expressed gene networks. To refine the gene list, we applied additional filtering based on expression fold change and module membership, resulting in a final set of 602 high-confidence genes for downstream functional analysis. Gene Ontology enrichment revealed that these genes were enriched in processes such as RNA splicing, nuclear speck (Fig. 2F). KEGG pathway analysis further highlighted significant enrichment in spliceosome, cell cycle, apoptosis (Fig. 2G), indicating that transcriptional and post-transcriptional regulation are hallmarks of LAC. We next constructed a correlation network among hub genes, integrating DEG status and copy number variation (CNV) profiles. This revealed a panel of candidate risk genes (e.g., GAPDH, FXR1, MAP2K1) with strong positive correlation to LAC, and protective genes (e.g., BTG1, ILF3, TRA2A) negatively associated with disease status (Fig. 2H). CNV analysis confirmed frequent amplifications in several risk genes (e.g., FXR1, DDX18) and deletions in protective genes (e.g., BTG1, TRA2A) (Fig. 2I), supporting their potential role in disease progression. Together, these findings indicate that LAC is driven by complex transcriptional remodeling, involving RNA processing, metabolic stress, and structural genomic alterations, and provide a focused set of candidate genes for prognostic modeling.Fig. 2Identification of LAC-related modules and candidate genes. (A) Sample clustering annotated by LAC status.(B) WGCNA dendrogram showing gene modules.(C) Module–trait correlations, with red and yellow modules most associated with LAC.(D) Volcano plot of DEGs between LAC and non-LAC groups.(E) Venn diagram of DEGs and module genes (927 overlaps).(F) GO enrichment of 602 refined genes highlights RNA splicing-related terms.(G) KEGG analysis shows enrichment in spliceosome, cell cycle, and apoptosis pathways.(H) Correlation network of hub genes with DEG and CNV status.(I) CNV alterations in key genes: amplifications in risk genes (e.g., FXR1), deletions in protective genes (e.g., BTG1). Identification of LAC-related modules and candidate genes. (A) Sample clustering annotated by LAC status.(B) WGCNA dendrogram showing gene modules.(C) Module–trait correlations, with red and yellow modules most associated with LAC.(D) Volcano plot of DEGs between LAC and non-LAC groups.(E) Venn diagram of DEGs and module genes (927 overlaps).(F) GO enrichment of 602 refined genes highlights RNA splicing-related terms.(G) KEGG analysis shows enrichment in spliceosome, cell cycle, and apoptosis pathways.(H) Correlation network of hub genes with DEG and CNV status.(I) CNV alterations in key genes: amplifications in risk genes (e.g., FXR1), deletions in protective genes (e.g., BTG1). Construction and validation of a prognostic model based on LAC-associated gene signatures To assess the prognostic value of LAC-associated genes in HNSCC, we systematically constructed risk prediction models using a panel of machine learning algorithms. A total of 101 algorithm combinations were evaluated, including Lasso, Ridge, stepwise Cox regression (forward, backward, both), CoxBoost, RSF, GBM, and SuperPC. The predictive performance of each model was quantified using the concordance index (C-index) across the TCGA-HNSC training set and the GSE41613 validation cohort (Fig. 3A). Among them, the StepCox (forward) + Ridge combination achieved the highest C-index (0.7) while retaining comparatively strong interpretability, and was therefore selected as the final model. The risk score of each patient was calculated based on the linear combination of gene expression values weighted by their regression coefficients (Fig. 3B). Genes such as PGK1, SPCS3, and MAP2K1 showed positive coefficients, suggesting association with poor prognosis, while genes like LAT, ELAVL1, and PRKCH were assigned negative coefficients. Univariate Cox regression analysis further confirmed that several genes were significantly associated with overall survival (p < 0.05). High-risk genes (HR > 1) included PGK1, MAP2K1, and KPNA2, whereas low-risk (protective) genes (HR < 1) included BTG1, LAT, and TRA2A (Fig. 3C). Kaplan–Meier survival curves showed that the risk score effectively stratified patients into prognostically distinct groups. In the TCGA-HNSC cohort, high-risk patients had significantly worse overall survival compared to low-risk patients (p < 0.0001; Fig. 3D). This result was independently validated in the GSE41613 dataset (p = 0.025; Fig. 3E), supporting the reproducibility of the model within the available cohorts. Time-dependent ROC curve analysis demonstrated favorable predictive performance. In the TCGA cohort, the AUCs for predicting 1-, 3-, and 5-year survival were 0.75, 0.80, and 0.82, respectively (Fig. 3F). In the GSE41613 validation set, the AUCs were 0.69, 0.64, and 0.64 (Fig. 3G), indicating that the model retained reasonable discriminatory power across datasets. Together, these results support the prognostic relevance of LAC-associated gene expression and suggest that the integrated machine-learning model may be useful for survival risk stratification in HNSCC, although broader validation is still needed.Fig. 3Construction and validation of a LAC-based prognostic model. (A) Heatmap of C-index values for 101 machine learning combinations; Lasso + StepCox selected for highest performance. (B) Bar plot showing model gene coefficients. (C) Forest plot of univariate Cox analysis for 37 genes. (D–E) Kaplan–Meier survival curves stratified by risk score in TCGA and GSE41613 cohorts. (F–G) Time-dependent ROC curves for 1-, 3-, and 5-year survival prediction in training and validation sets. Construction and validation of a LAC-based prognostic model. (A) Heatmap of C-index values for 101 machine learning combinations; Lasso + StepCox selected for highest performance. (B) Bar plot showing model gene coefficients. (C) Forest plot of univariate Cox analysis for 37 genes. (D–E) Kaplan–Meier survival curves stratified by risk score in TCGA and GSE41613 cohorts. (F–G) Time-dependent ROC curves for 1-, 3-, and 5-year survival prediction in training and validation sets. Clinical relevance and stratified survival analysis of the prognostic signature To further explore the clinical relevance of the constructed risk model, we assessed its association with patient characteristics in the TCGA-HNSC cohort. Heatmap visualization showed distinct gene expression patterns between high- and low-risk groups, alongside differences in T stage, N stage, age, and survival status (Fig. 4A). Donut plots illustrated the distribution of clinical variables between risk groups. The risk score was significantly associated with age (p = 0.0208), T stage (p = 0.0554), and N stage (p = 0.0014), while no significant differences were found for M stage (p = 1), overall AJCC stage (p = 0.1104), or gender (p = 0.4168) (Fig. 4B). Violin plots further confirmed that patients with more advanced T and N stage had significantly higher risk scores (Fig. 4C and D), and that older patients also exhibited higher risk levels (Fig. 4E). Although the overall difference in risk scores across AJCC stages was not significant, pairwise comparison between Stage I and Stage IV revealed a significant difference (p = 0.017), indicating some prognostic stratification at extreme stages (Fig. 4F). Kaplan–Meier analysis within clinical subgroups showed that the risk model maintained its prognostic power across nearly all stratifications. High-risk patients had significantly worse survival in both advanced T stage (T3–4) and early T stage (T1–2) subgroups (Fig. 4G and H), as well as across N0–N1 and N2–N3 nodal status groups (Fig. 4I and J). The model also robustly predicted outcomes in patients without distant metastasis (M0; Fig. 4K) and in those with advanced disease (Stage III–IV; Fig. 4L), with all comparisons yielding p < 0.0001. These findings confirm that the risk score not only reflects key clinical features, but also serves as a robust prognostic indicator across diverse clinical contexts in HNSCC.Fig. 4Clinical relevance and subgroup validation of the prognostic model. (A) Heatmap showing risk score, gene expression, and clinical traits. (B) Donut charts comparing clinical features between high- and low-risk groups. (C–F) Violin plots of risk scores by T stage, N stage, age, and AJCC stage. (G–L) Kaplan–Meier survival analysis stratified by T, N, M stages and AJCC stages, confirming model robustness across subgroups. Clinical relevance and subgroup validation of the prognostic model. (A) Heatmap showing risk score, gene expression, and clinical traits. (B) Donut charts comparing clinical features between high- and low-risk groups. (C–F) Violin plots of risk scores by T stage, N stage, age, and AJCC stage. (G–L) Kaplan–Meier survival analysis stratified by T, N, M stages and AJCC stages, confirming model robustness across subgroups. Construction of a nomogram and assessment of predictive accuracy for clinical application To assess whether the risk score serves as an independent prognostic factor, we first performed univariate and multivariate Cox regression analyses. In univariate analysis (Fig. 5A), age, gender, T stage, N stage, and risk score were significantly associated with overall survival (p < 0.05). In multivariate analysis (Fig. 5B), the risk score remained statistically significant (HR = 2.0231, p < 0.001), suggesting it is an independent prognostic indicator. A nomogram was subsequently constructed by integrating risk score, T stage, and N stage to predict 1-, 3-, and 5-year overall survival probabilities (Fig. 5C). Each variable was assigned a corresponding point score, and the total score was mapped to individual survival probability. Moreover, the nomogram exhibited superior discrimination ability, with time-dependent AUC values reaching 0.79, 0.81, and 0.78 for 1-, 3-, and 5-year survival, respectively (Fig. 5E), outperforming traditional clinical indicators. Calibration curves showed strong concordance between predicted and actual survival outcomes across all timepoints (Fig. 5D), confirming the nomogram's predictive accuracy. Decision curve analysis (DCA) indicated that the nomogram provided the greatest net clinical benefit over a wide range of threshold probabilities when compared to individual clinical variables (Fig. 5F). Furthermore, time-dependent C-index analysis demonstrated that the nomogram maintained superior predictive accuracy over T stage, N stage, overall stage, and risk score alone throughout the follow-up period (Fig. 5G). Together, these results support the clinical utility of the nomogram as a robust and individualized tool for survival prediction in HNSCC.Fig. 5Development and evaluation of a prognostic nomogram. (A–B) Univariate and multivariate Cox analyses showing risk score as independent prognostic factor. (C) Nomogram integrating risk score, T stage, and N stage. (D) Calibration plots for 1-, 3-, and 5-year survival prediction. (E) Time-dependent AUC curves demonstrating superior accuracy. (F) Decision curve analysis (DCA) comparing net clinical benefit. (G) C-index comparison showing nomogram outperforms other clinical variables. Development and evaluation of a prognostic nomogram. (A–B) Univariate and multivariate Cox analyses showing risk score as independent prognostic factor. (C) Nomogram integrating risk score, T stage, and N stage. (D) Calibration plots for 1-, 3-, and 5-year survival prediction. (E) Time-dependent AUC curves demonstrating superior accuracy. (F) Decision curve analysis (DCA) comparing net clinical benefit. (G) C-index comparison showing nomogram outperforms other clinical variables. Functional characterization of the risk signature through hallmark pathway enrichment To gain mechanistic insights into the biological processes underlying the risk signature, we performed Gene Set Variation Analysis (GSVA) and Gene Set Enrichment Analysis (GSEA) using hallmark gene sets. GSVA revealed that multiple oncogenic pathways, including hypoxia and glycolysis, were significantly upregulated in the high-risk group compared to the low-risk group (Fig. 6A). These pathways are commonly associated with tumor progression, immune evasion, and metabolic reprogramming. GSEA analysis further confirmed that high-risk patients showed significant enrichment in pathways related to epithelial–mesenchymal transition (EMT), angiogenesis, hypoxia, and allograft rejection (Fig. 6B and C), suggesting an immunosuppressive and pro-metastatic microenvironment. In contrast, the low-risk group exhibited enrichment in bile acid metabolism, peroxisome activity, and other metabolic processes associated with differentiated cell states and less aggressive phenotypes.Fig. 6Functional enrichment of risk-related pathways. (A) GSVA scores indicating upregulation of hypoxia and glycolysis in high-risk group. (B–C) GSEA plots showing EMT, angiogenesis, and immune suppression pathways enriched in high-risk tumors. (D–F) Kaplan–Meier curves stratified by hallmark gene sets (hypoxia, glycolysis, angiogenesis), confirming their prognostic value. Functional enrichment of risk-related pathways. (A) GSVA scores indicating upregulation of hypoxia and glycolysis in high-risk group. (B–C) GSEA plots showing EMT, angiogenesis, and immune suppression pathways enriched in high-risk tumors. (D–F) Kaplan–Meier curves stratified by hallmark gene sets (hypoxia, glycolysis, angiogenesis), confirming their prognostic value. Survival analysis stratified by hallmark gene set enrichment levels highlighted that patients with high expression of HALLMARK_HYPOXIA and HALLMARK_GLYCOLYSIS had significantly worse prognosis (p = 0.00083 and p < 0.0001, respectively; Fig. 6D and E), reinforcing the relevance of metabolic dysregulation in poor outcomes. Notably, HALLMARK_ANGIOGENESIS also predicted poor survival (p = 0.049; Fig. 6F), underscoring the involvement of tumor vascularization in high-risk tumors. Collectively, these findings demonstrate that the risk score not only predicts clinical outcomes but also reflects key tumor-promoting pathways, particularly those related to inflammation, metabolism, and immune modulation. Genomic landscape and tumor heterogeneity associated with the risk signature To explore the genomic characteristics associated with different risk groups, we first evaluated tumor mutational burden (TMB) and mutant-allele tumor heterogeneity (MATH) scores. High-risk patients exhibited significantly elevated MATH scores compared to the low-risk group (p = 0.00014; Fig. 7A), indicating greater intratumoral heterogeneity. Survival analysis further revealed that patients with high MATH scores had significantly worse overall survival than those with low MATH (p = 0.00064; Fig. 7B). Stratified analysis combining MATH and risk score classification showed that the H-MATH + high-risk subgroup had the poorest prognosis, whereas the L-MATH + low-risk subgroup had the best outcomes (p < 0.0001; Fig. 7C). Somatic mutation analysis demonstrated distinct mutation profiles between risk groups. In the high-risk cohort, the most frequently mutated genes included TP53 (80%), TTN (37%), CDKN2A (24%), and FAT1 (24%), while in the low-risk group, TP53 (60%), TTN (38%), and PIK3CA (20%) were most prominent (Fig. 7D and E). To further characterize the genomic alterations associated with different risk groups, we investigated somatic mutation co-occurrence and copy number variation (CNV) profiles. A co-occurrence matrix of the top 20 mutated genes revealed distinct patterns between high- and low-risk groups (Fig. 7F). Notably, strong co-mutation relationships were observed among high-risk-enriched genes such as TP53, CDKN2A, FAT1, and NOTCH1, while the low-risk group displayed more scattered and weaker co-occurrence signals. Mutually exclusive relationships were less frequent but also observed, suggesting potential functional redundancy among some driver events. In addition, CNV analysis showed that PIK3CA, CSMD3, and TTN exhibited high frequencies of amplification (GAIN), especially in the high-risk group, whereas CDKN2A and LRP1B were frequently deleted (LOSS) (Fig. 7G). These CNV alterations overlapped with several key mutated genes, suggesting a dual-hit mechanism contributing to pathway dysregulation in tumor progression. The convergence of co-occurring mutations and CNVs in oncogenes (e.g., PIK3CA, NOTCH1) and tumor suppressors (e.g., CDKN2A, TP53) underscores the genomic instability and aggressive phenotype of high-risk tumors. Together, these findings highlight the profound genetic heterogeneity in cervical cancer and reveal that high-risk patients harbor more extensive genomic alterations, including both mutations and structural variants, which may contribute to immune escape and treatment resistance.Fig. 7Genomic landscape and heterogeneity across risk groups. (A) Box plot of MATH scores between groups. (B–C) Survival curves based on MATH score and its combination with risk score. (D–E) Oncoplots showing top mutations in high-vs. low-risk groups. (F) Mutation co-occurrence matrices. (G) CNV profile highlighting high-risk-enriched amplifications and deletions. Genomic landscape and heterogeneity across risk groups. (A) Box plot of MATH scores between groups. (B–C) Survival curves based on MATH score and its combination with risk score. (D–E) Oncoplots showing top mutations in high-vs. low-risk groups. (F) Mutation co-occurrence matrices. (G) CNV profile highlighting high-risk-enriched amplifications and deletions. Distinct tumor immune microenvironment features between risk groups To evaluate the immune contexture associated with the risk signature, we first analyzed immune, stromal, and ESTIMATE scores derived from transcriptomic data. The high-risk group showed significantly elevated stromal scores (p = 0.046), immune scores (p = 0.0016), and total ESTIMATE scores (p = 0.37) compared to the low-risk group (Fig. 8A), suggesting a more complex and immune-infiltrated tumor microenvironment. We next assessed immune pathway activity using KEGG-based enrichment analysis. Immune-related pathways such as antigen processing and presentation, Fc gamma R-mediated phagocytosis, B cell receptor signaling, T cell receptor signaling, and natural killer cell-mediated cytotoxicity were significantly upregulated in the low-risk group (Fig. 8B), indicating more active antitumor immune responses. CIBERSORT deconvolution analysis revealed significant differences in immune cell infiltration patterns between groups. The low-risk group was characterized by increased infiltration of CD8+ T cells, activated CD4+ memory T cells (all p < 0.05), whereas the high-risk group exhibited higher proportions of M0 macrophages and M2 macrophages, indicative of an immunosuppressive microenvironment (Fig. 8C and D). To further dissect the relationship between individual model genes and immune cell infiltration, we performed correlation analysis between the 37 signature genes and 22 immune cell subtypes. The results revealed that several genes, including LAT was positively correlated with the infiltration of CD8+ T cells, M1 macrophages (Fig. 8E). These findings further support the dual immunological roles embedded within the risk model—reflecting both antitumor and immunosuppressive cellular programs. Furthermore, correlation analysis between the risk score and immune cell infiltration revealed that T follicular helper cells, and regulatory T cells (Tregs) were strongly negatively correlated with the risk score, suggesting a more immunoreactive phenotype in low-risk tumors (Fig. 8F). In contrast, M0 macrophages, resting CD4+ memory T cells, and activated mast cells were positively correlated with the risk score, reflecting a more immunosuppressive or less differentiated immune context in high-risk tumors. These findings support the immunological relevance of the risk model and its potential for guiding immune-based stratification.Fig. 8Immune microenvironment features in high- and low-risk tumors. (A) Immune, stromal, and ESTIMATE scores between groups. (B) KEGG pathway enrichment showing immune pathway upregulation in low-risk group. (C–D) Immune cell infiltration patterns by CIBERSORT. (E) Correlation heatmap between model genes and immune cell types. (F) Spearman correlations between risk score and immune cell subtypes. Immune microenvironment features in high- and low-risk tumors. (A) Immune, stromal, and ESTIMATE scores between groups. (B) KEGG pathway enrichment showing immune pathway upregulation in low-risk group. (C–D) Immune cell infiltration patterns by CIBERSORT. (E) Correlation heatmap between model genes and immune cell types. (F) Spearman correlations between risk score and immune cell subtypes. Prediction of immunotherapy response and drug sensitivity based on the risk model To evaluate the predictive capacity of the risk model for immunotherapy response, we analyzed immune checkpoint gene expression and clinical response data. Key immune checkpoint molecules including PDCD1, CTLA4, LAG3, TIGIT, CD27, TNFRSF25, and TNFRSF18 were significantly overexpressed in the low-risk group (all p < 0.01; Fig. 9A), suggesting a more active immune context and potential responsiveness to immune checkpoint blockade (ICB) therapy. Using the IMvigor210 cohort, we further validated the association between risk score and immunotherapy efficacy. A significantly lower proportion of CR/PR (complete or partial response) cases was observed in the high-risk group compared to the low-risk group (Fig. 9B). In addition, patients with CR or PR exhibited significantly lower risk scores than those with PD (progressive disease) or SD (stable disease) (p < 0.0001; Fig. 9C and D), indicating a strong inverse correlation between risk score and treatment responsiveness. Survival analysis further confirmed that low-risk patients had markedly improved overall survival under anti-PD-L1 treatment compared to high-risk patients (p < 0.0001; Fig. 9E), highlighting the clinical relevance of the risk model in predicting immunotherapy benefit. Additionally, we assessed drug sensitivity using pharmacogenomic profiling. Negative correlations were observed between risk score and sensitivity to SB505124 (TGF-β inhibitor, R = −0.22) and Staurosporine (broad-spectrum kinase inhibitor, R = −0.15), indicating that low-risk patients may be more responsive to these agents (Fig. 9F and G). Conversely, cisplatin and 5-Fluorouracil sensitivity were positively associated with risk score (R = 0.21 and 0.22, respectively), suggesting that high-risk patients may be less responsive to these conventional chemotherapies (Fig. 9H and I).Fig. 9Prediction of immunotherapy and drug sensitivity. (A) Expression of immune checkpoint genes across risk groups. (B–D) Association between risk score and ICB response in IMvigor210 cohort. (E) Survival analysis of high-vs. low-risk patients under anti-PD-L1 treatment. (F–I) Correlation between risk score and sensitivity to SB505124, Staurosporine, cisplatin, and 5-FU. Prediction of immunotherapy and drug sensitivity. (A) Expression of immune checkpoint genes across risk groups. (B–D) Association between risk score and ICB response in IMvigor210 cohort. (E) Survival analysis of high-vs. low-risk patients under anti-PD-L1 treatment. (F–I) Correlation between risk score and sensitivity to SB505124, Staurosporine, cisplatin, and 5-FU. Collectively, these findings demonstrate that the prognostic model not only predicts survival outcomes but also provides valuable guidance for individualized therapy selection, including immunotherapy and chemotherapy strategies. Discussion In this study, we conducted an integrative analysis combining single-cell and bulk transcriptomic data to comprehensively dissect the heterogeneity, prognostic biomarkers, and therapeutic implications in head and neck squamous cell carcinoma (HNSCC). Through the identification of lesion-associated changes (LAC)-related gene modules, construction of a robust prognostic risk model, and detailed characterization of the immune microenvironment and genomic alterations, our findings provide novel insights into HNSCC biology and therapeutic strategies. Single-cell RNA sequencing (scRNA-seq) analysis revealed substantial intratumoral heterogeneity and distinct immune activation profiles within malignant lesions. By annotating major cellular populations such as epithelial cells, fibroblasts, macrophages, T cells, and dendritic cells, we highlighted immune-related functional activation predominantly within macrophages and T cells in cancerous tissues, underscoring their potential roles in tumor immune modulation. These results align with recent scRNA-seq studies demonstrating immune infiltration diversity and functional specialization in multiple cancer types, reinforcing the relevance of targeted immunotherapeutic approaches in HNSCC. Our study further employed weighted gene co-expression network analysis (WGCNA) to identify LAC-specific transcriptional modules significantly correlated with tumor progression. The red and yellow modules emerged as crucial networks, with prominent enrichment in pathways such as RNA splicing, cell cycle regulation, apoptosis, and extracellular matrix remodeling. These findings echo previous reports implicating RNA-processing alterations and extracellular matrix dynamics in tumorigenesis and metastasis. Integrating differential expression and copy number variation data, we pinpointed key risk-associated genes (e.g., GAPDH, FXR1, MAP2K1) and favorable prognostic genes (e.g., BTG1, ILF3, TRA2A) that may serve as potential targets for therapeutic intervention and prognostic evaluation. Leveraging advanced machine learning algorithms, we developed a prognostic signature that effectively stratified HNSCC patients into distinct risk groups. Notably, the final Lasso + Stepwise Cox model demonstrated robust performance (highest average C-index = 0.7), validated across independent cohorts (TCGA-HNSC and GSE41613). Our signature significantly outperformed traditional clinical variables, accurately predicting survival outcomes at multiple timepoints. Moreover, clinical subgroup analyses confirmed the broad applicability and stability of the model across different stages and nodal statuses, reflecting its potential utility in personalized clinical management. To translate these findings into clinical practice, we constructed an individualized nomogram integrating risk scores with clinical parameters (T and N stages). The nomogram exhibited excellent predictive accuracy (AUC values consistently around 0.79) and superior clinical benefit over conventional indicators, supporting its practical application in risk assessment and therapeutic decision-making. Recent studies have similarly validated nomograms as effective tools for individualized prognosis in various cancers [21], suggesting broad translational relevance. Mechanistically, functional enrichment analyses identified hallmark pathways, including hypoxia, glycolysis, angiogenesis, and epithelial–mesenchymal transition (EMT), preferentially activated in high-risk tumors. Such pathways have been widely implicated in aggressive tumor behavior, metabolic reprogramming, and immune evasion [[22], [23], [24]]. Our survival analyses further emphasized their prognostic significance, reinforcing the biological robustness of our risk stratification model. The genomic landscape characterization further highlighted the distinct genetic instability and heterogeneity profiles between risk groups, evidenced by differences in tumor mutational burden (TMB), mutant-allele tumor heterogeneity (MATH) scores, and mutation/CNV co-occurrence patterns involving pivotal cancer driver genes (e.g., TP53, PIK3CA, CDKN2A). Such genomic complexity has previously been linked to poor prognosis and therapeutic resistance in cancers, and our findings underscore the potential therapeutic vulnerabilities associated with genomic instability in high-risk HNSCC patients. Importantly, our risk model demonstrated predictive value for immunotherapy and chemotherapy response, with significant associations between immune checkpoint gene expression and clinical outcomes under anti-PD-L1 treatment (IMvigor210). Notably, high-risk patients were predicted to benefit less from conventional chemotherapies such as cisplatin and 5-Fluorouracil, aligning with clinical observations of treatment resistance. These findings suggest that our prognostic model could aid clinicians in tailoring immunotherapy or chemotherapy strategies, optimizing therapeutic efficacy and patient outcomes. However, these findings should be interpreted as hypothesis-generating rather than practice-guiding evidence. Despite these strengths, several limitations warrant consideration. First, the study was based on retrospective public cohorts, and prospective validation remains necessary before clinical translation. Second, although an external cohort was used, the validation sample size remained relatively modest, and broader multi-center validation is still needed to establish generalizability. Third, while multiple algorithmic combinations were compared to reduce model-selection bias, the risk of overfitting cannot be fully excluded. Fourth, the functional and mechanistic roles of the identified genes and pathways require further experimental validation in vitro and in vivo. Finally, the immunotherapy and drug-sensitivity analyses were exploratory and correlation-based, and therefore should not be interpreted as direct therapeutic guidance. In conclusion, our comprehensive integrative analysis elucidates the complex molecular and immune landscapes underlying HNSCC progression, providing a robust prognostic signature and nomogram for individualized clinical management. These findings not only advance our understanding of HNSCC biology but also pave the way for personalized therapeutic strategies, potentially improving patient prognosis and treatment efficacy. Consent for publication Not Applicable. Ethics declaration Not applicable. Funding This research was not supported by any funds. CRediT authorship contribution statement Meihua Xiong: Writing – original draft. Yong Luo: Investigation, Methodology, Software. Shiping Hu: Writing – review & editing. Declaration of competing interest This study does not involve any conflict of interest.