Development and validation of a four-gene prognostic signature derived from disulfidptosis- and cuproptosis-related genes in oral squamous cell carcinoma Authors: Yao Yao, Xiaoshan Li, Ziyan Pang, Wenming Kang, Liqiang Gong, Qi Zhou, Chen Mao Published: 2026 ABSTRACT BackgroundDisulfidptosis and cuproptosis are recently identified forms of regulated cell death whose component genes may harbor prognostic information in oral squamous cell carcinoma (OSCC). However, no validated multi-gene prognostic signature derived from these pathways has been established.MethodsDifferentially expressed genes from two GEO discovery cohorts (GSE30784 and GSE37991) were intersected with 28 disulfidptosis- and cuproptosis-related genes. A LASSO-Cox model was developed in a TCGA oral HNSC training cohort (n = 268) and validated in GSE41613 (n = 97), GSE65858 (n = 270), and GSE42743 (n = 55). An extended clinical-genomic model incorporating age, AJCC stage, histological grade, alcohol use, radiation, and chemotherapy was evaluated in a complete-case subcohort (n = 212). Single-cell RNA-seq data (GSE103322; 5,902 cells) were analyzed to characterize hub gene expression across cell types. Hub gene expression was validated by qRT-PCR and ELISA in OSCC cell lines.ResultsA four-gene signature (CDKN2A, FLNA, GLS, TLN1) was identified. The risk score was an independent prognostic factor (HR = 1.36 per SD, 95% CI: 1.08–1.72, P = 0.009) after adjustment for six clinical covariates. The extended combined model achieved a C-index of 0.669 and a 5-year AUC of 0.805, outperforming the extended clinical model (C-index 0.650; 5-year AUC 0.722) and gene-only model (C-index 0.563; 5-year AUC 0.684). The signature achieved significance in GSE41613 (P = 0.046, HR = 1.81) and GSE65858 (P = 0.0013, HR = 1.99), with a concordant trend in GSE42743 (P = 0.079, HR = 2.10). Single-cell analysis showed that CDKN2A was preferentially expressed in malignant cells, while FLNA and TLN1 were enriched in fibroblasts, consistent with ECM-mediated tumor-stromal crosstalk. CellChat revealed that high-activity malignant cells exhibited enhanced LAMININ, COLLAGEN, and THBS signaling. qRT-PCR and ELISA confirmed upregulation of all four hub genes in CAL-27 and SCC-9 cells (P < 0.01).ConclusionsA four-gene prognostic signature derived from disulfidptosis- and cuproptosis-related genes provides independent prognostic value in OSCC and improves stratification when combined with standard clinical variables.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s12672-026-05478-z. Introduction Oral squamous cell carcinoma (OSCC) is the most common malignancy of the oral cavity, accounting for approximately 90% of oral cancers and representing a substantial fraction of head and neck squamous cell carcinomas (HNSCC) worldwide [1]. Despite advances in surgical techniques, radiotherapy, and systemic chemotherapy, the five-year overall survival (OS) rate for OSCC has remained around 50–65% over the past several decades, particularly for patients diagnosed at advanced stages [2–3]. Locoregional recurrence and cervical lymph node metastasis continue to be the primary drivers of poor outcomes. Consequently, there is an urgent need for reliable prognostic biomarkers that can stratify patients by risk and guide individualized treatment strategies. Regulated cell death (RCD) pathways have emerged as central mechanisms governing tumor initiation, progression, and therapeutic response [4]. Beyond the well-characterized apoptosis and necroptosis pathways, several novel forms of RCD have been described in recent years. Among these, disulfidptosis was identified by Liu et al. in 2023 as a distinct form of cell death triggered by disulfide stress in cells with high expression of the cystine transporter SLC7A11 under glucose deprivation [5]. In this process, abnormal accumulation of intracellular disulfide molecules induces aberrant disulfide bonding among cytoskeletal proteins—including actin, myosin heavy chains, filamins, and talin—leading to actin network collapse and subsequent cell death [5]. A set of 15 genes has been defined as core regulators of disulfidptosis, including SLC7A11, FLNA, FLNB, MYH9, MYH10, TLN1, ACTB, MYL6, CAPZB, DSTN, IQGAP1, ACTN4, PDLIM1, CD2AP, and INF2. Cuproptosis, reported by Tsvetkov et al. in 2022, is a copper-dependent form of regulated cell death that operates through direct binding of excess copper to lipoylated components of the mitochondrial tricarboxylic acid (TCA) cycle [6]. Copper accumulation leads to aggregation of lipoylated proteins such as DLAT and destabilization of iron–sulfur cluster proteins, ultimately causing proteotoxic stress and cell death [6]. Thirteen genes have been identified as key cuproptosis regulators: FDX1, LIAS, LIPT1, DLD, DLAT, PDHA1, PDHB, MTF1, GLS, CDKN2A, ATP7B, SLC31A1, and ATP7A. Both disulfidptosis and cuproptosis are linked to cellular metabolic states—disulfidptosis to glucose metabolism and redox homeostasis, and cuproptosis to mitochondrial respiration and the TCA cycle. The tumor microenvironment (TME), particularly the extracellular matrix (ECM), plays a central role in mediating cancer cell behavior and therapeutic response [7]. The ECM serves as a dynamic scaffold that transmits biochemical and mechanical signals between malignant cells and their surrounding stroma. In HNSCC, ECM remodeling driven by laminins, collagens, and thrombospondins has been implicated in promoting invasion, immune evasion, and metastasis [8]. In this study, we aimed to develop and validate a prognostic gene signature derived from disulfidptosis- and cuproptosis-related genes in OSCC. We first identified differentially expressed genes from the two DC gene sets, constructed a LASSO-Cox model in the TCGA training cohort, and validated it in three independent external cohorts. We then assessed the incremental value of the signature over an expanded clinical model, characterized hub gene expression at single-cell resolution to clarify the biological context, and performed CellChat analysis to examine TME communication patterns associated with risk stratification. Hub gene expression was verified by qRT-PCR and ELISA in OSCC cell lines. Methods Data collection and preprocessing Gene expression data and clinical information were obtained from publicly available databases. Two GEO datasets were used as discovery cohorts: GSE30784 (GPL570, Affymetrix HG-U133 Plus 2.0; 167 OSCC and 45 normal samples) [9] and GSE37991 (GPL6883, Illumina HumanRef-8 v3.0; 40 matched pairs of OSCC tumor and adjacent normal tissues) [10]. For the training cohort, we downloaded RNA-seq data (STAR-FPKM) and clinical annotations for TCGA-HNSC from the UCSC Xena platform (https://xenabrowser.net/). We restricted the TCGA cohort to oral cavity subsites (tongue, floor of mouth, gum, hard palate, and lip), yielding 268 tumor samples with survival data. Disease-specific survival (DSS) data were obtained from the TCGA Pan-Cancer Clinical Data Resource [11]. Three independent GEO datasets served as external validation cohorts: GSE41613 (GPL570; 97 OSCC patients with OS data) [12], GSE65858 (GPL10558, Illumina HumanHT-12 v4; 270 HNSCC patients, of which 83 were classified as oral cavity) [13], and GSE42743 (GPL570, Affymetrix HG-U133 Plus 2.0; 103 HNSCC patients from the validation set of Lohavanichbutr et al., of which 55 were oral cavity tumors with survival data) [12]. For single-cell RNA sequencing (scRNA-seq) analysis, we used GSE103322 (Smart-seq2; 5,902 cells from HNSCC tumors) [8]. The clinical characteristics of the TCGA oral HNSC training cohort are summarized in Table 1. A combined set of 28 unique genes was compiled from 15 established disulfidptosis regulators [5] and 13 established cuproptosis regulators [6], with no overlap between the two gene sets. Identification of disulfidptosis- and cuproptosis-related differentially expressed genes (DC-DEGs) Differential expression analysis was conducted for each discovery cohort using the limma package (version 3.56.2) in R [14]. For GSE30784, robust multi-array average (RMA) normalization was applied, and genes with an adjusted P-value (Benjamini–Hochberg) < 0.05 and |log2FC| > 0.5 were considered differentially expressed. For GSE37991, paired-sample analysis was performed using a nominal P-value < 0.05 with the same |log2FC| > 0.5 threshold, given the smaller paired sample size. The intersection of DEGs from both datasets with the 28 DC genes was computed to obtain the final set of DC-DEGs. This DEG pre-filtering step was designed to restrict model inputs to genes that are genuinely dysregulated in OSCC tissue relative to normal epithelium, thereby excluding non-differentially-expressed genes whose prognostic associations may reflect non-pathway-specific confounding (see sensitivity analysis, Sect. 2.11). Construction of the prognostic risk model In the TCGA oral HNSC training cohort (n = 268), we first performed univariate Cox proportional hazards regression for each of the 13 DC-DEGs. Only genes reaching a pre-specified threshold of P < 0.1 in univariate analysis were retained for LASSO penalization. The least absolute shrinkage and selection operator (LASSO) penalized Cox regression was then applied using the glmnet R package [15] with 10-fold cross-validation to select the optimal penalty parameter (λ). The model yielding the minimum partial likelihood deviance (λmin 0.0453) was selected, retaining four genes. Each patient’s risk score was calculated as a linear combination of the expression levels of the selected genes weighted by their respective LASSO coefficients:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Risk\:Score}=\sum\:_{i=1}^{4}{\beta\:}_{i}\times\:{\mathrm{Expr}}_{i}$$\end{document} where β_i is the LASSO coefficient and Expr_i is the normalized expression value of gene i. Patients were stratified into high- and low-risk groups using the median risk score as the cutoff. Cox regression, extended combined model, and nomogram Univariate and multivariate Cox proportional hazards regression analyses were conducted to assess whether the risk score was an independent prognostic factor. To facilitate clinical interpretation and avoid scale-dependent inflation of hazard ratios, the risk score was z-score standardized before entering Cox models; all reported HRs therefore reflect the change in hazard per one standard deviation (SD) increase in risk score. To evaluate the incremental value of the gene signature beyond an extended set of clinical variables, three models were compared: (1) an extended clinical model incorporating age, AJCC stage, histological grade (G1–G2 vs. G3–G4), alcohol use history (yes vs. no), receipt of radiation therapy (yes vs. no), and receipt of chemotherapy (yes vs. no); (2) a gene-only model (standardized risk score alone); and (3) an extended combined model (all six clinical variables plus the risk score). Smoking history was excluded as a covariate because pack-year data were unavailable for 118 of 268 TCGA samples (44.0%), and complete-case analysis with smoking would have reduced the analytical set to n = 111. Analysis of the three-model comparison was restricted to patients with complete data for all six clinical covariates (n = 212, 79.1% of the training cohort). Model performance was assessed by concordance index (C-index) and time-dependent AUC at 1, 3, and 5 years. A prognostic nomogram integrating the risk score with significant clinical variables was constructed using the rms R package [16]. Calibration curves and decision curve analysis (DCA) were performed. Disease-specific survival analysis To assess the robustness of the prognostic model across different survival endpoints, disease-specific survival (DSS) data were obtained from the TCGA Pan-Cancer Clinical Data Resource [11]. Kaplan–Meier analysis and Cox regression were performed for DSS using the same risk stratification. Time-dependent ROC curves were computed for 1-, 3-, and 5-year DSS. Immune infiltration analysis The single-sample gene set enrichment analysis (ssGSEA) algorithm implemented in the GSVA R package [17] was employed to quantify the relative infiltration levels of 27 immune cell types in each TCGA sample. Differences in immune cell enrichment scores between the high- and low-risk groups were assessed by the Wilcoxon rank-sum test. The ESTIMATE algorithm [18] was used to infer stromal scores, immune scores, and tumor purity for each sample. Consensus clustering and pathway analysis Unsupervised consensus clustering was performed on the 13 DC-DEG expression values using the ConsensusClusterPlus R package [19] with 1,000 iterations, a resampling rate of 80%, and agglomerative hierarchical clustering with Ward linkage and Pearson correlation distance. Gene set variation analysis (GSVA) [17] was performed using 50 Hallmark gene sets from MSigDB [20] to identify differentially enriched pathways between the clusters (Wilcoxon test, FDR < 0.05). External validation The four-gene risk score formula derived from the TCGA training cohort was applied to three independent validation cohorts. Because cross-platform expression scale differences preclude direct application of the training median, the surv_cutpoint function [21] was used to determine the optimal cutoff in each cohort, with a minimum group proportion of 25%. Continuous Cox regression was performed as a cutoff-independent complement. For GSE42743, raw CEL files were downloaded from GEO and normalized by the RMA method using the affy R package with the hgu133plus2cdf Brainarray CDF. Probe-to-gene mapping was performed using the hgu133plus2.db annotation package, with the most variable probe retained for each gene. Only samples annotated as oral cavity tumor with available OS data were retained (n = 55). Kaplan–Meier analysis, time-dependent ROC curves [22], and associations with lymph node metastasis and HPV status were also evaluated. Sensitivity analysis: LASSO without DEG pre-filtering To evaluate whether the DEG pre-filtering step influenced the gene selection outcome, a sensitivity analysis was performed in which all 28 DC genes (without requiring differential expression in OSCC) were carried forward directly to univariate Cox screening (P < 0.1) and LASSO-Cox. This analysis was conducted using the same TCGA training cohort, survival endpoint, LASSO parameters, and random seed as the primary analysis, with the sole difference being the omission of the DEG filter. Results are reported in Supplementary Figure S1m. Single-cell RNA sequencing analysis Processed scRNA-seq data from GSE103322 [8] comprising 5,902 cells (2,539 malignant and 3,363 non-malignant) from HNSCC tumors were obtained. Cell type annotations and malignant cell classifications from the original publication were used. Data were processed using the Seurat R package [23]. Two complementary analyses were conducted. First, to characterize the cell-type-specific expression of the four hub genes, dot plots and violin plots were generated using normalized expression values across all annotated cell types. Malignant cells were identified using the is_cancer field from the original publication. Mean normalized expression was computed per cell type. This analysis was designed to determine whether hub genes are intrinsically expressed in malignant cells or predominantly in stromal/immune cell types, thus providing biological context for the bulk prognostic signal. Second, to stratify malignant cells by their disulfidptosis-cuproptosis gene set activity, we employed AUCell [24], an algorithm that evaluates gene set activity in individual cells by testing whether genes in a signature are enriched among the top-ranked genes in each cell’s expression profile. The four hub genes (CDKN2A, FLNA, GLS, TLN1) were used as the input gene set. We emphasize that this AUCell stratification reflects gene-set activity within each cell’s transcriptome; it does not apply the bulk LASSO risk score weights to single-cell data and is not intended as a cell-state classifier derived from the bulk prognostic model. Among 2,539 malignant cells, median-split stratification yielded 502 high-activity and 2,037 low-activity subpopulations. To assess inter-tumor heterogeneity in the scRNA-seq data, only patients with at least 10 malignant cells were included in tumor-level analyses (n = 17 of 20 tumors), following the recommendation of the original publication [8]. Per-tumor cell-type balance is provided in Supplementary Figure S1n. CellChat cell–cell communication analysis To investigate how risk-stratified malignant cells communicate with the TME, we performed cell–cell communication analysis using CellChat (version 2.1) [25]. The malignant cell population was split into “Mal_High-risk” and “Mal_Low-risk” groups based on AUCell stratification, while non-malignant cell types retained their original annotations. The CellChatDB.human database was used as the ligand–receptor reference. Communication probabilities were computed using the trimean method, and interactions supported by fewer than 10 cells per group were filtered. Drug–gene interaction analysis Drug–gene interactions for the four hub genes were queried from the Drug Gene Interaction Database (DGIdb, version 4.0) [26]. Only interactions with established pharmacological evidence or clinical relevance were retained. Cell culture Human oral keratinocytes (HOK; ScienCell Research Laboratories, Carlsbad, CA, USA) were cultured in oral keratinocyte medium (OKM, ScienCell) supplemented with oral keratinocyte growth supplement (OKGS) and 1% penicillin/streptomycin. The OSCC cell lines CAL-27 (ATCC CRL-2095) and SCC-9 (ATCC CRL-1629) were obtained from the American Type Culture Collection (ATCC, Manassas, VA, USA). CAL-27 cells were maintained in Dulbecco’s modified Eagle’s medium (DMEM; Gibco) supplemented with 10% fetal bovine serum (FBS; Gibco) and 1% penicillin/streptomycin. SCC-9 cells were cultured in DMEM/F-12 (1:1; Gibco) with 10% FBS, 400 ng/mL hydrocortisone (Sigma-Aldrich), and 1% penicillin/streptomycin. All cells were maintained at 37°C in a humidified atmosphere containing 5% CO_2 and were confirmed to be mycoplasma-free. Quantitative real-time polymerase chain reaction (qRT-PCR) Total RNA was extracted using TRIzol reagent (Invitrogen, Carlsbad, CA, USA) and quantified with a NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific). Complementary DNA (cDNA) was synthesized from 1 µg of total RNA using the PrimeScript RT Master Mix (Takara Bio, Shiga, Japan). qRT-PCR was performed on an ABI 7500 Real-Time PCR System (Applied Biosystems) using TB Green Premix Ex Taq II (Takara Bio). GAPDH served as the endogenous reference gene. The relative expression of target genes was calculated using the 2-ΔΔ Ct method with HOK as the calibrator. Three biological replicates were assayed for each cell line, each in triplicate. Enzyme-linked immunosorbent assay (ELISA) Cell lysates were prepared using RIPA buffer (Beyotime, Shanghai, China) supplemented with protease inhibitor cocktail (Roche). Total protein concentrations were determined by BCA assay (Thermo Fisher Scientific). Human GLS (glutaminase) and human CDKN2A/p16 protein levels were measured using commercial ELISA kits (Elabscience, Wuhan, China) according to the manufacturer’s instructions. Absorbance was read at 450 nm (with 570 nm correction) on an Epoch 2 microplate reader (BioTek, Winooski, VT, USA). Standard curves were fitted using four-parameter logistic (4PL) regression (R2 > 0.999). Three biological replicates per cell line were assayed in duplicate. Statistical analysis All bioinformatics analyses were conducted in R (version 4.3.1). Continuous variables were compared between two groups using the Wilcoxon rank-sum test. Survival analyses were performed using the Kaplan–Meier method with log-rank tests. Cox proportional hazards models were used for univariate and multivariate regression. Time-dependent ROC analysis was performed using the timeROC package [22]. Machine learning model comparison was conducted using 5-fold cross-validation for support vector machine (SVM), ridge regression, random forest (RF), elastic net, and LASSO models, with 3-year OS as the endpoint. For in vitro experiments, differences between two groups were assessed by unpaired Student’s t-test. All experimental data are presented as mean ± standard deviation (SD) from three independent experiments. A two-sided P < 0.05 was considered statistically significant unless otherwise specified. Results Identification of 13 disulfidptosis- and cuproptosis-related DEGs in OSCC We first assessed the overall transcriptomic differences between OSCC and normal tissues. Principal component analysis (PCA) of GSE30784 demonstrated clear separation between tumor and normal samples (Fig. 1A). Differential expression analysis of GSE30784 identified 2,736 upregulated and 2,393 downregulated genes (adjusted P < 0.05, |log2FC| > 0.5), visualized by volcano plot (Fig. 1B). A parallel analysis of the paired GSE37991 dataset yielded a complementary set of DEGs (Fig. 1C). We then intersected the DEGs from both discovery cohorts with the combined set of 28 disulfidptosis and cuproptosis genes. This yielded 13 DC-DEGs: CDKN2A, DLAT, FLNA, FLNB, GLS, LIAS, LIPT1, MTF1, MYH10, MYH9, PDHB, SLC7A11, and TLN1 (Fig. 1D). A heatmap of the 13 DC-DEGs across GSE30784 samples confirmed consistent differential expression patterns between OSCC and normal tissues (Fig. 1E). Gene ontology (GO) enrichment analysis revealed significant associations with actin cytoskeleton organization, mitochondrial metabolic processes, and copper ion homeostasis pathways (Fig. 1F), consistent with the known biological functions of disulfidptosis and cuproptosis regulators. Construction of a four-gene prognostic risk model Univariate Cox screening of the 13 DC-DEGs (threshold P < 0.1) identified four candidate genes for LASSO penalization. LASSO-Cox regression applied to these four candidates in the TCGA oral HNSC training cohort (n = 268) yielded a four-gene model after 10-fold cross-validation (λmin =_= 0.0453; Fig. 2A, B). The retained genes were CDKN2A (β = -0.050, protective), FLNA (β = +0.031, risk), GLS (β = +0.064, risk), and TLN1 (β = +0.026, risk). Patients were stratified into high-risk (n = 134) and low-risk (n = 134) groups by the median risk score. The risk score distribution, survival status, and expression heatmap of the four hub genes are shown in Fig. 2C. Kaplan–Meier analysis demonstrated significantly worse OS in the high-risk group (P < 0.01; Fig. 2D). Time-dependent ROC analysis yielded AUC values of 0.61, 0.63, and 0.60 for 1-, 3-, and 5-year OS, respectively (Fig. 2E). A machine learning comparison using 5-fold cross-validation confirmed consistent model performance, with SVM achieving the highest AUC (0.592; Supplementary Fig. S1f). Independent prognostic value, extended combined model, and nomogram After z-score standardization, univariate Cox regression identified the risk score (HR = 1.27 per SD, 95% CI: 1.02–1.59, P = 0.037), age (HR = 1.02 per year, P = 0.024), and AJCC stage (HR = 3.15, P = 0.001) as significant prognostic factors. Multivariate Cox regression confirmed the risk score as an independent prognostic factor after adjustment for age, AJCC stage, histological grade, alcohol use, radiation, and chemotherapy (HR = 1.36 per SD, 95% CI: 1.08–1.72, P = 0.009; n = 212 complete cases; Table 2). Of the clinical covariates, age (P = 0.022), AJCC stage (P < 0.001), and receipt of radiation therapy (protective, HR = 0.52, P = 0.018) also reached significance. Grade, alcohol use, and chemotherapy were retained in the model for completeness but did not reach significance in multivariate analysis (P > 0.1). Extended combined clinical-genomic model Three models were compared in the complete-case subcohort (n = 212; Fig. 3A–C). The extended combined model (age + stage + grade + alcohol + radiation + chemo + risk score) outperformed both the extended clinical model and the gene-only model across all time horizons (3-year AUC: combined 0.739 vs. clinical 0.706 vs. gene 0.642; 5-year AUC: combined 0.805 vs. clinical 0.722 vs. gene 0.684; Fig. 3A, B). The C-index for the combined model was 0.669 (95% CI: 0.608–0.731), compared with 0.650 (0.588–0.712) for the extended clinical model and 0.563 (0.503–0.623) for the gene-only model (Fig. 3C). A nomogram integrating the risk score, age, and AJCC stage was constructed (Fig. 2G), with calibration curves demonstrating good agreement between predicted and observed survival (Fig. 2H) and DCA confirming clinical net benefit (Fig. 2I). DSS analysis showed a consistent trend (HR = 1.23 per SD, 5-year AUC = 0.67; Fig. 3D, E), supporting the robustness of the model across survival endpoints. Immune microenvironment characterization ssGSEA analysis of 27 immune cell types revealed distinct infiltration patterns between risk groups (Fig. 4A). Four immune cell types showed significantly different infiltration levels between high- and low-risk groups (Wilcoxon P < 0.05): immature B cells (higher in low-risk), macrophages, monocytes, and neutrophils (all higher in high-risk; Fig. 4B). Correlation analysis between hub genes and immune cell infiltration revealed distinct patterns, with GLS and TLN1 showing the most prominent correlations (Fig. 4C). ESTIMATE analysis demonstrated that the high-risk group had lower tumor purity and higher stromal and immune scores (Fig. 4D), suggesting a more complex TME in high-risk OSCC. Consensus clustering reveals two molecular subtypes Unsupervised consensus clustering based on the 13 DC-DEGs identified two optimal clusters. Cluster C1 (n = 184) and C2 (n = 84) showed a trend toward differential OS (P = 0.069; Supplementary Fig. S1o) with distinct DC-DEG expression patterns (Fig. 4E) and risk score distributions (Fig. 4F). GSVA pathway analysis revealed 42 of 50 Hallmark gene sets differentially enriched between clusters (FDR < 0.05; Fig. 4G). Cluster C2 showed enrichment in oxidative phosphorylation, fatty acid metabolism, and MYC targets, while C1 was enriched in epithelial–mesenchymal transition, inflammatory response, and angiogenesis—hallmarks of ECM remodeling and microenvironment reprogramming. Notably, C1 harbored significantly higher risk scores (Fig. 4F), establishing a link between the prognostic model and an ECM-remodeling transcriptional program. External validation of the four-gene signature In GSE41613 (n = 97 OSCC patients), the risk score achieved significant prognostic discrimination (log-rank P = 0.046, HR = 1.81, 95% CI: 1.00–3.28; Fig. 5A). Time-dependent ROC curves are shown in Fig. 5B. In GSE65858 (n = 270 HNSCC patients), the signature also demonstrated significant prognostic value (log-rank P = 0.0013, HR = 1.99, 95% CI: 1.30–3.04; Fig. 5C). Continuous Cox regression, independent of any cutoff choice, confirmed prognostic significance in both validation cohorts (GSE41613: P = 0.035; GSE65858: P < 0.001). A sub-analysis restricted to oral cavity cases in GSE65858 (n = 83) showed a concordant direction (HR = 1.56, 95% CI: 0.68–3.57) but did not reach significance (P = 0.286; Fig. 5D). Post-hoc power analysis for this sub-analysis indicated a statistical power of 16.5% to detect the observed HR, with a minimum detectable HR of 3.53 at 80% power given this sample size. This result likely reflects insufficient power rather than genuine absence of effect, and is discussed further in Limitations. Fig. 1Identification of disulfidptosis- and cuproptosis-related DEGs in OSCC. (A) PCA of GSE30784 showing tumor–normal separation. (B) Volcano plot of DEGs in GSE30784. (C) Volcano plot of DEGs in GSE37991. (D) Venn diagram of the intersection between differentially expressed genes from the two discovery cohorts and the 28 disulfidptosis/cuproptosis genes, yielding 13 DC-DEGs. (E) Heatmap of the 13 DC-DEGs. (F) GO enrichment analysis of the 13 DC-DEGs Identification of disulfidptosis- and cuproptosis-related DEGs in OSCC. (A) PCA of GSE30784 showing tumor–normal separation. (B) Volcano plot of DEGs in GSE30784. (C) Volcano plot of DEGs in GSE37991. (D) Venn diagram of the intersection between differentially expressed genes from the two discovery cohorts and the 28 disulfidptosis/cuproptosis genes, yielding 13 DC-DEGs. (E) Heatmap of the 13 DC-DEGs. (F) GO enrichment analysis of the 13 DC-DEGs Fig. 2Construction and evaluation of the four-gene prognostic risk model. (A) LASSO coefficient profiles. (B) Ten-fold cross-validation curve. (C) Risk score distribution, survival status, and hub gene expression. (D) Kaplan–Meier survival curves. (E) Time-dependent ROC curves for OS. (F) Forest plot of univariate and multivariate Cox regression (extended model with six clinical covariates; standardized HR per SD; n = 212 complete cases). (G) Nomogram for 1-, 3-, and 5-year OS. (H) Calibration curves. (I) Decision curve analysis Construction and evaluation of the four-gene prognostic risk model. (A) LASSO coefficient profiles. (B) Ten-fold cross-validation curve. (C) Risk score distribution, survival status, and hub gene expression. (D) Kaplan–Meier survival curves. (E) Time-dependent ROC curves for OS. (F) Forest plot of univariate and multivariate Cox regression (extended model with six clinical covariates; standardized HR per SD; n = 212 complete cases). (G) Nomogram for 1-, 3-, and 5-year OS. (H) Calibration curves. (I) Decision curve analysis Fig. 3Extended combined clinical-genomic model and disease-specific survival validation. (A) Time-dependent AUC comparison at 3 years: Extended clinical model (Age + Stage + Grade + Alcohol + Radiation + Chemo), Gene model (risk score), and Extended combined model (n = 212 complete cases). (B) Time-dependent AUC at 5 years. (C) Concordance index comparison with 95% confidence intervals. (D) Kaplan–Meier curves for disease-specific survival (DSS). (E) Time-dependent ROC curves for DSS Extended combined clinical-genomic model and disease-specific survival validation. (A) Time-dependent AUC comparison at 3 years: Extended clinical model (Age + Stage + Grade + Alcohol + Radiation + Chemo), Gene model (risk score), and Extended combined model (n = 212 complete cases). (B) Time-dependent AUC at 5 years. (C) Concordance index comparison with 95% confidence intervals. (D) Kaplan–Meier curves for disease-specific survival (DSS). (E) Time-dependent ROC curves for DSS Fig. 4Immune microenvironment characterization and consensus clustering. (A) Heatmap of ssGSEA immune cell enrichment scores. (B) Four immune cell types with significantly different infiltration between risk groups: immature B cells, macrophages, monocytes, and neutrophils (Wilcoxon P < 0.05). (C) Correlation between hub genes and immune cell infiltration. (D) ESTIMATE scores between risk groups. (E) DC-DEG expression across clusters. (F) Risk score distribution between clusters. (G) GSVA heatmap of differentially enriched Hallmark pathways Immune microenvironment characterization and consensus clustering. (A) Heatmap of ssGSEA immune cell enrichment scores. (B) Four immune cell types with significantly different infiltration between risk groups: immature B cells, macrophages, monocytes, and neutrophils (Wilcoxon P < 0.05). (C) Correlation between hub genes and immune cell infiltration. (D) ESTIMATE scores between risk groups. (E) DC-DEG expression across clusters. (F) Risk score distribution between clusters. (G) GSVA heatmap of differentially enriched Hallmark pathways Fig. 5External validation and clinicopathological associations. (A) KM curves in GSE41613 (P = 0.046, HR = 1.81, 95% CI: 1.00–3.28). (B) Time-dependent ROC in GSE41613. (C) KM curves in GSE65858 (P = 0.0013, HR = 1.99, 95% CI: 1.30–3.04). (D) KM curves in GSE65858 oral subset (P = 0.286, HR = 1.56, 95% CI: 0.68–3.57; n = 83; post-hoc power = 16.5%). (E) Risk scores by lymph node status. (F) Risk group proportions in N0 vs. N+ groups. (G) ROC for predicting lymph node metastasis. (H) KM curves stratified by HPV and risk. (I) Hub gene expression by lymph node status External validation and clinicopathological associations. (A) KM curves in GSE41613 (P = 0.046, HR = 1.81, 95% CI: 1.00–3.28). (B) Time-dependent ROC in GSE41613. (C) KM curves in GSE65858 (P = 0.0013, HR = 1.99, 95% CI: 1.30–3.04). (D) KM curves in GSE65858 oral subset (P = 0.286, HR = 1.56, 95% CI: 0.68–3.57; n = 83; post-hoc power = 16.5%). (E) Risk scores by lymph node status. (F) Risk group proportions in N0 vs. N+ groups. (G) ROC for predicting lymph node metastasis. (H) KM curves stratified by HPV and risk. (I) Hub gene expression by lymph node status Fig. 6Hub gene cell-type expression, malignant cell activity stratification, ECM-TME communication, and drug prediction. (A) UMAP of 5,902 cells from GSE103322 colored by cell type. (B) AUCell gene-set activity scores for the four hub genes projected on UMAP. (C) AUCell-based activity stratification of malignant cells (502 high-activity vs. 2,037 low-activity). (D) Hub gene expression in high- vs. low-activity malignant cells. (E) CellChat analysis: differentially active outgoing signaling pathways in high- vs. low-activity malignant cells. LAMININ, COLLAGEN, and THBS are the top upregulated pathways. (F) CellChat interaction strength heatmap across cell types. (G) Drug–gene interaction counts. (H) Drug–gene interaction bubble plot Hub gene cell-type expression, malignant cell activity stratification, ECM-TME communication, and drug prediction. (A) UMAP of 5,902 cells from GSE103322 colored by cell type. (B) AUCell gene-set activity scores for the four hub genes projected on UMAP. (C) AUCell-based activity stratification of malignant cells (502 high-activity vs. 2,037 low-activity). (D) Hub gene expression in high- vs. low-activity malignant cells. (E) CellChat analysis: differentially active outgoing signaling pathways in high- vs. low-activity malignant cells. LAMININ, COLLAGEN, and THBS are the top upregulated pathways. (F) CellChat interaction strength heatmap across cell types. (G) Drug–gene interaction counts. (H) Drug–gene interaction bubble plot Fig. 7Experimental validation of hub gene expression in OSCC cell lines. (A–D) qRT-PCR analysis of (A) CDKN2A, (B) FLNA, (C) GLS, and (D) TLN1 mRNA expression in HOK (normal oral keratinocytes), CAL-27, and SCC-9 cells. Expression was normalized to GAPDH and calculated by the 2-ΔΔ Ct method relative to HOK. (E–F) ELISA quantification of (E) GLS and (F) CDKN2A (p16) protein concentrations. Data are presented as mean ± SD from three independent experiments. **P < 0.01; ***P < 0.001 Experimental validation of hub gene expression in OSCC cell lines. (A–D) qRT-PCR analysis of (A) CDKN2A, (B) FLNA, (C) GLS, and (D) TLN1 mRNA expression in HOK (normal oral keratinocytes), CAL-27, and SCC-9 cells. Expression was normalized to GAPDH and calculated by the 2-ΔΔ Ct method relative to HOK. (E–F) ELISA quantification of (E) GLS and (F) CDKN2A (p16) protein concentrations. Data are presented as mean ± SD from three independent experiments. **P < 0.01; ***P < 0.001 In a third OSCC-specific validation cohort (GSE42743, n = 55 oral cavity tumor samples, 26 events), the signature showed a consistent prognostic direction with borderline significance (log-rank P = 0.079, HR = 2.10, 95% CI: 0.90–4.91; Supplementary Fig. S1k). Across all four validation datasets, the risk score was associated with shorter OS in the expected direction. Clinicopathological associations showed that N+ patients had higher risk scores (Fig. 5E), the proportion of high-risk patients was greater in the N+ group (Fig. 5F), and the risk score showed moderate ability to predict lymph node status (Fig. 5G). HPV-stratified analysis is shown in Fig. 5H, and hub gene expression by lymph node status in Fig. 5I. Table 1Clinical characteristics of the TCGA oral HNSC cohort (n = 268)CharacteristicCategoryn (%)Median [IQR]Age (years)60 [52–69]GenderMale187 (68.5)Female86 (31.5)AJCC StageaI19 (7.7)II37 (15.0)III54 (22.0)IV136 (55.3)T classificationbT1–T2108 (41.1)T3–T4142 (54.0)N classificationbN0101 (38.4)N+136 (51.7)Primary siteTongue131 (48.0)Floor of mouth56 (20.5)Oral cavity NOS43 (15.8)Base of tongue24 (8.8)Gum11 (4.0)Other8 (2.9)Vital statusAlive159 (58.2)Dead114 (41.8)Follow-up (months)20.7 [12.4–34.1]aStage information available for 246 patients (10.1% missing). bT/N classification available for 263 patients; TX (n = 13) and NX (n = 26) not shown. NOS, not otherwise specified Table 1Clinical characteristics of the TCGA oral HNSC cohort (n = 268)CharacteristicCategoryn (%)Median [IQR]Age (years)60 [52–69]GenderMale187 (68.5)Female86 (31.5)AJCC StageaI19 (7.7)II37 (15.0)III54 (22.0)IV136 (55.3)T classificationbT1–T2108 (41.1)T3–T4142 (54.0)N classificationbN0101 (38.4)N+136 (51.7)Primary siteTongue131 (48.0)Floor of mouth56 (20.5)Oral cavity NOS43 (15.8)Base of tongue24 (8.8)Gum11 (4.0)Other8 (2.9)Vital statusAlive159 (58.2)Dead114 (41.8)Follow-up (months)20.7 [12.4–34.1]aStage information available for 246 patients (10.1% missing). bT/N classification available for 263 patients; TX (n = 13) and NX (n = 26) not shown. NOS, not otherwise specified Clinical characteristics of the TCGA oral HNSC cohort (n = 268) aStage information available for 246 patients (10.1% missing). bT/N classification available for 263 patients; TX (n = 13) and NX (n = 26) not shown. NOS, not otherwise specified Table 2Univariate and multivariate Cox regression analyses for overall survival (extended model; n = 212 complete cases)2–4 (lr)5–7 VariableUnivariate analysisMultivariate analysisHR95% CI P HR95% CI P Risk score (per SD)a1.271.02–1.590.0371.361.08–1.720.009Age (per year)1.021.00–1.040.0241.021.00–1.040.022Stage (III–IV vs. I–II)3.151.58–6.290.0013.911.89–8.07< 0.001Grade (G3–G4 vs. G1–G2)1.240.75–2.050.3991.410.84–2.370.190Alcohol (Yes vs. No)1.090.69–1.740.7021.170.73–1.880.519Radiation (Yes vs. No)0.870.55–1.360.5290.520.30–0.890.018Chemotherapy (Yes vs. No)1.350.88–2.090.1691.500.88–2.550.135HR, hazard ratio; CI, confidence interval; SD, standard deviation. aThe risk score was z-score standardized before analysis; HRs reflect the change in hazard per one SD increase. Smoking history was excluded due to 44.0% missing data in the TCGA cohort. Complete-case analysis: n = 212 (79.1% of the training cohort of n = 268) Table 2Univariate and multivariate Cox regression analyses for overall survival (extended model; n = 212 complete cases)2–4 (lr)5–7 VariableUnivariate analysisMultivariate analysisHR95% CI P HR95% CI P Risk score (per SD)a1.271.02–1.590.0371.361.08–1.720.009Age (per year)1.021.00–1.040.0241.021.00–1.040.022Stage (III–IV vs. I–II)3.151.58–6.290.0013.911.89–8.07< 0.001Grade (G3–G4 vs. G1–G2)1.240.75–2.050.3991.410.84–2.370.190Alcohol (Yes vs. No)1.090.69–1.740.7021.170.73–1.880.519Radiation (Yes vs. No)0.870.55–1.360.5290.520.30–0.890.018Chemotherapy (Yes vs. No)1.350.88–2.090.1691.500.88–2.550.135HR, hazard ratio; CI, confidence interval; SD, standard deviation. aThe risk score was z-score standardized before analysis; HRs reflect the change in hazard per one SD increase. Smoking history was excluded due to 44.0% missing data in the TCGA cohort. Complete-case analysis: n = 212 (79.1% of the training cohort of n = 268) Univariate and multivariate Cox regression analyses for overall survival (extended model; n = 212 complete cases) HR, hazard ratio; CI, confidence interval; SD, standard deviation. aThe risk score was z-score standardized before analysis; HRs reflect the change in hazard per one SD increase. Smoking history was excluded due to 44.0% missing data in the TCGA cohort. Complete-case analysis: n = 212 (79.1% of the training cohort of n = 268) Hub gene cell-type expression and AUCell risk stratification in single-cell RNA-seq Cell-type-specific expression of hub genes To clarify where the hub genes are expressed within the OSCC microenvironment, we characterized their expression across annotated cell types from GSE103322 (Fig. 6A; Supplementary Fig. S1l). CDKN2A showed the highest mean normalized expression in malignant cells (cluster “0”, mean = 0.344), exceeding expression in all non-malignant cell types including T cells (0.086), macrophages (0.025), and fibroblasts (0.023). This pattern is consistent with the known accumulation of p16/CDKN2A in HNSCC cells through Rb pathway inactivation and oncogenic stress responses [27–28]. In contrast, FLNA and TLN1—core cytoskeletal genes—were most abundant in fibroblasts (FLNA: 1.312; TLN1: 0.664), reflecting their established roles in ECM-mediated cell adhesion; notably, both genes were also expressed at appreciable levels in malignant cells (FLNA: 0.888; TLN1: 0.292), consistent with cytoskeletal remodeling in cancer cells. GLS showed highest expression in dendritic cells (1.221), reflecting the glutamine metabolic demands of antigen-presenting cells; expression in malignant cells was modest (0.254). These patterns suggest that the four hub genes exert their prognostic influence through a combination of malignant cell-autonomous mechanisms (CDKN2A) and tumor-stromal crosstalk (FLNA, TLN1, GLS). AUCell activity stratification and per-tumor balance Beyond characterizing hub gene expression (UMAP in Fig. 6A), we applied AUCell to quantify disulfidptosis-cuproptosis gene set activity in individual cells (Fig. 6B). AUCell scores varied markedly across cell types, with the highest activity in malignant and fibroblast populations (Supplementary Fig. S1h). Among 2,539 malignant cells, median-split stratification yielded 502 high-activity and 2,037 low-activity cells (Fig. 6C). Differential expression analysis between high- and low-activity malignant cells confirmed significant upregulation of GLS, FLNA, and TLN1 in the high-activity subpopulation, while CDKN2A was downregulated (Fig. 6D), mirroring the protective and risk coefficients in the LASSO model. To assess inter-tumor heterogeneity, we restricted tumor-level analyses to patients contributing at least 10 malignant cells (n = 17 tumors); per-tumor composition is shown in Supplementary Fig. S1n. CellChat reveals ECM-mediated TME communication remodeling To investigate how AUCell-stratified malignant cells communicate with the TME, we performed CellChat analysis. We found that high-activity malignant cells had considerably stronger outgoing signaling than low-activity cells, especially through ECM-related pathways (Fig. 6E). The most differentially active outgoing pathways were LAMININ (Δ = +4.20), COLLAGEN (Δ = +2.16), THBS (thrombospondin, Δ = +1.17), and APP (Δ = +0.66; Fig. 6E). On the incoming side, high-activity malignant cells received stronger signals through LAMININ (Δ = +2.93), COLLAGEN (Δ = +1.52), VISFATIN (Δ = +0.49), and EGF (Δ = +0.41). The interaction strength heatmap showed that high-activity malignant cells had more extensive communication with fibroblasts, endothelial cells, and macrophages than their low-activity counterparts (Fig. 6F). These results suggest that malignant cells expressing higher levels of the DC gene program may contribute to active ECM remodeling through LAMININ and COLLAGEN signaling, consistent with the role of FLNA and TLN1 as cytoskeletal integrin-activating proteins in ECM adhesion. Drug–gene interaction prediction Query of DGIdb identified six drug–gene interactions (Fig. 6G, H). CDKN2A was associated with three CDK4/6 inhibitors: palbociclib, ribociclib, and abemaciclib [29]. GLS was linked to three glutaminase inhibitors: CB-839 (telaglenastat), BPTES, and DON [30]. Experimental validation of hub genes at mRNA and protein levels To provide independent experimental evidence supporting the biological relevance of the four-gene signature, we performed qRT-PCR and ELISA in OSCC cell lines. qRT-PCR analysis revealed that all four hub genes were significantly upregulated in both CAL-27 and SCC-9 cells compared with normal oral keratinocytes (HOK; Fig. 7A–D). CDKN2A mRNA was elevated 3.21-fold in CAL-27 (P < 0.001) and 1.92-fold in SCC-9 (P < 0.01) relative to HOK. FLNA showed 2.73-fold (P < 0.001) and 2.05-fold (P < 0.001) upregulation, respectively. GLS, the gene with the largest LASSO coefficient, exhibited 3.08-fold (P < 0.001) and 2.32-fold (P < 0.001) increases. TLN1 was upregulated 2.34-fold (P < 0.01) and 1.76-fold (P < 0.01). At the protein level, ELISA quantification confirmed elevated GLS and CDKN2A (p16) concentrations in OSCC cell lines (Fig. 7E, F). GLS protein was 3.10-fold higher in CAL-27 (P < 0.001) and 1.95-fold higher in SCC-9 (P < 0.001) compared with HOK. CDKN2A protein was 2.80-fold higher in CAL-27 (P < 0.001) and 1.52-fold higher in SCC-9 (P < 0.01). The three risk-associated genes (FLNA, GLS, TLN1; positive LASSO coefficients) were upregulated in both OSCC cell lines, concordant with their risk roles in the prognostic model. CDKN2A, which carries a negative (protective) coefficient, was also elevated in cancer cells, reflecting known compensatory p16 upregulation in response to oncogenic stress [27]. This pattern is consistent with its established role in HNSCC, where p16 overexpression can coexist with functional impairment through downstream pathway disruption [28]. Discussion In this study, we developed and validated a four-gene prognostic signature (CDKN2A, FLNA, GLS, TLN1) derived from genes that regulate disulfidptosis and cuproptosis. The signature provides independent prognostic value in OSCC after adjustment for six clinical covariates and incrementally improves prognostic stratification when combined with standard clinical staging. The four-gene model showed a moderate gene-only AUC (0.60–0.66), which is consistent with published expression-based prognostic models in HNSCC [12–13]. The extended combined model (C-index 0.669; 5-year AUC 0.805) outperforms the extended clinical model (C-index 0.650; 5-year AUC 0.722) and the gene-only model, suggesting that the molecular information complements standard staging. This improvement, though modest, is within the range typically reported for combined clinical-genomic models in head and neck cancer. The consensus clustering of the 13 DC-DEGs identified a cluster (C1) enriched for EMT, inflammatory response, and angiogenesis pathways with higher risk scores, providing biological context for the prognostic association. At the single-cell level, CellChat analysis showed that high-activity malignant cells had markedly increased LAMININ and COLLAGEN outgoing signaling compared with low-activity cells. This is biologically plausible given the functions of our hub genes: FLNA crosslinks actin filaments and connects the cytoskeleton to ECM-binding integrins [31], and TLN1 directly activates integrins to mediate cell–ECM adhesion [32]. The single-cell hub gene expression analysis further showed that CDKN2A is preferentially expressed in malignant cells while FLNA and TLN1 are most abundant in fibroblasts, suggesting that these genes may contribute to prognosis through complementary malignant-cell-autonomous (CDKN2A) and tumor-stromal crosstalk (FLNA, TLN1) mechanisms. Among the four hub genes, GLS had the largest LASSO coefficient (β = +0.064). GLS-derived glutamate feeds into glutathione synthesis via the xCT antiporter (SLC7A11), which is directly involved in the disulfide stress underlying disulfidptosis [5]. Glutamine-derived α-ketoglutarate also enters the TCA cycle, whose lipoylated components are the targets of copper-induced cell death [6]. This dual metabolic involvement makes GLS a plausible link between the metabolic requirements of both pathways, and glutaminase inhibitors such as CB-839 may be worth investigating in this context. Limitations Several limitations should be noted. First, the sub-analysis in GSE65858 oral subset (n = 83) did not reach statistical significance (P = 0.286; HR = 1.56). Post-hoc power analysis indicated that this sub-analysis had only 16.5% power to detect the observed HR, and would require a minimum HR of 3.53 to achieve 80% power at this sample size. We therefore interpret this null result as reflecting inadequate statistical power rather than true absence of prognostic effect, which is supported by the concordant trend direction and by significant findings in both GSE41613 and the full GSE65858 cohort. A well-powered OSCC-specific prospective cohort would be needed to confirm the signature’s generalizability. Second, the extended combined model achieved a C-index of 0.669, which is modest but comparable to values reported for other multi-gene prognostic models in HNSCC (typical range 0.62–0.68) [12–13]. Third, the scRNA-seq dataset (GSE103322, Smart-seq2) covers 20 HNSCC tumors (17 with ≥ 10 malignant cells), which limits conclusions about inter-tumor heterogeneity; results may differ with droplet-based platforms or larger cohorts. Fourth, our experimental validation was limited to cell lines; immunohistochemistry in clinical OSCC specimens would provide stronger evidence. Fifth, the present study employed complete-case analysis for clinical covariates, retaining 79.1% of the training cohort (n = 212); smoking history could not be included due to 44.0% missing data. Future directions The present findings establish the prognostic utility of this four-gene signature at the transcriptomic and cell-line level. Future work should address the functional significance of these genes in OSCC through immunohistochemical validation in paired clinical specimens, and gain- and loss-of-function experiments to determine sensitivity to disulfidptosis and cuproptosis inducers. The observation that high-activity malignant cells exhibit enhanced LAMININ and COLLAGEN signaling also warrants investigation through ECM co-culture and invasion assays. Combined metabolic targeting (GLS inhibition) with disruption of ECM-mediated communication represents a potential therapeutic strategy that merits further preclinical investigation. Conclusions We developed and validated a four-gene prognostic signature derived from disulfidptosis- and cuproptosis-related genes in OSCC. The signature provides independent prognostic value after adjustment for six clinical covariates and improves stratification when combined with standard clinical variables. Validation across three external cohorts (GSE41613, GSE65858, and GSE42743) supports generalizability. Single-cell analysis revealed that CDKN2A is preferentially expressed in malignant cells while FLNA and TLN1 are enriched in fibroblasts, consistent with tumor-stromal crosstalk. qRT-PCR and ELISA in OSCC cell lines confirmed upregulation of all four hub genes. These findings may inform risk stratification in OSCC and suggest targets for further biological investigation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Figure S1. Integrated supplementary analyses. (a) Time-dependent ROC curves for individual hub genes. (b–e) Kaplan–Meier survival curves for CDKN2A, FLNA, GLS, and TLN1. (f) AUC comparison across five machine-learning algorithms. (g) Hub gene detection rate across cell types in GSE103322. (h) AUCell score distribution across cell types. (i) Feature plots of hub gene expression on UMAP. (j) Dot plot of hub gene expression across cell types. (k) External validation in GSE42743 (n = 55 OSCC oral cavity patients). (l) Hub gene expression by cell type in GSE103322, with malignant cells distinguished from non-malignant populations. (m) Sensitivity analysis comparing LASSO coefficients between the primary DEG-filtered pipeline and the no-DEG-filter pipeline. (n) Per-tumor malignant-cell composition and AUCell activity balance across 17 HNSCC tumors contributing ≥ 10 malignant cells. (o) Kaplan–Meier survival curves for the two consensus clusters (log-rank P = 0.069). Supplementary Material 1. Figure S1. Integrated supplementary analyses. (a) Time-dependent ROC curves for individual hub genes. (b–e) Kaplan–Meier survival curves for CDKN2A, FLNA, GLS, and TLN1. (f) AUC comparison across five machine-learning algorithms. (g) Hub gene detection rate across cell types in GSE103322. (h) AUCell score distribution across cell types. (i) Feature plots of hub gene expression on UMAP. (j) Dot plot of hub gene expression across cell types. (k) External validation in GSE42743 (n = 55 OSCC oral cavity patients). (l) Hub gene expression by cell type in GSE103322, with malignant cells distinguished from non-malignant populations. (m) Sensitivity analysis comparing LASSO coefficients between the primary DEG-filtered pipeline and the no-DEG-filter pipeline. (n) Per-tumor malignant-cell composition and AUCell activity balance across 17 HNSCC tumors contributing ≥ 10 malignant cells. (o) Kaplan–Meier survival curves for the two consensus clusters (log-rank P = 0.069).