An immune related cell death index for prognostic stratification and immunotherapy response prediction in oral cancer Authors: Jinglin, Jiao Lu, Guihong Liu, Peijun Qi, Yingying Zong, Hang Yu, Zhaochen Hu, Jing Ding, Zhao Liu, Liangsicheng Tang Published: 2026 ABSTRACT Highlights•Cell death is a key biological process to maintain tissue homeostasis and prevent abnormal cell accumulation.•NETosis is mediated by the release of extracellular traps (NETs) from neutrophils and, although this is an immune defense mechanism, is associated with pro-inflammatory, metastatic, and thrombotic processes in some cancers.•IRCDI is expected to be an important biomarker to guide individualized treatment of oral cancer.Keywords: Oral cancer, Immune-related cell death, Prognostic model, Biomarker Introduction Oral cancer (OC) is the most common and highly aggressive oral malignancy worldwide, accounting for the vast majority of oral cancer cases [1,2]. Its pathogenesis is complex and involves extensive genomic mutations, epigenetic regulatory disturbances, as well as altered gene expression, often evolving from oral potentially malignant diseases (OPMDs) [3]. Although high-throughput technologies such as RNA-Seq have driven in-depth dissection of their molecular basis, no significant progress has been made in the clinical treatment outcome and long-term prognosis of OC, especially in advanced patients [4]. The disease is highly prevalent in the tongue, lips, and floor of the mouth, not only affecting basic functions such as chewing and swallowing, but also seriously impairing the quality of life of patients [5]. Risk factors include tobacco, alcohol, poor diet, viral infections, and environmental exposures, and are more prevalent in people of low socioeconomic background and men [5]. In recent years, the study of cell death mechanism has brought new hope for the prevention and treatment of oral cancer [6]. A variety of programmed cell death modes are becoming the key to identify the vulnerability of tumor cells and develop novel therapeutic targets, and different types of cell death not only affect the occurrence and development of tumors, but also regulate the tumor immune microenvironment and therapeutic response [7,8]. Therefore, it is necessary to deeply understand the specific mechanism of cell death pathway in OC, which is expected to reveal new molecular markers and treatment strategies and provide theoretical support and practical paths for improving the precise diagnosis and individualized treatment of oral cancer. Cell death is a key biological process to maintain tissue homeostasis and prevent abnormal cell accumulation, and programmed cell death (PCD) is particularly important in the development and progression of cancer [9,10]. Multiple cell death modes play different biological functions in the tumor microenvironment and have a profound impact on cancer progression, metastasis and treatment response. Anoikis is a form of death triggered by the loss of cell adhesion to the extracellular matrix, and its escape is an important marker for tumor cells to acquire migration and invasion [11]. Apoptosis is a classical programmed cell death mode, and its loss often leads to tumor cells evading clearance and continuous proliferation [12]. Autophagy has a dual role in cancer, both inhibiting tumorigenesis by removing abnormal proteins and damaging organelles, and may provide a survival advantage for tumor cells under stress conditions such as nutritional deprivation [13]. Ferroptosis is a mode of cell death characterized by iron-dependent lipid peroxidation and is considered a potential therapeutic target in multiple tumor types [14]. Immunogenic cell death can release a variety of injury-related molecular patterns (DAMPs), thereby stimulating anti-tumor immune responses, which is an important direction of cancer immunotherapy research [15,16]. In addition, Lysosome-dependent cell death involves lysosomal enzyme leakage and elicits cell damage, and its activation is closely associated with therapeutic resistance in certain tumor types [17]. Mitochondrial permeability transition-driven necrosis is a form of nonapoptotic cell death that relies on the opening of mitochondrial permeability transition pores and may be activated in response to tumor hypoxia and metabolic stress [18]. Necroptosis is a mode of death characterized by inflammation that can be activated when tumor cells evade apoptosis, and its immunomodulatory effects make it a potential target for anticancer therapy [19]. NETosis is mediated by the release of extracellular traps (NETs) from neutrophils and, although this is an immune defense mechanism, is associated with pro-inflammatory, metastatic, and thrombotic processes in some cancers [20,21]. Paraptosis is a caspase-independent mode of cell death characterized by organelle swelling and vesicular degeneration and has received attention in the treatment of drug-resistant tumors in recent years [22]. Pyroptosis is a mode of cell death characterized by cell membrane perforation and release of inflammatory factors that enhances immune responses and causes rapid destruction of tumor cells [23]. In summary, these forms of cell death play diverse and critical roles in the development, progression, treatment response, and immune regulation of cancer, and a deep understanding of their molecular mechanisms will provide an important theoretical basis for the development of novel anticancer strategies. In this study, to deeply dissect the heterogeneity of the immune microenvironment in oral cancer and optimize its individualized treatment strategy, we systematically analyzed the immune-related cell death (IRCD) pattern and constructed a feature collection consisting of immune-related cell death genes (IRCDG) based on data from TCGA and multiple other oral cancer cohorts. We used the ConsensusClusterPlus algorithm to cluster tumor samples to explore immune differences among different subtypes of tumors. Subsequently, we constructed an immune-related cell death index (IRCDI) derived from critical IRCDGs that showed stable prognostic predictive ability in multiple independent cohorts through multiple statistical and machine learning methods. Further combined with single cell transcriptome analysis and immune subtype profiles, we found that low IRCDI was strongly associated with stronger immunocompetence and a more favorable tumor microenvironment. Although IRCDI has limited predictive power in the TIDE algorithm, it shows good efficacy predictive value in multiple real immunotherapy datasets and reveals significant differences associated with the sensitivity of specific drugs, suggesting that IRCDI is expected to be an important biomarker to guide immunotherapy versus drug screening in oral cancer patients. Methods Acquisition and processing of transcriptome data RNA expression profiles and corresponding clinical information of oral cancer patients (N = 278) were obtained from the TCGA-HNSC dataset for model development. Gene expression profiles were normalized to TPM values and subsequently subjected to log2 transformation for further bioinformatics analysis. For external validation, microarray datasets GSE41613 (N = 97) and GSE65858 (N = 83) were retrieved from the GEO database [24,25]. Additionally, seven independent immunotherapy cohorts were included: Melanoma-phs000452 (N = 153), GSE100797 (N = 25), GSE78220 (N = 28), GSE91061 (N = 109), GSE93157 (N = 22), Melanoma-PRJEB23709, and IMvigor210 (N = 348) [[26], [27], [28], [29]]. Acquisition and processing of scRNA-seq The single-cell RNA sequencing dataset GSE181919, comprising 20 tumor samples, was obtained from the GEO database for downstream analysis [30]. Data preprocessing and analysis were performed using R software (version 4.1.3), primarily utilizing the Seurat package. For quality control, cells were retained if they met the following criteria: mitochondrial gene expression <10%, hemoglobin gene expression <3%, UMI counts between 200 and 80,000, and detected genes ranging from 200 to 8000. Normalization was performed using the NormalizeData function, followed by identification of 2000 highly variable genes via FindVariableFeatures. Cell cycle-related variations (S and G2M phase scores) were regressed out during data normalization using the ScaleData function. To mitigate batch effects, the Harmony algorithm was used. Dimensionality reduction and clustering were performed using UMAP and Louvain algorithms within Seurat. DEGs were identified using the FindAllMarkers function with thresholds of an adjusted p-value < 0.05, log2 fold change > 0.25, and expression in at least 10% of the cells. Cell annotation analysis Cell annotation was performed using manual annotation with reference cell marker genes as follows: Epithelial cell markers: EPCAM, KRT18, KRT19, CDH1. Fibroblast markers: DCN, THY1, COL1A1, COL1A2. Endothelial cell markers: PECAM1, CLDN5, FLT1, RAMP2. T cell markers: CD3D, CD3E, CD3G, TRAC. NK cell markers: NKG7, GNLY, NCAM1, RD1. B cell markers: CD79A, IGHG3, IGHG2. Plasma cell markers: JCHAIN. Myeloid cell markers: CD68, FCGR3A. The analysis results were then visualized using UMAP, histograms, and heatmaps. Immune-related cell death gene acquisition Immune-related cell death gene (IRCDG) was obtained from a previous study by Xie et al. [31]. Specifically derived from Anoikis, Apoptosis, Autophagy, Ferroptosis, Immunogenic cell death, Lysosome-dependent cell death, MPT − necrosis, Necroptosis, NETosis, Paraptosis, and Pyroptosis driven related biological processes. Specific gene lists are deposited in Table S1. Establishment of immune-related cell death index (IRCDI) To construct prognostic models, we employed a comprehensive framework integrating 20 different machine learning (ML) algorithms. Each algorithm was used to calculate an individual risk score for patients. Based on the optimal cutoff value identified by the surv_cutpoint function, patients in the TCGA and external cohorts were stratified into high- and low-risk groups. Subsequently, we evaluated differences in survival outcomes between these groups and assessed the predictive performance of each model. A total of 20 ML algorithms for scoring were used, and the specific list was deposited in Table S2. The model development workflow proceeded as follows: (a)Univariate Cox regression was first applied to the TCGA dataset to identify prognostically relevant genes. Univariate Cox regression was first applied to the TCGA dataset to identify prognostically relevant genes. (b)These candidate genes were then incorporated into each of the 20 ML algorithms to construct risk prediction models using a leave-one-out cross-validation (LOOCV) strategy within the TCGA cohort. These candidate genes were then incorporated into each of the 20 ML algorithms to construct risk prediction models using a leave-one-out cross-validation (LOOCV) strategy within the TCGA cohort. (c)All trained models were subsequently validated across multiple external datasets. All trained models were subsequently validated across multiple external datasets. (d)For each model, the Harrell's concordance index (C-index) was calculated across all validation cohorts, and the model with the highest average C-index was chosen as the optimal prognostic signature. For each model, the Harrell's concordance index (C-index) was calculated across all validation cohorts, and the model with the highest average C-index was chosen as the optimal prognostic signature. Mutation level analysis CNV was analyzed for each patient using GISTIC 2.0 software, and mutations were calculated for each patient using the R maftools package. Functional enrichment and pathway analysis Gene functional annotation was conducted using the clusterProfiler R package, specifically with the enrichGO function, to explore enriched biological pathways [32]. Protein-protein interaction networks were analyzed using Metascape with default parameters. Gene Set Enrichment Analysis (GSEA) was conducted using the GSEA function (with default settings) to explore functional differences between the high and low IRCDI groups. The outcomes were visualized with the gseaplot2 function from the enrichplot package. Subsequently, Spearman correlation analysis was conducted between the quantified pathway enrichment scores and IRCDI values to identify pathways most significantly associated with the index. Analysis of correlation between prognostic model and tumor immunity The degree of immune infiltration for each patient in the TCGA database was determined using IOBR software containing the results of seven assessment methods [33]. These data were used to create heat maps to quantify the relative proportion of immune cell infiltrates in the TME. The immune response of the TCGA dataset was predicted using the TIDE algorithm (Tumor Immune Dysfunction and Exclusion (TIDE) to evaluate differences in response to immunotherapy. Drug sensitivity analysis Drug sensitivity analysis was used to assess the value of predicting IRCDI in the clinical treatment of oral cancer. From the Cancer Genome Drug Sensitivity (GDSC) website, the TCGA project determined the half maximal inhibitory concentration (IC50) of the compound based on the oral cancer dataset. We predicted IC50 of compounds obtained from GDSC website in oral cancer patients using R pRRophetic software package [34]. Cell culture The cell lines used in this study were human normal oral epithelial cells GMSM-K and human buccal mucosa squamous carcinoma cells TR146, respectively. Cells were cultured in RPMI-1640 medium supplemented with 10% FBS at 37 °C in 5% CO₂. Western blot analysis Total protein was extracted from cultured cells and paired clinical tissue samples using RIPA lysis buffer (Beyotime, China) supplemented with protease and phosphatase inhibitors (Roche, Switzerland). Protein concentration was determined using the BCA Protein Assay Kit (Thermo Fisher Scientific, USA). Equal amounts of protein (30 μg per lane) were separated by 10% SDS‑polyacrylamide gel electrophoresis (SDS‑PAGE) and subsequently transferred onto polyvinylidene difluoride (PVDF) membranes (Millipore, USA). The membranes were blocked with 5% non‑fat milk in Tris‑buffered saline containing 0.1% Tween‑20 (TBST) for 1 hour at room temperature, followed by overnight incubation at 4 °C with primary antibodies against PDK4 (1:5000, Proteintech), and TUBULIN (1:10,000, Proteintech) as a loading control. After washing with TBST, the membranes were incubated with horseradish peroxidase (HRP)‑conjugated secondary antibodies (1:50,000, Proteintech) for 1 hour at room temperature. Protein bands were visualised using an enhanced chemiluminescence (ECL) substrate (Millipore, USA) and imaged with a chemiluminescence imaging system (Tanon, China). Immunofluorescence (IF) staining For immunofluorescence analysis, formalin‑fixed paraffin‑embedded (FFPE) clinical samples were cut into 4‑μm sections, deparaffinised in xylene, and rehydrated through graded ethanol series. Antigen retrieval was performed by heating the sections in citrate buffer (pH 6.0) at 95 °C for 20 min. Both cells and tissue sections were permeabilised with 0.1% Triton X‑100 in PBS for 15 min and blocked with 5% bovine serum albumin (BSA) in PBS for 1 hour at room temperature. Samples were then incubated overnight at 4 °C with primary antibodies against PDK4 (1:200, Proteintech), pan‑cytokeratin (pan‑CK, 1:200, Abcam), and Ki67 (1:400, Cell Signaling Technology). Statistical analysis All data processing, statistical analyses, and visualizations were performed using R software (version 4.1.3). Pearson's correlation coefficient assessed associations between continuous variables. For categorical variables, the Chi-square test was applied, while continuous variables were analyzed using either the Wilcoxon rank-sum test or Student's t-test, depending on data distribution. The optimal cutoff value for survival analysis was determined using the survminer package. Cox proportional hazards regression and Kaplan-Meier survival analyses were conducted using the survival package. Result Patients with oral cancer tumors based on cold and hot typing have different immune profiles First, we clustered the TCGA dataset based on the ConsensusClusterPlus algorithm (Fig. 1A, B). By combining the results of M3C, CDF, and PAC, we selected an optimal K value of 2 for clustering classification and divided OC patients into 2 groups (Fig. 1C, D; Supplementary Figure 1). Subsequently, according to the level of immune cells in the two groups of patients, the group with higher level of immune cells was divided into Hot group and the group with lower level of immune cells was divided into Cold group (Fig. 1E). The Hot group showed a higher level of immune cell infiltration, had an immunocompetent microenvironment, and may be more sensitive to immunotherapy; while the Cold group showed an immune silent state and may have immunosuppressive mechanisms (Supplementary Figure 2A, B). In addition, we predicted differences in immune cells between the Cold and Hot groups by MCPcounter, quantiseq, TIMER, CIBERSORT, CIBERSORT − ABS, and xCell algorithms, and the results were also consistent with our typing results, with the Hot group possessing higher levels of immune cells (Supplementary 2C—H). This classification reveals significant differences in the immune microenvironment among different tumor subtypes of OC, providing a reference for subsequent individualized treatment strategies. Subsequently, we performed IRCDG-based analysis of differentially expressed genes in two groups of OC tumors to obtain the IRCDG gene associated with Hot tumors (Supplementary 3A). By GO enrichment analysis, we found that differentially up-regulated genes were mainly enriched in T cell activation and positive regulation as well as immune-related pathways (Supplementary 3B). In addition, these genes were mainly located on chromosomes 1, 2, 5, 10, 11, 12, and 17 (Supplementary 3C). Classification analysis of cold and hot tumors. (A) Clustering heatmap for ConsensusClusterPlus algorithm. (B) Cluster comparison plot of ConsensusClusterPlus versus M3C. (C-D) CDF and PAC profiles corresponding to each K value. (E) Heatmap of ssGSEA scores for 28 immune cells. (F-H) IRCDG (immune-related cell death gene) expression plots and PCA plots for TCGA, GSE41613, GSE65858. (I) Circle plot of subtype of TCGA and content of cold and hot tumors in immunophenotyping. To validate the generality of typing, we clustered them in GSE41613, GSE65858 using the ConsensusClusterPlus algorithm and found that they shared the same characteristics of Hot tumor and Cold tumor typing (Fig. 1F–H). Finally, we compared the subtype of TCGA with the content of hot and cold tumors in immunophenotyping, and the results showed that the composition of hot and cold tumors was different in patients with different types, and patients with C4 type belonged to Cold tumors, while patients with C6 type belonged to Hot tumors (Fig. 1I). Genomics functional mapping of ircdg and establishment of ircdi A total of 63 genes were included in IRCDG, which were derived from Anoikis, Apoptosis, Autophagy, Ferroptosis, Immunogenic cell death, Lysosome − dependentcelldeath, MPT − driven necrosis, Necroptosis, NETosis, Paraptosis and Pyroptosis (Fig. 2A). We used Metascape for online analysis of its function and found that IRCDG is mainly involved in intracellular immune-related functional pathways such as leukocyte activation, positive regulation of immune response as well as positive regulation of cytokine production (Fig. 2B). In OC cancers, mutations are mainly derived from missense mutations, and picture types are mainly SNPs and C mutations are T types (Fig. 2C). Genes predominantly mutated in IRCDG are PIK3CG and STAT1, PIK3CG is predominantly missense, and STAT1 is predominantly nonsense (Fig. 2D). According to copy number variation (CNV) analysis, chromosome 1, 7, 16 and 17 were mainly copy number amplifications, while chromosome 2, 3, 4, 5, 10 and 11 were mainly copy number deletions (Fig. 2E). Analysis of OC mutation. (A) Tree diagram of cell death composition of IRCDG. (B) Reticulated graph showing functional analysis of Metascape. (C-D) Mutation statistics for TCGA-OC and IRCDG mutation waterfall plots. (E) Histogram showing CNV frequency for IRCDG. Subsequently, we used univariate Cox analysis to screen genes with prognostic roles in IRCDG (Fig. 3A). The number of genes was further reduced by using the Lasso algorithm (Fig. 3B–D). We employed a competitive machine learning framework to identify the most influential genes and corresponding models for constructing the IRCDI. A total of 20 algorithms were tested for modeling, with the ObliqueRSF algorithm chosen based on the highest mean C-index. The C-index was also computed to evaluate the models' performance across multiple datasets (Fig. 3E). Multivariate Cox analysis showed that IRCDI (the index calculated by the prognostic model) and both model genes were predictive of OC prognosis (Fig. 3F). Heat maps visually depict the proportional expression patterns of model genes and the clinical characteristics of each HCC patient (Fig. 3G). In addition, the results of survival analysis of the model genes PDK4, CCR7 in the TCGA dataset showed (Supplementary Figure 4) no significant correlation between PDK4 expression of the model gene and clinical parameters, while CCR7 expression had a significant correlation with stage and T stage (Supplementary Figure 5). Establish IRCDG prognostic model. (A) Forest plot showing prognostic analysis of IRCDG. (B-D) LASSO regression analysis of IRCDG. On the horizontal axis, the logarithm of the independent variable λ is represented, while on the vertical axis, the partial likelihood of λ is shown for each independent variable. Lambda coefficients represent gene confidence values. (E) Heat map showing C-index indices after modeling for 20 machine learning algorithms. (F) Forest plot showing multivariate Cox analysis for IRCDI (index calculated from prognostic model), PDK4 and CCR7 genes. (G) Heat map of PDK4 and CCR7 gene expression in the TCGA dataset. Clinical associations and survival prediction performance of ircdi Once the IRCDI was constructed, we first explored the relationship between clinical features and IRCDI in the TCGA cohort. The results showed that there was no significant correlation between IRCDI and gender and age as well as M stage (Fig. 4A–C). Interestingly, patients with a high IRCDI index were significantly associated with mortality (P = 2.2e-13) (Fig. 4D), and IRCDG was significantly associated with tumor type (Stage, N and T stage), with high-grade tumors having higher IRCDI values (Fig. 4E–G). To further evaluate the prognostic role of the model for oral cancer, we found that patients with a low IRCDI value had a better prognosis than those with a high IRCDI value by using Kaplan-Meier survival analysis (p < 0.0001) (Fig. 4H). In addition, multivariate Cox regression analysis showed that IRCDI had a significant advantage for oral cancer prognosis and could be used as an independent prognostic indicator (Fig. 4I). In addition, we validated the prognostic role of IRCDI in the external validation set, and the results showed that oral cancer patients with a low IRCDI had a better prognosis in both the GSE41613 cohort and the GSE65858 cohort (Fig. 4J, K). Supplementary Figure 6 shows the ROC analysis at 1, 2, and 3 years for the TCGA cohort, GSE41613 cohort, and GSE65858 cohort. The area under the receiver operating characteristic curve (AUC) for predicting 1-, 2-, and 3-year survival in the TCGA cohort was 0.80, 0.75, and 0.76, respectively. AUC area was greater than 0.5 in both the GSE41613 cohort as well as the GSE65858 cohort and was predictive. Analysis of clinical features and prognosis of OC. (A-G) Violin plots showed differences in IRCDI among patients with different age, sex, survival, stage, M, N, and T. (H) Kaplan-Meier survival curves were used to compare survival probabilities between high and low groups. (I) Forest plot showing multivariate Cox regression analysis by age, sex, survival, stage and N stage. (J-K) Kaplan-Meier survival curves were used to compare survival probabilities between the high and low groups in GSE41613 and GSE65858. Genomic pathway characteristics associated with ircdi To further explore the potential relevance of IRCDI, patients were initially stratified into two groups based on the median IRCDI value. We showed mutations in the top ten genes in the high and low IRCDI groups, with mutations in TP53 occurring in 66 percent of patients with low IRCDI and 78 percent of patients with high IRCDI (Fig. 5A, B). Subsequently, we combined TP53 mutation with IRCDI index to analyze its impact on prognosis. The results showed that patients with low IRCDI index as well as TP53 negativity had the best prognosis, and TP53 gene mutations showed different risk layers in combination with IRCDI (Fig. 5C, D). In addition, we noticed that PIK3CA mutations were higher in the low IRCDI group, and TP53 mutations were combined with IRCDI index to analyze their impact on prognosis. The results showed that patients with low IRCDI index and PIK3CA negative had the best prognosis, and PIK3CA gene mutation combined with IRCDI also showed different risk layers (Fig. 5E, F). Mutation, Function, and WGCNA Analysis. (A) Waterfall plots showing the top 10 mutated genes in patients in the low IRCDI group. (B) Waterfall plots showing top 10 mutated genes in patients in the high IRCDI group. (C) Bar plots show mutations in TP53. (D) Survival analysis showed the impact of TP53 mutations and IRCDI on the prognosis of OC patients. (E) Bar plots show mutations in PIK3CA. (F) Survival analysis showed the impact of PIK3CA mutations and IRCDI on the prognosis of OC patients. (G) GSEA enrichment analysis showed pathway results enriched for upregulated genes in both groups. (H) Volcano plot showing spearman correlation of IRCDI with the HALLMARK pathway. (I-J) Violin plots showed the activation of four positively and four negatively correlated pathways in the high and low groups. (K) Heat map of correlation between gene modules analyzed by WGCNA and clinical parameters. (L) Enrichment analysis results of GO databases for Magenta module genes and Pink module genes. We subsequently performed pathway enrichment analysis between the high and low IRCDI groups. GSEA results revealed that the MYOGENESIS pathway was upregulated in the high IRCDI group, while the ALLOGRAFT REJECTION and KRAS SIGNALING UP pathways were enriched in the low IRCDI group (Fig. 5G, H). Moreover, notable differences in oncogenic pathway activation were observed between the two groups: MYOGENESIS, GLYCOLYSIS, and UNFOLDED PROTEIN RESPONSE were activated in the high IRCDI group, whereas ALLOGRAFT REJECTION, IL6/JAK/STAT3 SIGNALING, IL2/STAT5 SIGNALING, and INTERFERON GAMMA RESPONSE were significantly enriched in the low IRCDI group (Fig. 5I, J). To further explore key gene modules related to IRCDI, we conducted WGCNA using the TCGA dataset (Supplementary Figure 7A, B). A total of 13 modules were identified and analyzed for their associations with IRCDI and clinical traits (Fig. 5K). Among them, the Pink module exhibited the strongest positive correlation with IRCDI (p = 2e-11), while the Magenta module showed the most significant negative correlation (p = 4e-13). Functional enrichment analysis revealed that Magenta module genes were primarily involved in immunoglobulin complex formation and antigen binding, whereas Pink module genes were enriched in muscle system processes and contractile fiber-related pathways (Fig. 5L; Supplementary Figure 7C, D). Low ircdi indicates an immune-active tumor microenvironment To examine the distribution of IRCDI across various cell types in oral cancer, we conducted a scRNA-seq analysis specific to oral cancer. Oral cancer cells were divided into eight populations by cell annotation (Fig. 6A). Fig. 6B calculates and shows IRCDI values for each cluster. The results showed that IRCDI was mainly expressed in B cells, T cells as well as myeloid (Fig. 6C). Subsequently, we investigated whether IRCDI could distinguish immune subsets using the IMvigor 210 cohort. Although the IRCDI score did not clearly distinguish between immune subtypes, patients in the low IRCDI group generally had higher immune cell content. Correlation analysis revealed a significant negative association between IRCDI and immune cell infiltration, particularly Activated B cells (r = −0.51), Activated CD4 T cells (r = −0.41), and Activated CD8 T cells (r = −0.42), all with p-values < 0.0001 (Fig. 6F–H). OC single cell as well as immune infiltrate analysis. (A) UMAP plot showing cell types after annotation of single cell data. (B) UMAP plot showing IRCDI expression in single cell data. (C) Violin plot showing IRCDI values for each cell type. (D) Violin plots showing IRCDI differences for each immunophenotyping in IMvigor210 data. (E) Box plots show the differential results for 28 immune cells in the high and low IRCDI groups. (F-H) Scatter plots showed the correlation between activated B, CD4, CD8 cell content and IRCDI, respectively. (I-K) Box plots showed differential expression of PD-1 in high and low IRCDI groups in TCGA, GSE41613, GSE65858. In addition, we detected better levels of PD-1 expression in patients in the low IRCDI group and validated them in the GSE65858 cohort (Fig. 6I–K). These results suggest that the IRCDI index reflects to some extent the activity of the tumor immune microenvironment. Although IRCDI failed to clearly distinguish immune subtypes, patients in the low IRCDI group generally had higher levels of immune cell infiltration, especially activated B cells, CD4 T cells, and CD8 T cells, and were significantly negatively correlated with IRCDI. In addition, PD-1 expression levels were higher in patients in the low IRCDI group, suggesting that they may have stronger immunocompetence and are more likely to benefit from immune checkpoint inhibition therapy. These results further support IRCDI as a potential indicator of tumor immune status and can be used as an aid in judging the potential for immunotherapy response. IRCDI predicts clinical benefit in immunotherapy cohorts and drug sensitivity Subsequently, we compared IRCDI to TIDE using the TIDE algorithm to predict immunotherapy in the TCGA dataset (Fig. 7A, I). Although no significant differences were found between the high and low IRCDI groups in TIDE, patients in the low IRCDI group showed better outcomes in the immunotherapy cohorts GSE100797, GSE91061, PHS000452, PRJEB23709, GSE93157, and IMvigor210 (Fig. 7B–H). In addition, IRCDI did not affect response to immunotherapy (Fig. 7J–P). Next, we performed drug sensitivity analysis on the TCGA dataset using oncoPredict software and GDSCv2 database to screen for drugs with significant association with IRCDI (| R | > 0.3 & P < 0.05) (Fig. 8A). We found that patients with high IRCDI were more sensitive to BI-25,361,086, and the IC50 of BI-25,361,086 showed a negative correlation with IRCDI values (Fig. 8B). Patients with low IRCDI were more sensitive to talazoparib1259, and the IC50 of talazoparib1259 was positively correlated with IRCDI values (Fig. 8C). Although IRCDI did not show a significant immunotherapeutic predictive advantage in the TIDE algorithm, patients with low IRCDI showed better treatment response and survival outcomes in multiple actual immunotherapeutic cohorts (including GSE100797, GSE91061, etc.), suggesting that IRCDI has potential clinical value in real-world immunotherapeutic evaluation. In addition, IRCDI was not directly associated with response to immunotherapy, but showed differences in drug sensitivity analysis: patients with high IRCDI were more sensitive to BI-25,361,086, while patients with low IRCDI were more sensitive to talazoparib1259, indicating that IRCDI can not only be used for patient prognosis assessment, but may also assist in guiding the selection of individualized therapeutic agents. Immunotherapy analysis. (A) Box plots showing differences in TIDE values between high and low IRCDI groups for the TCGA dataset. (B-H) Survival analysis showed the relationship between IRCDI and OC prognosis in 7 immunotherapy datasets. (I) Scatter plots show the correlation of IRCDI and TIDE values in the TCGA dataset. (J-P) Box plots show IRCD results for response to immunotherapy across 7 immunotherapy datasets. Drug sensitivity analysis. (A) Histograms showed the correlation between IRCDI and drug sensitivity (IC50) (| R | > 0.3 & P < 0.05). (B) Box plots showing differences in susceptibility to BI-25,361,086 between IRCDI high and low groups. Correlation plots showed the correlation between IRCDI values and BI-25,361,086 drug. (C) Box plots showing differences in susceptibility to Talazoparib_1259 between IRCDI high and low groups. Correlation plots showed the correlation between IRCDI values and Talazoparib_1259 drug. (D) Western blot showing PDK4 expression in normal oral epithelial cells (GMSM‑K) and oral cancer cells (TR146). (E) Western blot showing PDK4 expression in paired patient tissues. (F) Immunofluorescence staining of Ki67 in normal and tumour tissues. (G) Immunofluorescence co‑staining of PDK4 and pan‑CK in normal and tumour tissues. Finally, we performed WB analysis of PDK4 expression in normal oral epithelial cells GMSM-K and oral cancer cells TR146. PDK4 expression was significantly higher in tumor cells (Fig. 8D), which displayed a similar pattern in the tissue of oral cancer patient (Fig. 8E). By resorting to IF analysis, we observed a higher expression of Ki67 in the tumor tissue (Fig. 8F), which was observed along with higher PDK4 expression and pan-CK expression (Fig. 8G). These findings suggest that PDK4 may have important roles in the development and progression of oral cancer and have the possibility of deserving further investigation as potential therapeutic targets. Discussion In this study, based on multi-omics data and a variety of bioinformatics algorithms, we systematically analyzed the characteristics of immune-related cell death in oral cancer, and constructed the immune-related cell death index (IRCDI), comprehensively revealed the heterogeneity of the tumor immune microenvironment and its association with prognosis and treatment response, providing new ideas for individualized treatment of OC. First, we used the ConsensusClusterPlus algorithm to cluster oral cancer patients in the TCGA cohort and clearly divided them into two major subtypes: cold tumor and hot tumor. Hot tumor subtypes are rich in immune cell infiltration, have a more active immune microenvironment, and may be more sensitive to immunotherapy; while cold tumors present an immune silent state and may have immune escape mechanisms. We further confirmed the reliability of this typing by cross-validation with multiple immune cell assessment algorithms such as MCPcounter, CIBERSORT, etc. This typing method not only reveals significant differences in the immune microenvironment of OC, but also provides a theoretical basis for the development of differentiated immunotherapy strategies. Based on differential expression analysis of immune-related cell death genes (IRCDG), we selected a series of immune activation pathway genes significantly associated with thermal tumors and confirmed their important roles in T cell activation and immune response regulation by functional enrichment analysis. Subsequently, we constructed the IRCDI scoring system and selected ObliqueRSF as the optimal model by multi-algorithm cross-modeling, combined with Lasso and Cox regression to establish a stable prognostic model. We found that IRCDI had good prognostic predictive power in multiple independent external validation cohorts (GSE41613, GSE65858). In addition, oral cancer patients with low PDK4 and high CCR7 expression had a better prognosis in single-gene analysis. PDK4 has been shown to promote tumorigenesis and progression in a variety of cancers. It promotes metabolic reprogramming and growth of breast cancer cells by inhibiting PDH activity and enhancing aerobic glycolysis [35]. In ovarian cancer, high PDK4 expression is closely related to enhanced cell proliferation and poor prognosis [36]. Meanwhile, PDK4 can also enhance tumor cell tolerance to chemotherapeutic agents by maintaining glycolysis and reducing ROS-induced apoptosis in bladder and ovarian cancer [36,37]. These studies suggest that PDK4 is not only a critical metabolic regulator, but may also be a potential therapeutic target and prognostic biomarker in a variety of solid tumors. In the clinical correlation analysis, we observed that IRCDI was strongly correlated with tumor stage, with higher IRCDI values typically found in high-grade tumors, and they presented as poor prognostic factors in survival analysis. Further mutation analysis showed that the mutation status of TP53 and PIK3CA synergistically affected the prognosis of patients with IRCDI score, suggesting that IRCDI is not only an immunocompetent indicator, but also closely related to tumor genomic instability. Notably, GSEA enrichment analysis showed that there were significant differences in multiple tumor-related pathways between the high and low IRCDI groups, especially in the low IRCDI group enriched in immune-related pathways such as interferon gamma response and IL6/JAK/STAT3 signaling pathway, further strengthening its theoretical basis as an indicator of immune activity. The IL-6/JAK/STAT3 signaling axis is recognized as a key oncogenic driver across multiple malignancies. It contributes extensively to tumor initiation and progression by modulating cellular proliferation, suppressing apoptotic processes, promoting neovascularization, and facilitating immune evasion [38]. Moreover, the IL-6/JAK/STAT3 axis remains persistently activated in numerous solid tumors, playing a pivotal role in sustaining malignant phenotypes. Targeted inhibition of this pathway has been shown to effectively hinder tumor progression, highlighting its potential as a promising therapeutic target [39]. Subsequent investigations revealed that interferon-gamma (IFN-γ) secreted by tumor-associated macrophages can upregulate PD-L1 expression through activation of the JAK/STAT3 and PI3K/AKT signaling pathways, thus facilitating immune evasion and advancing lung cancer progression [40]. In conclusion, the IL-6/JAK/STAT3 pathway plays a pivotal role in regulating both tumor cell behavior and the immune microenvironment. By integrating single-cell data, we analyzed IRCDI expression across various cell types and found it to be predominantly expressed in B cells, T cells, and myeloid cells, with a significant negative correlation to activated CD4/CD8 T cells. Additionally, patients with low IRCDI levels exhibited higher PD-1 expression, suggesting that they may benefit more from immune checkpoint inhibitors. While IRCDI did not show significant immune response prediction capability in the TIDE algorithm, it demonstrated strong treatment benefit prediction in multiple real-world immunotherapy cohorts, such as IMvigor210 and GSE91061, indicating that it reflects actual treatment outcomes. Furthermore, in terms of drug sensitivity, IRCDI exhibited opposing sensitivity patterns to agents like BI-25,361,086 and talazoparib1259, highlighting its potential not only for guiding immunotherapy but also for tailoring the selection of traditional therapies. In addition, our study has limitations. In this study, bioinformatics analysis was mainly performed based on a publicly available database, which lacks experimental validation, and the clinical applicability of the results still needs further assessment. In addition, the tumor microenvironment is highly heterogeneous and may not reflect its complexity comprehensively by a single data type alone. In the future, multi-omics data and experimental validation need to be combined to enhance the reliability and persuasion of the study. We acknowledge that the single‑cell RNA‑seq dataset GSE181919 includes samples from multiple head and neck subsites and is not exclusively derived from oral cavity tumours. Although we selected samples annotated as oral‑cavity origin based on the original metadata, we cannot fully exclude potential confounding effects from anatomical heterogeneity. Consequently, our cell‑type‑specific IRCDI findings should be interpreted with caution. Dedicated single‑cell studies using exclusively oral‑cavity cohorts are warranted to further validate the distribution and functional relevance of IRCDI in oral cancer. Conclusion In summary, IRCDI, as a comprehensive scoring system based on immune-related cell death characteristics, has both prognostic evaluation, immune status judgment and potential therapeutic guidance. It can reflect the activity status of the tumor immune microenvironment in OC patients, assist in identifying patients who may benefit from immunotherapy, and suggest differences in drug sensitivity, with good prospects for clinical translation. Consent for publication Written informed consent for information about the patients to be published was obtained from all the study participants. Ethics approval and consent to participate Not applicable. Clinical trial number Not applicable. Funding Not applicable. CRediT authorship contribution statement Jinglin: Formal analysis, Data curation, Conceptualization. Jiao Lu: Data curation, Conceptualization. Guihong Liu: Formal analysis, Data curation. Peijun Qi: Methodology, Investigation. Yingying Zong: Investigation, Data curation. Hang Yu: Formal analysis, Data curation. Zhaochen Hu: Formal analysis, Data curation. Jing Ding: Visualization. Zhao Liu: Writing – review & editing, Methodology, Investigation. Liangsicheng Tang: Writing – review & editing, Supervision, Conceptualization. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments Thanks to all participants and staff for their contributions. Footnotes Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.tranon.2026.102987. Contributor Information Zhao Liu, Email: xyfylz@163.com. Liangsicheng Tang, Email: tlsctlsc@163.com. Supplementary materials References Associated Data Supplementary Materials