2.1 Source of data and pre-processing

Seven publicly accessible datasets (GSE27020,¹⁵ GSE31056,¹⁶ GSE30784,¹⁷ GSE39366,¹⁸ GSE41613,¹⁹ GSE65858 [LHNG cohort],²⁰ and TCGA-HNSC cohort) were utilized to gather messenger RNAs and clinical information. Data of TCGA-HNSC datasets obtained from The Cancer Genome Atlas (TCGA) website were downloaded, specifically RNA sequencing (RNA-seq; fragments per kilobase million value) data. The R package GEOquery.²¹ was used to download the microarray data for GSE27020, GSE31056, GSE30784, GSE39366, GSE41613, and GSE65858 from Gene Expression Omnibus (GEO). To remove nonbiological technical biases, the sva package utilized the “ComBat” algorithm.²² The OS data were available for TCGA-HNSC, GSE27020, GSE31056, GSE41613, and LHNG cohorts. Data on somatic mutation and copy number variation (CNV) were obtained from the TCGA database. To depict the CNV pattern of 33 genes involved in pyroptosis on human chromosomes, the R package “Rcircos” was utilized. The quantification of tumor mutation burden (TMB) involved calculating the number of somatic alterations per megabase within the genome, as stated by Hellmann et al. in 2018.²³

2.2 Unsupervised clustering analysis using a consensus algorithm

To identify differentially expressed PRGs, a collection of 33 PRGs was extracted from six integrated GEO datasets. Pyroptosis modification patterns were identified using unsupervised clustering analysis, which was based on the 33 genes regulating pyroptosis. This analysis classified patients into three distinct clusters. To ensure classification stability, the “Pam” method underwent unsupervised clustering analysis using the “ConsensuClusterPlus” R package,²⁴ which was repeated 1000 times.

2.3 Functional enrichment analysis and gene set variation analysis

To examine the diversity in biological processes among various pyroptosis modification patterns, we utilized the gene set variation analysis (GSVA)²⁵ technique with the R package “GSVA.” For GSVA analysis, the MSigDB database was used to download the gene sets labeled as “c2.cp.kegg.v6.2.-symbols.” The R package “clusterProfiler” was used to perform KEGG annotation for genes relevant to pyroptosis, with a threshold of FDR < 0.05.²⁶

2.4 Assessment of the infiltration of cells in the tumor microenvironment

Quantification of the infiltration levels of various immune cells in HNSC was conducted through a single-sample gene set enrichment analysis (ssGSEA) as described by Barbie et al. in 2009.²⁷ Recently published studies^{27, 28} provided gene panels specifically designed for 28 different types of immune cells. The ssGSEA analysis was used to calculate enrichment scores, which represented the relative abundance of tumor microenvironment (TME)-infiltrating cells in each sample.

2.5 Differential expression analysis between distinct phenotypes

Using the “Limma” R package,²⁹ differential expression analysis was conducted to identify differentially expressed genes (DEGs) among different subtypes.

2.6 Principal component analysis dimensionality reduction and establishment of the PPI index

I represent the manifestation of genes related to the m6A phenotype.

2.7 Somatic alteration data

Data on genetic mutations in patients from the TCGA-HNSC group were obtained from the TCGA site (https//www.cancer.gov/tcga/). The calculation of TMB in HNSC was determined by considering the overall count of non-synonymous mutations. HNSC driver genes were analyzed using the ComplexHeatmap package's “oncoplot” function between high and low immune cell infiltration (ICI) scores.³² We identified the 25 most frequently altered genes with the highest mutation rate.

2.8 Prediction of response to chemotherapy/immunotherapy

While immune checkpoint blockade therapies that inhibit T-cell suppressor molecules in cancer treatment have demonstrated remarkable outcomes and the ability to enhance the advancement of advanced cancers, they may not be appropriate for every patient.^{33, 34} The therapeutic benefit of the PPI score was further analyzed in two immunotherapy cohorts: the IMvigor 2100 cohort, which involved the intervention of advanced urothelial cancer with the anti-PD-L1 antibody atezolizumab,³⁵ and the GSE78220 cohort from GEO, which focused on the treatment of metastatic melanoma with the anti-PD-1 antibody pembrolizumab.³⁶

To forecast clinical responses to immune checkpoints, the TIDE algorithm³⁷ and Subclass mapping³⁸ were employed for the estimation of tumor immune dysfunction and exclusion. Predictions of the response to chemotherapy were also made for every sample using the most extensive pharmacogenomic database accessible to the public (Genomics of Drug Sensitivity in Cancer [GDSC], https//www.cancerrxgene.org/).³⁹ Table S1 listed a total of 149 medications with the capability to address cancer. The R software's “pRRophetic” package was utilized for making predictions. The package utilized ridge regression to calculate the IC50 values of the samples and evaluated the prediction accuracy through 10-fold cross-validation using the GDSC training set.⁴⁰

3.1 The genetic variation of PRGs in head and neck cancer is reflected in the landscape

In this study, a total of 33 PRGs were ultimately discovered. Out of the 506 samples, a total of 127 exhibited mutations in m6A regulatory genes, accounting for a prevalence rate of 25.10%. In HNSC samples (Figure 1A), CASP8 displayed the highest mutation frequency, with NIRP3 following closely. Notably, TNF, PJVK, and GSDME, which are involved in prognosis regulation, exhibited no mutations. In the examination of CNV modification occurrence, a dominant CNV modification was observed in 33 regulatory genes, with the majority of genes being linked to copy number amplification. However, cysteinyl aspartate protease (caspase or CASP) family genes like CASP8, CASP4, and CASP1 exhibited a widespread occurrence of CNV deletion (Figure 1B). Figure 1C displays the position of CNV modification of PRGs on chromosomes. Figure 1D clearly showed that HNSC samples were completely differentiated from normal samples based on the expression of the 33 PRGs. An abnormal elevation of most PRGs was observed in tumor samples when compared to normal samples, except for CASP9, ELANE, IL18, and IL6 (Figure 1E). To assess the reciprocal control among the PRGs, a Spearman correlation analysis was conducted (see Figure S1). Other PRGs exhibited a notable and positive correlation with CASP5. In general, the findings indicated that there is significant diversity in the genetic and expression patterns of PRGs between normal and HNSC samples, suggesting that the expression of PRGs plays a crucial part in the onset and progression of HNSC.

3.2 Identification of pyroptosis-related molecular subtypes

In the meta-GEO cohort, we deemed six GEO datasets (LHNG cohort/GSE65858, GSE41613, GSE39366, GSE31056, GSE30784, and GSE27020) as suitable, as they had clinical data available. The prognostic values of the 33 PRGs in patients with HNSC were revealed using a univariate Cox regression model and the Kruskal–Wallis test (Figure S2A,B; Table S2). The PRG network (Figure 2A) demonstrates the synthesis of the connections, interactions, and prognostic importance of the 33 PRGs in HNSC patients. The discovery indicated that the interaction among PRGs might have a significant impact on the development of pyroptosis alteration patterns and the characterization of TME cell infiltration in distinct tumors. Using the aforementioned assumptions, the consensus cluster analysis was employed to categorize samples into distinct groups with varying pyroptosis alterations, depending on the 33 PRGs' expression. Figures 2B and S3A–C revealed the identification of three separate clusters of modification patterns, consisting of 197 patients in Cluster I, 727 samples in Cluster II, and 286 cases in Cluster III. The three clusters were compared using Kaplan–Meier survival analysis, and the log rank test was employed to determine the significant disparity in survival time. The study revealed that individuals belonging to gene cluster III exhibited a more favorable prognosis, whereas those in gene cluster I had an unfavorable prognosis for OS (based on the log rank test, with a significance level of p < 0.05; as shown in Figure 2C). In addition, the TCGA-HNSC group was confirmed and yielded comparable outcomes (log rank analysis, p < 0.05; Figure 2D). Using PCA, the expression profiles of the 33 PRGs were analyzed to identify three subgroups in the meta-GEO cohort (Figure 2E).

3.3 The infiltration properties of immune cells from the TME in the three clusters associated with pyroptosis

To reveal the biological role of the pyroptosis-related clusters, an analysis of GSVA enrichment was performed. According to GSVA, Cluster A showed significant enrichment in pathways associated with energy metabolisms, such as CITRATE_CYCLE_TCA_CYCLE, PEROXISOME, and FATTY_ACID_METABOLISM. On the other hand, Cluster B exhibited significant enrichment pathways linked to cancer activation and stromal pathways, including the NOD_LIKE_RECEPTOR_SIGNALING_PATHWAY, RIG_I_LIKE_RECEPTOR_SIGNALING_PATHWAY, T_CELL_RECEPTOR_SIGNALING_PATHWAY, and NATURAL_KILLER_CELL_MEDIATED_CYTOTOXICITY. Interestingly, Cluster C exhibited a significant enrichment in metabolic regulation and signaling pathways related to the stroma (Figure 3A–C).

Using the ssGSEA algorithm, we systematically assessed the TME cell infiltration models and TME signatures in the three ICI clusters. Figure 3D displayed the heatmap of 28 subpopulations of immune infiltrating cells within pyroptosis-associated clusters. Cluster A exhibited a significant presence of B cells that were extremely inexperienced, CD8 T cells, follicular helper T cells, regulatory T cells (Tregs), activated natural killer (NK) cells, monocytes, M0 macrophages, and resting mast cells among the three subtypes. Group B patients exhibited a notably elevated abundance of T cells CD4 memory resting, M1 and M2 macrophages, dendritic cells resting, stromal score, and immune score. Cluster C subjects exhibited a notable rise in the infiltration of plasma cells. In addition, CIBERSORT, an algorithm that employs support vector regression to assess immune cell subpopulations in the TME, was employed to provide further insights into immune infiltration within the three subpopulations. Figure S4 also showed similar findings. Hence, there is speculation that the activation of stroma in Cluster B hinders the immune cells' ability to suppress tumor growth. The additional examination also verified that stromal stimulation in the Cluster B subtype was greatly intensified, and the mechanisms associated with cell death, the pathway of signaling through toll-like receptors, and the pathway of signaling through T cell receptors were demonstrated, providing further validation of our supposition.

Spearman's correlation analyses (Figure 3E) were conducted to investigate the precise association between every type of TME infiltration cell and each PRG. Enhanced ICI showed a significant correlation with the high expression of several PRGs like CASP family members and NOD2, while neutrophil infiltration levels were negatively correlated with the expression of GPX4, CASP3, and CASP6. NLRC4 caught our interest among these PRGs due to its notable association with prognosis and immune infiltration, as depicted in Figure 3F.

3.4 DEGs associated with pyroptosis in head and neck cancer

The “limma” package of the R software was used to determine the transcriptomic variation between different pyroptosis phenotypes. The empirical Bayesian approach (Figure 4A) was used to analyze the 1276 DEGs that were common to the three clusters. Figure 4B,C illustrates that DEGs were concentrated in immune-related Gene Ontology categories, such as T cell stimulation, release of neutrophil contents, immune receptor function, and neutrophil activation in immune response. In addition, they were also found in pathways associated with immune/cancer interactions, including the TNF signaling pathway, the IL-17 signaling pathway, the NF-kappa B signaling pathway, the PI3K-Akt signaling pathway, and the differentiation of Type 17 T helper cells (Th17).

Out of these DEGs, a total of 545 DEGs were recognized as prognostic genes according to Table 1. The prognostic DEGs were then analyzed by the unsupervised clustering method. In the meta-GEO cohort, three genomic clusters, specifically gene Clusters A, B, and C, were discovered. Figure 5A displayed the transcriptome profiles of the DEGs within the three genomic clusters, as depicted by the heat map. In gene cluster C, among the 33 PRGs, AIM2, CASP1, CASP4, CASP8, GSDMD, GSDME, IL1B, IL6, NLRP1, NOD1, PLCG1, and TNF showed significant upregulation. On the other hand, gene cluster A exhibited upregulation of CASP3, CASP6, CASP9, GPX4, GSDMB, and PRKACA. Aberrant expression of GSDMA, GSDMC, IL18, NOD2, PYCARD, and SCAF11 was observed in gene Cluster B (Figure 5B). Next, the examination focused on the particular association between every type of TME infiltration cell and each PRG. Gene Cluster A exhibited elevated levels of stimulated B cell, eosinophil, stimulated CD8 T cell, and regulatory T cell. Figure 5C showed that Gene cluster C exhibited enrichment of activated CD4 T cell, activated dendritic cell, CD56 bright NK cell, CD56 dim NK cell, gamma delta T cell, immature B cell, MDSC, macrophage, mast cell, monocyte, NK T cell, NK cell, regulatory T cell, Th1 cell, Th17 cell, and Th2.

According to survival analysis, patients with gene cluster C exhibited a more favorable prognosis, whereas those with gene cluster A showed an unfavorable prognosis for OS (log rank test, p < 0.001; Figure 5D). Furthermore, it was noted that individuals with a progressed clinical phase were characterized by the gene Cluster B subgroup, while those exhibiting decreased PD-L1 expression were predominantly found in gene Cluster B and C subgroups (Figure 5E). The results also indicated that stratifications exhibited distinct clinicopathologic characteristics and were identified as gene Clusters A, B, and C (Figure 5F,G,I). The regulatory genes showed notable variations among the three subgroups of PRGs signature, aligning with the anticipated outcomes of the pyroptosis modification pattern.

3.5 Establishment of the PPI index and exploration of its clinical relevance

While our results validate the involvement of pyroptosis in prognosis and regulation of immune invasion, it is important to note that these analyses were solely conducted on patient populations and did not provide an accurate prediction of the alteration of pyroptosis in an individual tumor. Hence, an indicator of future outcomes score was formulated and designated as the PPI index relying on the PCA algorithm. The meta-GEO cohort patients were divided into two categories based on their PPI scores, using the most accurate threshold values. Figure 6A illustrates the allocation of patients among the three gene clusters. Figure 6B also assessed the correlation between established immune signatures and the PPI score to provide a clearer depiction of the PRG signature's characteristics. In Figure 6C, pyroptosis Cluster A exhibited a greater PPI compared to Cluster C and Cluster B, while in Figure 6D, gene Cluster A displayed a higher PPI than Cluster C and Cluster B. Spearman's correlation analysis was utilized to investigate the association between identified biological traits and PPI scores. The scatter diagram indicated a significant positive correlation between PPI scores and cell cycle regulators (Figure 6E) as well as the WNT pathway (Figure 6F) while displaying a negative correlation with mismatch repair (Figure 6G) and angiogenesis (Figure 6H).

Following the assessment of the predictive significance of PPI scores, the meta-GEO cohort was examined to analyze immune activity and tolerance conditions. In Figure 6I, the Wilcoxon test indicated that the high PPI group exhibited significant overexpression of the majority of genes related to immune checkpoints and immune activity. Gene set enrichment analysis (GSEA) revealed that GLYCEROPHOSPHOLIPID_METABOLISM, FATTY_ACID_METABOLISM, and the ERBB_SIGNALING_PATHWAY were significantly enriched in the higher PPI score group, whereas ECM_RECEPTOR_INTERACTION, the JAK_STAT_SIGNALING_PATHWAY, TOLL_LIKE_RECEPTOR_SIGNALING_PATHWAY, NATURAL_KILLER_CELL_MEDIATED_CYTOTOXICITY, and the B_CELL_RECEPTOR_SIGNALING_PATHWAY were enriched in the lower PPI group (Figure 6J).

According to the Kaplan–Meier plotter, patients in the lower PPI score group exhibited significantly reduced OS rates compared to those in the higher PPI score group in the meta-GEO cohort (log rank test, p < 0.05; Figure 7A), as well as in GSE31056 (log rank test, p < 0.05; Figure 7B), GSE41613 (log rank test, p < 0.05; Figure 7C), GSE65858 (log rank test, p < 0.05; Figure 7D), and TCGA HNSC (log rank test, p < 0.05; Figure 7E) cohorts.

3.6 Relationship between PPI scores and somatic mutations

Increasing evidence suggests a link between genetic mutations in the tumor genome and the response to immunotherapy. Apart from the expression of PD-L1, TMB serves as a separate biomarker for the response to ICI.^{41, 42} The presence of genetic mutations in tumors is linked to the effectiveness of immunotherapy in various types of tumors and treatment methods, such as CPIs and cellular-based therapy.^{42, 43} This is a metric for genetic changes in a tumor and is a significant genomic indicator that is strongly linked to immunotherapy and predicting survival.^{44, 45} To understand the genetic impact of each PPI subgroup, the relationship between TMB and PPI scores was investigated, given the significant clinical significance of TMB. According to Figure 8A, individuals belonging to the high PPI score subgroup exhibited a notably greater TMB compared to those in the low PPI score subgroup (Wilcoxon test p < 0.05). Furthermore, the correlation analysis revealed a significant and positive correlation between PPI scores and TMB (Spearman coefficient R = 0.35, p < 0.05; Figure 8B). According to the findings in Figure 8C, patients with a low TMB exhibited a superior PPI compared to individuals with a high TMB (log rank test, p < 0.05). The analysis of survival, stratified by TMB status, revealed an association between predictions made using PPI scores. Both the high and low TMB subgroups showed significant differences in survival (log rank test, p < 0.05 for high TMB and high PPI scores (HH) compared to high TMB and low PPI scores (HL), and low TMB and high PPI scores (LH) compared to low TMB and low PPI scores (LL); Figure 8D). In general, these discoveries indicate that the PPI score could be unrelated to TMB and a reliable gauge of possible factors that can determine the reaction to immunotherapy.

Moreover, maftools³² were utilized to analyze significantly mutated genes in HNSC samples, comparing the low and high PPI score subgroups. Using a chi-square test, Figure 8E and Table 2 helped identify the 25 driver genes that had the most frequent alterations. Significant differences were observed between the low and high PPI score subgroups for NSD1, PTPRZ1, GRM1, ELMO1, FLNA, VPS13C, CUX2, and PKD1L1 among these genes. The discovery of these findings could offer fresh perspectives on the processes of tumor PPI composition and gene mutation in immune checkpoint blockade treatment.

3.7 PPI score for response to immunotherapy

We assessed the correlation between clinical characteristics and established PPI scores. The advanced TNM stage (Figure 9A,B) was correlated with high PPI scores. The analysis of survival, stratified by PPI scores, indicated a correlation with survival outcomes (Figure 9C,D). Furthermore, there was a positive correlation between higher PPI scores and objective response to therapy in the meta-GEO cohort, as depicted in Figure 9E,F. In the meantime, stratified survival analysis indicated a correlation between the PPI and the outcome success of the primary therapy. Among patients with partial/complete response, high PPI scores were related to worse survival, similar results were obtained among patients with stable/progressive disease (Figure 9G,H).

Despite not being approved as a conventional treatment for HNSC, immunological checkpoint medications were assessed using the submap algorithm to estimate the probability of immunotherapy response in the meta-GEO cohort. The findings revealed that individuals with high PPI scores were more likely to exhibit a favorable response to immunotherapy compared to those with low PPI scores (p < 2.2e−16). Furthermore, immunotherapy yielded more favorable outcomes in individuals with high PPI scores as per the TIDE algorithm (p < 0.05). The findings indicated that individuals with elevated PPI scores exhibited strong responsiveness to programmed cell death 1 (PD1) immunotherapy (p < 0.05 after Bonferroni correction) (Figure 10A). Treating patients with HNSC often involves the use of anticancer medication as a fundamental therapeutic approach. The purpose of the PPI model was to evaluate the reaction to antitumor medications in two molecular subcategories using the GDSC cell line dataset. The accuracy of the method's prediction was validated using 10-fold cross-validation, while the IC50 was utilized to estimate the sensitivity of the response. Out of the 130 categories of anticancer medication reactions, 74 drugs exhibited variances among the two PPI clusters as indicated in Table 3. The findings indicated that elevated PPI scores exhibited greater responsiveness to 68 different types of medications in comparison to lower PPI scores. Figure 10B showed that temsirolimus, camptothecin, pazopanib, parthenolide, and docetaxel exhibited significant promise in treating HNSC (p < 0.05). Specifically, camptothecin and docetaxel have been extensively utilized as primary chemotherapy in clinical settings.