DNA-methylome-derived epigenetic fingerprint as an immunophenotype indicator of durable clinical immunotherapeutic benefits in head and neck squamous cell carcinoma

Abstract

Background

Cancer immunotherapy provides durable response and improves survival in a subset of head and neck squamous cell carcinoma (HNSC) patients, which may due to discriminative tumor microenvironment (TME). Epigenetic regulations play critical roles in HNSC tumorigenesis, progression, and activation of functional immune cells. This study aims to identify an epigenetic signature as an immunophenotype indicator of durable clinical immunotherapeutic benefits in HNSC patients.

Methods

Unsupervised consensus clustering approach was applied to distinguish immunophenotypes based on five immune signatures in The Cancer Genome Atlas (TCGA) HNSC cohort. Two immunophenotypes (immune ‘Hot’ and immune ‘Cold’) that had different TME features, diverse prognosis, and distinct DNA methylation patterns were recognized. Immunophenotype-related methylated signatures (IPMS) were identified by the least absolute shrinkage and selector operation algorithm. Additionally, the IPMS score by deconvolution algorithm was constructed as an immunophenotype classifier to predict clinical outcomes and immunotherapeutic response.

Results

The ‘Hot’ HNSC immunophenotype had higher immunoactivity and better overall survival (p = 0.00055) compared to the ‘Cold’ tumors. The immunophenotypes had distinct DNA methylation patterns, which was closely associated with HNSC tumorigenesis and functional immune cell infiltration. 311 immunophenotype-related methylated CpG sites (IRMCs) was identified from TCGA-HNSC dataset. IPMS score model achieved a strong clinical predictive performance for classifying immunophenotypes. The area under the curve value (AUC) of the IPMS score model reached 85.9% and 89.8% in TCGA train and test datasets, respectively, and robustness was verified in five HNSC validation datasets. It was also validated as an immunophenotype classifier for predicting durable clinical benefits (DCB) in lung cancer patients who received anti-PD-1/PD-L1 immunotherapy (p = 0.017) and TCGA-SKCM patients who received distinct immunotherapy (p = 0.033).

Conclusions

This study systematically analyzed DNA methylation patterns in distinct immunophenotypes to identify IPMS with clinical prognostic potential for personalized epigenetic anticancer approaches in HNSC patients. The IPMS score model may serve as a reliable epigenome prognostic tool for clinical immunophenotyping to guide immunotherapeutic strategies in HNSC.

Graphical Abstract

Supplementary information

The online version contains supplementary material available at 10.1007/s13402-024-00917-x.

Introduction

Head and neck squamous cell carcinomas (HNSC) are the sixth most common malignancies worldwide with increasing incidence [1, 2]. Most patients present with locally advanced stage with a high risk of recurrence, and the median overall survival (OS) for patients with recurrent/metastatic (R/M) disease is only 10–13 months [3]. Recently, immune checkpoint inhibitors (ICIs) alone or in combination with platinum-based chemotherapy has shown improvements in OS in the treatment of R/M HNSC [4–7]. Nonetheless a considerable proportion of patients do not benefit from ICI therapy [8].

Chemical carcinogenesis including tobacco and alcohol and human papillomavirus (HPV) are the two major etiologies of HNSC patients [9]. HPV-positive (HPV+) and HPV-negative (HPV−) HNSC patients have distinct prognosis and biological characters, with HPV+ patients typically associated with favorable survival and high response rate to multiple therapy including immunotherapy [10–12]. Therefore, HPV status needs to be considered when in the study of immune microenvironment and prognosis in HNSC.

Unlike targeted therapies where the biomarker is usually a single genetic aberration in the target itself (e.g. EGFR mutations), DCB for immunotherapy requires an orchestrated immune microenvironment of antigen-presenting cells as well as tumor-infiltrating lymphocytes (TILs). Great efforts have been made to search for molecular biomarkers for the efficacy of immunotherapy in HNSC, including tumor mutation burden (TMB) [13], and programmed death ligand-1 (PD-L1) expression [14]. However, previous studies showed that even for patients with positive PD-L1 expression, the response rate for ICI therapy remains no more than 30% [15]. New biomarkers that reflect TME still needed to be identified to guide clinical treatment decision.

Comparisons of global gene-expression profiles of different disease states and treatment outcomes enabled a comprehensive understanding of tumor microenvironment, expression-based immune signatures such as transforming growth factor β (TGF-β) response, interferon γ (IFN-γ) response have been implemented to analyze tumor immunophenotypes. But none has been implemented in routine clinical use yet for the lack of a universal cut-off value and the susceptible degradation nature of mRNA [16]. Within the past decade, aberrant DNA methylation pattern has gained considerable attention, because their alterations are relatively stable and potentially reversible. As for immune response, recent studies demonstrated thar T cell exhaustion results in wholescale epigenetic remodeling, and epigenetic immunoediting may drive an acquired immune evasion in glioblastoma [17, 18]. Thus, unraveling DNA methylation profiles between different immunophenotypes may provide a useful biomarker for ICI intervention decision-making and might offer potential therapeutic targets for HNSC patients.

In this study, we used the gene expression of TCGA-HNSC dataset to identify the ‘Hot’ and ‘Cold’ immunophenotypes through consensus clustering, we found the immune ‘Hot’ tumors presented global hypermethylation than ‘Cold’ phenotype and then depicted the DNA-methylome-derived epigenetic fingerprint between immunophenotypes. Subsequently, an immunophenotype-related methylated signature (IPMS) score was developed to classify the immunophenotypes in TCGA-HNSC dataset and was validated in five independent HNSC datasets. It was also validated as an immunophenotype classifier for predicting DCB in patients who received distinct immunotherapy.

Materials and methods

Data processing

The procedure of this study is depicted in Fig. 1. RNA-sequencing (RNA-seq) and DNA methylation data for TCGA-HNSC and TCGA skin cutaneous melanoma (TCGA-SKCM) datasets along with the related clinical data were downloaded from the UCSC Xena browser (https://xenabrowser.net) and the therapeutic information was obtained from TCGA (https://portal.gdc.cancer.gov/). RNA-seq data for non-small cell lung cancer (GSE135222, n = 27, part of patients from GSE119144), DNA methylation data for head and neck carcinomas (GSE75537, n = 53, 52 with survival outcome; GSE38271, n = 72, 42 normalized patient samples used; GSE79556, n = 82; GSE124464, n = 64; GSE178218, n = 20) and DNA methylation data for non-small cell lung cancer (GSE119144, n = 60, 58 with survival and therapeutic response outcomes) were downloaded from Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/). RNA-sequencing data (fragments per kilo base of transcript per million mapped fragments) from TCGA were transformed into transcripts per kilobase million [19]. All used cohorts were listed in Table S1.

Fig. 1.

Flow chart of the current study. The immunophenotypes were identified and IPMS score as the immunophenotype-classifier was established based on TCGA-HNSC cohort by deconvolution algorithm. IPMS score was validated in the GSE75537, GSE38271, GSE79556, GSE124464, and GSE178278 HNSC GEO cohorts and two cohorts received immunotherapy (GSE119144, TCGA-SKCM). The Graphical abstract was created by Figdraw

Consensus clustering analysis

Five immune signatures (‘macrophages/monocytes’, ‘lymphocyte infiltration’, ‘transforming growth factor β [TGF-β] response’, ‘interferon γ [IFN-γ] response’, and ‘wound healing’) based on a previous study were used to represent tumor immunity in TCGA [20]. The single-sample gene set enrichment analysis (ssGSEA) scores of the five immune expression signatures for TCGA-HNSC primary tumor samples were calculated using R package gene set variation analysis (GSVA) [21, 22]. Next, we grouped TCGA-HNSC cancer samples by consensus clustering based on ssGSEA scores using the ConsensusClusterPlus package [23]. The proportion of ambiguously clustered pairs (PAC) [24] was used to determine the optimal number of clusters (k = 2). We also applied the Uniform Manifold Approximation and Projection algorithm (UMAP) [25] using the R package umap to visualize the HNSC samples in two dimensions.

Immunological and molecular features

We evaluated the immunological features of the two immunophenotypes (IMC1 and IMC2): (i) tumor-infiltrating immune cell abundances were estimated using the R package CIBERSORT [26]; (ii) ESTIMATE scores were computed using the R package ESTIMATE [27]; (iii) the scores for the major histocompatibility complex (MHC) were calculated by average gene expression levels of the core MHC-I set [28]; (iv) cytolytic activity scores (CYT score) computed as the geometric means of GZMA and PRF1 genes [29]; (v) immune score calculated using the R package xCell [30]; and (vi) ssGSEA scores of the five immune signatures.

The molecular features of two immunophenotypes were assessed using: (i) The tumor mutational burden (TMB) per megabase was calculated by dividing the total number of mutations per sample by the size of the coding region in the genome [31].; (ii) aneuploidy scores based on a previous study [32]; and (iii) average levels of DNA methylation were measurable by calculating the average values of methylation for all CpG sites in each patient. The correlation between average DNA methylation and immune-related traits was assessed using Pearson’s correlation analysis (R package corrplot).

Differential methylation analysis of the immunophenotype-related methylated CpG sites (IRMCs)

DNA methylation data (Illumina methylation450 BeadChip; Illumina, Inc., San Diego, CA, USA) for TCGA-HNSC was normalized using the R package ChAMP [33]; samples with > 20% missing values were filtered out and 468 TCGA-HNSC samples were used. We also compared the β-value of each CpG site between the Hot and Cold immunophenotypes to obtain the differential methylation probes using the Limma package [34] and the Bumphunter algorithm based on the ChAMP package. CpG sites with p-values < 0.01 and |Δβ| > 0.2 (Δβ = βHot – βCold) were defined as IRMCs. Epigenome-Wide Association Study (EWAS) analysis was performed using the EWAS open platform (https://ngdc.cncb.ac.cn/ewas/index) [35]. The target genes of up/down-regulated CpG sites in the Hot immunophenotype were selected for gene ontology (GO) analysis using the R package clusterProfiler.

Model construction and efficacy assessment

We obtained 35 IPMS using the Least Absolute Shrinkage and Selector Operation (LASSO) algorithm [36] with penalty parameter tuning conducted by 10-fold cross-validation in TCGA dataset from 311 IRMCs. Next, we created the IPMS matrix by using the median beta-value of the IPMS sites for each tumor immunophenotype in the TCGA train dataset. The robust partial correlations algorithm was used to deconvolute intrasample heterogeneity [37, 38]. Based on the IPMS matrix (C) representing the tumor immunophenotypes, the DNA methylation profile (m_c) of a specific sample (y) can be expressed as:

where ω_c denotes the weight coefficient of each immunophenotype. The model determines ω_c using a least-squares approach but within the constraints of the IPMS. ω_c represents the probability that sample y belongs to immunophenotype c. Deconvolution was conducted using the R package EpiDISH [39]. Furthermore, we evaluated tumor immunoactivity by computing an IPMS score using the formula:

where IPMS score = [−1, 1]; IPMS scores > 0 were classified as the ‘Hot’ immunophenotype, whereas IPMS scores < 0 were classified as the ‘Cold’ immunophenotype. Thus, we validated the efficacy of IPMS in five independent HNSC cohorts and two immunotherapy cohorts. We first determined the immunophenotype of each HNSC sample and then measured the infiltration levels of the four immune cells (CD8+ cytotoxic T-lymphocytes, CD19+ B-lymphocytes, CD56+ NK cells, and CD14+ monocyte lineage) using the R package MethylCIBERSORT [40]. To confirm the associations, we conducted Spearman’s correlation analysis between immune cell infiltration and IPMS scores. The p-values were adjusted using the Benjamin–Hochberg procedure.

Statistical analysis

The R software (v4.1.0) was utilized for all statistical analyses. Unpaired Student’s t test and Mann–Whitney U test were employed for variables with a normal and non-normal distribution, respectively. The categorical variables were compared using the Chi-square test. Survival analysis was conducted using the log-rank test. The survminer package was used for the univariate and multivariate Cox regression models to identify independent prognostic factors. The ggforest package was utilized to present the results visually. Two-tailed p-values < 0.05 were considered statistically significant.

Results

Identification of ‘Hot’ and ‘Cold’ HNSC immunophenotypes

First, the five core immune signatures were used to characterize intratumoral immune states [20] and their ssGSEA score were used to form two main clusters (IMC1 and IMC2) to further depict intratumoral immune heterogeneity and microenvironment of TCGA-HNSC primary tumor tissues (n = 493) (Fig. S1A). UMAP analysis demonstrated that the two immunophenotypes were also categorized into two dimensions (Fig. 2A), which indicated the efficiency of consensus clustering. The immunological and molecular characteristics were also compared to understand the inherent biological differences related to different clinical presentations. CIBERSORT [26] demonstrated that CD8 T cells, CD4 T cells, M1 macrophages, and B cells were more enriched in IMC2, while M0 macrophages were more common in IMC1 (Fig. 2B). Furthermore, IMC2 had significantly higher immune score, including ESTIMATE, MHC, CYT activity, and Xcell immune score (Fig. 2D). IMC2 centered on four immune-related signatures, including macrophages, lymphocyte infiltration, IFN-γ response, and TGF-β response, but performed poorly in wound healing (Fig. 2D). Therefore, IMC1 and IMC2 subtypes were defined as the ‘Cold’ and ‘Hot’ immunophenotypes, respectively.

Fig. 2.

Multi-omics characteristics of two cluster HNSC immunophenotypes. A UMAP plot of two immune immunophenotypes in HNSC tumor samples. B Boxplot of relative immune cell abundance for two immunophenotypes in TCGA-HNSC (ns: not significant, **p < 0.01, ***p < 0.001, ****p < 0.0001). C Kaplan–Meier survival analysis of the two cluster immunophenotypes in TCGA cohort (p = 0.00055). D The immunological characteristics of the HNSC immunophenotypes. E–G Molecular characteristics of the two immunophenotypes. Differences in E tumor mutation burden (TMB), F aneuploidy score, and G average methylation levels were assessed using the Wilcoxon test (*p < 0.05, ****p < 0.0001)

Molecular features and prognostic significance of the two immunophenotypes

To explore the prognostic implications of the immunophenotypes, a multi-Cox regression model was built and corrected for age, sex, disease stage, HPV status, smoking and anatomical location. The results showed that the ‘Hot’ immunophenotype was a significant protective factor compared to the ‘Cold’ immunophenotype (hazard ratio [HR] = 0.65, p = 0.005; Fig. S1B). Meanwhile, the two immunophenotypes showed significantly different overall survival (p = 0.00055; Fig. 2C) and disease-free survival (p = 0.017; Fig. S1C) rate. Multi-Cox and chi square test showed HPV positivity was a significant protective factor and enriched in Hot immunophenotype, which was consistent with previous studies and confirmed the accuracy of cluster immunophenotypes [10–12] (Fig. S1B, Table 1). Subsequent IPMS scores analysis between HPV+ and HPV− patients also proved this (Fig. S1D). Furthermore, we next did stratified analysis in order to explore whether immunophenotype provides additional biological and prognostic information besides HPV status. The result revealed that the HPV status did not interfere with the prognostic and immune stratification of immunophenotypes. The result revealed that HPV+ patients with different immunophenotypes presented distinct immune cell abundance, so as for HPV− patients (Fig. S1E). Kaplan–Meier survival analysis indicated that patients with Hot immunophenotype had longer overall survival in both HPV+ and HPV− group (p = 0.00028, Fig. S1F). Collectively, these results shows that immunophenotype provides additional immunological and prognostic information besides HPV status.

Table 1.

Clinicopathological characteristics of patients for two HNSC immunophenotypes in TCGA

Characteristics	Immunephenotype		P-value
	Cold (N = 237)	Hot (N = 256)
Sex			0.127
Female	55	76
Male	182	180
Age			0.515
≤60	119	120
>60	118	136
Stage			0.622
Stage I-II	51	61
Stage III-IV	179	188
HPV			<0.001
Negative	200	189
Positive	14	46
Radiation_therapy			0.069
NO	108	95
YES	129	161
Chemotherapy			0.138
NO	165	161
YES	72	95
Smoking			0.133
NO	116	107
YES	121	149
Location			0.053
Hypopharynx	7	3
Larynx	54	56
Oral	161	164
Oropharynx	15	33

The molecular differences were further analyzed between the two immunophenotype. Among the genetic variants, both ‘Hot’ and ‘Cold’ immunophenotypes had low overall TMB levels (< 10 [31]) (Fig. 2E). Comparing to the ‘Cold’ immunophenotype, the ‘Hot’ immunophenotype showed a lower aneuploidy score (Fig. 2F), which was associated with cancer progression and immune evasion [32, 41]. Among epigenetic alterations, the ‘Cold’ immunophenotype had significantly lower global DNA hypomethylation compared to the ‘Hot’ immunophenotype (Fig. 2G), which has previously been linked to the carcinogenic process [42]. This showed that global hypermethylation in the ‘Hot’ immunophenotype was highly associated with better overall survival in HNSC patients.

DNA methylation patterns of two HNSC immunophenotypes

Based on methylation differences between the two immunophenotypes, the association between methylation and immune-related traits was explored. The correlation analysis showed that the methylation level was significantly associated with improved immune characteristics and negatively correlated with tumor aneuploidy that was also associated with the markers of immune evasion and reduced response to immunotherapy [43] (Fig. 3A). This suggested that DNA methylation played an important role in HNSC immunophenotyping. The DNA methylation differential analysis was further performed to determine methylation variations between the two immunophenotypes using the ChAMP package [33]. This resulted in the identification of 311 IRMCs (7 down-regulated and 304 up-regulated CpG sites in the ‘Hot’ group; Fig. 3B, Table S2). The heatmap (Fig. 3C) showed that the two immunophenotypes had unique methylation patterns and that the most common IRMC location was chromosome 2; Only 1.6% of the IRMCs were localized within the CpG island, while the remaining IRMCs were on the shore (17.7%), shelf (8.4%), and opensea (72.3%) (Fig. 3D). IRMCs were mapped onto the genomic locations; 42.7% were localized within promoter regions (1stExon, 5ʹUTR, TSS200, and TSS1500), while the remaining were in the gene body (34.4%) and intergenic regions (35.4%) (Fig. 3E). EWAS analysis was subsequently performed in the IRMCs [35]. In the trait-related analysis, cancer-, HNSC-, and immune-related characteristics were considerably enriched in these IRMCs (Fig. 3F, Table S3). Some IRMCs were linked to HNSC risk variables, including air pollution [44] and smoking [45], suggesting that IRMCs played an important role in HNSC tumorigenesis. A critical role of epigenetic TME alterations in cancer cell metabolism has been reported [46]; IRMCs were also related to the metabolic trait, suggesting a relationship between IRMCs and TME. The GO enrichment analysis was performed for 311 IRMCs (Table S4), which indicated that the target genes of the seven down-regulated CpG sites in the ‘Hot’ immunophenotype were immune-related, while the 304 up-regulated CpG sites were not immune-related (Fig. 3G). This indicated that immune-related genes were dysregulated in different immunophenotypes. These findings highlighted the significance of IRMCs in HNSC development, immunoregulation, and TME.

Fig. 3.

DNA methylation patterns of the two HNSC immunophenotypes. A The correlation between mean methylation levels and the 11 signatures. Blue, negative correlation; Orange, positive correlation (*P < 0.05; **P < 0.01; ***P < 0.001). B Volcano plots of beta-value profiles in TCGA-HNSC. Red/blue symbols classify upregulated/downregulated CpG sites based on the criteria. C Heatmap representing methylation levels in 311 IRMCs in the two immunophenotypes along with CpG sites annotated using different colors, Chr: chromosome. D, E The CpG-site neighborhood (D) and genomic location (E) of IRMCs. TSS: transcription start site, UTR: untranslated region, IGR: intergenic region. F IRMC traits were analyzed using EWAS. G Gene Ontology (GO) enrichment analysis of two regulatory genes of IRMCs: genes regulated by 7 down-regulated (DOWN) and 304 up-regulated (UP) CpG sites in ‘Hot’ immunophenotype on the left- and right-hand sides, respectively. The x-axis indicates the number of genes within each GO term

Construction of IPMS score for tumor immunophenotyping

To develop a brief quantitative classifier by DNA methylation in individual HNSC immunophenotyping, the LASSO algorithm was used to perform dimension reduction for IRMCs in TCGA dataset and obtained 35 CpG sites (Fig. 4A). As the heatmap shows, 6 CpG sites was hypomethylated, while the remaining sites were hypermethylated in the ‘Hot’ Immunophenotype of TCGA dataset (Fig. 4B).The correlation between the 35 CpG sites and the immune features was further explored; The hypermethylation CpG sites in ‘Cold’ immunophenotype were significantly and positively related to the immune-deficient features, wound healing and M2 macrophages, but negatively associated with other immune features, including macrophages, lymphocyte infiltration, IFN-γ response, CD8 T cells, and M1 macrophages, whereas the hypermethylation CpG sites in ‘Hot’ immunophenotype showed converse results (Fig. 4C-D). These suggested the methylation level of IPMS was significantly related with TME. Therefore, 35 CpG sites were identified as IPMS to obtain score based on the deconvolution algorithm in the TCGA train dataset [39] and distinguish between the immunophenotypes. The ‘Hot’ immunophenotype based on consensus clustering showed a higher IPMS score compared to the ‘Cold’ immunophenotype in both train and test datasets (Fig. 4E, F). IPMS-predicted immunophenotypes had accuracies of 85.9% and 89.8% in TCGA train and test datasets, respectively (Fig. 4G). The above results revealed the tremendous potential of the IPMS score model as a DNA methylation classifier in tumor immunophenotyping clinically.

Fig. 4.

Construction of IPMS score model. A Identification of the optimal penalization parameter λ using the LASSO algorithm. B IPMS and clinical features of the two immunophenotypes in TCGA-HNSC. C, D Heatmap showing the absolute value of pearson’s correlation coefficient between β-values of IPMS and immune features in TCGA-HNSC. E, F Violin plot showing IPMS scores of the two immunophenotypes by consensus clustering in TCGA train (E) and test sets (F), compared using the Wilcoxon test (****p < 0.0001); The two figures are aim to show the accuracy of IPMS score for predicting two cluster immunophenotype. G Receiver operating characteristics curve based on IPMS scores for distinguishing cluster HNSC immunophenotypes in TCGA train and test sets

Biological functions of model genes involved in IPMS

Next, the biological functions of the 35 IPMS was investigated. A fifth of the IPMS localized within the promoter regions (1stExon, 5ʹUTR, TSS200, and TSS1500), while approximately 3/5 IPMS were located in the open sea. The 35 IPMS were located on 22 genes and 13 intergenic regions (Fig. 5A, Table S5). EWAS analysis showed that the cancer-, HNSC-, and immune-related characteristics were enriched in the 35 CpG sites (Fig. 5B, Table S6). The DNA methylation levels of cg17626301/cg01673307, cg09059466 and cg10082398 were significantly associated with the expressions of TAP1, HLA-DPB1 and CLDN20 respectively (Fig. 5A). The GO and KEGG analysis of the three target genes were further performed and found multiple enriched immune-related GO and KEGG terms (Fig. 5C-D, Table S7). These findings suggested that the IPMS target genes perform key functions in tumorigenesis and tumor immunology.

Fig. 5.

Biological functions of model genes involved in IPMS. A Correlation analysis between IPMS methylation levels and their target genes in TCGA cohort. B IPMS traits assessed by EWAS. C Gene Ontology and D Kyoto Encyclopedia of Genes and Genomes enrichment analyses of three target genes

Validation of IPMS model for immunophenotyping prediction

Robustness of IPMS score was tested in 260 HNSC patients in five external validation DNA methylation cohorts (GSE75537, GSE38271, GSE79556, GSE124464, and GSE178278) by determining the relationship between IPMS score and the levels of immune cell infiltration using MethylCIBERSORT [40]. All these cohorts showed significant positive associations between IPMS score and two immune cell (CD8+ T cells and CD19+ B cells) infiltration levels. There was also a partial correlation between IPMS score and infiltration levels of CD14+ monocytes and CD56+ NK-cells (Fig. 6A).

Fig. 6.

IPMS score validation for immunophenotype prediction in the GEO cohort. A The plot (left) shows the correlation analysis between immune cell infiltration level (MethylCIBERSORT) and IPMS scores; the violin picture (right) shows differences between the immunophenotypes (ns not significant, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001). B, C Kaplan–Meier survival analysis of overall and disease-free survival in the two immunophenotypes in GSE75537 (B) (p = 0.042) and TCGA radiation-therapy cohort (C) (p = 0.047)

In subsequent analyses, the GSE75537 (n = 52) dataset was used to confirm the prognostic capacity of the IPMS score (p = 0.042; Fig. 6B). In addition, the predictive value of IPMS score in radiation therapy and chemotherapy was evaluated using TCGA-HNSC cohort. The ‘Hot’ immunophenotypes had better survival outcomes in TCGA radiotherapy (p = 0.047; Fig. 6C) and chemotherapy (Fig. S2A, B) cohorts. The multi-Cox regression models for GSE75537 and TCGA radiotherapy cohorts also verified the independent prognostic value of IPMS predicted immunophenotypes (Fig. S2C, D). These results suggest that IPMS score model is a robust classifier for distinguishing HNSC immunophenotypes and may guide personalized treatment.

The role of immunophenotypes in predicting immunotherapeutic benefits

ICIs used in cancer treatment block T cell inhibitory molecules and have shown remarkable results, but not in all patients [6, 8, 13]. Our previous results revealed the correlation between IPMS score and immune cell infiltration levels, as well as the ability of IPMS score in identifying tumor immunophenotypes. To further explore the relationship between the IPMS-predicted immunophenotypes and the benefits of immunotherapy, two DNA methylation cohorts (GSE119144/GSE135222 and TCGA-SKCM with immunotherapy) were collected and subsequently analyzed. We found that the ‘Hot’ immunophenotype had higher immune cell infiltration levels (CD8+ T cells and CD19+ B cells) using MethylCIBERSORT [40] and increased immune signatures (CD8+ T effectors and immune checkpoints) [47] using ssGSEA in the two immunotherapy cohorts (Fig. 7A, B). This suggested that the ‘Hot’ immunophenotype may play a key role in treatment outcomes of immunotherapy. Notably, patients with the ‘Hot’ immunophenotype had a significantly better prognosis (Fig. 7C, F). In addition, the ‘Hot’ immunophenotype also had higher response rate to immunotherapy in GSE119144 and TCGA-SKCM (Fig. 7D, G). Notably, 12 of 14 DCB patients were identified as the ‘Hot’ immunophenotype in GSE119144 and 13 of 19 CR/PR(DCB) patients were identified as the ‘Hot’ immunophenotype in TCGA-SKCM (Fig. 7E, H). Receiver operating characteristic curves showed that there was a predictive advantage of IPMS score when compared with the expression levels of PDL1 in both GSE119144 and TCGA-SKCM (Fig. S3A, B). These findings strongly indicate that the IPMS score model may evaluate immunophenotypes of patients and predict the response to immunotherapy.

Fig. 7.

The role of immunophenotypes in predicting immunotherapeutic benefits. A, B The plot (left) shows the correlation analysis between immune cell infiltration levels (MethylCIBERSORT) (A) or ssGSEA scores of immune signatures (B) and IPMS scores; the violin picture (right) shows differences between the immunophenotypes (ns not significant, *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001). B Correlation analysis between PD1, PDL1, and CTLA4 (from left to right) and IPMS scores in GSE135222 cohort. C Kaplan–Meier survival analysis of the two immunophenotypes in the GSE119144 cohort (p = 0.017). D Durable clinical benefit rate of immunotherapy in the two immunophenotypes in the GSE119144 cohort (DCB: Durable clinical benefit; NDB: non-durable clinical benefit). E Proportion of immunophenotypes in GSE119144 DCB patients. F Kaplan–Meier survival analysis of the two immunophenotypes predicted by IPMS in TCGA-SKCM cohort (p = 0.033). G Response rate for various immunotherapies in the two immunophenotypes in TCGA-SKCM cohort. H Proportion of immunophenotypes in TCGA-SKCM CR/PR patients (CR complete response, PR partial response, SD stable disease, PD progressive disease)

Discussion

HNSC is among the most inflamed, immune-infiltrated cancers, particularly with CD8+ lymphocytes and natural killer (NK) cells [48]. Although anti-PD-(L)1 antibodies have been recommended as front-line treatment for R/M HNSC patients, the response rate are no more than 20%. Even for patients with positive PD-L1 expression, the response rate remains no more than 30% [15]. As for genomic stability marker like tumor mutation burden, previous studies indicated that HNSC is characterized by the presence of inflammation without high TMB [13]. Epigenetic modifications such as PD-L1 methylation, disruption of methylcytosine dioxygenase 2 were relevant to ICI response and may be a promising route for therapeutic intervention [49]. Although a few methylation sites have been used as prognostic indices [50, 51], the predictive value of DNA methylation profiles for immunophenotypes, prognosis and response of immunotherapy in HNSC has not been investigated.

In the present study, we categorized 493 primary HNSC patients into ‘Hot’ and ‘Cold’ immunophenotypes based on the ssGSEA scores of five immune signatures in TCGA-HNSC cohort. As expected, there were differences in immune cell infiltration levels and immune characteristics between the ‘Hot’ and ‘Cold’ immunophenotypes. The ‘Hot’ immunophenotype showed more immune-activity with better overall and disease-free survival rates while the ‘Cold’ immunophenotypes had tumorigenesis features, including high chromosome aneuploidy and hypomethylation [32, 52].

We then investigated epigenetic alterations and the correlation between methylation levels and immune signatures in the two immunophenotypes. The methylation level was significantly associated with immune characteristics and response to immunotherapy. Consistent with previous studies in lung cancer and melanoma, global loss of DNA methylation in HNSC also correlated with immune evasion signatures and poor prognosis [52]. We subsequently explored differentially methylated DNA sites between the two immunophenotypes to obtain the IRMCs. We then performed EWAS analysis [35]; the target genes of down-regulated methylated CpG sites were enriched in immune-related pathways on GO analysis, which demonstrated the crucial role of IRMCs in oncogenesis, gene expression regulation, and TME. IPMS were obtained from IRMCs using the LASSO algorithm in TCGA training dataset and the scores were calculated using the deconvolution algorithm. IPMS was verified in TCGA test dataset and five HNSC GEO validations. IPMS exhibited promising efficiency for classifying immunophenotypes at different immune cell infiltration levels and further analysis of IPMS-predicted immunophenotypes in TCGA radiotherapy and chemotherapy cohorts revealed its potential as an indicator for HNSC clinical treatment.

Additionally, 35 CpGs of the IPMS were significantly related to cancer and immune characteristics in EWAS. DNA methylation levels of cg17626301/cg01673307, cg09059466 and cg10082398 were significantly associated with the expressions of TAP1, HLA-DPB1 and CLDN20 respectively, suggesting a regulatory role of DNA methylation and gene expression. TAP1 had key functions related to MHC protein binding, antigen processing and presentation, and primary immunodeficiency and CLDN20 gene was one of the most important cross-talk genes between periodontitis and IgA nephropathy [53]. HLA-DPB1 had key functions related to lymphocyte proliferation, antigen processing and presentation, and INF-gamma-mediated signaling pathway. This may suggest that these immune-related characteristics were subject to epigenetic regulation and were involved in HNSC immunophenotyping. CLDN20 gene also was associated with metastasis and poor survival in ovarian cancer [54]. The function of IPMS and target genes further supported the validity of IPMS selection.

To further evaluate the relationship between IPMS-predicted immunophenotypes and immunotherapy benefits, we explored the importance of IPMS-predicted immunophenotypes in GSE119144/GSE135222 and TCGA-SKCM cohorts, consisting of patients treated with PD1 inhibitors and distinct immunotherapeutic agents, respectively. IPMS scores were positively correlated with the number of CD8+ T effector cells and immune checkpoints. The Hot immunophenotype with higher immune signatures identified using IPMS led to a greater benefit with immunotherapy and there was a predictive advantage of IPMS score when compared with the expression levels of PDL1. The best advantage of IPMS score was the universal cut-off value in distinct cohorts compared to previous tumor immune signatures [55, 56]. Furthermore, Our IPMS model demonstrated broader applicability based on 450 K chip than other model based on EPIC chip with a substantial number of missing values [57, 58]. The limitation to our study was a lack of large HNSC sample with immunotherapy dataset, the rapid growth of DNA methylation array datasets may solve these problems.

In conclusion, we established IPMS as an immunophenotype classifier to quantify clinical HNSC immunophenotypes. IPMS with the universal cut-off value was an efficient prognostic biomarker and predictor for TME, immunophenotypes, and response to immunotherapy. These results provide a novel insight into immunophenotyping and immunotherapy in HNSC, it has the potential to direct immunotherapy strategies and may facilitate the development of personalized epigenetic anticancer methods for different HNSC immunophenotypes.

Conclusions

By depicting DNA methylation patterns of distinct HNSC immunophenotypes, we constructed an epigenome prognostic tool for clinical immunophenotype prediction in HNSC patients. This model with a universal cut-off value has the potential to guide immunotherapeutic strategies and provides IPMS as potential therapeutic targets for epigenetic anticancer therapies.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Study design: X.T, R.Li, R.Lv and X.W; Data curation and quality control: X.T, R.Li and Y.W; Formal analysis: R.Li, R.C, W.H, X.R and X.T; Underlying data validation: X.T, R.Li, R.Lv and X.W; Manuscript draft: R.Li, X.W and X.T; Review and editing: X.T, R.Li, B.C, R.Lv and X.W; Funding acquisition: X.T, X.W and X.R. All authors read and approved the final manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (Grant No. 81903134, 82172673, and 82273148), the Natural Science Foundation of Guangdong Province (Grant No. 2019A1515110076) and the Outstanding Youths Development Scheme of Nanfang Hospital, Southern Medical University (Grant No. 2018J005).

Data availability

RNA-seq and DNA methylation data for TCGA data was downloaded from the UCSC Xena browser (https://xenabrowser.net) and GEO data was downloaded from downloaded from Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/).

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have read and agreed to the published version of the manuscript.

Competing interests

The authors declare no potential conflicts of interest.