Background

Head and neck cancer has a yearly incidence of over 800,000 new cases worldwide, with a mortality rate of more than 7%. Head and neck squamous cell carcinoma (HNSCC) accounts for 90% of head and neck cancers [1]. HNSCC are related to continuous exposure to tobacco, tobacco products, and alcohol. In recent years, an increasing proportion of HNSCC caused by human papillomavirus (HPV) infection has been reported [2, 3]. Patients with HPV-associated HNSCC are younger and more responsive to chemotherapy and radiotherapy [4,5,6]. However, HNSCC is prone to regional lymph node metastasis and distant metastasis, leading to a relatively poor prognosis [7].

HNSCC is treated with a comprehensive treatment strategy, including radiotherapy (RT), surgery, chemotherapy, and targeted therapy et al. However, patients who have significant benefit from therapy remain unknown. Studies have shown that genetic alterations and immune characteristics in the tumor microenvironment (TME) are closely associated with therapy sensitivity and effectiveness [8,9,10]. Therefore, the identification of potential biomarkers is of great importance in predicting the benefit of therapy, especially radiotherapy, enabling precision and individualization of treatment in HNSCC patients.

In this study, we sought to explore a prognostic risk biomarker for predicting therapeutic efficacy and prognosis in HNSCC. We focused on differentially expressed therapeutic response-related genes (TRRGs) between remission and non-remission patients by bioinformatics. We found that the prognostic risk signature and Nomogram based on TRRGs was an effective tool for predicting therapy response and prognosis, which is of great importance for individualized radiotherapy in HNSCC.

Materials and methods

Clinical samples and data collection

Data of The Cancer Genome Atlas (TCGA)-HNSC was from University of North Carolina TCGA genome characterization center and downloaded from website (https://tcga-xena-hub.s3.us-east-1.amazonaws.com/download/TCGA.HNSC.sampleMap%2FHiSeqV2_exon.gz). The GSE65858 Chip dataset was from Wichmann G’ study [11], and downloaded from website (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE65858). The GSE67614 Chip dataset was from Ding L’ study [12], and downloaded from website (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE67614). mRNA expression profiles and clinical information related to HNSCC are publicly available. Therefore, this study does not need the ethical approval of the local ethics committee. The prognosis information of gene protein was downloaded from the Human Protein Atlas (HPA) database (https://www.proteinatlas.org/). Immunotherapy data were obtained from the immunotherapy cohort data (29,443,960) of urothelial carcinoma, which was stored in the “Imvigor210CoreBiologies” R package.

Sample grouping and differentially expressed therapeutic response-related genes (TRRGs) analysis

Patients in TCGA-HNSC were divided into two groups based on their response to therapy. Complete response (CR) or partial response (PR) was set as the remission group. Stable disease (SD) or progressive disease (PD) was set as the non-remission group. Differentially expressed TRRGs between those two groups were analyzed using the “limma” R package. p value < 0. 05 and |Fold Change|> 1.5 were chosen as the cut-off values for differential expressed TRRGs analysis. The survival difference was analyzed by Kaplan‒Meier analysis. Multiple testing was not considered since the ranking of the TRRGs in the log-rank test was our focus.

Functional enrichment analysis

To explore the potential biological functions of differentially expressed TRRGs, Gene Ontology (GO) functions and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were performed by Metascape (https://metascape.org). The formal permission of using KEGG and GO was obtained from Kanehisa laboratories[13].

The mutation types and immune infiltration between the two groups

The mutation type maps of remission and non-remission groups were drawn using the “ComplexHeatmap” R package. The proportion of immune cells was calculated by CIBERSORT software (https://cibersort.Stanford.Edu/), and the infiltration proportion of 22 immune cells in the tumor expression profile was downloaded from PRECOG (http://precog.Stanford.Edu/) and compared by the Wilcoxon rank sum test.

Construction of the TRRGs signature

First, univariate Cox regression analysis was used to screen 206 TRRGs that were significantly related to overall survival (OS). Log rank p < 0.05 was set as the threshold. Then, 20 TRRGs were obtained using Least absolute shrinkage and selection operator (LASSO) regression analysis, which was useful to establish the most feasible and predictable TRRGs signature. Finally, a prognostic risk model was established based on the 20 TRRGs. The risk score of each patient relied on the therapy response and the regression coefficient obtained from the LASSO regression analysis. The risk score model was constructed by the following formula:

$${{\varvec{R}}{\varvec{i}}{\varvec{s}}{\varvec{k}}{\varvec{S}}{\varvec{c}}{\varvec{o}}{\varvec{r}}{\varvec{e}}}_{{\varvec{i}}}=\sum_{{\varvec{j}}=1}^{{\varvec{n}}}{{\varvec{e}}{\varvec{x}}{\varvec{p}}}_{{\varvec{j}}{\varvec{i}}}\times {{\varvec{\theta}}}_{{\varvec{j}}}$$

where i is the number of samples, j is the name of TRRGs, exp is the expression of the corresponding genes, and θ is the regression coefficient of TRRGs in the multivariate Cox regression analysis. Wu and Ren’ articles and their methods to accommodate strong correlations among features are also considered[14, 15].

Verification and evaluation the prognostic value of the TRRGs signature

To verify the characteristics of the TRRGs signature, we divided the patients into a high-risk group (Risk-H) and a low-risk group (Risk-L) by taking the average risk score value as the cut-off, which was further confirmed by Kaplan‒Meier analysis. A time-dependent receiver operating characteristic (ROC) curve was conducted to estimate the sensitivity and specificity of each risk group based on the area under the curve (AUC) value. We further used the GSE65858 dataset as a validation cohort to evaluate the prognostic value of the TRRGs signature. Univariate and multivariate Cox regression analyses were performed to assess the prognostic capability of the TRRGs-based risk signature for HNSCC patients.

Gene set enrichment analysis (GSEA) and construction of the nomogram

The expression and prognosis of HNSC-related model genes were queried from the HPA database (https://www.proteinatlas.org/). The differences in tumor mutation burden (TMB), immune checkpoint expression, stromal score, immune score, tumor purity, and immune cell proportion between the Risk-H and Risk-L groups were analyzed by the Wilcoxon test. The signaling pathways involved in TRRGs were determined by gene set enrichment analysis (GSEA). A nomogram based on the results of Cox regression analysis was established to predict the survival rate.

Statistical analysis

All the above statistical analysis was computed by R software. p-value < 0.05 (two-sided) was used as the statistically significant threshold. The survival difference between groups was analyzed by Kaplan‒Meier analysis. For the significance analysis between different characteristics, the Wilcoxon test was used to compare the differences between the two groups of samples, and the Kruskal test was used to compare the differences between multiple groups. Other statistical methods and algorithms used in this article are described in the corresponding steps.

Results

Identification of differentially expressed therapeutic response-related genes (TRRGs)

The workflow of the study is shown in Fig. 1. For the TCGA-HNSC training set dataset, 445 patients who received therapy and had survival information were enrolled. Patients were divided into a remission group (CR/PR) and a non-remission group (SD/PD) according to therapy response, which was further confirmed by Kaplan‒Meier analysis (Fig. 2A). The remission group and non-remission group contained 394 and 51 patients, respectively. A total of 1896 differentially expressed TRRGs were obtained (Supplemental Table 1), of which 1530 were upregulated and 366 were downregulated (Fig. 2B). The top 200 differentially expressed TRRGs are ranked by log fold-change (FC) values (Fig. 2C), and the distinct gene expression patterns in remission (Fig. 2D) and non-remission groups (Fig. 2E) are also displayed. TP53 and TTN are genes with the highest mutation rates in both groups. The mutation rates of PCLO, RYR2, and AHNAK were significantly higher in the remission group (Fig. 2F), while the mutation rates of CASP8, DNAH5, and XIRP2 were higher in the non-remission group (Fig. 2G). Of note, differentially expressed TRRGs were enriched in tube morphogenesis, chemotaxis, regulation of cell adhesion, etc., using GO and KEGG analysis (Fig. S1).

Fig. 1
figure 1

Complete workflow of our research. “n” denotes sample size. “p < 0.05” denotes the statistically significant threshold

Full size image
Fig. 2
figure 2

Differentially expressed therapeutic response-related genes (TRRGs) between the remission and non-remission groups in TCGA-HNSC training set data. A Survival difference between remission and non-remission groups by Kaplan‒Meier analysis. B Volcano plot of 1896 differentially expressed TRRGs, of which 1530 were upregulated and 366 were downregulated. C-E Heatmap of the top 200 TRRGs ranked by log fold-change values in remission and non-remission groups. F, G Mutation types and rates in remission and non-remission groups

Full size image

To address the potential underlying mechanisms that may affect therapy response, clinical characteristics and infiltrating immune cells were investigated. We found that HPV status by p16 testing was significantly associated with improved therapy sensitivity (Fig. 3A). Besides, data showed that plasma cells, T regulatory cells and resting mast cells were higher in the remission group, while activated mast cells were lower in the non-remission group (Fig. 3B).

Fig. 3
figure 3

Differences in infiltrating immune cells and clinical characteristics between the remission and non-remission groups. A Differences in infiltrating immune cells. B Differences in clinical characteristics. ns, not significant; *, p < 0.05; **, p < 0.01

Full size image

Construction of TRRGs signature with LASSO regression analysis

To determine the independent prognostic TRRGs, 206 genes associated with OS were chosen and identified using univariate Cox regression (Supplemental Table 2). The top 10 prognostic TRRGs and their prognostic values involved in OS are shown (Fig. 4A). Of the top 10 TRRGs, MK2N1 and AREG were significantly associated with worse outcomes, while the MASP1, CD79A, IGJ, CD19, MS4A1, ZNF541, KIAA0125, and CELSR3 genes showed the opposite trend (Fig. 4B-K). We found that AREG, E1CAM, and GPRC5D had prognostic efficacy in the Human Protein Atlas (HPA) (Fig. S2 A-C), while their mRNA expression was not significantly different between cancer and normal tissues in the TCGA-HNSC database (Fig. S2 D-F).

Fig. 4
figure 4

Survival analysis of the top 10 significantly prognostic TRRGs. A The random forest of the top 10 TRRGs by univariate Cox analysis. B-K Kaplan‒Meier analysis of the top 10 TRRGs

Full size image

To explore robust TRRGs and potentially prognostic models, TRRGs were further verified by LASSO regression (Fig. 5A). The data showed that with the increase in log lambda, the number of independent coefficients tended to 0 to lower the survival probability (Fig. 5B). The regression coefficient and Cox analysis of the 20 TRRGs were analyzed (Fig. 5C, D). Heatmap further displayed 20 candidate TRRGs (Fig. 5E), and distinct TRRGs patterns in the remission group (Fig. 5F) and non-remission groups (Fig. 5G).

Fig. 5
figure 5

Using LASSO regression analysis to construct the most predictable TRRGs signature. A LASSO coefficients distribution of 206 significantly prognostic TRRGs. B Tuning parameter (λ) selection cross‐validation error curve. Vertical lines were drawn at the optimal values. C Regression coefficient corresponding to the 20 screened TRRGs are shown. A larger absolute value of the coefficient represents a higher correlation. D Random Forest of 20 TRRGs. E Heatmap of the 20 TRRGs between remission and non-remission groups. F-G Different gene expression between remission and non-remission groups

Full size image

Validation of the prognostic value of signature

For guiding a more accurate treatment strategy, a risk prognostic model was constructed based on the 20 candidate TRRGs, then risk score of each sample was calculated (Supplemental Tables 3 and 4). Patients with HNSCC were divided into a high-risk (Risk-H) group and a low-risk (Risk-L) group according to the average risk score of all samples. Next, the TCGA-HNSC cohort (training set) and GSE65858 cohort (validation set) were used to demonstrate the value of the model. Kaplan‒Meier analysis showed that Risk-H was associated with worse OS than Risk-L both in the TCGA-HNSC cohort (p < 0.0001) and in the GEO cohort (Fig. 6A, D). The risk score distribution analyses and survival status also illustrated a higher risk of death in the Risk-H group (Fig. 6B, E). The AUCs for 1-, 3-, and 5-years were 0.753, 0.798, and 0.749 in the TCGA-HNSC cohort and 0.570, 0.568, and 0.716 in the GEO cohort, respectively (Fig. 6C, F).

Fig. 6
figure 6

Validation of the prognostic risk model in the TCGA-HNSC cohort (training set) and GSE65858 cohort (validation set). A, D Survival curves for HNSCC patients with high- and the low-risk score. B, E Survival status of patients with high- and the low-risk score. C, F Receiver operating characteristic (ROC) curves for 1-, 3- and 5-year overall survival in HNSCC patients. Up panels are TCGA-HNSC cohort and lower panels are GSE65858 cohort

Full size image

As radiotherapy is a crucial therapy in the management of HNSCC. To further evaluate the value of our prognostic model in HNSCC patients treated with radiotherapy, 275 patients from TCGA-HNSC were included. Data showed that Risk-L patients had better OS than Risk-H patients (Fig. 7A). What’s more, for patients treated with post-operative radiotherapy, Risk-L patients had a lower recurrence than Risk-H patients in the GSE67614 database from GEO (Fig. 7B, Supplemental Table 5). Above data indicating that risk score of TRRGs-based signature was a promising predictive factor for radiotherapy efficacy for HNSCC patients.

Fig. 7
figure 7

Evaluation of the prognostic value in HNSCC patients who received radiotherapy. A Kaplan‒Meier curves based on the TCGA-HNSC dataset. B Risk score for recurrent and non-recurrent groups in the GSE65858 cohort

Full size image

Taken together, the risk prognostic signature based on TRRGs was a potential indicator for prognosis and radiation sensitivity in HNSCC.

Correlations between risk score and clinical characteristics

To study the relationships between risk score and clinical features, the Wilcoxon test and Kruskal test were used. We found that the risk score was associated with HPV status, primary therapy outcome, clinical M, clinical T, neoplasm histologic grade, targeted molecular therapy, and age. No significance was found among the risk score and gender, clinical N, clinical TNM stage, radiation therapy and smoking (Fig. 8).

Fig. 8
figure 8

Correlation analysis between risk score and clinical characteristics. The Wilcoxon test was used for double terms and the Kruskal test for multiple terms

Full size image

Relationships between risk score and immune status and signaling pathways

As we described above, the distribution of infiltrating immune cells is significantly different in radiosensitive and resistant HNSCC. To further address the underlying mechanisms between risk score and immune status, tumor mutational burden (TMB), immune checkpoint blockade (ICB)-associated genes, immune score, matrix score, tumor purity and proportion of immune cells were analyzed. We found that there was no significant correlation between risk score and tumor mutational burden (Fig. 9A), while it was significantly associated with numerous immune checkpoint genes. Risk score was positively associated with PVR, CD276, CD274, and was negatively associated with PDCD1, CTLA4, LAG3, etc. (Fig. 9B-O). However, the risk score and nonsilent mutation rates, LAG3, HAVCR2, CD80, CD86, and LGALS3 were not significantly correlated (Fig. S3).

Fig. 9
figure 9

Associations of risk score with tumor mutational burden and immune checkpoint genes. A Risk score did not correlate strongly with tumor mutational burden. B-O Risk score was significantly associated with multiple immune checkpoint genes

Full size image

No significant difference was found among the immune score, matrix score, tumor purity and risk score. However, M0 macrophages, M2 macrophages, activated mast cells, and resting natural killer (NK) cells were more abundant in Risk-H patients, while naive B cells, plasma cells, CD8 T cells, T follicular helper (Tfh) cells, regulatory T cells (Tregs) and resting mast cells were enriched in Risk-L patients (Fig. 10A). These findings suggest that the risk score may be a potential biomarker for immunotherapy response.

Fig. 10
figure 10

Immune and molecular characteristics of different risk score subgroups. A Immune cells between the high-risk score (Risk-H) and low-risk score (Risk-L) groups. B Gene sets enriched in the Risk-H group. C Gene sets enriched in the Risk-L group. ns, not significant, *, p < 0.05, **, p < 0.01, ***, p < 0.001, ****, p < 0.0001)

Full size image

Gene set enrichment analysis (GSEA) was conducted to determine the gene sets and pathways in different risk scores. Data showed that Gene sets of Risk-H were enriched in the calcium and MAPK signaling pathway, cytokine‒cytokine receptor interaction and proteoglycans in cancer. (Fig. 10B, C). Gene sets of Risk-L were enriched in drug and retinol metabolism, metabolism of xenobiotics by cytochrome P45, etc.

Predictive capabilities of risk score in immunotherapy

As the risk score is somewhat related to immune status, we investigated whether it could predict patients’ response to immunotherapy. The IMvigor210 cohort is an open-label, multicenter, single-arm phase II clinical study that evaluates the safety and efficacy of atezolizumab in patients with metastatic urothelial carcinoma. We found that nonresponse (SD/PD) to immunotherapy was positively correlated with Risk-H (r = 0.6, p < 0.001), while the response was closely correlated with Risk-L (r = 0.5, p < 0. 001) (Fig. 11A, B). However, the risk score was not significantly different between the response and nonresponse groups (Fig. 11C). Of note, Risk-H had better OS than Risk-L (p = 0.045, Fig. 11D). The AUCs of the risk scores for 1 year and 3 years were 0.586 and 0.575, respectively (Fig. 11E).

Fig. 11
figure 11

Verified the predictive efficacy of the risk score based on the data from the IMvigor210 cohort. A Correlation map of response (complete response / partial response) and nonresponse (stable disease / progressive disease) to immunotherapy in different risk score groups. B Relative percent of response and nonresponse to immunotherapy in the Risk-H and Risk-L groups. C Comparison of the difference between response and nonresponse groups. D Kaplan–Meier curves of overall survival time of the high- and low-risk score groups in the metastatic urothelial carcinoma (mUC) sample. E The survival ROC curves of risk score in the mUC sample

Full size image

Establishment of the nomogram model

Univariate and multivariate Cox analysis showed that risk score was significantly related to survival (Fig. 12A, B), indicating that the risk score model based on TRRGs was an independent prognostic biomarker in the clinical practice of HNSCC.

Fig. 12
figure 12

Establishment of the nomogram model. A, B Univariate and multivariate Cox analyses of clinical features and risk score. C Nomogram model of overall survival based on Cox regression analysis. D AUC of the risk score, targeted molecular therapy and age. E Calibration curves for predicting overall survival at 1 and 3 years

Full size image

To better predict the prognosis of HNSCC, a nomogram was established using five independent prognostic variables based on Cox regression analysis (Fig. 12C). The nomogram showed that the risk score had the highest AUC, followed by age and targeted molecular therapy (Fig. 12D). The calibration curve was then drawn to evaluate the prediction accuracy of the nomogram. The calibration curves of 1 and 3 total OS rates were close to the standard line, suggesting that the nomogram has good performance in predicting survival probability (Fig. 12E).

Discussion

With the development of the new TNM staging system, transoral robotic surgery, intensity-modulated radiotherapy and immunotherapy, the survival of HNSCC has been greatly improved [16]. Radiotherapy is an indispensable treatment, especially for early-stage and locally advanced disease[17, 18]. However, the prognosis remains poor, as a substantial portion of HNSCC is resistant to radiotherapy[19, 20]. In this study, we conducted a comprehensive bioinformatics analysis to explore a potential biomarker for predicting therapy response and prognosis based on differentially expressed TRRGs. Our data suggested that TRRGs were significantly associated with therapy response, especially radiotherapy. More importantly, the risk score and nomogram model based on TRRGs were promising biomarkers for therapy response and prognosis in HNSCC. Additionally, risk score may be a prognostic biomarker for immunotherapy.

There are various factors that affect the sensitivity of radiotherapy in HNSCC [21,22,23,24,25,26,27,28,29], including genes, cytokeratin, miRNA, lncRNA, tumor molecular subtypes and anticancer drugs. Increasing prognostic models, based on gene-based signatures, risk score or mutations of key genetic abnormalities, have been explored to predict the response to radiotherapy [30,31,32]. Fortunately, we observed key differentially expressed TRRGs in HNSCC that affected patients’ response to radiotherapy. We found that CELSR3, AREG and MASP1 were among the top 10 differentially expressed TRRGs and were chosen to construct the risk score model. Of note, high expression of CELSR3 and MASP1 was significantly associated with superior OS, while high expression of AREG was just the opposite in HNSCC. AREG is a predictive biomarker in the treatment of colorectal cancer [33] and promotes the progression of cancers via the AREG/EGFR pathway in pancreatic cancer, breast cancer and ovarian cancer [34,35,36], suggesting that AREG may be an attractive target in HNSCC. However, CELSR3 has been reported to be highly expressed in hepatocellular carcinoma, prostate cancer, and lung adenocarcinoma and indicates a poor prognosis [37,38,39,40]. MASP1 is related to immune cell infiltration in head and neck cancer [41, 42] and could be a candidate target gene in lung cancer and cervical cancer [43, 44]. However, the functions of CELSR3, MASP1 and AREG remains unclear in HNSCC, and the mechanisms how these TRRGs affect the radiotherapy sensitivity and prognosis of HNSCC need to be explored in the future.

Currently, emerging data have demonstrated that HPV + oropharyngeal cancer (OPC) is more sensitive to radiotherapy and has prolonged OS and a reduced risk of death compared to HPV- OPC [45, 46]. HPV status tested by p16 was a reliable and surrogate marker for HPV infection in HPSCC [47]. Consistent with the above studies, our study found that the HPV-positive rate was remarkably higher in the remission group than in the non-remission group, indicating that HPV status was a biomarker for radiotherapy sensitivity. Additionally, the risk score was closely associated with clinical characteristics. Univariable and multivariable analyses suggested that HPV status and risk score were independent prognostic factors in HNSCC. Therefore, the HPV status test is strongly recommended before treatment, and the risk score is also a valuable biomarker, which remains to be confirmed in clinical practice.

To further address the mechanisms that affect therapy response, patients’ immune cells, TMB, ICB-associated genes, etc., and their correlations with risk score were explored. The subtypes of immune cells in the tumor microenvironment varies between the Risk-H and Risk-L groups. We found that CD8 T cells, Tfh cells, and naïve B cells were more abundant in the Risk-L group, indicating an immune hot microenvironment. M2 macrophages were more common in the Risk-H group. M2 macrophages have been proven to be related to chronic inflammation and promote tumor growth and metastases [48]. TMB has served as a valuable biomarker for predicting immune checkpoint inhibitor (ICI) efficacy [49]. In this study, we found that the risk score was not significantly associated with TMB, while it was significantly associated with multiple ICB-associated genes, including CD 274 and CTLA4 LAG3, suggesting that the tumor immune microenvironment (TIME) and ICB-associated genes were closely associated with radiotherapy sensitivity. Of note, the risk score may be a potential tool to predict the immune response and prognosis of patients with metastatic urothelial carcinoma by analyzing the IMvigor210 cohort. Due to the heterogeneity and complexity of the TIME, only a small number of HNSCC patients benefit from immunotherapy. Thus, more studies need to be performed to explain why and how the TIME affects radiotherapy sensitivity. In addition, the relationships between the risk score and immune response remain to be seen. Taken together, our study provides a new idea for exploring the interactions between radiotherapy, immunotherapy, and tumor immune microenvironments in HNSCC.

Last, we found that the nomogram established using the risk score and other clinical features, including age, HPV status, radiation therapy and targeted molecular therapy, had good prognostic efficacy in predicting survival in HNSCC. Nomograms that predict patients’ response to therapy and prognosis in cancers are increasing [50, 51]. Of note, nomograms need to be confirmed in multicenter clinical trials.

Although we identified candidate TRRGs in a large sample through bioinformatics technology, some of the limitations of this study are noteworthy. First, the primary source of clinical features in our dataset is the TCGA-HNSC database. Most of the patients were from North America, so we should be very careful to extend our findings to other geographic and ethnic groups. Second, few clinical characteristics, like p16 status and upfront surgery, were not complete, which may affect the analysis to some extent. Third, our study found 20 TRRGs that are significantly related to the radiotherapy sensitivity and prognosis of patients with HNSCC, while the biological mechanisms are not well understood. Therefore, further studies need to be conducted to confirm our findings.

Conclusion

In summary, TRRGs play an important role in HNSCC. A risk prognostic signature and Nomogram based on TRRGs is a promising biomarker for predicting therapy response in HNSCC. The underlying regulatory mechanisms of TRRGs may be related to immune cells, ICB-associated genes and the TIME.