A novel hypoxia-associated gene signature for prognosis prediction in head and neck squamous cell carcinoma

The raw RNA sequence (RNA-seq) data and corresponding clinical parameters of 502 HNSCC patients and 44 adjacent normal tissues were obtained from the TCGA database on 23 May 2022. A total of 499 HNSCC patients with complete clinical information and survival data were included for further analysis. For external validation purposes, RNA-seq data and clinical parameters of HNSCC patients were obtained from HNSCC-related mRNA datasets (GSE27020, GSE41613, GSE42743, and GSE117973) in the GEO database. The clinical baseline data of all included patients are shown in Additional file 1: Table S1 (Patient baseline information table) and Additional file 2 (Details of patients in the TCGA database).

A total of 200 HAGs were downloaded from the Molecular Signatures Database (https://​www.​gsea-msigdb.​org/​gsea/​msigdb/​cards/​HALLMARK_​HYPOXIA.​html). The same approach was used to obtain HAGs in the previously published study [14]. The "limma" package of R software was performed to distinguish the differentially expressed HAGs between HNSCC and adjacent normal tissues [15]. The FDR was adjusted by the Benjamini–Hochberg method. FDR < 0.05 and |logFC|> 1.0 was set as the cut-off criteria of differently expressed HAGs. To explore the biological functions of the hypoxia-associated gene signature, GO and KEGG enrichment analyses were performed by the R-package "Clusterprofiler" (version 4.0) [16, 17].

To establish a prognostic model, the R package "glmnet" was used to perform the univariate Cox regression and LASSO analyses on the differentially expressed genes screened previously. The penalty factor λ was identified by the minimum parameters. Next, a risk prognostic model was developed by multi-factor Cox regression, and the following formula was used to calculate the risk score: Risk Score = Gene1 CoefixExpi + Gene2 CoefixExpi + …GeneN CoefixExpi (Coef: coefficients, Exp: gene expression levels). 499 HNSCC patients from the TCGA database were classified into low and high risk subgroups based on the median risk score. Subsequently, the overall survival curves of different subgroups were compared by Kaplan–Meier analysis, the overall survival at 1, 3, and 5 years was described by time-dependent receiver operating characteristic (ROC) analysis, and the area under the curve (AUC) was used to access the model’s predictive power. Finally, this prognostic gene signature was also demonstrated to have prognostic value in predicting OS in HNSCC patients using the datasets GSE27020, GSE41613, GSE42743, and GSE117973.

The TCGA-HNSCC samples were randomized into two groups to clarify the association between the prognostic model and various clinical characteristics, such as staging, grade, age, gender, T stage, N stage, and M stage. There were two subgroups of patients: stage I/II and III/IV subgroups, grade I/II and III/IV subgroups, age < 60 and age ≥ 60 subgroups, male and female subgroups, T0-T2 and T3/4 subgroups, N0 and N + subgroups, and M0 and M1 + Mx subgroups, respectively. To confirm the 8-gene prognostic features’ independent prognostic value, Kaplan–Meier survival analysis was performed on specific subgroups with various clinical characteristics.

Immune differences between the two groups of the TCGA database were synthesized by using computational methods for assessing immune infiltration and function, including ESTIMATE [18], TIMER [19], MCP-counter [20], CIBERSORTx [21], and single-sample gene set enrichment analysis (ssGSEA). A two-sample Wilcoxon test was applied to compare immune infiltration and immune-related functions between the high and low risk groups.

Based on the TCGA HNSCC cohort, a nomogram containing characteristic risk scores and other clinical characteristics was developed to better predict HNSCC prognosis. Using univariate Cox regression analysis, clinical characteristics with significant associations with HNSCC prognosis were screened. Then, the variables screened in the previous step were included in a multivariate Cox regression analysis to search for independent predictive variables for OS in HNSCC patients and to construct the nomogram by statistically significant variables. Finally, the predictive performance of the established nomogram, risk score, and clinical characteristics was compared by using the decision curve analysis (DCA), calibration curves, ROC curves, and consistency index (C-index).

To better understand the biological processes of hypoxia-associated genes, the most relevant genes (Pearson |R|> 0.5, P < 0.05) were identified in the TCGA database, and functional enrichment analysis was performed using the R-package "Clusterprofiler". Subsequently, gene set variation analysis (GSVA) enrichment analysis [22] and Gene set enrichment analysis (GSEA) [23] were used to screen the HNSCC cohort from TCGA for signaling pathways significantly associated with risk groupings and to screen out significant pathways according to a false discovery rate (FDR) < 0.05. Benjamini–Hochberg method was used to correct the FDR.

The Zhongnan Hospital of Wuhan University provided nine pairs of HNSCC and adjacent non-cancerous tissues for this study, all of which were authorized by the ethics committee. Surgically resected patients with HNSCC who did not receive chemotherapy or radiotherapy were recruited for the study. Additional file 1: Table S2 presents the baseline information of the included clinical samples. Total RNA was extracted from tissues using TriQuick Reagent (Solarbio, Beijing, China, R1100); reverse transcription was carried out using a Prime Script RT kit (TaKaRa, Dalian, China, RR037A); and quantitative PCR was performed using standard protocols from the SYBR Green PCR kit (Toyobo, Osaka, Japan, QPK-201). The primer sequences for PCR are shown in Additional file 1: Table S3. Candidate genes’ relative mRNA levels were normalized to GAPDH mRNA expression, while differences were compared using paired t-tests. The changes were calculated using the 2^−∆∆Ct method. Data were presented as the mean values ± standard error of the mean (SEM) from at least three independent experiments. To further verify the differences in eight core genes between high and low risk groups for HNSCC. Tumor samples from 16 HNSCC patients were relatively quantified using qRT-PCR. The relative expression levels of 8 genes were input into the risk model and the risk score was calculated. The patients were divided into high and low risk groups according to the median risk score. qRT-PCR results were used to further compare the gene expression differences between high and low risk groups.

First, differential expression analysis of 200 HAGs was performed to identify differentially expressed HAGs between HNSCC (n = 502) and normal samples (n = 44). The results showed that the RNA expression of 54 HAGs was significantly different, which are presented as heatmap and volcano plot (adjusted p < 0.05, Fig. 1A-B). Then, GO enrichment analysis was performed on these 54 HAGs, which identified the main pathways involved in these 54 genes: response to hypoxia/response to oxygen levels/cell chemotaxis/glucose metabolic process (Fig. 1C). Finally, the KEGG enrichment analysis was also performed on these 54 HAGs, and the findings revealed that they were involved in the following signaling pathways: MAPK signaling pathway, HIF-1 signaling pathway, insulin signaling pathway, and focal adhesion (Fig. 1D).

A univariate Cox regression approach was applied to screen HAGs significantly associated with OS in HNSCC. The analysis showed that 49 of 200 HAGs were highly related to OS in HNSCC, and 16 of these 49 genes were significantly different between HNSCC tissues and adjacent paracancerous tissues (Fig. 2A). To further narrow down the gene selection, LASSO analysis was performed on the 16 genes screened in the previous step and selected 12 genes [ Serpin family E member 1 (SERPINE1), Homeobox B9 (HOXB9), Selenium-binding protein 1 (SELENBP1), C-X-C motif chemokine receptor 4 (CXCR4), Dystrobrevin Alpha (DTNA), Interferon-stimulated exonuclease gene 20 (ISG20), Stanniocalcin 2 (STC2), Heparan Sulfate-Glucosamine 3-Sulfotransferase 1 (HS3ST1), Adrenomedullin (ADM), Stanniocalcin 1 (STC1), Cysteine and glycine-rich protein 2 (CSRP2), and Transforming growth factor beta-induced protein (TGFBI)] as candidate genes (Fig. 2B-C). Next, multifactorial Cox regression analysis was performed on the 12 candidate genes. Finally, eight genes (HOXB9, SELENBP1, DTNA, ISG20, STC2, HS3ST1, CSRP2, and TGFBI) were screened for the construction of prognostic risk models; among them, SELENBP1, CSRP2, and ISG20 were considered as biomarkers indicative of good prognosis (HR < 1), while TGFBI, STC2, HOXB9, DTNA, and HS3ST1 were indicative of poor prognosis (HR > 1). A risk model of HNSCC based on eight prognostic genes was generated (risk score = (0.199 × HOXB9 exp.) + (-0.238 × SELENBP1 exp.) + (0.359 × DTNA exp.) + (-0.134 × ISG20 exp.) + (0.193 × STC2 exp.) + (0.418 × HS3ST1 exp.) + (-0.159 × CSRP2 exp.) + (0.094 × TGFBI exp.)). Based on this risk model, the Kaplan–Meier method was applied to further investigate patients’ survival in HNSCC. 499 patients with HNSCC were distributed equally into two subgroups based on median risk scores, compared with the high-risk subgroup, those in the low-risk subgroup lived longer (Fig. 2D-E). Then, the model’s capability of predicting future HNSCC risk was evaluated using a time-dependent ROC analysis. The risk model predicted 1-year, 3-year, and 5-year survival with AUC values of 0.660, 0.714, and 0.672, respectively (Fig. 2F).

The 8-gene prognostic risk model was validated in four GEO datasets (GSE27020, GSE41613, GSE42743, and GSE117973). According to the same risk score construction method in TCGA database, the four GEO datasets were classified into high and low risk groups based on the median risk scores and evaluated the performance of the risk model. The results of the analysis are shown in Figs. 3 and Additional file 3: Figure S1. The AUC for 1-year, 3-year, and 5-year OS was calculated to present the accuracy of the model prediction. In GSE27020 dataset, the AUC was 0.739, 0.701, and 0.687 respectively (Fig. 3A); In GSE41613 dataset, the AUC was 0.674, 0.733, and 0.680 respectively (Fig. 3B); In GSE42743 dataset, the AUC was 0.763, 0.717, and 0.789 respectively (Additional file 3: Figure S1A); In GSE117973 dataset, the AUC was 0.663, 0.748, and 0.736 respectively (Additional file 3: Figure S1B). Meanwhile, in all four GEO datasets (GSE27020, GSE41613, GSE42743, and GSE117973), patients in the low-risk group survived longer than those in the high-risk group (Fig. 3C-F, Additional file 3: Figure S1C-F). Taken together, the AUC of the risk score model for predicting the prognosis of patients in 3 years reached more than 0.7, indicating that the risk score model had a good value in predicting the prognosis of patients in three years.

Additionally, to explore the model’s prognostic value in HNSCC patients after stratification based on clinicopathological variables in the TCGA database, patients were classified into subgroups according to their Age, Gender, Grade, M stage, N stage, Radiation therapy, stage, and T stage to plot Kaplan–Meier survival curves. The sample was divided into 16 subgroups: Young (< 60 years) and Old (≥ 60 years), Female and Male, Grade I-II and Grade III-IV, Mstage M0 and Mstage M1 + Mx, Nstage N0 and Nstage N + , Radiation therapy NO and Radiation therapy YES, Stage I-II and Stage III-IV, Tstage T0-T2, and Tstage T3-T4. In each subgroup, the previous thresholds were selected and the patients were further allocated into the low and high groups. OS was significantly shorter in the high-risk group compared to the low-risk group across all subgroups (Additional file 3: Figure S2A-H, Additional file 3: Figure S3A-H). The results showed that the risk model could accurately predict the prognosis of HNSCC patients within the same clinicopathological subgroup. In conclusion, risk characteristics are crucial in determining a patient's prognosis for HNSCC.

To investigate the correlations between the risk scores model and immune status, immune cell infiltration and immune-related function scores were quantified by various algorithms (ESTIMATE, CIBERSORTx, TIMER, MCP counter, and ssGSEA). Figure 4 A demonstrates that the low-risk group's Immune Scores were significantly higher than those of the high-risk group, while the ESTIMATE Score and Stromal Score did not differ significantly, indicating a higher level of immunity in the low-risk group (p < 0.01). The variation of tumor microenvironment cells may be the cause of risk score heterogeneity. Then, based on TCGA-HNSCC data, analyses of the difference in immune cell subpopulations revealed significant differences between the two subgroups for scores of immune cells (B cells, T cells, Myeloid dendritic cells, plasma cells, Macrophage M0, T cell CD4 memory resting, T cell follicular helper, Tregs, NK cells resting, dendritic cells resting, Mast cells, iDCs, pDCs, Th2 cells, and TIL) (p < 0.001; Fig. 4B-E). Furthermore, immune function scores (Checkpoint, Cytolytic activity, HLA, Inflammation promoting, T cell co-inhibition, T cell co-stimulation, and Type I IFN response) were significantly different (p < 0.05; Fig. 4F). These results indicate that the types of immune infiltrating cells are different in the high and low risk groups, and the difference in immune microenvironment may be an important factor affecting the difference in the prognosis of patients.

In univariate Cox regression analyses, risk score (HR = 1.822, 95% CI = 1.549–2.142, p-value < 0.001), Age (HR = 1.378, 95% CI = 1.053–1.804, p-value = 0.0193), Stage (HR = 1.421, 95% CI = 1.187–1.702, p-value < 0.001), Tstage (HR = 1.296, 95% CI = 1.124–1.495, p-value < 0.001), and Nstage (HR = 1.541, 95% CI = 1.303–1.822, p-value < 0.001) were correlated to OS in patients with HNSCC (Fig. 5A). In multivariate Cox regression analysis, characteristic risk score (HR = 1.830, 95% CI = 1.479–2.265, p-value < 0.001), Nstage (HR = 1.480, 95% CI = 1.186–1.847, p-value < 0.001) and Age (HR = 1.449, 95% CI = 1.041–2.016, p-value = 0.0279) were shown to be independent predictors of OS in HNSCC patients (Fig. 5A). Then, the risk score, Age, and N stage were combined to construct a nomogram for predicting OS of 1-year, 3-year, and 5-year in HNSCC patients (Fig. 5B). The performance of the nomogram was evaluated, and it showed superior predictive potential compared to the risk score and other clinical characteristics (Age, Gender, Stage, T stage, Nstage, and Grade) for OS at 1, 3, and 5 years (AUC = 0.704, 0.758 and 0.733 respectively, Fig. 5C). The calibration curves for the nomogram demonstrated good agreement between the actual OS probabilities and the OS probabilities predicted by the nomogram (Fig. 5D). The nomogram had good discrimination compared to the risk score, Age, and Nstage, with a C-index close to 0.7 (Fig. 5E). The DCA curves showed that the nomogram had more net benefit than the risk score, Age, and Nstage (Fig. 5F).

To clarify the signaling pathways related to the hypoxia 8-gene signature, we screened the genes with the highest correlation coefficients with risk scores (Pearson |R|> 0.5, P < 0.05) in the TCGA database and applied functional enrichment analysis to them. A total of 50 negatively and 51 positively correlated genes were screened by correlation analysis, and the correlation heatmap was presented in Fig. 6A. Next, GO enrichment analysis was performed on these 101 genes, which revealed that these genes were mainly involved in signaling pathways including divalent inorganic cation homeostasis, response to insulin stimulus, fibroblast proliferation, positive regulation of hemostasis and positive regulation of coagulation (Fig. 6B). Subsequently, KEGG enrichment analysis showed that they were involved in signaling pathways including Axon guidance, Inflammatory mediator regulation of TRP channels, Rap1 signaling pathway, and EGFR tyrosine kinase inhibitor resistance (Fig. 6C). Additionally, the signaling pathways involved in the hypoxia 8-gene signature were investigated using GSVAs and the results showed that in the high-risk subgroup, the top 5 signaling pathways significantly activated included Glycolysis, Adipogenesis, MYC Targets V1, MTORC1 Signaling, and PI3K-AKT-mTOR Signaling (Fig. 6D). Finally, the signaling pathways involved in the hypoxia 8-gene signature were screened by GSEA enrichment analysis, which showed that the significantly enriched signaling pathways in the high-risk subgroup included: Dilated cardiomyopathy, ECM Receptor interaction, Focal adhesion, Hypertrophic cardiomyopathy and pathways in cancer (Fig. 6E). Significantly enriched signaling pathways in the low-risk subgroup included Alpha-linolenic acid metabolism, Autoimmune thyroid disease, Oxidative phosphorylation, Parkinson disease, and Ribosome (Fig. 6F).

To determine the expression of 8 hypoxia-related genes in HNSCC tumor tissues, we collected 9 pairs of tissues from oral squamous cell carcinoma patients for qRT-PCR verification. There was no difference in HOXB9, SELENBP1, ISG20, HS3ST1 and CSRP2 expression between tumor and normal tissues (Additional file 3: Figure S4A-E). Three genes (DTNA, STC2, and TGFBI) were differentially expressed between HNSCC tumor tissue and normal tissue, and STC2 and TGFBI were significantly up-regulated in tumor tissue, which was consistent with the analysis results (Fig. 7A-C). Furthermore, the risk model was used to divide 16 HNSCC patients into high and low risk groups. qPCR detection found that STC2, TGFBI and HOXB9 were significantly up-regulated between the high and low risk group (Fig. 7D-F), while SELENBP1, DTNA, ISG20, HS3ST1 and CSRP2 had no significant differences (Additional file 3: Figure S4F-J).