Development and validation of a fourteen- innate immunity-related gene pairs signature for predicting prognosis head and neck squamous cell carcinoma

In April 30, 2019, RNA-seq data and the latest clinical follow-up information were downloaded from TCGA using GDC API, including 612 RNA-seq data samples. Similarly, a set of chip data set GSE65858 in MINiML format were downloaded from NCBI, including the expression profile data and clinical follow-up information of 270 HNSCC sample. All patients underwent surgery with a negative surgical margin, receive no adjuvant or neoadjuvant therapy. A total of 1039 immune-related genes (removing the name-repeated gene) were downloaded from the InnateDB database (https://​www.​innatedb.​com/​).

For the TCGA RNAseq data, we screened 517 tumor samples with follow-up information and OS > 0, extracted the expression profile of the immune-related gene set and removed the gene with 0 expression level in 50% of the samples. For chip data sets, we screened samples with follow-up information and OS > 0, R package GEOquery was used to map the chip probes to GeneSymbol, the probes was mapped to multiple genes were removed, multiple probes were mapped to a single gene to take the median, gene expression profile were obtained, and the expression profile of the immune gene set were extracted. The clinical information of TCGA and GEO patients is shown in Table 1. The workflow is shown in Fig. 1.

For better model building and validation, we randomly divided the TCGA data set into two groups, one as a training set (N = 260), one as an internal validation set (N = 259), and the GSE65858 data set as an independent external validation set. During the random grouping process of TCGA, we kept the two groups of samples similar in age distribution, clinical stage, follow-up time, and proportion of patient deaths, while the number of samples after clustering the gene expression profiles of the two groups was close to each other, and the statistical characteristics of the two samples are shown in Table 2.

A total of 539,241 gene pairs were obtained by randomly permutation and combination of 1039 immune genes. For arbitrary gene i (IRG_i) and gene j (IRG_j), IRGP_ij,were calculated. The IRGP values were defined as follows:

Where IRG indicates the amount of gene expression, we calculated all IRGPs values for all samples and further filtered IRGPs with a standard deviation of 0, a total of 18,182 IRGPs were obtained.

Univariate Cox proportional hazard regression analysis was performed on each IRGPs as in Jin-Cheng et al. [12] to screen for those genes that were significantly associated with OS in the training data set, with p < 0.05 as the threshold.

LASSO is a popular method for regression modeling with a large number of potential prognostic features, because it can perform automatic feature selection in a manner that results in signatures with generally good prognostic performance [13]. The LASSO method has been extended to the Cox model for survival analysis and has been successfully applied for the purpose of building sparse signatures for survival prognosis in many application areas including oncology [14‐16], First, we used training set samples to conduct univariate Cox proportional risk regression analysis for each IRGPs, with log rank p < 0.05 as the threshold, 669 IRGPs with significantly correlated prognoses were identified. Furthermore, R software package glmnet [17] was used to screen robust prognostic immune-related gene pairs, and 3-fold cross validation was used to evaluate the optimal characteristics. The degree of LASSO regression complexity adjustment is controlled by the parameter λ, where the larger λ is, the greater the penalty for a linear model with more variables, so that a model with fewer variables is eventually obtained. In this study, the optimal model is obtained when λ = 0.1218186, and we choose the features incorporated in the model at this time as the optimal combination of features, i.e., 14-IRGPs. Multivariate Cox regression analysis was conducted using the stepwise regression method to determine the coefficient of each IRGPs in the 14-IRGPs, and the following risk score model was constructed: Where N is the number of prognostic IRGPs, Exp_k is the IRGP value of prognostic IRGPs, and e^HR_k is the estimated regression coefficient of IRGPs in the multivariate Cox regression analysis.

To validate the IRGPs signature, patients in test datasets were divided into low risk and high risk group according to the median value of the risk score, which calculated according to the prognostic signature. The log-rank test and Cox regression analysis were conducted to evaluate overall survival difference between the low risk and high risk groups. Receiver operating characteristic curve (ROC) curve was used to assess the categorization of IRGPs signature. The IRGPs signature was also compared with the published signature by KM survival curve, ROC curve, and C-index.

In order to observe the relationship between riskScore and clinical phenotype, the samples were divided into two groups based on the riskScore median of the samples, and the prognosis differences between high riskScore and low riskScore were compared respectively. Similarly, the relationship Grade, Age and Stage in High and Low TMEScore was analyzed.

We used R package clusterProfiler, v3.8 [18] for GO and KEGG enrichment analysis with a p value of less than 0.05 as the threshold. GSEA [19] was performed by R package GSVA using the MSigDB [20]. Gene sets with a false discovery rate (FDR) value less than 0.05 after performing 1000 permutations were considered to be significantly enriched.

For the TCGA training set samples, we used a univariate Cox proportional hazard regression model to establish the relationship between patient overall survival and immune-related gene expression, and obtained 164 prognostic genes. According to the calculation rule of the IRGPs value, a total of 7374 IRGPs are obtained. The univariate Cox proportional hazards regression model was used to establish the relationship between IRGPs and overall patient survival. Finally, we obtained 699 IRGPs with significant prognostic differences (Fig. 2a). In order to screen robust immune-related prognostic gene pairs, we used lasso regression to perform dimensionality reduction analysis on these 699 IRGPs. The results show that as the lambda increases, the number of independent coefficients tends to 0 (Fig. 2b), 3-fold cross-validation was used to build the model, and the model is optimal when lambda = 0.1218186 (Fig. 2c). We select the model when lambda = 0.1218186 as the final model, which contains a total of 14 IRGPs, 19 genes (Table 3). The risk scores of these 14 IRGPs in each sample are shown in Fig. 2d. Furthermore, we calculated the Risk Score of each sample based on the Risk model, and the formula is as follows:

Multivariate regression analysis was used to establish a risk model for 14 IRGPs in the training set, validation set, TCGA dataset and independent test set data (GSE65858 dataset) for 1, 3, and 5 years. The results suggest that the average AUC of the training set is 0.758, the average AUC of the validation set is 0.659, the average AUC of the TCGA dataset is 0.709, and the average AUC of the independent test set data is 0.685 (Fig. 3a-d). With the median risk score as the threshold, the training set samples were divided into risk-H and risk-L, and the KM survival curves of 14-IRGPs in training set, validation set, all data sets of TCGA and one independent GEO testsets (GSE65858 dataset) were drawn. The results showed that the prognosis of the risk-L group of all data sets was significantly better than that of the risk-H group (Fig. 3e-h). In summary, IRGPs have great potential as prognostic markers.

In order to observe the robustness of risk models in different clinical characteristics, we observed the predictive power of risk models in different TNMstages, Age, gender and smoking history. We found that the 14-IRGPs signature model can be significantly distinguished into high-risk group and low-risk group not only in early patients and late-stage patients (Fig. 4a, b) (log rank p = 0.00023, log rank p < 0.0001), but also in young data sets and elderly data sets (Fig. 4c, d) (logrank p < 0.0001, log rank p < 0.0001), and in female data sets and male data sets (Fig. 4e, f) (log rank p = 0.01, log rank p < 0.0001). Finally, our analysis of the samples with and without smoking history shows that 14-IRGPs signature can also significantly distinguish the high-risk group from the low-risk group (Fig. 4g, h) (log rank p = 0.002, log rank p < 0.0001), those results indicated that our model has a very stable predictive power in patients of different ages, stages and genders.

In order to identify the independence of 14-IRGPs signature model in clinical application, we used univariate and multivariate COX regression analysis to analyze relevant HR, 95%CI of HR, p value in TCGA training set, TCGA verification data set and all data of TCGA. We systematically analyzed clinical information recorded by TCGA patients, including age, T, N, AJCC Stage, Grade, Smoking, and our 14-IRGPs signature grouping information (Table 4).

The above conditions indicate that our model 14-IRGPs signature has a good predictive performance in terms of clinical application value in TCGA data set, and our model may be a prognostic indicator independent of other clinical factors and has an independent predictive performance in terms of clinical application value.

In order to further analyze the functions of 14-IRGPs, we first used clusterProfiler to conduct GO and KEGG enrichment analysis on 19 genes, and finally retained the results of p < 0.05. The results showed that these pathways were enriched to 356 GO BP, which were mainly T cell receptor signaling pathway, stress-activated MAPK cascade and other biological processes, and we show the most significant top 20 (Fig. 5a), Furthermore, we found that these 19 genes were significantly enriched in 39 GO CCs and 73 GO MFs, the most prominent of which were the top 20 (Fig. 5b, c).

ssGSEA was used to analyze the enrichment scores of each sample in each pathway in the TCGA data set, calculate the correlation between these pathways and risk scores, and 26 pathways with correlation > 0.25 were selected (Fig. 5d), we found that most of the samples with risk score present negative correlation, a small number of positively related with risk score. Cluster analysis was conducted according to the 26 KEGG pathway enrichment scores (Fig. 5e), it can be seen that among the 26 pathways, B CELL SIGNALING PATHWAY, PRIMARY IMMUNODEFICIENCY and other pathways increase with the increase of RiskScore, and FOCAL ADHESION, GALACTOSE_METABOLISM and other metabolism-related pathways decrease with the increase of RiskScore. Results suggested that the imbalance of these pathways is closely related to the development of tumors.

In addition, we obtained the signature genes of three immune cells, Activated B cell, Activated CD4 T cell, and Activated CD8 T cell, from a previous study [21], and calculated enrichment scores in each sample using the method of ssGSEA to assess the sample’s corresponding Immune cell scores. The differences in these three immune cell scores in the high and low risk groups of patients were analyzed and observed that B cell and T cell activation scores were all significantly lower in high risk patients with poor prognosis (Fig. 5f). Current immunotherapy-related datasets are rare. We found a cohort of PD-L1-treated patients with metastatic uroepithelial carcinoma shared by Sanjeev Mariathasan et al [22] and analyzed the differential expression of 19 genes in 14 IRGPs in patients with different response states after PD-L1 treatment. We observed a significant differential expression of 14 (73.6%) genes (Fig. 5g), suggesting that these genes are associated with immunotherapy.

In addition to 14-IRGPs, clinical features Stage and Age are also independent prognostic factors, indicating that they have complementary values. In order to further improve the accuracy of prediction, a new nomogram was established by integrating Stage, Age and 14-IRGPs using Cox model. According to this model, 14-IRGPs contribute the most to OS, followed by Age and Stage (Fig. 6a). By calculating the total score, oncologists could easily obtain the OS probability predicted by the nomogram of an individual patient. Furthermore, we used the calibration curve to evaluate the prediction accuracy of the model (Fig. 6b), results show that the predicted calibration curves of the three calibration points in 1, 3, and 5 years were close to the standard curve, which indicated that the model has good prediction performance. In addition, we also used DCA (Decision curve) to evaluate the reliability of the model (Fig. 6c). It was observed that RiskScore (14-IRGPs) and nomogram benefit significantly higher than the extreme curve, and nomogram is higher than RiskScore, and Age and Stage are close to the extreme curve. This suggests that RiskScore (14-IRGPs) and nomogram have good reliability.

In order to observe the performance of 14-IRGPs, the prognostic signature of three head and neck cancers reported in the past (3-gene signature of Cui L et al [23], 6-gene signature of Weidong Zhang et al [24] and 3-gene signature of Hongbo Zhou et al [25]) and four clinical features of T, N, age and Stage were selected. In order to make the model comparable, we calculated the Risk score of each head and neck cancer sample in the TCGA training set with the same method according to the corresponding genes in the 3 models, evaluated the ROC of each model and C-index (Table 5). We observe that the 6-gene signature model has the highest AUC above among the three models, while the average AUC for 1, 3 and 5 years is 0.63. The 1, 3, and 5 years AUC of 14-IRGPs signature were all above 0.73. In addition, in the C-index of all models, 14-IRGPs was significantly higher than other clinical features and models, indicating that our model has good application value.