Novel immune-related gene signature for risk stratification and prognosis prediction in ovarian cancer

The following 9 ovarian cancer (OC) expression chip datasets: GSE14407, GSE6008, GSE14001, GSE16708, GSE26712, GSE29450, GSE38666, GSE66957, and GSE105437, were downloaded from the Gene Expression Omnibus (GEO) database (www.​ncbi.​nlm.​nih.​gov/​geo/​) to compare the expression of 2498 immune-related genes (IRGs) in 428 OC and 77 normal ovary tissue specimens. 2498 IRGs were derived from the ImmPort database (https://​www.​ImmPort.​org/​home). The transcriptome RNA-sequencing data and corresponding clinical information of 374 and 93 OC patients were extracted from TCGA database (https://​portal.​gdc.​cancer.​gov/​) and ICGC data portal (https://​dcc.​icgc.​org/​), respectively. The TCGA data and ICGC data were applied as a prognostic model training set an external validation set, respectively. Over 20,000 primary cancer samples and matched normal samples were included in the TCGA database, which held over 2.5 petabytes of genomic, epigenomic, transcriptomic, and proteomic data. The ICGC Data Portal is a collaborative effort to describe genetic anomalies in 50 different cancer types.

Expression matrix data from GEO database containing 9 datasets from different labs were normalized with Limma R package. We also removed the batch effect between TCGA and ICGC datasets by SVA package in R. The differentially expressed IRGs (DEIRGs) were identified in 428 OC and 77 normal ovary tissue specimens from 9 independent GEO datasets using the limma R package, and the cutoffs were |log₂FoldChange| (|log₂FC|) > 1 and p < 0.05.

The biological function of the DEIRGs was investigated using GO term enrichment analysis and KEGG pathway enrichment analysis through “clusterProfiler” R package. The protein–protein interaction (PPI) network was determined among all DEIRGs using STRING database (https://​string-db.​org/​).

First, DEIRGs with prognostic values were screened via univariate cox analysis in the TCGA training cohort. Then, to avoid overfitting, we performed the least absolute shrinkage and selection operator (LASSO) Cox regression analysis with the identified prognostic genes using R package “glmnet” to construct an immune gene signature for OC patients in the TCGA training cohort. The independent variable in the LASSO analysis was the standardized expression matrix of prognostic DEIRGs identified before, and the response variables were survival status and OS of OC patients in the TCGA training cohort. Finally, a prognostic immune gene signature for assessing survival risk of OC patients was finally constructed using the standardized expression levels of independent prognostic DEIRGs and their corresponding regression coefficients. The formula of the risk score for each patient was:

\(\mathrm{The risk score}=\sum_{\mathrm{i}-\mathrm{1,2},\dots ,\mathrm{n}}\mathrm{regression coefficient }(\mathrm{genei})\times \mathrm{expression value of }(\mathrm{genei})\). OC patients were divided into low- and high-risk groups by the median risk score as the threshold.

The immune gene-based risk model divide OC patients in the TCGA training cohort and the ICGC validation cohort into a high-risk group and low-risk group respectively. The Kaplan–Meier survival curves were plotted by R package “survminer” to compare the survival differences of OC patients in different risk groups. Five-year receiver operating characteristic (ROC) curves of risk score and clinical features were plotted via “survival ROC” package to describe accuracy and performance of the model in the TCGA training cohort and ICGC validation cohorts. The larger of the area under the curve (AUC) of ROC curve, the more accurate of the model. We also exhibited the relationship between survival status of OC patients and risk scores in the training cohort and validation cohort, respectively. Principal component analysis (PCA) and t-SNE analysis were performed by R package “stats” and “Rtsne” to verify the distribution of high-risk and low-risk group patients.

The risk score was compared with the clinical traits to determine whether the risk score was associated with the clinical characteristics of OC patients in both TCGA training cohort and ICGC validation cohorts. Age, grade, pathological stage, and overall survival (OS) time were among the OC clinical data that were obtained from the TCGA database. The age and OS of patients were obtained from ICGC data portal. The relationship between the risk score ang these clinicopathological indexes were evaluated. To identify the independence of our risk score signature, univariate and multivariable Cox regression analyses were performed with R package “survival” in TCGA cohorts to identify independent prognostic indicators among risk score and clinical factors. Age, grade, pathological stage and Risk score were included in TCGA.

A nomogram was generated using the R package “rms” to predict the probability of 1-, 3-, 5- and 10-year OS of OC patients based on the independent prognostic DEIRGs that screened out for building the risk model.

The immune infiltration of OC patients was derived from Tumor Immune Estimation Resource (TIMER) website (https://​cistrome.​shinyapps.​io/​timer/​). The association between the abundance of 6 immune infiltrates cells (CD4+ T cells, dendritic cells, CD8+ T cells, B cells, macrophages, and neutrophils) and the immune gene-based risk model were analyzed using R.

After quality assessment and normalized of GEO sequencing data from 9 independent labs (Table 1, Fig. 1A), a total of 1211 gene were found to be differentially expressed genes (DEGs) in 428 OC tissues compared with 77 adjacent tissues, including 567 upregulated and 644 downregulated DEGs (Fig. 1B), we then analyzed expression of 2498 IRGs and obtained 129 differentially expressed IRGs (DEIRGs), including 79 DEIRGs and 50 DEIRGs that were upregulated and downregulated respectively (Fig. 1C).

A protein–protein interaction (PPI) network of DEIRGs was established and visualized in Fig. 2A. Enrichment analysis of DEIRGs showed that biological processes (BP), mainly leukocyte migration and cell chemotaxis were primarily enriched whereas the main molecular function (MF) consists of signaling receptor activator activity and receptor ligand activity. The most enriched cellular components (CC) were external side of plasma membrane (Fig. 2B). KEGG pathway indicated that the DEIRGs were mainly involved in Epstein Barr virus infection and antigen processing and presentation (Fig 2C).

We put the 129 DEIRGs obtained in the previous step into the TCGA training cohort (n=374) for identifying DEIRGs corelated with overall survival (OS) of OC patients. Univariate Cox regression analysis revealed 12 prognostic DEIRGs were significantly related to OS in OC patients (Fig. 3A). For avoiding overfitting, the 12 prognostic DEIRGs were included in the least absolute shrinkage and selection operator (LASSO) analysis to construct a prognostic risk model. According to the penalty parameter (Lambda) obtained in LASSO analysis, there are 5 independent prognostic DEIRGs (ANGPT4, PLTP, A2M, CXCR4 and MIF) were used to establish a risk score model in the TCGA training set (Fig. 3B-C). The correlation coefficient of each DEIRGs make up the risk model was shown in Table 2. Risk score = (0.3697* ANGPT4) +(0.0851* PLTP) +(-0.1118* A2M) +(-0.0423* CXCR4) +( -0.0492* MIF).

We calculated the risk score for 374 OC patients in the TCGA cohort and divided the patients into low-risk (n=187) and high-risk (n=187) groups based on the median cutoff value. The 93 OC patients in the ICGC cohort were also divided into low-risk (n=57) and high-risk (n=36) groups based on the same median cutoff value.

Figure 4A exhibited that in 374 OC patients from TCGA cohort, the OS of low-risk patients was markedly higher compared to that of high-risk patients. The 5-year ROC curve of Risk score and other clinical parameters were plotted to assess the reliability of the risk model, and the areas under the curve (AUCs) of risk score is 0.759, higher than any other clinical parameters (Fig. 4B). The risk score distribution of OC patients in the TCGA training cohort is shown in Fig. 4C, as the risk score increased, increase in number of deaths occurred (Fig. 4D). The expression patterns of 5 prognostic DEIRGs that composed the risk model in the TCGA training cohort was visualized (Fig. 4E).

The stability of the immune gene-based risk model was also examined in the ICGC-validation cohort. Consistent with the TCGA training cohort, OC patients in the low-risk group showed a substantially higher survival rate (Fig. 4F). In the results of the 5-year ROC curve of the ICGC validation cohort, the AUC value of the risk score was 0.778 and the significance for evaluating the prognosis far exceeded other clinical indicators (Fig. 4G). The risk score performed well not only in the training cohort but also in the validation cohort (TCGA-AUC = 0.759, ICGC-AUC = 0.778). Risk score distribution and corresponding survival status of OC patients in the ICGC validation cohort were presented in Fig. 4H-I. The expression profile of 5 risk genes in ICGC cohort was also exhibited in Fig. 4J.

All above results suggested that the model had good accuracy and general applicability.

Complete clinicopathological data were extracted and integrated with risk score of OC patients in TCGA cohort and ICGC cohort respectively, and evaluated whether the risk score could independent irrespective of other clinical features to be a prognostic factor. We used principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) analysis for data dimensionality reduction and found that patients in high-risk and low-risk groups in TCGA dataset (left panel) and ICGC dataset (right panel) were distributed in 2 discrete directions (Fig. 5A). The corresponding scatter diagrams determined by the Wilcoxon Rank Sum Test showed that age and pathological stage of patients in the TCGA training cohort were significantly related to the risk score. No significant differences were observed between various grade groups in the TCGA dataset or different age groups in ICGC dataset (Fig. 5B). All clinicopathological parameters and risk score were subjected to a univariate independent prognostic analysis, the result reflected age and risk score were significantly associated with the OS of the patients in the TCGA training cohort (Fig. 5C, left panel). Multivariate analysis including age, grade, pathological stage, and riskScore. The multivariate independent prognostic analysis results exhibited that the risk score was the only one independent prognostic judgment factor (Fig. 5C, right panel). To construct and visualize a survival prediction method for OC patients, a prognostic nomogram including 5 risk genes which made up the risk model was developed (Fig. 5D).

To evaluate whether our risk score model could reflect the tumor immune microenvironment in OC patients, we analyzed the relationship between risk score and immune cell infiltration in the TCGA training dataset. As show in Fig. 6, CD4+ T cells, CD8+ T cells, B cells, dendritic cells, and neutrophils were negative correlated with risk score. It indicated that the level of immune cell infiltration was downregulated in high risk OC patients.