Identification of an immune signature predicting prognosis risk of patients in lung adenocarcinoma

The expression data were downloaded from the TCGA database and the GEO database. The RNA-seq data of 500 LUAD patients were collected from the TCGA database and used as the training set, which were downloaded from University of California Santa Cruz (UCSC) Genome Browser (https://​xena.​ucsc.​edu/​public-hubs/​). GSE 81089 was the other RNA-seq data of 108 LUAD patients downloaded from the GEO database (http://​www.​ncbi.​nlm.​nih.​gov/​geo), which was used as one of the testing sets for constructing this signature. Fragments per kilobase of exon per million fragments mapped (FPKM) value was used to measure all of the RNA-seq data. The microarray data from GSE30219 (N = 85), GSE31210 (N = 226), GSE3141 (N = 57) were also collected from the GEO database and used as testing sets, respectively. A total of 976 patients were included for analysis. The clinical and survival information of the included datasets were summarized in Table 1. The comprehensive list of immune related genes containing a total of 1534 genes were downloaded from the ImmPort database (https://​immport.​niaid.​nih.​gov) [15].

The cases from the TCGA database were used as the training set to develop the immune signature. Univariate analysis and logRank test were used to identify immune related genes with prognostic ability. For the genes with prognostic ability, Cox proportional hazards model (iteration = 1000) with an lasso penalty was used to find the best gene model utilizing a R package called “glmnet” [16]. The best gene model was used to establish the immune signature. Then, the concordance (c)-index proposed by Harrell et al. [17] was applied to validate the predictive ability of the signature in all of the five datasets, by using the “survcomp” R package [18]. The larger c-index indicated the more accurate predictive ability of the model.

The Kaplan–Meier (K–M) survival curves were generated to graphically demonstrate the overall survival (OS) of the high-risk group and low-risk group which were stratified by the immune signature. The univariate and multivariate analyses of survival were conducted for both the immune signature and clinicopathologic factors. The R package called “survival” was utilized to perform the survival analysis.

Mutation data that contained somatic variants were stored in Mutation Annotation Format (MAF) form and were downloaded from Genomic Data Commons (GDC) (https://​portal.​gdc.​cancer.​gov/​).

Nonsynonymous mutations were used for our investigations, considering the uncertainty of functional consequences of synonymous mutations. And nonsynonymous mutations were potential sources of neoantigen epitopes. Nonsynonymous mutations included missense mutation, nonsense mutation, splice site mutation, frameshift mutation, and inframe mutation. The total number of nonsynonymous mutations were utilized as the mutation burden of LUAD patients to investigate the relationship between the immune signature and patients’ mutation load. The single nucleotide polymorphism (SNP) was also analyzed for its association with our signature. The number of neoantigens was cited from a published study [19] so as to figure out the correlation of the signature with the number of neoantigens in LUAD patients.

Student’s t test was conducted to make statistical comparison. The “ggplot” R package was used to generate boxplots. “ComplexHeatmap” R package was applied to generate heatmaps [20]. Two-tailed p values less than 0.05 were thought to be statistically significant. All of our analyses were conducted using R software version 3.5.1 (https://​www.​r-project.​org/​).

To make our investigations clearer, a workflow that illustrated the generation of the signature was demonstrated in Fig. 1. The univariate analysis was performed in all of the 1534 immune related genes for TCGA LUAD datasets. There were 144 genes with prognostic ability after the univariate analysis and logRank test (P < 0.05). The 144 immune related genes then underwent the Cox proportional hazards regression with tenfold cross-validation to generate the best gene model. We totally performed 1000 iterations and included 10 gene groups for further screening. The gene lists of the 10 gene groups were shown in Additional file 1: Table S1. As illustrated in Fig. 2a, a gene model with 30 immune related genes was with the highest frequencies of 211 compared to other nine gene models. Thus, this gene model became the most suitable role to generate the immune signature for LUAD. Therefore, we utilized the 30 immune related genes in this gene model to construct our immune signature, as listed in Additional file 1: Table S1. The coefficient value of the 30 genes were listed in Additional file 2: Table S2. The prognostic ability of the 30 immune related genes in LUAD patients was confirmed in the training set (The TCGA dataset, Additional file 3: Figure S1), which showed that all of the 30 genes were able to predict survival outcome of LUAD patients. However, the prognostic ability of the 30 genes was not consistent in the four testing sets (The GEO datasets, Additional files 4, 5, 6, 7: Figures S2–S5).

To validate our signature, we firstly calculated the c-index for the prediction of OS. The c-index for TCGA dataset, GSE30219, GSE31210, GSE3141, and GSE81089 were 0.723, 0.657, 0.7061, 0.641, and 0.619 respectively (P < 0.05, Fig. 2b), which indicated the high predictive accuracy of the signature for survival. Then, the risk score for each patient was calculated according to the coefficient value of the 30 genes. Patients were divided into high-risk and low-risk groups with the median risk score utilized as the cutoff value, as demonstrated in Fig. 3a–e. Patients of high-risk were with poor OS compared with those of low-risk in both TCGA and GEO datasets (Fig. 4a–e, P < 0.05). We further validated the prognostic ability of the signature in subgroups of LUAD, and we found the immune signature could also predict the survival outcome of patients in clinically important subgroups. In TCGA datasets, patients in high risk group demonstrated poor prognosis in T1–3 stage, N0–3 stage, M0–1 stage, stage I–IV, recurrence, and no recurrence (P < 0.05, Additional file 8: Figure S6). In GSE30219 datasets, patients in high risk group showed poor prognosis in T1 stage, N0 stage (P < 0.05, Additional file 9: Figure S7). In GSE31210 dataset, high risk patients in stage I group demonstrated poor survival outcome (P < 0.05, Additional file 10: Figure S8). In GSE81089 datasets, patients in stage III exhibited a negative correlation between the risk score and patients’ OS (P < 0.05, Additional file 11: Figure S9). The univariate Cox analysis of the immune signature also indicated the significant association of the signature with LUAD patients’ OS in both TCGA and GEO datasets (P < 0.05, Fig. 5). Multivariate Cox analysis further exhibited that our signature could serve as an independent predictor of patients’ survival outcome after adjusted by clinicopathologic factors including age, TNM stage, recurrence, and gender in TCGA cohort [Hazard ratio (HR) = 2.1868, 95% confidence intervals (95% CI) 1.7612 to 2.7152, P < 0.001], GSE30219 cohort (HR = 1.6354, 95% CI 1.1632–2.2993, P = 0.0047), and GSE81089 (HR = 1.5156, 95% CI 1.1425–2.0106, P = 0.0039), as demonstrated in Fig. 6. As for the prognostic ability of clinical factors, we found only tumor recurrence and stage IV could serve as independent predictors for patients’ OS, which indicated the strong prognostic ability of our signature.

To further validate the clinical value of the 30-genes immune signature, we evaluated the relationship between the signature and clinicopathologic factors. In TCGA cohort, patients of high-risk were tended to have advanced T stage, N stage, M stage, pathological stage and were under high risk of recurrence (P < 0.05, Additional file 12: Figure S10). In GSE30219 cohort, high-risk score was associated with higher T stage and N stage (P < 0.05, Additional file 13: Figure S11). In GSE31210 cohort, the risk score was only positively related to advanced pathological stage (P < 0.05, Additional file 14: Figure S12). And we did not find the association of the risk score and pathological stage in GSE81089 (Additional file 15: Figure S13).

Higher nonsynonymous mutation burden load and neoantigen number have shown associations with clinical efficacy of immune checkpoint inhibitor therapy [21, 22]. Therefore, we would like to investigate whether our immune signature could affect mutation load and number of neoantigen of LUAD for the possibility of the risk score to be the predictor of response to immune checkpoint inhibitor. Patients with high risk score exhibited higher nonsynonymous mutation load than those with low risk score (P = 0.0112, Fig. 7a). To further explore which types of nonsynonymous mutation were the major contributors to this relationship, we evaluated the association of the signature with different types of nonsynonymous mutation. High-risk group patients had higher missense mutation (P = 0.0098, Fig. 7b), nonsense mutation (P = 0.0166, Fig. 7c), splice site mutation (P = 0.0217, Fig. 7d), and inframe deletion (P = 0.0085, Fig. 7g). We did not find this relationship in frameshift mutation (Fig. 7e, f), inframe insertion (Fig. 7h), total deletion mutation (Fig. 7j), and total insertion mutation (Fig. 7k). Besides, we found there demonstrated a positive correlation between the signature and the number of SNP (P = 0.0098, Fig. 7i). However, we did not find correlation between the signature and the number of neoantigens in LUAD (Fig. 7l).