The Combined Detection of Immune Genes for Predicting the Prognosis of Patients With Non-Small Cell Lung Cancer

Abstract

Lung cancer is one of the leading causes of cancer-related death. In recent years, there has been an increasing interest in the fields of tumor and immunity. This study focused on the possible prognostic value of immune genes in non-small cell lung cancer patients. We used The Cancer Genome Atlas (TCGA) to download gene expression data and clinical information of lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). The immune gene list was downloaded from the Immport database. We then constructed immune gene prognostic models on the basis of Cox regression analysis. We further evaluated the clinical significance of the models via survival analysis, receiver operating characteristic (ROC) curves, and independent prognostic factor analysis. Moreover, we analyzed the associations of prognostic models with both mutation burdens and neoantigens. Using the Gene Expression Omnibus (GEO) and Kaplan–Meier plotter databases, we evaluated the validity of the prognostic models. The prognostic model of LUAD included 13 immune genes, and the prognostic model of LUSC contained 10 immune genes. High-risk patients based on prognostic models had a lower 5-year survival rate than did low-risk patients. The ROC curve analysis demonstrated the prediction accuracy of the prognostic models, as the area under the curve (AUC) was 0.742, 0.707, and 0.711 for LUAD, and 0.668, 0.703, and 0.668 for LUSC, when the predicted survival times were 1, 3, and 5 years, respectively. The mutation burden analysis showed that mutation level was associated with the risk score in patients with LUAD. The analysis based on GEO and Kaplan–Meier plotter demonstrated the prognostic validity of the models. Therefore, immune gene-related models of LUAD and LUSC can predict prognosis. Further study of these genes may enable us to better distinguish between LUAD and LUSC and lead to improvement in immunotherapy for lung cancer.

Introduction

It was estimated that lung cancer would lead to the largest number of cancer-related deaths in both male and female patients by 2020.¹ The 1-year survival rate of lung cancer is less than 50%. For patients in the early stage, the 5-year survival rate can reach 56%, but is only 5% for patients in the advanced stage.² Therefore, it is necessary to update the knowledge on lung cancer to help patients achieve a better prognosis.

There is a large volume of published studies describing the role of the immune system in lung cancer initiation and progression.³ At the same time, the immune system is also considered as an important component of the tumor microenvironment, which mainly includes various stromal cells (fibroblasts and endothelial cells), immune cells (T cells, B cells, dendritic cells, macrophages, and neutrophils), various factors secreted by cells (cytokines, chemokines, hormones, etc.), extracellular matrix, and the vascular system.⁴

These immune-related cancer studies have mainly focused on the following aspects: interaction of tumor cells and tumor-infiltrating immune cells in the tumor microenvironment, especially via exosomes, which further contribute to form the pre-metastatic niche⁵; immune cells’ influence on cancer activity, such as eosinophils playing an anticancer role in cryothermal treatment⁶ and reactivation of dysfunctional natural killer (NK) cells inhibiting tumor growth⁷; levels of immune cells being related to cancer subtypes and prognosis, which is not only demonstrated in research based on cancer databases but also experimentally confirmed for certain cancer types^8,9; and some immune factors being identified as prognostic indicators.^10,11 Moreover, studies on the progress of immune checkpoint blockers have aroused great interest in researchers studying immune-related cancer therapy.¹²

The immune system plays a significant role in lung cancer, and its state largely determines the response to treatment.¹³ Therefore, in this study, we used The Cancer Genome Atlas (TCGA) database to establish immune gene-related prognosis models in lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). We evaluated the prognostic effect of models by TCGA, Gene Expression Omnibus (GEO), and Kaplan–Meier plotter databases. This work may provide new ideas for in-depth mechanistic research and immunotherapy.

Materials and Methods

Data Source

The gene expression data (FPKM values) and clinical data as the training set were downloaded from the TCGA database (https://portal.gdc.cancer.gov). The immune gene list was acquired using the Immport database (https://www.immport.org/shared/genelists). The perl script (perl version 5.28.1) was used to process expression data to obtain mRNA matrix, convert Ensembl ID into gene symbols, and extract relevant clinical information (including survival time, survival status, age, gender, TNM stage, and recurrence) from the downloaded clinical data. The test set GSE3141, including the microarray data, was downloaded from the GEO database. R package sva, which removes batch effects and other unwanted variations in high-throughput experiments, was used to calibrate TCGA and GEO data. We summarized the clinical information in Table 1. In the training set, the gene expression data of LUAD involved 535 tumor samples and 39 normal samples, and for LUSC, 502 tumor samples and 49 normal samples.

Table 1.

The Clinical Information of Lung Cancer Patients From TCGA and GEO Databases.

Database	TCGA
Platform	Illumina HiSeq2000 RNA sequencing platform
Histology types	LUAD	LUSC
Age
≤60 years	160 (30.7%)	108 (21.4)
>60 years	343 (65.7.0%)	387 (76.8)
Unknow	19 (3.6%)	9 (1.8)
Gender
Male	242 (46.4%)	373 (74.0)
Female	280 (53.6%)	131 (26.0)
TNM stage
Stage I	279 (53.4%)	245 (48.6)
Stage II	124 (23.8%)	163 (32.3)
Stage III	85 (16.3%)	85 (16.9)
Stage IV	26 (5.0%)	7 (1.4)
Unknow	8 (1.5%)	4 (0.8)
T stage
T1	172 (33.0%)	114 (22.6%)
T2	281 (53.8%)	295 (58.5%)
T3	47 (9.0%)	71 (14.1%)
T4	19 (3.6%)	24 (4.8%)
TX	3 (0.6%)	0 (0.0%)
Unknow	0 (0.0%)	0 (0.0%)
N stage
N0	335 (64.2%)	320 (63.5%)
N1-3	175 (33.5%)	178 (35.3%)
NX	11 (2.1%)	6 (1.2%)
Unknow	1 (0.2%)	0 (0.0%)
M stage
M0	353 (67.6%)	414 (82.1%)
M1	25 (4.8%)	7 (1.4%)
MX	140 (26.8%)	79 (15.7%)
Unknow	4 (0.8%)	4 (0.8%)
Recurrence
Yes	109 (20.9%)	84 (16.7%)
No	197 (37.7%)	193 (38.3%)
Unknow	216 (41.4%)	227 (45.0%)
Survival status
Alive	355 (68.0%)	304 (60.3%)
Dead	167 (32.0%)	200 (39.7%)
Database	GEO (GSE3141)
platform	Affymetrix Human Genome U133 Plus 2.0 Array (GPL570)
Histology types	LUAD	LUSC
Survival status
Alive	26 (44.8%)	27 (50.9%)
Dead	32 (55.2%)	26 (49.1%)

Identification of Differentially Expressed Immune Genes

Using the R (https://www.r-project.org/, version 3.6.1) software, the shared genes between the mRNA matrix and immune gene lists were obtained. Through this step, we obtained expression data for immune genes. Differential expression analysis of immune genes was performed using the R package limma. The fold change (FC) of genes was indicated as the base-2 logarithm of FC (logFC). The Benjamini–Hochberg method was used to correct the P value. We considered differentially expressed immune genes as |logFC| > 1 and adjusted P value (adj. P or P. adjust) < 0.01. Adj. P is also called the false discovery rate (FDR or fdr) value.

Analysis of Volcano Plots and Heatmaps

We used heatmaps and volcano plots to present the results of differentially expressed immune genes. The R package pheatmap was used to plot the heatmap.

Establishment of Immune Gene Prognostic Model based on Cox Proportional Hazards Regression Model

First, for clinical information, we removed the patients without survival information or with a survival time of less than 90 days. Next, perl scripts were used to merge the data of differentially expressed immune genes with the corresponding clinical information. Then, the R package survival was used to conduct univariate and multivariate Cox regression analyses. The results of the univariate regression analysis are displayed in the forest plot. The differentially expressed immune genes meeting the filtering criteria (P < 0.01) in univariate Cox regression analysis were further used to build the immune gene prognostic model by multivariate Cox regression analysis. Using the model, we calculated risk scores that equaled the sum of the products of gene expression levels and the corresponding coefficients (∑expression levels * coefficients). The median risk score as the cutoff value was used to divide patients into high- and low-risk groups.

Evaluation of Prognostic Model Based on Survival Curve, ROC Curve, and Independent Prognostic Factor Analysis

R package survival and survminer were used to plot survival curves. R package survivalROC was applied to plot the ROC curve. Univariate and multivariate Cox regression analyses were used to determine independent prognostic factors for lung cancer patients, and the factors meeting P values of less than 0.05 in both univariate and multivariate analyses were considered as independent prognostic factors.

Mutation Burden and Neoantigen Analyses

Mutation data processed by the workflow called VarScan2 Variant Aggregation and Masking were downloaded from the TCGA database. The mutation burden was evaluated by the mutation count excluding synonymous mutations per million bases. We obtained neoantigen data based on the TCGA database from a published study.¹⁴

GEO and Kaplan–Meier Plotter Database

GSE3141 was used to validate the prognostic ability of models by survival curve analysis. R package survival and survminer were used to plot survival curves. We obtained survival curves of the single immune gene using the Kaplan–Meier plotter database (https://kmplot.com). We considered that an immune gene was related to survival rate when all probe sets per gene met the criterion of P < 0.05.

Statistical Analysis

For the statistical analysis of the difference in mutation burden and neoantigen levels between the high- and low-risk groups, the nonparametric test (Mann-Whitney U test) was used. The Kaplan–Meier method was used to generate survival curves, and the log-rank test was used to determine statistical differences. P < 0.05 (2-tailed) was considered statistically significant.

Results

Screening of Differentially Expressed Immune Genes and Presentation of Volcano Plots and Heatmaps

Figure 1 shows the workflow of the construction and validation of the prognostic models. A total of 474 differentially expressed immune genes were selected for LUAD (Table SI), including 312 up-regulated genes and 162 down-regulated genes, and 565 genes were selected for LUSC (Table SII), including 279 up-regulated genes and 286 down-regulated genes. We then displayed the gene expression difference of these immune genes in volcano plots (Figure 2A for LUAD and Figure 2B for LUSC) and heatmaps (Figure 2C for LUAD and Figure 2D for LUSC).

Figure 1.

The workflow of the construction and validation of the prognostic models.

Figure 2.

The volcano plot and heatmap of differentially expressed immune genes. In the volcano plot, red points represent up-regulated genes (logFC > 1 and adj. P < 0.01) and green points represent down-regulated genes (logFC < -1 and adj. P < 0.01), while black points indicate genes without significant differential expression (|logFC| < 1 or adj. P > 0.01). In the heatmap, genes with higher expression are shown in red and genes with lower expression are shown in green, while genes with the same expression level are in black. (A) Volcano plot for LUAD; (B) Volcano plot for LUSC; (C) Heatmap for LUAD; (D) Heatmap for LUSC.

Construction of an Immune Gene Prognostic Model

A total of 21 genes were identified in the univariate Cox regression analysis in LUAD (Figure 3A, P < 0.01) and 23 genes in LUSC (Figure 3B, P < 0.01). There were no common genes between LUAD and LUSC. Based on these possible prognostic genes, we built immune gene prognostic models by multivariate Cox regression analysis, as detailed in Table 2 for LUAD and Table 3 for LUSC. Although some genes did not meet the criterion of P < 0.01, it was significant to include these genes in this model when 2 types of genes (P < 0.01 and P >= 0.01) as a whole were statistically related to the prognosis. The genes meeting P < 0.01 in both univariate and multivariate Cox regression analyses were considered as independent prognostic factors.

Figure 3.

Forest plots of univariate Cox regression analysis. In forest plots, the vertical dotted line represents hazard ratio (HR) = 1; the green box represents HR < 1.0, which indicated that the immune gene was a favorable prognostic biomarker; Conversely, red box represents HR > 1.0, which identified the immune gene as a poor prognostic indicator. The length of the horizontal line represents the 95% confidence interval for each immune gene. (A) Forest plot for LUAD; (B) Forest plot for LUSC.

Table 2.

Immune Gene Prognostic Model of LUAD.

Symbol	Coefficient	HR	HR.95L1	HR.95H1	P value
S100A16*	0.001608	1.00161	1.000242	1.002979	0.021052
CRABP1	0.004246	1.004255	1.001423	1.007095	0.003214
RBP2	0.063071	1.065102	1.030106	1.101287	0.000215
FGF2	0.294875	1.342959	1.141534	1.579924	0.000376
BTK*	-0.06332	0.93864	0.858823	1.025875	0.162546
SEMA4B*	0.005656	1.005672	1.000444	1.010928	0.033433
IL11*	0.12218	1.129957	1.010837	1.263115	0.031587
INHA*	0.006085	1.006104	1.000331	1.01191	0.038193
ANGPTL4*	0.004722	1.004733	1.000056	1.009432	0.04731
LGR4*	0.013986	1.014084	0.99732	1.03113	0.10009
TNFRSF11A	0.215939	1.241026	1.116963	1.37887	5.86E-05
VIPR1*	-0.10309	0.902048	0.812785	1.001113	0.052497
SHC3*	-0.15857	0.853363	0.701788	1.037676	0.111995

¹HR95 L and HR95 H indicated the lower and upper limits of the 95% confidence interval; * These genes had P values that did not meet the standard P < 0.01 in the multivariate Cox regression analysis.

Table 3.

Immune Gene Prognostic Model of LUSC.

Symbol	Coefficient	HR	HR.95L1	HR.95H1	P value
CXCL5*	0.007292	1.007319	1.001712	1.012957	0.010455
PLAU	0.002959	1.002963	1.001316	1.004612	0.000417
RNASE7*	0.011302	1.011366	1.002049	1.02077	0.01669
IGKV1-6	0.000471	1.000471	1.000227	1.000715	0.000153
SEMA4C	0.013973	1.014071	1.00448	1.023753	0.003953
APLN*	0.042824	1.043754	1.002141	1.087095	0.039111
TSLP*	-0.20922	0.811214	0.686553	0.958509	0.013981
FGFR4*	0.042192	1.043095	1.003405	1.084354	0.03303
TRAV39	0.344197	1.410856	1.13394	1.755397	0.002019
JUN*	0.002498	1.002501	0.999229	1.005785	0.134278

¹HR95 L and HR95 H indicated the lower and upper limits of the 95% confidence interval; * These genes had P values that did not meet the standard P < 0.01 in the multivariate Cox regression analysis.

Clinical Significance of the Immune Gene Prognostic Model

According to the prognostic model, we calculated the risk scores for patients. Based on median risk scores, 456 patients with LUAD and 431 patients with LUSC were divided into high- and low-risk groups. To evaluate the prognostic model, the Kaplan–Meier method was used to generate survival curves and the log-rank test was used to determine the statistical difference (Figure 4A for LUAD and Figure 4B for LUSC). The results showed that the survival rate was significantly different between the high- and low-risk groups of LUAD and LUSC (P = 1.099e-07 and P = 2.082e-05, respecitvely). For LUAD, the 5-year survival rate in the low-risk group was 50% and 20% in the high-risk group. The 5-year survival rate of patients with LUSC in the low-risk group was 60% and that in the high-risk group was 36%.

Figure 4.

Survival curves analysis. At the bottom of survival curves, 2 lines of figures represent the number of survivors in high- and low-risk groups, which decreases gradually with follow-up time. (A) Survival curve for LUAD; (B) Survival curve for LUSC.

The area under the curve (AUC) of the ROC analysis was used to reflect the prediction accuracy of the prognostic model. When the predicted survival time was 1 year, 3 years, and 5 years, the corresponding AUCs were 0.742, 0.707, and 0.711 for LUAD (Figure 5A), and 0.668, 0.703, and 0.668 for LUSC, respectively (Figure 5B).

Figure 5.

ROC curves analysis. (A) ROC curve for LUAD; (B) ROC curve for LUSC.

In univariate and multivariate Cox regression analyses of LUAD (Figure 6A of univariate analysis and Figure 6B of multivariate analysis), the results showed that recurrence and risk scores were considered as independent prognostic factors (P < 0.05). Likewise, for LUSC (Figure 6C of univariate analysis and Figure 6D of multivariate analysis), recurrence and risk scores were also considered as independent prognostic factors (P < 0.05).

Figure 6.

The univariate and multivariate Cox regression analyses. (A) The univariate Cox regression analysis of LUAD; (B) The multivariate Cox regression analysis of LUSC; (C) The univariate Cox regression analysis of LUSC; (D) The multivariate Cox regression analysis of LUSC.

Associations of Prognostic Models With Non-Synonymous Mutations and Neoantigens

Somatic mutations and neoantigen production are associated with cancer immunity and immunotherapy.¹⁵ In our study, we analyzed the difference in non-synonymous mutation burden between the low- and high-risk groups, based on prognostic models, as well as the difference in predicted neoantigens between the 2 groups. In LUAD, patients in the high-risk group had higher mutation burdens than patients in the low-risk group (Figure 7A, P = 0.0005). However, regarding LUSC, the difference in mutation burdens between the 2 groups was not statistically significant (Figure 7B, P = 0.1241). In LUAD and LUSC, there was no statistical difference in the predicted antigens between the 2 groups (Figure 7C, LUAD, P = 0.7559; Figure 7D, LUSC, P = 0.8353).

Figure 7.

Mutation burden and neoantigen analyses. (A) Mutation burden analysis for LUAD; (B) Mutation burden analysis for LUSC; (C) Neoantigen analysis for LUAD; (D) Neoantigen analysis for LUSC.

Validation by GEO and Kaplan–Meier Plotter Databases

We used GSE3141 from the GEO database to validate the prognostic ability of the models. For both LUAD and LUSC, there were significant differences in survival rates between the high- and low-risk groups (Figure 8A, LUAD, P = 4.682e-03; Figure 8B, LUSC, P = 3.836e-02). In addition, to analyze the prognostic effect of the single immune gene included in prognostic models, we used the Kaplan–Meier plotter database, which included lung cancer data mainly from the GEO database. Figures S1 and S2 show the survival curves of all probe sets per immune gene. We found that for LUAD (Figure S1A-1, S1A-2, S1B-1 and S1B-2), S100A16 (probe set, 227998_at; HR = 2.25; P = 1.4e-10), CRABP1 (probe set, 205350_at; HR = 1.61; P = 6.8e-05), BTK (probe set, 205504_at; HR = 0.7; P = 0.0028), SEMA4B (probe set, 234725_s_at; HR = 1.45; P = 0.0025), INHA (probe set, 210141_s_at; HR = 2, P = 7.2e-09), ANGPTL4 (probe set, 223333_s_at; HR = 1.69; P = 2.6e-05), and ANGPTL4 (probe set, 221009_s_at; HR = 1.63; P = 3.7e-05) were closely related to prognosis, which reflected the effectiveness of the model indirectly. However, in LUSC (Figure S2A-1, S2A-2, S2B-1 and S2B-2), the single immune gene in the model had no significant connection with prognosis.

Figure 8.

Survival curves analysis based on GEO database. At the bottom of survival curves, 2 lines of figures represent the number of survivors in high- and low-risk groups, which decreases gradually with follow-up time. (A) Survival curve for LUAD based on GEO database; (B) Survival curve for LUSC based on GEO database.

Discussion

In this study, immune genes S100A16, CRABP1, RBP2, FGF2, BTK, SEMA4B, IL11, INHA, ANGPTL4, LGR4, TNFRSF11A, VIPR1, and SHC3 were included in the prognostic model of LUAD, while CXCL5, PLAU, RNASE7, IGKV1-6, SEMA4C, APLN, TSLP, FGFR4, TRAV39, and JUN were included in the model of LUSC.

S100A16, which is associated with poorer survival, is considered to be a prognostic marker for platinum-based adjuvant chemotherapy in LUAD after resection.¹⁶ CRABP1, which is associated with antimicrobial immunity according to the immune gene classification from the Immport database, is closely related to immune cell proliferation and apoptosis via the ERK signaling pathway. A study showed that mRNA and protein levels of CRABP1 were increased in 42% and 50% of NSCLC patients, respectively.¹⁷ To date, no studies have described in detail the action mechanisms of CRABP1 in lung cancer. RBP2 was found to increase the expression of IFN-γ in NK cells by interacting with the P50 and Socs1 promoters as well as to cause the demethylation of H3K4me3 in the Socs1 promoter, further upregulating IFN-γ levels.¹⁸ Moreover, RBP2 decreased the expression of E-cadherin by binding to its promoter, which was induced by TGF-β1, and promoted epithelial-to-mesenchymal transition (EMT) in gastric cancer.¹⁹ An in-depth exploration of RBP2 function in lung cancer is warranted, as RBP2 is involved in cancer progression by not only influencing cancer-related pathways but also by regulating the innate immune response. Serum FGF2 levels are related to poor prognosis in advanced NSCLC patients by promoting angiogenesis.²⁰ It was well-proven that BTK, as a crucial effector to promote B cell development, played an oncogenic role in B cell malignancies²¹; however, recent studies have shown that BTK enhances the functions of tumor suppressors, including p53 and p73, in LUAD (H1299) and colon cancer (HCT116) cell lines.²² It has been reported that SEMA4B inhibits tumor cell growth and metastasis in NSCLC by suppressing the PI3K-Akt signaling pathway.²³ The function of IL-11 in lung cancer has not been extensively studied. Only 1 article has shown that IL-11 promotes tumor cell growth, invasion, and metastasis in LUAD.²⁴ Singh et al. demonstrated that INK, as a good diagnostic and prognostic marker of ovarian cancers, also plays a role in promoting tumor metastasis and angiogenesis in other cancers, which may offer new vascular targets for cancer therapy.²⁵ ANGPTL4 has an effect on enhancing lung cancer cell invasion and migration partially through the ERK signaling pathway.²⁶ LGR4 belongs to a G-protein coupled receptor and is involved in activating the Wnt signaling pathway to influence tumor progression.^27-29 The NCBI gene database (https://www.ncbi.nlm.nih.gov/gene/) shows that TNFRSF11A regulates the interaction between dendritic cells and T cells to change adaptive immune responses, and activates the NF-kappa B and MAPK8/JNK signaling pathways. Despite its possible immense impact on anti-tumor immune responses and cancer development, there are no studies related to TNFRSF11A in LUAD. It has been experimentally shown in a recent study that VIPR1 serves as a tumor suppressor in LUAD, which is consistent with our results.³⁰ Our results showed that SHC3 served as a possible favorable factor for patients with LUAD. The Immport database revealed that SHC3 is associated with the function of NK cells. Meanwhile, the NCBI gene database also showed that SHC3 is present at relatively high levels in normal lung tissue. However, studies on SHC3 in lung cancer are lacking.

Although the single immune gene in the LUSC model had no significant association with prognosis, by retrieving literature in PubMed, we found that CXCL5 could promote tumor progression in colorectal cancer,³¹ prostate cancer,³² osteosarcoma,³³ papillary thyroid carcinoma.³⁴ As a chemokine, CXCL5 recruits neutrophils and promotes angiogenesis. PLAU, which converts plasminogen to plasmin and increases the migration ability of tumor cells, was found to be a positive regulatory factor of colorectal cancer.³⁵ RNASE7 is a possible tumor suppressor in cutaneous squamous cell carcinoma.³⁶ SEMA4C, as the target of cancer-related miRNAs, is down-regulated, which reverses EMT, in lung cancer.^37,38 Serum APLN levels increased significantly in LUSC compared to those in other lung cancer types or control groups.³⁹ The expression of TSLP in breast cancer tissue was higher than that in normal tissues and benign tumors.⁴⁰ JUN is considered an immune-related biomarker in hepatocellular carcinoma, which influences the active states of B cells and T cells.⁴¹ The NCBI gene database shows that IGKV1-6 and TRAV39 are related to the functions of B cells and T cells, respectively. However, at present, a search on the corresponding studies of the 2 genes does not display information in PubMed. Compared to FGF2 with high expression levels in LUAD, FGFR4 was more highly expressed in LUSC. Moreover, the NCBI gene database shows that FGFR4 is the gene with the highest expression in normal lung tissue compared to other tissues. At present, considerable research has been focusing on the development of FGFR inhibitors for cancer treatment.⁴²

The significant expression difference in our results and important roles in other cancers encouraged us to investigate the possible impact of these immune genes on lung cancer, although some genes have not been explored in such studies. In LUSC, the inconsistent results are due to the difference in data sources, as prognostic models were constructed using TCGA data, while survival curves of per immune gene were mainly plotted on the basis of the GEO data. The 2 databases contain different patient populations and use distinct detection methods, which led to validation differences. Moreover, whether distinct immune genes in prognostic models of LUAD and LUSC are differentially expressed in the 2 pathological types remains to be verified.

There have been some studies on the construction of prognostic models for lung cancer. However, these studies used different methodologies. Li et al. constructed an 8-gene prognostic signature for NSCLC. In their study, they selected genes that were not limited to immune genes by univariate Cox regression analysis based on TCGA and GEO databases.⁴³ In another study, the prognostic model was constructed by selecting differentially expressed genes based on the ESTIMATE algorithm-derived immune scores.⁴⁴ Although we used the same database compared with these studies, different analysis methods have generated different models, which will require further experimental validation. Recently, Shi et al. completed a similar study that used the lasso algorithm and multivariate Cox regression analysis to construct a prognostic model of immune genes in LUAD. Their results also demonstrated that ANGPTL4 is a promising immune gene for LUAD prognosis.⁴⁵ Although there was a relatively large number of immune genes in our models, we suggest that it is appropriate to retain these genes in the models as they play an important role in the immune system and cancer progression. Above all, we also note that the genes in the prognostic models regulate innate and adaptive immune responses in various ways, which inspires us to uncover the interaction between these immune genes and tumor-related immune responses. These prognostic models will be better applied in the clinic to evaluate patient prognosis and guide immunotherapy. Kunimasa et al. systematically summarized the lung cancer-related immune responses. The interaction of the immune system and tumor cells can be divided into 3 stages: the elimination, equilibrium, and escape phases.⁴⁶ Thus, we may change tumor progression by interfering with related immune cells and molecules in these stages.

Conclusion

We obtained immune gene prognostic models for LUAD and LUSC based on the TCGA database. Using the GEO and Kaplan–Meier plotter databases, we evaluated the validity of the prognostic models. The risk score based on prognostic models of LUAD and LUSC can serve as an independent prognostic factor, and in LUAD, the risk score was related to the mutation burden. Finally, further investigation of these genes can provide novel insights into the potential association between the immune system and lung cancer.

Abbreviations

adj. P or P. adjust

adjusted P value

AUC

Area under the curve

logFC

Base-2 logarithm of fold change

BP

Biological process

CC

Cell component

EMT

Epithelial-to-mesenchymal transition

FDR or fdr

False discovery rate

FC

Fold change

GEO

Gene Expression Omnibus

GO

Gene ontology

HR

Hazard ratio

KEGG

Kyoto Encyclopedia of Genes and Genome

LUAD

Lung adenocarcinoma

LUSC

Lung squamous cell carcinoma

MF

Molecular function

NK cell

Natural killer cell

NSCLC

Non-small cell lung cancer

ROC curve

Receiver operating characteristic curve

TCGA

The Cancer Genome Atlas