Integrated analysis of single-cell and bulk RNA-sequencing identifies a signature based on B cell marker genes to predict prognosis and immunotherapy response in lung adenocarcinoma

Abstract

As an essential component of the tumor microenvironment, B cells exist in all stages of tumor and exert important roles in anti-tumor immunity and shaping tumor development. We aimed to explore the expression profile of B cell marker genes and construct a prognostic signature based on these genes in Lung adenocarcinoma (LUAD). A total of 1268 LUAD patients from different cohorts were enrolled in this study. We performed an analysis of single-cell RNA-sequencing (scRNA-seq) data from Gene expression omnibus (GEO) database to identify B cell marker genes in LUAD. TCGA database was used to construct signature, and six cohorts from GEO database were used for validation. We also investigated the association between this signature and immunotherapy response. Based on 258 B cell marker genes identified by scRNA-seq analysis, a nine-gene signature was constructed for prognostic prediction in TCGA dataset, which classified patients into high-risk and low-risk groups according to overall survival. The multivariate analysis demonstrated that the signature was an independent prognostic factor. The signature's predictive power was verified in other six independent cohorts and different clinical subgroups. Analysis of immune profiles showed that high-risk groups presented discriminative immune-cell infiltrations and immune-suppressive states. More importantly, risk scores of the signature were closely correlated with PD-L1, tumor mutation burden, neoantigens, and tumor immune dysfunction and exclusion score. Our study proposed a novel prognostic signature based on B cell marker genes for LUAD patients. The signature could effectively indicate LUAD patients’ survival and serve as a predictor for immunotherapy.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00262-022-03143-2.

Introduction

Lung cancer is the second most common cancer and the leading cause of cancer-related death worldwide [1]. Non-small cell lung cancer (NSCLC) accounts for nearly 85% of lung cancer cases, and Lung adenocarcinoma (LUAD) represents the most common histological subtype of NSCLC [2–4]. In the past decades, molecular-targeted therapy has brought substantial clinical benefits for patients with LUAD [5]. Unfortunately, after receiving targeted treatments, such as EGFR tyrosine kinase inhibitors, most patients eventually become resistant to targeted therapy and their prognosis is still poor [6, 7]. Recently, immune checkpoint blockade therapy has dramatically changed the therapeutic strategy and become a frontline treatment in patients with advanced LUAD, but the overall response rate of Immune checkpoint inhibitors (ICIs) is relatively low, and only a subset of LUAD patients could benefit from ICI treatment [8, 9]. The great challenge of tumor immunotherapy is to identify prognostic biomarkers to determine therapeutic effect and predict prognosis [10].

Accumulating evidence has shown that immune responses within the Tumor microenvironment (TME) are important determinants of tumor behavior, progression, and aggressiveness, as well as the TME exerts a crucial impact on patients’ survival and response to immunotherapy [11–13]. In the TME, tumor-infiltrating B lymphocytes (TIBs) can be observed in all stages of human lung cancer development [14], and their presence differs between stage and histological subtypes [15, 16], suggesting a critical role for B cells during lung tumor progression. TIBs exert anti-tumor immunity by production of tumor-specific antibodies, secretion of interferon-γ, enhancing T cell and nature killer cell responses, and maintaining the structure and function of Tertiary lymphoid structure (TLS), all of which are associated with beneficial outcomes for lung cancer [17–19]. However, as multifaceted effectors, B cells can promote tumor progression by inducing tumor angiogenesis and immunosuppression through production of IL-10 [20, 21]. Moreover, B cells were reported to be correlated with prolonged prognosis, and high B cell infiltration was related to elevated PD-L1 expression in lung cancer patients [22–25]. Given the roles of B cells in anti-tumor immunity, it is necessary to better understand the gene expression profiles of B cells and their associations with prognosis and immunotherapeutic prediction.

Single-cell RNA-sequencing (scRNA-seq) provides a potent approach to explore the mechanisms of heterogeneity and evolution of tumors and pave the way for individualized treatment [26, 27]. Moreover, scRNA-seq analysis of immune cells in the TME contributed to dissect molecular characteristics of immune cells, which provided novel insights for cancer immunity [28, 29]. Establishing gene signatures based on molecular characteristics of immune cells might be an effective way to predict immunotherapeutic effect and prognosis of cancer patients. In this study, we performed integrated analysis of scRNA-seq and bulk RNA-seq of LUAD to identify the marker gene of B cells and construct B cell marker genes signature (BCMGS) for prognosis prediction of LUAD. Then, we validated the prognostic value of BCMGS and further analyzed the relationship between this B cell-based signature and PD-L1 expression, TMB, neoantigens, TIDE score, and landscape of immune cell infiltration to evaluate the association of BCMGS and immunotherapy response in LUAD.

Materials and methods

Data source and acquisition

Single-cell transcriptome file of three LUAD samples of GSE117570 was downloaded from the Gene expression omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/) and used to screen B cell marker genes of LUAD. The Cancer Genome Atlas (TCGA) bulk tumor transcriptome data and clinical information of 500 LUAD patients were downloaded from UCSC Xena (https://xenabrowser.net/) for further survival-related genes screening and signature construction. To validate the prognostic capability of the constructed signature, other six independent cohorts were downloaded from GEO datasets, including GSE11969 (n = 90), GSE31210 (n = 226), GSE37745 (n = 106), GSE50081 (n = 127), GSE30219 (n = 83), and GSE42127 (n = 133). The study utilized publicly available datasets with preexisting ethics approval from original studies. Informed consent was obtained for each participant. The study was conducted in accordance to the Helsinki declaration.

Identification of B cell marker genes by scRNA-seq analysis

We performed scRNA-seq data analysis by using “Seurat” (version 4.0.5), “SingleR” (version 1.6.1) packages [30]. We removed cells with more than 5% of mitochondrial gene. Cells with the number of gene mapped less than 50 and clusters with cell counts less than 3 were also removed. The top 2,000 variably expressed genes were used to perform principal component analysis (PCA). The top 15 principal components were selected for cell clustering analysis using the algorithm of T-distributed stochastic neighbor embedding (t-SNE). Genes that exhibited a |log2 (fold change)|> 1 and adjusted p value < 0.01 were considered as the marker genes.

Construction and validation of prognostic signature based on B cell marker genes

A univariate Cox regression analysis was performed using the R package “survival” (version 3.2–13) (https://CRAN.R-project.org/package=survival) to investigate correlations between the expression levels of B cell marker genes and the overall survival (OS) of patients with LUAD in TCGA dataset. Genes with p-value < 0.05 by univariate Cox regression analysis were identified as prognostic. Next, we used the Least absolute shrinkage and selection operator (LASSO) method for variable selection in a Cox regression model to determine significant prognostic genes, and one Standard error (SE) above the minimum criteria was selected. A stepwise multivariate Cox regression analysis was then used to determine the prognostic values of specific gene signatures. The selected genes formed a risk formula that was determined by a linear combination of the gene expression levels and weighted with the corresponding regression coefficients from the stepwise Cox proportional hazards regression model. The samples were sorted into either a high-risk or a low-risk cohort according to the median cutoff. Then, a Receiver operating characteristic (ROC) curve was generated by the R package “survivalROC” (version 1.0.3) [31]. The Kaplan–Meier survival curve of the risk scores was generated using the R package “survminer” (version 0.4.9) (https://CRAN.R-project.org/package=survminer). The predictive power of this risk model was verified using Kaplan–Meier survival analysis in six independent GEO datasets.

Pathway and function enrichment analysis

Pathway and functional enrichment analysis of Kyoko encyclopedia of genes and genomes (KEGG) and Gene ontology (GO) were conducted by using R packages: “ClusterProfiler” (version 4.0.5), “org.Hs.eg.db” (version 3.13.0), “ggplot2” (version 3.3.5), “enrichplot” (version 1.12.3) [32].

Immune cell infiltration analysis and gene sets variation analysis (GSVA)

CIBERSORT was used to estimate the abundance of immune cell infiltration in different risk groups in this study. We uploaded the FPKM of RNA-seq data of LUAD to the CIBERSORT web portal (http://cibersort.stanford.edu/), and the fractions of 22 immune cell types using 1,000 permutations. The profiling of multiple immune cell types was performed through the leucocyte gene signature matrix, termed LM22, for the CIBERSORT. Using R package of “GSVA” (version 1.40.1) [33], we performed the GSVA analysis to research the relationship between metagenes of inflammatory and immune response and the signature based on B cell marker genes.

The diversity of B cell receptor (BCR) and T cell receptor (TCR) repertoire

The richness and Shannon diversity index were used to characterize the diversity of the BCR and TCR repertoire. The richness measures the number of unique BCRs and TCRs in the sample, while the Shannon diversity index reflects the relative abundance of the different BCRs and TCRs. The richness values and Shannon diversity index valves of BCR and TCR in TCGA LUAD patients were collated from the Pan-Cancer Atlas study of Thorsson et al. [34].

Immunotherapeutic response prediction

The potential immunotherapy response prediction performance of the risk model was estimated with the following biomarkers: PD-L1 protein expression, Tumor mutation burden (TMB), neoantigens, and Tumor immune dysfunction and exclusion (TIDE) score. Tumor mutation data of patients with LUAD were downloaded from TCGA database (https://portal.gdc.cancer.gov/), which were analyzed by R package of “maftools” (version 2.8.05) [35]. Tumor mutational burden (TMB) was defined as the number of somatic, coding, base substitution, and indels per megabase of genome examined, and was calculated as previously described [36]. The neoantigen data of LUAD patients in the TCGA dataset were collected from The cancer immunome atlas (TCIA) (https://tcia.at/home). The Somatic copy number alteration (SCNA) burden was determined by fraction of genome altered. TIDE is a computational method that models two primary mechanisms of tumor immune escape—T cell dysfunction and T cell exclusion, which could be used to predict ICI response in cancer patients [37]. We obtained TIDE scores, T cell dysfunction scores, and T cell exclusion scores from the TIDE web (http://tide.dfci.harvard.edu) after following the instructions on the website to upload gene expression data.

Statistical analysis

Pearson correlation analysis was used to evaluate the correlation between continuous variables. Variables between groups were compared by Wilcoxon t test. Independent prognostic factors were calculated by Cox proportional hazards regression model. P < 0.05 was set as a significant difference in all statistical methods. R software version 4.1.0 (http://www.R-project.org) was used for data analysis and generation of figures.

Results

Identification of B cell marker genes expression profiles

After data processing and screening, we obtained gene expression profiles of 3583 cells from three LUAD samples for subsequent analysis. We performed PCA to reduce the dimensionality by using the 2000 variable genes (Fig. 1a) and identified nine cell clusters using Seurat (Fig. 1b). The annotations of cell identity on each cluster were defined by cross-referencing differentially expressed genes in each cluster with canonical marker genes, and cells in clusters 4 and 6 were classified as B cells (Fig. 1c). We also found that these two clusters had distinct gene expression profiles with genes differentially expressed between the nine clusters (Supplementary Fig. 1a). As a result, we found out 258 B cell marker genes of LUAD. Then we performed GO and KEGG analysis to explore the biological functions of these marker genes. The GO analysis showed that B cell marker genes were mostly related to the biological process of neutrophil activation and neutrophil degranulation (Supplementary Fig. 1b). The KEGG analysis revealed that B cell marker genes were primarily involved in the pathway of phagosome and antigen processing and presentation (Supplementary Fig. 1c).

Fig. 1.

Identification of B cell marker genes by single-cell RNA-sequencing analysis. a PCA plot colored by various samples. b t-SNE plot colored by various cell types. c The cell types identified by marker genes

Construction of prognostic signature based on B cell marker genes

To identify a prognostic signature based on B cell marker genes, we used TCGA cohort as the training set. We first performed a univariate Cox proportional regression analysis and found that 72 B cell marker genes were statistically significantly correlated with OS (P < 0.05) (Supplementary Table 1). Next, a LASSO Cox regression model was used to calculate the most useful prognostic genes, and one SE above the minimum criteria was chosen, resulting in a model with 27 genes: CD69, NFKBID, ID2, LDHA, IL7R, CD9, VAMP8, CTSL, MS4A7, AVPI1, TXN, KLF4, FBP1, SERPINH1, ALOX5AP, SSR4, HERPUD1, UBC, HSP90AA1, PFN1, PPIA, PABPC1, NUCB2, TGIF1, HLA-DMB, FOSL2, HLA-DQA1 (Supplementary Fig. 2a and 2b). To optimize this model to contain only the most predictive genes, a stepwise Cox proportional hazards regression model was used, which identified a final set of nine genes (Supplementary Fig. 2c). Subsequently, a risk score was built: Risk score = (− 0.261 × CD69 expression) + (0.425 × LDHA expression) + (− 0.165 × CD9 expression) + (0.112 × CTSL expression) + (0.312 × KLF4 expression) + (0.350 × SERPINH1 expression) + (0.306 × PPIA expression) + (− 0.265 × NUCB2 expression) + (− 0.106 × HLA-DQA1 expression). The risk score for each patient was calculated using this formula, and patients were divided into high-risk group (n = 250) and low-risk group (n = 250) based on the median cutoff point (cutoff value = 0.943). The distribution of risk scores and survival status of each patient are shown in Fig. 2a. The heatmap exhibited detail expression level of the enrolled nine genes (Fig. 2b).

Fig. 2.

Construction of B cell marker genes signature in TCGA database. a The distribution of risk score and survival status. b A heatmap generated on the basis of identified gene expression. c Kaplan–Meier curves of survival analysis in TCGA LUAD patients based on risk score. d ROC analysis of BCMGS for predicting the risk of death at 2, 3 and 5 years

Kaplan–Meier survival analysis showed that high-risk patients had significantly worse overall survival than low-risk patients (P < 0.001; Fig. 2c). We assessed the prognostic accuracy of the prognostic signature using the time-dependent ROC curves for OS after 2, 3, and 5 years. The AUC values at these times were 0.722, 0.718, and 0.693, respectively (Fig. 2d). Additionally, we explored its ability to predict disease-free survival (DFS) in TCGA LUAD cohort and found that high-risk patients had a significantly shorter DFS than the low-risk group (Supplementary Fig. 3, P < 0.001).

To further explore whether the signature-based risk score was an independent factor in patients with LUAD, univariate and multivariate Cox regression analyses in the TCGA database were conducted. The results of the multivariate Cox regression model confirmed that the risk score was a significant factor (HR:2.120, 95%CI:1.538–2.923, P < 0.001) independent of age, gender, smoking history, and clinical stage (Supplementary Table 2).

Validation of the prognostic value of the BCMGS in independent cohorts

To validate prognostic predictive power of the BCMGS, we first assessed its performance in six independent GEO cohorts. All the demographics of these public GEO cohorts are listed in Table 1. We calculated the risk score for each patient in five independent GEO cohorts using the same formula. The optimal cutoff point for risk score was calculated using the “survminer” package according to the expression value, the survival time, and the survival status, and patients in each cohort were separated into the high-risk group and low-risk group based on the optimal cutoff points.

Table 1.

Clinical characteristics of lung adenocarcinoma from multiple cohorts

Characteristics	TCGA N = 500	GSE11969 N = 90	GSE31210 N = 226	GSE37745 N = 106	GSE50081 N = 127	GSE30219 N = 83	GSE42127 N = 133
Age (year)
Median	66	62	61	64	70	60	66
Range	33–88	32–84	30–76	40–83	40–86	44–84	42–86
Gender
Male	230	47	105	46	65	65	68
Female	270	43	121	60	62	18	65
Smoking
Yes	415	45	111	/	92	/	/
No	71	45	115	/	23	/	/
NA	14	0	0	/	12	/	/
TNM stage
I and II	387	65	226	89	127	83	111
III and IV	105	25	0	17	0	0	21
NA	8	0	0	0	0	0	1
OS status
Alive	318	50	191	29	76	40	90
Death	182	40	35	77	51	43	43

NA, Not available; OS, Overall survival

Kaplan–Meier survival analysis revealed that patients in the low-risk group had significantly better OS than the high-risk group, either in GSE11969 (Fig. 3a, HR: 2.308, 95%CI: 1.201–4.436, P = 0.010), GSE31210 (Fig. 3b, HR: 3.228, 95%CI: 1.624–6.419, P < 0.001), GSE37745 (Fig. 3c, HR: 2.225, 95%CI: 1.322–3.743, P = 0.002), GSE50081 (Fig. 3d, HR: 2.074, 95%CI: 1.189–3.617, P = 0.009), GSE30219 (Fig. 3e, HR: 2.872, 95%CI: 1.543–5.344, P = 0.001), and GSE42127 (Fig. 3f, HR: 3.000, 95%CI: 1.266–7.113, P = 0.008). The ROC curves of risk score in each GEO cohort are shown in Supplementary Fig. 4. Moreover, a meta-analysis was performed to determine the integrated predictive significance of BCMGS in these six cohorts (n = 765). We conducted the meta-analysis by R package “meta” (version 5.1–0) [38], and the fixed effects model was selected based on the assessment of heterogeneity. Our result indicated that BCMGS was a significant risk factor for LUAD patients (HR: 2.488, 95%CI: 1.927–3.212, P < 0.001) (Fig. 3g).

Fig. 3.

Validation of B cell marker genes signature in six independent GEO datasets. a GSE11969 (n = 90). b GSE31210 (n = 226). c GSE37745 (n = 106). d GSE50081 (n = 127). e GSE30219 (n = 83). f GSE42127 (n = 133). g A meta-analysis based on GEO cohorts

Validation of the BCMGS in different clinical subgroups

We further validated the signature in LUAD patients stratified by different clinical characteristics and stages in TCGA training cohort. The predictive ability of the signature was first evaluated in patients with different genders, ages, and smoking histories and stages. Similarly, the predictive power of this risk model was validated in different clinical subgroups. The results showed that all the high-risk groups had significantly worse OS compared with the low-risk groups, across all clinical subgroups (P < 0.05) (Supplementary Fig. 5).

EGFR and KRAS mutations are two of the most frequent oncogene alterations in lung adenocarcinoma. Next, we further explored the performance of the signature in TCGA LUAD patients with or without KRAS/EGFR mutation. Likewise, we found that no matter in EGFR wild-type (WT), EGFR mutation (MUT), KRAS WT, or KRAS MUT subgroup, this risk model exhibited a robustly predictive power (P < 0.05) (Supplementary Fig. 6a-d). Besides, we divided patients into high SCNA and low SCNA subgroups based on the median value of SCNA burden and investigated the ability of the signature to predict the prognosis of patients in each subgroup. Consistent with the expected result, patients with the low-risk score showed an obvious survival advantage in both subgroups (P < 0.05) (Supplementary Fig. 6e and 6f).

Biological pathways related to the BCMGS

The outstanding predictive performance of the signature inspired us to explore the underlying mechanism. We first performed the correlation analysis to determine the genes that strongly correlated with the risk score (Pearson |R|> 0.5, P < 0.001). As shown in Fig. 4a, there were 28 positively related genes and 42 negatively related genes that were screened out. Then, using R package, we performed GO and KEGG analysis for these related genes. The result revealed that these genes were more involved in the biological processes of mitotic division in GO analysis (Fig. 4b). KEGG analysis further exhibited that these genes were more related to cell cycle and glycolysis/gluconeogenesis pathways (Fig. 4c).

Fig. 4.

Functional enrichment analysis of the genes that are closely correlated with the B cell marker genes signature. a Heatmap showed the 70 genes that had the most significant correlations with the signature (Pearson |R|> 0.5, P < 0.05). GO b and KEGG c analysis of the identified genes

Correlation of the BCMGS with immune cell infiltration and lymphocyte receptor repertoire diversity

As B cells played an important role in anti-tumor immune response, we explored the relationship of the BCMGS with immune cell infiltration and lymphocyte receptor diversity in LUAD patients. We first explored the relationship between the signature and immune cell infiltration. CIBERSORT analysis demonstrated that LUAD patients with high-risk score had higher proportion of T cell CD4 memory activated, NK cells resting, macrophages M0, mast cells activated, and neutrophils, but had lower proportion of T cell CD8, T cell CD4 memory resting, monocytes, dendritic cells resting, and mast cells resting (Fig. 5a). Then, the diversity of tumor-infiltrating B cell receptor (BCR) and T cell receptor (TCR) repertoire was evaluated by using the richness and the Shannon diversity index (Fig. 5b–e). We found an increase of the richness and the Shannon diversity index of BCR and TCR repertoire in LUAD patients with low-risk score, which indicated that LUAD patients with low-risk score had higher BCR and TCR repertoire diversity than patients with high-risk score.

Fig. 5.

Correlation of the B cell marker genes signature with immune cell infiltration and lymphocyte receptor repertoire diversity in TCGA cohort. a The comparison of 22 immune cells’ infiltration level between high-risk and low-risk groups. The BCR richness b, TCR richness c, BCR Shannon diversity index (d) and TCR Shannon diversity index e were compared between high-risk and low-risk groups

Inflammatory and immune profiles of the BCMGS

We analyzed the associations between the signature and seven clusters of metagenes (HCK, IgG, interferon, LCK, MHC-I, MHC-II, and STAT1), representing different types of inflammatory and immune responses. The expression pattern of these metagenes in the TCGA dataset is presented in Fig. 6a. Then, we performed Gene sets variation analysis (GSVA) to calculate the expression of seven gene clusters and investigate the correlation between the signature and each cluster of metagenes. The result indicated that the risk score had negative correlation with HCK, IgG, LCK, MHC-I, and MHC-II (Fig. 6b).

Fig. 6.

Inflammatory and immunologic profiles of the B cell marker gene signature in TCGA cohort. a Heatmap showed the relationship between risk score and inflammatory metagenes. b Correlogram was generated based on Pearson R-value between risk score and metagenes

Relationship between the signature and immunotherapy response

TIBs exert anti-tumor immunity in lung cancer, and the B cell-based immunotherapeutic strategies may provide a possible new option to improve cancer therapy [39]. Therefore, we further evaluated the association of the BCMGS and immunotherapy response by analyzing the correlation of the signature and widely recognized immunotherapy biomarkers in TCGA cohort. First, the protein expression of PD-L1, Tumor mutation burden (TMB), and number of neoantigens were compared between high-risk and low-risk groups. The results showed that high-risk patients harbored a significantly higher level of PD-L1 protein expression, TMB and neoantigens than low risk patients (Fig. 7a–c). Next, to get a comprehensive evaluation, we further enrolled the TIDE score, the T cell dysfunction score, and the T cell exclusion score, which are more accurate biomarkers into our analysis. As expected, high-risk patients were characterized by a significantly lower TIDE score, lower T cell dysfunction scores, and higher T cell exclusion scores (Fig. 7d–f), which indicated that high-risk patients tended to be more sensitive to immunotherapy response. Collectively, these results revealed that patients with high-risk score are more likely to benefit from immunotherapy, and the BCMGS may be a useful biomarker to identify LUAD patients who may benefit from immunotherapy.

Fig. 7.

The comparison of immunotherapy response biomarkers between high-risk and low-risk groups in TCGA cohort. The comparison of PD-L1 protein expression (a), Tumor mutation burden (TMB) (b), number of neoantigens (c), Tumor immune dysfunction and exclusion (TIDE) score (d), T cell dysfunction score (e) and T cell exclusion score (f) between high-risk and the low-risk groups

Discussion

With the advancement of cancer immunotherapy, increasing numbers of predictive biomarkers of immunotherapy response have been continually identified [40]. The impact of the TME on efficacy of cancer immunotherapy has been intensively studied, and much more attention has been paid to TME-related biomarkers. However, reliable biomarkers—based on the TME of tumorigenesis for the immunotherapy response and prognosis in LUAD—are still very rare. The development of scRNA-seq technologies provided us a way to dissect molecular characteristics of tumor-infiltrating immune cells in the TME. In the present study, we performed scRNA-seq analysis of LUAD and found out B cell marker genes in the LUAD tissue. Based on the B cell marker genes, we further constructed a novel prognostic prediction signature (BCMGS) for LUAD patients in the TCGA database, and the signature was identified as an independent risk factor for LUAD patients. We then validated the prognostic performance of BCMGS in six independent cohorts from GEO dataset. Analysis of biological pathway and immune profiles revealed that the signature-related genes were prominently enriched in cell cycle pathway, and patients with a high-risk score were characterized by distinctive immune cell proportions and immune-suppressive states. Additionally, we found that high-risk scores of BCMGS were positively related to different immunotherapy biomarkers, which indicates that immune checkpoint blockade therapy may be more appropriate for high-risk patients. According to our findings, BCMGS might be a useful tool to predict prognosis and immunotherapeutic effect in LUAD patients after further validation.

In this study, BCMGS was composed of nine B cell marker genes, which were CD69, LDHA, CD9, CTSL, KLF4, SERPINH1, PPIA, NUCB2 and HLA-DQA1, respectively. In the signature model, LDHA, CTSL, KLF4, SERPINH1, PPIA were the unfavorable genes for the outcome, whereas other genes presented protective function on the prognosis of LUAD patients. CD69 is a type-II C-lectin transmembrane receptor, which is a classical early marker of lymphocyte activation and involves in modulating immune responses [41]. CD69 was reported to be downregulated in LUAD patients’ cancer tissue compared with the normal tissue [42], and CD69 expression was inversely correlated with survival in melanoma [43]. LDHA, a crucial enzyme of energy metabolism, has been proved to be significantly upregulated in LUAD, and increased LDHA expression was an independent prognostic indicator for overall survival and recurrence-free survival in LUAD [44]. Previous studies showed that LDHA could promote tumor cells’ proliferation, invasion, migration, and metastasis, and might be a potential therapeutic target [45, 46]. CD9, a member of tetraspanin family, which is widely expressed on the surface of many types of malignant tumor cells, involves in several physiological and pathological processes, such as cellular adhesion, growth, differentiation, and tumor invasion [47]. Preclinical studies of cancer cell lines or mouse tumor xenograft models indicated that CD9 may serve as a potential therapeutic target for cancer [48]. CTSL is a ubiquitously expressed lysosomal endopeptidase that is primarily implicated in terminal degradation of intracellular and endocytosed proteins [49]. CTSL has been shown to proteolytically degrade the extracellular domain of E-cadherin and abolish its adhesive property, which play a key role in cancer metastatic aggressiveness [50]. KLF4, a DNA-binding transcriptional regulator, regulates a wide range of biologic activities: cell proliferation and differentiation, as well as the self-renewal of stem cells [51]. As a bifunctional transcription factor, KLF4 activates or inhibits transcription according to different target genes and utilizing different mechanisms. KLF4 can play an oncogenic or anti-cancer role, depending on the type of cancer involved [52]. Additionally, it is reported that KLF4 exerts immunosuppressive function through regulating macrophage polarization [53]. SERPINH1, also called heat shock protein 47 (HSP47), is a collagen-specific molecular chaperone, which participates in numerous steps of collagen synthesis. The expression levels of SERPINH1 can be regulated by microRNA (miR)-29, and the elevated expression of SERPINH1 is frequently found in a variety of cancers [54]. SERPINH1 can promote tumor growth and invasion, probably through regulation of the extracellular matrix (ECM) network, and may be a possible biomarker and therapeutic target [55]. PPIA catalyzes the cis–trans isomerization of prolyl acyl peptide bonds in oligopeptides and accelerates protein folding. PPIA has been reported to play a role in cyclosporin A-mediated immunosuppression and involve in the immune microenvironment-associated pathogenesis of lung adenocarcinoma or suggested to be significant predictors of recurrence-free or overall survival [56, 57]. NUCB2, located on chromosome 11, encodes a neuropeptide that serves important roles in regulating food intake and energy homeostasis [58]. A study confirmed that NUCB2 was a downstream target of miR-335-5p and that high NUCB2 expression was detected in LUAD and functional assay confirmed that NUCB2 played the oncogenic role in LUAD [59]. HLA-DQA1 is one of the MHC Class II family members and locates on chromosome 6p21. A major function of MHC Class II molecules in the immune system is to present antigens, and aberrant expression of MHC Class II may result in insufficient immune response or autoimmunity reaction, which may result in lots of diseases including cancers [60, 61]. Signature genes identified in this study could provide underlying targets for experimental design in the laboratory to elucidate molecular mechanisms in LUAD.

The predictive power of the BCMGS was validated in six independent cohorts and different clinical subgroups. Strikingly, our signature exhibited statistically significant OS stratification for LUAD patients in all independent cohorts and subgroups, and we further confirmed the robust performance of the signature by meta-analysis. The strong predictive capability of the BCMGS prompted us to explore the potential underlying mechanism. We determined 70 genes strongly correlated with risk score by correlation analysis and found that these correlated genes were mainly enriched in biological process of cell division and pathway of cell cycle through GO and KEGG analysis. It is well acknowledged that dysregulation of cell cycle can promote cancer proliferation and progression [62], which may explain the significant difference of prognosis between high-risk and low-risk groups. In addition, considering the impact of tumor microenvironment on the prognosis of tumor [63], we investigated the discrepancy of immune cell infiltration in high-risk and low-risk LUAD patients. The results revealed that high-risk patients had lower proportion of CD8 + T cells, monocytes, and dendritic cells. Moreover, we evaluated the diversity of the BCR and TCR repertoire, which could reflect the functionality of the immune system, including the spectrum of antigen recognition and the quantity of available T cell responding [64]. We found that high-risk LUAD patients were associated with the lower level of BCR and TCR repertoire diversity, which suggested an immunosuppressive TME in high-risk LUAD patients. Consistently, inflammatory metagene analysis further confirmed that high-risk patients were under an immune-suppressive state with a low-level function of T cell, monocyte, and antigen-presenting (LCK, HCK and MHC-I/II). All our findings indicated that the potential mechanism of the BCMGS’ predictive power may lie in dysregulation of cell cycle and TME heterogeneity.

Currently, PD-L1 expression, TMB, and neoantigens are well recognized to be predictive biomarkers of response to PD-1/PD-L1 inhibitors [65]. TIDE is a newly discovered immunotherapy predictor and has been proven to have better predictive performance than other biomarkers or indicators [37]. To prove that this signature can be a biomarker of immunotherapy response, we explored the relationship between this signature and the biomarkers mentioned above. We found that BCMGS high-risk patients had significantly higher levels of PD-L1 protein expression, TMB, neoantigens, and T cell exclusion scores, as well as lower level of TIDE scores and T cell dysfunction scores, which suggested that tumors with high-risk score had high immunogenicity to activate the immune cells to recognize tumors, but high PD-L1 expression induced the suppression of anti-tumor immunity in high-risk patients. These findings were consistent with the results of seven clusters of metagenes and lymphocyte receptor repertoire diversity that high-risk patients were under an immunosuppressive state—risk score had negative correlation with LCK, IgG, MHC-I/II, TCR richness and BCR richness. Hence, high-risk patients were more likely to benefit from anti-PD-1/PD-L1 immunotherapy that can relieve immune cells from the suppression in the TME. With further validation, BCMGS might act as reliable biomarker for predicting immunotherapy response.

The present study has several limitations. First, this signature was constructed based on public datasets. The predictive capability needs further verification in large-scale prospective clinical studies. Second, patients treated with immunotherapy were not examined in this study, so the ability of the BCMGS in predicting the response to immunotherapy was evaluated indirectly. Lastly, this study does not include any in vitro or in vivo experiments to further explore the potential molecular mechanism of the BCMGS in predicting prognosis and immunotherapy response.

Conclusively, we developed and validated a novel prognostic signature comprising of nine B cell marker genes by integrated analysis of single-cell and bulk RNA-sequencing, which could serve as a potent prognostic biomarker and might predict patients' response to immunotherapy in LUAD. Our study provides a new insight into the role of immune cell marker genes in prognosis and immunotherapy response of LUAD patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

SG and JH supervised the project and designed this study. ZQ organized the public data and prepared all the figures and tables. XW conducted the data analysis. PS and WL drafted the manuscript. JY revised the manuscript. All the authors reviewed and approved the final manuscript.

Funding

The study was supported by National Key R&D Program of China (2020AAA0109500); R&D Program of Beijing Municipal Education commission (KJZD20191002302); Beijing Municipal Science & Technology Commission (Z191100006619119); Beijing Hope Run Special Fund of Cancer Foundation of China (LC2019A11); Beijing Nova Program of Science and Technology (Z191100001119095); and The Institutional Fundamental Research Funds (2018PT32033).

Declarations

Conflict of interest

The authors declare no competing interests.

Informed consent

All the analyzed data used in our research were collected from a public database, such as TCGA and GEO; therefore, informed consent was not required for this analysis.

Ethical approval

Since this was a retrospective study and all the data used in our study were obtained from public databases TCGA and GEO, therefore, ethical approval was not required.