Identification of cancer stem cell-related biomarkers in lung adenocarcinoma by stemness index and weighted correlation network analysis

Abstract

Background

Accounting for tumor heterogeneity, cancer stem cells (CSC) are involved in tumor metastasis, relapse, and drug resistance. Genes regulating CSC characteristics in lung adenocarcinoma (LUAD) were explored and validated in this study.

Methods

The mRNA stemness index (mRNAsi) of more than 500 LUAD cases from The Cancer Genome Atlas database were calculated using a one-class logistic regression machine learning algorithm based on the mRNA expression of pluripotent stem cells and their differentiated progeny. mRNAsi-related key genes were identified by weighted correlation network analysis. The expression levels and prognostic roles of key genes were analyzed in Oncomine, PrognoScan, and Kaplan–Meier plotter databases, and validated using data from our center.

Results

The mRNAsi was significantly higher in LUAD compared with normal lung tissues. LUAD patients of advanced stage exhibited a higher mRNAsi and worse overall survival (OS). Eight key genes were identified: heat shock 70 kDa protein 4 (HSPA4), cell division cycle associated 7 (CDCA7), cell division cycle 20 (CDC20), cyclin-dependent kinase 1 (CDK1), CAP-GLY domain containing linker protein 1 (CLIP1), cyclin B1 (CCNB1), H2A histone family, member X (H2AFX), and Bloom syndrome, RecQ helicase-like (BLM). These genes were differentially expressed in various types of malignancies and validated in the LUAD cases. LUAD patients with low expression of CDC20, CDK1, CCNB1, H2AFX, or BLM had a significantly better OS, whereas OS was reduced for patients with low expression of CLIP1. In addition, the expression of CDCA7 did not significantly impact the OS of LUAD patients. The protein–protein interaction networks evaluated by STRING demonstrated strong relationships between these key genes, which were validated in our cases.

Conclusions

The mRNAsi was significantly higher in LUAD compared with normal samples. Eight mRNAsi-related key genes were associated with prognosis and the cell cycle, and were strongly correlated with each other and differentially expressed in tumor and normal samples. We provide a new strategy for exploring stemness-related genes in LUAD cases.

Introduction

Lung cancer remains the leading cause of cancer-related deaths worldwide; lung adenocarcinoma (LUAD) is the most common subtype (GBD Mortality and Causes of Death Collaborators 2016; Saito et al. 2016). Diagnosed at advanced stages in most cases, LUAD patients often do not have surgical options. Therefore, there are growing concerns regarding radiotherapy, chemotherapy, and immunotherapy, especially in patients of advanced stage. In addition, even small lung cancer may spread locally or systemically (Zhao et al. 2019). A better molecular understanding will improve the existing therapeutic strategies with the aim of achieving long-term control like a chronic disease in LUAD patients.

Cancer stem cells (CSC), a subset of cancer cells, are involved in tumor metastasis, relapse, and drug resistance. Similar to normal stem cells, CSC exhibit self-renewal and differentiation (Singh and Chellappan 2014). CSC were first identified in leukemia and breast cancer (Bonnet and Dick 1997; Singh and Chellappan 2014). They have now been identified in many types of malignancies, including brain, skin, intestinal, and lung tumors. Carney et al. first identified CSC in small-cell lung cancer in 1980 and LUAD in 1982. Some genes that regulate the self-renewal of normal stem cells were found to be oncogenic in many tumors, including notch, wnt, sox2, and shh. Furthermore, notch and sox2 were upregulated in non-small cell lung CSC, suggesting that some pathway may also play a crucial role in the maintenance and self-renewal of CSC, which exhibited a strong association with the initiation and progression of lung cancer.

Malta et al. 2018 reported that the mRNA expression based-stemness index (mRNAsi), calculated using a one-class logistic regression (OCLR) machine learning algorithm based on the mRNA expression of pluripotent stem cells and their differentiated progeny, was an effective tool to evaluate the degree of differentiation of a specific cancer. Weighted correlation network analysis (WGCNA) has been used for the identification of modules of highly correlated genes and their relationships with external sample traits. In this study, we investigated the mRNAsi and its related genes by WGCNA in LUAD from The Cancer Genome Atlas (TCGA) database and validated the results in our own cases.

Materials and methods

Ethics statement

This study was conducted with approval from the Ethics Committee of Zhongshan Hospital, Fudan University, Shanghai, China (Approval no. B2017-042). Written informed consent was obtained from all patients.

Data collection

The RNA-sequencing (RNA-seq) profiles and corresponding clinical phenotypes of 660 human LUAD and normal lung samples were identified from the TCGA database (https://portal.gdc.cancer.gov). We excluded 88 cases because of a lack of survival data.

The mRNAsi calculation and analysis

The “TCGAbiolinks” R package was applied to calculate the mRNAsi of each case based on their mRNA expression levels using the OCLR machine learning algorithm (Colaprico et al. 2016; Malta et al. 2018; Mounir et al. 2019). The associations of mRNAsi with stage, histological subtype, age, sex, and prior malignancy were analyzed using the Wilcoxon rank-sum test because of non-normal distribution. Overall survival (OS) was estimated with the Kaplan–Meier method and compared with the log-rank test for significance.

WGCNA

Differentially expressed genes (DEGs) between LUAD and normal lung samples were screened using the “edgeR” R package with false discovery rate < 0.05 and |log₂ fold change|> 2. Genes were removed from the list if their expression levels were < 1. The “WGCNA” R package was used to perform WGCNA to identify DEG co-expression modules as previously reported (Langfelder and Horvath 2008; Pan et al. 2019).

Identification of significant module and key genes

The association of external traits with gene expression profiles was defined as gene significance (GS). Module membership (MM), highly related to the correlation between the module expression mode and external traits, was determined as the average absolute gene significance in a module. Statistical significance was calculated by the relevant p values. A cutoff (< 0.25) was adopted to merge modules of similar heights. The associations of gene modules with the mRNAsi were then analyzed. After the most significant module was found, MM > 0.8 and GS > 0.6 were used to screen key genes in the module.

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis of the key genes

The involved pathways and biological functions of the key genes were analyzed with GO and KEGG pathway enrichment analysis using the “clusterProfiler” R package.

Expression and prognosis analysis of the key genes

We used the Oncomine database (https://www.oncomine.org) to identify the expression levels of the key genes in various types of cancers (Rhodes et al. 2007). The threshold limits were as follows: p value of 0.001, fold change of 1.5, and gene ranking of all.

Prognosis analysis of LUAD patients was executed by the PrognoScan database (https://www.abren.net/PrognoScan/) (Mizuno et al. 2009) and the Kaplan–Meier plotter database (https://kmplot.com/analysis/) across a large collection of publicly available cancer datasets (Lanczky et al. 2016). The threshold was adjusted to a Cox p value < 0.05.

Interaction and co-expression analysis of the key genes

The protein interaction network was analyzed by STRING (https://www.string-db.org) version 11.0 (Szklarczyk et al. 2019). The “corrplot” R package was used to calculate Pearson correlations between the key genes based on expression data.

Data validation of the key genes

We collected 30 LUAD and 23 corresponding normal lung tissues from the same patients, which were sequenced to obtain mRNA expression profiles. The expression, co-expression, and clinical associations of the key genes were investigated.

Results

The mRNAsi characteristics and its prognostic roles in LUAD

As shown in Fig. 1a, the mRNAsi was significantly higher in LUAD than that in normal lung tissues. However, it was similar between two groups as the aspect of age and prior malignancy (Fig. 1b, c). Interestingly, male LUAD patients were more likely to suffer from higher mRNAsi tumors (Fig. 1d). In terms of tumor stage, there was an increase in mRNAsi in advanced T stage, N stage, M stage, and TNM stage, although no significance was found perhaps due to the limited number of samples (Fig. 1e–h). Unexpectedly, we did not observe any significant difference of mRNAsi in the five subtypes of invasive adenocarcinoma (Fig. 1i). Nevertheless, the mRNAsi was higher in non-mucinous LUAD compared with mucinous adenocarcinoma (Fig. 1j).

Fig. 1.

mRNAsi distribution in a all TCGA lung adenocarcinoma cases, b age, c prior malignancy, d sex, e T, f N, g M, h TNM stage, i five histological subtypes, j mucinous vs. non-mucinous adenocarcinoma

Of all the cases, patients with low mRNAsi had better OS despite no significance, which resembled the findings in each stage, with the exception of stages I and IV (Fig. 2). The lack of significance may be explained by the limited number of cases.

Fig. 2.

Overall survival of low and high mRNAsi in a all cases, b stage I, c stage II, d stage III, and e stage IV

Most significant modules and key genes in LUAD

We screened 4304 DEGs (3323 upregulated and 981 downregulated) between LUAD and normal lung samples (Fig. 3); the DEGs were subjected to WGCNA to identify gene co-expression modules. According to the scale-free topology criterion, the power β = 3 was adopted to determine a scale-free topology index (R²) of 0.93. Eventually, 13 modules with different colors were generated. Meanwhile, the genes in the gray module were set as background genes (Fig. 4a). Then, the relevance of eigengenes in the 14 modules with mRNAsi was evaluated using the module-trait relationships method (Fig. 4b). The yellow module, which contained 355 genes, exhibited the highest negative correlation with mRNAsi. The correlation (r = 0.94; p = 6.2e−167) between MM and GS of each gene in the yellow module was obvious. Thus, the yellow module was selected for subsequent analysis.

Fig. 3.

Differentially expressed genes between lung adenocarcinoma and normal samples

Fig. 4.

Weighted gene co-expression network of lung adenocarcinoma. a Correlation between the gene modules and mRNAsi. The correlation coefficient decreased in size from red to green. The corresponding p value was also annotated. b Scatter plot of module eigengenes in the yellow module

The criteria MM > 0.8 and GS > 0.6 were used to screen key genes in the module. We screened eight genes from the yellow module: heat shock 70 kDa protein 4 (HSPA4), cell division cycle associated 7 (CDCA7), cell division cycle 20 (CDC20), cyclin-dependent kinase 1 (CDK1), CAP-GLY domain containing linker protein 1 (CLIP1), cyclin B1 (CCNB1), H2A histone family, member X (H2AFX), and Bloom syndrome, RecQ helicase-like (BLM). Data from the Oncomine database revealed that the expression levels of the eight genes in tumor and normal samples differed greatly in many types of malignancies (Fig. 5).

Fig. 5.

mRNA expression patterns of key genes in overall cancers in Oncomine database. The number in the colored cell represents the number of analyses meeting these thresholds. The color depth was determined by the gene rank. The red cells indicate that the mRNA levels of target genes are higher in tumor tissues than in normal tissues, while blue cells indicate that the mRNA levels of target genes are lower in tumor tissues than in normal tissues

Enrichment analysis and interactions of the key genes

To explore the biological features and significance of the key genes, the “clusterProfiler” R package was used for GO and KEGG pathway enrichment analyses. As shown in Fig. 6, GO analysis showed that the key genes were enriched in regulation of cell cycle checkpoint, negative regulation of mitotic cell cycle checkpoint, damaged DNA binding, etc. KEGG pathway analysis indicated that the key genes were enriched in cell cycle, oocyte meiosis, etc. (Fig. 6d). The protein–protein interaction networks evaluated by STRING demonstrated a strong relationship between the key genes (Fig. 6e).

Fig. 6.

GO (a–c) and KEGG (d) pathway analysis of eight key genes. e Protein–protein interaction between key genes. The number of the solid line represents the strength of the relationship

Prognostic roles of key genes in LUAD

The PrognoScan and Kaplan–Meier plotter databases were used to analyze the prognostic potential of the eight key genes. However, the Kaplan–Meier plotter database did not include the OS data of CDK1. As shown in Fig. 7, LUAD patients with low expression of CDC20, CDK1, CCNB1, H2AFX, or BLM exhibited significantly better OS. However, OS was worse for patients with low CLIP1. In addition, expression of CDCA7 did not significantly impact the OS of LUAD patients. Unexpectedly, the influence of HSPA4 on OS was contradictory in two databases, which may be due to different probes used to detect its expression.

Fig. 7.

Overall survival of low and high expression of key genes in a–h prognoscan database and i–o Kaplan–Meier plotter database

Validation and analysis of the key genes

The mRNA expression levels were detected and analyzed in 30 LUAD and 23 corresponding normal lung tissues from 21 LUAD patients at our center (Table 1). mRNA expression levels of the key genes, except CLIP1, were higher in tumor samples (Fig. 8a). Since the patients were all alive in the validation cohort, their OS could not be analyzed. However, a strong correlation was found among CDC20, CDK1, CCNB1, and H2AFX, indicating that the strategies employed in this study are feasible for the identification of key genes involved in CSC characteristics (Fig. 8b).

Table 1.

Characteristics of patients with lung adenocarcinoma in the validation cohort

Characteristics	Number	Percent (%)
Age (years)
Mean (range)	57 (33–81)
Sex
Male	6	28.6
Female	15	71.4
Histology
Adenocarcinoma in situ	3	14.3
Minimally invasive adenocarcinoma	5	23.8
Invasive adenocarcinoma	12	57.1
Invasive mucinous adenocarcinoma	1	4.8
TNM stage
0	4	19.0
Ia1	4	19.0
Ia2	7	33.4
Ia3	1	4.8
Ib	4	19.0
IIb	1	4.8

Fig. 8.

a mRNA level of eight key genes in lung adenocarcinoma and normal samples. b Correlation between eight key genes. The correlation coefficient greater than zero means positive correlation, coefficient less than zero means negative correlation

Discussion

In this study, we calculated the mRNAsi of 513 LUAD and 59 normal lung samples based on the mRNA expression profiles from the TCGA database using the OCLR machine learning algorithm. Eight key genes associated with CSC characteristics were identified by WGCNA. The prognostic potential and correlations of the eight key genes were explored with the PrognoScan and Kaplan–Meier plotter databases and STRING, and were also validated in 30 LUAD and 23 corresponding normal lung samples from 21 LUAD patients from our center. Our results may provide a reference for the identification of CSC-related key genes.

Despite significant advances in target therapy and immunotherapy, the survival of LUAD patients remains poor, especially for advanced cases. Drug resistance, loco-regional relapse, and distant metastasis often lead to failure of surgery and other therapies. An improved understanding of the cellular and molecular heterogeneity of LUAD could help with the advent of novel therapeutic strategies.

CSC has been reported to be involved in tumor heterogeneity, drug resistance, recurrence, and metastasis of lung cancer. CSC is thought to be derived from the oncogenic transformation of normal stem cells, which results in dysregulation of their self-renewal. The progeny of different CSC constitutes the different clones in the tumor bulk, which leads to lung cancer heterogeneity. Mutant KRAS and EGFR pathways have been shown to be involved in the transformation of alveolar epithelial cell type II into LUAD CSC during injury (Xu et al. 2012; Desai et al. 2014). However, basal cells are thought to be the origin of squamous cell carcinoma stem cells due to their similarities of molecular markers (Ochieng et al. 2014; Ferone et al. 2016).

Tyrosine kinase inhibitor (TKI) resistance of non-small cell lung cancer (NSCLC) has been reported to be associated with stem-like phenotypes, because some stem cell markers, such as ALDH, POU5F1, and SOX2, were upregulated following the acquisition of such resistance (Li et al. 2017). In addition, the aberrant activation of CSC-involved pathways promoted the epithelial–mesenchymal transition in TKI-resistant NSCLC. Some studies targeting lung CSC and their tumor microenvironment have shown favorable benefits (Heng et al. 2019).

Consistent with our findings, Pan et al. revealed that advanced T stage bladder cancer had a higher corrected mRNAsi, which also had a worse survival. However, the survival differences between low and high mRNAsi groups were not observed in all cases, such as T1 and T2 stage LUAD in the present study. This may be due to the limited sample size. Furthermore, compared with four other types of invasive adenocarcinoma, micropapillary predominant adenocarcinoma had a higher mRNAsi, which has been reported to be associated with a poor prognosis (Travis et al. 2011).

Eight mRNAsi-related key genes were found to be differentially expressed between tumor and normal tissues in various types of malignancies. These eight key genes were highly correlated with one another, and mainly participated in the cell cycle. CDC20 is the target of the spindle assembly checkpoint, which plays an important role in chromosome segregation and mitotic exit. Downregulation of CDC20 in lung cancer inhibited cell proliferation and induced G2/M cell cycle arrest (Kapanidou et al. 2017). Kato et al. reported that high expression of CDC20 was significantly associated with advanced T stage and pleural invasion in NSCLC patients. Interestingly, LUAD patients with high levels of CDC20 had better OS compared with those with low CDC20, which were similar between low and high level in lung squamous cell carcinoma. We also observed that CDC20 was upregulated in LUAD (Kato et al. 2012). It has also been reported that depletion of CDC20 enhanced tumor sensitivity to chemotherapy and radiotherapy (Kapanidou et al. 2017). CDK1 is crucial in the control of cell division (Malumbres 2014). Consistent with our study, Liu et al. revealed that CDK1 and CDC20 are co-expressed in LUAD, and were significantly associated with poor prognosis (Liu et al. 2018). The combination of CCNB1 and CDK1 is essential for mitosis (Gavet and Pines 2010). BLM is a member of the RecQ helicase family and plays a critical role in DNA repair, recombination, and replication. Patients with BLM mutations develop cancer early and tend to get multiple primary tumors. Cancer is the leading cause of death for these patients, often in childhood (German 1997). BLM has also been implicated in the BRCA1 and Fanconi’s anemia pathway (Croteau et al. 2014).

The phosphorylation of histone H2AX is a sensitive marker for DNA double-strand breaks, which initiate genomic instability and can lead to cancer. Mutation and deletion of histone H2AX are frequently detected in malignancies, and contribute to tumor development, progression, and resistance (Bonner et al. 2008). CDCA7, a new member of the cell division cycle-associated family of genes, is a direct target of c-Myc (RC et al. 2005). CDCA7 has been reported to be overexpressed in lymphoid tumors and promoted migration and invasion through cytoskeleton dynamics (Martín-Cortázar et al. 2019). Guiu et al. demonstrated that CDCA7, as an evolutionary conserved Notch target, was involved in hematopoietic stem cell emergence (Guiu et al. 2014). HSPA4, a mammalian nucleotide exchange factor, has been shown to promote migration, invasion, and transformation activity in lung cancer (Wu et al. 2011). Held et al. revealed that HSPA4 is required for normal spermatogenesis (Held et al. 2011). CLIP1, a microtubule binding protein, regulates nucleus positioning and shape (D'Alessandro et al. 2015).

There are some limitations of this study. First, the cases from the TCGA database did not include tumor grade information, which is a direct reflection of stemness. Second, there was a lack of prognosis information in cases from our center, which made it impossible to analyze the prognostic roles of the key genes.

In conclusion, we calculated and analyzed the mRNAsi of 513 LUAD samples from the TCGA database based on their mRNA expression profiles, which was significantly higher in LUAD than in normal lung samples. In addition, we identified and validated eight key genes related to stemness, prognosis, and cell cycle pathways by WGCNA. We provide a new strategy for exploring stemness-related genes in LUAD cases.

Abbreviations

LUAD

Lung adenocarcinoma

CSC

Cancer stem cells

mRNAsi

MRNA expression-based-stemness index

OCLR

One-class logistic regression

WGCNA

Weighted correlation network analysis

TCGA

The Cancer Genome Atlas database

OS

Overall survival

DEGs

Differentially expressed genes

GS

Gene significance

MM

Module membership

GO

Gene Ontology analysis

KEGG

Kyoto Encyclopedia of Genes and Genomes analysis

CDCA7

Heat shock 70 kDa protein 4 (HSPA4), cell division cycle associated 7

CDC20

Cell division cycle 20

CDK1

Cyclin-dependent kinase 1

CLIP1

CAP-GLY domain containing linker protein 1

CCNB1

Cyclin B1

H2AFX

H2A histone family, member X

BLM

Bloom syndrome, RecQ helicase-like

TKI

Tyrosine kinase inhibitor

NSCLC

Resistance of non-small cell lung cancer

Funding

This work was supported by the National Natural Science Foundation of China (Grant No.: 81672268).

Compliance with ethical standards

Conflict of interest

The authors have no conflicts of interest to declare.

Ethics approval

This study was conducted with approval from the Ethics Committee of Zhongshan Hospital, Fudan University, Shanghai, China (Approval No. B2017-042).

Consent to participate

Written informed consent was obtained from all patients.

Consent for publication

Not applicable.

Availability of data and material

The RNA-sequencing (RNA-seq) profiles and corresponding clinical phenotypes of human LUAD and normal lung samples were identified from the TCGA database (https://portal.gdc.cancer.gov).

Code availability

Not applicable.