Identification of cancer stemness and M2 macrophage-associated biomarkers in lung adenocarcinoma

Abstract

Objective

Cancer stemness and M2 macrophages are intimately linked to the prognosis of lung adenocarcinoma (LUAD). For this reason, this investigation sought to identify the key genes relevant to cancer stemness and M2 macrophages, explore the relationship between these genes and clinical characteristics, and determine the potential mechanism.

Methods

LUAD transcriptomic data was analyzed from The Cancer Genome Atlas (TCGA) as well as the Gene Expression Omnibus databases. Differential expression analysis was performed to discern abnormally expressed genes between LUAD and control samples in TCGA cohort. The Cell type Identification by Estimating Relative Subsets of RNA Transcripts (CIBERSORT) algorithm was applied to determine varyingly infiltrated immune cells in LUAD compared with the control samples in TCGA cohort. Weighted correlation network analysis (WGCNA) was performed to identify genes associated with mRNA expression-based stemness index (mRNAsi) and M2 macrophages. Least absolute shrinkage and selection operator (LASSO), RandomForest (RF) and support vector machine-recursive feature elimination (SVM-RFE) machine learning methods were conducted to detect gene signatures. Global survival evaluation (Kaplan-Meier curve) was applied to investigate the relationship between gene signatures and the survival time of LUAD patients. Receiver operating characteristic (ROC) curves were produced to define biomarkers relevant to diagnosis. Gene Set Enrichment Analysis (GSEA) was performed to probe the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways related to diagnostic biomarkers. The public single-cell dataset of LUAD (GSE123902) was used to investigate the expression differences of diagnostic biomarkers in various cell types in LUAD. Real-time quantitative PCR (qRT-PCR) was performed to confirm key genes in lung adenocarcinoma cells.

Results

A total of 5,410 differentialy expressed genes (DEGs) as well as 15 differentially infiltrated immune cells were identified between LUAD and control sepcimens in TCGA cohort. Thirty-seven DEGs were associated with both M2 macrophages and mRNAsi according to the WGCNA analysis. Sixteen common gene signatures were obtained using three diverse algorithms. CBFA2T3, DENND3 and FCAMR were correlated to overall and disease-free survival of LUAD patients. ROC curves revealed that CBFA2T3 and DENND3 expression accurately classified LUAD and control samples. The results of single cell related analysis showed that two diagnostic biomarkers expressions were differed between the different tissue sources in M2-like macrophages. QRT-PCR demonstrated the mRNA expressions of CBFA2T3 and DENND3 were upregulated in lung adenocarcinoma cells A549 and H2122.

Conclusion

Our study identified CBFA2T3 and DENND3 as key genes associated with mRNAsi and M2 macrophages in LUAD and investigated the potential molecular mechanisms underlying this relationship.

Abbreviation

NSCLC

non–small cell lung cancer

LUAD

lung adenocarcinoma

CSC

cancer stem cell

TCGA

The Cancer Genome Atlas

GEO

Gene Expression Omnibus

DEGs

differentially expressed genes

mRNAsi

mRNA expression-based stemness index

EREG-mRNAsi

epigenetically regulated mRNAsi

CIBERSORT

Cell type Identification By Estimating Relative Subsets Of RNA Transcripts

WGCNA

Weighted correlation network analysis

LASSO

Least absolute shrinkage and selection operator

RF

RandomForest

SVM-RFE

Support vector machine-recursive feature elimination

ROC

receiver operating characteristic

AUC

area under the ROC curve

KEGG

Kyoto Encyclopedia of Genes and Genomes

UMAP

uniform manifold approximation and projection

t-SNE

t-distributed stochastic neighbor embedding

1. Introduction

Lung cancer has the greatest incidence and fatality rates of all cancers, endangering human health in a serious way [1]. The prognosis for lung cancer is poor and the mortality rate is high because lung cancer is prone to metastasis, and therefore, most lung cancer patients do not have the option for radical resection when they are diagnosed [2]. About 80–85% of all lung cancer pathology types are caused by non–small cell lung cancer (NSCLC), with lung adenocarcinoma (LUAD) being the majority of NSCLC. Therefore, LUAD biomarkers, therapeutic targets and new treatments are still widely needed.

The tumor microenvironment has been the subject of considerable investigation in recent years as evidence of its significance in tumor growth has emerged. The circumstances for tumour progression are created by the tumor microenvironment [3]. Macrophages are found in high frequency in tumor tissue and are the richest type of inflammatory cells in the surrounding environments around the tumor [4]. The tumor microenvironment is linked to many processes, such as tumor growth, invasion, and metastasis, and is tightly relevant to treatment resistance and poor prognosis of various solid tumors, including NSCLC, ovarian cancer, breast cancer, and gastric cancer [[5], [6], [7], [8]]. According to Wei K et al. [9], LUAD metastases and M2 macrophage infiltration are positively correlated. By combining M2 macrophage related genes and WGCNA, Xu C et al. [10] further investigated important genes connected to LUAD prognosis. Based on WGCNA, Zhang Y et al. [11] investigated the expression of stem cell-related genes in lung adenocarcinoma (LUAD), and they preliminary assessed the relationship between gene expression and patient survival using the K-M curve. Between high and low mRNAsi subtypes in LUAD, Li N et al. [12] assessed the predictive significance of differentially expressed immune-related genes (DEIRGs). The probable processes were subsequently investigated by Chen M et al. [13] by focusing on pharmacological targets, gene alterations, and additional common genes associated to immunology and stemness.

Moreover, it has been published in the literature that M2 macrophages are closely related to tumor stem cells. For example, in glioblastoma, M2 macrophages can secrete PTN protein in large quantities, which acts on the cancer stem cell (CSC) receptor PTPRZ1 on glioblastoma [14]. PTPRZ1 activation leads to a series of signaling pathways, generates and maintains CSCs, promotes tumor growth and progression, and leads to increased mortality in patients. In breast cancer, binding of LSECtin on the surface of macrophages to the cancer cell receptor BTN3A3 subsequently activates downstream signaling, promotes breast cancer development and increases the stemness of cancer cells [15]. At present, the regulatory role and link between M2 macrophages and CSCs has not yet been clearly defined in LUAD. Therefore, this study aimed to explore the biomarkers related to the immune microenvironment and tumor cell stemness in LUAD progression for the potential regulatory role between M2 macrophages and CSCs based on the public datasets. Various bioinformatics methods as well as gene expression confirmation in vitro were conducted, such as receiver operating characteristic (ROC), survival curves, and functional enrichment analysis. In Fig. 1, the experimental design is displayed.

Fig. 1.

Experimental design in the current study.

2. Materials and methods

2.1. Data origin and collection

LUAD RNA-Seq data were analyzed from The Cancer Genome Atlas (TCGA) (http://cancergenome.nih.gov) and the Gene Expression Omnibus (GEO) (http://www.ncbi.nlm.nih.gov/geo/) databases. Totally, 515 LUAD and 59 adjacent control specimens in the TCGA were enrolled into picking differentially expressed genes (DEGs). The clinical record and profile of the LUAD patients in TCGA were downloaded from cBioPortal (http://cbioportal.org). The GSE13213 database containing 117 LUAD samples was incorporated to verify the relevance between biomarkers and overall survival (OS). LUAD (n = 3) and adjacent control (n = 3) samples in the GSE121397 were deployed to define the diagnostic effect of biomarkers. The results of the mRNA expression-based stemness index (mRNAsi) as well as the epigenetically regulated mRNAsi (EREG-mRNAsi) of TCGA-LUAD were extracted from a precedent investigation. Moreover, GSE123902 datasets was downloaded to investigate the expressions of the biomarkers in single cell RNA-seq data from LUAD samples, which included seven primary LUAD tissue samples and four healthy lung tissue samples. The clinical information of tumor and normal samples in mentioned datasets involved in the study see Table S1.

2.2. Identification of DEGs in TCGA-LUAD

DEGs between 515 LUAD and 59 adjacent control specimens were detected via the DESeq2 R package (criteria: |log₂FC| > 1 and p < 0.05) [16]. Volcano plots showing the distribution of DEGs were3 plotted using the ggplot2 (R package) [17]. In the meantime, the pheatmap (R package) was applied to plot the heatmap of the expressions of DEGs [18].

2.3. Immune infiltration in TCGA-LUAD

The Cell type Identification By Estimating Relative Subsets Of RNA Transcripts (CIBERSORT) algorithm was conducted to detect the infiltrations of 22 immune cells in TCGA-LUAD (LUAD vs control specimens) [19], which were shown as the coxcomb diagram. The differential contents of 20 immune cells (The infiltrations of T cells CD4 naive and T cells gamma delta were minimal or even absent, so as to be deleted) in LUAD or adjacent control specimens were investigated and displayed as the violin plot.

2.4. Develop the network relying on the soft threshold

A specimen clustering graph was created to identify outliers. After that, a weighted correlation network analysis (WGCNA) based on the tumor samples in TCGA-LUAD with p < 0.05 calculated with CIBERSORT was run depending on DEGs and specimen traits (the variously infiltrated immune cells detected by CIBERSORT). The pickSoftThreshold function was deployed to compute β from 1 to 20 for the purpose of defining the optimal soft threshold to ensure a scale-free network. Next, the pairwise gene correlation matrix was transformed into a weighted matrix based on the optimal soft threshold and further to produce an adjacency matrix.A topological overlap matrix (TOM) representing the overlap in shared neighbors, was calculated using the adjacency matrix. Additionally, the gene dendrogram as well as module color were determined according to the TOM-based dissimilarity measure with a minimum size (gene number) of 100. Furthermore, we segmented the primary module by using Dynamic Tree Cut, and combined similar modules. Following, the correlation coefficient (pearson method) between the modules and specimen traits were computed to mine the remarkable modules linked to M2 macrophages. Similarly, WGCNA was run to search for the significant modules relevant to mRNAsi. Genes in the significant modules linked to M2 macrophages and mRNAsi were combined for the following study.

2.5. Identification of key genes in LUAD

To identify the key M2 macrophages-mRNAsi-related biomarkers that were associated with LUAD survival, LUAD patient samples from TCGA were separated into training and validation sets randomly (with a ratio of 7:3) to construct a gene signature. Least absolute shrinkage and selection operator (LASSO) algorithm was enrolled into screening genes with the parameter as follow: family = “binomial”, type.measure = “class” via the glmnet R package [20] and the predictive accuracy of LASSO model was checked via ROC curve in both the training and validation datasets. The randomForest R package was then applied in the training dataset to perform RandomForest (RF) analysis to screen gene signature [21]. The importance of the gene signatures was ranked according to the Mean Decrease Accuracy. An ROC curve was plotted for the validation set to check the precision of the RF model. Support vector machine-recursive feature elimination (SVM-RFE) was then run via the e1071 (R package) [22], which identified the optimal variables by removing SVM-generated eigenvectors combined with a 5‐fold cross‐validation. Key genes in LUAD were defined by intersecting gene signature extracted by above three distinctive algorithms.

2.6. The clinical significance of key genes in LUAD

LUAD cases in TCGA were categorized into two expression subgroups (high- and low-expression) relying on the expression of key genes. K-M curve as well as log-rank tests were performed to compare the OS between two expression subgroups, and to screen key genes associated with OS. Furthermore, the relationship between key genes and OS was validated in the GSE13213 dataset. Disease-free survival (DFS) was also compared between two expression subgroupss. Thereafter, the clinical significance of key genes associated with both OS and DFS in distinguishing LUAD and normal specimens were evaluated by ROC curves. Specifically, key genes with an area under the ROC curve (AUC) larger than 0.7 were defined as candidate diagnostic biomarkers. Finally, the relationship between biomarkers relevant to diagnosis and clinical characteristics, including sex, age, Tumor stage, Metastasis stage, Nodal stage, cancer stage, smoking and race were investigated.

2.7. Functional enrichment of diagnostic biomarkers

In order to further explore the biological pathways related to the diagnostic biomarkers, the correlation between diagnostic biomarkers and all other genes in 574 samples (59 normal samples and 515 tumor samples) were calculated, then all sorted genes that were ranked based on the correlation coefficients from high to low were used as the gene set to be tested. The “C2: KEGG subset” was used as a reference gene set to detect the enrichment of the KEGG signaling pathways in the tested gene set. Gene Set Enrichment Analysis (GSEA) on diagnostic biomarkers was incorporated using clusterProfiler and org.Hs.eg.db packages [23,24]. The threshold was set to |NES| > 1, p.adjust <0.05.

2.8. Processing and analysis of single-cell transcriptomic profiles in LUAD

Single-cell analysis using the R package “Seurat” was performed to assess the expressions of the diagnostic biomarkers in single cell RNA-seq dataset from LUAD samples (GSE123902) [25]. The parameter of “min.cells = 3,min.features = 200” was firstly set in the “CreateSeuratObject” functions. According to the recommendation of the 10x Genomics websites, the “scDblFinder” were performed for doublet detection [26], where the multiplet rate for cell concentration greater than 10,000 was set to 8%, and a multiplet rate of 5% was set for that less than 10,000. Following removing the data from low quality cells (retained cells that showed <20% mitochondrial genome reads), the raw counts were normalized through the LogNormalize method, and highly variable genes were identified using the FindVariableFeatures function.The transcriptome data was integrated using the “IntegrateData” function, the principal component analysis (PCA) was conducted and examined via a ElbowPlot. Next, cells were clustered with the FindNeighbor and FindClusters function and the cell clusters were annotated by the CellMarker database. Differences in each cell type content between the normal and LUAD samples were compared, and the expressions of key genes in different cell types were explored using the Analysis of variance (ANOVA) test. Moreover, considering the important significance of M2 macrophages, suppressor myeloid cells were further extracted based on the annotation results to check the expression levels of diagnostic biomarkers using Wilcoxon test.

2.9. Cell cultures

Three distinctive cell lines (A549, H2122 and BEAS-2B) were sourced from Shanghai Yuchun Biotechnology Co., Ltd. BEAS-2B cells were cultivated in Dulbecco's Modified Eagle's Medium (Cat #12800-017, Gibco, USA) comprised of 10% fetal bovine serum (Cat #10270-106, Gibco, USA) and 1% antibiotic-antimycotic reagent (Cat #15140-122, Gibco, USA). A549 and H2122 cells were cultivated in RPMI- 1640 medium (Cat # 11875-101, Gibco, USA). The cells were grown in a 5% CO2 incubator at 37 °C.

2.10. Real-time quantitative PCR

For qRT-PCR, cells were collected and total RNA was isolated via Trizol reagent (Cat #15596026,Invitrogen, USA). After that, cDNA was generated by using a PrimeScript reverse transcriptase (RT) reagent kit (Cat #G492,ABM, Canada). Gene expression was detected with SYBR Green PCR mix (Cat #G891,ABM, Canada) on an ABI 7500 system (ABI, Foster city, CA, USA) according to manufacturers’ instructions. Actin was considered as an internal control for mRNA normalization. The primer sequence is listed in Table 1.

Table 1.

Primer sequence of this study.

Primer name	Sequence (5′-3′)
H-DENND3-F	GAGTGCGGTACTGGCGG
H-DENND3-R	CTGCTCGAGACTTCGGAGA
H-FAM155B–F	TGCCCCTTTATACTCCCCGA
H-FAM155B–R	GGGCCTCAGACTCCTTGAAC
H-ITM2A-F	GTGACTGAGGAGGCTGACAT
H-ITM2A-R	AGCAACAAGTCCAGGTAAGCA
H-MAP3K8–F	AAACATGAGGCCCTGAACCC
H-MAP3K8-R	GATTCCTCGGTGCTTCCTGT
H-OTX1-F	ACCCCCATACGGCATGAAC
H-OTX1-R	GACTCCGGCAGGTTGATCTT
H-SCARF1-F	CCCAAAGGGCAGCACGTC
H-SCARF1-R	ACAGATGGGGATGGTGCATTC
H-ZMYND15-F	AGCGACGAGCAGCATTTTAC
H-ZMYND15-R	CCCCTTTCTCCTTAGTGCCAG
H-CBFA2T3-F	CCACTTCCGAGATGCCTACC
H-CBFA2T3-R	CCACTCTTCTGCCCACTCAC
H-FCAMR-F	GCTCATTGCTGGCACTTTACC
H-FCAMR-R	TTGTGGAAGGGCGAAAGAAGA
H-ISL2-F	CCTTTTCTGGGTGCTATGGGT
H-ISL2-R	CGACACCCGCAGGATAAACT
H-MC4R–F	CGTGGGATGCACACTTCTCT
H-MC4R–R	CTCGTAGCACCCTCCATCAG
H-β-actin-F	CATGTACGTTGCTATCCAGGC
H-β-actin-R	CTCCTTAATGTCACGCACGAT

2.11. Statistical analysis

Statistical data were shown as mean ± standard deviation using GraphPad 8.0 (La Jolla, CA, USA). The distinctions among three groups were computed using a one-way analysis of variance followed by a Turkey test. The standard for significant differences were *P < 0.05, **P < 0.01, and ***P < 0.001.

3. Results

3.1. Appraisal of DEGs and differentially infiltrated immune cells in LUAD

A total of 5,410 DEGs were uncovered between LUAD and adjacent control cases in TCGA (Table S1), including 1,968 down-regulated and 3,442 up-regulated and genes in LUAD relative to control samples (Fig. S1A). The expression patterns of DEGs in every specimens were illustrated in the heatmap (Fig. S1B). The 5,410 DEGs were used in the following WGCNA analysis.

Immune infiltration of 22 immune cells in 219 LUAD and 25 control samples was shown in the coxcomb diagram (Fig. 2A). We noted that the proportions of M2 macrophages was in tumor groups higher than in controls. Further, the contribution ratio of the remaining 20 immune cells were compared between LUAD and control samples (Naive CD4 T cells were detected in only one control sample, and no gamma delta T cells were detected in either LUAD or control samples). It is found that the infiltration of 15 out of 20 immune cells, such as activated memory CD4 T cells, M0 macrophages, and M2 macrophages, were significantly different between LUAD and control samples (Fig. 2B). Even the abundance of resting memory CD4 T cells were highest in both LUAD and control samples, there was no significant difference in infiltration of which between the two groups.

Fig. 2.

Immune cell differential analysis. (A) Proportion of 22 types of infiltrating immune cells (TIICs) (cancer and control samples). (B) The variation in the proportion of 21 TIICs in cancer and control samples.

3.2. Identifying genes associated with M2 macrophages and cancer cell stemness in LUAD

Considering the significance of changes in M2 macrophage infiltration levels in the tumor group, we adopted WGCNA analysis using 219 LUAD samples in TCGA-LUAD with p < 0.05 after filtering by CIBERSORT and 5,410 DEGs to screen genes relevant to M2 macrophages. Depending on the sample clustering results, no outlier sample were found. By using the pickSoftThreshold function, we identified the optimal soft threshold power. When the power = 3, the R² was ∼0.85 (Fig. 3A). By using the Dynamic Tree Cut package, we obtained 10 modules (Fig. 3B). On the basis of module-trait correlation shown in Fig. 3C, we noted that green and magenta modules had the highest negative and positive correlations with M2 macrophages, respectively. Thus, 375 genes (green module) and 117 genes (magenta module) were considered M2 macrophage-related genes. Finally, we drew scatter plots to visualize the correlation between modules and genes, and the correlation between and traits and genes. As shown in Fig. 3D, only the correlation in the macrophage M2 phenotype was shown.

Fig. 3.

Identification of modules associated with M2 macrophages and tumor cell stemness. (A) Determination of soft thresholds. (B) The correlation of modules with grouped traits. (C) Heatmap for module-trait correlations showed that green and magenta modules were associated with M2 macrophages. Higher positive and negative correlations were observed. (D) Correlations with the M2 macrophage phenotype. (E) Determination of soft thresholds in the data. (F) The correlation of modules with grouped traits. (G) Heatmap for module-trait correlations showed that the turquoise module and the blue module have higher mRNAsi levels. Positive and negative correlations. (H) Correlations to the mRNAsi phenotype.

Similarly, to screen genes associated with cancer cell stemness, tumor stemness data downloaded from the literature were matched. A total of 13 tumor samples in TCGA-LUAD without cancer cell stemness data were filtered out and the remaining 502 LUAD samples with cancer stemness data were used in the WGCNA analysis. Similarly, the sample clustering showed no outlier sample. The pickSoftThreshold function determined the optimal soft threshold to equal three, whereby R² was ∼0.85 (Fig. 3E). By using the Dynamic Tree Cut package, we obtained eight modules (Fig. 3F). Relying on the module-trait relevance in Fig. 3G, we noted that the turquoise and blue modules had the highest positive and negative correlations with mRNAsi, respectively. Thus, 2,154 genes (turquoise module) and 995 genes (blue module) were considered mRNAsi-related genes. Finally, we generated scatter plots to visualize the correlation between modules and genes (module membership), and the correlation between traits and genes (gene significance). As shown in Fig. 3H, only correlations in the mRNAsi phenotype was shown.

3.3. Identification of 16 key gene signatures in LUAD

The 37 overlapping genes of the 492 module genes related to M2 macrophages and 3149 module genes relevant to phenotype of mRNAsi were obtained and input into the LASSO, RF and SVM-RFE to uncover gene signatures (Fig. 4A). The LASSO algorithm was run and 20 signature genes were obtained at a 0.02872899 lambda.min (Fig. 4B). The AUCs from the LASSO analysis were 0.999618 and 0.999472 in the training cohort and validation set, separately, illustrating the high precision of the LASSO model (Fig. 4C). Using RF, we selected the top 30 genes according to the Mean Decrease Accuracy plot (Fig. 4D) to construct the logistic regression model. The AUC in the validation set was 0.970588 (Fig. 4E), demonstrating the reliability of the logistic regression model. SVM-RFE resulted in a total of 29 gene signatures identified at the 5-fold cross validation with an accuracy of 0.9895 (Fig. 4F). Finally, BMP8B, DENND3, FAM155B, IL22RA2, ITM2A, MAP3K8, OTX1, SCARF1, TPO, ZMYND15, ABCB4, CBFA2T3, FCAMR, ISL2, MC4R and TAS1R3 was identified as a key gene signature in LUAD through overlapping genes identified by LASSO, RF and SVM-RFE (Fig. 4G).

Fig. 4.

Identification of key genes. (A) Venn diagram for 30 genes common to M2 macrophages and mRNAsi. (B) 20 signature genes were selected by the least absolute shrinkage and selection operator (LASSO) Cox models, and cross-validation results for tuning parameter selection in the LASSO model was displayed. (C) Receiver operating characteristic (ROC) curves for the LASSO model. (D) RandomForest (RF) model to screen out the top 30 genes from 37 intersecting genes. (E) ROC curve for RF. (F) Plot of Support vector machine-recursive feature elimination (SVM-RFE) for prediction accuracy and generalization error changed with number of features after 5-fold cross-validation. (G) Venn diagram of LASSO, RF, and SVM, and 16 key genes were obtained.

3.4. Authentication of key genes associated with OS and DFS of LUAD patients and identification of CBFA2T3 and DENND3 as diagnostic biomarkers

To elucidate the prognostic role of key gene signature in LUAD, we divided LUAD patients in the TCGA and GSE13213 into low- and high-expression groups, respectively. In the TCGA-LUAD cohort, we found that patients with higher expression of CBFA2T3, DENND3, FCAMR, ITM2A, MAP3K8, SCARF1, ZMYND15, and MC4R, and lower expression of FAM155B, ISL2, OTX1 had better OS (p < 0.05, Fig. 5A). In the GSE13213 cohort, we detected 10 key gene signatures, and observed that patients with higher expression of CBFA2T3, DENND3, FCAMR, and SCARF1 had better OS (p < 0.05, Fig. 5B). Further, the expression of four shared genes with distinct survival difference in two datasets (CBFA2T3, DENND3, FCAMR, and SCARF1) were significantly different between tumor and normal samples (Fig. 5C). Meanwhile, a Kaplan-Meier graphs showed that the expression of CBFA2T3, DENND3, and FCAMR were also correlated to DFS except for SCARF1 (Fig. 5D). Given the link between CBFA2T3, DENND3 and FCAMR expression with both OS and DFS, these genes were considered for further analyses.

Fig. 5.

Survival analysis of key genes and identification of diagnostic biomarkers. (A) Validation of overall survival (OS) curves for 11 key genes with significant difference (p < 0.05). (B) Validation of OS curves for four genes in the GSE13213 dataset (p < 0.05). (C) Expression of DENND3, SCARF1, CBFA2T3, and FCAMR in samples. (D) Disease-free survival (DFS) curve for four key genes. (E) ROC curve of the AUC value for key genes in the TCGA dataset. (F) Exclusion of the FCAMR gene based on the AUC value being less than 0.7 in the GSE121397 datasets.

We next investigated the diagnostic value of CBFA2T3, DENND3 and FCAMR. In TCGA, the AUCs of CBFA2T3, DENND3 and FCAMR were 0.878, 0.964 and 0.647, respectively (Fig. 5E). In the GSE121397 dataset, the AUCs of CBFA2T3, DENND3 and FCAMR were 1, 1 and 0.667, respectively (Fig. 5F). The above results suggest that CBFA2T3 and DENND3 have high accuracy in classifying LUAD and normal samples. Therefore, CBFA2T3 and DENND3 are potential diagnostic biomarkers in LUAD.

In addition, we investigated the relevance between diagnostic biomarkers and clinical characteristics. We found that the expression of CBFA2T3 was significantly different between groups stratified by sex (male vs. female), Nodal stage (N0 vs. N2), Tumor stage (T1 vs. T2, and T1 vs. T3), cancer stage (stage Ⅰ vs. stage Ⅲ) and race (Asian vs. White) (Fig.S2A). DENND3 expression was significantly different between groups divided by N stage (N0 vs. N1, and N0 vs. N2), T stage (T1 vs. T2), cancer stage (stage Ⅰ vs. stage Ⅲ, and stage Ⅰ vs. stage Ⅱ) and smoking status (level 1 vs. level 3) (Fig. S2B). Stratified survival analysis of smoking history and gender traits was used to detect the gender bias of CBFA2T3 and the smoking bias of DENND3, as displayed in Fig. S3A-D, there was a significant difference between high and low expression groups of CBFA2T3 only in survival rates of non-smokers, suggesting a independently prognostic significance of CBFA2T3 and DENND3 genes in LUAD cohorts. Further, a nomogram based CBFA2T3 and DENND3 was built for clinical utilize (C-index = 0.595) (Fig. 6A), and calibration curve suggested the excellent efficacy of nomogram for predicting LUAD (Fig. 6B).

Fig. 6.

Nomogram was constructed based on CBFA2T3 and DENND3 for predicting LUAD. (A) Establishment of survival nomogram (C-index = 0.595). Pr: probability, the asterisk after the gene tag represents significance, and the density curve represents the gene score. (B) Calibration curve for the nomogram.

Moreover, GSEA results revealed that both CBFA2T3 and DENND3 expression was related to the calcium signaling pathway, cAMP signaling, cytokine-cytokine receptor interactions, the MAPK signaling pathway, neuroactive ligand-receptor interactions, the PI3K-AKT signaling and the Ras signaling pathways (Fig. 7A–B).

Fig. 7.

Gene Set Enrichment Analysis (GSEA) of two diagnostic biomarkers based on the “C2: KEGG subset”. (A) CBFA2T3, (B) DENND3. The vertical coordinates represent the enrichment scores, with positive ES indicating that a functional gene set is enriched in the front of the sequence and negative ES indicating that a functional gene set is enriched in the back of the sequence. The horizontal coordinates represent the sorted genes based on the correlation coefficients of diagnostic biomarker and all other genes in 574 samples from high to low, and each small vertical line represents a gene.

3.5. Analysis of key genes and diagnostic biomarkers at the level of single cell

To investigate the expression differences of mRNAsi-M2 macrophages-associated key genes in various cell types in LUAD, the public single-cell RNA-seq data of LUAD (GSE123902) was analyzed. After quality-controlled and normalized, highly variable genes were identified and shown in Fig. S4, and the top 30 principal components (PCs) were selected for the PCA elbowplot and categorized into seven cell types: B cells, T cells, Myeloid cells, Fibroblast cells, NK cells, EPCs and ENCs cells (Fig. 8A–C). It can be further seen that the contents of NK cells were significantly different in the tumor and normal samples, where the NK cells infiltration was lower in tumor group than normal samples (Fig. 8D). Results in Fig. 8E (gene expression data for FAM155B, FCAMR, and MC4R are lacking in the GSE123902 dataset) showed that compared with other key genes, the expression of ITM2A was higher in ENCs, and MAP3K8 expression was higher in Myeloid cells, and Fig. S5 suggested that ITM2A within ENCs cells, and MAP3K8 within Myeloid cells significantly expressed lower in tumor groups than in normal groups. For the other gene expression in Myeloid cells and ENC across tumor and normal condition, the downregulated mRNA levels of ABCB4, CBFA2T3, DENND3, ZMYND15 and the up-regulated expression of IL22RA2, ITM2A, SCARF1 were observed in Myeloid cells from the tumor groups. Otherwise, both MAP3K8 and SCARF1 expressions were decreased in ENCs cells within the tumor groups. Besides, only the expression of ISL2 in NK cells from the tumor tissues was considerably down-regulated.

Fig. 8.

Analysis for key genes at the level of single cell. (A) The elbowplot visualized the standard deviation of each principal component (PC). (B) The uniform manifold approximation and projection (UMAP) plot showed that all the cells in the 11 LUAD samples could be classified into 25 clusters. And the distribution of each sample was showed. (C) The UMAP plot showed that 25 clusters could be classified into seven cell types. And the distribution of each sample was showed. (D) Differences in the cell type fractions between the tumor and normal samples using analysis of variance (ANOVA). (E) Expression of key genes in seven cell types.

Next, the Myeloid cells were extracted and re-clustered into three clusters (M1−like macrophages, M2−like macrophages and Mast cells) using the ‘Seurat’ package as well (Fig. 9A–D). Cell type fraction between the tumor and normal samples was presented in Fig. 9E, while there was no distinct differences in three cell types expressions. It was worth noting that the expressions of the two diagnostic biomarkers were differed between the different tissue sources in M2-like macrophages (Fig. 9F–G). On the other hand, considering the distinct expression of MAP3K8 in Myeloid cells between tumor and normal groups, the spearman correlation analysis of MAP3K8 and CBFA2T3, DENND3 were conducted in Table 2, while there was weak correlations.

Fig. 9.

Analysis for diagnostic biomarkers at the level of single cell. (A) Visualization of highly variable genes in Myeloid cells. (B) The elbowplot visualized the standard deviation of each PC in Myeloid cells. (C) The t-distributed stochastic neighbor embedding (t-SNE) plot showed that all the cells could be classified into 11 clusters. And the distribution of each sample was shown. (D) The t-SNE plot showed that 11 clusters could be classified into three cell types. And the distribution of each sample was shown. (E) Differences in the cell type fractions between the tumor and normal samples (ANOVA). (F) Differences in DENND3 expression in each cell type from the tumor and normal samples (Wilcoxon test). (G) Differences in CBFA2T3 expression in each cell type from the tumor and normal samples (Wilcoxon test). **p < 0.01, *p < 0.05.

Table 2.

Spearman correlation results of MAP3K8 and CBFA2T3, DENND3.

Gene	immune_cells	cor	p.value
DENND3	MAP3K8	0.379939229	0
CBFA2T3	MAP3K8	0.238200463	8.43E-09

3.6. Authentication of two diagnostic biomarkers in lung adenocarcinoma cells

Finally, we utilized qRT-PCR to validate the mRNA expression of diagnostic biomarkers in the A549 and H2122 cells (lung cancer cell lines), indicating that the expression results in vitro were consistent with that in public expression. The expression of both CBFA2T3 and DENND3 were higher in lung cancer cells (Fig. 10). The other fourteen key genes expressions were exhibited as well, it could be found that the expression of FCAMR, MAP3K8, SCARF1, MC4R, FAM155B, ISL2,BMP8B,IL22RA2,TPO, ABCB4 and TAS1R3 were up-regulated in lung cancer cells versus the normal lung bronchial epithelial cells, BEAS-2B. The expression of ZMYND15, ITM2A were down-regulated in lung cancer cells. There was no difference in the expression of OTX1.

Fig. 10.

The mRNA expression of two diagnostic biomarkers and partial key genes was detected using qRT-PCR in lung adenocarcinoma cells. ***p < 0.001, **p < 0.01, *p < 0.05.

4. Discussion

Lung adenocarcinoma (LUAD) is a prevalent type of lung cancer presenting with a remarkable fatality profile [1]. LUAD does not have noticeable symptoms in the early stage, and the majority of individuals do not receive an early diagnosis. Therefore, the optimal time for the most effective treatment has often already passed once the patient is diagnosed. Radical surgery still results in a high recurrence and metastasis rate. Therefore, improving early diagnosis rates and overall prognosis of LUAD is of immediate importance for benefiting the lives of LUAD patients.

As is widely known, the two main causes of tumor therapy failure are recurrence and metastasis. There is mounting evidence that cancer stem cell (CSCs), or tumor-initiating cells, are linked to tumor growth and recurrence [27]. Self-renewal, differentiation, tumorigenesis, and tumor heterogeneity are among the characteristics of CSCs [28]. So according to reports, CSCs, along with radiotherapy or chemotherapy, are the primary causes of tumor recurrence [29]. Lung CSCs play a crucial role in lung cancer metastasis and treatment resistance. Pluripotent stem cell proteins are frequently expressed at high levels in LUAD stem cells, especially Nanog and Oct-4 [30]. In lung cancer cells, Oct-4 expression retains CSC-like characteristics and is linked to a poor prognosis in LUAD [31]. Teng et al. [32] discovered that the lung CSC marker Oct-4 is controlled by canonical Wnt signaling, which also plays a significant role in the biology of lung CSCs.

One of the crucial immune cell subsets connected to cancer metastasis are tumor-associated macrophages (TAM). TAMs are the richest inflammatory cell type infiltrating the tumor microenvironment. The development of lung cancer depends on the polarization of macrophages. TAMs come in two different varieties: classically activated macrophages (M1) and alternatively activated macrophages (M2) [33]. Tumor necrosis factor alpha (TNF-α), interleukin 6 (IL-6), interleukin 12(IL-12), and other pro-inflammatory substances can all be produced by M1 macrophages. Additionally, these cells have high levels of CD86 and inducible nitric oxide synthase (INOS), which function to destroy and stop tumor cells.Some characteristic phenotypic markers of M2 macrophages are CD163, CD204, and CD206. Transforming growth factor beta (TGF-β) and IL-10 can be released by these cells, which is beneficial to tumor growth and infiltration [34,35]. M2 TAMs promote cancer cell migration and metastasis. A549 cells overexpressing Oct-4 express high levels of macrophage-colony stimulating factor, which helps to increase M2 macrophages and enhance tumor migration. M2 macrophages can promote tumorigenesis. M2 macrophage infiltration into tumor islets is associated with a poor outcome in NSCLC [36]. Yoshida et al. [37] showed that a higher M2 macrophage density is a standalone predictor of a higher start of dissemination via airspace. Additionally, for stage 0-I LUAD, a high CD25+/high CD163+ immune cell infiltration ratio is a good prognostic indicator. Prostate cancer is aided by the STAT3/MALAT1 pathway when M2 macrophages release IL-8 [38].

This study used bioinformatics to explore biomarkers related to the immune microenvironment and tumor cell stemness in samples from patients with LUAD. This was a preliminary study on the regulatory role between M2 macrophages and cancer stem cells. In this investigation, the gene expression matrices of 574 specimens (59 normal specimens, 515 cancer specimens) were normalized. A differential analysis of normal and tumor samples was performed using the DESeq2 package. A total of 5,410 DEGs were identified. The differential immune cells in LUAD patient cancer tissues and control tissues were evaluated using the CIBERSORT algorithm, and the differences were statistically assessed. WGCNA was used to identify modules related to M2 macrophage and tumor cell stemness. The related modules were intersected to obtain 37 overlapping genes related to immune microenvironment, tumor cell stemness, and M2 macrophages. Cross-validation of intersecting genes was performed using three machine learning algorithms. Among these overlapping genes, 20 characteristic genes were screened using LASSO, 30 genes were screened using RF, and 24 genes were screened using SVM. The intersection of the three methods resulted in 16 genes as key genes. Plotting OS and DFS curves was performed for key genes, which were validated on independent GEO datasets. Three key genes with significant differences in expression were CBFA2T3, DENND3, and FCAMR. ROC curves were drawn for these three key genes, and two key genes with AUC >0.7 were obtained. The genes with an AUC >0.7 and a correlation with OS and DFS were selected as candidate genes. The correlation of candidate genes CBFA2T3 and DENND3 with clinicopathological features was analyzed. Gender, age and race were significantly correlated with the CBFA2T3. The Nodal stage and cancer stage were significantly correlated with DENND3.

Known also as CBFA2T3, the myeloid translocation gene 16 (MTG16) is a transcriptional co-repressor [39,40]. Chen et al. [41] uncovered that CBFA2T3 was significantly associated with lung cancer survival. A protective gene and an important factor in the development and spread of lung cancer is CBFA2T3. Zhang et al. [42] discovered that the tumor suppressor gene CBFA2T3 can be used as a particular biomarker for the identification of NSCLC. Studies have revealed that CBFA2T3 also exerts a role in suppressing other tumor types. By blocking the E protein transcription factor, CBFA2T3 also controls carcinogenesis, colitis, and colonic epithelial differentiation [43]. CBFA2T3 is also involved in cancer and leukemia development and has been discovered to demonstrate leukemia and breast cancer loss of heterozygosity aberrations as well as repetitive translocations [44]. The leukemia fusion protein E2A-Pbx1 as well as wild-type E2A-mediated transcription are both inhibited by CBFA2T3, according to studies. This shows that CBFA2T3 may be a viable therapeutic target for acute lymphoblastic leukemia. MTG8 and CBFA2T3 are transcriptional co-repressors that control how intestinal stem cells migrate and differentiate from their niche into intestinal epithelial cells [45]. Intestinal crypt hyperproliferation and expansion of intestinal stem cells was observed in MTG8 and CBFA2T3 knockout mice. LGR5 and ASCL2, two genes that are particularly expressed by stem cells, are repressed from transcribed by CBFA2T3 binding to the promoters of these genes.The function of CBFA2T3 in macrophages is currently unknown.

A region known as the PHenn domain, which is unstructured, houses an actin-binding site on the guanine nucleotide exchange factor DENND3.For DENND3 to engage in autophagy, the PHenn domain must bind to microfilaments [46]. Xu et al. [47] demonstrated that autophagosome trafficking is promoted by activated Rab12 through its ability to bind to LC3 and associate with autophagosomes that express LC3. As a result, the ULK/DENND3/Rab12 axis is essential for the autophagy that results from starvation. The function of DENND3 in LUAD is still presently unknown, and in vivo and in vitro experiments are required to further elucidate DENND3 biology in LUAD.

Most B cells and macrophages contain FCAMR, a membrane surface protein that binds to the Fc fragment of IgA and IgM and causes endocytosis [48,49]. Li et al. discovered that immunotherapy helps LUAD or lung squamous cell carcinoma (LUSC) patients live longer and have higher quality of life. As a result, distinct searches were made for immune-related gene signatures in the LUAD and LUSC subtypes. Based on the prognostic model for the immune-related gene FCAMR, high- and low-risk groups were discovered to be significantly relevant to LUSC OS and clinical traits. This aids in determining immunotherapy response and keeping track of treatment tactics [50]. Feng et al. used LPS to activate the p38 MAPK and NF-kappaB pathways through the TLR4 receptor, as well as the FCAMR gene's up-regulated expression in human macrophages. This makes it easier for IgM/CuoxLDL complexes to attach to macrophages and for foam cells to develop. The pathogenesis of bacterial infection exacerbating atherosclerosis was preliminarily explored [51]. There is limited literature on FCAMR in tumors, and more experiments are therefore needed to study it.

To achieve further insight into the mechanism by which CBFA2T3 and DENND3 regulate LUAD, we performed a GSEA. Both genes were discovered to be associated with the Ras signaling pathway, PI3K-AKT signaling pathway, MAPK signaling route, calcium signaling pathway, cAMP signaling, and cytokine-cytokine receptor interaction. By influencing the Calcium signaling system, Li R et al. [52] shown that the immune gene CBFA2T3 may help Dehong Hump Cattle (DHH) become more resilient to the effects of heat and disease. DENND3 was identified as one of the genes influencing Ca absorption, according to Reyes Fernandez PC et al. [53].According to Bohlouli M et al. [54], DENND3 can function as a heat stress response (HSR) related gene to control disease resistance in dairy cows. This shows that CBFA2T3 and DENND3 may share a same targeting mechanism through the calcium signaling pathway, and more functional studies are required to confirm this.

Further, Chen et al. [55] investigated the heterogeneity between M2 macrophages and their differentiation related genes in LUAD patients at the single-cell level. Liu et al. [56] also used single cell sequencing analysis, transcriptome sequencing and a series of in vitro experiments to explore the expression of key gene EPHB2 in LUAD. Inspired by these studies, we used different datasets to explore the expression levels of biomarkers in different types of cells, especially M2 macrophages, and tissues from different sources in LUAD patients, thereby demonstrating the innovation of this study. It was discovered by Zhang Q et al. [57] that ISL2 may inhibit EphB2.In this study, tumor tissue had considerably lower ISL2 levels and significantly reduced NK cell expression. This shows that LUAD progression may also be influenced by the interaction mechanism between low ISL2 expression and high EphB2 expression. Both EPHB2 and MAP3K8 may function as kinase genes to encourage ccRCC cancer spread, according to Ghatalia P et al. [58]. According to our research, NK cells were markedly downregulated in tumor tissue. A novel gene feature for predicting the prognosis of lung adenocarcinoma and immune response based on NK cell marker genes was discovered and verified by Song P et al. [59] in conjunction with the thorough examination of single cell and bulkRNA sequencing data. The loss of MAP3K8 specifically targets macrophages, which promotes lung fibrosis and inflammation [60]. By controlling M2 macrophage activation, MAP3K8 can prevent TH2 driven inflammation.A crucial step in the progression of pulmonary fibrosis is the reduction of MAP3K8 expression [61].However, we discovered that there was a poor association between MAP3K8 and two important genes, CBFA2T3 and DENND3.We found the mRNA expression of CBFA2T3 and DENND3 were upregulated in A549 and H2122 (lung adenocarcinoma cells). This study systematically studied the regulatory related genes between M2 macrophages and tumor stem cells in LUAD, reflecting the innovation of this study.

5. Limitations

There are some limitations in the study. Due to the difficulty in collecting samples from LUAD patients, current work solely used mRNA data from lung cancer cells to build a predictive model, the clinical efficacy of which as well as the expression patterns of CBFA2T3 and DENND3 in M2 macrophages need to be further evaluated by collecting a large number of cohorts with LUAD. Further, the GSEA results need to be further confirmed and clarified in combination with functional experiments such as knockout or overexpression of CBFA2T3 and DENND3 genes in vitro and in vivo. In future research, we will also further validate the expression levels of key genes by targeting macrophages isolated from patients. And we will also detect the expression of key genes in lung cancer patients through methods such as flow cytometry and immunohistochemistry.we hope to provide new theoretical basis for the diagnosis of LUAD.

6. Conclusions

This study identified two key LUAD candidate diagnostic genes–CBFA2T3 and DENND3–found to be associated with M2 macrophages and tumor cell stemness. The clinical relationship and potential molecular mechanism of LUAD were also investigated. This research identifies a possible target for lung cancer targeted therapy, but our research still requires additional clinical validation to confirm its results.

Availability of data and materials

The data used to support the findings of this study are included within the article.

Author contributions

XiaoFang Wang; Xuan Luo: Conceived and designed the experiments; Performed the experiments; Analyzed and interpreted the data; Wrote the paper.

Li Bian: Conceived and designed the experiments; Contributed reagents, materials, analysis tools or data.

ZhiYuan Wang; YangHao Wang: Performed the experiments; Analyzed and interpreted the data.

Juan Zhao: Performed the experiments; Wrote the paper.

Data availability statement

Data associated with this study has been deposited at https://github.com/wangxiaofangkmmu/steam/blob/main/code.R.

Ethics approval and consent to participate

Not applicable.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Regional Fund Project of the National Natural Science Foundation of China and Yunnan Provincial Department of science and Technology-Kunming Medical University applied basic research joint special major project (Serial number: 82060423; 202001AY0700 1–007).

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Li Bian reports financial support was provided by the Regional Fund Project of the National Natural Science Foundation of China. Li Bian reports financial support was provided by Yunnan Provincial Department of science and Technology-Kunming Medical University applied basic research joint special major project.

Appendix A. Supplementary data

The following is the Supplementary data to this article: