In total, four bulk datasets and one single-cell dataset for NSCLC were collected and combined in this research.

The TCGAbiolinks R package (version: 2.25.0) was used to download the clinical characteristics and gene expression data of LUAD and LUSC patients from the TCGA database [32]. We retained only the expression data that had matching clinical traits. GSE13213 and GSE31210 are two external datasets in this study. The GSE13213 dataset, based on the Agilent platform, contains gene expression microarray profiles and clinical information for 117 patients with LUAD [33]. The GSE31210 dataset, based on the Affymetrix platform, contains gene expression microarray profiles and clinical information for 226 LUAD samples included in our study [34, 35]. The removeBatchEffect() function in the limma R package (version: 3.52.1) was used to remove the batch effect among the four bulk RNA-seq datasets (TCGA-LUAD, TCGA-LUSC, GSE13213, and GSE31210).

There are 190 single-cell datasets, comprising 6,297,320 cells, available from TISCH2 (Tumor Immune Single-Cell Hub, http://tisch.comp-genomics.org/) [36].

We processed data from the study with accession number GSE117570, which includes 11,485 cells from four patients [37, 38].

To assess the proportions of M0, M1, and M2 macrophage subclasses in NSCLC samples from four bulk datasets—two TCGA RNA-seq datasets and two external GEO microarray datasets, the CIBERSORTx R tool (https://cibersortx.stanford.edu/) was used to estimate their abundance. This tool, designed and developed by Prof. Alizadeh and Prof. Newman, is a popular method for providing the percentages of 22 different immune cell types by analyzing gene expression profile data [39, 40].

Samples were divided into two groups based on the median immune score of each macrophage subcluster type as follows (Algorithm 1).

First, we filtered the genes based on the following criteria: genes in the top 75% of the median absolute deviation (MAD) and with MAD greater than 0.01. We removed genes that exhibited minimal variation across all samples because they were considered to have little effect on distinguishing between high- and low-M1 macrophage patient groups. Then, samples were screened to remove outliers. Finally, WGCNA was performed using the WGCNA R package (version 1.71) to identify gene modules most associated with M1 macrophage after dividing patients into high- or low-M1 macrophage content groups based on the median M1 infiltration score [14].

The pickSoftThreshold() function of WGCNA was used to determine the soft power value (beta). The next step was to use hierarchical clustering to find gene modules, with a minimum of 100 genes per module (minModuleSize = 100). Subsequently, the relationships between modules and sample characteristics were analyzed. Finally, the genes in the most relevant module were identified and defined as M1-related genes.

The R package Seurat (version: 4.1.1) was used to conduct scRNA-seq data analysis [41–44]. Genes detected in fewer than three cells and cells with fewer than 200 detectable genes were disregarded, and the proportion of mitochondrial genes was restricted to under 30%. The doubletFinder_v3() function of the DoubleFinder R package (version: 2.0.3) was used to identify and remove doublet cells in the scRNA-seq data [45]. The LogNormalize method was selected for normalizing data using the NormalizeData() function. We analyzed the four scRNA-seq datasets from four patients separately and then merged them.

The criteria for removing principal components (PCs) were as followings: (1) a cumulative contribution rate greater than 90%; (2) each PC contributing less than 5% to the variance; (3) a difference of less than 0.1% between two consecutive PCs. The resolution was determined using the clustree R package (version: 0.5.0). Increasing the resolution resulted in more interlacing of clusters, indicating that our classification was becoming excessive.

Annotation was performed using the SingleR (version: 1.10.0) and celldex (version: 1.6.0) R packages [46]. The HumanPrimaryCellAtlasData() dataset served as the reference. tSNE was used for dimensionality reduction. Marker genes in subpopulations were identified using the FindAllMarkers() function in the Seurat R package with the Wilcox test. Cutoff values were set at log fold change threshold (logfc.threshold) = 0.25 and minimum percent (min.pct) = 0.25. Finally, only positively expressed genes were retained (only.pos = TRUE).

A similar protocol was used for recluster analysis of the macrophage subpopulation. The difference was in the annotation process, where cell types were assigned based on a combination of marker genes and functional analysis results, rather than auto-annotation. To assess the precision of our reclustering system, we utilized AUCell (version: 1.18.1) in our study [47]. AUCell enables the identification of cells with active, specific gene expression signatures. RUNX3, ADAM19, LIMD2, PRKCB, CYTIP, FNBP1, ICAM3, WIPF1, TAP1, PSMB9, and LAP3 were defined as M1 marker genes. Additionally, the topGO R package (version: 2.48.0) was employed to explore the functional characteristics within each cluster.

The M1-related risk genes were defined as a class of group genes that do not exhibit linear relationships. They are overlapping genes between the first gene list from the M1-related gene co-expression module and the second gene list from marker genes in the M1 subcluster. First, we intersected the two gene lists and then removed co-linear genes based on LASSO analysis. LASSO analysis was performed using the glmnet (version: 4.1.4) and survival (version: 3.4.0) R packages to avoid overfitting problems.

By concentrating on significant M1-related macrophage risk genes, identified from the overlapping genes between the pink module from WGCNA and maker genes from the M1 subcluster, we created a prognosis-associated signature for NSCLC patients. Overall survival (OS) was the outcome of this study, and prognostic M1-related risk genes were examined using a multivariate Cox regression model on the training dataset. Genes were considered candidate prognostic genes if they had a p-value < 0.1 in the regression model.

The mathematical formula used to calculate the M1-related prognostic signature was:

where β_i represents the coefficient value from Cox regression analysis, exp_i represents the expression levels for gene i, and n represents the number of significant risk genes.

The schematic workflow for designing and validating the prognostic signature in NSCLC is shown in Figure 1.

Two R packages, survival (version: 3.4.0) and survminer (version: 0.4.9), were used to determine the best cutoff of risk scores. NSCLC patients were then divided into high- and low-risk groups based on the optimal risk score cutoff.

Univariate and multivariate Cox regression analyses were conducted to evaluate whether age, gender, stage, and risk score are independent risk factors. Two R packages, survival (version: 3.4.0) and forestplot (version: 2.0.1), were used for assessment and visualization. Risk factors with p-value < 0.05 and HR > 1 were regarded as significant independent risk factors for poor prognosis in NSCLC patients.

The protein expression levels of risk genes in the risk score model were obtained and compared using the HPAanalyze R package (version: 1.14.0) [48]. HPAanalyze provides an interface for accessing and visualizing data from the HPA database (https://www.proteinatlas.org/). HPA is an open-source website for downloading and visualizing human protein data in cell lines, normal tissues, and tumors [49–55]. The protein expression profiles of risk genes in lung cancer and normal lung tissue were calculated and visualized separately using the hpaVisPatho() and hpaVisTissue() functions from quantified IHC staining results. A heatmap displaying the expression of relevant proteins in low, medium, and high levels was created.

To comprehensively validate the functional differences between high- and low-risk score groups, we included three popular functional enrichment analysis methods in our study: Metascape [56], GSEA, and ssGSEA.

DEGs between high- and low-risk score groups were identified using the limma R package (version: 3.52.1) [57]. Genes with |Fold change| > 1.5 and p-value < 0.05 were treated as significant DEGs. Up-regulated and down-regulated genes in the high-risk score group were separately uploaded to Metascape (https://metascape.org/gp/index.html#/main/step1). Enriched terms were downloaded in a bar graph summary format.

KEGG pathways gene sets were downloaded from the MSigDB database (https://www.gsea-msigdb.org/gsea/msigdb) [58–60]. Enrichment analysis and figure plotting were performed using the GSEA() and gseaplot2() functions from the clusterProfiler (version: 4.4.4) and enrichplot (version: 1.16.1) R packages.

Cancer hallmark gene sets were collected from the MSigDB database. The gsva() function in GSVA (version: 1.44.2) and GSEABase (version: 1.58.0) packages were used to calculate the enrichment score of each gene set in each sample using the ssGSEA algorithm . The fortify_mantel() function in the ggcor R package (version: 0.9.7) was employed to visualize the relationship between enrich score and risk score.

This study utilized R language (version 4.2.2) for bioinformatics and statistical analysis. The significant cutoff for p-values was set at 0.05. Kaplan-Meier analysis was performed to assess survival differences between patient groups. Additionally, a t-test was employed to analyze gene scores among M0, M1, and M2 groups. The nomogram illustrating the relationship between the sum of risk factors scores and survival time was generated using the regplot R package (version: 1.1). Genomic variation analysis included the collection of copy number variation (CNV) data in NSCLC from the cBioPortal database (https://www.cbioportal.org/). The frequencies of CNV changes for risk genes were shown by the ggplot2 R package (version: 3.3.6). The VennDiagram R package (version: 1.7.3) was utilized to compute and visually represent overlapping genes between the M1 characteristics-related gene list and the M1 marker gene list.

In total, the GSE13213 dataset contained 117 samples and 19,412 genes, while the GSE31210 dataset included 226 samples and 20,857 genes. The TCGA-LUAD dataset comprised 596 samples and 59,427 genes, and the TCGA-LUSC dataset had 551 samples and 59,427 genes. These four datasets were integrated to mitigate batch effects (Figure S1). Finally, after removing samples without clinical information and excluding normal samples, we obtained data from 1,475 samples representing 1,432 patients and 15,596 genes for further prognostic analysis.

The expression profile was input into the CIBERSORT() function of the CIBERSORTx R tool (https://cibersortx.stanford.edu/). This allowed the calculation of the immune score for each cell type in each sample. Subsequently, samples were grouped based on the median immune score for M0, M1, and M2 categories individually. The survival curve in the M1 classification system exhibited the most significant differences between high and low immune score groups compared to other macrophage subclusters (Figure 2A–C). Therefore, our focus shifted to identifying M1-related risk genes for subsequent prognostic analysis in NSCLC.

After filtering genes based on the cutoff of MAD across all samples, 11,697 genes were retained. Outliers were removed, resulting in 1,773 samples. As visualized in Figure 2D, a soft power beta value of 3 was selected as optimal. Setting the minimum module size to 100 (Table 1) yielded 12 modules; genes in the grey module, which could not be classified into color modules, were ignored (Figure 2E). The pink module showed the strongest correlation with the M1 trait and no significant association with other modules (Figure 2F, G). Therefore, 222 genes in the pink module were identified and defined as M1 macrophage-related genes for NSCLC patients.

The identification of M1 macrophage-related marker genes from scRNA-seq data in NSCLC patients involves two main steps. First, macrophage subpopulations are isolated from all single cells within NSCLC patients. Second, M1 macrophage-related marker genes are identified.

After filtering doublets, genes, and samples, we obtained the following datasets: 6,610 genes across 1,788 cells for the first patient, 9,317 genes across 1,250 cells for the second patient, 3,205 genes across, 319 cells for the third patient, and 7,454 features across 1,364 cells for the last patient (Figure S2A–C). The optimal number of PCs and resolution were set at 14 and 0.6, respectively ( Figure S2D–G). Four Seurat objects were normalized and integrated, resulting in 21 clusters (Figure 3A, Figure S2H). Additionally, seven main cell subpopulations were identified through SingleR annotation: epithelial cells (1,851 cells), monocyte (957 cells), NK cells (676 cells), macrophage (980 cells), B cells (170 cells), fibroblasts (52 cells), and endothelial cells (71 cells), as shown in Figure 3B. The tSNE plot depicted the distribution of immune cells in the scRNA-seq data (Figure 3C). The classification of immune cells into various subtypes (e.g., T cells, B cells, macrophages) is crucial for understanding the composition of immune cell populations in samples. This classification is particularly important in studying diseases such as cancer, where the immune response plays a critical role. By categorizing immune cells, we gain insights into how the immune system is responding. Single-cell analysis enables the identification of rare or novel immune cell subpopulations that may significantly impact disease progression or treatment response. Furthermore, immune cell classification provides crucial context for interpreting other study results. For instance, variations in immune cell composition may correlate with clinical outcomes or other molecular findings. To validate our classification approach, a bubble plot featuring three marker genes per cell type was generated (Figure 3D). The distribution of each cell type was visualized in a stacked percentage bar plot (Figure 3E), revealing similar proportions across patients. Additionally, enriched functional pathways associated with different cell types were illustrated in Figure S3A–B (GO-MF and KEGG).

The macrophage subpopulation was isolated from all single cells for further analysis (Figure 4A). We determined the optimal number of PCs and resolution to be 11 and 1.6, respectively. Figure 4B depicts the cell distribution of M0, M1, and M2 using tSNE visualization. Each cell’s AUCell score was calculated using the AUCell R package, with cells from the M1 cluster exhibiting higher scores than those from the other clusters (Figure 4C, D). Figure 4E illustrates the distribution of individual heterogeneity within cell clusters. Additionally, six marker genes and the top six DEGs are presented in Figure 4F. Enriched pathways for M0, M1, and M2 are detailed in Figure S3C–D, encompassing GO-MF and KEGG pathways. These findings collectively affirm the reliability of our cell type classification system.

Eleven M1 macrophage-related risk genes were identified by intersecting the genes from the M1 trait-related and M1 marker gene lists (Figure 5A). LASSO regression analysis was subsequently applied to these 11 genes (RUNX3, ADAM19, LIMD2, PRKCB, CYTIP, FNBP1, ICAM3, WIPF1, TAP1, PSMB9, and LAP3) to mitigate model overfitting issues (Figures 5B, C). After eliminating genes with collinear associations, namely ADAM19, ICAM3, WIPF1, and LAP3, four risk genes remained significant with a p-value < 0.1. Table 2 summarizes the results from the Cox regression model. The risk score was defined based on the outcomes of the Cox regression analysis.

Subsequently, the risk score for each NSCLC patient was computed using the formula above. The risk plot demonstrated a correlation between risk score and gene expression levels, highlighting that elevated expression levels of ADAM19 and LAP3 are linked to poor prognosis in NSCLC patients, whereas lower expression levels of ICAM3 and WIPF1 are associated with favorable prognosis (Figure 5D). Figure 5E further illustrates a significant difference in survival time between low- and high-risk score groups among NSCLC patients.

To integrate multi-omics perspectives, we incorporated protein expression data from the HPA database using the HPAanalyze R package. Figure 6A presents an accumulated percentage bar chart for three of the four prognostic risk genes (the protein expression data of ADAM19 is unavailable from the HPA database.). Consistent with dysregulation observed in transcriptomic analysis, WIPF1 protein levels were notably decreased in tumor lung samples compared to normal lung samples. Additionally, genomic variations were detected in these prognostic genes. For instance, we observed an increased frequency of copy number amplification for the LAP3 gene in the LUAD and LUSC cohort from cBioPortal (Figure 6B). Independent prognostic analysis was conducted to validate whether clinical characteristics and risk scores could serve as independent prognostic factors for NSCLC patients (Figure 6C, D). These results confirmed that age, gender, stage, and risk score are indeed independent risk factors. This suggests that these factors alone can predict outcomes in NSCLC patients without needing additional factors. A nomogram was generated using age, gender, stage, and risk score factors (Figure 6E). The nomogram facilitates the assessment of 3-, 5-, and 10-year survival rates for each NSCLC patient.

The differential expression analysis between high- and low-risk groups was conducted, and the results are presented in Figure 7A. The high-risk group exhibited 86 up- and 239 down-regulated genes. The up-regulated genes showed enrichment in pathways related to cell structure development, immune responses, and several signaling pathways, including skin and cartilage development, as well as cellular responses to interferon-gamma, IL-17, and Wnt signaling pathways (Figure 7B). Conversely, the down-regulated genes were enriched in pathways associated with disorders such as dopaminergic neurogenesis (Figure 7B). Additionally, GSEA corroborated these findings, highlighting enriched functional pathways among up-regulated genes, including the cell cycle, cytokine receptor interaction, antigen processing and presentation, and toll-like receptor signaling pathway.

In contrast, down-regulated genes were enriched in pathways related to drug metabolism and steroid hormone biosynthesis (Figure 7C, D). To investigate the dysregulation pathways influenced by the risk score, we analyzed the relationship between the risk score and 50 cancer hallmarks. Figure 7E indicates that the risk score is significantly positively associated with DNA repair, adipogenesis, PI3K-AKT-mTOR signal, fatty acid metabolism, heme metabolism, bile acid metabolism, and oxidative metabolism. Conversely, the risk score is significantly negatively associated with G2M checkpoint, protein secretion, and peroxisome functions.