Normalized expression matrices processed as Transcripts Per Kilobase of exon model per Million mapped reads (TPM) from scRNA-seq performed using a 10 × genomics platform were downloaded from the Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/) database. The sequencing analyses were conducted on samples from nine fresh GBM tumors from adult patients in Massachusetts General Hospital (MGH) (GSE131928) (4). A total of 16,201 cells were acquired. A seurat file was created by importing the expression matrix data into the version 3.6.2 R studio program and saving the data in RDS format.

The data were normalized using the global-scaling method “LogNormalize” and preprocessing the data by linear conversion. To detect the most variable genes used for dimensionality reduction, we employed the “FindVariableFeatures” function in R studio. Next, we utilized PCA to reduce the dimensional linearity based on the variable genes. JackStraw plots and Elbow plots were used to decide the number of dimensionalities. Finally, graph-based unsupervised clustering was conducted and visualized using a non linear t-SNE plot, with a resolution of 0.4, defined by the functions of “FindNeighbors” and “FindClusters”. All the operations were processed by using the seurat R package in R studio.

To identify the cell types, we first analyzed the Spearman correlation coefficient between the transcriptomes of each cell across all nine tumors in our dataset and each cell type-specific gene expression profile in the HPCA reference set using SingleR package. Then, the expression levels of distinct cell type-specific marker genes based on the CellMarker databases (24) were utilized to further confirm the cell types. Markers for macrophage were C1QA, C1QB, C1QC, TYROBP, and CD68. For T lymphocyte, the markers were CD3G, GZMH, IL2RB, PRF1, and ICOS. TMEM119 was employed to distinguish the microglia from TAM clusters. The relatively positive expressions of MHC-II molecules in TAMs were considered as distinct features of primed state. The cell type information was visualized by t-SNE plot in R studio.

The marker genes of distinct clusters were identified using the “FindAllMarkers” function of the seurat package and visualized as a heat map, violin plots, and feature plots using the ggplot2 package in R studio. The positive percentage of the cells where the gene is defined as the marker is at least 25%.

The gene regulatory network and specific transcription factors were identified by employing the Single Cell Regulatory Network Inference and Clustering (SCENIC) tool (25) using default parameters. The SCENIC analysis was run as described on the cells that passed the filtering using the 20,000 motifs database for RcisTarget and GRNboost R packages. The regulated genes and AUCell score of each transcription factor were calculated to estimate the specificity to each cell type. The AUCell scores were visualized by t-SNE plot and heat map. The top three regulons in each cell type were acquired by calculating the RSS. The relationships among transcription factors were shown in heat map by connection specificity index (CSI).

DAVID (https://david.ncifcrf.gov/) is an open website providing a comprehensive set of functional annotation tools to extrapolate biological meaning from an extensive list of genes. Using DAVID, we identified the main cellular physiological activities reflected by the top 500 differentially expressed genes of each cluster. The KEGG pathways were displayed as a bubble plot by the ggplot2 package in R studio.

The expression data of SPI1 and the clinical information of glioma samples were obtained from TCGA, CGGA (26), and Rembrandt databases. The data were sorted into three subgroups according to the WHO grade and used to calculate the mean and 95% confidence interval (CI) of SPI1 expression. The results were visualized as scatter plots.

We utilized TIMER2.0 (Tumor Immune Estimation Resource version 2.0) online web (http://timer.cistrome.org/) (27) to analyze the correlation between the expression of SPI1 and other genes. We input “SPI1” in the “Interested Gene” search box of the “Cene_Corr” module in TIMER2.0 and “C1QA”, “C1QB”, “C1QC”, “SRGN”, “TYROBP”, “GAP43”, “GPM6B”, “SEC61G”, “PTN”, “CX3CR1”, “CSF1R”, and “IRF8” in the “Gene Expression” box, and obtained the correlation in all tumor types. We picked up the GBM results and saved the scatter plots with rho and p value.

The tumor purity, macrophage infiltration, and survival prognosis analysis were performed as previously described (28).

To examine the immune microenvironment of GBM, we obtained 10 × Genomics scRNA sequencing datasets derived from fresh GBM tissues that spanned nine patients, accessed from the Gene Expression Omnibus (GEO) database (accession number GSE131928) (4). A total of nine datasets were integrated and comprehensively analyzed to assess GBM heterogeneity and determine the nature of various associated immune microenvironments. We used t-distributed stochastic neighbor embedding (t-SNE), the unsupervised non linear dimensionality reduction algorithm, to differentiate cell types based on their relative gene expression values. Overall, single-cell transcriptomes for a total of 16,201 cells were retained after initial quality controls and were differentiated into 22 different clusters (cluster 0~-21), visualized as bidimensional t-SNE maps ( Figure S1A ). The distribution of each dataset is displayed in Figure S1B . The clusters were sorted into four main cell types by comparing their gene expression characteristics for tumor (light gray), macrophages (green), microglia (blue), and T lymphocyte (purple) ( Figure 1A ). Surprisingly, the number of malignant cells (tumor) accounted for 62.31% (10,095 cells) of the cell population, suggesting that numerous tumor-associated non-malignant cells are a prominent feature of the GBM microenvironment. We observed that TAMs (macrophage and microglia) comprised 36.39% of the tumor tissue cells (the numbers of macrophage and microglia were 3,288 and 2,608 cells, respectively), and the percentage of T cells was only 1.30% (210 cells) of the total cell count ( Figure 1B ). The percentages of cell types in each sample are shown in Figure S2 .

Heat mapping was performed to examine the expression levels of the top 10 genes in the different clusters. As shown in Figure 1C and Table S1 , each cluster displayed distinct gene expression features. The clusters in the TAM group (particularly clusters 2, 3, 6, 7, 9, and 19) shared a number of similar marker genes, such as complement c1q C chain (C1QC), complement c1q B chain (C1QB), Metallothionein 1G (MT1G), serglycin (SRGN), c-c motif chemokine ligand 3 like 3 (CCL3L3), major histocompatibility complex class II DR beta 5 (HLA-DRB5), and s100 calcium binding protein A8 (S100A8). By analyzing the expression of different genes, we observed that TAMs could be identified by the expression of Arachidonate 5-Lipoxygenase Activating Protein (ALOX5AP), complement c1q A chain (C1QA), C1QB, C1QC, SRGN, and TYRO protein tyrosine kinase binding protein (TYROBP), whereas malignant cells expressed the marker genes growth-associated protein 43 (GAP43), glycoprotein M6B (GPM6B), SEC61 translocon subunit gamma (SEC61G), and pleiotrophin (PTN) ( Figure 1D ). The non malignant cells were differentiated into T cells and TAMs by the expression of the genes CD3 gamma chain (CD3G), granzyme H (GZMH), interleukin 2 receptor subunit beta (IL2RB), perforin 1 (PRF1), and inducible T cell costimulator (ICOS) ( Figure S1C ). The non malignant and malignant cells were also easily separated in bidimensional PC maps, which rely on linear principal component analysis (PCA) to reduce dimensionality ( Figure S1D ).

Altogether, our data suggest that TAMs are the most common immune cell type in GBM, comprising over half of the tumor cell population.

To comprehensively characterize the signatures of TAMs in GBM, we subclustered and re analyzed the cells in the macrophage and microglia group. In total, 5,896 cells were sorted and hierarchically sorted into 14 clusters ( Figure S3A ). CD68 has been widely recognized as a pan-macrophage marker, and we therefore used feature and violin plots to visualize the expression and distribution of CD68 in these clusters. As shown in Figures S3B, C , CD68 was highly expressed in almost every cluster, except for cluster 12, which may be tumor cells misdefined as TAMs according to the gene signature above. Then, we further analyzed the CD68-positive clusters, which we considered to be TAMs. The macrophage compartment in GBM contains both brain-resident microglia and bone marrow–derived macrophages. TMEM119 is a reliable microglia marker that discriminates inherent and extrinsic macrophages in the brain (29). Using the violin plot of TMEM119 expression, we separated the macrophage clusters into two groups. The TMEM119-positve clusters (clusters 4, 5, 6, 7, 10, and 11) were identified as resident microglia, and the TMEM119-negative clusters (clusters 0, 1, 2, 3, 8, 9, and 13) were defined as bone marrow–derived macrophages ( Figure S3D ). By examining the marker genes defining the TMEM119-positive clusters, we selected cytoplasmic fmr1 interacting protein 1 (CYFIP1), ectonucleoside triphosphate diphosphohydrolase 1 (ENTPD1), and vav guanine nucleotide exchange factor 1 (VAV1) as possible microglia marker genes ( Figure S3D ). MHCII is a glycoprotein present on specialized antigen-presenting cells, including macrophages. The elevated expression of MHCII indicates that macrophages have transitioned to a primed state, in which they have acquired the ability to present tumor-specific antigens (30). Macrophages otherwise maintain a repressed state. We therefore examined the expression of MHCII by assessing the levels of the human genes, which were HLA-DQ alpha 2 (HLA-DQA2) and HLA-DQ beta 2 (HLA-DQB2). As shown in Figure S2E , HLA-positive expression was confined to clusters 1, 2, 4 5, 6, 8, and 9. Together with their expression of TMEM119, the cells were subdivided into four distinct groups as follows: TMEM119⁺-HLA⁺ cells (clusters 4, 5, and 6), TMEM119⁺-HLA⁻ cells (clusters 7, 10, and 11), TMEM119⁻-HLA⁺ cells (clusters 1, 2, 8, and 9), and TMEM119^—HLA⁻ cells (clusters 0, 3, and 13) ( Figure S3F ).

Next, we wished to determine the precise nature and activation status of TAMs in GBM. In our analysis of the distinctive gene expression profiles of microglia, we found that the cell population of cluster 11 charactered by matrix metallopeptidase 9 (MMP9), MMP25, and periostin (POSTN), which are associated with cell migration, should be more likely microglia with naïve status as an immunological defender in the brain, which has migration and chemotaxis ability. We defined cells with this expressions profile as “microglia: unactivated”. Similarly, the expression of cytokine-like 1 (CYTL1) and early growth response 3 (EGR3) were used to define primed microglia. Interferon-induced protein with tetratricopeptide repeats 3 (IFIT3) and interferon alpha inducible protein 27 (IFI27) distinguished microglia in the repressed state. The gene profile analysis of macrophages suggested that cluster 8 represented priming macrophages, owing to the expression of cell cycle-associated genes [aurora kinase B (AURKB), cell division cycle associated 3 (CDCA3), and assembly factor for spindle microtubules (ASPM)]. HLA-positive macrophages could be classified as primed, and HLA-negative macrophages (with the exception of cluster 8) were categorized as repressed, due to the expression of MT1G and ankyrin repeat domain 28 (ANKRD28) ( Figures S4A, B ).

With all of the above characteristics in mind, we redefined the original 14 clusters as seven distinct subsets by comparing the expression of TMEM119 and MHCII with the expression profiles of different clusters ( Figures 2A and S4C ). TAMs in GBM consist primarily of bone marrow–derived macrophages, accounting for 60.28% of the cell population, of which 4.55% are undergoing priming, 27.51% are in the primed state, and 28.22% are repressed. Of the remaining cell population, 37.57% consists of resident microglia in unactivated (2.17%), primed (23.74%), and repressed states (11.65%) ( Figure 2B ).

To investigate the relationship between the infiltrated TAMs and different GBM subtypes, we assessed the distribution of the cells covered by the nine tumor datasets and selected the MGH105 (40.28%), MGH124 (18.00%), and MGH115 (9.94%) datasets for analysis, as together they covered over half of the total cell number and seven different clusters ( Figures 3A and S5 ). Fittingly, we found that these datasets represented diverse tumor subtypes previously defined by TCGA (MGH105 was classified as “mixed”, MGH115 as “classical”, and MGH124 as “proneural”) (4), and we therefore examined the characteristics of these distinct GBM subtypes using these three datasets. In the TCGA-mixed tumor, almost all TAMs in the GBM were bone marrow–derived macrophages, with primed state cells occupying 54.95% of the total number (clusters 1 and 2) and repressed state cells accounting for 41.56% (cluster 0) ( Figure 3B ). The TAMs in the TCGA-proneural tumor were characterized by primed microglia (cluster 4; 45.24%), repressed macrophages (cluster 3; 41.19%), and priming macrophages (cluster 8; 10.65%) ( Figure 3C ). A different composition was observed in the TCGA-classical tumor, with repressed microglia (cluster 7) comprising 73.21% of the total cell number and other TAM subtypes, such as repressed macrophages (cluster 3; 15.01%) and priming macrophages (cluster 8; 7.00%), making up the rest of the cell proportion ( Figure 3D ).

Heat map analysis showed that macrophages/microglia in the same activation state possessed distinct gene expression profiles in the different subtypes of GBM ( Figure 3E and Table S2 ). Each of the clusters included cells present in at least three of the nine datasets ( Figure 4A ). To further clarify the main cellular physiological activities of TAMs, the top 500 differentially expressed genes in each cluster were identified and evaluated by KEGG pathway analysis. As shown in Figures 4B, E , “oxidative phosphorylation” was one of the most repressed macrophages from the TCGA-mixed tumor (cluster 0), but “protein processing in ER” was the most enriched set of pathways in the same macrophages from the TCGA-proneural and classical tumors (cluster 3). Primed state macrophages in the TCGA-mixed tumor (clusters 1 and 2) were enriched in the “ribosome” pathway ( Figures 4C, D ). Cluster 4, which represents primed microglia that make up the highest percentage of the cells in the TCGA-proneural tumor, was characterized by the KEGG enrichment term “phagosome” ( Figure 4F ). The pathways associated with “leishmaniasis” were enriched prominently in repressed microglia from the TCGA-classical tumor (cluster 7) ( Figure 4G ). The only TAM cluster present in all three subtypes of GBM was priming macrophages (cluster 8), which were enriched for pathways involved in “DNA replication” and the “cell cycle” ( Figure 4H ).

To identify the distinct key transcription factors of different states of macrophages/microglia and compare the regulon differences, Single-Cell rEgulatory Network Inference and Clustering (SCENIC) was employed to draw the gene regulatory network. Based on the calculated regulon activity score (RAS) of transcription factors, we could identify the differences in the activities of regulons among the cell types ( Figure 5A ). The relevance among the regulons was evaluated by connection specificity index (CSI) and visualized in Figure S6 . Some regulons, such as E2F transcription factor 7 (E2F7) and Cone-Rod Homeobox (CRX), had elevated activities in primed macrophage but decreased activities in repressed macrophage. Meanwhile, some others including SMAD Family Member 1 (SMAD1) and JunD proto-oncogene (JUND) showed low activities in primed but a high activities in repressed state of macrophages. Similarly, high activities of cAMP-responsive element modulator (CREM) and interferon regulatory factor 3 (IRF3) were calculated in primed microglia, but Transcription Factor 7 (TCF7) and interferon regulatory factor 9 (IRF9) in repressed microglia. By calculating the regulon specificity score (RSS), we examined the crucial regulons of each cell type and visualized the top three regulons in sequence diagrams and t-SNE plots ( Figures 5B, C and S7A–I ). Surprisingly, we found that SPI1 is picked out in top three of priming macrophages and all states of microglia, and also important for primed and repressed macrophages ( Table S3 ). SPI1 is known as a crucial transcription factor for the development of macrophage (31, 32). Our results suggested that the transcription activity of SPI1 is essential for macrophages and microglia maturation and polarization, leading to a tumor-harmful microenvironment formation.

In order to further clarify the vital role of SPI1 in GBM, the expression levels in different grades were measured by analyzing the transcriptome data in TCGA, CGGA, and Rembrandt databases. As shown in Figure 6A , grade IV (GBM) samples had the highest level of SPI1. As previously mentioned that marker genes of the TAM group included SPI1, C1QA, C1QB, C1QC, SRGN, and TYROBP ( Figure 1D ), we utilized Spearman’s rho value to interrogate the coexpression relationship between the expression level of SPI1 and these genes in GBM. After being analyzed in the TCGA-GBM dataset, the statistical positive correlations were observed between the SPI1 level and the genes expressed in the TAM group (C1QA: r = 0.823; C1QB: r = 0.873; C1QC: r = 0.905; SRGN: r = 0.687; TYROBP: r = 0.774) ( Figure 6B ), but not in the Tumor group (GAP43: r = 0.043; GPM6B: r = −0.135; SEC61G: r = −0.117; PTN: r = −0.298) ( Figure S8 ). Meanwhile, we employed the TIMER algorithm to explore if there were potential relationships between the macrophage infiltration and the expression of these genes. We found that the elevated expression level of SPI1 had lower tumor purity and a higher infiltrating level of macrophage (Tumor purity: r = −0.528; macrophage infiltration: r = 0.687) ( Figure 6C ), so did TAM genes ( Figure S9A ). The tumor genes had no statistical correlation with tumor purity and macrophage infiltration ( Figure S9B ). Moreover, survival analysis using the Kaplan-Meier plotter tool displayed that the upregulated expression of SPI1 in tumor mass was linked to short overall survival (OS) and disease-free survival (DFS) time in patients with GBM ( Figure 6D ). Moreover, we inspected the likely downstream target genes of SPI1 in TAMs. By analyzing the glioma data in the TCGA database, the expression level of SPI1 displayed a significant positive correlation with target genes (CX3CR1: r = 0.53; CSF1R: r = 0.876; IRF8: r = 0.678) ( Figure S10A ). The co-expression signature was also observed in violin plots in single-cell data ( Figure S10B ).

The above results implied that tumor treatment targeting SPI1 could probably remodel the immune microenvironment of tumor and prolong the life span of GBM patients.