Interrogation of the Microenvironmental Landscape in Brain Tumors Reveals Disease-Specific Alterations of Immune Cells

The immune composition of brain malignancies

We first determined the broad immune cell abundance in the brain TME by analyzing the pan-leukocyte marker CD45 through immunofluorescence (IF) staining of whole tissue sections and flow cytometry (FCM) analyses of non-tumor brain tissue, IDH mut low-grade and IDH wt high-grade gliomas, and BrMs originating from different primaries including breast cancer, lung cancer and melanoma (Figures 1A, 1B and S1A). This showed a leukocyte abundance from ~20–40% across the cancer samples. Stratification of CD45⁺ cells into myeloid and lymphoid lineages revealed a significant increase in myeloid cells in IDH mut and IDH wt gliomas, and of lymphocytes in IDH wt tumors and BrMs compared to non-tumor tissue (Figure 1B, p<0.05, one-sided Student’s t-test). We used multicolor fluorescence-activated cell sorting (FACS) to analyze 14 major immune cell populations across 100 clinical samples (Figure S1A, Tables S1 and S2), and collected cells for RNA sequencing (RNA-seq) from 48 patients (Table S3, full clinical annotation).

By incorporating cell lineage tracing and mouse models of high-grade gliomas and BrM, we previously identified the cell surface marker integrin alpha 4, ITGA4/CD49D, as a means to discriminate tumor-associated MG (T-MG) from tumor-associated MDM (T-MDM) (Bowman et al., 2016), which we integrated into clinical sample analyses herein. This enabled sorting of CD45− non-immune cells, CD49D^low MG, CD49D^high MDM, neutrophils, CD4⁺ and CD8⁺ T cells (Figure S1A, Tables S2 and S3A) for transcriptome analysis by RNA-seq. We assessed sorting fidelity by FCM re-analysis of the sorted CD49D^low and CD49D^high TAM populations (purity 98.4–99.8%), and by investigating the frequency of the canonical IDH codon 132 missense mutation in the RNA-seq reads from CD45− cells, CD49D^low and CD49D^high TAM populations. While we observed a mean mutated allele frequency of 0.43 in CD45− cells from IDH mut gliomas (range 0.3 – 0.61), this was very rare in TAMs (mean = 0.01, range 0.0 – 0.09), indicating a reliable separation of cell populations. In a t-distributed stochastic neighbor embedding (t-SNE) visualization of sorted populations, samples clustered mostly by cell type (Figure S1B) with gliomas and BrMs discernible as separate groups in the CD45− population.

In this global expression analysis within the context of the other major brain TME components, CD49D^low and CD49D^high TAM populations clustered closely, suggesting a broad transcriptomic similarity. We thus further interrogated the utility of CD49D to differentiate between TAM populations by analyzing association of MG- and MDM-specific ontogeny core gene sets, previously identified from lineage-tracing studies (Bowman et al., 2016), in human CD49D^low and CD49D^high cells sorted from non-malignant and brain cancer tissues. This revealed an enrichment of ontogeny core gene sets in the corresponding cell type (Figure 1C), demonstrating our ability to accurately distinguish MG and MDM in human samples across different disease entities. Interestingly, these core signatures were influenced within certain tumor types, with T-MDM showing an increased MG core gene set signal in IDH mut gliomas, and T-MG acquiring MDM features in BrMs, suggesting a tissue-dependent transcriptional programming of these cells, as further interrogated below.

We next assessed the landscape of intratumoral immune cell populations (Figure S1A and Table S2) using clustering analysis to identify patterns of cellular abundance (Figure 1D, Chi-square test for independence, p value < 0.0001). This revealed three major clusters: (1) non-tumor samples and IDH mut gliomas characterized by a dominance of MG with low numbers of other immune cells; (2) IDH wt gliomas, and several BrMs, with an influx of MDM and to some extent neutrophils into the tumor, while mostly excluding lymphocytes; (3) predominantly BrMs, and few IDH wt gliomas, exhibiting the most diverse immune cell landscape with a substantial infiltration of T cells and neutrophils. Certain tumors contained CD14^low/CD16⁺ non-classical monocytes, CD14⁺/CD16⁺ intermediate monocytes, CD16⁻ granulocytes, dendritic cells (DC), or immature myeloid cells. Across all samples the lymphocyte compartment was mostly composed of T cells with fewer NK cells and B cells.

Principal component analysis (PCA) of the relative abundance of all investigated populations confirmed that MG, MDM, neutrophils, CD4⁺ and CD8⁺ T cells are the major immune cell determinants of the brain TME landscape (Figure 1E). PC1 separated non-tumor tissue and IDH mut gliomas from IDH wt gliomas and BrMs, while PC2 distinguished IDH wt gliomas and BrMs. Further analysis stratifying for IDH status in gliomas and the primary tumor site in BrMs verified a substantially higher proportion of lymphocytes in BrMs (Figure 1F, mean_{lymphocytes %CD45}⁺ = 46.23 %, SEM = 4.15, t-test, p < 0.0001). Melanoma-BrMs exhibited the most abundant lymphocyte infiltrate with a sizeable CD8⁺ T cell fraction (mean_CD8⁺ _%CD45⁺ = 33.01 %, SEM = 5.82, one-way ANOVA, p < 0.01). Regulatory T cells (Tregs) were detected in certain BrMs (mean_{Treg %CD45}⁺=1.2 %, SEM = 0.36), while rare in gliomas (mean_{Treg %CD45}⁺=0.25 %, SEM = 0.05, t-test, p < 0.05).

Because of the prominence of T-MG and T-MDM in the myeloid compartment of brain malignancies we used IF staining and deconvolution analyses to independently validate their presence. Commonly employed MG markers such as P2RY12, TMEM119, and SALL1, and MDM-associated genes such as AHR and VDR, showed varying RNA expression levels across different brain malignancies while maintaining their cell type-specificity (Figure S2A), in a similar manner as observed for the ontogeny core gene sets (Figure 1C). An equivalent pattern was observed at the protein level (Figure S2B), where P2RY12 showed the highest expression in non-tumor tissue and CD68 was most abundant in BrM TAM populations. This necessitated the use of both markers, complemented with CD49D, to reliably identify MG and MDM in IF analyses (Figure S2C). We used this strategy to interrogate a cohort of non-tumor, glioma and BrM samples by whole-section quantification, confirming MDM accumulation in IDH wt gliomas and BrMs (Figure 2A–C). Furthermore, comparison of tissue processed independently for IF and FCM from the same individual samples demonstrated significant concordance (Figure S2D).

We queried the sorted cell populations for T-MG- and T-MDM-specific differentially expressed genes (DEG) that separate these two populations from the most abundant other cell types, i.e. CD45− cells, neutrophils and T cells (Figure S2E). Several of the genes highly expressed in T-MG are well-established MG markers (P2RY12, TMEM119, TAL1), while genes highly expressed in T-MDM include markers of alternative macrophage polarization (FCGR2B, CLEC10A) and DC-like phenotypes (CD1C, CD1B, CD207) with increased phagocytic- and antigen cross-presentation ability (CD209). These gene sets also allowed us to utilize a publicly available integrated dataset (Vivian et al., 2017) containing bulk expression data of healthy cortical brain tissue from the Genotype-Tissue Expression project (GTEx, The GTEx Consortium, 2013), and low- and high-grade glioma samples from The Cancer Genome Atlas (TCGA, Ceccarelli et al., 2016) in a bulk tissue transcriptome deconvolution approach (Racle et al., 2017). The estimates obtained of MG and MDM proportions in this external dataset (n=711 samples) verified the prevalence of MG in IDH mut gliomas and MDM enrichment in IDH wt gliomas (Figure 2D).

MG and MDM exhibit a multifaceted polarization phenotype in brain malignancies

We next employed PCA to specifically focus on TAMs and analyze genes whose expression was influenced by tissue type (i.e. reference MDMs, non-tumor brain, gliomas and BrMs) and cell type (i.e. MG and MDM) (Figure 3A). Within the first two PCs, MG and MDM projected into different spaces, with in vitro differentiated MDMs distinct from tissue-derived samples. We observed a gradient across PC1 with non-tumor brain tissue at one end, traversing IDH mut and IDH wt gliomas, and ending with BrMs. Thus, TAM transcriptomic changes are influenced by the brain TME per se and also by the specific type of malignancy.

We contrasted T-MG and T-MDM from BrMs or gliomas (regardless of IDH mutation status) to MG from non-tumor brain or in vitro differentiated MDM from healthy donors respectively (Figure S3A, Table S3A, Table S4). This revealed profound expression changes in both populations, with T-MDM exhibiting a higher magnitude in their transcriptional response compared to T-MG (Figure S3A). The intersect of DEGs in gliomas and BrMs was highest in T-MDM (Figure S3B), potentially reflecting the greater changes experienced by these cells upon entering the completely foreign environment of a brain tumor. This was also evident when focusing on genes upregulated in both glioma and BrMs that are private to either T-MG or T-MDM (Figure 3B). In both T-MG and T-MDM the number of shared genes was higher across different diseases than between these two cell populations within the same tumor type. Consequently, only a small number of genes (n=137) showed concordant upregulation across a comparison of all diseases and TAM types (Figure 3B).

To explore the underlying biological processes conserved in both gliomas and BrMs we examined the intersect of upregulated genes (Figure S3B) in T-MG or T-MDM using gene set over-representation analysis (ORA). In the Molecular Signature Database (MSigDB, Liberzon et al., 2015) “hallmark” collection of major biological categories both T-MG and T-MDM showed pathway enrichment in (i) modeling of the TME (“Angiogenesis”, “Hypoxia”), (ii) inflammation (“Inflammatory Response”, “Allograft Rejection”), and (iii) immune cell activation states (“TNFα Signaling via NFκB”, “Interferon α Response”, “Interferon γ Response”, “IL2 STAT5 Signaling” and “IL6 JAK STAT3 Signaling”) (Figure S3C).

We also assessed the M1 and M2 polarization status of T-MG and T-MDM using a panel of marker genes (Murray et al., 2014). However, no evident pattern emerged of a defined M1 vs. M2 phenotype in either glioma- or BrM T-MG or T-MDM (Figure S3D). To further explore the activation state of T-MG and T-MDM, we subjected their respective upregulated genes to ORA of macrophage stimulus-specific programs (Xue et al., 2014). This revealed a multifaceted response (Figure 3C) incorporating both canonical M1 (IFNγ) and M2 polarization (IL4), including expression changes associated with chronic inflammatory stimuli (TNFα + prostaglandin E (PGE₂), and TNFα + PGE₂ + Pam3CysSerLys4 (TPP)) and exposure to free fatty acids (oleic acid (OA), palmitic acid (PA)), which have been implicated in modulating myeloid cell function (Thapa and Lee, 2019). This indicates a diverse transcriptional programming of T-MG and T-MDM in gliomas and BrMs extending beyond a simple M1 vs. M2 polarization.

To understand which processes are linked to, and potentially driving, these responses we identified the gene set enrichment (GSEA, Subramanian et al., 2005) leading edge genes in T-MG and T-MDM in gliomas and BrMs, and clustered them into leading edge metagenes (LEMs) with non-negative matrix factorization (Godec et al., 2016). This identified up to 5 distinct LEMs per cell type and comparison that were tested for significant overlap in a pairwise fashion (Figure S3E), and annotated using Gene Ontology (GO) terms (Figure 3D). LEMs associated with mitosis and cell proliferation were present in both T-MG and T-MDM in gliomas and BrMs (Figure 3D, Group 1). The biological validity of these LEMs were verified by staining for Ki-67, a marker of cell proliferation, in non-tumor, glioma and BrM tissue sections (Figure 3E), showing increased proliferation in both T-MG and T-MDM in IDH wt gliomas and BrMs, and in T-MG in IDH mut gliomas.

Interestingly, LEMs enriched for Type I Interferon (IFN) signaling were detected in glioma and BrM T-MDM and in BrM T-MG, but not in glioma T-MG (Figure 3D, Group 2). Sustained Type I IFN signaling has been implicated in mediating immune suppression and ICB resistance (Benci et al., 2016). The stringency of these Group 2 LEMs was validated by building a protein-protein interaction (PPI) network of the shared LEM genes (Figure S3F). Beyond their role in antiviral responses, the genes highlighted at the center of the PPI network (Figure S3F, red nodes) have been implicated in a variety of both tumor promoting and suppressing roles (Benci et al., 2016). Similarly, the more peripheral network nodes IL15 and TNFSF10 are potentially able to modulate an effective immunological anti-tumor response or induce apoptosis in cancer cells respectively (Bouralexis et al., 2005; Santana Carrero et al., 2019). We asked if these genes were directly induced by secreted factors in the brain TME, and established cell-based assays to expose MDMs to TME-conditioned medium (CM) generated from single cell suspensions of freshly-isolated glioma or BrM samples in culture. All genes analyzed were upregulated by BrM-TME-CM, and to a lesser extent by Glioma-TME-CM (Figure 3F). We also detected induction of Inflammation- and NFκB-signaling-associated LEMs in BrM-MG, glioma-MDM and BrM-MDM (Figure 3D, Group 3). LEMs that point towards a Th17 response (Group 4), and the recruitment of immune cells and interactions between different immune cell compartments were exclusively detected in MDM (Group 5). Collectively, these analyses reveal an acquisition of a multifaceted activation state of MG and MDM upon their integration into the TME of brain malignancies.

IDH mutation status associated with changes in glioma TAM activation

We next asked whether MG and MDM occupy distinct regions within the TME of IDH wt gliomas. Spatial analysis of tissue sections showed significant enrichment of both populations in the perivascular niche (Figures 4A and S4A). Analysis of their distribution relative to CD31⁺ vascular structures showed a closer proximity of T-MDM compared to T-MG (Figures 4B and S4A). Interrogation of anatomical transcriptome data from the Ivy GAP study (Puchalski et al., 2018) also demonstrated an enrichment of T-MDM in the microvascular compartment (Figure S4B). This enrichment coincided with CD4⁺ and CD8⁺ T cells, indicating further spatial TME organization in IDH wt gliomas.

We assessed if the distinct T-MG and T-MDM distribution and cell number is paralleled by their activation state. In the LEM analysis we had detected a Type I IFN response in glioma MDM but not MG (Figure 3D); we therefore queried the FCM data to analyze levels of MHC class II HLA-DR expression. This showed significantly increased HLA-DR in T-MDM compared to T-MG in both IDH mut and IDH wt tumors (Figure 4C). We screened the associated RNA-seq data for antigen processing and presentation pathway gene sets using GSEA and gene set variation analysis (GSVA) (Figure 4D). Interestingly, we found evidence for increased expression of MHC class II antigen presentation gene sets in IDH wt glioma MDM, and also antigen processing-associated pathways (Figure S4C), and MHC class I presentation gene sets (Figure 4D). While these findings suggest the potential of TAMs, particularly T-MDM, to initiate an immune response, this potential is generally not realized in the glioma TME - based on the current status of ICB trials in this disease, and we thus asked if there was also evidence of pro-tumor states in these cell populations.

We compared T-MG and T-MDM from IDH wt gliomas to T-MG from IDH mut gliomas, as they constitute the most abundant TME cell types in these tumors respectively (Figure 1F and 2C, Table S5). This revealed 489 DEGs in T-MG (Figure 4E, Table S5, 406 up-, 83 down-regulated), and 1478 DEGs in T-MDM (Figure 4F, Table S5, 903 up-, 575 down-regulated). While these gene lists were generated by comparing T-MDM from IDH wt gliomas vs. T-MG from IDH mut gliomas, they similarly separated T-MDMs in IDH mut vs. IDH wt disease in a clustering analysis (Figure 4F), indicating they indeed reflect T-MDM alterations based on the IDH status of the tumor. 421 genes exhibit a similar pattern across both TAM cell types (343 up-, 78 down-regulated), suggesting that T-MG and T-MDM can also acquire a common transcriptional pattern in IDH wt tumors. Among the shared genes were several encoding extracellular matrix (ECM) proteins (Figure 4G, FN1 and VCAN) and ECM-associated matricellular proteins (THBS1, TGFBI, LGALS3 and ANGPTL4), that regulate the availability of ECM-sequestered ligands, angiogenesis, and tumor immunity (Mushtaq et al., 2018). This suggests that TAMs help shape the composition and effector functions of ECM proteins in IDH wt tumors. We also found the anti-inflammatory molecules ANXA1 and GPNMB (Figure 4G), previously implicated in pro-tumorigenic macrophage polarization and inhibition of T cell activation (Kobayashi et al., 2019; Ripoll et al., 2007), upregulated in T-MG and T-MDM.

We next investigated inflammation mediators within the CD45− population of IDH wt tumors in parallel with their corresponding receptors in TAMs. TGFB2 expression was elevated compared to IDH mut CD45− cells and the accessory TGFβ receptor ENG was highly expressed in IDH wt TAMs (Figure 4H). TGFB2 has pleiotropic effects in inflammation and tissue remodeling during wound healing, and has been implicated in an autocrine signaling loop in glioblastoma cells (Rodon et al., 2014). The neuroinflammatory cytokine MDK, which modulates TAM polarization to a M2-like phenotype in glioma (Meng et al., 2019), was upregulated in CD45− cells from IDH wt tumors, and its receptors SDC4 and ITGA4/CD49D were differentially expressed in T-MDM versus T-MG (Figure 4H), suggesting cell type-specific effects of this inferred signaling loop.

We asked if a T-MDM-specific gene set generated from IDH wt gliomas was associated with a survival difference in patients. By logistic regression we derived a representative signature, consisting of 36 genes (Figure S4D), from the total number of genes upregulated in TAMs in brain malignancies (Figure 3B). This included the macrophage marker RUNX3, the atypical chemokine receptor ACKR3 which can regulate CXCL12-CXCR4 signaling, the ER-stress protein HERPUD1 and the inhibitory Fc receptor FCGR2B, which can modulate macrophage activation (Bournazos et al., 2016; Li et al., 2018), and the cytokine IL19 that affects angiogenesis and macrophage polarization (Richards et al., 2015). The signature was used to classify patients in a merged TCGA dataset of low- and high-grade gliomas (Figure 4I, Figure S4E). In IDH mut patients a decrease in median overall survival was associated with enrichment of the T-MDM IDH wt signature, while IDH wt patients with a low enrichment score showed increased survival. This was confirmed in a multivariate Cox proportional hazard model that included the transcriptomic glioma subtypes (as annotated in the TCGA dataset) and IDH status (Figure 4J). To verify this effect did not simply reflect changes in T-MDM number, we classified the TCGA cohort based on the enrichment of the T-MDM-specific gene set used for deconvolution, which showed low impact on survival (Figure S4F).

In light of disappointing outcomes from PD1/PDL1 ICB trials in glioblastoma to date, we queried whether the abundant T-MG and T-MDM could contribute to the limited therapeutic efficacy. We performed ORA of a panel of 20 gene sets previously associated with “innate anti-PD1 resistance” (IPRES, Hugo et al., 2016) in the TAM DEGs of IDH wt gliomas, and found a sizeable fraction to be upregulated in both T-MG and T-MDM (Figure S4G). We then included the CD45− population and interrogated enrichment of IPRES gene sets on the single sample level by GSVA (Figure S4H). This yielded a diverse picture, with both tumor cells and TAMs enriched for IPRES gene sets to varying degrees. Therefore, both TAMs and CD45− cells from IDH wt gliomas may contribute to mediating innate ICB resistance.

The immune contexture influences the tumor microenvironment on a global level

Through the integrated analysis of protein and gene expression data we next explored the impact of immune cell infiltration. From 200 inflammation-associated proteins assessed, 55 were differentially detected in our sample cohort (clinical information, Table S3B). Unsupervised clustering analysis revealed distinct clusters with abundant inflammatory proteins in tumors (Figure 5A). The profile of IDH wt gliomas and BrMs showed a sizeable overlap (protein cluster 1), encompassing angiogenic factors (VEGFA, ANG), growth factors (PDGFA, TGFB1, SPP1 and GDF15), several proteases and protease inhibitors (SERPINE1, CTSS, TIMP1), the proteolysis cascade regulator PLAUR, and the cytokines CCL2 and CCL5 (Figure 5A, Figure S5A). However, we also found distinct protein patterns between gliomas and BrMs. The neurotrophic growth factor FGF2 and neuronal cell adhesion molecules including ALCAM, which regulates immune cell infiltration during neuroinflammation (Lecuyer et al., 2017), were highly expressed in non-tumor brain, IDH mut and IDH wt samples (protein cluster 3, Figure S5A). Conversely, BrM samples had abundant immune-regulatory molecules affecting myeloid and lymphocytic cells and their heterotypic signaling (protein cluster 2, Figure S5A), such as CD40L, IL6R, INHBA and AREG (Morianos et al., 2019; Zaiss et al., 2015) possibly reflecting the greater immune cell diversity in BrMs. This orthogonal dataset reinforces the RNAseq analyses showing inflammatory signaling pathways are highly enriched in brain tumors.

We integrated the cell type-specific RNA-seq data and bulk protein data to distinguish proteins with a more restricted expression versus those that are expressed across a range of cell types. Transcriptome data from CD45− cells, TAMs, neutrophils, CD4⁺ and CD8⁺ T cells from all tumor samples was clustered using a self-organizing map (SOM). This yielded 6 SOM spots, i.e. metagenes of co-expressed genes (Figure 5B), which recapitulated the respective cell lineages (Figure S5B). The CD45− populations were assigned to three distinct spots which were associated with more aggressive IDH wt gliomas and BrMs (spot VI), or reflected the brain-intrinsic or -extrinsic tumor origin (spots I and V). These cell type-associated SOM spots overlapped considerably with the protein data (30/55 proteins, Fisher’s exact test, p < 0.0001, Figure 5C). While VEGFA, ANG and TGFB1 were expressed by diverse cell types in gliomas and BrMs, other genes such as GDF15 and IGFBP2 showed a more CD45− cell-restricted expression (Figure 5D). The significant contribution of TAMs to production of key inflammatory proteins including SPP1 and IHNBA is reflected by TAM SOM spot III constituting the largest group of proteins with cell type-specific expression (Figures 5C and 5D).

Myeloid cells show a distinct phenotype in BrMs

Our global analysis juxtaposing the expression patterns of TAMs in gliomas (regardless of IDH status) to BrMs showed disease- and cell type-specific transcriptomic changes. We thus explored BrM-specific alterations by focusing on genes upregulated only in relation to both the corresponding reference and to IDH wt gliomas (Figure S6A and S6B, Table S6). Various cytokines, chemokines and pro-inflammatory molecules were elevated in both BrM-MG and BrM-MDM (Figure 6A), including potent mediators of autoimmune neuroinflammation CSF2 and IL23A (Zhao et al., 2017), and the pattern-recognition receptor MARCO. Intriguingly, antibody-mediated MARCO-targeting in extracranial tumors increases M1-like polarization of TAMs and enhances ICB efficacy (Georgoudaki et al., 2016). These effects relied on interaction of the antibody with FCGR2B, which is also part of the T-MDM IDH wt signature (Figures S2E and S4C). Finally, RETN, which is involved in systemic inflammatory disorders (Filkova et al., 2009), was upregulated in BrM-TAMs (Figure 6A).

Analysis of individual BrM-TAM populations uncovered distinct expression patterns. BrM-MG showed restricted upregulation of IL6 (Figure 6A), which exerts immunosuppressive effects on T cell function in cancer and mediates ICB resistance (Tsukamoto et al., 2018), and the receptor TREM1 which modulates pro-inflammatory responses both in MG and systemically in myeloid cells during neuroinflammation (Liu et al., 2019; Xu et al., 2019). Among upregulated chemokines, we found increases in both TAM cell types (CCL23), and either BrM-MG- (CXCL5, CXCL8) or BrM-MDM-restricted increases (CCL8, CCL13, CCL17 and CCL18) (Figure 6A). These results reveal distinct contributions of TAM populations to the inflammatory TME milieu in a disease-specific manner.

GSEA identified additional cell type-specific enrichment patterns: BrM-MG showed evidence of IL6 pathway activity (Figure S6C), and in BrM-MDM, the “Naba core matrisome” gene set was significantly enriched (Figure S6D). This prompted us to assess expression of genes encoding ECM and matricellular proteins in BrM-MDM vs. BrM-MG, which revealed genes encoding matrix proteins including type III and IV collagens, FN1, the proteoglycans LUM and OGN and the matricellular proteins ECM1, SPARC, and SPARCL1 as highly expressed in BrM-MDM (Figure 6B). While ECM remodeling has been implicated in tumor progression, LUM, OGN, SPARCL and SPARCL1 exhibit both pro- and anti-metastatic properties which underscores the complex context-dependent role of the ECM (Kai et al., 2019). We also found high expression of cathepsin proteases CTSB and CTSW in BrM-MDM (Figure 6B), which participate in multiple tumor-promoting processes including invasion and metastasis (Olson and Joyce, 2015). The hyaluronan receptor HMMR, involved in macrophage chemotaxis and fibrosis in lung injury (Cui et al., 2019), was also higher in BrM-MDM (Figure 6B). Together these data suggest that the ECM is not only shaped by macrophages at the primary site (Afik et al., 2016), but that T-MDM may also play a pivotal role in ECM niche construction in BrM that is distinct from IDH wt gliomas (Figure 4G).

Given the upregulation of CXCL8, a key neutrophil chemoattractant, by BrM-MG (Figure 6A) we explored the TME contribution to recruitment of neutrophils, which were highly abundant in BrM (Figure 1F). Analysis of major neutrophil-recruiting chemokines and their receptors showed broad expression across all interrogated myeloid cells (Figure S6E). To explore the phenotype of BrM-associated neutrophils we queried the RNA-seq data which revealed BrM-specific upregulation of ITGA3 (Figure 6C), which is involved in neutrophil tissue infiltration in sepsis, and CXCL17 previously implicated in neutrophil- and macrophage recruitment in cancer (Li et al., 2014). We also observed upregulation of the adenosine receptor ADORA2A (Figure 6C), that attenuates the phenotype of pro-inflammatory neutrophils (Barletta et al., 2012). Furthermore, we found increased expression of CD177 (Figure 6C), a cell surface receptor that modulates neutrophil migration and activation and serves as a marker for PR3-positive neutrophils that in turn negatively affect T cell proliferation (Yang et al., 2018). Notably, MET, which has been linked to recruitment of immunosuppressive neutrophils in cancer (Glodde et al., 2017), was upregulated in neutrophils in a BrM-specific manner (Figure 6C). In sum, we have uncovered multiple disease-specific alterations of myeloid cells extending beyond BrM-TAMs to neutrophils, which has potential implications for the recruitment and activation of other cell types within the TME, including T cells.

TAMs are poised towards an immunomodulatory phenotype in BrMs

While we found a significant accumulation of CD4⁺ and CD8⁺ T cells in BrM vs. IDH wt gliomas by FCM, this analysis of dissociated tissue samples lacks structural information. We thus performed neighborhood analysis of IF-phenotyped IDH wt and BrM tissue sections to elucidate whether there is a spatial relationship between TAMs and CD3⁺ T cells in BrM. In IDH wt gliomas, both T-MG and T-MDM mostly neighbored homotypic cells while lacking T cells in their close vicinity (Figure 7A–B, S7A), possibly reflecting the general T cell sparseness in these tumors. By contrast, both TAM populations neighbored T cells far more frequently in BrM, indicating the potential for interaction (Figure 7A–B, S7A).

We thus investigated the T cell activation state in BrMs in relation to unmatched healthy donor blood and also juxtaposed them to the corresponding populations from IDH wt gliomas. Compared to controls, CD4⁺ T cells from BrM showed evidence of a hyporesponsive, anergic phenotype (Figure 7C) while CD8⁺ T cells exhibited an exhaustion signature (Figure 7D) which usually occurs upon chronic activation, resulting in upregulation of inhibitory receptors. These defective T cell states can be caused by aberrant activation or T cell inhibition by both tumor cells and antigen-presenting cells in the TME and constitute a major obstacle in treating cancers,.

To delineate putative mechanisms in the BrM TME that may drive these alterations, we probed the RNAseq data from CD45− cells, TAMs, and T cells (Figure 7E) for expression of activating and inhibitory immunomodulatory signals (Wei et al., 2018). This revealed upregulation of various canonical T cell activators and co-activators but also mediators of inhibition in T cells (PDCD1/PD1, CD28, CTLA4), while T cell-inhibiting and activating signals were detected in both TAM populations (CD274/PD-L1, PDCD1LG2/PD-L2). Notably, we found upregulated CD80, which has diverse roles in T cell activation as it heterodimerizes with CD274, provides co-stimulatory signals to T cells via CD28, and exerts inhibitory effects via interaction with CTLA4 (Zhao et al., 2019), in both TAM populations when compared to their normal references and IDH wt tumor populations (Figure 7E). The potential contribution of TAMs to metabolic immune evasion is also suggested by high expression of IDO1 and IDO2 (Zhai et al., 2018) in BrM (Figure 7E).

We investigated additional immunomodulatory mediators using weighted gene correlation network analysis (WGCNA, Langfelder and Horvath, 2008), and correlated the resulting expression patterns with paired FCM abundance of CD4⁺ and CD8⁺ cells in a disease and cell type-specific manner. We identified 15 unique co-expression modules showing significant correlation (p < 0.05) of their eigengenes (i.e. the first PC of the module expression data) with any of the provided sample traits (Figure S7B). Among these the “brown” WGCNA module correlated with both a specific BrM-MDM annotation and CD4⁺ T cell abundance. ORA of this module revealed signals for pathways such as coagulation and ECM modulation (Figure S7C) that impact the availability and activity of growth factors and cytokines within the TME (Mohan et al., 2020). We ranked genes by module membership strength, and correlation with CD4⁺ T cell abundance, which identified several factors with opposing immunomodulatory functions (Figure 7F). While the receptors CD300E and BST1 promote monocyte motility and survival (Isobe et al., 2018; Ortolan et al., 2019), we also detected effectors of immunosuppression such as actin-associated regulatory protein CNN2 which negatively regulates macrophage motility and phagocytic activity (Huang et al., 2008). The leukocyte immunoglobulin-like receptor subfamily B members LILRB2 and LILRB3, which attenuate myeloid cell activation (van der Touw et al., 2017), are also highly ranked genes within this module. Interestingly, LILRB2 was identified as a novel myeloid immune checkpoint that limits antitumor immunity (Chen et al., 2018). We also found evidence for effects on T cells as CD52, which in its soluble form inhibits T cell function, was among the BrM-MDM module genes. The notion that BrM-MDM undergo disease-specific alterations distinct from the primary extracranial tumor is supported by the upregulation of these genes (Figure 7G) in our analysis of an external cohort of BrM samples compared to their matched primary tumor tissue (Vareslija et al., 2019).

Resource availability:

Experimental model and subject details

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Informed consent was obtained from all individual participants included in this study. The collection of non-tumor and tumor tissue samples at the Centre Hospitalier Universitaire Vaudois (CHUV, Lausanne, Switzerland) was approved by the Commission cantonale d’éthique de la recherche sur l’être humain (CER-VD, protocol PB 2017–00240, F25 / 99). Sample collection at Memorial Sloan Kettering Cancer Center (MSKCC, New York, NY, USA) was approved by the institutional review board (IRB, protocols #IRB #06–107, #14–230). Non-tumor samples of cerebral cortex tissues were collected at CHUV during medically indicated surgical treatment of refractory epilepsy patients, and at MSKCC in normal brain distant from the tumor in patients with low-grade glioma or from post-mortem samples collected through the rapid autopsy program with no history of brain malignancy.

Tissue specimens were immediately collected from the operating room and processed as described below. All patient-related data and unique identifiers were removed so that human samples were anonymized before any further processing.

Pathological review and molecular analysis of tumor samples was performed as part of standard clinical care at the respective locations (CHUV or MSKCC). In all glioma samples subjected to RNA sequencing, the IDH1 and IDH2 mutation status was verified by inspection of the reads from the CD45− population aligning to the IDH1 and IDH2 loci with the Integrative Genomics Viewer (IGV, Robinson et al., 2017). For immunofluorescence sections the tumor diagnosis was confirmed independently, for all non-tumor samples, the absence of malignancy was equally confirmed by a pathologist.

Peripheral blood and buffy coats were obtained from the Transfusion Interrégionale, Croix-Rouge Suisse (Epalinges/Lausanne, Switzerland), the New York Blood Center (New York, NY, USA), and healthy donors.

Method Details

Quantification and statistical analysis

Summary data are presented as mean ± standard error of the mean (SEM) or Tukey boxplots using “ggplot2”. Numerical data was analyzed using the statistical tests noted within the corresponding sections of the manuscript. Hierarchical clustering was performed using Ward’s method with 1-Pearson correlation coefficient as the distance metric unless noted otherwise. P values were annotated as follows: * <0.05, ** <0.01, *** <0.001, **** <0.0001, ns >0.05.