Targeting Microglial Metabolic Rewiring Synergizes with Immune-Checkpoint Blockade Therapy for Glioblastoma

Glioblastoma (GBM) constitutes one of the most lethal cancers (1, 2). Despite multimodal treatment, the median survival time of GBM patients under 70 remains only 12 to 18 months (3). Although the molecular basis for gliomagenesis has been extensively investigated, targeted therapies have largely failed in clinical trials (4, 5). Although immunotherapies have improved patient outcomes in other cancer types, numerous immune-oncology efforts—cell-based therapies, oncolytic viruses, immune-checkpoint inhibition, etc.—have offered limited benefit against GBM in clinical trials (6–9). As the brain immune environment is unique and GBMs display low tumor mutational burden (10), modeling the specific regulation of immune responses in the normal and neoplastic brain has been a high priority to design targeted therapies to improve GBM care through combinatorial paradigms (11).

Microglia are primary immune cells in the brain, uniquely derived from yolk sac hematopoietic stem cells during development (12), whereas systemically derived macrophages serve both overlapping and complementary functions (13–15). In GBM, microglia and macrophages constitute the major nontumor stromal cells and impact on tumor growth (16, 17). Microglia enhance GBM growth, at least in part, by regulating mTOR-mediated immunosuppression of the tumor microenvironment (18). Therefore, microglia have been proposed as a potential target for GBM immunotherapies (19–21).

Metabolic reprogramming of tumor and stromal cells is prevalent in many cancers (22, 23). A key histologic criterion of GBM is necrosis, which is characterized by immune infiltration, hypoxia, and nutrient restriction. Single-cell analyses have revealed that glioma cells under environmental stress display epigenetic plasticity (24). The tumor immune microenvironment (TIME) in many cancers reflects the effects of hypoxia, nutrient deprivation, and other stressors, suggesting cross-talk between tumor metabolism and immune responses through immunometabolic rewiring (25–27). Oxidative stress plays a critical role in a wide spectrum of cancers (28, 29). For example, oxidative stress facilitates the transition of breast cancer stem cells from a ROS^lo mesenchymal subtype to a ROS^hi epithelial state (30). Oxidative stress induces an immunosuppressive tumor milieu; for example, tumor regulatory T cells (Treg) undergo apoptosis to impair spontaneous and PD-L1 blockade–mediated antitumor T-cell immunity (31). Treg apoptosis derives from weak nuclear factor erythroid 2–related factor 2 (NRF2)–associated antioxidant responses and high vulnerability to free oxygen species in the TIME. However, the value of targeting oxidative stress–induced TIME immunometabolic reprogramming as a potential Achilles’ heel for GBM, either as a monotherapy or combined with immune-checkpoint blockade, remains unresolved. Here, we probed the immune landscape of gliomas to identify targetable molecular mechanisms that could be interrogated as novel therapeutic approaches in the management of GBM.

Based on the hypothesis that the TIME landscape in GBM informs therapeutic targets, we interrogated a syngeneic murine glioma in vivo model to characterize the TIME using comparative single-cell RNA sequencing (scRNA-seq) analysis between tumor-bearing and tumor-free brains (Fig. 1A and B). Similar to previous reports (32), microglia constituted the majority of CD45⁺ cells in tumor-bearing brains (Fig. 1C) and the cluster identified as microglia expressed microglial markers, including Tmem119, Cx3cr1, P2ry12, and Crybb1, at higher levels compared with other cell types (Supplementary Fig. S1A; ref. 33). Among the CD45⁺ cells, eight clusters of microglia with different functions were identified in glioma-bearing mice and control, respectively (Supplementary Fig. S1B and S1C). We performed Venn diagram analysis to identify overlapping gene ontology (GO) terms among the differential gene expression of the microglia cluster in the single-cell analysis of our GL261 scRNA-seq dataset as well as two glioma patient sample single-cell analysis datasets (GSE117891 and GSE162631; refs. 34, 35; Fig. 1D). Nine GO terms overlapped among the three datasets, with “Response to oxidative stress” ranked topmost (Fig. 1E). The upregulated expression of oxidative stress genes was more prominent in microglia compared with other immune cell types in the glioma microenvironment (Supplementary Fig. S1D). Most microglia clusters in glioma-bearing mice displayed an elevated expression of an oxidative stress gene signature compared with controls (Fig. 1F). In addition, reanalysis of public dataset GSE162631, which contains scRNA-seq analysis of freshly isolated cells in four GBM patients from the tumor core and paired surrounding peripheral tissues, showed that oxidative stress gene expression was elevated in most microglia clusters of the tumor cores compared with that of tumor peripheral tissues (Supplementary Fig. S1E and S1F). In addition, we isolated microglia from tumor tissues and surrounding healthy brain parenchyma in GL261-bearing mice and sent them for RNA-seq analysis (Supplementary Fig. S1G); oxidative stress gene expression was elevated in the microglia from tumor tissues compared with surrounding healthy brain parenchyma (Supplementary Fig. S1H). These data suggested that tumor-infiltrating microglia within gliomas endure oxidative stress.

As oxidative stress has been associated with the induction of oncogenic effects in multiple cancer types (36–38), we interrogated the functional impact of oxidative stress on the function of microglia. We isolated primary microglia from the cerebral cortices of neonatal C57BL/6 mice, treated them with hydrogen peroxide (H₂O₂) to induce oxidative stress, and then measured their effects on tumor cell growth in coculture experiments (Fig. 1G). H₂O₂-treated primary murine microglia enhanced tumor cell proliferation and migration compared with control cells (Fig. 1H), which was attenuated by the coadministration of N-acetyl-cysteine (NAC), a potent antioxidant. In validation experiments, H₂O₂ treatment enhanced oncogenic effects of murine BV2 microglia cocultured with murine GL261 glioma cells (Fig. 1I) and human HMC3 microglia cells cocultured with patient-derived GSC2907 human GBM cells (Fig. 1J), as measured by cell proliferation and migration, compared with the control groups. These effects were impaired by coadministration with NAC (Fig. 1I and J). Similar results were also observed in the functions of primary microglia from the cerebral cortices of adult C57BL/6 mice (Supplementary Fig. S1I). Moreover, we cocultured GL261 cells with microglia isolated from either control brains or GL261-bearing brains and found that similar to glioma growth-promoting functions of H₂O₂-treated microglia in vitro, microglia derived from GL261-bearing brains enhanced GL261 growth compared with that from normal brains (Supplementary Fig. S1J). However, other myeloid cells resident in the central nervous system, such as neutrophils, did not show growth-promoting effects for glioma, even after oxidative stress challenge (Supplementary Fig. S1K). Collectively, our results support that oxidative stress reprograms microglia toward an alternatively activated state, which promotes the proliferation and aggressiveness of glioma cells.

To investigate the impact of oxidative stress on microglial function in vivo, primary microglia were treated with H₂O₂ and then coimplanted with murine glioma cells that had been stably transfected with a firefly luciferase reporter gene (GL261-Luc or CT2A-Luc cells) into syngeneic hosts (Fig. 2A). The elevated ROS levels of microglia lasted for at least 13 days after coimplantation with murine glioma cells in vivo (Supplementary Fig. S2A and S2B). The apoptosis-inducing effects of H₂O₂ treatment on microglia were minimal (Supplementary Fig. S2C), suggesting that peroxide stress altered tumor growth through cell–cell interactions. H₂O₂-pretreated primary microglia cells induced more rapid tumor growth than either tumor cells implanted alone or tumor cells together with untreated primary microglia cells in vivo (Fig. 2B and C). In parallel to in vitro results, NAC treatment reversed the glioma growth-promoting effects of H₂O₂-pretreated primary microglia (Fig. 2B and C). Differences in tumor size reflected differential survival times of tumor-bearing mice; mice bearing glioma cells together with H₂O₂-pretreated primary microglia had the shortest survival compared with mice bearing tumor cells alone or with untreated primary microglia cells (Fig. 2D). NAC treatment selectively attenuated the survival effects of H₂O₂-pretreated primary microglia (Fig. 2D). CT2A murine glioma cells displayed similar results to GL261 cells (Fig. 2E–G), suggesting that these results were not model specific.

We hypothesized that oxidative stress induced microglia to reprogram the TIME, specifically tumor-infiltrating T cells. We examined the functional states of T-cell subsets in the context of oxidative stress–challenged primary microglia in vivo (Fig. 2H). Tumors grown in mice coimplanted with H₂O₂-treated primary microglia and GL261-Luc cells displayed reduced numbers of CD8⁺ T cells and increased numbers of CD4⁺ T cells (but not Tregs) compared with tumors from mice coimplanted with control primary microglia and GL261-Luc cells (Fig. 2I–K; Supplementary Fig. S2D). Cotreatment with NAC blocked the changes in T-cell numbers, supporting an on-target effect of oxidative stress (Fig. 2I–K). H₂O₂-treated primary microglia increased expression of the immune-checkpoint marker PD-1 on CD8⁺ T cells in tumor-bearing mice compared with mice coimplanted with tumor cells and control microglia, which was inhibited by cotreatment with NAC (Fig. 2L and M). IFNγ was also reduced on CD8⁺ T cells in the mice bearing tumor cells coimplanted with H₂O₂-treated primary microglia compared with mice bearing tumor cells and control microglia, which could be reversed by cotreatment with NAC (Fig. 2N). In CT2A in vivo allografts, tumors grown in mice coimplanted with H₂O₂-treated primary microglia and CT2A-Luc cells displayed reduced numbers of CD8⁺ T cells and nonsignificant changes of CD4⁺ T cells compared with tumors from mice coimplanted with control primary microglia and CT2A-Luc cells (Fig. 2O). Cotreatment with NAC blocked the changes in CD8⁺ T-cell numbers (Fig. 2O). H₂O₂-treated primary microglia increased the expression of PD-1 on CD8⁺ T cells in tumor-bearing mice compared with mice coimplanted with tumor cells and control microglia, which was inhibited by cotreatment with NAC (Fig. 2P). IFNγ was reduced in CD8⁺ T cells in the mice bearing tumor cells coimplanted with H₂O₂-treated primary microglia compared with mice bearing tumor cells and control microglia, which could be rescued by cotreatment with NAC (Fig. 2P). These results demonstrate that oxidative stress–mediated immune reprogramming of the brain TIME favors tumor growth.

Based on the effects of oxidative stress on the TIME, we hypothesized that antigen presentation is influenced by oxidative stress. Therefore, we generated a glioma line transfected with the ovalbumin gene (GL261-OVA) to model specific in vivo antigen presentation in OT-I [C57BL/6-Tg(TcraTcrb)1100Mjb/J] mice, which results in MHC class I–restricted, ovalbumin-specific, CD8⁺ T-cell responses (Fig. 3A). Primary microglia induced the proliferation of CD8⁺ T cells in vitro, confirming antigen presentation functions of oxidative stress–challenged microglia (Fig. 3B). However, H₂O₂ treatment impaired antigen presentation by microglia, which was partially rescued by NAC treatment (Fig. 3B). Secretion of IFNγ and TNFα by CD8⁺ T cells in OT-I mice was reduced in cocultured H₂O₂-pretreated primary microglia and GL261-OVA compared with cocultured untreated primary microglia and GL261-OVA (Fig. 3C). These effects were reduced by NAC cotreatment (Fig. 3C).

We next investigated the biological functions of microglia in the TIME under oxidative stress. Oxidative stress reduced the expression of major histocompatibility complex I (MHC-I) in primary murine microglia, which was rescued after cotreatment with NAC (Fig. 3D). To identify the cellular mechanisms through which these effects were mediated, we analyzed the effects of oxidative stress on the expression of markers for classically activated microglia (CD86) and alternatively activated microglia (CD206) in primary microglia derived from GL261-Luc tumor–bearing brains (Fig. 3E). Oxidative stress increased CD206 expression (Fig. 3F and G), whereas CD86 expression was reduced (Fig. 3H and I). These effects could be rescued by cotreatment with NAC (Fig. 3F–I). Similar results were also observed in CT2A-Luc tumor–bearing mice (Supplementary Fig. S2E and S2F).

We next examined the expression of several myeloid-related functional genes in microglia. In murine BV2 microglia and human HMC3 microglia, oxidative stress downregulated the expression of NOS2 and IL6 (markers of classic activation of myeloid cells) and upregulated the expression of ARG1 and IL10 (markers of alternative activation of myeloid cells), which was rescued by cotreatment with NAC (Supplementary Fig. S2G and S2H). Next, we isolated primary microglia from normal murine brains and assessed the impact of oxidative stress on microglia plasticity (Supplementary Fig. S2I). Similar to the results from GL261-Luc–bearing mice, oxidative stress induced the expression of CD206 in primary microglia (Supplementary Fig. S2J). Expression analysis revealed that oxidative stress decreased mRNA levels of immune activation–related genes, including TNFα, IFNγ, and IL1β, and increased the expression of immune suppression–related genes, including IL10, CXCL1, CXCL5, CXCL12, CX3CL1, and CCL22, in primary murine microglia (Supplementary Fig. S2K). Comparative RNA-seq of control and H₂O₂-treated BV2 microglia as well as control and H₂O₂-treated primary microglia isolated from glioma-bearing mice revealed that oxidative stress induced a specific transcriptional pattern related to either immunoactivation or immunosuppression (Supplementary Fig. S2L). Oxidative stress induced biological processes that included angiogenesis, ROS metabolic process, and response to TGFβ, as well as immune activation–related processes, such as positive regulation of leukocyte proliferation, positive regulation of lymphocyte activation, and cellular response to TNF (Supplementary Fig. S2M). However, oxidative stress primarily induced the upregulation of immune activation–related genes, including Nos2, IFNγ, and IL1β, and downregulation of immune suppression–related genes, including Arg1, TGFβ1, CXCL1, CXCL5, and CX3CL1, in neutrophils (Supplementary Fig. S2N).

These findings led us to explore the effects of oxidative stress–challenged microglia on T cells in the brain by coculturing CD8⁺ T cells with microglia in the presence of OVA antigen ex vivo (Fig. 3J), which increased PD-1 expression with microglia that were pretreated with H₂O₂ compared with control conditions, and this increase was blocked by NAC cotreatment (Fig. 3K and L). Thus, oxidative stress induces microglia toward an alternatively activated state and disrupts their antigen presentation functions for CD8⁺ T cells.

To identify key regulators in the phenotypic switch of oxidative stress–challenged microglia in GBM, we performed chromatin immunoprecipitation enrichment analysis master regulator discovery (39) to identify potential transcription factors mediating transcriptional alterations in microglia. A systems network biology approach was used to construct the transcription factor–based regulatory network, which identified 15 transcription factors potentially causal in the transcriptome-wide alterations in oxidative stress–challenged microglia (Fig. 4A). We prioritized nuclear receptors, as they are more amenable to therapeutic targeting. Among the 15 transcription factor candidates, only two were nuclear receptors, hepatocyte nuclear factor 4α (HNF4α) and NR4A2 (Fig. 4B). Validation with quantitative PCR (qPCR) supported only Nr4a2 as upregulated in microglia upon oxidative stress treatment (Fig. 4C), with NAC cotreatment downregulating Nr4a2 levels (Supplementary Fig. S3A). We next checked the expression changes of Nr4a2 in our RNA-seq data of both the BV2 cell line and primary microglia. Nr4a2 expression was increased after H₂O₂ challenge in BV2 compared with control (Fig. 4D), and, similarly, Nr4a2 expression was also increased in primary microglia isolated from mouse glioma samples compared with that from surrounding healthy parenchyma (Fig. 4E). Interrogation of in silico scRNA-seq analysis of human patient GBM surgical specimens (GSE84465; ref. 40) revealed that a population of myeloid cells displayed high expression of NR4A2 (Fig. 4F). Further reanalysis of both our GL261 scRNA-seq analysis data and GSE162631 human glioma scRNA-seq analysis data (35) demonstrated that NR4A2 expression was upregulated in the microglia of tumor samples compared with control brain (Fig. 4G and H). We next examined the expression of NR4A2 in the microglia of clinical samples and found that the abundance of NR4A2⁺ microglia was higher in GBM (grade 4 glioma) patient tumor tissues compared with that in the nonneoplastic brain and grade 1–3 glioma patient tumor tissues (Supplementary Fig. S3B and S3C). These data led us to further characterize the functional roles of NR4A2 in microglia response to oxidative stress.

As the subcellular localization of nuclear receptors is stimulus responsive, we examined the effects of oxidative stress on the subcellular localization of NR4A2 in microglia. In primary murine microglia, H₂O₂ treatment induced the upregulation of NR4A2 expression and nuclear translocation of NR4A2, whereas NAC treatment blocked these effects (Fig. 4I and J). These results were validated in murine BV2 and human HMC3 microglia (Supplementary Fig. S3D–S3F). Quantification of subcellular NR4A2 protein levels in microglia by immunoblot analysis revealed that H₂O₂ treatment increased total NR4A2, whereas NAC cotreatment downregulated NR4A2 levels (Fig. 4K; Supplementary Fig. S3G), primarily due to an increase in nuclear NR4A2, which was blocked by cotreatment with NAC, whereas cytoplasmic NR4A2 levels did not substantially vary between treatment groups (Fig. 4K; Supplementary Fig. S3G). These data suggest that oxidative stress induces NR4A2 nuclear entry in oxidative stress microglia.

To determine the functional importance of NR4A2 in oxidatively stressed microglia, we quantified NR4A2⁺ IBA1⁺ cells in surgical biopsies from a cohort of grade I–IV glioma patient specimens, revealing a positive correlation with the abundance of 8-hydroxy-2′-deoxyguanosine (8-OHdG)–positive Iba1⁺ cells (8-OHdG is a specific marker to measure oxidative stress levels), which marks microglia with high oxidative levels (r = 0.648, P < 0.001; Fig. 4L). In contrast, the abundance of CD8⁺ T cells negatively correlated with the abundance of NR4A2⁺ IBA1⁺ cells in both IDH wild-type and mutant patients (r = −0.2834, P = 0.01; Fig. 4L). Therefore, we cocultured murine CD8⁺ T cells with primary murine microglia under different conditions. OT-I mice CD8⁺ T cells cocultured with H₂O₂-treated primary murine microglia pulsed with OVA displayed increased PD-1 expression, which could be reversed when cocultured with primary microglia with NR4A2 knockdown (Fig. 4M; Supplementary Fig. S3H and S3I). Concordantly, H₂O₂ treatment reduced MHC-I expression on primary murine microglia, which could be rescued by the knockdown of NR4A2 in microglia (Fig. 4M; Supplementary Fig. S3J). We next cocultured GL261-OVA cells with primary murine microglia under different conditions and then isolated primary microglia using magnetic beads. Isolated primary microglia were cocultured with OT-I mice CD8⁺ T cells. Proliferation of CD8⁺ T cells was measured by carboxyfluorescein succinimidyl amino ester (CFSE) staining, which fluorescently labels live cells intracellularly. With H₂O₂-treated primary microglia, CD8⁺ T-cell proliferation was reduced, with effects reversed by the knockdown of NR4A2 in microglia (Fig. 4N; Supplementary Fig. S3K). OT-I mice CD8⁺ T cells cocultured with H₂O₂-treated primary microglia displayed reduced secretion of both IFNγ (Fig. 4O) and TNFα (Fig. 4P), which was blocked by NR4A2 knockdown in microglia. We explored the potential mechanisms of the increased expression of PD-1 in CD8⁺ T cells after coculture with H₂O₂-treated primary microglia. As previous reports have shown that IL10 could enhance the expression of PD-1 in CD8⁺ T cells (41), we considered whether the impact of H₂O₂-treated primary microglia on PD-1 expression in CD8⁺ T cells was also mediated by IL10. Blockade of IL10 using a neutralization antibody decreased the H₂O₂-treated primary microglia induction of PD-1 expression in CD8⁺ T cells (Supplementary Fig. S3L). Moreover, although conditioned media from H₂O₂-treated microglia promoted cell proliferation and migration of GL261 tumor cells, NR4A2 knockdown in H₂O₂-treated microglia blocked these effects (Supplementary Fig. S3M and S3N). NR4A2 knockdown also decreased the expression of CD206 in H₂O₂-treated microglia and increased the expression of CD86 in H₂O₂-treated microglia (Supplementary Fig. S3O).

Based on the goal to translate our findings into potential therapeutic paradigms with clinical significance, we leveraged a pharmacologic NR4A2 inhibitor, Bay-11-7082 (42). In murine GL261 allografts, we examined the mRNA expression of Nr4a2 in microglia and other cell types and found that the microglial Nr4a2 mRNA level was higher compared with other cell types (Supplementary Fig. S3P). We hypothesized that the pharmacologic targeting of NR4A2 primarily alters the antigen presentation functions of microglia in GBM. Bay-11-7082 treatment decreased Nr4a2 mRNA levels in H₂O₂-challenged primary microglia, but did not alter Nr4a2 mRNA levels in untreated primary microglia (Fig. 4Q). Bay-11-7082 administration reduced the growth of GL261-Luc–derived tumors in vivo, as measured by both luciferase imaging of tumor-bearing mice and histologic staining (Fig. 4R; Supplementary Fig. S3Q). The effects of Bay-11-7082 treatment on tumor size translated into prolonged survival of GL261-Luc–bearing mice (P = 0.001; Fig. 4S). In tumors derived from the Bay-11-7082 treatment group, CD206⁺ microglia were reduced in frequency compared with tumors treated with vehicle control, whereas the abundance of MHC-I⁺ microglia and CD8⁺ T cells was increased, as measured by flow cytometry (Supplementary Fig. S3R and S3S). Bay-11-7082 treatment increased the expression of effector molecules, including TNFα, IFNγ, and granzyme B on CD8⁺ T cells and decreased expression of immune-checkpoint molecules, including PD-1 and TIM3 on CD8⁺ T cells (Supplementary Fig. S3T). To address NR4A2 function in tumor cells rather than immune cells, we knocked down NR4A2 in CT2A cells but did not find changes in tumor growth and tumor-bearing mouse overall survival in vivo (Supplementary Fig. S3U–S3X), suggesting a specific role of NR4A2 in the TIME. Collectively, NR4A2 pharmacologic inhibition blocks the oxidative stress impairment of antigen presentation and mediates the phenotypic switch of microglia toward alterative activation, manifesting as reduced tumor growth.

As NR4A2 is an orphan nuclear receptor superfamily and a transcription factor, we performed chromatin immunoprecipitation followed by deep sequencing (ChIP-seq) to identify its target DNA elements. Among the peaks identified, the top genes identified included serine protease inhibitor A3N (Serpina3n), WAP, Kazal, immunoglobulin, Kunitz and Netrin domain-containing protein 1 (Wfikkn1), protein SFI1 homolog (Sfi1), dynamin-3 (Dnm3), Dixin (Dixdc1), squalene monooxygenase (Sqle), high mobility group protein HMG-I/HMG-Y (Hmga1), and T-cell activation GTPase-activating protein 1 (Tagap1; Fig. 5A). GO enrichment analysis revealed that genes transcriptionally regulated by NR4A2 were primarily involved in glycosylation, phosphorylation, cell motility, kinase signaling, and gene transcription (Fig. 5B). Comparative RNA-seq of NR4A2-knockdown and control microglia was used to construct a Venn diagram to identify overlapping genes that met the three criteria: (i) genes on whose promoter regions bound NR4A2 identified by ChIP-seq; (ii) genes upregulated in microglia under oxidative stress identified by RNA-seq; and (iii) genes downregulated in microglia after NR4A2 knockdown identified by RNA-seq. Only three genes overlapped satisfying the criteria: Sqle, Tagap1, and Hmga1 (Fig. 5C). We examined which of these genes were bona fide downstream targets transcriptionally regulated by NR4A2 in microglia. NR4A2 knockdown reduced mRNA levels of both Nr4a2 and Sqle in primary microglia, but did not cause a substantial change in the mRNA levels of Hmga1 or Tagap1 (Fig. 5D). Aligned with the reported functions of SQLE (43), NR4A2 knockdown in microglia downregulated the expression of genes primarily involved in lipid metabolism (Supplementary Fig. S4A) and upregulated the expression of genes involved in MHC-I complex and antigen presentation (Supplementary Fig. S4B).

To genetically study NR4A2 function, we constructed Nr4a2-knockout mice using CRISPR–Cas9 strategies (Fig. 5E). As previous studies reported that homozygous Nr4a2-knockout mice die within 12 hours after birth (44), we used heterozygous Nr4a2-knockout mice (Nr4a2^+/−) for functional studies. Primary microglia were isolated from Nr4a2^+/− mice (Fig. 5F), which expressed lower levels of both Nr4a2 and Sqle compared with wild-type mice (Fig. 5G). Under oxidative stress, binding of NR4A2 to the promoter region of the Sqle gene was elevated compared with control by ChIP-qPCR analysis (Fig. 5H). Confirmatory luciferase reporter assays were used to explore transcriptional regulation of Sqle by NR4A2. We cloned the Sqle-3′-UTR into the pGL4.10 basic vector and produced a pGL4.10 vector harboring the Sqle promoter region. Cotransfection of pGL4.10-Sqle and Nr4a2-overexpressing plasmid increased luciferase activity compared with control, suggesting that NR4A2 directly targets Sqle (Fig. 5I). To validate that NR4A2 specifically controls Sqle gene expression by the predicted binding sites in the promoter region, we generated a mutant construct, pGL4.10-Sqle-Mut, in which the NR4A2 binding site sequences on Sqle-3′-UTR were mutated. Cotransfection of this mutant construct with the NR4A2 plasmid into cells demonstrated that mutation of the NR4A2 binding sites from the 3′-UTR of Sqle attenuated the effects of NR4A2 on luciferase activity (Fig. 5J). Immunoblotting demonstrated that H₂O₂ treatment induced SQLE protein expression, which was reversed by either cotreatment with NAC or NR4A2 knockdown (Fig. 5K). In glioma patient tumor samples, NR4A2⁺IBA1⁺ microglia coexpressed SQLE, whereas NR4A2⁺SQLE⁺IBA1⁺ microglia were rarely observed in the nonneoplastic brain (Fig. 5L and M). Moreover, we observed that the abundance of SQLE⁺ microglia was higher in GBM patient samples compared with grade 1–3 glioma or nonneoplastic patients (Supplementary Fig. S4C and S4D). Therefore, Sqle is a bona fide transcriptionally regulated downstream target of NR4A2 in GBM microglia.

SQLE catalyzes the stereospecific oxidation of squalene to 2,3-oxidosqualene and is considered to be a rate-limiting enzyme in steroid biosynthesis (43). Therefore, we examined the impact of oxidative stress on cholesterol metabolism in oxidative stress microglia. In microglia, neither H₂O₂ treatment nor NR4A2 knockdown affected the levels of the substrate, squalene; however, H₂O₂ treatment increased 2,3-oxidosqualene levels, and this effect was reversed by NR4A2 knockdown (Supplementary Fig. S4E). Treatment with the SQLE-specific inhibitor NB-598 (45) alleviated the phenotypic alterations caused by H₂O₂ challenge in primary microglia, as measured by the mRNA levels of Il6, Nos2, Arg1, and Il10 (Fig. 5N). These data suggest that NR4A2 mediates cholesterol metabolic alterations in microglia.

To interrogate the function of NR4A2 in microglia, we generated both global heterozygous Nr4a2-knockout mice and CX3CR1⁺ myeloid cell–specific homozygous Nr4a2-knockout mice. We constructed Nr4a2-flox transgenic mice (Fig. 6A) and crossed them onto a CX3CR1⁺ myeloid cell–targeting Cre-expressing background, Cx3cr1-Cre, to generate CX3CR1⁺ myeloid cell–specific Nr4a2-knockout mice (Nr4a2^fl/flCx3cr1^Cre mice; Fig. 6B). Although no significant reduction of microglia numbers was observed in Nr4a2^fl/flCx3cr1^Cre mice compared with control mice (Supplementary Fig. S5A and S5B), NR4A2⁺IBA1⁺ cell numbers were reduced in the brains of Nr4a2^fl/flCx3cr1^Cre mice bearing GL261-Luc cells compared with those in Nr4a2^fl/fl mice by immunofluorescent analysis (Fig. 6C and D). Tumor sizes of GL261-Luc–derived tumors grown in Nr4a2^fl/flCx3cr1^Cre mice implanted were smaller than those grown in Nr4a2^fl/fl mice (Fig. 6E and F). Nr4a2^fl/flCx3cr1^Cre mice bearing intracranial GL261-Luc cells displayed prolonged survival compared with Nr4a2^fl/fl tumor–bearing mice (P = 0.005; Fig. 6G). The number of SQLE⁺IBA1⁺ cells in the brains of Nr4a2^fl/flCx3cr1^Cre mice bearing GL261-Luc cells was also reduced compared with those in Nr4a2^fl/fl mice (Fig. 6H and I). CD206⁺IBA1⁺ cell numbers in both tumor and peritumoral regions were deceased in Nr4a2^fl/flCx3cr1^Cre mice compared with those in Nr4a2^fl/fl mice (Fig. 6J–L). Flow cytometry analysis showed that MHC-I⁺ microglia and CD8⁺CD45⁺ T cells were increased in Nr4a2^fl/flCx3cr1^Cre mice compared with those in Nr4a2^fl/fl mice (Fig. 6M), whereas the frequency of CD206⁺ microglia was reduced (Supplementary Fig. S5C). Significant changes in the frequency of CD86⁺ cells and CD4⁺CD45⁺ T cells were not observed (Supplementary Fig. S5C). Expression levels of immune-checkpoint molecules, including PD-1, CTLA4, and TIM3, on CD8⁺ T cells were reduced in the brains of Nr4a2^fl/flCx3cr1^Cre mice implanted with GL261-Luc cells compared with those in Nr4a2^fl/fl mice, whereas expression levels of effector molecules, including TNFα and IFNγ, were increased in Nr4a2^fl/flCx3cr1^Cre mice (Fig. 6M). To prove that the impact of microglial NR4A2 on glioma cells was mediated through the immune system, we used a CD8 neutralization antibody to deplete CD8⁺ T cells in vivo. The prolonged survival of glioma-bearing Nr4a2^fl/flCx3cr1^Cre mice compared with control mice was abolished by administration of the CD8 neutralization antibody (Fig. 6N), indicating that this effect was indeed CD8⁺ T cell–dependent. To examine the possible effects of monocyte-derived macrophages on glioma tumor growth in the conditional knockout mouse models, we analyzed NR4A2 expression in the macrophages of either normal brain or tumor-bearing brain in our GL261 scRNA-seq dataset and found that different from microglia, there was no significant difference between the two groups (Supplementary Fig. S5D). In vitro coculture experiments revealed that H₂O₂-treated macrophages did not promote the tumor growth of GL261 cells (Supplementary Fig. S5E). We further isolated the primary microglia from either GL261-bearing brain or normal brains to detect the expression of Nr4a2 and found that Nr4a2 was prominently increased in the microglia of glioma-bearing brains compared with control brains (Supplementary Fig. S5F and S5G). In addition, a previous study suggested that glioma size increases in Ccr2 (a classic marker for monocyte-derived macrophages; refs. 46, 47)-knockout mice compared with wild-type, suggesting that monocyte-derived macrophages may play a tumor-inhibiting, instead of tumor-promoting, role in glioma growth (48). These data suggest that the primary tumor-suppressive effects in Nr4a2^fl/flCx3cr1^Cre mice compared with control mice were likely derived from NR4A2 inhibition in microglia.

In validation studies, tumors derived from GL261-Luc cells grown in Nr4a2^+/− mice were smaller than those grown in wild-type mice (Fig. 6O and P). Nr4a2^+/− mice bearing GL161-Luc cells also displayed improved survival (P = 0.01; Fig. 6Q). An immunofluorescent assay revealed that CD206⁺IBA1⁺ microglia were reduced in Nr4a2^+/− mice bearing GL261-Luc–derived tumors compared with wild-type mice (Supplementary Fig. S5H and S5I). mRNA levels of the classically activated myeloid marker Nos2 were slightly increased, whereas mRNA levels of alternatively activated myeloid markers, including Arg1 and Il10, were reduced in tumor-bearing Nr4a2^+/− mice (Supplementary Fig. S5J). Further flow cytometry analysis showed that CD206⁺ microglia were reduced in Nr4a2^+/− mice compared with wild-type mice, although minimal changes were observed for MHC-I⁺ and CD86⁺ microglia (Supplementary Fig. S5K). The abundance of CD8⁺ T cells was increased in Nr4a2^+/− mice compared with wild-type mice, and the expression of immune-checkpoint molecules, including PD-1 and TIM3, was reduced on CD8⁺ T cells, whereas effector molecules TNFα and IFNγ were increased (Supplementary Fig. S5L). Thus, genetic targeting of NR4A2 in microglia suppresses tumor growth due to reduced immunosuppression.

Immune-checkpoint blockade therapy has shown limited therapeutic efficacy against GBM (49, 50). We found that PD-1 antibody treatment synergized with CX3CR1⁺ myeloid cell–specific Nr4a2 knockout improved the survival of glioma-bearing mice compared with control, with anti–PD-1 monotherapy, or in CX3CR1⁺ myeloid cell–specific Nr4a2 knockout mice (Supplementary Fig. S6A). To translate our findings into a potential preclinical therapeutic paradigm, we interrogated the antitumor efficacy of inhibiting the NR4A2–SQLE axis using pharmacologic inhibitors of NR4A2 (Bay-11-7082; ref. 51) or SQLE (terbinafine; refs. 52, 53) in combination with the immune-checkpoint blockade. We first analyzed the alterations of components of TIME with each monotherapy and combination regimens. Combined treatment with Bay 11-7082 and anti–PD-1 increased the abundance of CD8⁺ T cells in tumor-bearing mice, whereas expression levels of immune-checkpoint molecules, including PD-1, CTLA4, LAG3, and TIM3, on CD8⁺ T cells were reduced compared with the other three treatment groups (Fig. 7A). In parallel, combined treatment with terbinafine and anti–PD-1 downregulated expression levels of immune-checkpoint molecules, such as PD-1, CTLA4, and TIM3, on CD8⁺ T cells compared with the other treatment groups (Fig. 7B and C). Although anti–PD-1 monotherapy showed limited efficacy in GL261-Luc–bearing mice, combined treatment using anti–PD-1 and Bay-11-7082 induced tumor regression and prolonged survival compared with monotherapies or treatment control (Fig. 7D and E; Supplementary Fig. S6B). Combinatory administration with terbinafine and anti–PD-1 also induced tumor regression and prolonged survival for GL261-Luc–bearing mice compared with each monotherapy or treatment control (Fig. 7F and G; Supplementary Fig. S6B). Similar synergistic effects were observed in CT2A-Luc–bearing mice (Supplementary Fig. S6C–S6J). These data indicated that pharmacologic inhibition of the NR4A2–SQLE axis augments the efficacy of immune-checkpoint blockade therapy.

To explore the clinical relevance as well as whether oxidative stress–challenged microglia informed patient outcome, we further included 125 patients with glioma based on tissue microarray and stratified the patients based on levels of oxidative stress, as measured by 8-OHdG⁺ microglia levels. Patients with glioma whose tumors had high levels of 8-OHdG⁺ microglia had worse overall survival compared with 8-OHdG⁺ microglia–low group patients, with survival differences in patients with both low-grade and high-grade glioma (Fig. 7H). Abundance of 8-OHdG⁺ microglia correlated with tumor grade (Fig. 7I), and the levels of ROS were higher in tumor regions compared with peritumor regions or nonneoplastic brains (Fig. 7J–L; Supplementary Fig. S7A and S7B). We next explored the spatial relationship between CD8⁺ T cells and oxidative stressed microglia in an in-house cohort of patient samples. Oxidative stress levels of microglia, assessed by 8-OHdG staining, were inversely correlated with the abundance of CD8⁺ T cells (Supplementary Fig. S7C–S7E). Contact between CD8⁺ T cells and microglia was reduced in patient tumors with high oxidative stress compared with that in low oxidative stress (Supplementary Fig. S7F). Regional variation within tumors showed that in areas of low oxidative stress, microglia were more likely to be in contact with CD8⁺ T cells, whereas in areas of high oxidative stress, microglia displayed reduced contact with CD8⁺ T cells (Supplementary Fig. S7G). Both the number of CD8⁺ T cells and the contact between CD8⁺ T cells and microglia were negatively correlated with the abundance of 8-OHdG⁺IBA1⁺ cells in patients with glioma (Supplementary Fig. S7H). Moreover, in glioma tumor regions from clinical samples, the abundance of 8-OHdG⁺CD206⁺IBA1⁺ microglia was higher than that from nonneoplastic brains (Supplementary Fig. S7I and S7J).

Univariate and multivariate Cox regression analyses of six factors [MGMT, IDH mutation, 8-OHdG, age, GFAP, and epithelial membrane antigen (EMA)] in the 125 patients with glioma demonstrated that high oxidative stress level of microglia was an independent factor of poor prognosis in patients with glioma (Supplementary Fig. S7K and S7L). As CX3CR1 is a well-established marker of microglia, we performed a subgroup analysis of The Cancer Genome Atlas (TCGA)-GBM dataset (54), which revealed that in patients with CX3CR1^hi GBM, high NR4A2 expression was associated with a worse prognosis compared with those with low expression NR4A2 (Supplementary Fig. S7M). In contrast, in patients with CX3CR1^lo GBM, NR4A2^hi and NR4A2^lo groups did not differ in prognosis (Supplementary Fig. S7M). Similarly, in patients with NR4A2^hi GBM, a high expression of CX3CR1 had a worse prognosis compared with low CX3CR1 expression (Supplementary Fig. S7N). However, in patients with NR4A2^lo GBM, CX3CR1^hi and CX3CR1^lo groups did not display differences in survival (Supplementary Fig. S7N). Thus, the oxidative stress–NR4A2–SQLE axis in microglia informs clinical outcomes for patients with GBM.

In this study, we found that in the TIME of GBM, microglia are under severe oxidative stress and upregulate NR4A2-dependent transcriptional machinery. NR4A2 regulates SQLE expression and cholesterol metabolism. Functional experiments in Nr4a2^+/− mice and Nr4a2^fl/flCx3cr1^Cre mice demonstrated that this metabolic reprogramming reshaped microglia phenotypic plasticity and disrupted their antigen presentation capacity for CD8⁺ T cells. Pharmacologic targeting of NR4A2 or SQLE was augmented by immune-checkpoint blockade therapy in vivo. Our findings support a new therapeutic paradigm in GBM through the inhibition of an oxidative stress–NR4A2–SQLE axis in microglia.

The blood–brain barrier (BBB) or the neurovascular unit limits the interplay between the central nervous system and the peripheral circulation, reflected in the GBM TIME with the majority of immune cells being tissue-resident myeloid cells and rare infiltration of adaptive immune cells (55, 56). Microglia and macrophages, as major tissue-resident myeloid cells, display extensive phenotypic plasticity in the context of GBM (16). When stimulated with protumorigenic external signals, macrophages and microglia usually assume alternatively activated and immunosuppressive states (17), whereas upon tumor treatment, they assume M1-like polarization to suppress tumor growth (19). Strategies to induce M1-like states in microglia and macrophages have been developed to treat patients with GBM (57–59).

Oxidative stress directly modulates the malignancy and metabolic state of cancer cells themselves (36, 60), as well as conditioning microenvironmental niche components, including immune cells and fibroblasts. For instance, tumor-infiltrating Tregs sustain and amplify their tumor suppressor capacity through death via oxidative stress (31). These findings suggest that the oxidative pathway may serve as a metabolic checkpoint that links the molecular interactions between cancer cells and the immune cells as well as the immune–immune interface within the TIME.

NR4A2, as a member of the nuclear receptor superfamily, functions in both physiologic (61–63) and pathologic circumstances (64, 65). NR4A2 also serves as a key factor in tumorigenesis and cancer progression. Jing and colleagues found that NR4A2–osteopontin–Wnt signaling contributes to hepatic stellate cell–induced cancer promotion in intrahepatic cholangiocarcinoma (66). Cyclooxygenase-2 inhibitors suppress colorectal cancer growth by downregulating osteopontin and NR4A2 (67). The role of NR4A2 in antitumor immunity has recently been investigated; whereas chimeric antigen receptor (CAR) T cells have been less effective against solid tumors, CAR T cells lacking all three NR4A transcription factors (Nr4a triple knockout, including Nr4a2) promoted tumor regression and prolonged the survival of solid tumor–bearing mice (68). Inhibition of NR4A2 receptors enhances chemotherapy-related antitumor immunity by breaking Treg-mediated immune tolerance (69). The reported toxicities of NR4A2 inhibitors have been minimal (70), suggesting that NR4A2 is a potentially safe and potent immunotherapeutic target.

Cholesterol homeostasis is an essential metabolic pathway for cancer cells. In GBM, tumor cells rely on exogenous cholesterol for survival, taking up cholesterol by suppressing liver X receptor (LXR) ligand synthesis, with the LXR–cholesterol axis revealed as a potential therapeutic target (71). Moreover, IDH mutation enhances 24(S)-hydroxycholesterol production and alters cholesterol homeostasis in glioma (72). SQLE catalyzes the stereospecific oxidation of squalene to (S)-2,3-epoxysqualene and is a rate-limiting enzyme in cholesterol metabolism. SQLE supports tumor growth in multiple cancer types (73, 74) and is a viable pharmaceutical target (75, 76). SQLE inhibitors demonstrate acceptable toxicities in preclinical cancer models (75, 77). However, SQLE activity in the immune system and its potential value in cancer immunotherapy have remained obscure. Our findings demonstrate that oxidative stress induced NR4A2 upregulation and nuclear translocation in microglia, transcriptionally activating downstream Sqle expression. Dysregulation of SQLE resulted in altered cholesterol metabolism in microglia and an immune-suppressive TIME, favorable for GBM growth and progression. However, how cholesterol synthesis is intricately linked to immunosuppressive functions of microglia warrants further in-depth exploration.

Although immunotherapy has transformed the care of patients afflicted with many cancer types, clinical trials of immunotherapies as monotherapy or combined with other regimens for patients with GBM have demonstrated infrequent successes (9, 49, 50)—with numerous negative studies—supporting the critical need to develop novel therapeutic combinations. TIM3 and PD-1 blockade shows survival benefit in preclinical GBM models (78), and a phase I clinical trial to evaluate the treatment efficacy of sabatolimab, a first-in-class anti-TIM3 mAb, in combination with anti–PD-1 and stereotactic radiosurgery in recurrent GBM is underway (ClinicalTrials.gov Identifier: NCT03961971). Antigen presentation in the immune system acts as a critical step to activate cytotoxic T lymphocytes to defend against both pathogens and tumors. Antigen-presenting cells include dendritic cells, macrophages, and neutrophils. Stimulated adult microglia cross-prime CD8⁺ T cells injected in the brain (79). Our data showed that NR4A2-regulated transcriptional machinery functionally impairs the antigen presentation capacities of microglia for CD8⁺ T cells and inhibition of NR4A2 synergizes with PD-1 therapy in GBM models. This finding potentially creates a new opportunity by increasing antigen presentation capacities of microglia to enhance immune-checkpoint therapy for patients with GBM.

Taken together, we leveraged a single-cell sequencing-based integrated strategy to interrogate the immune landscape of GBM to reveal an oxidative stress microglial immune state. These microglia displayed an NR4A2-regulated transcriptional machinery to rewire cholesterol metabolism, generating an alternatively activated phenotype and impaired antigen presentation capacity for CD8⁺ T cells in GBM. Inhibition of the NR4A2–SQLE axis in microglia augmented the therapeutic effects of immune-checkpoint blockade therapy in preclinical GBM models. Future prospective clinical trials to investigate the therapeutic benefit of this approach in patients with GBM may be warranted.

This study was approved by the Institutional Ethics Committee of Sichuan University and Sichuan Provincial People's Hospital. Written informed consents were obtained from all patients prior to analysis. All patient samples were archived in a snap-frozen or formalin-fixed, paraffin-embedded state. Paraffin-embedded or snap-frozen tumor tissues from patients with glioma were used for immunofluorescent staining. Glioma patient and tissue sample information of glioma tissue microarray, which was purchased from Shanghai Outdo Biotech Co., Ltd., is shown in Supplementary Table S1. Among them, samples from patients with glioma were used for ROS level evaluation by the Reactive Oxygen Species Assay Kit (Beyotime, S0033S) or anti–8-OHdG (GeneTex, GTX35250).

All animal studies were reviewed and approved by the Institutional Ethics Committee of Sichuan University. Heterozygous Nr4a2-knockout mice, Nr4a2-floxed mice, and OT-I mice were constructed and purchased from GemPharmatech Co., Ltd. CX3CR1-Cre and female C57BL/6 mice were purchased from Cyagen (China) and Chengdu Dossy Experimental Animals Co., Ltd. (China), respectively. For intracranial xenograft generation, mice (4–6 weeks of age) were anesthetized and inoculated with 2 × 10⁵ or 2 × 10⁴ tumor cells (GL261-Luc or CT2A-Luc)/3 μL into the frontal region of the cerebral cortex, a site 2.5 mm lateral (right), 1.5 mm anterior, and 3.5 mm ventral with respect to the bregma.

In oxidative stress experiments, primary microglia were treated with H₂O₂ (500 μmol/L) either in the absence or presence of NAC (5 mmol/L) for 12 hours. Murine glioma cells (GL261-Luc, 2 × 10⁵ cells; CT2A-Luc, 2 × 10⁴ cells) were incubated with conditional medium of treated primary microglia for 12 hours. Treated glioma cells and primary microglia (1:1, 2 × 10⁵ or 2 × 10⁴ cells) treated with H₂O₂ were inoculated together into the frontal region of the cerebral cortex. In vivo bioluminescence-based imaging was conducted with QuikView 3000 (Bio-Real) or IVIS Lumina III (PerkinElmer). The ROS level of microglia was evaluated with the Reactive Oxygen Species Assay Kit (Beyotime, S0033S).

For in vivo therapy, anti–PD-1 mAbs (BE0146, Bio X Cell) were administered intraperitoneally at a dosage of 200 to 500 μg per mouse on day 3 after tumor cell implantation and then every 2 days for the duration of the experiment. Anti-CD8 mAbs (BE0061, Bio X Cell) were given intraperitoneally at a dose of 800 μg per mouse after tumor inoculation and then twice a week for the duration of the experiment. Bay-11-7082 (Targetmol, T1902) was given intraperitoneally at a dose of 25 mg/kg per mouse on day 3 after tumor inoculation and then every 2 days for the duration of the experiment. Terbinafine (Targetmol, T6702) was given by oral gavage at a dose of 560 mg/kg per mouse on day 3 after tumor inoculation and then every 2 days for the duration of the experiment.

Mice with neurologic deficits or moribund appearance were sacrificed. Neurologic signs or signs of morbidity included hunched posture, gait changes, lethargy, and weight loss. Brains were harvested, fixed in 4% formaldehyde for 48 hours, and processed for paraffin-embedded blocks.

Primary microglia cultures were prepared as previously described (80). Briefly, brains from the cerebral cortex of neonatal C57BL/6 mice (P0–P2) were minced using a scalpel and dissociated in 1640 medium containing 0.125% trypsin and 50 μg/mL DNAse I at 37°C for 10 minutes. Cells were isolated by centrifugation at 300 × g for 10 minutes at 4°C and were seeded on the plate and cultured with Dulbecco's modified Eagle medium (DMEM)/F12 supplemented with 10% FBS and 50 ng/mL mouse M-CSF (PeproTech, 300-25). After 10 to 15 days in culture, microglia were isolated by mechanical shaking (400 rpm, 1 hour). The purity of microglia was further determined by FACS.

In adult mice, including Nr4a2^+/− mice and C57BL/6 mice, primary microglia were sorted by flow cytometry. Brains of 4- to 6-week-old mice were triturated and digested by 2 mg/mL Papain and 50 μg/mL DNAse I at 37°C for 10 minutes and then passed through a 70-μm cell strainer. Flow cytometry sorting was used to purify CD11b⁺CD45^lo cells from brain tissue. OT-I CD8⁺ T cells were isolated from the spleens of 6- to 12-week-old mice. Selection was carried out with a negative CD8 isolation kit (Miltenyi Biotec, 130-104-075) following the manufacturer's instructions. Neutrophils and macrophages were isolated from the bone marrow of 6- to 8-week-old C57BL/6 mice, which were harvested in RPMI 1640 medium. For neutrophils, bone marrow was laid on top of histopaque-1077 and histopaque-1119, followed by centrifugation at 600 × g for 20 minutes. Neutrophils were enriched in the interface of histopaque-1077/1119. For macrophages, bone marrow cells were stimulated with M-CSF (20 ng/mL) after being lysed with Red Blood Cell Lysis Buffer. Culture medium was half changed every 3 days. On day 7, macrophages were collected for further experiments.

For cell lines, the murine GBM cell line GL261, the murine GBM cell line CT2A, and the human embryonic kidney cell line 293T, as well as the human microglia cell line HMC3, were cultured in DMEM. The murine microglia BV2 cell line was cultured in DMEM: F-12 (DMEM/F-12). These cell lines were cultured in the indicated medium containing 10% FBS (Gibco) and 1:100 antibiotic–antimycotic (Gibco). GL261 and CT2A cell lines were transduced with a lentiviral vector containing a luciferase reporter together with the puromycin resistance gene (GL261-Luc, CT2A-Luc). Gl261 cells were transfected with OVA plasmid, which was gifted by Professor Yuquan Wei, Sichuan University, China. Conditioned media were collected from treated or untreated cells as indicated after culturing for 12 hours. GSC2907 was derived by our laboratory as reported previously (81). Glioma stem cells were cultured in the neurobasal media (Invitrogen) supplemented with B27 without vitamin A (Invitrogen), EGF and bFGF (20 ng/mL each; PeproTech), sodium pyruvate, and glutamax. All cells were confirmed to be Mycoplasma-free and maintained at 37°C and 5% CO₂.

Immunofluorescent staining of tissues was performed as described previously (82), using primary antibodies anti-IBA1 (Abcam, ab5076, 1:100), anti-CD8 (Abcam, ab217344, 1:100, or HuaBio, ER1905-98, 1:50), anti–8-OHdG (GeneTex, GTX35250, 1:100), anti-Nurr1 (Abcam, ab176184, 1:100; Proteintech, 10975-2-AP, 1:50; or Santa Cruz, sc-376984, 1:50), anti-Sqle (Proteintech, 12544-1-AP, 1:50), anti-CD206 (HuaBio, ET1702-04, 1:50, or Proteintech, 60143-1-Ig, 1:50), anti-SOD2 (Proteintech, 24127-1-AP, 1:100), anti-MGMT (Proteintech, 67476-1-Ig, 1:100), anti-IDH (Dianova, DIA-H09, 1:50) and secondary antibodies with Alexa Fluor dyes (Thermo Fisher Scientific). Detailed information for these antibodies is in Supplementary Table S2. Next, specimens counterstained with DAPI were imaged by laser-scanning confocal microscopy (LSM 880, Zeiss) at × 40 magnification with at least 10 fields per section. Isotype-matched antibodies were used for the negative control to adjust the immunofluorescence signal. Immunofluorescence images were analyzed by ZEN lite (blue edition) software to identify infiltrated CD8⁺ T cells and 8-OHdG⁺ cells in the samples.

Cell proliferation was assessed by counting tumor cell number under microscope or flow cytometry. Briefly, GL261 cells or GSC2907 cells were seeded into 96-well plates at a density of 3,000 cells/100 μL and incubated overnight at 37°C. Conditional medium was added in the respective group and incubated cells for 24 to 48 hours. GL261 cell number was counted with IncuCyte Live-Cell Analysis under 10 × lens, and GSC2907 cells were digested into single-cell suspensions and counted by flow cytometry.

Transwell 24-well chambers (Corning or Millipore) were applied for the in vitro cell migration assay as described previously (83). More than five contiguous fields of each sample were assessed to obtain a representative number of cells that had migrated across the membrane.

Mouse or human cells were stained with the following antibodies at a dilution ratio of 1:50: anti-CD45 PerCP-Cy5.5 (BD Biosciences, 550994), anti-CD11b FITC (BioLegend, 101206), anti-CD86 APC (BioLegend, 105011), anti-CD206 PE (BioLegend, 141706), anti-CD8a FITC (BioLegend, 100706), anti-CD4 APC (BioLegend, 100516), anti–MHC-I PE (BioLegend, 114607), anti-Foxp3 PE (BD Biosciences, 563101), anti-IFNγ PE (BD Biosciences, 554412), anti-TNFα PE/Cyanine7 (BioLegend, 506324), anti–Granzyme B PE/Cyanine7 (BioLegend, 396410), anti–PD-1 PE/Cyanine7 (BioLegend, 109110), anti-Tim3 PE (BioLegend, 134003), anti-CTLA4 PE (BioLegend, 106305), and anti-LAG3 PE (BioLegend, 125207). For intracellular staining, cells were fixed in the Fixation and Permeabilization Kit (eBioscience, 00-5521-00). After being washed with Fixation and Permeabilization Buffer, the cells were stained intracellularly for 30 minutes in the dark. All samples were acquired on a BD LSRFortessa (BD Biosciences) or ACEA NovoCyte (ACEA Biosciences) and were analyzed using FlowJo (FlowJo LLC) or NovoExpress software (ACEA Biosciences).

Primary microglia were prepared as described above. Microglia were treated with H₂O₂ (500 μmol/L) or NAC (5 mmol/L) for 12 hours and cocultured with GL261-OVA at a ratio of 1:1 for 12 hours. Subsequently, the CD11b⁺ microglia were positively purified by CD11b (Microglia) MicroBeads (Miltenyi Biotec, 130-093-634). The CD8⁺ T cells were isolated from OT-I mice as described above and were labeled with 2 mmol/L CFSE (eBioscience, 65-0850-84) by incubating at 37°C for 10 minutes. The CD11b⁺ microglia were incubated with CFSE-labeled CD8⁺ T cells at the ratio of 1:10 for 2 days. The supernatant was collected for cytokine quantitation, including IFNγ and TNFα, which were detected by the Mouse IFNγ ELISA Kit (Proteintech, KE10001) and the Mouse TNFα ELISA Kit (Proteintech, KE10002), respectively. T-cell proliferation was analyzed by FACS in terms of CFSE dilution. To detect PD-1⁺ T cells, microglia were treated with H₂O₂ (500 μmol/L) or NAC (5 mmol/L) for 12 hours. CD8⁺ T cells isolated from OT-I mice were cocultured with H₂O₂-treated microglia pulsed with OVA for 24 hours. The proportion of PD-1⁺ T cells was analyzed by FACS.

BV2 and primary microglia were transfected with Nr4a2 siRNA (5′-UCAACAAUGGAAUCAAUCCAUUCCC-3′; RiboBio) using Lipofectamine RNAiMAX (Invitrogen) according to the manufacturer's instructions. The concentrations were 50 nmol/L equally in both experimental and control groups. The short hairpin RNAs (shRNA) specific for mouse Nr4a2(pLKO.1 vector) were purchased from Tsingke. The target sequence was as follows: 5′-UCAACAAUGGAAUCAAUCCAUUCCC-3′. To generate lentivirus, 293T cells were cotransfected with pLKO.1-shRNA plasmid and packaging plasmids PSPAX2 and PMD2.G using transfection reagent (jetPRIME). CT2A cells were transduced by culturing with viral supernatants in the presence of polybrene (OBIO).

Whole-transcriptome sequencing libraries were constructed as described previously (84). The libraries were sequenced on the Illumina HiSeq platform (Novogen; Oebiotech). RNA integrity was assessed using the RNA Nano 6000 Assay Kit of the Bioanalyzer 2100 system (Agilent Technologies). A total amount of 3 μg RNA per sample was used as input material for the RNA sample preparations. Briefly, mRNA was purified from total RNA using poly-T oligo–attached magnetic beads. Fragmentation was carried out using divalent cations under elevated temperature in First-Strand Synthesis Reaction Buffer (5×). First-strand cDNA was synthesized using a random hexamer primer and M-MuLV Reverse Transcriptase (RNase H−). Second-strand cDNA synthesis was subsequently performed using DNA polymerase I and RNase H. Remaining overhangs were converted into blunt ends via exonuclease/polymerase activities. After adenylation of 3′ ends of DNA fragments, adapters with hairpin loop structure were ligated to prepare for hybridization. In order to select cDNA fragments of preferentially 370 to 420 bp in length, the library fragments were purified with AMPure XP system (Beckman Coulter). Then PCR was performed with Phusion High-Fidelity DNA polymerase, Universal PCR primers, and Index (X) Primer. At last, PCR products were purified (AMPure XP system), and library quality was assessed on the Agilent Bioanalyzer 2100 system. The clustering of the index-coded samples was performed on a cBot Cluster Generation System using the TruSeq PE Cluster Kit v3-cBot-HS (Illumina) according to the manufacturer's instructions. After cluster generation, library preparations were sequenced on an Illumina Novaseq platform, and 150-bp paired-end reads were generated. Raw data (raw reads) of fastq format were first processed through in-house perl scripts. In this step, clean data (clean reads) were obtained by removing reads containing adapter, reads containing ploy-N, and low-quality reads from raw data. All the downstream analyses were based on the clean data with high quality. For reads mapping to the reference genome, reference genome and gene model annotation files were downloaded from the genome website directly. The index of the reference genome was built using Hisat2 v2.0.5, and paired-end clean reads were aligned to the reference genome using Hisat2 v2.0.5. We selected Hisat2 as the mapping tool, as Hisat2 can generate a database of splice junctions based on the gene model annotation file and thus a better mapping result than other nonsplice mapping tools. For quantification of gene expression level, featureCounts v1.5.0-p3 was used to count the read numbers mapped to each gene. Then, the fragments per kilobase of exon model per million mapped fragments (FPKM) of each gene were calculated based on the length of the gene and read count mapped to this gene. FPKM, expected number of fragments per kilobase of transcript sequence per million base pairs sequenced, considers the effect of sequencing depth and gene length for the read count at the same time, and is currently the most commonly used method for estimating gene expression levels. For gene functional annotation, GO analysis was performed with the R package clusterProfiler (85), which supports statistical analysis and visualization of functional profiles for genes and gene clusters.

For scRNA-seq, the live CD45⁺ sorted cell population from control and glioma-bearing mice brain were run on the 10X Genomics Chromium platform (Oebiotech). scRNA-seq libraries were prepared with Chromium Single-cell 3′ Reagent v2 (or v3) Kits according to the manufacturer's protocol. Single-cell suspensions were loaded on the Chromium Single-Cell Controller Instrument (10X Genomics) to generate single-cell gel beads in emulsions (GEM). Briefly, CD45⁺ single cells were suspended in calcium- and magnesium-free PBS containing 0.04% weight/volume BSA. CD45⁺ cells were added to each channel with a targeted cell recovery estimate of cells. After the generation of GEMs, reverse transcription reactions were engaged in barcoded full-length cDNA formation, and then the emulsions were disrupted using the recovery agent and cDNA cleanup with DynaBeads Myone Silane Beads (Thermo Fisher Scientific). cDNA was then amplified by PCR with appropriate cycles that depend on the recovery cells. Subsequently, amplified cDNA was fragmented, end-repaired, A-tailed, index adapter ligated, and library amplified. Then, these libraries were sequenced on the Illumina sequencing platform (HiSeq X Ten) and 150-bp paired-end reads were generated. For scRNA-seq data preprocessing, the Cell Ranger software pipeline (version 3.0.2 or others) provided by 10X Genomics was used to demultiplex cellular barcodes, map reads to the genome and transcriptome using the STAR aligner, and downsample reads as required to generate normalized aggregate data across samples, producing a matrix of gene counts versus cells. The R package Seurat (version 4; ref. 86) was used for data scaling, transformation, clustering, dimensionality reduction, differential expression analysis, most visualization, alignment, and quality control of sequencing data. Briefly, sequencing reads were aligned against the mm10-3.0.0 mouse reference genome with STAR (87), and count matrices were built from the resulting BAM files. Quality of cells was then assessed based on two metrics step by step: (i) the number of detected genes per cell and (ii) the proportion of mitochondrial gene counts. The following criteria were then applied to filter low-quality cells: gene number <200 or >7,000, or mitochondrial gene proportion >0.25. For integration of cells from different samples, gene–cell matrix of all samples was integrated with Seurat to remove batch effects across different samples. In parameter settings, the first 15 dimensions of principal component analysis (PCA) were used. For dimensionality reduction and visualization, the filtered gene–cell matrix was first normalized using “glmGamPoi” methods in Seurat v.4 with default parameters. The top 3,000 variable genes were used in Seurat SelectIntegrationFeatures function. Then t-distributed stochastic neighbor embedding (t-SNE) was performed on the top 15 principal components for visualizing the cells. For clustering of cells, graph-based clustering was performed on the PCA-reduced data for clustering analysis with Seurat v.4. The resolution was set to 0.4 to obtain a finer result. Briefly, the first 15 principal components of the integrated gene–cell matrix were used to construct a shared nearest-neighbor graph (SNN; FindNeighbors() in Seurat), and this SNN was used to cluster the dataset (FindClusters()) using a graph-based modularity-optimization algorithm of the Louvain method for community detection. For cluster-specific biomarker analysis, PrepSCTFindMarkers () + FindAllMarkers () were used to perform differential gene expression analysis with array data of SCT.

For microglia reclustering analysis in GL261 scRNA-seq data, the filtered microglia gene–cell count matrix was first normalized by using “glmGamPoi” methods in Seurat v.4 with default parameters. The top 3,000 variable genes were used in Seurat SelectIntegrationFeatures function. Then, t-SNE was performed on the top 15 principal components for visualizing the cells. For clustering of cells, graph-based clustering was performed on the PCA-reduced data for clustering analysis with Seurat v.4. The resolution was set to 0.2 to obtain a finer result. Briefly, the first 15 principal components of the integrated gene–cell matrix were used to construct an SNN (FindNeighbors() in Seurat), and this SNN was used to cluster the dataset (FindClusters()) using a graph-based modularity-optimization algorithm of the Louvain method for community detection. For cluster-specific biomarker analysis, PrepSCTFindMarkers() + FindAllMarkers() were used to perform differential gene expression analysis with array data of SCT.

NR4A2 ChIP assays were performed with 10⁷ cells using SimpleChIP Enzymatic Chromatin IP Kit (Cell Signaling Technology, #9003) according to the manufacturer's instructions. Briefly, cells were fixed with 1% formaldehyde at room temperature for 10 minutes and subsequently quenched with 0.125M glycine. Then, we removed media and washed cells twice with ice-cold PBS, completely removing the wash from the culture dish each time and suspended with lysis buffer supplemented with Protease Inhibitor Cocktail II. Chromatin was sonicated to obtain DNA fragments with an average length of 150 to 900 bp for sequencing by Micrococcal Nuclease. A soluble fraction of sheared chromatin was diluted with ChIP dilution buffer and incubated with NR4A2 antibody (Abcam). Immunoprecipitation samples were incubated overnight at 4°C with rotation. Protein G magnetic beads were added to each immunoprecipitation reaction and incubated for 2 hours at 4°C with rotation.

Chromatin-captured beads were collected by Magnetic Separation Rack and washed by low-salt wash and high-salt wash at 4°C for 5 minutes with rotation. Beads were resuspended in elution buffer for 30 minutes at 65°C and removed by magnet, and eluted chromatin supernatants were collected by Magnetic Separation Rack. DNA was reverse cross-linked by adding 5M NaCl and Proteinase K and incubated for 2 hours at 65°C. DNA was purified using purification columns. Eluted DNA was used to perform qPCR or sequencing library preparation. ChIP quantitative RT-PCR was performed using ChamQ SYBR qPCR Master Mix (Vazyme) on a CFX96 Real-Time System (Bio-Rad). The sequences of the primers used for all qPCR assays are in Supplementary Table S3. The libraries were sequenced on the Illumina HiSeq platform (Novogen). Low-quality reads from raw reads were removed by SOAPnuke programs (88). The remaining reads were mapped to the reference genome (UCSC mm10) by Bowtie2 (89). To determine the target genes of Nr4a2, the model-based analysis of ChIP-seq (MACS) version 2.2.6 was used to identify the specific peak region. The region was defined as a peak with P < 0.05. ChIP peak was annotated by ChIPseeker, an R bioconductor package (90).

The mRNA level of each gene was measured via quantitative RT-PCR. Cellular RNA was isolated using the RNeasy kit and reverse-transcribed to cDNA using cDNA synthesis kit (Norgen) following procedures previously described (84). Then quantitative RT-PCR was measured using ChamQ SYBR qPCR Master Mix (Vazyme) on a CFX96 Real-Time System (Bio-Rad). Data were analyzed using the ΔΔCt method. The endogenous control transcripts were used for normalization. The sequences of the primers used for all qPCR assays are in Supplementary Table S3.

The 2-kb region directly upstream of Sqle was predicted using JASPAR software. The Sqle promoter construct contained the Nr4a2 binding sites 1 (Sqle site 1: –1,417 to –1,424), 2 (Sqle site 2: –1,280 to –1,287), 3 (Sqle site 3: –395 to –402), and 4 (Sqle site 3: –1,574 to –1,581). The dual-luciferase assay was performed using the Dual-luciferase Reporter Assay kit (Beyotime, RG089S). Briefly, 1 × 10⁵ cells were seeded in each well of a 24-well tissue culture plate. The 293T cells were incubated until 70% confluent. Cells in each well were transfected with pGL4.10 basic vector or the pGL4.10 vector harboring the Sqle promoter regions (2-kb region upstream of Sqle) or the pGL4.10 vector harboring the mutated Sqle promoter region (Sqle site 2: –1,280 to –1,287) using Lipofectamine 3000 reagent. To examine the relationship between NR4A2 and Sqle promoter activity, reporter plasmid was cotransfected with Nr4a2 or vector. The Renilla luciferase reporter pGMLR-TK plasmid, 0.01 μg per well, was cotransfected as the internal control. Forty-eight hours later, the cells were treated with passive lysis buffer. The Dual-luciferase Reporter Assay System (E1910, Promega) was used according to the manufacturer's instructions. Luciferase activities were normalized to the Renilla activity value, and promoter transcription activity was compared with basic pGL4.10 vector control.

For blots of NR4A2 and SQLE, microglia were treated with H₂O₂ or NAC for 24 hours. Whole protein was prepared in RIPA Lysis Buffer containing a protease inhibitor cocktail (Thermo Scientific Halt, 87786). Nuclear protein or cytoplasmic protein was prepared by Nuclear and Cytoplasmic Extraction Reagents (Solarbio, R0050), which were used according to the manufacturer's instructions. Cell lysates were sonicated and clarified by centrifugation. Supernatants were collected and quantified for protein concentration, and mixed with SDS-PAGE protein loading buffer (Beyotime, P0286) at 95°C for 10 minutes. The same amount of protein was separated by SDS-PAGE. Protein was transferred to 0.45 mm or 0.2 mm polyvinylidene difluoride membranes (Millipore, IPVH00010 or ISEQ00010). Five percent nonfat milk in TBST (TBS with 0.1% Tween 20) was used for blocking and dilution of primary and second antibodies. Membranes were blocked at room temperature for 1 hour and incubated with primary antibody overnight at 4°C. Secondary antibodies were applied after washing and incubated at room temperature for 1 hour. Membranes were then washed with TBST, and signals were developed with a Western ECL kit (Bio-Rad, #1705061) and detected with Thermo Fisher Scientific iBright.

For lipid metabolite profiling experiments, squalene and 2,3-oxidosqualene were measured as previously described (74). Briefly, BV2 cells were transfected with siNr4a2 for 48 hours and treated with H₂O₂ for 12 hours. Cells (1 × 10⁶) were collected and resuspended in 600 μL of cold LC/MS-grade methanol, and nonpolar metabolites were extracted by consecutive addition of 300 μL LC/MS-grade water followed by 400 μL of LC/MS-grade chloroform. After centrifugation for 10 minutes at 10,000 × g and 4°C, the lower lipid-containing layer was carefully collected and dried under nitrogen. Dried lipid extracts were reconstituted in 50 μL 65:30:5 acetonitril:isopropanol:water (v/v/v) and detected with a mass spectrometer. The XCalibur QuanBrowser 2.2 (Thermo Fisher Scientific) was used at a 5-ppm mass tolerance and referenced an in-house library of chemical standards. Total lipid signal calculated with LipidSearch was used to normalize raw peak areas to generate relative abundance.

The accession numbers for gene expression data and microglia cell ChIP short-read sequencing data reported in this article are PRJNA769541, PRJNA767182, PRJNA889561, and PRJNA890930 and GSE185758 and GSE185759, available at the Sequence Read Archive and Gene Expression Omnibus repository (GEO; http://www.ncbi.nlm.nih.gov/geo/), respectively. Biomarkers to identify specific cell types in single-cell sequencing data are listed in Supplementary Table S4.

Statistical analyses were performed using R (version3.4.3, www.r-project.org). GraphPad Prism 7.0 was used to analyze and present data. Cutoff values for Nr4a2 expression levels in tumor-infiltrating microglia cells were obtained with X-tile software version 3.6.1 (Yale University School of Medicine) for clinical sample grouping (TCGA-GBM). Kaplan–Meier curves were generated to show survival significance calculated based on log-rank tests. Pearson correlation was used to assess the association between the infiltration of CD8⁺ T cells and the percentage of 8-OHdG or Nr4a2 levels of microglia in tumors. Two-sided Student t tests were used to assess differences between two groups. One-way ANOVA was used for comparison of more than two groups. Univariate and multivariate Cox regression analyses were used to identify the factors related to overall survival. P < 0.05 were considered statistically significant. Statistical significance is shown as *, P < 0.05; **, P < 0.01; ***, P < 0.001.

K. Yang reports grants from the NCI, the ASCO Conquer Cancer Foundation, and the RSNA R&E Foundation outside the submitted work. J.N. Rich reports grants and personal fees from Synchronicity Pharma outside the submitted work. No disclosures were reported by the other authors.

Z. Ye: Data curation, validation, investigation, methodology, writing–original draft. X. Ai: Data curation, validation, investigation, methodology. K. Yang: Writing–review and editing. Z. Yang: Investigation, writing–review and editing. F. Fei: Resources. X. Liao: Resources. Z. Qiu: Software. R.C. Gimple: Writing–review and editing. H. Yuan: Writing–review and editing. H. Huang: Software. Y. Gong: Data curation. C. Xiao: Resources. J. Yue: Resources. L. Huang: Software, formal analysis. O. Saulnier: Software. W. Wang: Software, formal analysis. P. Zhang: Data curation, software. L. Dai: Supervision. X. Wang: Supervision. X. Wang: Resources. Y.H. Ahn: Investigation. C. You: Resources. J. Xu: Resources. X. Wan: Investigation. M.D. Taylor: Supervision. L. Zhao: Conceptualization, writing–review and editing. J.N. Rich: Conceptualization, supervision, project administration, writing–review and editing. S. Zhou: Conceptualization, supervision, project administration, writing–review and editing.

The authors thank Bo Peng from Fudan University, China, for helpful discussion and constructive suggestions. This work was supported by the National Natural Science Foundation of China (grants #82273255, #81822034, #81821002, and #81773119 to S. Zhou; grant #82103127 to Z. Ye; and grant #82202889 to X. Ai), the National Key Research and Development Program of China (grants #2022YFA1106600, #2017YFA0106800, and #2018YFA0109200 to S. Zhou), Sichuan Science-Technology Project (grants #22ZYZYTS0070 and #2019YFH0144 to S. Zhou), Direct Scientific Research Grants from West China Second Hospital, Sichuan University (grants #KS021 and #K1907 to S. Zhou), the NIH (grants CA197718, NS103434, CA268634, and CA238662 to J.N. Rich), the Computational Genomic Epidemiology of Cancer Program at Case Comprehensive Cancer Center (grant #T32CA094186 to K. Yang), a Young Investigator Award in Glioblastoma from the ASCO Conquer Cancer Foundation (to K. Yang), and an RSNA Research Resident Grant (to K. Yang).

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.