Immunologic Features in De Novo and Recurrent Glioblastoma Are Associated with Survival Outcomes

Glioblastoma (GBM) is the most common primary malignant brain tumor in adults and is associated with a poor prognosis (1). Over the past 15 years, treatment options for GBM have shown little advancement. Standard-of-care treatment is surgical resection with adjuvant chemoradiotherapy followed by maintenance chemotherapy with temozolomide, an alkylating agent. However, with this approach, the median overall survival is only 16.9 months, and, unfortunately, relapse is expected for nearly all patients (2).

Therapeutic resistance in GBM has been associated with tumor evolution. Genomic alterations are known to accumulate in response to temozolomide such that recurrent GBM (rGBM) displays a hypermutated phenotype compared with the initial tumor at diagnosis (de novo; refs. 3, 4). The quality of tumor-specific mutant peptides (neoantigens) present in rGBM along with T lymphocyte infiltrates identifies a subset of patients with a more favorable outcome (5). These findings raise the possibility that rGBM and de novo GBM differ immunologically, with potential implications for immunotherapy.

Immunotherapy has become a standard-of-care treatment for many solid cancers, yet therapeutic targeting of the immune system has not produced a significant clinical benefit for most patients with GBM (6). This treatment resistance has been attributed to the ability of GBM to orchestrate an immunosuppressive tumor microenvironment and to a lack of immunogenic neoantigens expressed by tumor cells that have a low mutational burden in de novo GBM (7). Overall, GBM is defined as an immunologically “cold” tumor characterized by a paucity of tumor-infiltrating CD3⁺ T cells but a strong presence of tumor-associated macrophages (8, 9). The abundance of tumor-associated macrophages correlates with worse survival (8, 9).

Several studies suggest that immune-checkpoint blockade with anti–PD-1 as well as vaccines might trigger productive T-cell responses in some patients with GBM (10–13). These studies have been conducted in patients with rGBM, and, thus, the contribution of standard-of-care treatment to any immunologic alterations detected remains undefined. Thus, there is a strong need to examine the immune contexture of both de novo GBM and rGBM to understand the immune response that occurs during GBM evolution.

Here, we performed deep cytometric profiling of tumor-infiltrating immune cells using mass cytometry and multiplex IHC (mIHC) in de novo GBM and rGBM surgical specimens to define the composition and distribution of immune-cell subsets detected in GBM tumor tissues. We found similar proportions of myeloid and T cells in de novo GBM and rGBM, but differences in subsets across tumors, and substantial intratumoral heterogeneity in the spatial distribution of immune infiltrates. We identified increased numbers of CD3⁺pSTAT1⁺ cells in perivascular regions as associated with a longer median time from relapse to death. Collectively, our findings indicate that although both de novo GBM and rGBM display potential targets for immunotherapy, rGBM might have a better potential for response to immunotherapies.

Tumor samples were obtained from patients with GBM after written informed consent and deidentified. Peripheral blood mononuclear cells (PBMC) from healthy volunteers were obtained from the Human Immunology Core at the University of Pennsylvania. Studies were conducted in accordance with the 1996 Declaration of Helsinki and approved by the University of Pennsylvania's Institutional Review Board. For analysis by cytometry by time of flight (CyTOF), tumor samples were obtained from 27 patients (14 with de novo GBM and 13 with rGBM) after confirmation of a GBM diagnosis by a board-certified neuropathologist. Whole tumor was processed by mechanical and enzymatic dissociation using a gentleMACS Octo Dissociator in combination with the Brain Tumor Dissociation Kit (Miltenyi Biotec, cat no. 130-095-942) and then filtered through a 75-μm strainer to generate a single-cell suspension. Samples were frozen in liquid nitrogen until further use. Characteristics of patients included in the CyTOF analysis are listed in Supplementary Table S1. For mIHC analyses, 17 patients who previously had surgery for de novo GBM and then rGBM performed at the University of Pennsylvania were identified through a medical records search. Unstained slides were cut at 5-μm thickness from formalin-fixed paraffin-embedded (FFPE) blocks for analysis. Characteristics of patients included in the mIHC analyses are listed in Supplementary Table S2, and the number of regions analyzed is detailed in Supplementary Table S3.

Thawed single-cell suspensions of tumor or healthy volunteer PBMC were assessed using CyTOF as previously described (14). Briefly, antibody panels (lymphoid, myeloid, and T cell) were designed to simultaneously measure the expression of molecules related to cell lineage, differentiation state, and function (Supplementary Table S4). Cells (3×10⁶) were incubated with cisplatin for 1 minute, then stained at room temperature for 30 minutes with metal-labeled antibodies to detect surface proteins. Cells were then permeabilized using Fixation/Permeabilization solution (eBioscience, Thermo Fisher, cat no. 00-5123-43) at room temperature for 30 minutes and then stained at room temperature for 60 minutes with metal-labeled antibodies to detect intracellular proteins. Cells were washed in permeabilization buffer and resuspended in PBS-1.6% paraformaldehyde before acquisition by CyTOF (Helios, Fluidigm). CD45⁺ cells were identified as Iridium⁺Beads^–Cisplatin^–CD45⁺. Monocytes were identified as CD3^–CD14⁺. B cells were identified as CD3^–CD19⁺. Natural killer cells were identified as CD3^–CD56⁺. Mucosal-associated invariant T (MAIT) cells were identified as CD3⁺CD161⁺TCRV7.2⁺. T-cell receptor gamma/delta (TCRγδ) cells were identified as CD3⁺TCRγδ⁺. Regulatory T cells (Treg) were identified as CD3⁺CD4⁺CD25⁺CD127^–. Follicular helper T (TFH) cells were identified as CD3⁺CD4⁺CXCR5⁺PD-1⁺. Natural killer T (NKT) cells were identified as CD3⁺CD161⁺TCRV24⁺. Analysis was performed using FlowJo v10 (TreeStar), as well as R v3.5.1 and RStudio v1.1.383. R scripts and other resources are available online at https://github.com/wherrylab/Cytof_analysis_calanio.

FFPE tumor specimens were sectioned onto positively charged glass microscope slides. Automated mIHC was performed on a Ventana Discovery Ultra automated slide staining system (Roche). Slides were deparaffinized at 70°C for 24 minutes. Following deparaffinization, heat-induced epitope retrieval was performed using an EDTA-based preconditioning solution (Ventana Medical Systems, cat no. 950-500) for 64 minutes at 100°C. Slides were incubated with the appropriate primary and secondary antibodies prior to chromogenic signal detection using the reagents and antibodies listed in Supplementary Table S5. Tissues were counterstained with hematoxylin (Ventana Medical Systems, cat no. 760-2021) and bluing reagent (Ventana Medical Systems, cat no. 760-2037). Whole slide scans of stained tissues were acquired at 40× using an Aperio CS2 scanner (Leica Biosystems).

Scanned brightfield images were digitally quantified using custom image analysis algorithms. All algorithms were created using Visiopharm Integrator System (VIS) software (Version 2019.07). A board-certified neuropathologist identified and annotated areas of dense and infiltrative tumor from hematoxylin- and eosin-stained slides. Annotations were then transcribed onto scanned multiplex images. Areas of necrosis, nonneoplastic brain, predominantly reactive/treatment-related changes, and artifactual staining were excluded from the analysis. Custom image analysis algorithms were applied to quantify the number of CD3⁺, CD3⁺Ki-67⁺, CD8⁺, Foxp3⁺, CD68⁺, CD163⁺, CD163⁺Ki-67⁺, EGFR⁺p53⁺, HLA-DR⁺, HLA-DR⁺CD68⁺, and Ki-67⁺ cells within the neuropathologist-annotated areas of GBM. Thresholds were similarly applied across all samples to reduce variability. Each Visiopharm algorithm was reviewed to ensure reproducibility and to minimize error and variability with cell segmentation and with detection of cytoplasmic and nuclear features. Absolute cell counts were normalized to the total area of the annotations and reported as densities (cells per mm²).

To visualize the quantity and distribution of CD8⁺ lymphoid aggregates in paired de novo GBM and rGBM specimens, heat maps based on the density of CD8⁺ cells present within a 100-μm drawing radius were generated on annotated tissues. Standardized minimum and maximum thresholds were applied during heatmap generation to allow for comparisons among specimens.

Separately, to facilitate the unbiased selection of T-cell high and T-cell low regions, individual regions of interest (ROI) were delineated around areas of low and high density identified from the heat map. ROI sizes were determined by the overall areas of the identified tumor. These ROIs were carried over to annotate the identical T-cell low and T-cell high regions on serial sections of tissues that were stained for additional immune markers. All quantitative measurements were normalized to the area of the ROIs. To quantify the number of T-cell high loci per case, 1,035 by 1,035-μm square grids were superimposed upon CD8 density heatmaps. The number of grids containing CD8-enriched hotspots was quantified and expressed as a percentage of the total number of grids.

For group comparisons and correlation analyses, testing was performed using PRISM 8.4 (GraphPad Software). Normality of distributions was assessed using D'Agostino–Pearson omnibus normality test, and variance between groups of data was determined using the F-test. For nonnormal data, nonparametric Mann–Whitney U tests were used for unpaired analyses. Descriptive statistics included mean, median, standard deviation, and range for continuous variables and frequency and proportion for categorical variables. For survival discrimination, unsupervised k-means clustering (k = 2) was used to stratify patients into two groups based on overall survival, via the kmeans command in the stats suite of commands on R (version 3.5.3). Ten random starts and 1,000 maximum iterations were used as parameters. K-means clustering demonstrated successful convergence, with stratification of patients into two distinct survival groups minimizing within-cluster variance (between-cluster sum of squares/total sum of squares = 0.701). Cluster analysis was also reperformed with kernel density estimation using the pdfCluster package on R, which determines the optimal number of clusters and partitions variables via nonparametric density estimation. The same survival group assignments were reached through kernel density estimation.

Analyses correlating gene expression and patient survival used the maximal selected rank statistics package (maxstat on R) according to prior studies (9). This methodology was used to select the optimal cutoff for dichotomizing a quantitative predictor, such as gene expression, into high and low groups, when analyzed in association with a quantitative response, such as survival. Subsequently, Kaplan–Meier curves were generated for high- and low-expression groups, and differences in survival were quantified via the maximally selected rank test and independent t tests. Statistical significance was defined as P < 0.05.

The data generated in this study are available within the article and its supplementary data files or from the corresponding authors upon reasonable request.

To examine the immune infiltrate of human GBM, we performed the deep single-cell immune profiling of tumors using mass CyTOF. We evaluated human GBM surgical specimens collected on unmatched patients with de novo GBM (n = 14) or rGBM (n = 13; Fig. 1A; Supplementary Table S1). All rGBM patients had received standard-of-care therapy, comprised of concurrent radio- and chemotherapy, followed by maintenance chemotherapy with temozolomide. Three rGBM patients (010, 012, and 065) received experimental treatment in a clinical trial after completing standard-of-care therapy (Supplementary Table S1).

We used three CyTOF antibody panels (myeloid, lymphoid, and T cell) to measure the expression of up to 40 molecules by tumor-infiltrating immune cells (Supplementary Table S4). All samples were analyzed using a T-cell panel, but due to insufficient material from some samples, analysis with either a myeloid or lymphoid panel was limited to only a subset of patients (de novo GBM, n = 5 and rGBM, n = 5 for the myeloid panel; de novo GBM, n = 10 and rGBM, n = 9 for the lymphoid panel). Using this approach, we found that the CD45⁺ leukocyte infiltrates seen in GBM tissues consisted primarily of myeloid cells (∼60% of CD45⁺ hematopoietic cells) and T cells (∼20%) (Fig. 1B–D). Proportions of myeloid cells and T cells were similar in de novo GBM and rGBM (Fig. 1B–D; Supplementary Table S6). Other cells detected included B cells (∼6%), NK cells (∼4%), MAIT cells (∼1%), TCRγδ cells (∼1%), and rare NKT cells (∼0.05%; Supplementary Fig. S1A).

A more detailed analysis of CD45⁺ myeloid cells identified three major subsets. The Myeloid I subset was the most abundant and expressed CD64, CD11b, HLA-DR, CD80, and low levels of CD14 but lacked expression of CD68 and CD163 (Fig. 1E; Supplementary Fig. S1B). As such, Myeloid I resembled tissue-resident microglial-derived tumor-associated macrophages (TAM; refs. 15, 16). The Myeloid II subset expressed CD68, HLA-DR, CD1c, CD14, CD15, CD163, CD40, and CD73, but lacked expression of costimulatory molecules including CD80 and CD86. As such, Myeloid II is consistent with a mixture of type-1 myeloid dendritic cells (cDC1) and CD163^hi monocyte-derived TAM (15, 16). Finally, the Myeloid III subset expressed CD68, CD24, CD14, CD80, and low CD163 but lacked expression of HLA-DR (Fig. 1E; Supplementary Fig. S1B). As such, Myeloid III resembled CD163^low monocyte-derived TAM (15, 16). Overall, the proportion of the Myeloid I, II, and III subsets varied among patients and between de novo GBM and rGBM tumors, with a tendency for more monocyte-derived (Myeloid II and III) and less tissue-resident (Myeloid I) TAM in the recurrent cases (Fig. 1E). The overall proportions of cells expressing CD163, HLA-DR, CD80, and CD86 were largely equivalent between de novo GBM and rGBM tumors except for a slight increase in the proportion of CD80-expressing myeloid cells detected in de novo GBM tumors (P = 0.01, Fig. 1F). The latter may relate to the increased abundance of CD80-expressing cells in the Myeloid I subset seen in the de novo setting (Fig. 1G). As such, although myeloid cells were equally abundant in de novo GBM and rGBM tumors, their subset composition appeared to differ with an increased abundance of monocyte-derived macrophages in rGBM tumors.

Given that T cells are key mediators of antitumor activity in response to immunotherapy, we next examined T-cell composition in GBM tumors. Overall, the frequency of CD8⁺ T cells was similar for de novo GBM and rGBM. However, de novo tumors contained a higher frequency of CD4⁺ T cells (∼35% vs. 20% in rGBM; Fig. 1H). CD4⁺ T cells for de novo GBM and recurrent tumors were comprised of similar proportions of CD25⁺CD127^– Tregs (∼15%; Supplementary Fig. S1C) and CXCR5⁺PD-1⁺ TFH cells (∼15%; Supplementary Fig. S1D), indicating that the increased frequency of CD4⁺ T cells may relate to an increase in conventional CD4⁺ T cells in de novo GBM tumors. The majority of conventional CD8⁺ T cells expressed PD-1 (∼80%), in both de novo GBM and rGBM tumors (Supplementary Fig. S2A and S2B). PD-1-expressing CD8⁺ T cells also coexpressed multiple coinhibitory immune checkpoints such as TIM-3, LAG-3, and CTLA4, consistent with a substantial proportion of exhausted CD8⁺ T cells (TEX) infiltrating human GBM (Fig. 1I; ref. 17). In both de novo GBM and rGBM, ∼60% of CD8⁺ T cells also expressed both PD-1 and CD39, a combination associated with TEX cells and previously implicated as identifying tumor-reactive T cells in other settings (ref. 18; Supplementary Fig. S2C and S2D).

In settings of chronic infection or cancer in mice, populations of progenitor TEX cells (CXCR5⁺CD69^+/–), transitional or intermediate TEX cells (CXCR5^–CD69^–), and terminal TEX cells (CXCR5^–CD69⁺) have been defined with distinct functional properties (19–23). For example, progenitor TEX cells have a key role in response to immune-checkpoint blockade (24–26). In addition, progenitor TEX cells contain the highest frequency of IFNγ and TNF producers and are the subset of TEX cells that undergoes proliferative expansion in response to PD-1 pathway blockade (19). In contrast, intermediate TEX cells contain more migratory cells that are in the cell cycle and are derived from progenitor TEX cells. Intermediate TEX cells eventually transition to terminal TEX cells that display features of tissue residency (19, 20, 23). Based on this biology, we evaluated expression of CXCR5 and CD69 on PD-1⁺CD8⁺ T cells in GBM tumors. In both de novo GBM and rGBM, approximately 60% to 75% of PD-1⁺CD8⁺ T cells expressed CXCR5, with about 50% expressing CD69. In addition, most CD69⁺CD8⁺ T cells coexpressed CXCR5 with 10% to 15% lacking CXCR5 (Supplementary Fig. S3A and S3B). This pattern is suggestive of a predominantly progenitor TEX cell phenotype detected in GBM tumors (19, 27). Thus, our data show that CD8⁺ T-cell infiltrates in GBM are likely to be tumor reactive based on high PD-1 and CD39 coexpression (18) and exist predominantly in a progenitor TEX cell state.

Although myeloid and T cells appeared to be similarly abundant phenotypically between de novo GBM and rGBM, we hypothesized that their spatial distribution may be altered. To explore this possibility, we performed mIHC on 17 paired de novo GBM and rGBM tumor tissues (Supplementary Table S2). Here, we focused our analyses on ROIs containing the tumor, an approach that contrasts with our CyTOF studies, which examined cells present throughout the entire tumor tissue specimen. Despite this difference, mIHC revealed a similar overall content of tumor-infiltrating CD8⁺ T cells in the paired de novo GBM and rGBM samples (Fig. 2A). CD8⁺ T cells were detected throughout the tumor microenvironment, but the density of CD8⁺ T cells was greatest in perivascular regions for both de novo GBM and rGBM (Fig. 2B), consistent with previous findings (28). Visual examination of CD8 staining showed qualitative differences in the distribution of CD8⁺ T cells between paired de novo GBM and rGBM samples (Fig. 2C). This difference was due to an increased number of perivascular regions associated with CD8⁺ T cells seen in rGBM samples compared with paired de novo GBM samples (Fig. 2D) and appeared to be independent of sex as a biological variable (Supplementary Fig. S4).

Next, we applied a series of mIHC panels and computer-assisted image analysis algorithms to define the immune microenvironment of perivascular T cell–enriched regions present in the setting of rGBM (Fig. 2E; Supplementary Fig. S5). Heatmaps displaying the density of CD8⁺ cells in tumor regions (Fig. 2C) were used to define CD8⁺ T-cell low (“T-cell low”) and CD8⁺ T-cell high (“T-cell high”) regions (Supplementary Table S7). This approach identified proliferating T cells, as defined by Ki67 expression, mainly in T-cell high regions in rGBM (Fig. 2F and G). In addition, pSTAT1 expression, a marker of interferon signaling, was frequently associated with T-cell high regions, although some rGBM samples lacked pSTAT1 expression, and not all T-cell high regions were associated with pSTAT1 (Fig. 2F and G). Similarly, the myeloid compartment in T-cell high regions was enriched in HLA-DR⁺ and pSTAT1⁺ myeloid cells (Fig. 2F and H). In contrast, the density of CD68⁺ macrophages that lacked HLA-DR was similar between T-cell low and T-cell high regions in rGBM samples (Supplementary Fig. S6).

We next asked whether immune features present within T-cell high regions differed between de novo GBM and rGBM. Perivascular T-cell high regions in rGBM displayed a lower density of FOXP3⁺ Tregs and a significantly higher ratio of CD8⁺ T cells to FOXP3⁺ Tregs compared with T-cell high regions in de novo GBM (Fig. 2I and J). Moreover, some perivascular T-cell high regions in rGBM also showed a higher expression of pSTAT1⁺ cells compared with perivascular T-cell high regions in de novo GBM, although this comparison did not meet statistical significance (P = 0.08; Fig. 2K and L). Collectively, these data suggest that rGBM displays features of an immunologically active tumor that are distinct from the de novo tumor, particularly in the perivascular regions where immune cells may accumulate.

A key question is whether intra- or peritumoral immune features have any relationship to disease outcomes. Thus, we next examined the relationship between immune features present in T-cell high regions in rGBM and survival outcomes. Patients were dichotomized into either short-term survivors (STS) or long-term survivors (LTS) using an unsupervised k-means clustering algorithm based on the median overall survival from the time of diagnosis (Fig. 3A). The median overall survival in the STS group was 11.5 months and in the LTS group was 27.5 months (P < 0.0001). In the LTS group, T-cell high regions contained significantly fewer cells expressing Ki-67 (Fig. 3B and C). To investigate whether this Ki-67 difference reflected changes in tumor-cell proliferation, we excluded proliferating T cells (CD3⁺Ki-67⁺) and macrophages (CD163⁺Ki-67⁺) from the total population of Ki-67⁺ cells detected in T-cell high regions. Consistent with prior work (29), the residual CD3^–CD163^–Ki-67⁺ population, presumably composed primarily of tumor cells, showed the same Ki-67 relationship with survival, with the LTS cohort exhibiting lower numbers of Ki-67⁺ tumor cells in T-cell high regions (Fig. 3C). T-cell high regions in LTS patients were also enriched in CD68⁺ and HLA-DR⁺CD68⁺ macrophages (Fig. 3D and E). Of note, LTS patients displayed a significantly higher density of pSTAT1⁺ cells, but the density of CD3⁺ cells coexpressing pSTAT1 did not reach significance (Fig. 3F and G). Taken together, these data suggest that the immunologic microenvironment of T-cell high regions in rGBM may harbor prognostic value for overall survival.

To further examine the relationship between spatial immune microenvironmental features and disease outcome, we evaluated immunologic features defined by the mIHC analyses and their association with clinical outcomes in rGBM. To do this, we stratified patients with rGBM (Supplementary Table S2) into two groups using the maximal selected rank statistics approach based on all expression features individually. We then tested their association with survival using log-rank tests and corrected for multiple testing. The CD3⁺pSTAT1⁺ binary classification (low vs. high) identified six patients with high expression and eight patients with low expression. Although no difference in overall survival was observed based on this classification, patients in the CD3⁺pSTAT1⁺ high group had a longer median time from relapse to death (high, 10 months; low, 4 months; P = 0.0223; Fig. 4). This extended time to relapse was confirmed in a Cox proportional hazards model including age, sex, Ki67, MGMT, EGFR, TP53, and extent of resection at de novo status as predictor variables (for CD3⁺pSTAT1⁺P = 0.0223). These data suggest that the T-cell compartment, especially in the perivascular regions of the tumor, in some patients with rGBM may be polarized toward an antitumor activated phenotype with clinical implications.

In this study, we examined contextual and spatial differences in the immune landscape of de novo GBM and rGBM tumors and identified immunologic features with potential clinical implications. Using mass cytometry, myeloid cells and T cells were found to be similarly present in de novo GBM and rGBM. Myeloid cells were the most abundant cells in the tumor microenvironment and were accompanied by infiltrating TEX cells, consistent with prior studies (30–32). This finding is also consistent with recent work comparing the microenvironmental landscape in glioma and brain metastases (15, 16). In these studies, the myeloid-cell composition of gliomas was shown to be markedly heterogeneous and was composed of tissue-resident and -infiltrating leukocytes. Our study is consistent with these reports and provides new insights into variations in the leukocyte infiltrate present in de novo GBM and rGBM. For example, we identified differences in the composition of myeloid subsets detected in de novo GBM and rGBM samples. Although we were not able to definitively distinguish between tissue-resident and monocyte-derived macrophages in our studies (15, 16), rGBM tended to have an increased frequency of myeloid cells that resembled monocyte-derived macrophages. Prior studies have shown that a high infiltration of monocyte-derived macrophages is associated with poor survival in GBM patients (15). In contrast, de novo GBM tended to have an increased frequency of myeloid cells that resembled tissue-resident macrophages. It is possible that these differences reflect a response to standard-of-care therapy (33). Alternatively, they may indicate cancer cell–intrinsic differences between de novo GBM and rGBM tumors that differentially shape the recruitment and in situ differentiation of monocyte-derived macrophages.

Differences in T-cell subsets were also observed between de novo GBM and rGBM. With deep profiling, TEX cells were detected in both de novo GBM and rGBM, with most of these cells expressing CXCR5, as well as potential therapeutic targets such as PD-1, LAG-3, and CTLA-4. Notably, these TEX cells expressed CD39 together with PD-1, suggesting that they may be tumor reactive (18). Using mIHC, we then further tested if the spatial distribution of infiltrating immune cells differed between de novo GBM and rGBM. T-cell high perivascular regions were more abundant in rGBM compared with de novo GBM and differentially displayed markers of immune activation (pSTAT1) and cellular proliferation (Ki67). Increased numbers of CD3⁺pSTAT1⁺ cells in perivascular regions associated with a longer median time from relapse to death. Taken together, our findings identify the spatial distribution of T cells rather than their abundance as a potential key immunologic determinant that is associated with the evolution and pathogenesis of GBM.

Immune infiltration is a recognized determinant of outcomes in cancer. In de novo GBM, myeloid-cell infiltrates are associated with an unfavorable prognosis (8, 9). In contrast, variable correlations with survival have been reported for T-cell infiltrates (9, 34). However, less is known about the immunologic features of rGBM, where tumors are known to have acquired molecular alterations, including increased alternative-splicing events, that may influence tumor immunogenicity (35). Here, our studies identified distinct spatial organization and activation characteristics of immune cells in de novo GBM versus rGBM tumors. Specifically, T cell–enriched perivascular regions were more abundant in rGBM compared with de novo GBM. Perivascular spaces have been recently identified as lymphopoietic niches at the central nervous system border (36, 37). It is also known that the perivascular space is important for maintaining the self-renewal and stemness of cancer cells (38, 39). Our findings suggest that enrichment of activated T cells in perivascular regions may be a determinant of longer survival in patients with rGBM. This observation raises the possibility that tumor-reactive T cells in perivascular regions may control the fate of glioma stem cells. Consistent with this idea, cytotoxic T cells engineered to recognize tumor antigens (e.g., HER2 and NKG2D ligands) or redirected against CD133, a marker of stem cells, have shown antitumor activity against glioma stem cells (40–42). However, it remains unclear whether endogenous tumor-reactive T cells are present in perivascular regions and capable of eliminating glioma stem cells in patients. In contrast, prior studies have shown that glioma stem cells can suppress T-cell proliferation and activation (43). Nonetheless, manipulating the perivascular niche may be a strategy to leverage the antitumor potential of infiltrating leukocytes (44).

Our findings of abundant T cell–enriched perivascular regions in rGBM may indicate their potential role in response to therapeutic interventions. In line with this idea, we found that T-cell activation in the perivascular region associated with increased time from relapse to death. Specifically, a high density of CD3⁺ T cells expressing pSTAT1 was found suggesting recent signaling from interferons or other inflammatory mediators. Interferons can have antitumor activity but may also support immune evasion by triggering the upregulation of immune-checkpoint molecules (45). In GBM, interferon signaling has been previously associated with a poor prognosis (46). However, our data differ from these analyses given that we focused on pSTAT1 expression in T cell–enriched regions of GBM. It will be important in subsequent analyses to better understand the signaling programs that distinguish T cell–enriched regions defined by high and low pSTAT1 expression.

Perivascular regions in GBM are associated with monocyte-derived macrophages, which portend a poor outcome (15). TAMs in these regions support tumor growth by producing soluble factors that sustain glioma stemness (47). However, we found that an increased density of CD68⁺ myeloid cells expressing HLA-DR associated with increased survival. These changes seen in the activation status of both myeloid cells and T cells may reflect antitumor immune activity and/or treatment-induced inflammation. In addition, we found that not all perivascular regions were infiltrated by T cells in rGBM, and the degree of infiltration was variable between patients, thereby illustrating both intra- and interpatient heterogeneity in the immune response to GBM. Analyses of intercellular communications and the phenotypical state of cells in perivascular regions after various treatment regimens are warranted.

Our studies were conducted in a small cohort of patients with de novo GBM and rGBM and, thus, may not be representative of all subsets of GBM. Our analyses also assessed only a small cross-sectional tumor region of GBM for each patient, raising the possibility of sampling bias due to the prominent intratumoral heterogeneity that is characteristic of GBM. In addition, our studies addressed defined time points, namely, surgical resection for de novo GBM and rGBM, and thus, did not address the dynamics of GBM pathogenesis. In this regard, diagnostic imaging strategies that incorporate tissue biomarkers may enable a whole lesion assessment and permit noninvasive longitudinal analyses.

Immunotherapy has yet to reproducibly demonstrate activity in GBM. For example, monotherapy with pembrolizumab has shown minimal benefit so far (48). However, recent work suggests that combining immune-checkpoint blockade (e.g., anti–PD-1) with tumor vaccines might trigger productive T-cell responses in some patients with GBM (10–13). These studies have all been conducted in patients with rGBM and whether targeting de novo tumors would be more impactful is unclear. Our findings revealed abundant T cell–enriched perivascular regions in rGBM and show that immune activation is associated with delayed disease progression. Thus, we propose that rGBM may have a greater potential for response to immunotherapies.

In summary, our study identifies the perivascular space as a potential immune determinant of GBM biology. Specifically, increased T-cell infiltration into perivascular regions was seen in rGBM compared with de novo GBM and T-cell activation in perivascular regions correlated with improved outcomes. Thus, the perivascular space in GBM may be a key immunologic target with therapeutic implications.

C. Alanio reports grants from Parker Institute during the conduct of the study and is a consultant at Biotherapy Partners. Z.A. Binder reports a patent for 17/005227 pending and licensed to Tmunity and a patent for 2021-560665 pending and licensed to Tmunity. E. Wherry reports grants from NIH and Parker Institute for Cancer Immunotherapy during the conduct of the study; personal fees from Merck, Marengo, Janssen, Related Sciences, Rubius Therapeutics, Synthekine, Surface Oncology, Arsenal Biosciences, and other support from Danger Bio outside the submitted work. D.M. O'Rourke reports grants from Tmunity Therapeutics during the conduct of the study; personal fees and other support from Prescient Therapeutics, other support from Implicyte, and personal fees from Century Therapeutics outside the submitted work; in addition, D.M. O'Rourke has a patent for compositions and methods comprising a high-affinity CAR with cross-reactivity to clinically relevant EGFR mutated proteins pending and licensed to Tmunity Therapeutics and a patent for combination cell therapy approaches for GBM targeting IL13Ra2 and EGFR variants pending and licensed to Tmunity Therapeutics. G.L. Beatty reports grants from NIH during the conduct of the study; personal fees from Seagen, Boehinger Ingelheim, Cour Pharmaceuticals, Adicet Bio, Aduro Biotech, AstraZeneca, BioMarin Pharmaceuticals, Cantargia, Opsona, Merck, Molecular Partners, Nano Ghosts, BiolineRx, and Shattuck Labs, grants and personal fees from Bristol-Myers Squibb, Genmab, Incyte, Janssen, Monopteros, Pancreatic Cancer Action Network, Verastem, grants from Hibercell, Arcus, Newlink, and Halozyme outside the submitted work; in addition, G.L. Beatty has a patent for human CAR T cells (10,640,569, 10,577,417) licensed to Tmunity Therapeutics and Novartis. No disclosures were reported by the other authors.

C. Alanio: Formal analysis, writing–original draft, writing–review and editing. Z.A. Binder: Formal analysis, writing–original draft, writing–review and editing. R.B. Chang: Formal analysis, writing–original draft, writing–review and editing. M.P. Nasrallah: Formal analysis, writing–original draft, writing–review and editing. D. Delman: Formal analysis, writing–review and editing. J.H. Li: Formal analysis, writing–review and editing. O.Y. Tang: Formal analysis, writing–review and editing. L.Y. Zhang: Formal analysis, writing–review and editing. J.V. Zhang: Formal analysis, writing–review and editing. E.J. Wherry: Supervision, writing–review and editing. D.M. O'Rourke: Supervision, writing–review and editing. G.L. Beatty: Formal analysis, supervision, writing–original draft, writing–review and editing.

The authors thank members of the Wherry, O'Rourke, and Beatty laboratories for critical feedback, the Pathology Clinical Service Center of Penn Medicine for sample acquisition and slide cutting, and the Glioblastoma Translational Center of Excellence for coordinating the work. The authors also acknowledge and thank the patients from whom tissue samples were donated and made this study feasible. The study and analyses were funded by R01-CA197916 (G.L. Beatty), The GBM Translational Center of Excellence (Z.A. Binder, L.Y. Zhang, J.V. Zhang, and D.M. O'Rourke), The Templeton Family Initiative in Neuro-Oncology (D.M. O'Rourke), The Maria and Gabriele Troiano Brain Cancer Immunotherapy Fund (D.M. O'Rourke), and the Neurosurgery Research and Education Foundation (O.Y. Tang). E.J. Wherry was supported by NIH grants CA016520, CA210944, AI149680, AI155577, AI108545, and AI082630. E.J. Wherry and C. Alanio were also supported by the Parker Institute for Cancer Immunotherapy which supports the cancer immunology program at the University of Pennsylvania.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.