Exploring the immune environment of glioblastoma in humanized mouse models

Abstract

Background:

Glioblastoma (GBM) is the deadliest primary brain tumor in adults, where current therapies fail to extend survival meaningfully. Available animal GBM tumor models, especially therapy-resistant and recurrent models with human tumor and human immune cell interactions, are limited, impeding innovative treatment research. To address this critical obstacle, we established a unique GBM mouse model using patient-derived xenografts (PDXs) in humanized mice.

Methods:

We selected two immunodeficient mouse models that express key human cytokines required for the proper reconstitution of myeloid lineage cells. After undergoing myeloablation, mice received CD34+ hematopoietic stem progenitor cells derived from human umbilical cord blood for humanization. Upon confirming the reconstitution of human blood cells, mice were xenografted with radiation-resistant PDXs. Tumor profiles and immune cell infiltration were analyzed via spectral flow cytometry, immunohistochemistry, and single-cell RNA sequencing (scRNA-seq). The results were benchmarked against scRNA-seq data from patients with recurrent human GBM.

Results:

A diverse range of human immune cells, including T cells, natural killer cells, and myeloid lineage cells, infiltrated PDX tumors in humanized mice. Notably, gene expression profiles in these immune cells resembled those of recurrent human GBM. Unlike conventional xenograft models, this model highlighted enhanced tumor diversity, particularly a high fraction of neural progenitor-like cells.

Conclusions:

Our humanized GBM mouse model exhibited an immune cell signature similar to that of human recurrent GBM. This model is a valuable resource for analyzing the tumor immune landscape and assessing new therapies, particularly immunotherapies.

Introduction

Glioblastoma (GBM; 2021 WHO CNS grade 4) is the most frequent and lethal type of malignant brain tumor in adults, with an average survival of 15 months.^1,2 Patients initially respond to standard therapies, including surgical excision, chemotherapy, and radiotherapy. Still, tumors recur and acquire resistance at a significant rate, leading to patient demise. In recent years, extensive genomic characterization of GBM has been performed, uncovering key oncogenic signaling pathways that drive the disease.³ However, strategies targeting these pathways have not significantly improved patient survival. Finding innovative approaches to lower mortality rates among GBM patients presents a significant challenge, which can only be overcome by enhancing our comprehension of the mechanisms that enable tumors to adapt to treatment and develop resistance.

GBM tumor tissue consists of a combination of neoplastic and stromal cells. The majority of non-neoplastic cells in tumors are immune cells, accounting for nearly 30% of the total volume,⁴ primarily consisting of microglia that reside in the brain and macrophages that originate in the bone marrow.⁵ Single-cell RNA-sequencing (scRNA-seq) analyses have shown that the immune populations present in the tumor microenvironment (TME) of GBM become more diverse upon tumor recurrence following radio- and chemotherapy.⁵ A higher percentage of T cells, natural killer (NK) cells, and B cells is found in recurrent tumors.⁵ However, these immune cells fail to control tumor growth due to various tumor-related immune suppressive mechanisms, including the recruitment of regulatory T cells (Tregs), production of immunosuppressive cytokines, and up-regulation of checkpoint ligands.⁶ Developing effective treatments for GBM and assessing their in vivo efficacy requires understanding the interactions between GBM cells and the immune system. Nevertheless, existing GBM tumor models, which involve transplanting human cancer cells into immunodeficient mice or syngeneic mouse tumor cells into immunocompetent mice, are insufficient for investigating the human immune microenvironment in GBM.⁷

A humanized mouse model reflecting human immune cell dynamics has been established and is widely used to investigate tumor biology, immunology, and therapeutic interventions.^8,9 This model humanizes the murine blood system through myeloablation and subsequent implantation of human hematopoietic stem progenitor cells (HSPCs), often sourced from umbilical cord blood.¹⁰ Such a model offers a human blood system, facilitating in vivo research into human immune responses, hematopoiesis, immunology, and related diseases, including cancer, within a mouse host. We have previously developed a human B-cell lymphoma humanized mouse model. It demonstrated its use in assessing a novel anticancer treatment, including its ability to promote antitumor memory responses.^8,9 Diverse humanized mouse systems have also been utilized to examine the pathophysiology of GBM and to assess the effectiveness of treatment modalities using xenografted human GBM cell lines.^11,12 These studies indicate that human GBM cells can generate tumors in humanized mice, which exhibit pathological attributes like those observed in patients with GBM. Flow cytometric analyses also indicated that various immune cell types, including macrophages, CD4+ T cells, CD8+ T cells, and NK cells, had infiltrated these tumors.¹¹ Nevertheless, earlier models employed standard NOD.Cg-Prkdc^scid Il2r^gtm1Wjl/SzJ (NSG) or the closely associated DRAG strain for humanization, leading to the inadequate progression of human myeloid lineage cells and Tregs because of the absence of vital human cytokines required for their development.^8,9,13–16 Myeloid cells are the most prominent immune cells in the patient’s GBM microenvironment, comprising 30–50% of the tumor mass.⁵ Several studies indicate that they make the tumor immunologically “cold”. Proficient myeloid lineage reconstitution is critical for adaptive immune responses¹⁷ as well as the differentiation of Tregs via enhanced reconstitution of dendritic cells (DCs).¹⁸ Several new strains of mice were created to overcome the limitations of first-generation humanized mice.^19,20 NSG-SGM3²¹ and NOG-EXL²² second-generation models, which cover almost the entire spectrum of human immune cells except microglial cells, are exciting. They still retain endogenous mouse CD45+ cells but lack a functional murine adaptive immune system.^21,22 Both strains are derived from NOD-SCID (NOD.Cg-Prkdc^scid) mice and differ in the degree of inactivation of the IL2 common γ chain (IL2rg) gene and in the presence of human cytokine transgenes.²³ The NSG-SGM3 expresses human interleukin-3 (hIL3), granulocyte-macrophage colony-stimulating factor (hGM-CSF), and stem cell factor (hSCF), while the NOG-EXL expresses only hIL3 and hGM-CSF. Moreover, the NOG-EXL strain has nearly 10 times lower cytokine production than NSG-SGM3, thereby suppressing aberrant macrophage activation and HSPC exhaustion, enabling prolonged animal monitoring in research contexts.^20,24

A thorough exploration of the immune cell presence in GBM tumors in these advanced humanized mouse models has not been performed. Furthermore, models of recurrent GBM, which exhibit a more profound immunosuppressive microenvironment, remain largely unexplored. To address this lack of knowledge, we examined the immune landscape of radio-resistant human GBM patient-derived xenograft (PDX) tumors in NSG-SGM3 and NOG-EXL humanized mice. We used immunohistochemistry, spectral flow cytometry, and single-cell transcriptomic analysis to characterize and quantify infiltrating immune populations. The collected data reveal that established human GBM PDX models in humanized mice create an immunosuppressive microenvironment and alter tumor cell characteristics. These models will be valuable for investigating immune-tumor interactions and guiding the development and optimization of therapies for recurrent GBM.

Materials and Methods

Mouse humanization

The NSG-SGM3 mice were purchased from the Jackson Laboratory (Bar Harbor, ME). The NOG-EXL mice were obtained from Taconic Biosciences (Germantown, NY). Humanization of NSG-SGM3 and NOG-EXL mice was performed as reported previously.¹³ Before humanization with human CD34+ cells, 1–3-day-old mouse pups were irradiated with X-rays at 90 cGy a day in advance. Alternatively, adult NOG-EXL mice (8 weeks old) were conditioned by intraperitoneal injection of busulfan (35 mg/kg).^8,9 Purified CD34+ cells (0.1 × 10⁶ cells/mouse) were suspended in 20 μL of AutoMACS buffer (Miltenyi Biotec, Gaithersburg, MD) and injected into the mice intravenously.²⁵ Human CD45 levels in peripheral blood (PB) served as the metric for assessing human blood cell reconstitution 12–14 weeks following humanization.

Single-cell RNA Sequencing (scRNA-seq)

The samples of PB, tumor-infiltrating lymphocytes (TILs), and tumor cells were obtained from NOG-EXL mice bearing EGFP+ JX-14P-RT cells.²⁶ Following the manufacturer’s protocol, scRNA libraries were prepared using the Chromium Single Cell 3’ Reagent Kits (10x Genomics, Pleasanton, CA). Gene expression matrices were generated by the 10x Genomics CellRanger v6.1. The detailed methodology is described in the Supplementary Methods.

A further detailed description of these procedures is provided in the Supplementary Methods, including isolation of CD34+ cells from umbilical cord blood, cell culture, mice, conventional flow cytometry, spectral flow cytometry, immunohistochemistry, immunofluorescence, single-cell RNA sequencing (scRNA-seq), public single-cell data analysis, isolation of neural-progenitor (NPC)-like tumor cells from tumors, cell proliferation assays, sphere and colony formation assays, and statistical analysis.

Results

Humanized mice efficiently support GBM-PDX tumor growth.

To inventory the array of human immune cells infiltrating GBM tumor tissues in a xenograft humanized mouse model, we used two independent immunodeficient mouse strains that support enhanced reconstitution of the human myeloid lineage. The NSG-SGM3 strain supports enhanced myeloid differentiation, while the NOG-EXL strain enables extended monitoring of humanized mice post-tumor grafting.²⁴ Humanized mice were prepared as previously reported (Figure 1A).^8,9,13 The average frequency of humanization in mouse PB was 63.9% (50.1–73.5%) in NSG-SGM3 and 66.1% (45.2–97.0%) in NOG-EXL (Figure 1B).

Figure 1. Human PDX GBM cells formed aggressive tumors in humanized mice.

(A) Schematic showing the experimental design to establish humanized mice. Human hematopoietic stem progenitor cells (HSPCs) were isolated from human cord blood. NSG-SGM3 or NOG-EXL newborn mice were irradiated with 90 cGy for successful engraftment. Alternatively, we used 8 week old NOG-EXL for humanization after 35 mg/kg busulfan treatment. 0.1 × 10⁶ HSPCs were injected intravenously. HSPCs were engrafted in the mice, and the human immune system was reconstructed. HSPCs differentiate into lymphoid and myeloid lineage cells. (B) The ratio of humanization in peripheral blood in humanized mice on week 9 (NSG-SGM3; n = 7, NOG-EXL; n = 32). (C) Histopathologic analysis of brain tumors in naive and humanized mice. Scale bars = 1 mm (left), 100 μm (middle), and 50 μm (right). (D) Survival curves for mice implanted with 4 × 10⁵ tumor cells (JX39P-RT) in NSG-SGM3 naive and humanized mice. (7 mice/group; log-rank test). (E) Bioluminescence images of mice 2 weeks after tumor implantation. (F) Quantification data of (E) (Mann-Whitney U-test). (G) Histopathologic analysis of whole spinal cord in naive and humanized mice. Disseminated tumor growth is identified in the area indicated by the arrows. Scale bars = 1 mm.

We initially investigated the capacity of NSG-SGM3 and NOG-EXL mice to support human GBM PDX growth in the presence of myeloid cells. Understanding this is key for GBM, where myeloid cells drive tumor progression.²⁷ Flow cytometric analysis of PB confirmed that the NSG-SGM3 strain had stable engraftment of human myeloid lineage cells, including monocytes, macrophages, and granulocytes, in addition to NK cells. In contrast, these human immune cell populations were absent in the naive mice (Figure S1).

As a source of human GBM, we selected the highly malignant PDX line, JX39P-RT, which recapitulates the characteristics of recurrent GBM.²⁶ The JX39P-RT line was developed from the parental line, JX39P, by subcutaneous implantation in nu/nu mice, followed by fractionated irradiation (2 Gy × 6 fractions) across multiple rounds with regrowth between treatments.^26,28 These cells harbor genetic alterations in EGFRvIII and CDKN2A, commonly observed in primary GBM and often retained at recurrence.²⁹

Following orthotopic xenografting into the right forebrain, the JX39P-RT cells formed aggressive tumors in both naive and humanized mice in two strains (Figure 1C and Figure S2A). Pathological examination of the tumors revealed the pathognomonic features of human GBM,³⁰ represented by pseudopalisading, hemorrhage, microvascular proliferation, necrosis, active proliferation, and cellular heterogeneity. Tumor histopathology was nearly identical between naive and humanized mice. Humanized mice showed a modest trend toward shorter survival compared with naive mice; however, in NSG-SGM3 mice, this difference did not reach statistical significance (Figure 1D and Figure S2B). JX39P-RT recapitulated the recurrent GBM tumor with a unique pattern of selective spinal cord metastasis following intracranial implantation. Intracranial tumor burden was comparable, but spinal cord dissemination decreased in humanized mice in both models, especially in the NSG-SGM3 strain (Figure 1E–G and Figure S2C–E). Despite these differences in metastatic spread, humanization did not alter the infiltrative invasion pattern of brain tumors (Figure S3).

Similar findings emerged from additional radioresistant GBM cells, JX14P-RT (PDGFRA amplification)²⁶, in the NOG-EXL strain (Figure S4). The results indicate that both human PDX lines develop aggressive intracranial GBM tumors in severely immunocompromised NSG-SGM3 or NOG-EXL mice, irrespective of human immune cell presence. However, human immune cells can modulate select aspects of tumor behavior.

A diverse range of human immune cells infiltrates GBM-PDX tumors in the brains of the xenografted humanized mice.

The restricted distribution of GBM-PDX tumor cells in the spinal cords of humanized mice suggested that human immune cells play a role in the in vivo behavior of GBM tumors. We next examined the proportions of human CD45+ immune cells (125,151 cells total) in the brains and blood of humanized NOG-EXL mice transplanted with JX39P-RT (n=6 mice) or JX14P-RT (n=4 mice) using spectral flow cytometry with a pan-immune panel (Table S1). Human CD45+ immune cells in tumor tissue were compared with those in PB. Unsupervised clustering and Uniform Manifold Approximation and Projection (UMAP) visualization revealed robust reconstitution of diverse human immune cell populations in both the brain and PB (Figure 2A). Each cluster was annotated by expression of lineage markers (Figure 2B,C). Distinct cell distribution patterns emerged from comparing these compartments. The brain had enriched Tregs, exhausted CD4+ and CD8+ T cells; conversely, PB largely comprised B cells, naive CD4+ T cells, and CD4+ T cells (Figure 2D,E). These findings indicated preferential accumulation of immune-suppressive and dysfunctional T cell subsets within the TME.

Figure 2. Spectral flow cytometry analysis of tumor-infiltrating human immune cells in humanized GBM-PDX models.

(A) UMAP visualization of major human immune cell populations isolated from tumors [JX39P-RT (n = 6) and JX14P-RT (n = 4)] and peripheral blood [JX39P-RT (n = 5) and JX14P-RT (n = 4)] in humanized mouse models, identified using pan-immune spectral flow cytometry. Clusters include CD4+ and CD8+ T cells (naive and exhausted states), regulatory T cells (Tregs), natural killer (NK) cells, B cells, monocytes/macrophages (Mono/Macro), and granulocytes. (B) Dot plot of all marker expression across immune subsets. Dot size indicates the fraction of cells expressing each marker, and color represents mean expression. (C) UMAP plots displaying marker expression levels for hCD3, hCD4, hCD8, hCD62L, hPD1, hCD25, hCD14, hCD19, hCD56, hCD66b, hCD127, and hCD69. (D) UMAP plots comparing CD45+ immune cell compositions between JX14P-RT and JX39P-RT tumors and peripheral blood in NOG-EXL humanized mice. (E) Quantification of immune subset frequencies in tumors and peripheral blood. Percentages represent the proportion of each subset within the total human CD45+ population.

Flow cytometry quantification further confirmed that CD3+ T cells and CD14+ monocytes/macrophages were the major immune subsets in the tumors, comprising approximately 20–80% and 10–45% of CD45+ cells, respectively (Figure S5A,B). Immune cell compositions were broadly similar between the JX39P-RT and JX14P-RT models, differing only in the proportion of CD45+ cells among total live cells (Figure S5B).

Spectral flow cytometry utilizing a myeloid panel further dissected the myeloid compartment (Table S2). The gating strategy for myeloid panel analysis is shown in Figure S6A. Unsupervised clustering and UMAP visualization of human CD45+ myeloid cells (20,151 cells total) revealed distinct myeloid subpopulations in both the brain and PB (Figure 3A). Each cluster was annotated based on the expression of canonical myeloid markers (Figure 3B,C). Comparative analysis between these compartments demonstrated striking differences in myeloid composition: the brain TME was predominantly composed of macrophages with a modest M2 polarization bias, whereas PB was comprised mainly of monocytes (Figure 3D,E). These findings suggest active recruitment and differentiation of monocytes into tumor-associated macrophages within the GBM microenvironment.

Figure 3. Characterization of tumor-associated myeloid populations in humanized GBM-PDX models.

(A) UMAP visualization of human myeloid cells isolated from tumors [JX39P-RT (n = 6) and JX14P-RT (n = 4)] and peripheral blood [JX39P-RT (n = 5) and JX14P-RT (n = 4)] in humanized mouse models, identifying major subclasses including polymorphonuclear myeloid-derived suppressor cells (PMN-MDSCs), classical monocytes, monocytic myeloid-derived suppressor cells (M-MDSCs), intermediate monocytes (Int. Monocytes), monocyte-derived dendritic cells (moDCs), M1-like macrophages, M2-like macrophages, conventional dendritic cells type 1 (cDC1s), and conventional dendritic cells type 2 (cDC2s). (B) Dot plot of canonical markers used to define each myeloid population (hCD14, hHLA-DR, hCD16, hCD33, hCD64, hCD206, hCD86, hCD11c, hCD141, hCD1c, hCD123, and hCD15). (C) UMAP plots showing expression distribution of selected markers across the UMAP clusters. (D) Distribution of myeloid subsets in tumors and matched blood samples from JX14P-RT and JX39P-RT models. (E) Relative abundance of each myeloid population in tumors and peripheral blood. Percentages represent the proportion of each subset within the total human CD45+ CD11b+ population.

A comparative analysis between JX14P-RT and JX39P-RT models in NOG-EXL mice revealed that JX14P-RT tumors had significantly higher numbers of human CD45+ cells (Figure S6B), consistent with the pan-immune panel results (Figure S5B). Despite certain variations observed in myeloid subsets, most of them showed no significant difference (Figure S6B). To better understand myeloid cell function in humanized mice, we analyzed the residual mouse microglia in NOG-EXL mice, as their entire myeloid compartment is impaired. Flow cytometry using a mouse microglia panel (Table S3) showed no increase in mouse microglia numbers or upregulation of P2RY12 or CD86 expression in humanized mice (Figure S7A,B), suggesting that mouse humanization contributes minimally to mouse microglia within GBM PDX. The findings indicate that humanized immune systems successfully engrafted into human GBM xenograft models, distributing diverse human myeloid cells in the TME.⁵

The immune landscape within the tumor evolves during tumor progression.

To understand immune profile differences between humanized mouse strains and inter-individual variability, we analyzed immune cells in tumor and PB samples from NSG-SGM3 (n=7) and NOG-EXL (n=6) mice bearing JX39P-RT tumors via flow cytometry. PB samples from both models showed abundant B and CD4+ T cells (Figure S8A,B). Monocytes tended to be more frequent in NOG-EXL mice. Within tumors, CD4+ and CD8+ T cells were the predominant cell populations in most NSG-SGM3 mice, although some mice exhibited increased infiltration of NK cells, monocytes/macrophages, or granulocytes (Figure S8A). B cells were rarely detected. A similar trend was observed in NOG-EXL mice, though monocyte/macrophage infiltration was more pronounced (Figure S8B). To determine whether our findings matched human GBM immune traits, we analyzed a public GBM scRNA-seq dataset (GSE211376),³¹ extracting immune cell proportions from 53 primary and 26 recurrent cases. Consistent with prior reports^5,32, recurrent tumors had higher levels of lymphoid infiltration than primary tumors (Figure S9A,B). In humanized NOG-EXL mice, our models largely recapitulate key immune features of human recurrent GBM despite limitations in sample preparation (Figure S9C). Immune cell infiltration within GBM-PDX tumors was verified by immunofluorescence histological staining (Figure S10). Additionally, immune cell infiltration was detected in the spinal cords (Figure S11), suggesting that immune surveillance at this site may contribute to the suppressed spinal dissemination of tumor cells observed in humanized mice (Figure 1E–G and Figure S2C–E).

To explore how tumor size influences immune cell composition within the TME, we categorized NOG-EXL mice bearing JX14P-RT or JX39P-RT tumors into small, medium, and large groups using bioluminescence imaging (Figure S12A). This analysis demonstrated that larger tumors had significantly higher T-cell infiltration, evidenced by increased hCD3+ and hCD8+ T-cell frequencies (CD3: small vs large, p = 0.018; CD8: small vs medium, p = 0.038; small vs large, p = 0.006) (Figure S12B). These results indicate that humanized mice recapitulate the increased T-cell infiltration observed during tumor growth in syngeneic mouse models.³³

Single-cell RNA sequencing reveals diverse immune populations in the TME.

To understand the immune landscape of robust TILs identified in humanized mouse GBM-PDX tumors, scRNA-seq was performed on human CD45+ cell populations from PB and GBM-PDX tumors, using JX14P-RT in humanized NOG-EXL mice (Figures 4–6). Human CD45+ TILs were isolated from three GBM-PDX tumor tissues 49 days post xenografting, corresponding to a time point when large tumors were established, and subjected to scRNA-seq. 16,847 human CD45+ cells, consisting of TILs and peripheral blood leukocytes (PBLs), passed quality control (Figure S13A,B and Table S4). These cells were integrated to capture broad immune cell subsets and visualized using UMAP dimensionality reduction, yielding 21 distinct clusters based on a specific gene expression pattern (Figure 4A,B and Tables S5). Using PanglaoDB as an external reference of curated cell-type markers, we calculated gene-set scores and verified that the highest scores aligned with our annotations (Tables S6–8).³⁴ Overall, TILs displayed higher levels of several immune cell types compared to PBLs (Figure 4C): TILs had increased proportions of Tregs (cluster 0), CD8+ T cells (cluster 2), effector CD8+ cells (cluster 3), activated CD4+ cells (Th17-like) (cluster 6), GZMK+ CD8+ cells (cluster 9), and naive / memory T cells (cluster 10). Although not statistically significant, Tregs were more abundant in TILs than in PBLs (median 14.15% [interquartile range, 10.96–19.91%] vs. 8.34% [IQR, 6.75–8.88%]; p = 0.202) (Figure 4D and Table S9), as known in patient GBM tumors.³⁵ There was a reduction in various populations of human PBLs, including naive / memory CD4+ and CD8+ T cells (clusters 4 and 7, p = 0.00902 and p = 0.0328), B cells (cluster 11, p = 0.0163), and B cell-like (clusters 17, p = 0.0398) (Figure 4D and Table S9). The overarching trends in immune profiles were consistent among individual mice despite differences in the donor mice’s blood cell population ratios (Figure 4E,F and Figure S14A,B).

Figure 4. Single-cell RNA-seq analysis of JX14P-RT tumor in NOG-EXL humanized mouse model.

(A) UMAP plot of cell types clustered by single-cell transcriptional analysis of CD45+ cells (n = 16,847) from peripheral blood leukocytes (PBLs) and tumor-infiltrating lymphocytes (TILs) isolated from three NOG-EXL humanized mice bearing transplanted JX14P-RT human glioblastoma cells. (B) Feature plots showing expression of canonical marker genes used to define immune cell clusters. (C) UMAP plot showing the origin of CD45+ cells as PBLs or TILs. (D) Comparison of cell type percentages between PBLs and TILs. The box represents the median with quartiles at the midpoints between minimum-median and median-maximum. Whiskers span the full range of data. Two-tailed Student’s t-test. *P < 0.05, **P < 0.01. Comparisons without significance indicators are non-significant. (E) Distribution of 21 cell types in PBLs between three mice. (F) Distribution of 21 cell types in TILs between three mice.

Figure 6. Comparison of brain tumor characteristics between humanized and naive mouse models.

(A) UMAP plots of tumor cells from individual humanized mice (n = 3) and naive mice (n = 2), highlighting 18 distinct cell types. (B) UMAP plots illustrate differences between humanized and naive mouse tumors. (C) UMAP plots showing tumor cell distribution for each humanized and naive mouse. (D) Effective number of clusters per mouse (exp(H’)), derived from the Shannon diversity index (H’) calculated from the distribution of tumor cells across the 18 clusters. Values are shown as fold change relative to the mean of naive mice (mouse4 and mouse5; reference line = 1.0). (E) Distribution of tumor cell distances to group centroid in PCA space comparing humanized (n = 5,730) and naive mice (n = 5,966). Mann-Whitney U-test, ***P < 0.001. (F) Two-dimensional representation of cellular states where four regions indicate NPC, OPC, AC, and MES states. Each dot represents the exact position of an individual tumor cell, reflecting its relative meta-module score. Dot colors distinguish between tumor cells derived from humanized or naive mice. (G) Proportional distribution of tumor cell origins from humanized (orange) and naive (green) mice across astrocyte (AC), mesenchymal (MES), neural-progenitor (NPC), and oligodendrocyte-progenitor (OPC) states. Kruskal-Wallis rank sum test, ***P < 0.001. (H) Distribution of stemness scores among tumor cells within AC, MES, NPC, and OPC cellular states. Kruskal-Wallis rank sum test, ***P < 0.001. (I) Distribution of apoptosis scores across four distinct cellular states. The color intensity represents the apoptosis score of individual cells within each state.

The T cells that infiltrate GBM-PDX tumors exhibit immunosuppressive properties.

GBM-PDX tumors in humanized mice promoted the establishment of M2-type macrophages (Figure 3E) and Tregs (Figure 2E), suggesting a more immunosuppressive TME. To determine whether T cells that have infiltrated GBM-PDX tumors in humanized mice exhibit an immunosuppressive state similar to that observed in recurrent GBMs, 8,650 CD3+ (CD3D+) T cells were selected (Figure 5). The UMAP algorithm reduced dimensionality, enabling them to be sorted into 14 clusters (Figure 5A) based on a specific gene expression pattern (Tables S10–12).³⁶ The identified clusters correspond to various CD4+ and CD8+ T cell subsets, including naive/central memory (CM), resident memory (RM) cells, and Tregs. Two independent Treg populations (1 and 2) suggest recurrent GBM might harbor diverse functional Tregs in TILs, similar to lung cancer.³⁷ Additionally, GZMK+ CD8+ T cell clusters were observed, exhibiting high GZMK expression in the GBM TME. These GZMK+ CD8+ T cells are a newly identified population in the GBM TME, characterized by low classical cytotoxic markers (PRF1, GNLY, GZMB) and an exhausted phenotype (Figure 5B,C).³⁶ Their presence in GBM-PDX tumors is vital for interpreting the human GBM TME in animal models. Overall, T cell exhaustion was confirmed by four markers: CTLA4, PD1, LAG3, and TIGIT (Figure 5D). Tregs (clusters 12 and 13) showed high CTLA4 and TIGIT expression, while CD8+ T cells (clusters 7–11) displayed high PD-1 and LAG3 expression, consistent with prior research.³⁸

Figure 5. Single-cell transcriptomic characterization of tumor-infiltrating T cells and regulatory T cells in humanized GBM-PDX model.

(A) UMAP plot of tumor-infiltrating T cells (n = 8,650) isolated from JX14P-RT tumors (n = 3) in NOG-EXL humanized mice, showing 14 transcriptionally distinct subsets.
(B) Feature plots of canonical T-cell markers (CD4, CD8A, GZMK, FOXP3, CTLA4, PDCD1, NR4A2, IL7R, and KLRB1) used to define functional phenotypes.
(C) Dot plot illustrating expression patterns of T-cell–associated genes across the 14 subsets.
(D) Violin plots comparing immune checkpoint gene expression (CTLA4, PDCD1, LAG3, and TIGIT) among T-cell subsets. (E) UMAP plot of tumor-infiltrating Tregs (n = 1,528) showing five subpopulations, including activated OX40^hi/GITR^hi clusters.
(F) Feature plots displaying expression of activation and lineage-defining genes in Treg subsets.
(G) Dot plot summarizing differential expression of immunoregulatory and activation-related genes among Treg clusters.
(H) Violin plots highlighting surface checkpoint expression (CTLA4 and LAG3) and suppressive-function genes (TNFRSF18/GITR and MAGEH1) in Treg subsets.

To assess whether the GBM-PDX CD3+ T cell composition matched that of patient GBMs, we analyzed scRNA-seq data and the recurrent GBM dataset.³¹ A total of 44,560 cells were visualized via UMAP (Figure S15A), with cluster genes/markers in Figure S15B,C. The TME of recurrent GBM contains clusters of CD4+ T naive/CM, CD4+ Tregs, CD8+ GZMK+, CD8+ cytotoxic RM, and CD8+ exhausted T cells, aligning with humanized mice TME. Using pseudo-bulk Spearman correlation on shared HVGs, T-cell states from humanized mouse tumors showed overall transcriptional similarity to best-matching clusters in publicly available recurrent human GBM datasets (Table S13).

Humanized mouse models revealed significant infiltration of Tregs in GBM-PDX tumors.

Tregs are vital for the immunosuppressive TME in recurrent GBM.³⁹ We next conducted a detailed study of these cells in humanized mouse GBM-PDX tumors, identifying 1,528 Tregs among 8,650 CD3+ T cells. The UMAP algorithm classified Tregs into five distinct clusters (Figure 5E): four “activated” clusters (C1-C4) characterized by high expression of IL2RA, CCR8, MAGEH1, and LAG3, which are associated with Treg activation (Figure 5F,G;Table S14–16), and one resting cluster (C5) lacking these activation signatures.^37,40 The first subcluster (C1/Activated 1/OX40^hiGITR^hi) exhibited a gene expression profile indicative of a distinct activated state. This was characterized by elevated expression levels of tumor necrosis factor receptor superfamily (TNFRSF) 4 (also known as OX40) and TNFRSF18 (also known as GITR) (Figure 5F,G).⁴¹ The expression levels of additional genes related to Treg activation increased, with some genes being prevalent across multiple activated clusters; LAG3 (C2/Activated 2), ICOS, and TNFRSF9/4–1BB (C3/Activated 3), and IKZF2/Helios (C4/Activated 4). To assess Treg activation, we analyzed markers such as CTLA4, LAG3, TNFRSF18, and MAGEH1 (Figure 5H).^40,41 All five clusters showed high expression of CTLA4 and TNFRSF18. Activated clusters C1–3 also highly expressed LAG3 and MAGEH1, indicating a potent suppressive phenotype.⁴⁰ Activated C4 showed lower LAG3 and MAGEH1 expression than other active clusters, suggesting it might be peripheral effector Tregs (Figure 5E).⁴⁰ These results indicate that GBM-PDX tumors harbor heterogeneous Tregs with diverse phenotypes, each exhibiting distinct immunosuppressive functions.^40,41

We further assessed the above outcomes using the human recurrent GBM dataset,³¹ following the same approach as in Figure 5. A total of 2,661 Tregs were extracted and visualized using UMAP dimensionality reduction (Figure S16A,B). This dataset also contained five distinct clusters, which were slightly different from those in GBM-PDX tumors. Activated clusters displayed expression of activation genes, with OX40, GITR (Activated 1), LAG3 (Activated 2), and TNFRSF9 (Activated 3) exhibiting particularly high levels (Figure S16C,D). All five clusters exhibited high CTLA4 and TNFRSF18 expression; however, MAGEH1 was highly expressed solely in activated clusters (1–3) (Figure S16E). Notably, Treg activation in humanized mouse GBM-PDX tumors was comparable to that of recurrent human GBM (Table S17).

The humanized mouse model demonstrated infiltration of myeloid-derived suppressor cells (MDSCs).

Myeloid-derived suppressor cells (MDSCs) play a crucial role in sustaining immunosuppression in the TME by supporting the recruitment and inhibitory functions of Tregs.⁴² We then investigated MDSC populations in humanized mice, where 321 macrophages and monocytes were identified and grouped into four distinct clusters using UMAP (C1-C4, Figure S17A), based on the expression levels of CD14, FCGR3A, S100A8, S100A9, and MRC1 (CD206), which are associated with the immunosuppressive functions of myeloid cells (Figure S17B,C;Table S18).^5,43 The C1/monocytic myeloid-derived suppressor cell (M-MDSC) cluster exhibited an immunosuppressive gene expression profile, defined by high levels of CD14, S100A8, and MRC1 (Figure S17B,C).⁴¹ All four clusters exhibited high S100A9 expression, suggesting that macrophages (C2-C4) differentiated from M-MDSCs, not monocytes (Figure S17D).⁴³ Macrophages and monocytes extracted from the same human recurrent GBM dataset described above were re-analyzed,³¹ identifying 16 clusters, including two MDSC clusters (C5, C6) and two M2-like macrophage groups (C15, C16) with high S100A9 and MRC1 expression (Figure S18). We next projected 321 tumor-infiltrating macrophages and monocytes from our humanized mouse model onto the consensus immunomodulatory myeloid programs defined by Miller et al.⁴⁴ Of these, 216 cells (67.3%) expressed at least one of the four immunomodulatory programs (Figure S17E). Most cells localized to the immunosuppressive quadrant and none in the microglia-enriched region, consistent with the absence of human microglia in humanized mouse model. These data support partial overlap between myeloid immunomodulatory states in our humanized mouse model and those reported in human GBM.

The human immune microenvironment maintains the diverse characteristics of GBM-PDX tumors.

A hallmark of recurrent human GBM is its significant tumor diversity, which primarily drives treatment resistance.^32,45 However, achieving an accurate representation of tumor diversity in a standard GBM-PDX mouse model remains a significant challenge.⁷ GBM-PDX tumors from humanized mice exhibit a detailed TIL profile similar to that of recurrent human GBM. We therefore conducted a study comparing GBM-PDX tumor cells in mice, distinguishing between models with and without mouse-humanization. For scRNA-seq, JX14P-RT cells, EGFP-marked, were isolated from the brains of three humanized and two naive mice. After quality control (Figure S13C and Table S19–23), 5,730 cells from the humanized mice and 5,966 cells from the naive mice were identified, respectively, and classified into 18 clusters (Figure 6A). The UMAP plot revealed a distinct separation between tumor cells from humanized and naive mice (Figure 6B), which was reproducible across independent samples (Figure 6C). To quantify tumor cell heterogeneity, we calculated the Shannon diversity index (H’) and the effective number (exp(H’)) for each mouse. All three humanized mice showed higher values than the naïve group mean (n = 2), expressed as fold change relative to the naive mean (Figure 6D). In addition, we quantified transcriptional dispersion by measuring the Euclidean distance from each cell to its group centroid in PCA space (PC1-PC30), showing humanized mice had significantly higher distances than naive mice (Mann-Whitney U-test, p < 0.001) (Figure 6E). Tumor diversity is evident in cell cycle distribution⁴⁶; GBM-PDX tumors from humanized mice consistently showed fewer S-phase cells and more G2/M-phase cells across different samples (Figure S19). These results indicate heightened tumor variability in the humanized mouse model.

The human immune microenvironment enriches stemness-high tumor cells in human GBM PDX tumors.

Elevated stemness in GBM, especially in recurrent tumors, correlates with increased resistance to both radiotherapy and chemotherapy, as well as with immune system evasion.^32,47 Using Neftel et al.’s methods⁴⁸, we assessed GBM-PDX tumor differentiation computationally. Based on marker expression profiles from patient-derived GBM tumors (Table S24), GBM cells are grouped into four differentiation categories: neural-progenitor (NPC), oligodendrocyte-progenitor (OPC), astrocyte (AC), and mesenchymal (MES). GBM-PDX tumors showed distinct changes, including a significant increase in NPC-like cells, marked by enhanced SOX4 and CD24 expression (Figure 6F,G and Figure S20A).⁴⁸ This was confirmed in three humanized and two naive control mice (Figure S20B). Elevated markers were also validated for the four clusters (Figure S20C,D), and increased SOX4 expression was also confirmed in the JX39P-RT GBM-PDX humanized model (Figure S20E).

The presence of NPC-like cells in GBM-PDX tumors could represent a significant evolutionary adaptation to withstand immune pressure. To test this, we measured stemness and apoptosis across four tumor cell groups, using scores derived from canonical marker expression levels⁴⁹ and apoptosis markers (Table S25).⁵⁰ The NPC-like cell population predictably demonstrated the highest stemness score among all groups (Figure 6H). Regarding apoptosis, these cells scored lower than AC and MES populations, and comparable to OPC populations (Figure 6I and Figure S20F). Additionally, tumor cells from humanized mice had significantly lower apoptosis scores than those from naive mice (Mann-Whitney U-test, p < 0.001) (Figure S20G).

To confirm the stem-like nature of the NPC-like population, we examined their cell cycle. NPC-like cells showed more G1 phase enrichment than other cell states (Figure S21A). We confirmed that CD24+ tumor cells, sorted from humanized JX39P-RT tumors, proliferated slower than CD24-cells (p = 0.043) (Figure S21B,C). Furthermore, CD24+ cells exhibited enhanced radioresistance after 3 Gy radiation compared to CD24-cells (p = 0.0094) (Figure S21D,E), consistent with the stem-like characteristics of the NPC-like population. The findings indicate that NPC-like cells in humanized GBM models exhibit functional cancer stem cell traits, including slow cycling and therapy resistance.

The above data reveal NPC-like tumor cells in GBM-PDX tumors from humanized mice, whereas this phenotype is absent in immunodeficient models. This suggests that the human immune environment may influence these cell states. While extensive validation across additional models and larger cohorts is necessary, humanized GBM-PDX models hold promise for studying immune-tumor interactions and evaluating immunotherapies for recurrent GBM.

Discussion

Animal models are essential for cancer research, linking preclinical human cancer biology with the evaluation of new therapeutic strategies. Human GBM research often uses immunocompromised mice with human GBM PDX tumor cells. Advances in humanized mouse models now allow the study of the complex dynamics between human cancer cells and immune cells in the TME.^8,9 We utilized an advanced humanized mouse model using NSG-SGM3 and NOG-EXL strains implanted with radio-resistant GBM PDX tumor cells. These models feature radiation-resistant GBM PDX tumors, enabling in-depth examination of recurrent GBM within a reconstituted human immune microenvironment. Both models enhance myeloid lineage human cell reconstitution, including MDSCs,^51,52 a process inadequately supported by standard humanized NSG mice.¹³ These models provided three key insights into a recurrent GBM animal model: they mirrored its immune characteristics, showing human immune cell infiltration and transcriptional similarities to patient tumors; they recapitulated the immunosuppressive microenvironment’s temporal evolution; and they revealed significant infiltration of immunosuppressive cells (Tregs, exhausted T cells, MDSCs, M2 macrophages) exhibiting activation features similar to human recurrent GBM. Furthermore, human immune cells enhanced tumor variability and stemness in GBM PDX cells, fostering NPC-like cell development. These findings indicate that humanized GBM PDX mice effectively mimicked human recurrent GBM conditions and immune responses.

T-cell responses play a crucial role in tumor immunity, and their dysfunction is a hallmark of recurrent GBM.⁵³ Previous studies show GBM TME infiltration by Tregs and exhausted T cells causes local immunosuppression and treatment resistance.³⁹ Our humanized mouse models, developed with radiation-resistant human GBM PDX tumors, effectively mirrored key T cell responses in recurrent GBM patients. Despite human T cell infiltration into GBM PDX tumors, in-depth examination was limited by inadequate MDSC reconstitution, which is critical for creating a suppressive GBM TME and is strongly linked to T-cell exhaustion.⁴² In addition, MDSCs contribute to the recruitment and subsequent activation of Tregs in the TME through both direct and indirect mechanisms.^42,52 Using scRNA-seq, we first comprehensively characterized T-cell populations in humanized GBM PDX mouse models. Our study revealed a T-cell landscape remarkably similar to that of human recurrent GBM, identifying activated Tregs with high OX40/GITR expression and exhausted GZMK+ CD8+ T cells that matched patient tumors. We also discovered heterogeneous, activated Treg subsets with distinct molecular signatures within the TME.³⁹ The model well reflects human GBM’s complex immunoregulatory networks through diverse Treg populations and immunosuppressive markers. It further captures recurrent GBM’s T-cell dysfunction, as indicated by exhausted CD8+ T cells expressing multiple checkpoint molecules.

Recurrent GBM is characterized by therapy-resistant tumor cells arising from complex TME interactions, leading to increased stemness, altered cell cycle, and enhanced immune evasion. Our findings suggest that NPC-like cells arose in humanized GBM models and exhibited stem-like features, including slow cycling and enhanced therapy resistance. This highlights the distinct advantages of our model for studying GBM-immune system interactions, providing a strong basis to investigate immune evasion and therapy resistance. Moreover, this model is a critical resource for assessing novel therapies, particularly those that interact with the host immune systems.

Our humanized mouse model effectively recapitulates key features of human recurrent GBM, including tumor–immune integration, despite several limitations. Variability in immune cell composition was observed among mice, which may be partly attributable to differences in tumor size at the time of analysis (Figure S12). Notably, mice with comparable tumor sizes showed broadly consistent immune profiles, suggesting that at least some of this variability reflects dynamic tumor progression rather than intrinsic shortcomings of the model. This underscores the critical role of tumor size when interpreting immune responses. Nevertheless, additional analyses will be required to more rigorously define these relationships. In particular, because our single-cell RNA-seq dataset captured a limited number of myeloid cells, deeper profiling will be necessary to robustly characterize myeloid phenotypic states and their associations with tumor burden. Future studies should also explore additional modulators such as hypoxia, intratumoral hemorrhage, and reactive astrocyte accumulation.

Our current model fails to incorporate human microglia, which are essential for GBM biology,⁵⁴ and their successful reconstitution in mice is hampered by their yolk sac origin.⁵⁵ While a team led by Gorantla successfully engineered a humanized NOG mouse model that derives microglia via transgenic human interleukin-34 (hIL34),⁵⁶ integrating this into a GBM-PDX model could enhance understanding of microglia’s role in tumors. Moreover, current models use the human leukocyte antigen (HLA)-mismatched tumor and immune cells, potentially causing unintended immune responses that may not reflect natural GBM progression. Although scRNA-seq comparisons with original patient tumors are most physiologically relevant, producing patient-specific humanized mouse models is currently a challenging technical endeavor because HSPCs derived from adult peripheral blood exhibit limited engraftment and lineage reconstitution.⁵⁷ Future studies should focus on developing models that match patient HSPCs and tumor cells to address this. These matched models would better reflect GBM patients’ immune responses and provide a reliable platform for evaluating immunotherapeutic strategies.

Key Points.

• A wide array of human immune cells infiltrates GBM-PDX tumors in humanized mice.

• The immunosuppressive environment in human recurrent GBM was recapitulated in humanized mice.

• The tumor heterogeneity of GBM-PDX is more evident in humanized mouse xenografts.

Importance of the study.

Current GBM treatment development is primarily constrained by the absence of appropriate animal models that can enhance our understanding of human GBM biology and its immune dynamics. We created a unique GBM mouse model using myeloid-enhanced humanized mice with patient-derived xenografts resistant to radiation. Detailed immune cell analysis revealed various human blood cells in GBM-PDXs derived from humanized mice. Single-cell RNA sequencing uncovered crucial T-cell subgroups, notably regulatory and exhausted T cells, which matched those observed in recurrent GBM patient tumors. Moreover, the model outperformed standard xenograft approaches in tumor heterogeneity, resembling the intricate diversity observed in recurrent GBM samples. Our humanized GBM mouse model provides a valuable platform for studying interactions between human tumors and immune cells. The tool is designed to facilitate the identification of specific immunological characteristics and to accelerate the appraisal of immunotherapies for various brain tumor types.

Abbreviations:

GBM

Glioblastoma

scRNA-seq

single-cell RNA sequencing

TME

tumor microenvironment

NK

natural killer

Tregs

regulatory T cells

HSPCs

hematopoietic stem progenitor cells

DCs

dendritic cells

NSG

NOD.Cg-Prkdc^scidIl2rg^tm1Wjl/SzJ

NSG-SGM3

NOD.Cg-Prkdc^scidIl2rg^tm1WjlTg(CMV-IL3, CSF2, KITLG)1Eav/MloySzJ

NOG-EXL

NOD.Cg-Prkdc^scid Il2rg^tm1Sug Tg(SV40/HTLV-IL3, CSF2)10–7Jic/JicTac

hIL3

human interleukin-3

hGM-CSF

human granulocyte-macrophage colony-stimulating factor

hSCF

human stem cell factor

PDX

patient-derived xenograft

PB

peripheral blood

TILs

tumor-infiltrating lymphocytes

NPC

neural-progenitor

UMAP

Uniform Manifold Approximation and Projection

PBLs

peripheral blood leukocytes

TNFRSF

tumor necrosis factor receptor superfamily

MDSCs

myeloid-derived suppressor cells

M-MDSC

monocytic myeloid-derived suppressor cell

OPC

oligodendrocyte-progenitor

AC

astrocyte

MES

mesenchymal

hIL34

human interleukin-34

HLA

human leukocyte antigen

Data Availability:

Prior to manuscript publication, we will deposit our Single-cell RNA sequencing data in the Gene Expression Omnibus (GEO) repository for public access.