Glioblastoma Subtypes Exhibit Distinct Migration Mechanics and Immune Responses

Abstract

Glioblastoma (GBM) remains a deadly cancer driven in part by invasion of tumor cells into the brain. Transcriptomic analyses have identified distinct molecular subtypes, but mechanistic differences that account for clinical differences are not clear. In this study, we show that, as predicted by the motor-clutch model of cell migration, mesenchymal glioma cells are more spread, generate larger traction forces, and migrate faster in brain tissue compared with proneural cells. Despite their rapid migration and comparable proliferation rates in vitro, mice with mesenchymal tumors survive longer than those with proneural tumors. This improved survival correlated with an immune response in mesenchymal tumors, including T cell–mediated. Consistently, inducing mesenchymal tumors in immunodeficient mice resulted in shorter survival, supporting a protective immune role in mesenchymal tumors. Thus, mesenchymal tumors have aggressive migration but are immunologically “hot,” which suppresses net proliferation. These two features counteract each other and may explain the lack of a strong survival difference between subtypes clinically, while also opening up new opportunities for subtype-specific therapies.

Significance:

This study highlights new mechanical and immunologic insights into GBM molecular subtypes using an integrated modeling–genome engineering strategy, which can potentially facilitate GBM subtype-specific therapeutic strategies.

Introduction

Glioblastoma (GBM; World Health Organization grade IV primary brain tumor) progression can be characterized in terms of tumor growth and spreading, two key parameters which are influenced by many of the hallmarks of cancer (1). In GBM, tumor spreading is driven by tumor cells’ ability to infiltrate healthy brain parenchyma, which prevents complete surgical resection and results in tumor recurrence (2–4). Molecular and genetic analyses of human GBM have identified at least three distinct molecular subtypes: proneural, classic, and mesenchymal (5–7). These subtypes were shown to strongly correlate with specific genetic alterations (mesenchymal: NF1 loss; classic: EGFRvIII; and proneural: PDGFRA) and cellular developmental states (6–9). Despite accumulating evidence of distinct transcriptomic and genetic signatures, the characteristic mechanistic differences between such signatures, if any, have not been identified. As a result, it remains unclear how knowledge of the different subtypes should inform clinical decisions.

One intriguing correlate of subtype is the level of CD44 expression, a cell surface protein expressed on tumor and immune cells, which is known to play a role in cancer progression across a variety of cancers, including GBM (7, 9–17). In GBM, we previously showed that CD44 expression is a prognostic marker with a biphasic dependence: better outcomes are observed at both lower and higher levels of CD44, whereas poorer outcomes are observed at intermediate levels, an example of optimality and the “goldilocks” phenomenon (17). In animal models, CD44 expression has further been shown to correlate with glioma cell migration in a biphasic relationship with a peak migration rate at intermediate expression level, which also correlated with the minimum in survival in both the animal model and human GBM (17). In addition, CD44 transcript levels are shown to vary across GBM molecular subtypes with elevated expression in mesenchymal tumors (5, 6, 15). CD44 expression in the mesenchymal tumors is, on average, closer to the CD44 level that corresponds to the minimum in patient survival than the proneural subtype (17). As an adhesion molecule, CD44 engages the extracellular matrix with the actin cytoskeleton through adapter proteins to mediate cell migration (11). This suggests that mesenchymal cells have a near-optimal level of CD44 adhesion molecules to serve as molecular “clutches” that resist myosin II motor forces, allowing them to migrate faster than proneural cells which on average have a lower, suboptimal level of CD44 clutches [see Figs. 2C and 4E in Klank and colleagues (17)]. This could then explain the slightly worse outcomes for mesenchymal patients and higher cell migration and invasion (7, 18). In addition, it would predict that mesenchymal cells would have lager cellular spread area, be more polarized, and generate more traction force as they migrate. More generally, lower CD44 is indicative of an epithelial state, and higher CD44 indicative of a mesenchymal state (19, 20), and so an increase in myosin motors and adhesions, either integrin- or CD44-mediated, may be driving the epithelial-to-mesenchymal transition (EMT) in a variety of cancers, such as breast cancer (21).

Figure 2.

Simulations of cell migration as a function of subtype based on CD44-mediated cellular adhesive clutches. A, Schematic of proneural (low CD44) and mesenchymal (medium CD44, i.e., optimal) subtypes. B, Wind-rose plots of simulated cell trajectories on different stiffnesses (κ_sub = 0.1, 1, 10 pN/nm). C, Simulated cell migration random motility coefficients show higher migration with increased (optimal) adhesion across a range of substrate stiffnesses. D, Cell summed traction force exhibits higher traction forces with increased (optimal) adhesion across all stiffnesses. E and F, Cell spread area and aspect ratio increase with increased adhesion on substrate stiffnesses of 0.1 and 1 pN/nm. Error bars represent the SEM. +, P < 0.05; *, P < 0.01; **, P < 0.001; ***, P < 0.0001; ****, P < 0.00001, Statistical significance was assessed using the Kruskal–Wallis rank test with Dunn–Sidak correction.

Figure 4.

Cancer cell migration is subtype specific and independent of the tumor microenvironment and species. A, Six representative images of primary isolated cells from each of the two mouse tumor subtypes (3 NRAS and 3 PDGF). GFP-tagged tumor cells (green) and blood vessels staining using isolectin B4 (magenta). B, NRAS/mesenchymal primary isolated tumor cancer cells migrate faster than PDGF/proneural primary isolated cancer cells in normal mouse brain tissue as measured by random motility coefficient, see Supplementary Video S2. C and D, Quantification of cell spread area and cell aspect ratio showing that NRAS/mesenchymal cancer cells are more spread (C) and tend to be somewhat more polarized (D), P = 0.075, than PDGF/proneural cancer cells in healthy brain slices. E, Six representative images of proneural and mesenchymal PDX lines cells in healthy mouse brain tissue. Labeled tumor cells using DiO membrane dye (green) and blood vessel staining using isolectin B4 (magenta). F, Mesenchymal PDX cells migrate faster than proneural PDX cells in normal mouse brain tissue as measured by random motility coefficient, see Supplementary Video S3. G and H, Quantification of cell spread area and cell aspect ratio showing that mesenchymal PDX cells are more spread (G) and more polarized (H) than proneural PDX cells in healthy brain slices. Solid and dashed lines represent mean and median values, respectively. Error bars represent the SEM. +, P < 0.05; *, P < 0.01; **, P < 0.001; ***, P < 0.0001; ****, P < 0.00001. Statistical significance was assessed using the Kruskal–Wallis rank test.

Based on these previous results, we tested the hypothesis that a key mechanistic difference between GBM molecular subtypes is that proneural cells are slow migrating and mesenchymal cells are fast migrating. To address this question, we generated animal models recapitulating the transcriptomic signatures of human mesenchymal and proneural GBM in an immunocompetent background using perturbations of known GBM oncogenic pathways. Specifically, mesenchymal and proneural-like tumors were driven by SV40-large T (LgT) antigen, to mimic common inhibition of p53 and Rb signaling found in GBM (22, 23), in combination with either NRAS^G12V(NRAS) or PDGFB (PDGF), respectively, which resulted in mesenchymal or proneural transcriptomic features with only a single genetic change required to switch subtypes in a wild-type (WT) mouse background. As predicted, CD44 expression was higher in NRAS-driven tumors and, consistent with our simulation predictions, ex vivo brain slice live imaging showed that NRAS tumor cells migrate faster than PDGF tumor cells and exhibit greater spreading, polarization, and force generation as well. Despite increased migration, the NRAS cohort had better survival than PDGF, which was attributed to enhanced antitumoral immune response in NRAS tumors, consistent with increased immune cell infiltration in human mesenchymal GBM (7, 24–26). Consistently, NRAS-driven tumors had shorter survival in immunodeficient (Rag1^−/−) mice compared with immunocompetent mice, supporting a role for the immune system in limiting tumor progression in mesenchymal tumors. Overall our work identified a clinically actionable difference in migration mechanics between GBM subtypes and establishes an integrated biophysical modeling and experimental approach to mechanically parameterize and simulate distinct molecular subtypes in preclinical models of cancer.

Materials and Methods

Generation of mouse tumor models

All animal studies were conducted under an approved Institutional Animal Care and Use Committee (IACUC) protocol at the University of Minnesota. All animals were housed in a daily monitored animal facility. Both immunocompetent FVB/NJ WT (RRID: IMSR_JAX:001800) and immunodeficient FVB/N mice homozygous for the Rag1^tm1/Mom mutation (Rag1^−/−; RRID: IMSR_JAX:002216) strains of mice were used in this study. Malignant gliomas were induced in neonatal mice by DNA plasmid injection into the right lateral ventricle as described previously (27, 28). Briefly, neonatal mice were injected with 1 μg of plasmid DNA mixed with polyethyleneimine (jetPEI, Polyplus) and 5% dextrose in a total volume of 2 μL at a rate of 0.7 μL/minute. The following four plasmids were used in a 1:1:1:1 ratio: empty vector, pT2/C-Luc/PGK-SB100, pT/CMV-LgTAg-IRES-GFP, and pT2/Cag-NrasV12 or pT2/Cag-mPDGF. Tumor initiation and detection require stable integration of all plasmids, ensuring that only cells with successful delivery and expression of full construct set give rise to detectable tumors. Animals were monitored daily for morbidity by bioluminescent (BL) imaging.

Immunohistochemistry of mouse tumor sections

Formalin-fixed and paraffin-embedded (FFPE) mouse brain tissues were used to prepare 4-μm-thick slides. FFPE tissue slides were stained with hematoxylin and eosin or immunohistochemistry (IHC) using standard methods. Supplementary Table S1 contains a list of antibodies and reagents used for antigen retrieval, blocking, and detection.

Quantification of IHC staining of mouse tumor sections

IHC data were quantified by counting the number of pixels in an image that were positively 3,3′-Diaminobenzidine (DAB)-stained. To avoid user bias and subjective counting, k-means clustering was used to identify pixels representing areas of positive DAB, hematoxylin staining, and background. In this implementation, every pixel in an analyzed image is assigned to one of four clusters, each cluster representing a different component in the image: positive DAB-stained areas (brown), positive hematoxylin-counterstained areas (blue), unstained tissue (light blue), and background glass slide (beige). Digital images of equal sizes (2,000 × 2,000 pixels) of DAB-stained and hematoxylin-counterstained tumor samples were converted from Red-Green-Blue to the Hue-Saturation-Value (HSV) color model. Three fields were sampled from representative areas from each tumor.

Using a custom written MATLAB (RRID: SCR_001622) algorithm, user input is used to define areas representing the four components (positive DAB stain, hematoxylin counterstain, unstained tissue, and background glass slide). These points are used as the initial estimates for the centroid locations of each of the four clusters. The squared Euclidean distance between each pixel’s HSV coordinates and the HSV coordinates of each cluster’s centroid is computed. Each pixel is then assigned to the cluster with the minimum squared Euclidean distance to the cluster centroid. Cluster centroids are recalculated as the mean of the HSV coordinated of all current members. This process is iterated until the centroid of each cluster is stable. The number of pixels in the positive DAB-stained cluster is used to quantify percent positive pixels.

Transcriptional profiling of mouse tumors

Mice were euthanized in a CO₂ chamber and perfused transcardially with isotonic saline. Mouse brains were extracted, and green fluorescent protein (GFP) goggle (#FHS/EF-2G2; BLS-ltd) was used to dissect GFP-positive tumor tissues from NRAS (N = 3) and PDGF (N = 3) mouse brains. Two matched normal brain tissues were collected from brain regions away from the tumor, and an additional normal brain tissue sample was collected from a health FVB adult mouse (N = 3; normal brain tissue; NBT). All samples were immediately placed in RNALater solution (cat. #R0901; MilliporeSigma) for 24 hours and then flash-frozen in liquid nitrogen and stored at −80°C for downstream processing. RNA extraction and sequencing were performed at the University of Minnesota Genomics Center. RNA was extracted using RNAeasy Plus Universal Mini Kit (cat. #73404; Qiagen), and libraries were prepared using TruSeq stranded mRNA (cat. #20020594; Illumina).

Next-generation sequencing was performed on the prepared RNA libraries using an Illumina HiSeq 2500 device in high output mode and generated 51 bp reads with an approximate depth of 20 million paired reads per sample. Mapping and expression calculations were generated using the rnaseq-pipeline of Gopher-pipelines (https://bitbucket.org/jgarbe/gopher-pipelines), which executed TopHat2 (29) and Cuffnorm (30) using the UCSC mm10 version of the mouse reference genome. Fragments Per Kilobase of transcript per Million mapped reads (FPKM) RNA expression of mouse tumors and healthy brain tissues listed in Supplementary Table S2. Fastq files and the Cuffnorm output were deposited at Gene Expression Omnibus (GSE161154).

Human GBM transcriptomic data

UNC RNASeqV2 level 3 expression, normalized RNA-Seq by Expectation-Maximization (RSEM), profiles of 171 samples [The Cancer Genome Atlas (TCGA)-GBM)] were retrieved from Broad GDAC Firebrowse (31). IDH status and subtype information were added to each sample based on the Wang and colleagues classification (7). For downstream analysis, 147 IDH-WT samples were used (57 classic; 52 mesenchymal; and 38 proneural).

Clustering analysis of mouse and human expression profiles

To analyze the transcriptional profiles of mouse and human datasets, a value of 0.1 was added to all FPKM and RSEM values to minimize the impact of inaccurate low values (32). The expression data were log-transformed and mean-centered. Transcripts with a standard deviation greater than 1 were clustered using average linkage hierarchical clustering in MATLAB (RRID: SCR_001622), with Pearson correlation used as the similarity metric. A custom MATLAB script was used to systematically identify transcriptional clusters within each dataset. For the mouse dataset, clusters were defined by a correlation threshold >0.5 and a minimum of 100 transcripts. For the human dataset, a correlation threshold 0.2 and >100 transcripts were used. Fisher exact test was applied to compare cluster memberships. All genes identified within each cluster are listed in Supplementary Table S3.

To quantify the relative expression of gene clusters and subtype-specific gene signatures, we calculated the average expression of each gene set per sample. This was done by taking the simple average of mean-centered, log₂-transformed RNA expression values across all genes in the set. These per-sample averages were then plotted and used to calculate the mean relative expression for each mouse cohort and human GBM molecular subtype. Subtype gene signatures used for this analysis are listed in Supplementary Table S4 (7). Supplementary Table S4 also includes the subset of genes which were selected from the subtype gene signatures list based on our clustering criteria (standard deviation >1).

Generation of mouse primary tumor lines

For each line, a tumor-bearing mouse was euthanized and transcardially perfused with isotonic saline. Tumor tissue was collected and minced with a scalpel. Minced tumor fragments were incubated with PBS for 15 minutes at 37°C. Tumor fragments were further dissociated by mixing them up and down using a 1,000 μL micropipette. Finally, tumor suspension was passed through a 40 μm sterile cell strainer (Thermo Fischer Scientific), and filtrate was spun down and plated on Matrigel-coated (cat. #354230; Corning) T-75 flasks (Corning) using NSC media which consisted of DMEM/F12 (Gibco cat. #11320033; Thermo Fischer Scientific) with 1× B-27 supplement (Gibco cat. #12587010, Thermo Fischer Scientific) and 1× penicillin/streptomycin (cat. #30-002-CI Corning). Twenty ng/mL EGF (cat. #100-15; PeproTech) and FGF (cat. #100-18B; PeproTech) were added to the cell culture media every 2 to 3 days. Cells were cultured in a 37°C 5% CO₂ incubator. Once a confluent layer was achieved, cells were detached using 0.25% Trypsin–EthyleneDiamineTetraacetic Acid (EDTA; cat. #25-053-CI; Corning) and frozen down for later use.

Once tumor lines were established, cells were grown as neurospheres using NSC media and ultra-low attachment six-well plates (cat. #3471; Corning). Neurospheres were dissociated using accutase (cat. #AT104; Innovative Cell Technologies). In total, six different mouse primary tumor lines were established: three NRAS and three PDGF. As cells were primary, low-passage, mouse tumor–derived lines, rather than highly passaged established cell lines, cell authentication and Mycoplasma testing were not conducted on the cell lines.

Patient-derived xenograft cell line culture

The patient-derived xenograft (PDX) cell lines were taken from the Mayo Clinic GBM PDX collection (managed by J.N. Sarkaria, Mayo Clinic). The collection was established under an approved IACUC protocol and in accordance with the Belmont Report and US Common Rule with approval from the Mayo Clinic Institutional Review Board, and written informed consent was obtained for the patient studies, as previously described (33).

Three mesenchymal PDX lines (GBM 16, 39, and 44) and three proneural PDX lines (GBM 64, 80, and 85) were selected to study their migration in organotypic mouse brain slice. Cells were cultured on Matrigel-coated (354230; Corning) tissue culture flasks in a 37°C 5% CO₂ incubator, Supplementary Table S5. NSC media were used to culture PDX cell lines, and 20 ng/mL EGF and FGF were added to the cell culture media every 2 to 3 days. Cell line authenticity was assessed using short tandem repeat profiling (Genecopoeia), and Mycoplasma testing was performed routinely.

Ex vivo confocal imaging of tumor-bearing brain slices

Tumor-bearing mice were sacrificed when BL signals were around 5 × 10⁷ radiance (p/second/cm²/sr). Mice were euthanized in a CO₂ chamber and perfused transcardially with isotonic saline. Mouse brains were extracted and kept in chilled artificial cerebrospinal fluid (124 mmol/L NaCl, 2.5 mmol/L KCl, 2 mmol/L MgSO₄, 1.25 mmol/L KH₂PO₄, 26 mmol/L NaHCO₃, and 10 mmol/L glucose). Coronal brain slices of thickness 300 μm were prepared using a vibratome (Leica Biosystems). Only one slice was used for live-cell imaging. Isolectin GS-IB4 (Alexa Fluor 568 conjugate cat. #I21412; Molecular Probes) was used to label the vasculature.

Before imaging, the slice was washed and transferred into a No. 0 glass bottom 35 mm culture dish (cat. #P35G-0-20-C; MatTek) containing serum-free DMEM (Gibco, Thermo Fischer Scientific). A tissue culture anchor (cat. #SHD 42-15; Warner Instruments) was placed on top of the slice to prevent movement during imaging. The slice was then imaged on a Zeiss LSM 7 Live swept-field laser confocal microscope (Zeiss) at 15-minute intervals for up to 20 hours in humidified 5% CO₂ air at 37°C. Images were collected with a 20× objective lens (Plan-ApoChromat 20X, 0.8 NA, Zeiss). The number of Z stacks of several regions of interest was adjusted to ensure that the data acquisition of one frame in the time series was completed under 15 minutes (10–20 planes with 10 μm z-step was typically used). Maximum intensity projections from multiple Z stacks were used to generate 2D images for quantitative morphologic and trajectory analyses. Images were registered by an affine transformation using ImageJ (ImageJ, RRID: SCR_003070) StackReg plug-in (École Polytechnique Fédérale De Lausanne) to account for stage drift and tissue relaxation during time-lapse imaging.

Live-cell imaging of tumor cells in organotypic brain-slice culture

Healthy mouse brain slices were prepared using the same method as tumor-bearing brain slices above. For experiments using GFP-positive mouse primary tumor lines, neurospheres were dissociated using accutase (Innovative Cell Technologies). The protocol of grafting cancer cells into the brain slice was described in details in our previous publication (34). Briefly, after creating a single cell suspension, 300,000 cells in 3 mL of media were plated onto the brain slice. The cells were cocultured with the brain slice for 4 hours at 37°C and 5% CO₂ before imaging to promote cell infiltration into the brain slice. Phenol-free NSC media +2% FBS (Gibco, Thermo Fischer Scientific) was used. The slices (Isolectin GS-IB4 stained) were washed several times using cell culture media and transferred into a No. 0 glass bottom six-well plate (cat. #P06G-0-20-F; MatTek). The slices were imaged on a confocal microscope with a 10× objective lens (Plan-ApoChromat 10X, 0.45 NA, Zeiss). Similar imaging protocol as mentioned above was applied. Each primary mouse line was imaged in one slice and was repeated at least twice except for mGBM3102 where only one successful slice was imaged and analyzed.

For PDX cells, 500,000 to 800,000 cells were stained using DiO membrane dye (cat. #V22886; Thermo Fisher Scientific) for 5 minutes and then washed twice before plating onto the brain slice inside a 35 mm tissue culture dish. The grafting of the cells to the brain slice was similar to the mouse primary cells described above. The slices were imaged on a confocal microscope using a 20× objective lens (Plan-ApoChromat 20X, 0.8 NA, Zeiss). Similar imaging protocol was applied. Experiment repeated 2 to 3 times on new set of slices each time.

For both mouse primary tumor cells and PDX cells, we also acquired the maximum intensity projections from multiple Z stacks and performed image registration for further analysis for cell migration and morphology. Images were registered by an affine transformation using ImageJ StackReg plug-in (École Polytechnique Fédérale De Lausanne) to account for stage drift and tissue relaxation during time-lapse imaging.

Single-cell tracking and morphology analysis

Single-cell migration was tracked as previously described using a custom-written image segmentation algorithm in MATLAB (17, 35). Using cell centroid coordinates, the mean squared displacement (MSD) of the cell trajectories over time was calculated using the time interval overlap method (36). To quantify the dispersion of cells, the MSD over time was used to calculate the random motility coefficient μ according to the equation [MSD (t) = 4 μt; assuming 2-D geometry]. Using segmented cell regions, cell area and cell aspect ratio, defined as the ratio between the major and minor axis lengths of a fitted ellipse, were measured for each individual tracked cell. Average mean square displacement showed linear relation with time, and random motility coefficients were calculated from each individual trajectory. Distributions of random motility coefficients, cell area, and cell aspect ratio for the different conditions were compared using the Kruskal–Wallis test, which is a nonparametric rank-based test.

BL imaging and analysis

Animals were monitored for tumor development and progression using noninvasive BL imaging (BLI). Oncogene-injected animals were injected intraperitoneally with 100 μL of 28.5 mg/mL luciferin (cat. #LUCK-1G; GoldBio) prior to imaging. Mice were then anesthetized using 3% isofluorane and imaged on an IVIS50 or IVIS100 instrument (Xenogen). Images were acquired 10 minutes after injection with 5 minutes exposure time (Xenogen LivingImage Software). To avoid saturation, exposure time was reduced appropriately in fully grown tumors and accounted for in the analysis. BL images were processed using a custom written MATLAB (MATLAB, RRID: SCR_001622) algorithm where background signal was subtracted and pixels away from the tumor were set to zero. BL signal from each animal was then normalized to the initial time point when tumor was first detected.

Quantification of proliferation of mouse primary tumor line

To measure proliferation rate, 200,000 cells from each line were plated into an ultra-low adhesion six-well plate and grown as neurospheres. Growth factors were added every 2 to 3 days. At day six, neurospheres were dissociated and cells counted. After counting, the remaining cells were replated and resumed growing as neurospheres. Cells were also counted and replated at days 9 and 13. The experiment was repeated three times using each of five mouse primary tumor lines (three NRAS and two PDGF tumor lines). For each replicate, the average cell count for each cohort was calculated using the cell count from the different corresponding tumor lines.

Traction force measurements

Traction force measurements of mouse primary tumor lines were performed using traction force microscopy (TFM) on polyacrylamide gels embedded with 0.2 μm crimson fluorescent beads (cat. #F8806; Thermo Fisher Scientific) and coated with type-I collagen (cat. #354236; Corning). Collagen-coated polyacrylamide gels of varying Young’s modulus were prepared as previously described (35, 37). Briefly, 0.7, 4.6, and 9.3 kPa polyacrylamide polymer mixture with fluorescent beads were cast onto a No. 0 glass bottom dish then coated with type-I collagen using Sulfo-SANPAH (cat. #A35347; Thermo Fisher Scientific). Mouse primary tumor cells were dissociated from neurospheres and plated on prepared gels at low density (1–5 cells/mm²) using NSC media +2% FBS.

To measure force transmission, TFM was performed as previously described (35). Briefly, Nikon TiE and Ti2 epifluorescence microscopes were used to image fluorescent bead positions before and after cell detachment via trypsin. A Zyla 5.5 sCMOS camera (Andor Technology) and a 40x/0.95NA Ph2 lens with 1.5× intermediate zoom (60× total magnification, 110 nm spatial sampling) was used. Cells were maintained at 37°C and 5% CO₂ for the duration of imaging using an Oko lab Bold Line top stage humidified incubator (Okolab). At each stage position, a phase-contrast image of the cell was acquired. Next, an image of fluorescent beads at the top surface of the gel was captured using a 575/25 nm LED and eGFP/mCherry filter set with LED fluorescence illumination from a SpectraX Light Engine (Lumencor). Media in dishes were carefully removed, cells were detached with 0.25% trypsin/EDTA (Corning), and fluorescence images of beads in the absence of cells were acquired at saved stage positions.

Using a previously described method (35), the displacement field was determined using particle image velocimetry (PIV) using the before and after bead images. A window size of 80-pixels (8.8 μm) square was used in PIV, and a final lattice spacing of 20 pixels (2.2 μm) was achieved. Stress and displacement vectors were obtained by solving the inverse Boussinesq problem in Fourier space (38). By integrating the product of the stress and displacement vectors over the entire image, substrate strain energy was determined as previously described (35).

Vascular deformation analysis

To visualize vessel deformation during tumor cell migration, a montage of representative time frames was generated from the vessel channel. In addition, a color-coded maximum-intensity projection over time was created, in which the color of each pixel corresponds to the time point at which the maximum intensity occurred. In these projections, regions of the vessel that remained in the same position over the imaging period (i.e., were stationary) appeared gray, whereas regions that shifted position over time appeared colored. Deformation was therefore defined as time-dependent displacement or bending of the vessel structure adjacent to the migrating cell. No additional quantitative metric was applied; the images are presented as a qualitative representation of vessel deformation.

Stochastic cell migration simulator

The previously described (17, 35) cell migration simulator (CMS v1.0) was used to simulate cells migration dynamics in response to changes in cell adhesion. The parameters used in the simulations are presented in Supplementary Table S6. The number of adhesive clutches (N_c) was adjusted to model the change in adhesion observed between PDGF/proneural and NRAS/mesenchymal tumors. N_c of 2,500 and 7,500 clutches were used to simulate PDGF/proneural and NRAS/mesenchymal tumor cells, respectively. Four hours of cell dynamics were simulated and the first hour was excluded from analysis to allow the system to reach steady state. Analysis was performed using a 10-minute sampling interval as previously described (35).

Brownian dynamics tumor simulator

The Brownian dynamics tumor simulator (BDTS) was used as previously described with modifications (39, 40). In the present study, we extended the BTDS to 3-dimentional tumors and incorporated immune cells’ dynamics. Briefly, simulations started with 27 cancer cells, modeled as rigid sphere with radius (r_cancer), placed in a 3 × 3 × 3 grid where the distance between each cancer cell (center-center) is 3 × r_cancer. In simulation, including immune response, eight T cells, also modeled as rigid sphere with radius (r_CTL), were included, and each was placed 1.5 × r_cancer away from a randomly selected cancer cell. At each simulation time step of 1 minute, cancer and T cells are allowed to move randomly and grow as spheres with a linear volumetric growth rate. Movement and growth are rejected if the newly assigned space is already occupied by a like cell (i.e., no-overlap enforced between cancer cells and between T cells). However, a cancer cell and T cell contact occurs when a proposed cell movement put the distance between cell centers less than or equal (r_CTL + r_cancer). The duration of the contact is (1/k_dissoc), in this case 10 minutes. For every contact, both cancer cell and T cell take a “hit” that reduces their hit points (HP) by one. Both cancer cell and T cell have limited HPs, and once HP is depleted (equals 0), the cell dies or become exhausted. For NRAS simulations, T cells were added to the tumor simulator and T cell–mediated killing was simulated. For PDGF simulation, only cancer cells were simulated. Cancer cell motility was estimated from ex vivo brain slice imaging of tumor cells, and proliferation rate was estimated from the in vitro proliferation of mouse primary tumor lines. The rest of parameters were estimated based on previous published work or used as an adjustable parameter (see Supplementary Table S7).

Statistical analysis

Fisher exact test was used to compare the mouse and human transcriptomic clusters. One-way analysis of variance (ANOVA) was used to compare transcript levels. Analysis of covariance was used to compare between the two regression lines fit to the bioluminescent intensity time course data. Rank test, Kruskal–Wallis, one-way ANOVA was used to compare single-cell behaviors and IHC quantifications. Where appropriate, a subsequent Dunn–Sidak test for multiple comparisons was performed.

Results

Genetically induced high-grade glioma mouse models recapitulate the transcriptomic signatures of mesenchymal and proneural GBMs

To characterize the mechanics of GBM subtypes, we utilized the Sleeping Beauty (SB) transposon-based gene transfer system to induce high-grade gliomas in immunocompetent FVB/NJ strain mice (17, 27, 28, 41, 42). Plasmid constructs carrying SB transposons with oncogenic driver transgenes (SV40-LgTA + NRAS^G12Vor SV40-LgTA + PDGFB; here termed NRAS and PDGF, respectively) were used to model mesenchymal and proneural GBM tumors, respectively (Fig. 1A). Additional plasmids encoding firefly luciferase and GFP, as well as the SB transposase, were coinjected to enable stable genomic integration, confirm successful gene transfer, and allow monitoring of tumor development and growth via BLI) and single-cell tracking by fluorescence microscopy. Notably, tumor initiation and detection require stable integration of all plasmids, ensuring that only cells with successful delivery and expression of full construct set give rise to detectable tumors. Similar to human GBM, histologic sections from these tumors exhibited highly mitotic tumor cells, necrosis, anaplasia, and perivascular infiltration and proliferation (Fig. 1B).

Figure 1.

De novo induced GBM mouse models using immunocompetent mice recapitulate mesenchymal and proneural subtypes of human GBM. A, Schematic of mouse models. Plasmids encoding oncogenic drivers NRAS^G12V or PDGFβ in combination with SV40LgTA were injected into P1 FVB mice to induce mesenchymal and proneural high-grade gliomas, respectively. B, Hematoxylin and eosin (H&E) and Ki67 IHC staining of NRAS and PDGF tumor sections. Scale bars, 50 μm. C, Unsupervised hierarchical clustering of mRNA expression in induced mouse tumors and human IDH-WT GBM-TCGA. Arrows indicate conserved genes present in both mouse and human gene clusters as defined by systematic comparison of gene cluster membership between datasets. D, Quantification of relative expression of mouse gene clusters within the mouse dataset (top) and human GBM molecular subtypes (bottom). E and F, Relative expression of key mesenchymal and proneural genes within mouse tumors (E) and human tumors (F). Solid and dashed lines represent mean and median values, respectively. Error bar represents the SEM. +, P < 0.05; *, P < 0.01; **, P < 0.001; ***, P < 0.0001; ****, P < 0.00001, Statistical significance was assessed using one-way ANOVA. [A, Created in BioRender. Li, R. (2026) https://BioRender.com/49xmcik.]

To assess whether the NRAS and PDGF tumors recapitulated the mesenchymal and proneural subtypes, respectively, we performed cross-species transcriptomic analysis using bulk RNA sequencing (RNA-seq) data from mouse and human tumors. Bulk RNA-seq was performed on tumor tissues from both cohorts and on NBTs (NRAS N = 4, PDGF N = 4, and NBT N = 3). IDH-WT human GBM transcriptomic profiles were retrieved from Broad GDAC Firebrowse (31). Unsupervised hierarchical clustering of the mouse dataset (Supplementary Table S2) revealed clear differences between normal tissue and tumor tissue and between NRAS and PDGF tumors (Supplementary Fig. S1A). Not surprisingly, gene ontology enrichment analysis, using EnrichR (43), showed an enrichment of cell cycle–related processes in tumor tissue–specific gene cluster (817 genes; Supplementary Fig. S1B) and neuronal processes in normal tissue cluster (1722 genes; Supplementary Fig. S1D). Interestingly, the NRAS-specific cluster (1,327 genes) was enriched with cytokine-mediated signaling and inflammatory response processes (Supplementary Fig. S1C).

To determine whether any variations observed between the two mouse cohorts were also present in human tumors, unsupervised hierarchical clustering was performed on both mouse and human tumor datasets and clusters were independently identified in both datasets (Supplementary Table S3). Three and 10 gene clusters were identified in both mouse and human datasets, respectively, as shown in Fig. 1C. We identified mouse cluster MC1 (n = 1,534 genes) as being significantly enriched in genes found in human cluster HC1 (n = 1,186 genes, P < 1 × 10⁻¹⁵), MC2 (n = 414 genes) is significantly enriched in genes found in HC2 (n = 1,098 genes, P < 1 × 10⁻¹⁵), and MC3 (n = 232 genes) is significantly enriched in genes enriched in HC4 (n = 432, P < 1 × 10⁻¹⁵). These results show that conserved transcriptomic patterns distinguish subtypes of both mouse and human tumors.

To assess whether the transcriptional patterns present in the mouse tumor models represent previously described GBM subtypes, we compared the expression of identified mouse gene clusters within human GBM subtypes. We found that MC1, which is enriched in NRAS tumors, is significantly enriched in human mesenchymal GBM relative to proneural and classic GBMs (Fig. 1D). In contrast, MC3, which is enriched in PDGF tumors, is significantly enriched in proneural GBM relative to mesenchymal and classic GBMs (Fig. 1D). Furthermore, we found the expression of known mesenchymal and proneural genes are relatively elevated in NRAS and PDGF tumors, respectively (Fig. 1E and F). Additionally, using mesenchymal and proneural gene signatures, NRAS and PDGF tumors showed relatively elevated expression of mesenchymal and proneural gene signatures, respectively, similar to human mesenchymal and proneural tumors (Supplementary Fig. S2A–S2D). These results demonstrate that NRAS and PDGF tumors transcriptionally resemble mesenchymal and proneural GBMs, respectively, consistent with established subtype markers and gene signatures defined (5–7).

Motor-clutch modeling of cell migration predicts NRAS/mesenchymal tumor cells will migrate faster, have larger cell spread area, and generate more force than PDGF/proneural tumor cells

To examine tumor cell migration, we used our cell migration simulator (17, 35) to predict migration phenotypes in response to gene expression changes. The cell migration simulator is based on the motor-clutch model which incorporates actin-based protrusion dynamics, mass conservation, and force balances to reproduce cell polarization and random motility in 1D and 2D compliant microenvironments (17, 34, 35, 44–48). The number of adhesion/clutches and motors are key determinants of cell migration, with a relative balance being essential for efficient migration (35, 45, 49). Using a representative set of 54 cell migration genes previously identified to be expressed in the human U251 GBM cell line (35), NRAS tumors showed significantly increased expression of adhesion and adapter genes (Supplementary Fig. S3A). A similar pattern was observed in MES relative to PN tumors in the TCGA-GBM dataset (Supplementary Fig. S3B). Both NRAS and MES tumors upregulated CD44 and its cognate adhesion adapter gene moesin (MSN), which mechanically links the CD44 cytoplasmic tail to F-actin (11, 50–55). Notably, the levels of myosin motor genes were not differentially expressed in the mouse dataset, whereas, in the human dataset, MYH9 and MYO1C were modestly upregulated in MES tumors but to a lesser degree than adhesion molecules (Supplementary Fig. S3B). These results suggest that NRAS/mesenchymal tumor cells have a higher number of adhesion/clutches than PDGF/proneural tumor cells and little to no change in the number of motors.

Based on these results, we simulated the effect of CD44 expression level on cell migration by adjusting the number of adhesion bonds (number of clutches, N_c) in the model (17, 35). We used low clutches relative to motors (low adhesion) to simulate PDGF/proneural cells and a medium level of clutches that balanced the number of motors (optimal adhesion) to simulate NRAS/mesenchymal cells (Fig. 2A).

Our simulations show that lowering the number of adhesions, representing the PDGF/proneural case, results in reduced cell migration, force transmission, cell spread area, and cell polarization due to an insufficient number of clutches relative to the number of motors, as shown in Fig. 2B–F. In contrast, when clutch and motor numbers are balanced, representing the NRAS/mesenchymal case, simulated cells recover their ability to migrate, transmit forces, spread, and polarize across a range of substrate stiffnesses. Consequently, simulation results predict that NRAS/mesenchymal tumor cells will migrate faster than PDGF/proneural tumor cells due to increase of adhesion (i.e., CD44 expression) rather than small differences in motor gene expression (Supplementary Fig. S3A and S3B). In addition, with a higher number of clutches and a balanced motor-clutch ratio, NRAS/mesenchymal tumor cells are predicted to generate higher force, spread more, and display greater polarization.

NRAS/mesenchymal tumor cells migrate faster than PDGF/proneural tumor cells in brain tissue

To test our model prediction that NRAS/mesenchymal cells migrate faster than PDGF/proneural cells, we performed live-cell imaging on tumor-bearing mouse brain slices by confocal microscopy. Time-lapse images of GFP-positive tumor cells were used to track single-cell migration and generate single-cell trajectories. As shown in Fig. 3A and B and Supplementary Video S1, NRAS/mesenchymal tumor cells seemed qualitatively to move farther, have larger spread area, and polarize to a greater extent than PDGF/proneural tumor cells. Quantitative analysis of single-cell trajectories confirmed that NRAS/mesenchymal tumor cells have a higher random motility coefficient than PDGF/rponeural tumor cells (30.1 μm²/hour vs. 2.5 μm²/hour, P < 0.001; see Fig. 3C; Supplementary Fig. S4A). In addition, morphologic analysis of tumor cells revealed cell spread area and cell aspect ratio (i.e., polarization) are also higher in NRAS/mesenchymal tumor cells than PDGF/proneural tumor cells (406.6 μm² vs. 235.8 μm², P < 0.00001 and 2.1 vs. 1.7, P < 0.001, respectively, see Fig. 3D and E; Supplementary Fig. S4B and S4C). As predicted by our modeling, and the hypothesis that NRAS/mesenchymal has balanced motors and clutches whereas PDGF/proneural lacks sufficient clutches, we find NRAS/mesenchymal cells migrate faster, are more spread, and are more polarized than PDGF/proneural cells.

Figure 3.

NRAS/mesenchymal cells migrate faster, are more spread, and are more polarized than PDGF/proneural cells in ex vivo tumor-bearing brain tissue. A, Representative ex vivo fluorescent montage of GFP-tagged tumor cells (green) and blood vessels staining using isolectin B4 (magenta). B, Wind rose plots of NRAS/mesenchymal and PDGF/proneural migrating tumor cell migration trajectories (N = 46). C, NRAS/memsenchymal cancer cell motility is faster than for PDGF/proneural cancer cells. D and E, NRAS/mesenchymal cancer cells are more spread and polarized than PDGF/proneural cells as evidenced by larger spread area and aspect ratio. Solid and dashed lines represent the mean and median, respectively. Error bars represent the SEM. **, P < 0.001; ****, P < 0.00001. Statistical significance was assessed using the, Kruskal–Wallis rank test.

Migration phenotype is species- and tumor microenvironment– independent

To determine whether migration phenotype is cancer cell–intrinsic as predicted by our modeling and not due to microenvironment differences, we generated three primary mouse lines grown as neurospheres from each cohort to investigate their migration phenotype outside their tumor microenvironment. Organotypic mouse brain slice culture was used to image tumor cell migration in healthy mouse brain slices (34). Dissociated mouse tumor cells were plated and allowed to invade and migrate in healthy mouse brain slices. Figure 4A shows representative fluorescence images of primary isolated cells in organotypic slice culture. Time-lapse imaging was used to track single cells and quantify their migration rates. Consistent with the ex vivo migration in intact tumor-bearing brain slices, random motility coefficient in normal mouse brain tissue is higher for primary NRAS/mesenchymal tumor cells than primary PDGF/proneural tumor cells (131.4 μm²/hour vs. 31.3 μm²/hour, P < 0.00001; Fig. 4B; Supplementary Fig. S4D; Supplementary Video S2). The area of cell spreading is also higher in primary NRAS/mesenchymal tumor cells than PDGF/proneural (1,075.1 μm² vs. 838.8 μm², P < 0.001; Fig. 4C; Supplementary Fig. S4E). Furthermore, cell aspect ratio, the ratio of major and minor axes of a fitted ellipse, was trending higher in NRAS/mesenchymal tumor cells than PDGF/proneural but did not reach statistical significance (Fig. 4D; Supplementary Fig. S4F).

To assess the relevance of these results to human GBM, we tested the migration phenotype of six PDX lines (three mesenchymal and three proneural) using the organotypic mouse brain slice culture (Supplementary Table S5). Figure 4E shows representative fluorescence images of PDX cells in organotypic slice culture. We found mesenchymal PDX cells migrate faster than proneural PDX cells (43 μm²/hour vs. 8.8 μm²/hour, P < 0.00001; Fig. 4F; Supplementary Fig. S4G; Supplementary Video S3). In addition, similar to our mouse models, mesenchymal PDX cells have larger area of cellular spreading and aspect ratio relative to proneural PDX cells (701.7 μm² vs. 564.6 μm², P < 0.00001 and 2.3 vs. 2, P < 0.001, respectively, see Fig. 4G and H; Supplementary Fig. S4H and S4I).

Traction strain energy is larger for NRAS/mesenchymal cells than for PDGF/proneural cells, consistent with model predictions

In addition, cell migration simulations predict NRAS/mesenchymal tumor cells would have increased force generation as a result of higher number of clutches resulting in balanced myosin motors and clutches, relative to PDGF/proneural tumor cells which would have insufficient clutches relative to motors (Fig. 2D). Using the primary isolated mouse lines, TFM was used to measured traction strain energy generated by tumor cells on polyacrylamide hydrogels coated with type-I collagen (35, 38). Consistent with model predictions, NRAS/mesenchymal tumor cells generate higher traction strain energy than PDGF/proneural tumor cells across different substrate Young’s moduli (Fig. 5A and B). Cell spread area is also higher in NRAS/mesenchymal than PDGF/proneural on polyacrylamide hydrogels (Fig. 5C). In addition, NRAS/mesenchymal cells exhibit stiffness sensitive cell spreading; cells on stiff substrate (4.6 and 9.3 kPa) were more spread than on soft substrate (0.7 kPa, P < 0.00001; see Fig. 5C). Furthermore, we also examined force generation of mesenchymal and proneural PDX cells in mouse brain slices. Qualitative analysis of vasculature deformation is consistent with the model prediction that mesenchymal PDX cells generate larger deformations relative to proneural PDX cells. These deformations are identified from maximum-intensity projections of the vessel channel over time, where stationary regions remain gray and regions that shift over time appear colored. In this representation, deformation corresponds to a time-dependent displacement or bending of the vessel near the migrating tumor cell as shown in Fig. 5D and E and Supplementary Video S3.

Figure 5.

NRAS/mesenchymal cancer cells generate larger traction forces in vitro on 2D hydrogels and ex vivo in brain slices than PDGF/proneural cancer cells. A, Representative phase-contrast images and traction force magnitude heatmaps on type I collagen–coated polyacrylamide hydrogels across different substrate stiffnesses. B, NRAS/mesenchymal cancer cells (red) generate higher traction forces as measured by total strain energy relative to PDGF/proneural (blue) cancer cells. Measurement noise is presented (gray) across all stiffnesses. C, Similar to mouse brain slices, NRAS/mesenchymal cancer cells (red) cultured on collagen-coated polyacrylamide hydrogels have larger spread area than PDGF cancer cells (blue). D and E, Tissue deformation visualized using temporal color coding using 2-hours time-lapse and 15-minute time interval. D and E, Representative montage of three time points and corresponding color-coded maximum projection over time. Color indicates the time point at which the maximum intensity occurred; gray indicates no displacement; (D) proneural PDX cells and (E) mesenchymal PDX cells generate larger deformations of mouse brain capillaries (E) than proneural PDX cells shown in D. Arrows point to migrating cell direction. Scale bars, 20 μm, and inset scale bars, 10 μm. Solid and dashed lines represent the mean and median values, respectively. Error bars represent the SEM. +, P < 0.05; ****, P < 0.00001. Statistical significance was assessed using the Kruskal–Wallis rank test.

NRAS/mesenchymal mice have better survival and slower tumor growth rate

As NRAS/mesenchymal cells have nearly optimal CD44 expression, and therefore higher motility compared with PDGF/proneural cells which have a suboptimal low level of CD44 expression (17), we asked whether the differences in migration rate, morphology, and force generation correlate with disease progression and survival. Specifically, based on the faster migration in the NRAS/mesenchymal cohort, we expected that these mice would progress faster and die sooner than PDGF/proneural mice. To test this hypothesis, we measured survival times of tumor-bearing mice and found that, opposite to our expectation, the NRAS/mesenchymal cohort had better median survival than the PDGF/proneural cohort (NRAS N = 21, PDGF N = 24, 65 days vs. 35 days, log-rank test, P < 0.0001; Fig. 6A). To explain the difference in survival, we quantified in vivo tumor growth by BLI of tumor-bearing mice. Consistent with their shorter survival, we found that PDGF tumors grew twice as fast as NRAS tumors (slopes: 0.127 ± 0.01191 vs. 0.0716 ± 0.00343 P < 0.001; Fig. 6B and C). Using mouse tumor neurospheres, we quantified mouse primary tumor cell line proliferation rates in vitro and found no significant difference between NRAS/mesenchymal and PDGF/proneural cells (Fig. 6D). These results imply that an additional factor, besides proliferation or migration, enables the NRAS/mesenchymal mice to live longer and their tumors to grow slower in vivo than PDGF/proneural mice.

Figure 6.

NRAS/mesenchymal mice live longer than PDGF/proneural mice. A, Kaplan–Meier plot of animal survival for NRAS/mesenchymal and PDGF/proneural tumor-bearing mice shows extended survival of the NRAS/mesenchymal cohort. Log rank Mantel–Cox test P < 0.00001. B, BLI of NRAS/mesenchymal mice shows slower growing tumors relative to PDGF/proneural mice. C, Quantification of BLI integrated intensity for the two cohorts. Normalized BLI intensity for all data points and linear fits for each cohort. PDGF/proneural slope = 0.127 ± 0.01191 and NRAS/mesenchymal slope = 0.0716 ± 0.00343 D, Normalized cell count of primary isolated mouse tumor cells showing no significant difference in proliferation rate in vitro using neurosphere culture. Solid and dashed lines represent the mean and median values, respectively. Error bars represent the SEM. **, P < 0.001. Statistical significance was assessed using analysis of covariance to compare two regression lines.

NRAS/mesenchymal mice have increased immune response relative to PDGF/proneural mice

Because mesenchymal GBMs are known to be immunologically “hot”—which presumably confers a survival benefit due to an antitumoral immune response—relative to immunologically “cold” proneural GBMs (7, 9, 24), we assessed the extent to which mouse NRAS/mesenchymal tumors induce an immunologic response relative to PDGF/proneural tumors and normal brain tissues. Using our mouse transcriptomic dataset and previously published GBM immune gene sets (24), NRAS/mesenchymal tumors were found to have increased expression of both immune activators and suppressors gene signatures similar to human mesenchymal GBM (24), as shown in Fig. 7A and B. Specifically, expression of immune cell marker genes such as Aif1, Itgam (microglia/macrophages), and Cd3 (T cells) are elevated in NRAS/mesenchymal tumors relative to PDGF/proneural tumors [Fig. 7C (top)] and in mesenchymal GBMs relative to proneural and classic GBMs [Fig. 7C (bottom)]. Elevated expression of immune cell markers is indicative of increased immune cell infiltration in mesenchymal tumors. Consistent with the transcriptomic findings, IHC staining of NRAS/mesenchymal and PDGF/proneural tumor sections revealed significant levels of immune cell infiltration, including both microglia/macrophages and T-lymphocytes, in NRAS/mesenchymal but much less so in PDGF/proneural tumors (Fig. 7D). Furthermore, increased immune infiltration and activity in NRAS/mesenchymal tumors was accompanied by increased cell killing as measured by granzyme B and cleaved caspase-3 staining (Fig. 7D). Image clustering analysis was used to quantify CD3, IBA1, granzyme B, and cleaved caspase-3 staining, and statistically significant differences were observed between NRAS/mesenchymal and PDGF/proneural cohorts (Fig. 7E–H). This antitumor response was also evident in the three instances in which NRAS mice developed tumors and tumor regression was observed (Supplementary Fig. S5); whereas, in the long-surviving PDGF cohort, there was no evidence of tumor regression. Despite the antitumor immune response observed in NRAS/mesenchymal tumors, transcriptomic analysis also revealed elevated relative expression of immune checkpoint genes, including PDL1, CTLA4, and CD200R1 (Supplementary Fig. S6A and S6B). Thus, the NRAS/mesenchymal tumors, like human mesenchymal GBMs, are immunologically “hot” with evidence of both immune activation and immune suppression, as well as evidence of cell killing. Altogether, the enhanced immune cell–associated tumor cell killing provides a mechanism by which survival is extended in NRAS/mesenchymal tumors despite their enhanced migration speeds relative to PDGF/proneural tumors.

Figure 7.

NRAS/mesenchymal tumors are immunologically “hot” and PDGF/proneural tumors are immunologically “cold,” consistent with human GBM subtypes. A and B, Clustering analysis of immune activators and suppressors previously reported in human GBM (Doucette and colleagues; ref. 24). Similar to mesenchymal GBM, NRAS/mesenchymal tumors have elevated immune activator and suppressor expression relative to normal brain tissue and PDGF/proneural tumors. C, Expression of immune cell surface marker genes in mouse (top) and human (bottom) tumors shows elevated expression in NRAS/mesenchymal tumors and human mesenchymal GBM tumors relative to PDGF/proneural and human proneural GBM tumors. D, IHC confirms elevated immune cell infiltration in NRAS tumors and associated elevation of immune-mediated killing of tumor cells. E–H, Quantification of IHC images using a k-means clustering algorithm. Solid and dashed lines represent the mean and median values, respectively. Error bars represent the SEM. +, P < 0.05; *, P < 0.01; **, P < 0.001; ***, P < 0.0001; ****, P < 0.00001. Statistical significance was assessed using one-way ANOVA.

Brownian dynamics simulations and immunodeficient tumors model explain NRAS/mesenchymal and PDGF/proneural tumor progression

To quantitatively describe the interplay between tumor cell migration and proliferation and antitumoral immune response in tumor growth, we developed a 3D BDTS based on our original 1D Brownian dynamics simulator (39, 40). The simulator takes into account antitumoral immune cells that infiltrate tissue, migrate, proliferate, encounter cancer cells, deliver cytotoxic agents, dissociate from cancer cells, undergo exhaustion, and, eventually, undergo apoptosis (or egress to lymphatics) as shown in Fig. 8A. At the same time, cancer cells migrate, proliferate, and undergo CTL-mediated death in the presence of antitumoral immune cells in the case NRAS/mesenchymal tumors. In the case of PDGF/proneural tumors, no antitumoral immune cells were simulated. Figure 8B shows simulation output at days 0 and 16, which showed the observed behavior of overall faster growth of PDGF/proneural tumors. In NRAS/mesenchymal tumor simulations, cancer cells seem more dispersed, whereas, in PDGF/proneural simulations, cancer cells are less dispersed. Simulated tumor growths were plotted in Fig. 8C and shown in Supplementary Video S4. Using the parameters in Supplementary Table S7, including the experimentally observed single cell migration speeds and neurosphere proliferation rates, simulated tumors qualitatively recapitulate the in vivo growth profile of NRAS/mesenchymal and PDGF/proneural tumors without parameter adjustment (Fig. 8C).

Figure 8.

BDTS of 3D NRAS/mesenchymal and PDGF/proneural tumors. A, Schematic showing a diagram of the model. B, Simulator output at day 0 and day 16 for PDGF and NRAS simulations. In NRAS simulations, dead cancer cells are larger gray spheres and T cells are smaller black spheres, see Supplementary Video S4. C, Cancer cell count over time showing the difference in tumor growth in the presence of T cells with simulated PDGF/proneural tumors growing faster than NRAS/mesenchymal tumors. D, Kaplan–Meier plot for NRAS/mesenchymal and PDGF/proneural tumor-bearing mice in WT vs. immunodeficient (Rag1^−/−) background. NRAS/mesenchymal tumors show significantly reduced survival in Rag1^−/− mice compared with WT (log-rank Mantel–Cox test P = 0.012). PDGF/proneural survival is not affected. E, Summary of the main findings depicting the relationship between molecular subtypes, cellular adhesion, cell migration, and immune response.

Lastly, given the strong immune gene signature and immune cell infiltration observed in NRAS/mesenchymal tumors, we sought to determine whether this immune response functionally affects tumor progression. To test this, NRAS and PDGF tumors were induced in an immunodeficient (Rag1^−/−) mice to assess the role of the adaptive immune system in controlling NRAS/mesenchymal versus PDGF/proneural tumor growth. Compared with immunocompetent WT mice, NRAS tumor-bearing Rag1^−/− mice exhibited significantly shorter survival (WT n = 7, Rag1^−/−n = 10, 64 days vs. 45.5 days, log-rank test, P = 0.012). In contrast, PDGF tumor-bearing Rag1^−/− mice exhibited similar survival to WT controls (WT n = 8, Rag1^−/−n = 7, 33 days vs. 31 days, log-rank test, P = 0.952), Fig. 8D. This suggests that the adaptive immune system plays a crucial role in GBM progression in mesenchymal subtype. These combined findings are summarized in a schematic (Fig. 8E), which depicts the subtype-specific relationship between CD44-associated migration and immune infiltration and how these factors differentially influence tumor progression.

Discussion

Understanding glioma progression and the mechanism driving glioma cell migration is critical for the design of effective therapies. In this study, we developed high-grade glioma mouse models which capture the transcriptomic and the immune microenvironment changes associated with human proneural and mesenchymal GBMs. Using the mouse models and PDX lines, we defined a mechanistic difference in glioma cell migration which highlights a functional characteristic of GBM molecular subtypes. The migration difference was consistent with changes in cellular adhesion, notably by CD44, but not molecular motors such as myosin II motors. This finding points toward an antimigratory therapy approach targeted against cellular adhesion as opposed to myosin motors. With the negative phase III clinical trial of the integrin-inhibitor cilengitide for GBM (56), it is possible that integrins may not be the major adhesion molecules utilized by glioma cells to migrate but instead they could utilize CD44. Although anti-CD44 therapies have not been tried in GBM, an anti-CD44 monoclonal antibody therapy (RO5429083, Roche) has been investigated in phase I trials in patients with solid tumors and with AML (57, 58). Therefore, an anti-CD44 therapy could provide clinical benefits by slowing glioma migration.

Furthermore, the upregulation of CD44 in mesenchymal tumors is supportive of the existing literature which defines CD44 as a marker of cancer stem cell and EMT (19, 20, 53). During EMT, cancer cells take a more mesenchymal migratory phenotype to allow them to migrate through dense ECM and metastasize (59). Consistent with prior mechanistic studies demonstrating a direct role for CD44 in glioma cell migration (17), our results associate enhanced migration in mesenchymal glioma cells with increased traction forces due to increased adhesion molecules expression “clutches” such as CD44. Additionally, Anderson and colleagues (60) demonstrated a functional role for CD44 in glioma cell migration using brain tissue slice assays, in which both CD44 knockout cells and anti-CD44 monoclonal antibody treatment resulted in reduced migration. Similarly, in breast cancer cells, TGF-β–induced EMT is associated with increased traction forces and clutch number (21). Interestingly, downregulation of NF1, a negative regulator of Ras, in epithelial breast cancer cells and Schwann cells also induces expression of transcription factors related to EMT (61). Moreover, in our study, NRAS^G12V expression was used to mimic NF1 downregulation and inactivation in mesenchymal GBM (6, 62). Ras hyperactivation of the MAPK pathway is required for EMT but not PI3K activation by Ras (63). Altogether, our results implicate EMT in enhanced glioma cell migration and force transmission associated with increased molecular clutches through and suggest upregulation of clutches, either integrins or CD44, as a conserved feature of EMT across a range of cancers.

Despite the faster migration of NRAS/mesenchymal cells, the antitumoral immune response within the NRAS/mesenchymal mouse model is able to slow disease progression and improve survival despite enhanced migration relative to the PDGF/proneural mouse model. Such an antitumor response could potentially be used to slow disease progression and improve clinical outcome for patients with GBM. In both mouse and human GBMs, mesenchymal tumors are immunologically “hot” relative to the immunologically “cold” proneural tumors (7, 24). Despite the presence of immune cells within mesenchymal tumors, immune suppression leads to tumors ultimately prevailing against the antitumoral immune response. Based on these findings, we propose an immune checkpoint inhibition strategy, in combination with an antimigratory therapy, targeting mesenchymal GBM but not proneural GBM.

Our study utilizes an integrated, state-of-the art experimental approach to study GBM progression and model GBM molecular subtypes by switching a single oncogenic driver (NRAS^G12V ↔ PDGFB) in an immunocompetent background without the need for genetically engineered mouse strains or further breeding. Using live-cell and brain slice imaging, we identify key mechanical differences between mesenchymal and proneural tumor cells, with mesenchymal cells have larger cellular spread area, generate larger forces, and migrate faster. The functional differences were all predicted by a motor-clutch model for cell adhesion and migration where mesenchymal cells have an optimal level of CD44-mediated adhesion (clutches) relative to myosin motors, whereas proneural cells lack sufficient CD44 to match the myosin motor activity. Despite the faster migration, NRAS/mesenchymal mice live longer, consistent with the presence of an antitumoral immune response that is lacking in PDGF/proneural mice. This was further validated using immunodeficient mice: NRAS tumors showed worsened survival in Rag1^−/− mice, whereas PDGF tumors showed no difference in survival between WT and Rag1^−/− animals. Such dynamics are readily captured computationally with little parameter adjustment using a 3D BDTS. Our results support a model in which mesenchymal GBM exhibits high CD44 expression and strong immune infiltration, whereas proneural GBM shows lower CD44 and minimal immune activity (Fig. 8E). Overall, this work establishes an integrated in vivo genetic and biophysical modeling framework to connect animal model and human transcriptionally defined subtypes to fundamental mechanistic understanding, which has the potential to enable a new modeling-centric approach to clinical translation with application in a wide range of cancers (64). Although our model focuses on comparing mesenchymal and proneural GBM subtypes by altering a single oncogenic driver, it does not fully capture the cellular heterogeneity found in human GBM, in which different subtypes and cell states often coexist within the same tumor (8, 9). That said, this controlled system allows us to isolate and study the behaviors of each subtype in an immunocompetent setting. This approach is valuable for identifying subtype-specific mechanisms and for building predictive models. Future work incorporating additional mutations, lineage tracing, or single-cell analysis could help bridge the gap between this model and the cellular diversity seen in patients.

Supplementary Material

Table S1

Table S1: Related to Figures 1A and 7D. IHC antibodies and reagents

Table S2

Table S2: Related to Figure 1. FPKM expression of mouse tumors and healthy brain tissues.

Table S3

Table S3: Related to Figure 1C. List of genes identified in each cluster (MC1, MC2, MC3, HC1, HC2, HC4).

Table S4

Table S4: Related to Figure S2. List of previously published subtype gene signatures (CL, MES, and PN gene signatures) and subset with normalized standard deviation >1 in mouse and human datasets

Table S5

Table S5: Characteristics of patient-derived xenograft (PDX) lines used in this study

Table S6

Table S6: Cell migration simulator parameter values

Table S7

Brownian dynamics tumor simulator parameters values

Figure S1

Figure S1. Unsupervised clustering of mouse tumor and healthy brain tissue transcriptomic profiles and pathway enrichment analysis. A) Heatmap showing expression profile of NRAS and PDGF tumors and healthy mouse brain tissues. Heatmap shows existence of three gene clusters: Tumor-specific cluster, normal brain tissue specific cluster and NRAS tumor-specific cluster. B,C&D) Gene ontology analysis of gene clusters was performed using EnrichR.

Figure S2

Figure S2. Clustering analysis of mouse and human tumors using gene signatures associated with classical, mesenchymal, and proneural subtypes. A) Heatmap shows the clustering of human GBM samples using subtype-specific gene signatures. B) Heatmap shows the clustering of mouse tumors using subtype-specific gene signatures. C&D) Quantification of subtype gene signatures within human GBM subtypes and NRAS and PDGF mouse tumors. Solid and dashed lines represent mean and median values respectively. Error bars are S.E.M. +p <0.05, * p <0.01, ** p<0.001, *** p<0.0001, **** p<0.00001.

Figure S3

Figure S3. Differential expression analysis of cell migration genes. Volcano plot depicting –log10(adj p value) calculated using FDR adjusted Student’s t test versus log2(average FC) for individual genes in A) mouse tumors and B) human TCGA-GBM

Figure S4

Figure S4. Quantification of single cell migration and cell morphology. A, B, & C) Ex vivo migration and morphology data from each individual animal. D, E, & F) Migration and morphology data for each individual mouse primary tumor line in mouse organotypic brain slice. G, H, & I) Migration and morphology data for each individual PDX line in mouse organotypic brain slice.

Figure S5

Figure S5. BL image sequences showing tumor regression in some rare cases. All three surviving mice in NRAS cohort showed tumor regression.

Figure S6

Figure S6. Expression of immune checkpoint genes within mouse tumors. Expression of immune checkpoint genes were upregulated in NRAS tumors relative to PDGF tumors (B) similar to human mesenchymal and proneural (A).

Video S1

Related to Figure 3

Video S2

Related to Figure 4A

Video S3

Related to Figure 4D

Video S4

Related to Figure 8

Data Availability

Fastq files and the Cuffnorm output were deposited at Gene Expression Omnibus (GSE161154).

Code Availability

All data and codes are available on the Odde laboratory website (BDTS simulation and analysis codes are at https://drive.google.com/file/d/1T4FEdRsEchBCMSvWy8vhU4QvF7ABjVqn/view and https://drive.google.com/file/d/1MdPpuBp4HZFKuTBRAV028EHsrenII7-1/view, respectively). The CMS code is available at https://drive.google.com/file/d/1Qx0xvYhsWSrUPC8zXUmmtMW2ZTOzh-Tk/view. Both BDTS and CMS codes are also located at https://github.com/davidodde/shamsan_etal_codes. Codes are also available upon request to the corresponding author.

Authors’ Disclosures

J.N. Sarkaria reports grants from Rain Therapeutics, Sumitomo Dainippon Pharma Oncology, Boehringer Ingelheim, ModifiBio, ABL Bio, Otomagnetics, Inhibrx, Breakpoint Therapeutics, Bristol Myers Squibb, Cybrexa, Australis, Curtana, OncoMAGNETx, Reglagene, S1 Oncology, and SPARC outside the submitted work. H.B. Clark reports grants from the NIH during the conduct of the study. D.A. Largaespada reports other support from Luminary Therapeutics, Recombinetics, Styx Biotechnologies, ImmunSoft, and NeoClone outside the submitted work. No disclosures were reported by the other authors.

Authors’ Contributions

G.A. Shamsan: Data curation, formal analysis, methodology, writing–original draft, writing–review and editing. C.J. Liu: Data curation, writing–review and editing. B.C. Braman: Data curation. R. Li: Data curation. S.K. Rathe: Formal analysis. A.L. Sarver: Formal analysis. N. Ghaderi: Data curation. M. McMahon: Data curation. R.L. Klank: Conceptualization. B.R. Tschida: Conceptualization. S.J. McFarren: Data curation. P.C. Rosato: Resources. D. Masopust: Resources. J.N. Sarkaria: Resources. H.B. Clark: Formal analysis. S. Rosenfeld: Supervision, funding acquisition, writing–review and editing. D.A. Largaespada: Conceptualization, supervision, funding acquisition, writing–review and editing. D.J. Odde: Conceptualization, supervision, funding acquisition, writing–review and editing.