Tumor tissue was collected directly during surgery. Patients included in the study gave written informed consent and specimen collection was approved by the local Institutional Review Board of Charité University Medicine, Germany (EA4/019/12), the Ethics Committee of the University of Rostock, Germany (A 2009/34) and the Institutional Ethics Committee of the Medical University of Graz, Austria (24-402 ex 11/12).

Animal studies were performed in accordance with the German Animal Welfare Act and approved by local authorities (Landesamt für Gesundheit und Soziales, LaGeSo Berlin, Germany) under the permission H0308/18 for the PDX generation and proliferation in vivo and A0010/19 for in vivo therapy experiments. For establishment of s.c. PDX models, patient tumor specimens were cut into 3-4 mm sized fragments and transplanted s.c. on anesthetized NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ (NSG) mice (Charles River Laboratories, Sulzfeld, Germany). Health status of mice was checked daily, and body weight and tumor size were measured twice per week. When tumors successfully engrafted and reached a size of about 1 cm³, they were excised, cut into 3×3×3 mm sized fragments and transplanted s.c. on female NMRI Foxn1nu mice for the consecutive in vivo passage. Mice were housed under standard conditions in IVC caging systems with 22°C +/- 1°C, 12 h light-dark cycle, food, water and nesting material ad libitum.

PDX tumor tissue obtained from s.c. in vivo passage (passage numbers between 4 and 9) was used to prepare a single cell suspension via mechanical break up (gentleMACS Dissociator, Miltenyi Biotec, Teterow, Germany). Anesthetized mice were fixed in a stereotactic frame, the skin on the scull was opened and 2 µl cell suspension of 1×10⁵ tumor cells was injected intracerebral (i.cer.) into the cortex of the right hemisphere. After cell injection, the syringe was removed slowly, and the skin was closed using Histoacryl ^® tissue adhesive (B.Braun, Melsungen, Germany). The following day mice received meloxicam subcutaneously.

Animal health condition was checked daily and body weight was measured at least twice weekly. Mice were terminated for ethical reasons when showing behavioral abnormalities and body weight loss > 10%; both signs for progressive intracranial tumor growth. Brains were dissected and frozen in isopentane at -80°C. Sequential 10 µm cryostat microtome sections in coronal plane were prepared, cresyl violet-stained and the tumor area measured as described previously (18, 24).

For sensitivity testing of s.c. PDX models, 3×3×3 mm tumor fragments (passage numbers between 2 and 6) were transplanted subcutaneously onto female Rj : NMRI-Foxn1nu/nu nude mice (Janvier Labs, Le Genest-Saint-Isle, France) as described (see above). Animals were randomized and treated at palpable tumor size (0.08 to 0.2 cm³) with the respective drugs: bevacizumab (10 mg/kg intraperitoneally, 3× per week for 2 weeks; Roche, Basel, Switzerland), everolimus (5 mg/kg orally, (day 1-5)×2; Novartis, Basel, Switzerland), irinotecan (15 mg/kg intraperitoneally, (day 1-5)×2; Fresenius, Bad Homburg, Germany), salinomycin (10 mg/kg orally, day 1-14; Sigma-Aldrich, St. Louis, MI, USA), Sorafenib (80 mg/kg, orally, (day 1-5)×2; Bayer, Leverkusen, Germany) and temozolomide (90 mg/kg orally, day 1-5; MSD, Kenilworth, NJ, USA). Drug efficacy was determined by measurement of tumor volumes (TV). TV measurement was performed with a digital caliper and volumes were calculated using the formula:

Studies were terminated for ethical reasons when first animals reached a TV > 1.5 cm³. To describe therapeutic responses, the ratio (T/C, in %) of mean volumes of tumors in treatment (T) versus control groups (C), as well as relative tumor volumes (RTV) in treatment groups were used:

A T/C value > 50% was defined as no response, > 25% as minor response, > 10% as moderate response and T/C ≤ 10% a strong response. We furthermore considered a RTV value > 1.7 as progression, a RTV > 0.3% stable disease, and RTV ≤ 0.3% as partial remission.

For sensitivity testing of orthotopic GBM PDX models, tumor cells were inoculated, and animals randomly distributed into control and treatment groups. Treatment was started 6 to 7 days after tumor cell inoculation. Once first animals showed health impairments, the study was terminated and maximum tumor area in coronal plane measured microscopically and used to evaluate therapeutic responses.

For statistical analyses of tumor size differences between control and treatment groups at study end, one-way ANOVA followed by Dunnett’s multiple comparison test were performed. Prism software v5.02 (GraphPad Software, San Diego, USA) was used for all analyses.

For histopathological analysis, 5 µm sections of formalin fixed paraffin embedded (FFPE) tumor samples were deparaffinized in ROTICLEAR^® (Carl Roth GmbH, Karlsruhe, Germany) and rehydrated using ethanol and distilled water. Sections were then stained according to a standard hematoxylin-eosin protocol ​ (24)​. For immunohistochemistry (IHC) antigen retrieval was done in hot citrate buffer for 15 min, followed by cooling for 40 min and washing with PBS.

For MGMT staining, slides were blocked with 5% goat serum in PBS for 1 h at room temperature (RT), and subsequently incubated with primary antibody (5 µg/ml rabbit anti-MGMT, PA5-79668, ThermoFisher Scientific, Germany), for 1 h at RT. Sections were then washed twice in TBST buffer (20 mM Tris/HCl pH 7.5, 150 mM NaCl, 0.1% Tween-20) and incubated with the SuperVison 2 Single Species HRP-polymer rabbit kit (DCS GmbH, Hamburg, Germany) according to the manufacturer’s instructions. Sections were counterstained with hematoxylin solution.

For Ki-67 staining, slides were first blocked with 3% hydrogenperoxide for 30 min, washed twice with PBS and then with 5% normal goat serum in PBS for 30 min, both at RT. After blocking, sections were incubated with primary antibody (6.5 µg/ml rabbit anti-Ki-67 in PBS, ab15580, Abcam, Cambridge, UK) for 1 h at RT. Sections were then washed twice in TBST buffer and incubated with HRP-conjugated secondary antibody (#111035003, 1:200, Jackson ImmunoResearch Laboratory, West Grove, USA) for 30 min, RT. Sections were washed twice in TBST, followed by detection of secondary antibody using chromogen substrate buffer DAB+ (#K3468, Dako North America, Inc., Carpinteria, CA, USA) and DAB-chromogen (#K3468, Dako North America, Inc., Carpinteria, CA, USA). Sections were washed with PBS, then with distilled water and counterstained with hematoxylin solution.

All Images were acquired using the Axioskop 40 and AxioVision V3.5 (Zeiss, Jena, Germany).

RNA sequencing (RNASeq) of untreated s.c. tumor tissue samples was performed for 23 established GBM PDX models. Next-generation sequencing and processing of the raw data was performed by ATLAS Biolabs GmbH (Berlin, Germany).

About 50-100 mg snap frozen PDX tumor tissue was dissolved in 1.5 ml of TRIzol™ (ThermoFisher Scientific, Carlsbad, CA, USA) using a gentle MACS dissociator and M tubes (Miltenyi Biotec, Bergisch Gladbach, Germany). The integrity of the isolated total RNA was analyzed using an Agilent Bioanalyzer 2100 and the RNA 6000 Nano Kit (Agilent, Santa Clara, CA, USA). The Illumina TrueSeq Stranded mRNA Library Prep Kit was used for preparation of RNAseq libraries, followed by 100 bp PE-sequencing on an Illumina HiSeq 2500 device with a depth of 80–100 million reads (40–50 Mio cluster) (Illumina, Cambridge, UK).

Read quality was validated with FastQC v0.11.8 (25). Human (reference: ensembl hg38) and mouse (reference: ensembl mm10) reads were split with Xenome v1.0.1 (26). STAR aligner v2.6.1a (27), QualiMap v2.2.1 (28) and eXpress v1.5.1 (29) were used to map the human-specific reads and quantify transcript expression.

GATK 4.0.2.1 (https://github.com/broadinstitute/gatk/releases) (30) and the Ensemble Variant Effect Predictor (VEP), release 94 (https://www.ensembl.org/info/docs/tools/vep/index.html) (31) were used for variant calling and annotation of mapped reads. Additionally, to verify annotations and to further interpret variants of interest the openCRAVAT pipeline v2.2.7 was used (32). The mutational analysis was performed for a set of GBM-relevant genes (33). Quality-tested variant calls were filtered based on allele frequencies from gnomAD 3.1. to identify somatic alterations. Variants not included in gnomAD 3.1 or with a gnomAD allele frequency below 0.001 were considered ( Supplementary Table S1 ).

Copy Number Variations (CNV) were predicted from gene expression mutational data using RNAseqCNV (34) and CaSpER (35).

RNAseq raw count data were transformed to gene length corrected trimmed mean of M-values (GeTMM) (36) to perform single-sample gene set enrichment analyses (36) (https://github.com/broadinstitute/ssGSEA2.0) regarding gene sets of the molecular signature database (MsigDB v2022.1) (37). Subtype classification was performed based on published gene set expression profiles (12). Group-wise comparison of enriched gene sets between responding and resistant PDX models towards individual therapy was performed with gene set enrichment analyses v4.3.2 (37).

A total of 131 tumor tissues from patients diagnosed with GBM were used for PDX generation. Of these, 39 models (30%) engrafted and were consecutively passaged subcutaneously for a minimum of three times and were then considered stable, established s.c. PDX. Median age of patients at resection is 63, and 15 of the characterized GBM PDX were from male patients (58%) ( Table 1 ), reflecting the clinical situation of GBM grade IV. Until now, 26 GBM PDX models have been subject to chemosensitivity testing, 23 of which subject to molecular characterization by RNA sequencing. These 26 models include 19 GBM, 4 recurrences of GBM and 3 GBMs with unknown treatment history. In addition, 15 GBM PDX were established as orthotopic, intracranial models, of which 4 have been subject to sensitivity testing.

Our established GBM PDX models closely resemble the respective patient’s tumor histology, with similar morphology and expression patterns of MGMT and Ki-67 ( Figure 1A ). Expression of proliferation marker Ki-67 was comparable over several s.c. passages and after parallel orthotopic inoculation, as seen in immunohistological stainings of PDX tumor sample sections ( Figure 1E ). Individual orthotopic PDX models were able to retain the characteristic infiltrative growth of GBM ( Figure 1D ), compared to nodular growing s.c. tumors. Furthermore, each model displayed stable characteristic growth patterns over several s.c passages, with a slight increase in tumor growth rate in higher passage numbers ( Figure 1B ). Tumor doubling times varied highly between different s.c. models ( Figure 1C ), reflecting the heterogeneity of our GBM PDX panel.

PDX tumor bearing mice received monotherapies of everolimus, sorafenib, bevacizumab, irinotecan, salinomycin or standard of care temozolomide for up to 2 weeks. Tumor growth was monitored during therapies until the study was terminated for ethical reasons ( Figure 2A ). Of the administered compounds, salinomycin and sorafenib showed no efficacy in our panel of GBM PDX. The anti-VEGF antibody bevacizumab and mTORC1 inhibitor everolimus showed only limited efficacy, leading to moderate (T/C < 25%) to strong responses (T/C < 10%) and stable tumor volumes (RTV 0.34-1.7) in 7 and 4 out of 26 models, respectively. A better response rate was seen for topoisomerase I inhibitor irinotecan, which led to stable tumor volumes in 20 and tumor regression (RTV > 0.34) in 2 models. When compared compared to untreated PDX tumors we only observed a moderate to strong response in 17 models. The standard of care drug temozolomide showed the best tumor growth inhibition among the tested compounds. We observed moderate to strong responses in 21 models, with treatment leading to stable tumor volumes in 15 models and tumor regression in 7 models ( Figure 2B ). Interestingly, Glio11414, a PDX established from a recurrent GBM previously treated with temozolomide, showed resistance toward temozolomide.

The BBB and the tumor microenvironment within the brain parenchyma significantly contribute to tumor biology and eventually treatment outcome in GBM. We therefore compared efficacy of bevacizumab, everolimus, irinotecan and temozolomide in 4 orthotopic GBM PDX ( Figure 2C ). To evaluate treatment responses between s.c. and orthotopic PDX we used T/C at study end [%].

The targeted drugs bevacizumab and everolimus performed less efficiently in orthotopic PDX tumors compared with matching subcutaneously growing PDX tumors ( Figure S1 ). None of the orthotopic models were considered responders to everolimus (T/C > 50%), while it caused a reduction of tumor volumes (T/C < 50%) in two out of four matching s.c. models. For bevacizumab, differences in efficacy in s.c. versus orthotopic models were even more pronounced. All s.c. growing PDX models were considered minor to moderate responders (T/C 50 - 10%), but no response was seen in the four matching orthotopic GBM models. Regarding irinotecan, chemosensitivity of s.c. vs. orthotopic models was different. Irinotecan showed similar efficacy in orthotopic and matching s.c. Glio10618 and Glio12032 models. Compared to their respective vehicle treated control group, orthotopic tumors were significantly reduced in size ( Figures S1B, C ) and showed moderate responses (T/C 10% - 25%). However, in the orthotopic PDX Glio13066, irinotecan showed no efficacy (T/C 84%), whereas it caused pronounced tumor growth inhibition in the matching s.c. model (T/C 7%). Out of the four tested drugs, temozolomide showed best efficacy in the orthotopic PDX, where in all 4 tested models glioma growth was reduced (T/C 2% - 41%), with significant differences to control groups in 2 out of 4 tested models ( Figures S1A–D ). Temozolomide efficacy in s.c. models was slightly better in all models showing moderate to strong responses (T/C 1% - 13%). In conclusion, our diverging results point to the importance of orthotopic GBM PDX models to evaluate drug efficacy in a clinically relevant setting.

The mutational status of 23 GBM PDX was analyzed using the available RNA sequencing data regarding genes frequently mutated in GBM ( Table S2 ). In 23 of these genes, we could identify somatic mutations, with the most frequent mutations occurring in EGFR and PARP1 in 9 PDX models each. In addition, mutations in TP53 (7 PDX), FAT1 and PTEN (6 PDX each), as well as ATM, BRCA2 and MTOR (each in 5 PDX) were identified. Overall, in 12 PDX mutations within the PI3K/Akt/mTOR signaling pathway could be found ( Figure 3A ). We furthermore used available RNA sequencing data to visualize the expression of mutated genes, allowing for selection of PDX expressing specific targets for preclinical studies. Interestingly, 6 of the 9 PDX bearing mutations in EGFR showed a comparably high expression of the mutated receptor ( Figure 3B ).

We performed copy number variation analysis and principal component analysis to determine transcriptomic similarities within the cohort and to identify potential molecular subgroups. While clustering in copy number variation remained inconclusive ( Figure S2 ), the principal component analysis revealed two clusters of three and four models, and a third cluster of 16 models. Cluster I consists of Glio10995, Glio12421 and Glio11413, whereas cluster II contains Glio10535, Glio10888, Glio11368 and Glio13066 ( Figure 4A ). To clarify whether the observed clustering is in line with molecular GBM subtypes mesenchymal, classical and proneural, we analyzed a set of 150 genes described as characteristically expressed in these subtypes ( Figure 4B ) (12). Members of cluster I are enriched for mesenchymal signature genes, while members of cluster II are enriched for signature genes of the classical phenotype. The remaining 16 models formed a third cluster defined by comparably high enrichment of genes related to the proposed proneural subtype, but with enrichment in more than one molecular subtype in individual models.

To gain further insight into the tumor biology and cancer-related pathways of tumor progression in our PDX models, we compared their transcriptomes according to gene sets of cancer hallmarks ( Figure 4C ). Models of cluster I, the proposed mesenchymal subtype, also clustered in gene sets like epithelial mesenchymal transition, hypoxia, glycolysis and the P53 pathway, inflammation-related signatures like response to interferon alpha/gamma and TNF signaling, as well as angiogenesis. The two adjacent clusters of three models each showed comparably low enrichment scores in the mentioned hallmarks, but high scores in hallmarks G2M checkpoints, as well as E2F targets. The remaining PDX models showed individual expression profiles regarding the analyzed hallmarks, with three of the four PDX resembling the classical subtype grouping together.

Since we observed pronounced clustering in VEGF signaling and angiogenesis, the PI3K/Akt/mTOR pathway, and MAPK and JAK/STAT signaling ( Figure 4D ), all commonly deregulated in GBM, we analyzed expression of related gene sets in more detail. In VEGF and angiogenesis-related gene set analyses ( Figure S3A ), models resembling the mesenchymal subtype grouped together with two PDX mostly resembling the classical subtype, but with additional enrichment of the mesenchymal gene expression signature (Glio10535 and Glio13066). Another cluster of five models had comparably low enrichment of analyzed gene sets, while the remaining PDX showed different individual expression profiles. Clustering could also be observed in gene set enrichment analyses of PI3K/Akt/mTOR-related gene sets ( Figure S3B ), where four models were separated by comparably high enrichment scores, especially in PI3K/Akt dependent signaling. Another seven models showed comparable low enrichment scores for PI3K/Akt-dependent signaling, while the remaining 12 models showed different individual expression profiles regarding the analyzed gene sets. In analyses of MAPK/ERK and JAK/STAT signaling-related gene sets ( Figure S3C ), models of the mesenchymal subtype were again separate from the remaining models and characterized by overall high enrichment scores. The two PDX resembling the classical subtype the most pronounced (Glio10888 and Glio11368) were localized in a cluster of four models with comparably low enrichment scores, while remaining models showed individual expression profiles.

In a next step we analyzed potential correlations between gene expression profiles and treatment responses of PDX tumors via gene set enrichment analyses. Comparison of expression profiles of irinotecan responders and non-responders did not reveal significant differences between the two phenotypes. Despite the observed clustering of PDX models in single-sample gene set enrichment analyses regarding their expressions of different VEGF signaling and angiogenesis-related gene sets, there was no statistically significant difference in between bevacizumab non-responders and responders. For temozolomide, our analyses did reveal significant differences (p < 0.1 and FDR < 25%) in expression profiles between responding and non-responding PDX ( Figure 5A ). In temozolomide resistant models, expression of hallmark gene sets hypoxia and mTORC1 signaling were upregulated, while no correlation between the expression of MGMT, a known predictive marker for temozolomide resistance (5, 6), and monitored responses could be detected. However, only one of these models were sensitive towards the mTOR inhibitor everolimus. Comparing expression profiles of everolimus responders and non-responders, responders showed higher scores for hallmark gene sets epithelial mesenchymal transition and angiogenesis, as well as in inflammatory response, TNF alpha signaling via NF-kappaB, IL6/JAK/STAT3 signaling and the reactive oxygen species pathway ( Figures 5B , S4 ).