Intertumoral heterogeneity in patient-specific drug sensitivities in treatment-naïve glioblastoma

Glioblastoma biopsies were obtained from 12 informed patients with explicit written consent undergoing surgery for GBM at Oslo University Hospital, Norway as approved by The Norwegian Regional Committee for Medical Research Ethics (REK 2017/167). The GSC cultures were established both from several focal tumor biopsies and ultrasonic aspirate generated during surgery. The IDH status was evaluated by immunohistochemistry and sequencing, and the MGMT promoter methylation status was evaluated by methylation-specific quantitative PCR. Cell cultures were established and maintained in serum-free media containing bFGF and EGF (both R&D Systems), as previously described [14]. Differentiation was induced, and cells fixed and stained, as previously described [14]. Images were acquired using Olympus Soft Imaging Xcellence software v.1.1. The total number of cells from one passage to the next in serial passages was extrapolated using the formula (total number of cells from previous passage/cells plated) x (total number of cells from current passage). All experiments in this study have been performed within the 10th passage of individual GSC cultures. Patient characteristics are summarized in Additional file 1.

Cells were suspended in PBS with 2% fetal bovine serum (Biochrom) and stained with directly conjugated antibodies (CD15-PerCP, R&D Systems, CD44-APC, Thermo Fisher Scientific, CD133-PE, Miltenyi Biotec, CXCR4-PE, Miltenyi Biotec) according to the manufacturer’s instructions. Cells were washed three times before analysis by flow cytometer LSRII (BD Bioscience). FlowJo software v.10.4.1 was used for data analysis. Dead cells were identified by propidium iodine (Thermo Fisher Scientific), and doublets were excluded by gating.

The National Animal Research Authority approved all animal procedures (FOTS 8318). C.B.-17 SCID female mice (7–9 weeks old, Taconic) were anesthetized with an injection of zolazepam (3.3 mg/mL), tiletamine (3.3 mg/mL), xylazine (0.45 mg/mL) and fentanyl (2.6 μg/mL) and placed in a stereotactic frame (David Kopf Instruments). Cells were prepared and transplanted, as previously described [14]. The animals were regularly monitored for signs of distress and killed by cervical dislocation after 15 weeks or earlier if weight loss > 15% or neurological symptoms developed. The brains were harvested and further processed as previously described [14]. Images of brain sections were acquired using Axio Scan.Z1 (Carl Zeiss). Processing of images was performed using ImageJ 2.0.

The oncology drug collection consisted of 461 compounds and covered most U.S. Food and Drug Administration and European Medicines Agency (FDA/EMA)-approved anticancer drugs and investigational compounds with a broad range of molecular targets. The complete drug collection is listed in Additional file 2. The compounds were dissolved in 100% dimethyl sulfoxide (DMSO) and dispensed on 384-well plates using an acoustic liquid handling device, Echo 550 (Labcyte Inc). The pre-drugged plates were kept in pressurized Storage Pods (Roylan Developments Ltd.) under inert nitrogen gas until needed. The patient-derived GSCs were plated at a density of 3000 cells/well using a MultiDrop Combat (Thermo Scientific) peristaltic dispenser. The plates were incubated in a humidified environment at 37 °C and 5% CO₂, and after 72 h cell viability was measured using CellTiter-Glo® Luminescent Cell Viability Assay (Promega) with a Molecular Device Paradigm plate reader. The resulting data were normalized to negative control (DMSO) and positive control wells (benzethonium chloride). The quantification of drug sensitivity was utilized by the drug sensitivity score (DSS), as previously described [22, 23]. In brief, each drug was evaluated over a 5-point dose-escalating pattern covering the therapeutic range. The resulting dose-response was analyzed by automated curve fitting defined by the top and bottom asymptote, the slope, and the inflection point (EC₅₀). The curve fitting parameters were used to calculate the area defined as area of drug activity (between the 10 and 100% relative inhibition to positive and negative control) into a single measure as the DSS. The selective drug sensitivity score (sDSS) of each compound was calculated as the difference between the DSS in the individual culture and the average DSS of all screened GBM cultures. One culture (T1505) was excluded from the analysis of the overall drug sensitivity due to an error in the automatic seeding procedure for 29% (132/461) of the drug responses.

Cells were plated at 5000 cells/well in a 96-well plate (Sarstedt, Germany) under sphere conditions, cultured for 24 h before the addition of drugs and further incubated for 72 h. Viability was assessed using Cell Proliferation Kit II XTT (Roche) solution incubated for 24 h before analysis on a PerkinElmer EnVision. The viability is corrected for the background signal and reported relative to negative control (DMSO), as the mean and standard error to the mean of five independent experiments.

Next generation sequencing and gene expression microarray experiments were performed at the Genomics and Bioinformatics Core Facility at the Norwegian Radium Hospital, Oslo University Hospital (Norway). The library preparation for RNA sequencing was performed using the Truseq mRNA Illumina protocol, and the samples were sequenced on the Illumina HiSeq platform (paired end 2 × 75 bp). Normalized expression data was further analyzed in J-Express 2011. Subgrouping of the GSC cultures as proneural or mesenchymal was performed by analyzing gene expression microarray data using the HumanHT-12 chip (Illumina). Unsupervised hierarchical clustering was performed according to the gene panels described by Mao et al. and Phillips et al. [24, 25]. Quality issues led to one culture (T1461) not being successfully sequenced and could not be included in the gene expression analyses.

The robustness of the patient-derived GSC model system in preserving the tumorigenicity and molecular features of the parent tumor is well documented by us and others [12‐16, 26]. Such patient-derived GSCs, however, display considerable intertumoral differences in morphology and behavior in vitro and in vivo [12, 14].

In this sample cohort, eleven cultures formed free-floating tumorspheres, while one culture proliferated adherently (T1505). The individual cultures maintained their morphology upon serial passages and could be serially expanded. Intertumoral differences were observed in the in vitro spheroid and differentiation morphology, expression of GSC markers, total cell yield after serial passaging, and in vivo tumor formation characteristics (Fig. 1). Overall, the GSC cultures presented with considerable tumor-to-tumor variability in both morphology and behavior in vitro and in vivo, while maintaining culture specific characteristics.

Subsequently, we explored whether the intertumoral heterogeneity among GSC cultures is reflected in the sensitivity to a collection of 461 anticancer compounds using automated high-throughput technology. An overview of the drug collection is provided in Table 1. Reproducibility of the HTS was assessed by repeated screenings evaluated by a blinded investigator and displayed a ranked correlation of r = 0.823 (Spearman, p < 0.0001). The median passage number at the time of drug screening was 3 (range: 1–7).

A DSS ≥10 was defined as the threshold to classify a drug response as moderate to strong (Fig. 2a). Following DSRT, in total, 115 compounds (25% of the entire drug collection) displayed this response in the GSC culture cohort. The median was 33 drugs (range: 22–95). Two cultures, T1459 and T1506, clearly had higher number of drugs with a DSS ≥10, 79 and 95 drugs, respectively (Fig. 2b). The sensitivity to any given drug was, however, heterogeneous, as 93 of the 115 drugs (81%) with a DSS ≥10 displayed intersample differences equivalent to a moderate to strong difference in sensitivity (∆DSS ≥10, DSS_max - DSS_min). The overall sensitivity to the entire drug collection (n = 461) significantly differed among all GSC cultures (p < 0.0001). Based on the differences in the overall drug sensitivity, the cultures were broadly clustered into three major categories of most (T1459 and T1506), moderate (T1461, T1502, T1547, T1456, T1550) and least (T1454, T1561, T1549, T1548) sensitive cultures (Fig. 2c, Additional file 3). Correspondence analysis of the DSS to all drugs clustered the two most sensitive cultures distinctively apart along the first component variance (14.9%), while the second component variance (11.3%) spread the cultures without identifying any clear pattern of clustering (Fig. 2d).

Based on global gene expression profiling, the clustering of the GSC cultures differed from the clustering according to drug sensitivity, as the two most sensitive cultures clustered separately. We found more similarities in the gene expression between cultures categorized as moderate and least sensitive (T1456, T1454, T1548) than related to their overall drug sensitivity (Additional file 4). Further exploring selected gene panels involved in general drug resistance, drug metabolism, GSC related, and glioblastoma related genes did not identify any shared expression pattern of the most sensitive cultures compared to the others (Additional file 5).

The overall drug sensitivity only explained a small proportion of the variance, suggesting that tumors can be grouped into a few subtypes. As 81% of the drugs with a DSS ≥10 also displayed ∆DSS ≥10 among all cultures, we explored how the heterogeneity in the sensitivity to anticancer drugs distributed across different mechanistic classes and molecular targets. The 115 drugs with a DSS ≥10 in any GSC culture represented a wide range of drug classes, including apoptotic modulators, conventional chemotherapies and inhibitors of histone deacetylases, heat shock proteins, proteasomes and different kinases. Across all classes and molecular targets, the distribution of drug sensitivities largely displayed a continuum from insensitive to the most sensitive tumor (Fig. 3).

To explore whether the GSC model system preserves the individual biological consistency of drug sensitivities, we categorized drug sensitivity patterns based on the specific molecular target within a class of drugs (e.g., MEK1/2 inhibitors in the kinase inhibitor class). We found a clear pattern in which drugs with a specific target displayed the highest efficacy in the same tumor. For instance, among MEK1/2 inhibitors with a DSS ≥10 (n = 5) in any GSC culture, T1550 was the most sensitive culture to four of five MEK1/2 inhibitors (and the 2nd most sensitive to the final inhibitor). Correlation matrices displayed that the average (±standard deviation) ranked correlation of the sensitivity to MEK1/2 inhibitors was 0.61 (±0.18) (Fig. 3). Similarly, the GSC cultures most resistant to a specific class of drug displayed a clear pattern of broad resistance to all drugs targeting the same specific molecular target. While being the most sensitive to MEK1/2 inhibitors, T1550 was the most resistant culture to CDK inhibitors (n = 5). The correlation matrices displayed that the average correlation of sensitivity to CDK inhibitors was 0.82 (±0.11) (Fig. 3). This consistency of individual drug sensitivity and resistance patterns was found across all major classes within the drug collection (Fig. 3). This demonstrated that individual biological traits involved in drug sensitivities are preserved and consistent in patient-derived GSC cultures and display individual uniqueness. In the DSRT, none of the GSC cultures displayed sensitivity to the standard-of-care, temozolomide (TMZ, Additional file 3).

The heterogeneity of drug sensitivity patterns in individual GSC cultures demonstrated that DSRT could uncover patient-specific vulnerabilities and potential treatment options for functional precision medicine. However, for DSRT to guide decision-making in patient treatment, we investigated the manual reproducibility of selected compounds in an independent laboratory performed by different personnel. To obtain a closer description of the biologically relevant concentration range, we performed a narrower 5-point concentration range and defined reproducibility by the ability to capture the inflection range with similar levels of EC₅₀-calculation and maximal inhibition. The independent validation confirmed the reproducibility by quantifying EC₅₀ in similar low molar concentrations and reaching levels of maximal inhibition in different drugs across different tumors (Additional file 6).

As the drug sensitivity and resistance patterns were linked to drug classes and molecular targets, we stratified the GSC cultures according to similar drug sensitivity patterns. For the stratification into patient-specific drug sensitivity for any given drug, we calculated the differential response in an individual culture from the average response in all GSC cultures. Thus, we quantified each drug response in each individual culture as either increased (+) or decreased (−), defining this as the selective DSS (sDSS) (Additional file 7). Correspondence analysis of the sDSS to all drugs clustered the cultures according to the overall sensitivity along the first component variance (19.1%), while the second component variance (12.8%) clustered the cultures based on the similarities in the sensitivity and resistance patterns (Additional file 7). Unsupervised hierarchical clustering revealed that the relationships among similar drug sensitivity patterns were based on the mechanistic target (Fig. 4, Additional files 8 and 9). The two most sensitive cultures were of the proneural subtype; however, in the moderate to least sensitive tumors, the proneural and mesenchymal subtypes were evenly interspersed (Fig. 4). The MGMT promoter methylation of the parent tumor status was not concordant with the clustering as the two most sensitive tumors and two of the four least sensitive tumors were MGMT promoter methylated.

To comprehend the overall heterogeneity in drug sensitivities in the entire culture cohort, we calculated the enrichment of drugs with the same modes of action in individual cultures according to the ratio of observed versus expected (O/E, if expected number of drugs was < 1, the value was set to 1) (Fig. 5a). By selecting drugs that had at least moderate efficacy (DSS ≥10) increased patient-specificity (sDSS ≥3) and O/E ≥ 3 in individual cultures, we found eight different drug categories of various molecular targets to be enriched in the treatment-naïve GSC cultures (Fig. 5b). The stratification into patient-specific responses identified the GSC cultures with the highest vulnerability to any given drug or class of drug. The dose-response curves of drugs that have been investigated in clinical trials of GBM demonstrated the existence of both resistant and sensitive GSC cultures in the treatment-naïve disease (Fig. 5c). Similarly, drugs from various categories currently recruiting patients for trials in GBM displayed the same pattern including both existing resistant and sensitive GSC cultures in a heterogeneous GBM population (Fig. 5c).