Sensitivity to BUB1B inhibition defines an alternative classification of glioblastoma

Abstract

Glioblastoma multiforme (GBM) remains a mainly incurable disease in desperate need of more effective treatments. In this study, we develop evidence that the mitotic spindle checkpoint molecule BUB1B may offer a predictive marker for aggressiveness and effective drug response. A subset of GBM tumor isolates require BUB1B to suppress lethal kinetochore-microtubule attachment defects. Using gene expression data from GBM stem-like cells, astrocytes and neural progenitor cells that are sensitive or resistant to BUB1B inhibition, we created a computational framework to predict sensitivity to BUB1B inhibition. Applying this framework to tumor expression data from patients, we stratified tumors into BUB1B sensitive (BUB1B^S) or BUB1B resistant (BUB1B^R) subtypes. Through this effort, we found that BUB1B^S patients have a significantly worse prognosis regardless of tumor development subtype (i.e., classical, mesenchymal, neural, proneural). Functional genomic profiling of BUB1B^R versus BUB1B^S isolates revealed a differential reliance of genes enriched in the BUB1B^S classifier, including those involved in mitotic cell cycle, microtubule organization and chromosome segregation. By comparing drug sensitivity profiles, we predicted BUB1B^S cells to be more sensitive to type I and II topoisomerase inhibitors, Raf inhibitors and other drugs, and experimentally validated some of these predictions. Taken together, the results show that our BUB1B^R/S classification of GBM tumors can predict clinical course and sensitivity to drug treatment.

Introduction

Glioblastoma multiforme (GBM) is the most aggressive common form of brain cancer in adults. GBM is characterized by poor survival (i.e. overall median survival time is one year (1)), remarkably high tumors heterogeneity (2), and lack of effective therapies (3). Recent studies have revealed that GBM can be divided into distinct subtypes with prognostic relevance based on differences in gene expression (4,5), somatic mutations (5,6), DNA methylation (7), and DNA copy number (8). While genomic analyses have defined GBM subclasses based on transcriptional and epigenetic regulations, the subclasses have failed to suggest effective subclass-specific treatment strategies, with the exception of the subclass based on promoter methylation status of the DNA repair gene MGMT. In this case, a subset of GBM tumors with MGMT promoter methylated and transcriptionally down-regulated are more likely to benefit from the addition of temozolomide to radiotherapy (9). However, the majority of GBM patients show very little benefit from surgery, radiation, and temozolomide (i.e., standard of care therapy). Thus, to achieve better outcomes in the clinic, we need a better patient stratification and more in-depth understanding of the biology of these tumors.

Both adult and pediatric GBM tumors appear to be hierarchically organized suggestive of a cancer stem cell origin (10–13). Consistent with this notion, tumor-initiating GBM stem cells (GSCs) have recently been isolated that retain the development potential and specific genetic alterations found in the patient’s tumor (10,11,14,15). GSCs retain the specific genetic and epigenetic signatures found in the original tumor (10,11,15), and give rise to tumors with GBM patient-specific molecular signature and histologic features when implanted into the cortex of rodents (11,15). Importantly, treatments targeting GSCs may be more effective because GSCs are radio-resistant and chemo-resistant due to its preferential activation of the DNA damage response, which eventually results in tumor recurrence (16). Therefore, the use of patient derived GSC isolates can allow investigation of the molecular characteristics of subpopulation of tumors, and potentially develop more effective treatments.

Recently, we performed shRNA kinome screens in GSC isolates and non-neoplastic neural progenitor controls for genes required for GSC expansion (17). Combing these results with a GBM bionetwork created from patient molecular profiles, we identified BUB1B as the top GSC-specific hit. BUB1B encodes the highly conserved Bub1-like pseudo-kinase, BubR1, which possesses multiple functional domains implicated in mitotic checkpoint control, mitotic timing, and regulating KT-MT attachment (18). These include: N- and C-terminal KEN box domains required for Cdc20 binding and APC inhibition (19); a C-terminal kinase domain required for protein stability (20), and a GLEBS domain necessary for kinetochore localization during mitosis (21). While BUB1B is essential for mammalian development (22), its essential function is contained solely within the N-terminal KEN box (23), which enables BubR1 to act as a pseudo-substrate inhibitor of APC/C^Cdc20 during G2 and pre-anaphase mitosis, preventing premature anaphase onset (23). In contrast, we observed that in ~60% of GSCs, Ras-transformed cells, and HeLa cells its GLEBs domain becomes essential for viability to promote kinetochore-microtubule attachment (17). Mechanistic experiments demonstrated that oncogenic Ras signaling triggers alterations in kinetochore regulation resulting in added GLEBs domain requirement and the primary reason we observe differentially sensitivity to BUB1B knockdown (17,24).

BUB1B-inhibition sensitive (BUB1B^S) cells invariably have shorter metaphase inter-kinetochore distances (IKDs), or shorter average distances between sister kinetochores during mitosis when stable end-on microtubule attachments have formed (17,24). This serves as an indirect measure of the pulling forces generated by dynamic microtubules bound to kinetochores, such that stronger attachments lead to longer IKDs and weaker attachments produce shorter IKDs (24). Although IKDs are reliable predictors of BUB1B^R/S and in theory could be used to predict tumor sensitivity to BUB1B inhibition, in practice, taking IKD measurements is laborious and time consuming, requiring confocal microscopic z-sectioning of mitotic cells, and unlikely to be useful to "type" tumor samples. Here, we instead used gene expression signatures associated with BUB1B^S and BUB1B^R cells to create a computational framework for predicting BUB1B^S/R status for GSCs, which examines differential expression of 838 genes. We then applied our classifier to existing GBM patient tumor data, revealing that BUB1B^S GBM tumor patients have significantly worse prognoses, regardless of their molecular subtype. Examination of results from genome-wide RNAi screens in BUB1B^S/R isolates revealed differential reliance of genes enriched in mitotic cell cycle, and chromosome segregation, and suggested potential therapeutic treatments of BUB1B^S cells. Applying the classifier to drug sensitivity profiles confirmed BUB1B^S cells to be more sensitive to type I and II topoisomerase inhibitors as well as histone deacetylase (HDAC) inhibitors, and additionally uncovered BUB1B^S cells to be more sensitive to Raf inhibitors. Taken together our results suggest that BUB1B^S/R classifier is useful for predicting tumor aggressiveness and therapeutic responses of GBM tumors in additional to their sensitivity to BUB1B-GLEBs domain inhibition.

Materials and Methods

RNA-seq data

We collected 18 GSC cell lines that have been previously published, one neuronal stem cell (NSC), the normal human astrocyte cell line (i.e. NHA, STEMCELL Technologies), and RasV12 transformed NHA (Russell Pieper, University of California San Francisco, San Francisco, CA) (References in Supplementary Table S1). Three biological replicates for each cell line were sequenced using Illumina platform. 20 million reads per sample on average were mapped to transcripts. We removed one GSC cell line (i.e. G25) from our further analysis due to its low number of mapped sequences for all three replicates (Supplementary Table S1). Sequencing data can be accessed at NCBI Gene Expression Omnibus (GEO) under GSE89623. In addition, we collected 9 GBM tumor sample from Mount Sinai Hospital September 2014 through November 2015 and derived GSC isolates from each tumor sample. The tumor tissue samples and derived GSCs were sequenced using Illumina HiSeq2500. On average 42 million reads per sample were mapped to transcripts. Sequencing data can be accessed at GEO under GSE94874. The cell culture protocol and preprocessing of RNA-seq data is detailed in Supplementary Materials and Methods.

Prediction of BUB1B^S/R status

We first identified genes whose expression levels were significantly associated with BUB1B^S/R status. We designed a multiple linear model to explain the variation in mRNA expression level for each gene in terms of BUB1B^S/R status and cell types as covariate. The associated genes were defined by their significance level of the coefficients for BUB1B^S/R status. Based on these genes associated with BUB1B^S/R status, we trained an elastic net classifier, a regularization method to prevent overfitting problems in a generalized linear model (25). Then, the elastic net classifier was applied to gene expression profiles of new samples to predict BUB1B^S/R status. See Supplementary Materials and Methods for details.

Extracting molecular profiles of GBM tumor cells from bulk tumor tissue expression

We expected that molecular profiles of GSCs and GBM tumor cells were similar. Thus, we extracted molecular profile of GBM tumor cells from expression profiles of bulk tumor samples by following three steps. Frist, we determined gene expression profile of each normal brain cell type and GSC. The single cell expression levels of 6 different brain cell types (26) and our GSC data were merged, normalized (27), and averaged to define the profile of each cell type. Secondly, we deconvoluted bulk tissue expression data by using CIBERSORT (https://cibersort.stanford.edu/) with the expression levels of normal brain cells and GSC, resulting in relative proportion of 6 normal cell types and GSC. Finally, we extracted molecular profile of GBM tumor cells from the measurements of bulk tumor samples based on relative proportion and gene expression profile of each cell type. The extracted molecular profiles of GBM tumor cells were used to predict BUB1B^S/R status. See details in Supplementary Materials and Methods.

Association of the BUB1B^S/R status with survival rate

We characterized BUB1B^S/R subtypes in term of the clinical survival rate for three data sets including TCGA (28), Gravendeel et al. data set (29) and Joo et al. data set (30) (see Supplementary Materials and Methods for download and preprocessing of datasets). First, the overall survival was associated with the predictive values based on the elastic net classifier by fitting cox proportional hazards model (31). Additionally, the Kaplan-Meier curves for BUB1B^S and BUB1B^R subtypes that are defined as samples with the highest and lowest 25% predicted values were compared (32). The test of equality for survival distributions was performed using the log-rank test.

Identification of genes required for the expansion of BUB1B^S/R GSCs

We performed genome wide shRNA screen and Bar-code array analysis for three GSC cells and one NSC cell (CB660) as previously described (33). We detected each isolate’s candidate targets as shRNAs under represented in GSCs relative to NSCs. Based on these candidate targets, we identified genes only required for expansion of BUB1B^S and BUB1B^R GSCs. These genes were used to define BUB1B^S/R lethal sub-networks. See Supplementary Materials and Methods for details.

Drug discovery by using BUB1B^S lethal sub-networks

With genes in the two large BUB1B^S specific lethal sub-networks (see Supplementary Materials and Methods for details), we determined whether these genes were up-regulated or down-regulated in BUB1B^S GSCs by comparing their gene expression levels between BUB1B^S GSCs and BUB1B^R GSCs (Supplementary Fig. S7). Then, we investigated whether these genes were perturbed by drug treatments by using the tool (http://www.broadinstitute.org/cmap/) from the connectivity map (34). A negative enrichment score from the output of the tool represents the treatment of the drug might inhibit the BUB1B^S lethal genes (therapeutic to BUB1B^S group), resulting in selectively killing of BUB1B^S GSCs.

Association between drug sensitivity and BUB1B^S/R status in glioma cell lines

We predicted BUB1B^S/R status of glioma cell lines from Cancer Cell Line Encyclopedia (CCLE) and Cancer Genome Project (CGP) database (35,36). We collected gene expression levels of 59 and 43 glioma cell lines from CCLE and CGP database, respectively, and applied our classifier on gene expression profiles to predict BUB1B^S/R status. Next, we explored the drug sensitivity measurements of cell lines to identify drug that has different sensitivity on BUB1B^S/R subtypes. We used the concentration at which the drug response reached an absolute inhibition of 50% (IC50) and an ‘activity area’ to capture simultaneously the efficacy and potency of a drug (36). See details in Supplementary Materials and Methods. For each measurement, we calculated the correlation between drug sensitivity measurements and predictive values that determine BUB1B^S/R status across glioma cell lines to identify drugs that have different sensitivities between BUB1B^S and BUB1B^R cell lines.

Measure drug sensitivity of Etoposide and Irinotecan in GSCs

The effects of Etoposide and Irinotecan on GSC isolates were determined by treating them with various concentrations of each drug for incubation period 72h at concentrations ranging from 1µM to 4mM for Etoposide and 0.01µM to 1mM for Irinotecan. See details in Supplementary Materials and Methods.

Results

Construct a BUB1B^R/S status classifier for GSC isolates based on gene expression profiles

Based on our previous results, where we demonstrated that GBM tumor isolates are either BUB1B-inhibition sensitive (BUB1B^S) or resistant (BUB1B^R) (17) (Fig. 1A), we hypothesized that BUB1B^S/R could be captured by genome-wide gene expression analysis. To this end, we analyzed the RNA-seq profiles of 18 GSCs as well as astrocyte and neural progenitor cells, classifying them by BUB1B^R/S status. BUB1B^R/S status was determined by sensitivity to LV-shBUB1B and/or measurement of inter-kinetochore distance at metaphase (17,24) (Supplementary Table S1 and Supplementary Fig. S1). This analysis revealed 9 BUB1B^S isolates, 4 BUB1B^R isolates, and 5 isolates that had not been experimentally tested for BUB1B^R/S (Fig. 1B, one isolate was removed from further analysis due to low sequencing quality). Additionally, the neural progenitors and the normal human astrocyte cell line (i.e. NHA) were BUB1B^R, on the other hand, the RAS-treated astrocyte was BUB1B^S (Fig. 1B). To identify a BUB1B^R/S gene expression differences, we used a linear model to fit the variation in mRNA expression level for each gene in terms of BUB1B^R/S. We identified 838 genes whose expression is associated with BUB1B^S/R at false discovery rate (FDR) <1% corresponding to p<10^−3.9. Gene set enrichment analysis (37) of the 838 BUB1B^S/R-associated genes revealed that they were significantly enriched for cell cycle related pathways such as the mitotic cell cycle (Supplementary Fig. S2).

Fig. 1. The overview of BUB1B^S/R status prediction procedure.

(A) Determine BUB1B^S/R status. BUB1B^S/R status can be determined by measurement of inter-kinetochore distance (IKD) at metaphase. BUB1B^S and BUB1B^R subtype correspond to short and long IKD, respectively. (B–D) Prediction of BUB1B^S/R status by expression levels. (B) The initial BUB1B^S/R status for training the classifier. The red and blue indicated BUB1B^S subtype and BUB1B^R subtype, respectively. The white indicates unknown status. (C) The predicted values by applying the classifier to GSC isolates with known and unknown BUB1B^S/R status. The large predicted value indicates higher probability of being BUB1B^S subtype. G144 and 1406 were classified as BUB1B^R. 448T and 559T were classified as BUB1B^S. (D) Gene expression levels of GSCs used for prediction. (E) Validation of BUB1B^S/R status by measuring IKD. IKDs were measured for four GSC isolates for which we had not previously measured IKDs. The measured IKDs were compared to 0827, which have long IKDs, by t-test. The results show that G144 and 1406 cells did not differ from 0827 (i.e. not significant (n.s.)), while 448T and 559T were significantly shorter, consistent with our prediction.

We next developed an elastic net-based classifier (25) (see Methods and Materials for details) to predict sensitivity to BUB1B-inhibition using the 838 BUB1B^S/R–associated genes (Fig. 1C&1D). The cross-validation results indicate that the elastic net classifier can achieve low mean squared error (Supplementary Fig. S3). We applied the classifier to four GSC isolates of unknown BUB1B^R/S status, including G144, 1406, 448T, and 559T (i.e., white in Fig. 1B). The predictive values of new samples obtained by fitting elastic net classifier on its gene expression profiles were used to determine BUB1B^S/R (Fig. 1C&1D). We predicted that G144 and 1406 would be BUB1B^R, while 448T and 559T would be BUB1B^S. Consistent with these predictions, IKD measurements of metaphase cells confirmed that both G144 and 1406 have long IKDs (i.e., BUB1B^R), while 448T and 559T have short IKDs (i.e., BUB1B^S) (Fig. 1E). Thus, these results demonstrate that our classifier can accurately predict BUB1B^R/S status.

Application of the BUB1B^R/S classifier to GBM tumor data

We next tried to apply our classifier to GBM patient tumors to investigate the relationship between BUB1B^R/S status and GBM patient prognosis. However, tumor samples are complex mixtures of malignant and normal tissue cells; molecular signal might be averaged out between normal and cancer cells for tumor tissue. To investigate the relationship between tumor tissues and GSCs, we collected 9 pairs of GSCs and the corresponding tumor tissues from which they were derived, and measured gene expression levels. When considering the gene expression, GSCs, which were purely neoplastic, were clearly separated from the corresponding tumor tissues (Supplementary Fig. S4A). This indicates that molecular signature of tumor tissues likely contained features of both GBM tumor cells and non-tumor cells. The predicted value of the BUB1B^R/S status based on gene expression levels of tumor tissues and corresponding derived GSCs were not consistent (Supplementary Fig. S4B). A reasonable explanation is that the normal, non-transformed cells are predicted to be BUB1B^R (17) so that the prediction of the BUB1B^R/S status based on the mixed molecular signals of tumor tissue expression is confounded by proportion of normal cells in the tumor tissues. These results indicate needs of decomposition of the molecular signals of GBM tumor cells from normal brain cells in bulk tumor tissues in order to accurately predict the BUB1B^R/S status of tumor cells.

To extract the molecular signal of GBM tumor cells, we decomposed tumor tissue expression into expression of diverse cell types in the human brain (i.e. astrocytes, endothelial, microglia, neurons, oligodendrocytes, and Oligodendrocyte progenitor cells (OPCs)) (26) and our GSCs using CIBERSORT (38), and estimated relative fraction of diverse brain normal cell types and GMB tumor cell in each tissue sample (Supplementary Table S2) (see Methods and Materials for details). The relative fraction of GBM tumor cells varied from 19% through 55% (Supplementary Fig. S4C), and indeed, the predicted value of the BUB1B^R/S status based on tissue expression levels was highly associated with the predicted proportion of GBM tumor cells in the mixture (Supplementary Fig. S4C). The predicted cancer cell fractions were consistent with histological results. For example, the tumor tissue sample with the lowest GBM tumor cell fractions (i.e. 9217A), the histological staining of tumor tissue indicated lower percentage of cancer cells than others (Supplementary Fig. S4D). Based on the relative fraction and molecular profile of each cell type, we extracted the molecular signals of GBM tumor cells from bulk tumor tissue expression levels (see Methods and Materials for details). The deconvoluted expression profiles of GBM tumor cells were used to predict BUB1B^S/R, and resulted in the similar order of predictive values with those based on GSC expression levels (Supplementary Fig. S4E), Taken together, the deconvolution of bulk tumor tissue expression levels to extract molecular signatures of GBM tumor cells is the essential step to predict accurate BUB1B^S/R based on tumor tissue expression levels.

Prognostic value of BUB1B^R/S subtypes

We characterized BUB1B^S/R subtypes in term of the clinical survival rate for three data sets for GBM: the Cancer Genome Atlas (TCGA) (28,39), Gravendeel et al. data set (29), and Joo et al. data set (30). We applied the deconvolution procedure to these tissue gene expression datasets, and predicted BUB1B^S/R status based on extracted molecular profiles of GBM tumor cells. The overall survival was significantly associated with the predicted sensitivity value based on the elastic net classifier by fitting cox proportional hazards model (31) for TCGA data sets based all three platforms (e.g. Agilent platform p-value < 0.040 based on Cox regression) (Table 1). The BUB1B^S subtype showed worse survival rate than the BUB1B^R subtype (Fig. 2A), and the difference assessed by log-rank test (32) was significant. The application to the Gravendeel et al. data set and Joo et al. data set yielded similar results that BUB1B^S/R status was associated with survival rate (e.g. Gravendeel et al. p-value < 0.0097 and Joo et al. p-value < 0.003 based on Cox regression) (Fig. 2A and Table 1) with poor prognosis for patients in the BUB1B^S subtype. Note that the predicted sensitivity scores based on tissue expression levels without deconvolution were not consistently associated with survival rate (Table 1), underscoring the importance of extracting expression signatures of GBM tumor cells before applying our classification method. To be able to perform more robust exploration of the relationship of BUB1B^S/R subtype to survival rate, we incorporated the other clinical information into predicted sensitivity values. Both gender and the methylation status of MGMT promoter that is prognostic to temozolomide treatment (9) were found not to be significantly associated with predicted sensitivity values (Supplementary Fig. S5A) for any data sets, on the other hand, age at diagnosis was found to be significantly associated with survival rates for some of datasets (Supplementary Fig. S5B). Even though the inclusion of age as a covariate decreased the significance level of predicted sensitivity value to BUB1B-inhibition in cox regression model for most data sets, the BUB1B^S/R subtypes were still significantly associated with survival rate for most datasets (Supplementary Table S3).

Table 1.

Association of BUB1B^S/R subtype with survival rate for TCGA data(4), Gravendeel et al.,(29), and Joo et al., data set(30).

Study	Platform	# samples	p-value by Cox1		Coefficient by Cox2
			GBM tumor cells3	Tissue4	GBM tumor cells	Tissue
TCGA	Agilent	574	0.0186	0.2903	0.1360	−0.0382
	Affymetrix	529	0.0406	0.0286	0.1096	0.0541
	RNA-seq	154	0.0500	0.0747	0.0945	0.0387
Gravendeel et al.	Affymetrix	159	0.0097	0.0638	0.1761	0.0972
Joo et al.	Affymetrix	58	0.0035	0.0064	0.6480	0.1266

¹p-value and

²coefficients by fitting cox proportional hazards model (31). The sensitivity to BUB1B-inhibition was predicted by

³inferred molecular profiles of GBM tumor cells or

⁴bulk tissue gene expression levels.

Fig. 2. The BUB1B^S/R status within tumor tissue data sets.

(A) Association between the BUB1B^S/R status and patient survival. We applied our classifier to three independent cohort data sets and characterized association with survival rate. The survival rate was significantly associated with BUB1B^S/R status for all cohort data sets. The BUB1B^S patients has worse survival rate. (B) The invasiveness of patient’s and xenograft tumors was associated with BUB1B^S/R status. X-axis indicates the invasiveness of tumors and y-axis indicates the predicted values from our classifier. The invasiveness of patients’ tumors was defined as: 0=no invasion, 1=distance of invasion < 2× diameter of tumor mass, 2=2× diameter of tumor mass < distance of invasion <3× diameter of tumor mass. The invasiveness of xenograft’s tumor was defined as Y=distance of infiltration > 2× diameter of main mass in the paraffin sections. (C) Comparison our classification to the conventional classification. For each conventional class, we calculated the number of BUB1B^S/R subtypes. The y-axis indicates the fraction of BUB1B^S (red) or BUB1B^R (blue) tumors to BUB1B-inhibition.

We further examined the other clinical phenotypes associated with BUB1B^S. In particular, we analyzed Orthotopic glioblastoma xenograft data from 58 GBM patients (30) with expression profiles from GBM patients’ primary tumors and parallel in vivo xenograft tumors. We examined the invasiveness of GBM patients as well as xenograft inherited from these GBM patients, and found suggestive association between our predicted sensitivity values and invasiveness of patient’s tumor as well as the corresponding xenograft (Fig. 2B), indicating that patients with BUB1B^S tumors were likely to have more invasive disease, resulting in worse overall survival.

Comparison with the existing molecular subtyping method for GBM

We compared the existing molecular GBM tumor classifications (5,7) to our classification based on BUB1B^S/R. First, we compared the 838 genes associated with BUB1B^S/R to 840 genes that were used to determine the conventional subtypes, including Classical, Mesenchymal, Neural, and Proneural subtypes (5). Interestingly, only 35 genes were in common (p-value > 0.2), representing the uniqueness of our classification of GBM samples based on BUB1B^S/R.

In addition, we compared membership of our BUB1B^S/R subtypes to the previously identified molecular subtypes based on gene expression and CpG island methylation status including G-CIMP, Classical, Mesenchymal, Neural and Proneural subtypes (7). Interestingly, for TCGA data, Mesenchymal subtype was enriched for the samples predicted as a BUB1B^R subtype (Fig. 2C; p-values from Fisher exact test were 1.6×10⁻¹², 0.01, 0.31 for Affymetrix, Agilent, and RNAseq platform, respectively), on the other hand, Neuronal subtype was enriched for the samples predicted as a BUB1B^S subtype (Fig. 2C; p-values from Fisher exact test were 5.9×10⁻⁹, 5.7×10⁻⁹, 0.34 for Affymetrix, Agilent, and RNAseq platform, respectively). Other subtypes consisted of similar numbers of BUB1B^R/S classified samples.

We also compared our classification to the previously identified molecular subtypes for GBM patients of Gravendeel et al. data (29) and Joo et al. data (30). As there was no reported molecular subtype information associated with the data set, we classified the GBM samples into four subtypes (i.e. Classical, Mesenchymal, Neural and Proneural) based on the expression signatures previously defined (5). The result was similar to those from TCGA data: only Neural subtype (p-value from Fisher exact test are 7×10⁻⁴) and Mesenchymal subtype (p-value from Fisher exact test are 1.0×10⁻⁷) were enriched for samples classified as a BUB1B^S subtype and BUB1B^R subtype, respectively for Gravendeel et al. data set (Fig, 2C). Thus, the previously used molecular subtypes cannot be fully reflected in our classification based on BUB1B^S/R, demonstrating the uniqueness of our classification based on sensitivity to BUB1B-inhibition.

BUB1B^S tumor cells were differentially sensitive to perturbation of mitotic cell cycle, microtubule organization, and chromosome segregation gene networks

To further understand the relationship between the BUB1B^S versus BUB1B^R state in tumor cells, we analyzed the results from genome-wide RNAi screens in BUB1B^S and BUB1B^R GSC isolates, and neural progenitor isolates (33). This screening approach revealed 576 and 122 candidate targets required for expansion of BUB1B^R and BUB1B^S GSCs, respectively (see Materials and Methods for details). We prioritized these candidate therapeutic targets by identifying the biological pathways enriched for screen hits in each subtype, assuming that there are different sub-networks selectively lethal to each subtype. We projected these subtype specific screen hits onto a GBM network, constructed de novo from GBM samples of The Cancer Genome Atlas (TCGA) (4) by integrating gene expression and DNA copy number variation data (40,41). Examination of sub-networks in the GBM network revealed BUB1B^S and BUB1B^R-specific sub-networks, which were enriched for candidate lethal hits of BUB1B^S or BUB1B^R GSC isolates (Supplementary Fig. S6A&S6B), respectively. The projection of candidate lethal hits of BUB1B^R GSC isolates on GBM network failed to detect significantly large BUB1B^R specific sub-network (Supplementary Fig. S6B, see Supplementary Materials and Methods for details), on the other hand, we detected two significantly large BUB1B^S specific sub-network (Supplementary Fig. S6A and Fig. 3A). The two sub-networks that are specific to BUB1B^S cells (Fig. 3A), containing eleven screen hits (i.e. RACGAP1, POLE, CCNA2, MYBL2, RFWD3, ANXA2, KRR1, PRRP1R12A, NCL, and CPSF6), were significantly enriched for several biological processes related to mitotic cell cycle (p-value <1.3×10⁻¹¹), microtubule cytoskeleton organization (p-value < 8.9×10⁻⁶) and chromosome segregation (p-value < 3.4×10⁻⁶) (Fig. 3B). Additionally, the expression levels of these genes were mostly up-regulated for BUB1B^S GSCs compared to BUB1B^R GSCs (Supplementary Fig. S7), suggesting that BUB1B^S cells might experience the heightened mitotic stress and render the cell hypersensitive to perturbation of the mitotic machinery. Together, inhibition of these sub-networks will lead to selectively impair the viability of BUB1B^S GSCs. We focus on these selectively lethal sub-networks rather than the individual lethal gene, and further used the sub-networks to identify potential drugs that may be effective to BUB1B^S subtypes in the next section, because the perturbation of the sub-network tightly connected with the selectively lethal genes will be more efficient to kill cancer stem cells.

Fig. 3. BUB1B^S specific lethal sub-networks.

(A) The two largest sub-networks that are specifically lethal to BUB1B^S cells are shown. The red, blue, and yellow represent BUB1B^S lethal genes, BUB1B^R lethal genes, and cancer lethal genes, respectively. (B) The significant pathways enriched for genes within sub-networks specific to BUB1B^S cells.

Potential treatments effective for BUB1B^S tumors

We next wished to determine if we could use our BUB1B^R/S classifier as a predictor of sensitivity to existing drugs, since there are no existing therapeutics strategies for inhibiting BUB1B-GLEBS domain activity. To this end, we endeavored to systematically identify drugs that would selectively impair the viability of BUB1B^S cells using the Connectivity Map (CMAP), encompassing 7,000 genome-wide expression responses to 1309 different compounds (34). We analyzed the expression changes of genes in the sub-network specific to BUB1B^S cells (Fig. 3A) after the each drug’s treatment to identify the drug that inhibits genes within this sub-network. A negative enrichment score from connectivity map represents the treatment of the drug might inhibit the genes specifically lethal to BUB1B^S cells, resulting in selectively killing of BUB1B-inhibition sensitive GSCs. We identified several potential drugs, including MS-275, Etoposide, Camptothecin, Irinotecan, Thioridazine, and Azacitidine (Table 2). Interestingly, Etoposide targets DNA topoisomerase II whose disruption perturbs centromere organization leading to intense Bub1 on kinetochores (42), and both Camptothecin and Irinotecan target DNA topoisomerase I, whose treatment in GBM was beneficial (43). Also, Thioridazine, an antipsychotic drug, has been identified as an antiglioblastoma and anticancer stem cell agent (44,45).

Table 2.

The potential drugs that were predicted to be specifically cytotoxic to the BUB1B^S tumors, and their association with BUB1B^S subtypes in glioma cell lines based on CCLE or CGP database (35,36).

Drug name	Enrich ment1	p-value	FDR2	Target3	Tested in CCLE4	Tested in CGP5
GW-8510	−0.974	0	0	CDK2/3/5	-	-
Thioridazine	−0.562	0	0	Dopamine receptor	-	-
0175029-0000	−0.874	0.00002	0.0016		-	-
Camptothecin	−0.973	0.00006	0.0029	Topoisomerase I	-	0.783 (−0.04)
H-7 dihydrochloride	−0.922	0.00006	0.0029	PKC	-	-
Alsterpaullone	−0.969	0.00008	0.0032	CDK	-	-
Irinotecan	−0.964	0.00012	0.0041	Topoisomerase I	0.1002 (0.399)	-
MS-275	−0.988	0.00034	0.0103	HDAC	-	0.0084 (0.553)
Azacitidine	−0.907	0.0015	0.0397	DNA methyl transferase	-	-
Cloperastine	−0.701	0.00163	0.0397		-	-
Fluspirilene	−0.811	0.00251	0.0556		-	-
Etoposide	−0.803	0.00294	0.0593	Topoisomerase II	-	0.0096 (0.420)
Nordihydroguaiaretic acid	−0.449	0.00316	0.0593	mTORC1	-	-

¹Enrichment score calculated from Connectivity Map (34).

²False discovery rate.

³Targets of each drug.

⁴p-value indicating whether predictive score of cell lines in CCLE data were correlated with the drug sensitivity, and correlation in parenthesis.

⁵p-value and correlation based on CGP data. “−“ indicates no drugs were tested in each database.

To validate our candidate drugs that have specifically cytotoxic to BUB1B^S tumor cells, we analyzed cell line data sets, CCLE and CGP with gene expression profiles as well as drug sensitivity measurements (35,36). Application of our classifier to CCLE data yielded 12 and 14 BUB1B^R or BUB1B^S cell lines, respectively, out of 59 glioma cell lines (Supplementary Fig. S8A). For CGP data, 13 and 8 out of 43 glioma cell lines were classified into BUB1B^R or BUB1B^S subtypes, respectively (Supplementary Fig. S8B). Furthermore, we found 27 common cell lines within both CCLE and CGP data sets, and the predictive values of these common cell lines were similar (Fig. 4A) (correlation coefficient = 0.69 and p-value <5.7×10⁻⁵), resulting in consistent classification. In summary, we found that 5 cell lines are consistently classified into BUB1B^R subtype, and 4 are classified into BUB1B^S subtype in both data sets.

Fig. 4. Validation of drug sensitivity in BUB1B^S/R glioma cell lines.

(A) The predictive values of cell lines based on CCLE and CGP data sets (35,36) were shown. (B) Association between the predictive values and drug sensitivity for each drug. The yellow and red represents the association based on active area and IC50, respectively. −log(p-value) of correlation coefficients were shown. (C) Examples of drugs that were significant associated with the predictive values. The cell lines were sorted by the predictive values. The red and blue represents the cell lines classified as BUB1B^S and BUB1B^R subtypes, respectively. (D) IC50 for the cell lines from the previous studies (47–49). X-axis represents the predictive values from our classifier and y-axis represents the IC50 value. The red and blue line indicated the threshold for classification as BUB1B^S or BUB1B^R subtypes, respectively. The results from previous studies are consistent with our prediction.

We further investigated the drug sensitivity of cell lines from both CCLE and CGP data set. Based on the drug sensitivity measurements (i.e. the half maximal inhibitory concentration (IC₅₀) and activity area) for each glioma cell line and predicted BUB1B^S, we tested whether the two groups showed significantly distinct drug sensitivity. We measured the correlation between drug sensitivity measurements and the predictive value from the elastic net classifier. Encouragingly, 4 out of 13 potential drugs (Table 2) were tested in at least one of two data sets (CCLE or CGP) and the sensitivity of three out of 4 drugs was significantly associated with BUB1B^S status. We validated the BUB1B^S subtype showed more sensitive viability to Etoposide (p-value < 0.009) targeting topoisomerase II, MS-275 (p-value < 0.008) targeting histone deacetylase (HDAC) from CGP data set, and Irinotecan (p-value < 0.10) targeting topoisomerase I from CCLE data set (Fig. 4B&4C). Even though the sensitivity to Camptothecin targeting topoisomerase I was not significantly associated with BUB1B^S from CGP data, sensitivity to Topotecan and Irinotecan that both targeting topoisomerase I was different between BUB1B^S and BUB1B^R subtype from CCLE data (Fig. 4B&4C), suggesting the drug targeting topoisomerase I as the BUB1B^S subtypes specific drugs. This discrepancy can be explained by two potential reasons: i) the different drug dose usages of Camptothecin between CGP data and CMAP. CGP data uses 0.0003~0.1 µM concentration and CMAP uses 11.4µM concentration; ii) tissue specific effect. Only MCF7 and PC3 that are not glioma cell lines were tested from CMAP, on the other hand, we considered glioma cell lines for CGP data. Interestingly, the most glioma cell lines tested showed sensitive to Camptothecin compared to other drugs (Supplementary Fig. S9). Furthermore, we detected PLX4720 targeting BRAF as a potential drug that is more effective to BUB1B^S subtypes from both CCLE and CGP data sets. This suggests the potential importance of RTK-Ras signaling affecting KT function (46).

Additionally, we collected drug sensitivities in several glioma cell lines reported in published studies (Fig. 4D). We compared predictive scores and IC50 value from each paper. Interestingly, even though CGP data did not confirm the significant sensitivity difference of Campothecin, the previous studies showed that BUB1B^R cell line (i.e. U87) is less sensitive to Campothecin than BUB1B^S cell line (i.e. DBTRG-05) (47) and two cell lines (i.e. U118 and SW1783) which have higher predictive scores (48), consistent with our prediction. Also, Metz et al. (49) confirmed that BUB1B^R cell line (i.e. U87) is less sensitive to Irinotecan than BUB1B^S cell line (i.e. U251), and Ciesielski et al. (48) indicates the different sensitivity to Etoposide, consistent with our prediction.

Prediction and validation of drug sensitivity of Irinotecan and Etoposide on GSC isolates

Among the predicted drugs that selectively lethal to BUB1B^S GSCs we selected two drugs (i.e. Etoposide and Irinotecan) due to their previous treatment in GBM patients (50,51). To validate the drugs that show different cytotoxicity to subtypes, we collected 9 additional GSC isolates. We measured genome-wide expression levels for these GSCs, and applied our classifier to gene expression profiles of these GSC isolates to predict the sensitivity to BUB1B-inhibition resulting 4 BUB1B^S and 4 BUB1B^R samples and one sample ranked in between (Fig. 5A).

Fig. 5. Prediction and validation of drug sensitivity of Irinotecan and Etoposide on GSC isolates.

(A) Prediction of BUB1B^S/R status for additional GSC isolates (yellow), resulting in 4 BUB1B^S, 4 BUB1B^R, and one non-classified samples. (B) The exposure to Etoposide reduced cell growth of 9217B (BUB1B^S GSC) compared to that of 10647B (BUB1B^R GSC). The IC50 of Etoposide of 9217B and 10647B were 144µM and 4.857mM, respectively. (C) For treatment of Irinotecan, 9217B (BUB1B^S GSC) and and 9260B (ranked between BUB1B^S and BUB1B^R) showed more reduced cell growth compared to 10647B (BUB1B^R GSC).

These GSC isolates were treated with Etoposide and Irinotecan to determine their cytotoxicity (Fig. 5B&5C). As shown in Fig. 5B (listed in Supplementary Table S4), the exposure to Etoposide reduced cell growth of 9217B compared to that of 10647B: the IC50s of Etoposide for 9217B and 10647B were 144µM and 4.857mM, respectively. We detected the same pattern for Irinotecan as Etoposide. The BUB1B-inhibition sensitive GSC--9217B and the GSC ranked between sensitive and resistant--9260B showed more reduced cell growth compared to BUB1B-inhibition resistant GSC--10647B (Fig. 5C). These results indicate that Etoposide and Irinotecan have selective cytotoxicity to BUB1B^S GSCs, consistent with our predictions.

Discussion

One of the promises of precision medicine is the improved ability to predict which treatments will work best for specific patients. Our results suggested that GSCs can be grouped into BUB1B^S or BUB1B^R subtypes. Our BUB1B^S/R associated gene signatures are not associated with previously identified gene lists of GBM classification, suggesting the uniqueness of our classification. Our signature was highly enriched for mitotic cell cycle, supporting the association with short IKDs and lethal KT-MT instability. We extracted GBM tumor cell profiles deconvoluted from bulk GBM tissue expression levels to identify new prognostic subtypes of GBM that may have been obscured previously by confounding normal and GBM tumor cells. Our BUB1B^S/R subtype based on sensitivity to BUB1B-inhibition was associated with overall survival, and BUB1B^S subtype had poor survival rate presumably because patients with these tumors tended to have more invasive disease. We identified lethal genes to GSCs of each subtype, and further suggested potential treatments that would be more effective to BUB1B^S subtype by connectivity map database accompanied with siRNA screens and GBM causal networks.

The gene signature associated with the BUB1B^S/R subtype was able to classify GSCs and GBM tumors into two subtypes with clinical relevance. Previous studies showed that patient cohorts treated with Etoposide had significant improvement in median overall survival (50), and the treatment of Bevacizumab with Irinotecan was effective in recurrent GBM patients (43). We predicted and validated the effective responses to Etoposide and Irinotecan of BUB1B^S GSCs. We expect that classification of BUB1B^S/R subtype would improve the effectiveness of these drugs to GBM patients. Whether the patients with BUB1B^S tumors responding better to these treatments needs to be studied further. Also in the in future high through drug screening experiments, we may better focus on screening compounds for only BUB1B^S cells instead of screening compounds for all GBM tumor cells. On the other hand we identified several drugs, such as MS-275, that show differential sensitivity depending on BUB1B^R/S classification in vitro cell lines. However, drugs including MS-275 show marginal to moderate anti-glioma effects in clinical trials (52,53). Due to the blood-brain barrier (BBB) that protects tumors from exposure to many active drugs, drugs that are very effective in solid tumors frequently fail to demonstrate activity in brain tumors (54). Alternative strategies targeting GBM tumors are needed. Recently, Food and Drug Administration (FDA) approved electric fields therapy referred to as tumor treating fields (TTFields) as a treatment of recurrent glioblastoma. TTFields target rapidly dividing tumor cells, resulting in aberrant mitotic exit of tumor cells and disruption of cells during mitosis (55). The molecular mechanism targeted by TTFields is similar to the one of classifiying BUB1B^R/S cells. It will be useful to investigate the molecular targets of TTFields by genome-wide transcriptomes and compare them with our genes used to classify BUB1B^R/S as well as to investigate whether BUB1B^S GBM patients have better prognosis with TTFields treatment.

Even though targeting GSCs is a promising therapeutic option to treat GBM, most GBM tumor expression profiles measured in bulk tissues were resulted from a mixture of tumor as well as normal cells. The classifier based on GSC expression levels was not suitable to apply to tissue expression levels directly because normal cells were BUB1B^R (17). Our proposed procedure was not only applicable to classify GSC isolates but also to classify GBM tissue into BUB1B^S and BUB1B^R subtypes, based on the extracted GBM tumor cell profiles from the tissue expression decomposition. Our analysis of tissue and GSC expression indicated that the proportion of tumor cells was variable across samples. The heterogeneous nature of the tumors from the same patient needs to be taken into account to develop therapies. One challenge will be how to treat the patients with a mixture of tumor cells from distinct BUB1B^S/R subtypes. Indeed, our results indicate that some of tumors from the same patients have potentially distinct BUB1B^S/R subtypes (e.g. both 9260B and 10647A were derived from two different tumors occurred at different time in the same patient). Also, when considering tumor tissue expression of GBM patients, we detected non-clearly classified samples whose predictive scores were close to zero, indicating the potential subpopulations with distinct BUB1B^S/R subtypes in the tumor tissues.

To assess tumor cell heterogeneity with regard to BUB1B^S/R subtypes we applied our approach to a publically available single cell sequencing dataset (2) which consists of single cell transcriptomic profiles of stemlike tumor-propagating glioblastoma cells (GBM6 and GBM8) and single cells isolated from five tumors as well as population transcriptomic profiles of five tumor tissues and stemlikes cells from three tumors as controls. For each transcriptomic profile of either single cell or population controls, we applied our BUB1B^R/S classifier and calculated predicted values from elastic net. First, we analyzed single cell sequencing data for GSCs (GBM6 and GBM8), and expected homogeneous transcriptional profiles as well predicted BUB1B^R/S scores. However, we found that the predicted BUB1B^R/S score was associated with the quality of single cell data with the predicted score anti-correlated with the number of non-expressed genes in each single cell sequencing data (Supplementary Fig. S10). If the cells of low sequencing quality data were removed, the predicted scores for GSCs were indeed homogeneous as expected (Supplementary Fig. S10). The result suggests that we need to filter out low quality single cell sequencing data that contains a large number of non-expressed genes in further data analyses. Second, we examined intra-tumor heterogeneity of single cells from 5 individual tumor tissue samples. Again, we found that the predicted BUB1B^R/S score was associated with the quality of single cell data with the predicted score anti-correlated with the number of non-expressed genes in each single cell sequencing data (left panels in Supplementary Fig. S11A–E), which was consistent with the trend observed in GSCs (Supplementary Fig. S10). After removing cells of low quality sequencing data, there was still transcriptional heterogeneity within cells from the same individual tumor tissue sample (left panels in Supplementary Fig. S11A–E). Third, we used the data set to further test our deconvolution method for predicting transcriptional feature of extracted GBM tumor cells from bulk tumor samples. The predicted BUB1B^R/S values for GSCs were more similar to the values for the extracted GBM tumor cells than to the values for bulk tumor samples (right panels in Supplementary Fig. S11A–E and S12), and the predicted scores of the extracted GBM tumor cells were close to the average of scores of single cells. For example, the tumor MGH31 was predicted as BUB1B^R based on bulk tumor tissue transcriptional profile while predicted as BUB1B^S based on both GSC and extracted GBM tumor cells transcriptional profiles. The results of single cell data indicate intratumoral heterogeneity of tumors. Thus, for future personalizing treatments, it is necessary to explore intratumoral heterogeneity with respect to the BUB1B^S/R subtypes.

In conclusion, our results suggest a new way to classify GBM tumors into two molecular subtypes, which have different prognosis and response to potential drug treatments. Further understanding of these subtypes may provide critical information to develop precision medicine for the deadliest cancer, GBM.