Abstract
Glioblastoma is the most common yet deadliest primary brain cancer1. The neural behavior of glioblastoma, including the formation of synaptic circuitry and tumor microtubes, is increasingly understood to be pivotal for disease manifestation2–8. Nonetheless, the few approved treatments for glioblastoma target its oncological nature, while its neural vulnerabilities remain incompletely mapped and clinically unexploited. Here, we systematically survey the neural molecular dependencies and cellular heterogeneity across glioblastoma patients and diverse model systems. In 27 patient tumour samples taken directly after surgery, we identify a spectrum of cancer cell morphologies indicative of poor prognosis and discover a set of repurposable neuroactive drugs with consistent anti-glioblastoma efficacy. Glioblastoma cells exhibit functional dependencies on highly expressed neuroactive drug targets, while interpretable molecular machine learning (COSTAR) reveals their downstream convergence on AP-1-driven tumour suppression. This drug-target connectivity signature is confirmed by highly accurate in silico drug screening on >1 million compounds using COSTAR, as well as by multi-omic profiling of drug-treated glioblastoma cells. Thus, Ca2+-driven AP-1 pathway induction represents a tumour-intrinsic vulnerability at the intersection of oncogenesis and neural activity-dependent signaling. Opportunities for clinical translation of this neural vulnerability are epitomized by the antidepressant Vortioxetine synergizing with current standard of care treatments in vivo. Together, the results presented here provide a mechanistic foundation and conceptual framework for the treatment of glioblastoma based on its neural origins.
Introduction
Glioblastoma is a deadly brain cancer with limited treatment options, shaped by heterogeneous developmental programs, genetic drivers, and tumor microenvironments 9–13. Despite an increasing understanding of this heterogeneity, the alkylating agent Temozolomide (TMZ), which prolongs median survival from 12 to 15 months, remains the only first-line drug approved for glioblastoma 14–16. Targeted therapies have been largely unsuccessful, in part due to the blood-brain barrier (BBB) limiting tumor accessibility, the presence of treatment-resistant glioblastoma stem cells (GSCs), and the lack of clinically predictive models 17–22. Systemically addressing these therapeutic roadblocks is an urgent clinical need.
An emerging paradigm is to consider glioblastoma in the context of the nervous system. Single-cell RNA sequencing (scRNA-Seq) and lineage tracing studies of glioblastoma have identified stemness signatures resembling neural development 7, 11, 12, 23–28. At the brain-tumor interface, synaptic integration of cancer cells into neural circuits regulates tumor growth 5, 6. Within the tumor, the extension of microtubes akin to neuronal protrusions promotes the formation of treatment-resistant invasive networks 2, 4, 8. Furthermore, modulating specific neurotransmitter or other secretory pathways in the tumor microenvironment impairs glioblastoma metabolism and survival 3, 29–31. Such neural aspects of glioblastoma offer new clinically-targetable vulnerabilities that could be exploited by repurposing approved “neuroactive” drugs (NADs). Neuroactive drugs can cross the BBB and are routinely prescribed for indications such as psychiatric or neurodegenerative diseases. Yet, as neuroactive drugs are originally developed to modulate the nervous system, their anti-cancer activity in glioblastoma patients is largely unknown.
Several key questions arise. First, how does neural intratumor heterogeneity across glioblastoma patients relate to disease course and response to therapy? Second, are there tumor-intrinsic neural vulnerabilities that are therapeutically targetable? Third, if so, which molecular dependencies and associated pathways are involved?
Here, we find morphological and neural stemness features across glioblastomas that relate to disease prognosis and drug response. Using pharmacoscopy (PCY), an ex vivo imaging platform 32–34 that captures patient and tumor complexity, we screen repurposable neuroactive drugs and identify a set with potent anti-glioblastoma activity. Top neuroactive drugs work consistently across patients and particularly target GSCs with neural morphologies associated with invasion and poor prognosis. These top drugs are validated across multiple glioblastoma model systems including patient-derived cultures and orthotopic xenograft mouse models. Integration of anti-glioblastoma response with multiplexed RNA-Seq, reverse genetic screening, and machine learning of drug-target networks reveals convergence of neuroactive drugs with anti-glioblastoma activity on AP-1 transcription factor and BTG tumor suppressor gene families. Using this drug-target connectivity signature, we predict and validate new candidate drugs across >1 million compounds in silico. The antidepressant Vortioxetine is the top PCY-hit and inducer of the AP-1/BTG signature, synergizing with both first- and second-line glioblastoma therapies in vivo. Our study identifies clinically-actionable neuroactive drugs for the treatment of glioblastoma converging on a gene regulatory network involved in cell proliferation and neural activity.
Results
Glioblastoma stem cell morphologies prognostic of poor outcome
Glioblastoma cells adopt unique cellular morphologies and stemness properties to integrate and survive in the brain 2, 4, 8, 12, 35. To comprehensively profile morphological and molecular heterogeneity within and between glioblastoma patients, we performed high-content confocal imaging of freshly dissociated surgical samples across 27 patients (prospective cohort; Fig. 1a,b and Extended Data Fig. 1a). In parallel, for all patients, somatic genetic alterations were determined by targeted next-generation sequencing (NGS) and scRNA-Seq was performed for a subset of patients (n=4 patients, Fig. 1a, Extended Data Fig. 1a-f and Supplementary Table 1,2). Glioblastoma cells were identified and placed along a neural stemness gradient by scRNA-Seq and immunofluorescence against the neural progenitor marker Nestin and the mature astrocytic marker S100B (Fig. 1c,d and Extended Data Fig. 1c-e). Immune cells present in the tumor microenvironment (TME) were identified by the pan-immune marker CD45 and T cell marker CD3. Marker negative cells included additional TME cell types observed by scRNA-Seq or tumor cells with low protein levels of Nestin and S100B (Fig. 1c,d and Extended Data Fig. 1e,f). Patient samples were highly heterogeneous in composition, with Nestin+ or S100B+ glioblastoma cells ranging from 4-39%, CD45+ immune cells from 1-82%, and all marker negative cells 13-84% (Fig. 1d).
Nestin+ cells are a proliferative GSC subpopulation at the apex of the neural stemness gradient shown to sustain long-term tumor growth 17, 19, 36–38. Upon visual inspection of Nestin+ GSCs, we observed cellular “morphotypes” distinguishable by the presence of tumor microtubes (TMs) and other morphological features such as cell size and shape (Fig. 1b,e-g). Using deep learning on 51,028 manually curated single-cell image crops across all patient samples, we trained a convolutional neural network (CNN) to classify Nestin+ cells into four main morphotypes (M1-M4) with 84.3% accuracy (Fig. 1d,f and Extended Data Fig.2a,b). Single-cell feature maps extracted from the CNN and nuclei segmentation revealed a continuum of M1-M3 morphotypes and a distinct cluster of small M4 cells (Fig. 1f,g). M1 (PTM; polygonal TM-containing) and M2 (ETM; elongated TM-containing) GSC morphotypes had varying distributions of TMs per cell (Fig. 1e), while TM-absent M3 (RB; round big) and M4 (RS; round small) morphotypes were characterized by their roundness yet differed in cell size (Fig. 1f,g and Extended Data Fig. 2c). Nestin expression was higher in the more complex M1-M3 morphotypes (Fig. 1g).
GSC morphotype composition varied dramatically across patients (M1: 1.3-31.4%; M2: 0.5-43.4%; M3: 1.9-38.6%; M4: 14-95.1%, Fig. 1d). Abundance of morphotypes was associated with cell proliferation, where the fraction of complex morphologies (M1-M3) was correlated with Ki67 levels measured by pathology in patient-matched tissue sections (Fig. 1h, top left). Furthermore, patient stratification based on morphotype composition was significantly associated with progression-free survival (PFS): lower baseline abundance of complex morphologies (M1, M3) and higher abundance of the small M4 morphotype was associated with better clinical outcome (n=17 patients with annotated PFS, Fig. 1h and Extended Data Fig. 2d). We did not observe a survival difference based solely on Ki67 levels, sex, or the abundance of marker-defined populations (Extended Data Fig. 2e); implying a unique prognostic contribution of the aggressive GSC morphologies. While MGMT promoter methylation status, a prognostic factor associated with response to alkylating agents, stratified patient survival, complex morphotype (M1-M3) abundance was independent of MGMT status (Fisher’s test, P=0.19, Extended Data Fig. 2e). Thus, while tumor microtubes have been associated with glioblastoma grade 2, we now show that GSCs adopt a spectrum of complex morphologies prognostic of poor clinical response also among glioblastomas.
Clinical concordance of ex vivo single-cell drug profiling
The DNA-alkylating chemotherapy TMZ remains the only approved first-line drug for glioblastoma. To determine if the chemosensitivity of highly heterogeneous patient samples relates to clinical response, we tested first- and second-line glioblastoma chemotherapies (n=3 drugs) in two independent glioblastoma cohorts: our main prospective cohort (n=27 patients) and a bio-banked retrospective cohort (n=18 patients, Fig.1a,i-l, Extended Data Fig. 3a, and Supplementary Table 1). We utilized pharmacoscopy (PCY), a clinically predictive ex vivo image-based screening platform developed in-house 32–34, 39–42. Pharmacoscopy measures the drug-induced relative reduction of any marker- or morphology-defined cell population in heterogeneous patient samples. Here, a positive PCY score indicates a reduction of Nestin+ or S100B+ glioblastoma cells relative to non-malignant TME cells, while a negative score indicates higher toxicity in the TME (Fig. 1i). Glioblastoma patients exhibited a wide spectrum of ex vivo responses to the first-line therapy Temozolomide (50-500µM; 2 days of incubation; Fig. 1j). We performed survival analysis on 16 of the 27 prospectively tested patients, limited to those patients with primary tumors and having received first-line TMZ treatment with evaluable PFS (n=16 patients, with n=3 ongoing responses). Prospectively stratifying patients by their ex vivo TMZ sensitivity suggested improved PFS for ex vivo responders, reaching significance at 100µM TMZ (p = 0.041; Fig. 1k and Extended Data Fig. 3b). We repeated this PCY-based response stratification analysis in a retrospective validation cohort with documented PFS and OS (n=18 patients). Here, ex vivo TMZ responders also had significantly longer PFS and OS, evidenced across all four tested TMZ concentrations (Fig. 1l). In contrast, TMZ sensitivity of CD45+ immune cells or of GSC morphotypes was not associated with clinical response. Ex vivo TMZ response further related to MGMT status, as we observed higher sensitivities to low dose TMZ (50µM) in patient tumors with methylated MGMT promoter (Extended Data Fig. 3c). Taken together, this suggests that ex vivo TMZ response of glioblastoma cells, as well as the baseline abundance of GSC morphotypes at diagnosis, are both novel prognostic factors for first-line response. The concurrence of pharmacoscopy with clinical response establishes the utility of image-based drug screening in patient material for the discovery of new glioblastoma therapies.
Therapeutically targeting neural tumor heterogeneity
The question arises whether it is possible to pharmacologically target the heterogeneous spectrum of malignant cells, both in terms of their molecular composition and morphological nature. To find repurposable drug candidates that target glioblastoma cells and aggressive GSC morphologies, we tested both neuroactive and oncology drug libraries across the prospective glioblastoma cohort (Fig. 2a-c, Extended Data Fig. 3d,e, and Supplementary Table 3). The neuroactive drug (NAD) library consisted of drugs approved for neurological diseases such as depression, epilepsy, and Alzheimer’s disease (n=67 drugs; 20µM). The oncology drug (ONCD) library included therapies approved for solid tumor indications such as cyclin-dependent kinase (CDK) and receptor tyrosine kinase (RTK) inhibitors (n=65 drugs; 1 and 10µM). We screened the NAD library across the whole prospective cohort (n=27 patients) and additionally tested the ONCD library in a subset of patients when enough surgical material was available (n=12 patients). As primary PCY-readouts, we measured on-target reduction of Nestin+ or S100B+ cells and GSC morphotype composition changes across patients (Fig. 2b,c and Extended Data Fig. 3d). Drug responses for other TME populations including immune cells and marker negative cells were also measured (Extended Data Fig. 3e).
Despite limited success of targeted cancer therapies for glioblastoma, opportunities for molecularly-guided precision oncology and drug repurposing remain. Among ONCDs, the top mean ranking PCY-hit across patients was Elesclomol, a BBB-penetrant mitochondrial oxidative stress-inducer (Fig. 2c and Extended Data Fig. 3d) 43, 44. Elesclomol represents a compelling therapeutic candidate amidst growing evidence highlighting the importance of mitochondrial dynamics in brain tumorigenesis 45–47. Among all tested drugs, Elesclomol response showed the strongest positive correlation with age, suggesting a potential link with mitochondrial aging and benefit for elderly patients (Fig. 2d). Ranked 2nd and 6th among the ONCDs were the two structurally-related multi-tyrosine kinase inhibitors, Sorafenib and Regorafenib. Regorafenib has recently been proposed to mediate a survival benefit over Lomustine in a randomized phase II clinical trial for patients with recurrent glioblastoma (Fig. 2c and Extended Data Fig. 3d) 48. Across our cohort, Regorafenib showed one of the strongest positive correlations to Ki67 levels, which could be leveraged to inform future clinical trials (Fig. 2e). Conversely, both of the tested mTOR inhibitors Temsirolimus and Everolimus showed no ex vivo efficacy, in line with negative results from randomized clinical trials (Extended Data Fig. 3d) 49–51. As a tumor’s mutational profile often influences drug response, we examined associations between ONCD responses and genetic alterations measured by targeted NGS across the cohort. The strongest pharmacogenetic association for ONCDs was the markedly increased ex vivo sensitivity of patients carrying p53 mutations to the CDK4/6 inhibitor Abemaciclib (Fig. 2f and Extended Data Fig. 3f-g). Our exploratory analysis of oncological drugs approved for other indications thereby identifies unevaluated treatment opportunities in glioblastoma.
Neuroactive drugs, in contrast, are developed to cross the blood-brain barrier and act upon the nervous system. Among NADs, we identified 15 of 67 drugs (22%) with consistent anti-glioblastoma activity across patients (top NADs; PCY-hits; mean PCY score > 0.03). Top NADs effectively reduced fractions of aggressive GSC morphologies in patient samples (M1-M3), and showed activity across patient-derived cultures (PDCs, n=3 lines) and established glioblastoma cell lines (n=4 lines, Fig. 2a,b). Dose-response relationships were further confirmed in glioblastoma cell lines for selected top NADs (n=9 drugs, Extended Data Fig. 4a-d). The top mean ranking PCY-hit was Vortioxetine, a safe and novel class of antidepressant with previously unknown anti-glioblastoma activity (Fig. 2b,c). Strikingly, Vortioxetine was more potent in patient samples with higher baseline abundance of aggressive GSC morphologies (Fig. 2g). Vortioxetine was also the most consistently effective drug across all tested glioblastoma model systems, with ex vivo efficacy surpassing Regorafenib (Fig. 2c). Other clinically attractive NADs included Paroxetine and Fluoxetine, both antidepressants of the selective serotonin reuptake inhibitor (SSRI) class, and Brexpiprazole, an atypical antipsychotic used for the treatment of schizophrenia. Brexpiprazole ex vivo response was related to sex, where men responded better than women (Fig. 2h and Extended Data Fig. 3g). Not all top NADs were clinically suitable considering the historically reported side-effects of the cannabinoid receptor blocker Rimonabant and the antipsychotic Zotepine, yet these could still inform future drug design studies.
Notable exceptions to the pan-patient and model consistency of top NADs were Apomorphine, Olanzapine, and Sertindole (Fig. 2b). Apomorphine showed high TME toxicity in a subset of patient samples and diminished efficacy in PDCs and glioblastoma-initiating spheroid cell lines (ZH-161, ZH-562); Olanzapine showed diminished activity against aggressive GSC morphologies, PDCs, and cell lines; and Sertindole response was highly variable among patients, despite its potency in glioblastoma models. Patient variability in ex vivo Sertindole response was related to the loss of FGFR2 (copy number variation; CNV loss), representing the most significant pharmacogenetic NAD association (Fig. 2i and Extended Data Fig. 3f).
Here, by comprehensively screening across heterogeneous patients and model systems, we identify a set of repurposable neuroactive drugs that effectively target the neural heterogeneity of glioblastoma cells. The consistency of the anti-glioblastoma efficacy of these neuroactive drugs in diverse model systems even in the absence of the TME suggests that they are targeting tumor-intrinsic vulnerabilities.
Divergent functional dependencies on neuroactive drug targets
The multitude of neuroactive drugs with anti-glioblastoma activity was unexpected, prompting the question as to whether there could be shared underlying mechanisms. Top NADs represented diverse drug classes without significant enrichment, indicating that canonical mode-of-action did not explain drug efficacy (Fig. 3a). This is in line with a previous screen of neurochemical compounds in patient-derived stem cell lines, which found various neurochemical classes represented among their hits 29. Among our tested serotonin and dopamine pathway modulators, for example, only 4 out of 11 antidepressants (36%) and 6 out of 16 antipsychotics (38%) exhibited anti-glioblastoma activity in primary patient samples on average (Extended Data Fig. 4e). Such drug classifications, however, simplify the polypharmacological drug-target profiles of neuroactive drugs. The majority of NADs act on multiple primary target genes (PTGs). These include ion channels, GPCRs, and enzymes that modulate neurotransmission in the central nervous system, whose expression remains a largely unexplored dimension of glioblastoma heterogeneity. Dependency on neuroactive PTGs with high lineage specificity and consistent expression across patients could explain the activity of top NADs.
We first set out to determine the expression of NAD PTGs by scRNA-Seq. We integrated scRNA-Seq data from our glioblastoma patient cohort (n=4 patients) with two published data sets (Neftel et al.,n=9 patients; Yu et al., n=9 patients, Fig. 3b,c and Extended Data Fig. 5a,b) 12, 28. Collectively, we profiled 22 glioblastoma patient single-cell transcriptomes across 25,510 cells for their expression of PTGs with biochemical NAD-interactions reported in the Drug Targets Commons database (DTC) 52. Among these PTGs, certain classes of ion channels and GPCRs were enriched in neural lineage cells (e.g. potassium, glutamate, and cannabinoid), while other classes showed broader expression patterns (e.g. calcium, adrenaline, Extended Data Fig. 5a). To quantify PTG expression across cell types and patients, we defined neural- and patient-specificity scores (NS and PS; Fig. 3b, Extended Data Fig. 5b and Methods). For detected genes, a higher NS indicates relative enrichment in glioblastoma cells (range -1 to 1) and a higher PS (range 0 to 1) indicates more patient-specific expression, while both scores will be close to zero for low-abundance genes. We first benchmarked the neural- and patient-specificity scores against the whole transcriptome for each dataset, including key glioblastoma (Nestin, S100B, CD133, SOX2) and immune marker genes (CD45, CD3, CD14) that were on expected opposites of the NS spectrum (Extended Data Fig. 5b and Supplementary Table 4). Among PTGs, ion channels and receptors with high neural-specificity included the calcium signaling modulator SIGMAR1, glutamatergic AMPA receptor subunit GRIA2, and cannabinoid receptor CNR1 (Fig. 3c). Patient-specificity for neurological receptors SIGMAR1 and CNR1 were on average 1.7 to 3-fold lower than for oncogenic RTKs EGFR and PDGFRA, despite similar detection levels. Thus, we find abundant and consistent pan-patient expression of neuroactive drug targets on glioblastoma cells, confirmed across independent cohorts.
We next tested genetic dependencies on these NAD PTGs by performing an siRNA screen in LN-229 glioblastoma cells (n=59 genes, Fig. 3d and Extended Data Fig. 5c,d). LN-229 cells were confirmed to have patient-comparable PTG expression as well as neuroactive drug responses (Fig. 2b, Fig. 3d and Extended Data Fig. 5d). A significant decrease in cell viability was observed upon knockdown of 9 PTGs, including ADRA2B, SIGMAR1, DRD1, HTR3A, and MC3R (Fig. 3d and Extended Data Fig. 5c,d). Of these, ADRA2B and MC3R expression levels stratified glioblastoma patient survival in The Cancer Genome Atlas (TCGA) cohort (Extended Data Fig. 5e). However, these primary drug targets representing genetic dependencies were all annotated in DTC to predominantly interact with drugs that showed no activity by pharmacoscopy. Inversely, top neuroactive drugs had many annotated targets not representing genetic dependencies. For example, based on DTC, only 6 of the 17 NADs interacting with ADRA2B were PCY-hits, and only 1 out of 11 NADs interacting with HTR3A was a PCY-hit (Fig. 3e). Therefore, while presenting novel neural vulnerabilities, no PTG dependency uniquely explained the majority of our top neuroactive drugs.
Anti-glioblastoma activity explained by drug-target convergence
Despite their chemical and primary target diversity, our top NADs may still converge upon common downstream signaling pathways. To test this, we developed an interpretable machine learning approach called COSTAR: convergence of secondary drug targets analyzed by regularized regression. COSTAR is designed to identify the minimal drug-target connectivity signature predictive of efficacy.
We expanded the drug-target search space to include PTGs with any bioactivity annotated by DTC, termed extended PTGs (ePTGs). Secondary target genes (STGs) downstream of ePTGs were subsequently mapped by high-confidence protein-protein interactions annotated in the STRING database (Fig. 3f). This resulted in a drug-target connectivity map, or “COSTAR constellation”, of all DTC-annotated drugs in our NAD and ONCD libraries (n=127 of 132 tested drugs) with 975 extended primary targets, 10,573 secondary targets, and 114,517 network edges (Fig. 3f). Using logistic LASSO regression, we trained a multi-linear model that identifies the minimal set of STGs that maximally discriminates PCY-hit drugs (n=30; top-15 drugs from both NADs and ONCDs) from PCY-negative drugs (n=97; all other tested drugs) in a cross-validation setting (Fig. 3f,g Extended Data Figure 6a, and Methods). Thereby, COSTAR converged upon the most simplistic connectivity signature that was predictive of anti-glioblastoma drug efficacy (Fig. 3h and Extended Data Fig. 6a,b). Encouragingly, COSTAR identified a signature that classified the 127 drugs in our training data with 92.1% accuracy, correctly predicting 20/30 PCY-hits and 96/97 negative drugs (Fig. 3g).
The COSTAR connectivity signature predominantly linked PCY-hit NADs to the secondary target BTG2 through JUN/TP53 ePTGs (Fig. 3h and Extended Data Figure 6b). BTG2 and TP53 are both tumor suppressors that control cell cycle and differentiation, while JUN is a member of the AP-1 transcription factor (TF) family that regulates gene expression in response to stimuli such as neural activity 53, 54. Conversely, the majority of PCY-hit ONCDs were connected to the secondary target AP1S2, a protein involved in clathrin assembly, through the cyclin G-associated kinase GAK (Fig. 3h and Extended Data Figure 6b). A subset of PCY-hit ONCDs were also linked to BTG2 through cyclins CCND1 and CCNH, while a subset of PCY-hit NADs were linked to AP1S2 through RAB9A, a member of the RAS oncogene family (Fig. 3h). Taken together, this reveals pathway convergence on AP-1 transcription factors and cell cycle regulation as a unique signature predictive of anti-glioblastoma activity of neuroactive drugs.
COSTAR can be utilized as a powerful in silico drug screening tool. It can match the drug-target profile of any annotated compound with the identified connectivity signature to predict its hit probability (COSTAR score). To experimentally validate the COSTAR signature and find additional neuroactive drug candidates with anti-glioblastoma activity, we computationally screened across 1,120,823 DTC-annotated compounds and selected 48 previously untested repurposable and BBB-permeable drugs among the top and bottom scoring compounds (COSTAR-hits and COSTAR-negs resp.; Fig. 3i-k). All predicted COSTAR-hits (n=23 drugs) were linked to the secondary target BTG2. Conversely, none of the COSTAR-negs (n=25 drugs) had annotated connections to BTG2 (Extended Data Figure 6c). To validate these COSTAR predictions, we tested all 48 drugs across four GBM patient samples ex vivo (P030, P032, P034, P042), and observed excellent agreement between COSTAR predictions and experimental results (mean AUC=0.94, Fig. 3j,k). The new COSTAR-hits again represented diverse drug classes, including the antipsychotic Trifluoperazine, antiparkinsonian Ethopropazine, antidepressant Sertraline, and bronchodilator Salmeterol (Fig. 3j). These results validate our interpretable molecular machine learning approach for in silico drug discovery, and confirm AP-1 and cell cycle pathway convergence as a predictive signature of neuroactive drugs with ex vivo anti-glioblastoma activity.
From neural activity-dependent signaling to tumor suppression
Convergent drug targets of top NADs represent transcription factors and tumor suppressors, suggesting common gene regulatory networks (GRNs) underlying the anti-glioblastoma activity of NADs. To determine transcriptional changes upon drug perturbation, we profiled 20 drug responses across two time-points (6 and 22 hours) in LN-229 cells by DRUG-Seq 55, 56. These 20 drugs were selected to include PCY-hit NADs spanning diverse drug classes (n=11), PCY-hit ONCDs (n=7), PCY-negative NADs (n=2), and a DMSO control (Fig. 4a-d, Extended Data Fig. 7a,b, and Supplementary Table 3). Differential gene expression analysis upon PCY-hit NAD treatment compared to the PCY-negative NADs and DMSO control revealed a strong transcriptional drug response, with upregulated genes enriched in ‘MAPK signaling’ and ‘transcriptional misregulation of cancer’ pathways at both time-points (Fig. 4b and Extended Data Fig. 7c,d).
In remarkable alignment with COSTAR, AP-1 and BTG family member genes were strongly upregulated in response to the PCY-hit NADs (Fig. 4b,d and Extended Data Figure 7e). This upregulation was observed even for Vortioxetine and Brexpiprazole, both lacking DTC-annotations at the time of analysis and thus not contributing to the COSTAR training (Fig. 4d). We saw rapid and sustained upregulation of AP-1 TFs JUN and FOS, as well as BTG1, a close homologue of BTG2 identified by COSTAR 57, 58 (Fig. 4b,d). JUN and FOS are context-dependent oncogenic factors as well as canonical ‘immediate-early genes’ (IEGs) that are rapidly induced upon neural activity 53, 54, 59. The presence of other upregulated IEGs, including NR4A1, EGR1 and ARC, strengthened this surprising involvement of neural-activity dependent signaling in glioblastoma (Fig. 4b).
To find key upstream regulators mediating the transcriptional response to PCY-hit NADs, we performed transcription factor binding-site (TFBS) enrichment analysis of the upregulated genes (Fig. 4c and Extended Data Figure 7f). The most significantly enriched TF motifs at 6 hours were AP-1, ATF (a member of the AP-1 superfamily), and CREB (a calcium-activated regulator of AP-1 transcription) 53, 60, 61 (Fig. 4c). Not only were AP-1 TFs the most enriched binding domain, but their own expression was also directly induced upon PCY-hit NAD treatment (Fig. 4b,d and Extended Data Figure 7e). ATF3 was the most significantly upregulated gene across both time-points, while ATF4, JUN, JUNB, FOS, and FOSB were among top differentially expressed genes (Fig. 4b,d and Extended Data Figure 7c). Induction of AP-1 factors was primarily NAD-specific, where ONCD treatment did not elicit a similar global transcriptional response (Fig. 4d and Extended Data Figure 7c). Though NAD-induced AP-1 expression was sustained across both time-points, TFBS enrichment analysis of upregulated genes at the 22 hour time-point identified forkhead TF family members (e.g. FOXO1, FOXO3, FOXD3, HFH1) as a gene regulatory module succeeding AP-1 Extended Data Figure 7f). AP-1 and forkhead TFs have well-established roles in mediating immediate-early response to neural activity and regulating long-term cell differentiation, respectively 53, 54, 62–64.
Upstream of AP-1, a rapid Ca2+ influx and calcium-dependent signaling typically precede IEG expression and AP-1 activation 53, 61, 65, 66. We therefore measured both NAD-mediated endoplasmic reticulum (ER) calcium store release as well as extracellular calcium influx in LN-229 cells by high-throughput FLIPR calcium assays (n=17-18 drugs; Supplementary Table 3). While none of the tested drugs triggered ER Ca2+ store release, we observed an immediate and strong extracellular Ca2+ influx in response to 5 out of 8 of our PCY-hit NADs (Fig. 4e,f and Extended Data Fig. 8a,b). The strongest Ca2+ influx was observed upon Vortioxetine treatment, while two additional antidepressant hits, Paroxetine and Fluoxetine, also elicited strong responses (Fig. 4e,f). In contrast, the PCY-neg NADs (n=6) including two other antidepressants, and ONCD-hits (n=2; Elesclomol, TMZ) did not trigger calcium influx (Fig. 4e,f). These results demonstrate that, for the majority of our top NADs, a rapid drug-induced Ca2+ influx precedes IEG upregulation and subsequent anti-glioblastoma activity.
Downstream of AP-1, we evaluated whether BTG tumor suppressors could be direct effectors of the AP-1 gene regulatory network. BTG1 was one of the top 20 most significantly induced genes by the PCY-hit NADs (Fig. 4b,d and Extended Data Figure 7c). BTG2 was strongly induced in specific conditions, including Vortioxetine and Paroxetine treatment (Fig. 4d). To delineate regulators of BTG family genes, we leveraged genome-wide mapping of transcriptional regulatory networks by PathwayNet, a tissue-aware data integration approach that utilizes 690 ChIP-Seq datasets from the ENCODE project 67. The most enriched transcriptional regulators of BTG1/2 were members of the AP-1 TF network (e.g. JUN, ATF3, FOS), implying BTG tumor suppressor gene expression is directly mediated by AP-1 factors (Fig. 4g).
Congruence between NAD-induced AP-1 activation and its anti-glioblastoma activity would strengthen a causal role for this gene regulatory network. Remarkably, drug-induced expression of AP-1 TFs and BTG genes (the ‘COSTAR signature’) was highly correlated with a drug’s ex vivo anti-glioblastoma efficacy in patient samples (R=0.72, P=1.4e-05; Fig. 4h). We additionally performed BTG1/2 and JUN loss-of-function experiments by siRNA-mediated knockdown in LN-229 cells. Quantitative RT-PCR after 72 hours of gene silencing confirmed reduced expression of BTG1/2 and JUN and revealed interdependent regulatory interactions governing their expression (Extended Data Figure 8c). BTG1 but not BTG2 inhibition accelerated cell growth measured by live-cell imaging across 7 days (Fig. 4i). Increased cell proliferation upon BTG1 inhibition was also validated by endpoint pharmacoscopy after 3 days (Fig. 4j). By DRUG-Seq analysis, BTG1 induction represented a pan-NAD signature, while BTG2 was induced only in specific drug conditions such as Vortioxetine treatment (Fig. 4d). Vortioxetine also triggered the strongest calcium increase and represented the most effective NAD overall by pharmacoscopy. We therefore next evaluated the functional dependencies of Vortioxetine anti-glioblastoma efficacy on BTG1/2 and JUN. After two days of siRNA-mediated gene silencing and one subsequent day of drug treatment, BTG1 inhibition rescued Vortioxetine-mediated cell death (Fig. 4j).
Together, these results propose a model in which a subset of neuroactive drugs mediate anti-glioblastoma activity by triggering a rapid calcium influx, IEG and AP-1 transcription factor activation, followed by engagement of an antiproliferative program that includes BTG-driven tumor suppression (Fig. 4k).
Potent and rapid AP-1 induction by Vortioxetine
To closer examine the temporal dynamics of the most effective neuroactive drug, Vortioxetine, we performed in-depth transcriptomic and proteomic profiling at 3-6 time-points in LN-229 cells (Fig. 5a and Extended Data Fig. 8d-h). Gene expression change over time was the strongest driver of variance in Vortioxetine-treated RNA samples analyzed by principal component analysis (Extended Data Figure 8d). Among the top 100 gene loadings for PC1 (38.9% explained variance), we observed induction of AP-1 TFs FOS, JUNB, ATF4, and the AP-1 effector gene ARC already at 3 hours (Fig. 5a and Extended Data Figure 8e). At the RNA level, 10 AP-1 TFs were significantly upregulated at all time-points, with temporal dynamics depending on individual TFs (Fig. 5a and Extended Data Figure 8e). At the protein level, gene set enrichment analysis (GSEA) at 3 hours revealed DNA-binding transcription activator activity as the most over-represented gene ontology (GO) term (Extended Data Fig. 8f-h). Rapid and sustained protein expression of AP-1 TFs was observed between 3 and 9 hours, coinciding with upregulation of BTG1/2 and other negative regulators of cell cycle such as CDKN1B and PPM1B (Fig. 5a and Extended Data Fig. 8f,h). Significant activation of MAPK, ER stress response, and DNA damage pathways were also observed (Fig. 5a). Conversely, cytoskeletal components and oncogenic RTKs associated with the malignant phenotype of glioblastoma, including EGFR, NTRK2, and PDGFRA, were downregulated upon Vortioxetine treatment (Fig. 5a).
AP-1 activity and function in tumorigenesis can be context- and cell type-dependent 59. Consequently, we set out to confirm AP-1 induction upon NAD treatment in compositionally heterogeneous glioblastoma patient samples taken directly after surgery. We examined AP-1 expression both at the single-cell gene and protein level.
We first performed scRNA-Seq on cells from patient P024 following 3 hours of ex vivo Vortioxetine treatment (Fig. 5b,c and Extended Data Fig. 9a,b). Analysis of 1736 single-cell transcriptomes revealed 4 main clusters with intermixed Vortioxetine-treated and DMSO-control cells (Extended Data Figure 9a). Cluster 1 represented the most malignant cluster with highest expression levels of glioblastoma stemness markers such as Nestin, Ki67, CCND2, and VEGFA. Cluster 2 and 3 also expressed glioblastoma stemness markers, at lower levels than cluster 1. Cluster 4 represented an immune cluster, with the highest expression of the pan-immune marker CD45 (Fig. 5b and Extended Data Figure 9b). Per-cluster comparison of Vortioxetine-treated versus DMSO-control cells revealed glioblastoma-specific induction of AP-1 factors upon Vortioxetine treatment (Fig. 5c). For example, JUNB, JUND, and AP-1 effector gene ARC were upregulated in cluster 1, while ATF4 and MAF were induced in all three glioblastoma clusters (cluster 1-3), with a more pronounced induction of ATF4 in clusters 2 and 3 (Fig. 5c).
Next, we performed immunofluorescence against JUND, ATF4, FOS and the AP-1 effector gene HOMER1 in three additional glioblastoma patient samples (P039, P040, P042) after Vortioxetine treatment (Fig. 5d,e). AP-1 factor induction in Nestin+ glioblastoma cells was observed in a patient-, time-point, and concentration-dependent manner. The strongest induction of all tested AP-1 factors was seen in patient sample P040 that had high abundance of complex morphotypes (M1-M3), which were reduced upon Vortioxetine treatment (Fig. 5d,e). HOMER1 and ATF4 were induced in all three patient samples, while FOS and JUND exhibited more patient variability (Fig. 5d). Together, this detailed analysis of Vortioxetine response highlights the added layer of complexity relating to heterogeneous cellular states across cell types and patients, and supports our transcriptomic and functional data elucidating AP-1 induction to be a key neural vulnerability hijacked by PCY-hit NADs.
Preclinical translation of neuroactive drugs
Vortioxetine elicits a strong Ca2+ influx and subsequent AP-1 response, reduces aggressive GSC morphologies, and potently induces cell death of glioblastoma cells ex vivo. To determine if Vortioxetine affects other tumor properties that contribute to recurrence such as invasiveness and long-term growth, we performed a series of experiments to test the effect of Vortioxetine on cell invasion, clonogenic survival, and spheroid formation using established human long-term glioma cell lines (LN-229, LN-308) and glioma-initiating cells (ZH-161, ZH-562) (Extended Data Fig. 9c-f). In a collagen-based spheroid invasion assay, Vortioxetine decreased the average migration distance and reduced the number of invading cells in 3 out of 4 lines (Extended Data Fig. 9c,d). In a resazurin-based clonogenic survival assay, Vortioxetine strongly impaired long-term cell viability in a concentration-dependent manner across all cell lines (4 out of 4 lines, Extended Data Figure 9e). In a complementary 3D spheroid-formation assay, Vortioxetine potently inhibited tumor growth over time (Extended Data Figure 9f). These results demonstrate the multifaceted anti-tumor effects of Vortioxetine upon glioblastoma invasion, survival, and proliferation.
Finally, we tested the top ex vivo NAD hits including Vortioxetine (PCY-hit NADs; n=4 or 5 drugs across different drug classes) in orthotopic human-xenograft glioblastoma mouse models (Fig. 5f and Extended Data Fig. 10a-c). To account for the variability observed in different orthotopic models, we tested two different models (LN-229, ZH-161) across three independent trials (Trials I-III) of drug-testing in vivo (Fig. 5f and Extended Data Figure 10a). We included Temozolomide (TMZ) as a positive control, and as negative controls we tested one PCY-neg NAD Paliperidone and a vehicle control. Since all tested NADs have confirmed BBB-penetrance and are approved for other neurological disorders, in vivo concentrations were determined a priori based on literature and clinical evidence 68–73. Vortioxetine was consistently the most effective drug (in 3/3 trials) and showed significant survival benefit comparable to TMZ despite being tested at considerably lower dosage (Fig. 5f). Brexpiprazole was the 2nd-best PCY-hit NAD in vivo (2/3 trials), while other NADs showed a significant survival benefit in 1 out of 3 trials (Extended Data Figure 10a). Consistent with PCY, the negative control Paliperidone did not show a significant survival benefit (2/2 trials) (Fig. 5f). In the most aggressive orthotopic model with the shortest median survival of the vehicle control, Vortioxetine and TMZ were the only effective drugs (Trial II: ZH-161; Fig. 5f, right), whereas for the least aggressive model, all tested PCY-hit NADs (5/5 NADs) significantly prolonged survival (Trial III: ZH-161; Extended Data Figure 10a). MRI images of ZH-161 transplanted mice (Trial II) after 15 days of Vortioxetine, Apomorphine, and Temozolomide treatment showed marked reduction of tumor size (Extended Data Fig. 10b,c). The striking consistency of our patient ex vivo and mouse in vivo results demonstrate strong translatability of PCY-based NAD discovery and confirm Vortioxetine as the most promising clinical candidate.
Given the different mechanisms of neuroactive drugs and approved chemotherapies, their successful combination could facilitate the rapid adoption of NADs into clinical routine. Therefore, we tested the combination of Vortioxetine with either first- or second-line standard of care drugs for glioblastoma, TMZ and Lomustine (CCNU) in vivo (Trial IV: ZH-161; Fig. 5g). All three single agents significantly prolonged survival, with Vortioxetine results now confirmed in 4 out of 4 in vivo trials (Fig. 5f,g). Remarkably, compared to TMZ or CCNU single agents, the combination of Vortioxetine with either drug provided even further survival benefit, with a median survival increase of 20-30% compared to the single agents (Fig. 5g). Given the dire need for effective treatment options in glioblastoma, this strong preclinical evidence of the safe antidepressant Vortioxetine urges for clinical investigation in patients.
Discussion
Here we present the first therapeutic single-cell map across glioblastoma patient samples that reveals the morphological and neural molecular heterogeneity of glioblastoma. Glioblastoma stem cells adopt distinct cell morphological states that encode clinical prognostic value. While the presence of tumor microtubes has been associated with tumor grade 2, we show that even within glioblastoma, complex GSC morphologies are prognostic of shorter progression-free survival. Image-based drug screening in the context of heterogeneous patient samples predicted clinical response to chemotherapy and enabled the discovery of repurposable neuroactive drugs that target the spectrum of glioblastoma cells across 27 patients and various model systems, greatly expanding upon prior literature 74–76.
A number of new personalized therapeutic opportunities emerge from our exploratory drug response analysis across a genetically and clinically heterogeneous patient cohort. These include Elesclomol sensitivity increasing with age, Abemaciclib response associating with p53 mutant status, and higher Brexpiprazole response in men. Response to the antidepressant Vortioxetine, the most promising preclinical candidate, was particularly aligned with aggressive GSC morphotypes present in patients with poor prognosis. These efforts contribute to the nascent community of glioblastoma research focusing on the investigation of patient-derived tumor explants that facilitate translational investigation of complex tumor behavior, including the development of genetically characterized patient cultures, organoid biobanks, and regionally annotated samples 29, 77–82.
Our systematic analysis of the neuroactive drug mechanisms, drug target expression, and functional genetic dependencies indicated a diverse set of possible neural vulnerabilities of glioblastoma. Given this diversity, it is surprising that our interpretable machine learning approach COSTAR identified a simple drug-target connectivity signature predictive of anti-glioblastoma efficacy. COSTAR effectively applies Occam’s razor to the collective biochemical drug-protein-protein interaction networks, offering a novel conceptual framework applicable to all fields of drug discovery. Through COSTAR, we uncovered a convergence of AP-1 transcription factor activity and cell cycle regulation on BTG-mediated tumor suppression. AP-1 and BTG upregulation was a defining feature of the response to neuroactive drugs with anti-glioblastoma activity, where a growth-suppressing role for BTG1 was confirmed by functional genetics. While the defining pharmacological properties remain to be identified, our results indicate a plurality of drug mechanisms converging on this novel tumor-suppressing pathway.
Previous studies have demonstrated the role of neuronal activity in regulating glioblastoma growth at the brain-tumor interface, highlighting the influence of the tumor microenvironment in mediating the neural behavior of the tumor. Here we reveal that cell-intrinsic neural activity in glioblastoma mimics those of neural lineage cells, offering a therapeutic vulnerability that enables direct targeting of the tumor. In neurons, immediate early gene expression including AP-1 is typically a hallmark of neural activity or insult, but in cancer cells, AP-1 factors were originally described as context-dependent oncogenes. We now find that neuroactive drugs can hijack this activity-dependent neural signaling, triggering a strong transcriptional response that, in the context of glioblastoma cells, leads to rapid cell death. Treating glioblastoma tailored to the cellular history and lineage of the cancer rather than its unstably transformed state may represent new hope for this devastating disease.
Extended Data Figure Legends
Methods
Patient sample processing and drug testing
Surgically removed tumors were collected at the University Hospital of Zurich (Universitätsspital Zürich, USZ) with approval by the Institutional Review Board, ethical approval number KEK-StV-Nr.19/08, BASEC number 2008-02002. Metadata of the prospective and retrospective glioblastoma patient cohorts including clinical parameters, experiment inclusion, and genetics summary can be found in Supplementary Table 1. The prospective cohort consists of patients where fresh tissue was taken directly after surgery (n=27 patients for drug screening, plus an additional n=3 patients for validation experiments). The retrospective cohort (n=18 patients) consists of patients for which snap-frozen bio-banked tissue was available covering a broad spectrum of progression-free survival. Retrospective samples were further selected based on quality control measures including cell viability, cell number, and the amount of debris present in the sample.
Tissue samples were first washed with PBS and cut into small pieces using single-use sterile scalpels. Subsequent dissociation was performed in reduced serum media (DMEM media; #41966029 with 2% FBS; #10270106, 1% Pen-strep; #15140122, and 25mM HEPES; #15630056, all products from Gibco) supplemented with Collagenase IV (1mg/ml) and DNaseI (0.1mg/ml) using the gentle MACS Octo Dissociator (Miltenyi Biotec, 130-096-427). Homogenates were filtered through a 70um Corning cell strainer (Sigma-Aldrich, CLS431751) and washed once with PBS containing 2mM EDTA. Myelin and debris removal was performed by a gradient centrifugation of the cell suspension in a 7:3 mix of PBS:Percoll (Sigma-Aldrich, P4937) and washed again with PBS. Dissociated patient cells were seeded at 0.5-1.5×10^4 cells/well into clear-bottom, tissue-culture treated, CellCarrier-384 Ultra Microplates (Perkin Elmer, #6057300). Prior to cell seeding, tested drugs were re-suspended as 5mM stock solutions and dispensed into the 384 well plates using an Echo 550 liquid handler (Labcyte) at their respective concentrations in a randomized plate layout to control for plate effects. Information regarding drugs used in this study can be found in Supplementary Table 3. For drug library testing (glioblastoma drugs, GSDs; neuroactive drugs, NADs; oncology drugs, ONCDs) cells were incubated in reduced serum media at 37°C, 5% CO2 for 48 hours with drugs.
Targeted Next Generation Sequencing (NGS, Oncomine Comprehensive Assay v3)
Tissue blocks from patient-matched glioblastomas were used to determine genetic alterations including mutations, copy number variation and gene fusion. Formalin-fixed paraffin-embedded (FFPE) tissue blocks were collected from the Tissue Biobank at USZ archives. Tumor area was marked on the HE slide and relative tumor cell content was estimated by a trained pathologist. 1-3 cores cylinders (0.6 mm diameter) or 40 um from the tumor area of the FFPE blocks were used for DNA and RNA isolation. DNA was isolated with the Maxwell 16 FFPE Tissue LEV DNA Purification Kit (Promega, #AS1130) according to the manufacturer’s recommendations. The double-strand DNA concentration (dsDNA) was determined using the fluorescence-based Qubit dsDNA HS Assay Kit. RNA was extracted with the Maxwell 16 FFPE Tissue LEV RNA Purification Kit (Promega, #AS1260) according to the manufacturer’s recommendations. To avoid genomic DNA contamination, samples were pretreated with DNase1 for 15 min at room temperature (RT). Library preparation with 20 ng DNA or RNA input was conducted using the Oncomine Comprehensive Assay v3 following the manufacturer’s instructions. Adaptor/barcode ligation, purification and equilibration was automated with Tecan Liquid Handler (EVO-100). NGS libraries were templated using Ion Chef and sequenced on a S5 (Thermo Fisher Scientific), and data were analyzed using Ion Reporter Software 5.14 with Applied Filter Chain: Oncomine Variants (5.14) settings and Annotation Set :Oncomine Comprehensive Assay v3 Annotations v1.4.
For NGS data analysis, the Ion Reporter Software within Torrent Suite Software was used, enabling detection of small nucleic variants (SNVs), copy number variations (CNVs), gene fusions and indels from 161 unique cancer driver genes. Detected sequence variants were evaluated for their pathogenicity based on previous literature and the ‘ClinVar’ database 86. Gene alterations described as benign or likely benign were not included in our results. Non-pathogenic mutations harboring a Minor Allele Frequency higher than 0.01 were not selected. The Default Fusion View parameter was selected. For CNV confidence range, the default filter was used to detect gains and losses using the confidence interval values of 5% confidence interval for Minimum Ploidy Gain over the expected value and 95% confidence interval for Minimum Ploidy Loss under the expected value. CNV low confidence range was defined for gain by copy number from 4 to 6 (lowest value observed for CNV confidence interval 5%:2.9) and loss from 0.5 to 1 (highest value observed for CNV confidence interval 95%:2.43). High confidence range was defined by gain up to 6 copy number (lowest value observed for CNV confidence interval 5%:4.54) and loss below 0.5 copy number (highest value observed for CNV confidence interval 95%:1.37). 5% and 95% interval confidence of all selected loss and gain are available in Supplementary Table 2. The minimum number of tiles required was eight. Results are reported as detected copy number.
Cell culture
The adherent human glioblastoma cell lines LN-229 (ATCC, #CRL-2611) and LN-308 were cultured in Dulbecco’s modified Eagle medium (DMEM, #41966, Gibco) supplemented with 10% fetal bovine serum (FBS, #10270106, Gibco). LN-229 and LN-308 cells were passaged using Trypsin-EDTA (0.25%, Gibco, #25200056). For DRUG-seq, RNA-Seq, siRNA knockdown, and proteomics measurements using LN-229 cells, low-passage cells below passage 15 were used. The spheroid human glioblastoma-initiating cell lines ZH-161 and ZH-562 was generated from freshly isolated tumor tissue and cultured in Neurobasal medium (NB, #21103049, Gibco) supplemented with B27 (Gibco, #17504044), 20 ng/mL b-FGF (Peprotech, #AF-100-18B), 20 ng/mL EGF (Peprotech, #AF-100-15), 2 mM L-glutamine (Gibco, #25030081). ZH-161 and ZH-562 cells were passaged using Accutase (Stemcell Technologies, #07920). Cell lines were authenticated at the Leibniz Institute DSMZ (Braunschweig, Germany) and regularly tested negative for mycoplasma.
Immunocytochemistry
Cells were fixed with 4% PFA (Sigma-Aldrich, #F8775) in PBS and blocked in 5% FBS and 0.1% Triton containing PBS. For characterization of cellular composition across glioblastoma patient samples, cells were stained overnight at 4°C in blocking solution with the following antibodies and dilutions: Alexa Fluor® 488 anti-S100 beta (1:1000, Abcam, #ab196442, clone EP1576Y), PE anti-Nestin (1:150, Biolegend, #656806, clone 10C2), Alexa Fluor® 488 anti-CD3 (1:300, Biolegend, #300415, clone UCHT1), Alexa Fluor® 647 anti-CD45 (1:300, Biolegend, #368538, clone 2D1) and DAPI (1:1000, Biolegend, #422801, stock solution 10mg/ml). Due to the manufacturer discontinuation of the Alexa Fluor® 488 anti-S100 beta antibody, from patient sample P30 and onwards in the prospective cohort, samples were either stained with a self-conjugated Alexa Fluor® 488 anti-S100 beta antibody, where Alexa Fluor™ 488 NHS Ester (Thermo Scientific, #A20000) was conjugated to the anti-S100 beta antibody (Abcam, #ab215989, clone EP1576Y) or the following antibody combinations where the 488 and 555 channel markers were swapped: Alexa Fluor® 488 anti-Nestin (1:150, Biolegend, #656812, clone 10C2), Alexa Fluor® 555 anti-S100 beta (1:1000, Abcam, #ab274881, clone EP1576Y), PE anti-CD3 (1:300, Biolegend, #300441, clone UCHT1), Alexa Fluor® 647 anti-CD45 (1:300, Biolegend, #368538, clone 2D1).
Other antibodies used in this study include the following: Alexa Fluor® 647 anti-Tubulin Beta 3 (1:1000, Biolegend, #657406, clone AA10), Alexa Fluor® 555 anti-Cleaved Caspase-3 (1:500, Cell Signaling Technology, #9604S), Alexa Fluor® 546 anti-HOMER (1:300, Santa Cruz Biotechnology, #sc-17842 AF546, clone D-3), PE anti-CFOS (1:300, Cell Signaling Technology, #14609S, clone 9F6), FITC anti-ATF4 (1:300, Abcam, #ab225332), Alexa Fluor® 488 anti-JUND (1:300, Santa Cruz Biotechnology, #sc-271938 AF488, clone D-9), Alexa Fluor® 594 anti-CD45 (1:300, Biolegend, #368520, clone 2D1).
Confocal microscopy and image analysis
Imaging of the 384 well plates was performed with an Opera Phenix automated spinning-disk confocal microscope at 20x magnification unless otherwise specified (Perkin Elmer, HH14000000). Single cells were segmented based on their nuclei (DAPI channel) using CellProfiler 2.2.0. Downstream image analysis was performed with MATLAB R2019a-R2020a. Fractions of marker positive cells for each condition were derived for each patient sample based on the histograms of the local background corrected intensity measurements. Marker positive fractions were averaged across each well/condition and compared to the negative DMSO control.
Deep learning of glioblastoma stem cell morphologies
To generate a training dataset, Nestin+CD45-cells identified by immunofluorescence across the whole prospective glioblastoma patient cohort were cropped as 150×150 pixel images. These single-cell image crops were then manually-curated and labeled as four morphological classes (M1-M4 morphotypes) based on their shape, size, and presence of tumor microtubes. A convolution neural network (CNN) with a modified Alexnet architecture 85 as shown in Extended Data Figure 2a was then trained on this manually-curated training data with 12,757 images per class and 51,028 images in total. CNN training included usage of the Adam optimizer, with a mini-batch size of 150 and a maximum number of 30 epochs. The initial learning rate was set to 0.001 with a piecewise learning rate schedule and a drop factor of 0.01 every 6 epochs.
Network performance is shown as a confusion matrix in Extended Data Figure 2b. All Nestin+ single-cell images were subsequently classified by this pre-trained CNN to determine morphotype abundances across patients and drug conditions. For visualization of the CNN-based GSC morphotypes, UMAP plots were generated based on the CNN feature space that consists of ten dimensional activations taken from the 2nd-last fully connected layer of the network. The CNN feature space of 84,180 cells (maximally 1000 cells per class and patient, n=27 patients) was projected on the UMAP using the following parameters: distance metric, seuclidean; number of neighbors, 10; minimal distance, 0.06. Different morphological and marker-based features from the original cell segmentation determined by CellProfiler 2.2.0 such as cell area, eccentricity, and roundness, and mean marker intensity were selected for visualization.
siRNA knockdown and quantitative real-time PCR
All siRNAs used in the study were part of the MISSION® esiRNA (Sigma-Aldrich, Euphoria Biotech) library (Supplementary Table 5) and ordered as custom gene arrays (esiOPEN, esiFLEX). FLUC esiRNA (EHUFLUC) targeting firefly Luciferase was used as a negative control, and KIF11 esiRNA (EHU019931) was used as a positive control for transfection and viability. For all siRNA experiments, low-passage LN-229 cells were used. siRNAs were transfected at 10ng/well in 384 well plates and 40ng/well in 96 well plates using Lipofectamine RNAiMAX (Invitrogen, #13778075). Imaging and drug incubation experiments were conducted in 384 well plates, while Incucyte live cell imaging and cell lysis preparation for RNA extraction and quantitative real-time PCR was performed in 96 well plates. For 384 well plates, both the siRNAs and Lipofectamine transfection reagent were dispensed using a Labcyte Echo liquid handler in a randomized plate layout to control for plate effects when possible. For data presented in Figure 3d, Extended Data Figure 5c,d, and Extended Data Figure 8c, cells were incubated at 37°C, 5% CO2 for 48 hours following siRNA transfection before fixing, immunohistochemistry, and RNA extraction. For data presented in Figure 4j, following 48 hours of siRNA transfection, cells were incubated for an additional 24 hours with either DMSO vehicle control or Vortioxetine (10µM) before fixing and subsequent analysis.
siRNA knockdown efficiency and relative abundance for the following target genes; BTG1, BTG2, JUN, and MKI67 was measured by TaqMan™ Array plates (Applied Biosystems, Standard, 96-well Plate; Format 16 with candidate endogenous controls) using the TaqMan™ Fast Advanced Master Mix (Thermo Scientific, #A44360) on a QuantStudio™ 3 Real-Time PCR System (Applied Biosystems, #A28567). Total RNA from LN-229 cells was extracted using the Direct-zol RNA MicroPrep Kit (Zymo Research, #R2062), RNA concentration was measured using the Qubit 4 Fluorometer (Thermo Scientific), and cDNA synthesized with the iScript™ cDNA Synthesis Kit (Bio-Rad, #1708890). For each TaqMan biological replicate assay (n=3 replicates) 25ng of cDNA per sample was used. To calculate the relative abundance of each target gene, the geometric mean Ct value of four endogenous control genes (18s rRNA, GAPDH, HPRT, GUSB) was subtracted from each [sample-target gene] Ct value to derive the deltaCt (dCt) value. Then, the mean deltaCt value from FLUC negative control samples was subtracted from each [sample-target gene] deltaCt value to derive the delta-deltaCt (ddCt) value. Finally, relative abundance (fold-difference) of each [sample-target gene] was calculated as the 2^(-ddCt).
COSTAR: Convergence of secondary drug-targets analyzed by regularized regression
COSTAR is an interpretable molecular machine learning approach that utilizes logistic LASSO regression in a cross-validation setting to learn a multi-linear model that identifies the minimal set of drug-target connections that maximally discriminates PCY-hit drugs from PCY-negative drugs.
Drug-target connections were retrieved from the Drug Target Commons (DTC) 52. DTC is a crowd-sourced platform that integrates drug-target bioactivities curated from both literature and public databases such as PubChem and ChEMBL. Drug-target annotations (DTC bioactivities) listed as of August 2020 were included, with the target organism limited to Homo sapiens. Among PCY-tested drugs in our NAD and ONCD libraries, 127 out of 132 drugs had DTC ‘bioactivity’ annotations. PTGs with biochemical associations to a given drug correspond to bioactivities with the inhibitory constant ‘KI’ as the ‘End Point Standard Type’. Extended PTGs (ePTGs) include all annotated drug bioactivities. Secondary target genes (STGs) down-stream of ePTGs were retrieved by high-confidence protein-protein interactions annotated in the STRING database (interaction score≥0.6). The final drug-target connectivity map that was used for COSTAR consisted of 127 PCY-tested drugs, 975 extended primary targets, 10,573 secondary targets, and 114,517 network edges. The 127 drugs were labeled either as PCY-hits (n=30, equally split across NADs and ONCDs) or PCY-negative drugs (n=97) based on the ranked mean PCY score across patients.
A 20-fold cross-validated LASSO generalized linear model was trained in Matlab with the drug-target connectivity map as the predictor variable and PCY-hit status (hit vs. neg) as the binomially-distributed response variable to identify the optimal regularization coefficient (lambda) across a geometric sequence of 60 possible values. Final model coefficients were fitted using the lamba value corresponding to the minimum deviance in the cross-validation analysis shown in Extended Data Figure 6a. COSTAR performance was first evaluated on the training dataset, represented as a confusion matrix in Figure 3g. Using this trained linear model, COSTAR was next utilized as an in silico drug screening tool to predict the PCY-hit probability (COSTAR score) based on the connectivity of 1,120,823 compounds annotated in DTC (Supplementary Table 6). For interpretability, COSTAR subscores, defined as the individual connectivity to target genes multiplied by their respective coefficients (betas) in the linear model, can be investigated in Extended Data Figure 6b,c. COSTAR predictions from this in silico screen were further experimentally validated in glioblastoma patient samples on a set of new drugs predicted as either COSTAR-hits or COSTAR-negs (n=48 drugs total; n=23 COSTAR-hits; n=25 COSTAR-negs).
Single-cell RNA-Seq and analysis of other published datasets
Cryopreserved single-cell suspensions of glioblastoma patient samples that were part of the prospective cohort were thawed in reduced serum media (DMEM containing 2% FBS) and used for single-cell RNA-Seq experiments. Viability markers SYTOX Blue (1 μM, Thermo Fisher, #S11348) and DR (1 μM, Biolegend, #424101) were added to the cell suspension at least 15 minutes before sorting. FACS gates were set based on CD45 (Alexa Fluor® 594 anti-CD45, 1:20, Biolegend, #368520, clone 2D1), SYTOX Blue and DR intensities to isolate live CD45+ and CD45-populations as shown in Extended Data Figure 1b using the BD FACSAriaTM Fusion Cell Sorter. Cells were sorted into DNA LoBind® Eppendorf tubes (VWR, #525-0130), then CD45-cells were mixed with CD45+ cells at 2:1 to 10:1 ratios depending on cell availability to enrich for glioblastoma cells. For patient sample P024 that was used to measure the effect of Vortioxetine drug treatment, FAC-sorted cells were incubated for 3 hours with or without 20µM Vortioxetine before proceeding to library preparation. Single-cell RNA-Seq library preparation was performed using the Chromium Next GEM Single Cell 3’ v3.0 and v3.1 kits (10x Genomics) according to the manufacturer’s protocol. Libraries were sequenced on the Novaseq 6000 (Illumina). Read alignment to the GRCh38 human reference genome, generation of feature-barcode matrices, and aggregation of multiple samples were performed using the Cell Ranger analysis pipeline (10x genomics, versions 3.0.1 and 6.1.1). Four patient samples (P007, P011, P012, P013) were processed in November 2019 with the earlier version of 10x Genomics library prep kits and Cell Ranger analysis pipeline while the later sample (P024) was processed in September 2021. Quality control for this in-house dataset was performed by only analyzing high-quality cells with fewer than 10% of mitochondrial transcripts and genes that had at least a count of 2 in at least 3 cells.
To analyze additional glioblastoma patient samples by single-cell RNA-Seq, we utilized two published datasets: (Neftel et al. 2019) and (Yu et al. 2020). For Neftel et al., we removed cells with less than 2^9 detected genes and/or more than 15% of mitochondrial transcripts. For Yu et al. the data was already prefiltered, but patient samples (7-9, 14-15) that did not correspond to glioblastoma (grade IV astrocytomas) were not included. For both datasets only genes that had at a count of 2 in at least 2 cells were included in the analysis.
Neural patient specificity scores and patient specificity scores for each gene were defined as follows: using the in-house dataset, we identified putative cell types by unsupervised clustering using Monocle 87 and annotated the clusters based on known marker genes as being either immune or neural cells. We then obtained a list of differentially expressed genes between immune and neural cells using DESingle 88, using a logFC cutoff of 0.5. This yielded a list of 11571 neural-specific and 1157 immune specific genes. Using these lists as cell-type specific gene sets, we calculated an immune- and a neural score for each cell using singscore, and classified every cell in the additional datasets as either neural or immune based on a linear combination of both scores. To assess how specifically a gene is expressed in neural cells, we defined a ‘neural specificity score’ as follows: [neural specificity = fraction of neural cells expressing gene – fraction of immune cells expressing gene] where we define expression of a gene in a cell as having any non-zero count. Thus, a positive score indicates that a gene is more often found in neural cells than in immune cells, and vice versa for negative scores. This score ranges from -1 (gene is expressed in all immune cells and no neural cells) to =1 (gene is expressed in all neural cells and no immune cells). Note that for low expressed genes, this score will be close to 0, reflecting the fact that we cannot make clear statements about cell type specificity for genes with expression values close to the detection limit of scRNA-Seq. To assess how much gene expression for a single gene varies across patients, we defined a ‘patient specificity score’ as follows: First, for every gene gi and every patient pj we calculated a cell type composition independent fraction of cells expressing gene gi as [Fraction_expressing_ij = fraction_expressing_immune_ij + fraction_expressing_neural_ij]. We then defined patient specificity as the median absolute deviation (MAD) of fraction_expressing across all patients, thus defining [Patient_specificity_i = mad(Fraction_expressing_i,:)].
DRUG-Seq
High-throughput multiplexed RNA sequencing was performed with the Digital RNA with pertUrbation of Genes (DRUG-Seq) method as described in (Ye et al. 2018) with a few modifications. Modifications to the published method are the following: 1) extraction of RNA prior to cDNA reverse transcription with the Zymo Direct-zol-96 RNA isolation kit (Zymo, #R0256) 2) change of reverse transcription primers for compatibility with standard Illumina sequencing primers 3) cDNA clean-up prior to library amplification performed with the DNA Clean & Concentrator-5 kit (Zymo, #D4013) 3) tagmentation was performed with 2ng input and sequencing library generated using the Nextera XT library prep kit (Illumina, #FC-131-1024). In short, 1×10^4 LN-229 cells were plated in CellCarrier-96 Ultra Microplates (PerkinElmer, #6055302) and incubated overnight in reduced serum media at 37°C, 5% CO2 prior to drug treatment. A total of 20 drugs (Supplementary Table 3) were profiled across two different time-points (6 hours and 22 hours; n=4 replicates per drug and time-point). These 20 drugs were selected to include PCY-hit NADs spanning diverse drug classes (n=11), PCY-hit ONCDs (n=7), PCY-negative NADs (n=2), and a DMSO control. Cells in drug-treated 96-wells were lysed with TRIzol™ Reagent (ThermoFisher, #15596018) and then subsequent cDNA and library prep was performed as described above. 100bp (80:20) paired-end reads were generated using Illumina’s NextSeq 2000 platform.
Calcium assays on the FLIPR platform
For calcium assays, 24 hours prior to the experiment, LN-229 cells were seeded at a density of 70,000 cells/well on poly-D-Lysine-coated ViewPlateTM-96 F TC 96-well black polystyrene clear bottom microplates (PerkinElmer, #6005182) in 100µl full medium. Calcium 6 dye stock solution was prepared by dissolving a vial from Calcium 6 assay kit (Molecular Devices, #5024048) in 10 ml sterile-filtered nominal Ca2+ free (NCF), modified Krebs buffer containing 117mM NaCl, 4.8mM KCl, 1mM MgCl2, 5mM D-glucose, 10mM HEPES (pH 7.4) and 500µl aliquots were stored at -20°C. Before each experiment, the dye stock was freshly diluted 1:10 in NCF Krebs buffer and after removing the medium from the cells, 50µl of the diluted dye was applied per well. In order to allow the cells to absorb the dye into their cytosol, they were incubated at 37°C for 2 hours in the dark. The fluorescence Ca2+ measurements were carried out using FLIPR Tetra® (Molecular Devices) where cells were excited using a 470–495nm LED module and the emitted fluorescence signal was filtered with a 515–575nm emission filter according to the manufacturer’s guidelines.
In the ER Ca2+ store release assay, the stable baselines were established for 50 seconds before 50µl of 2µM (2X) Thapsigargin (Sigma-Aldrich, #T9033) or 40µM (2X) drug solutions freshly prepared in NCF Krebs buffer were robotically dispensed to the cells to determine whether the drugs impact the ER Ca2+ stores. Next, the cells were incubated and fluorescence was monitored in the presence of Thapsigargin or drugs for another 5 min. In the extracellular Ca2+ uptake assay, after initial recording of the baseline, 50µl of 4mM CaCl2 (2X) prepared in NCF Krebs buffer was dispensed onto the cells to re-establish a physiological 2mM calcium concentration and the fluorescence was monitored for 5 min. Next, 60µM (3X) drug solutions freshly prepared in Krebs buffer containing 2mM CaCl2 were robotically dispensed to the cells and the fluorescence was recorded for an additional 4 min. The raw data was extracted with the ScreenWorks software version 3.2.0.14. The values represent average fluorescence level of the Calcium 6 dye measured over arbitrary selected and fixed time frames.
Time-course RNA-Seq library preparation and sequencing
Low passage LN-229 cells (passage 5-6) were seeded at 2×10^5 cells/well into in 6-well Nunc™ Cell-Culture Treated Multidishes (ThermoFisher, #140675) and incubated overnight in reduced serum media at 37°C, 5% CO2 prior to drug treatment. The following day, Vortioxetine (Avachem Scientific, #3380) was manually added to each well at a final concentration of 20µM. At the start of the experiment, LN-229 cells that were not treated with Vortioxetine were collected as the 0 hour time-point. After 3, 6, 9, 12, and 24 hours following Vortioxetine treatment, drug-containing media was removed and cells were collected in TRIzol™ Reagent (ThermoFisher, #15596018). Total RNA was isolated using Direct-zol RNA MicroPrep Kit (Zymo Research, #R2062) according to the manufacturer’s protocol and RNA quality and quantity was determined with the Agilent 4200 TapeStation. Sample RIN scores ranged from 5.9-10 (mean: 9.33). RNA input was normalized to 300-400 ng and RNA libraries were prepared using the Illumina Truseq stranded mRNA library prep following manufacturer’s protocols. 100bp single-end reads were generated using Illumina’s Novaseq 6000 platform with an average sequencing depth of approximately 50 million reads per replicate. Reads were mapped and aligned to the reference human genome assembly (GRCh38.p13) using STAR/2.7.8a and counts were extracted using featureCounts. Subsequent read normalization (variance stabilizing transformation, vsd-normalized counts) and RNA-Seq analysis including differential expression (DE) analysis was performed with the R package ‘DESeq2’ 89.
Time-course Proteomics
Cell preparation and Vortioxetine treatment was performed as in ‘Time-course RNA-Seq library preparation’ except cell numbers were scaled to be seeded in T-150 culture flasks and 3 time-points were measured (0, 3, 9 hours). Peptides for mass spectrometry measurements were prepared using the PreOmics iST kit (PreOmics) on the PreON (HSE AG). The robot was programmed to process 8 samples in parallel. During the first step of processing, cell pellets were resuspended in 50ul of lysis buffer and denatured for 10 minutes at 70°C. According to the manufacturer’s protocol, this step was followed by 3 hours of digestion with trypsin and Lys-C. Peptides were dried in a speed-vac (Thermo Fisher Scientific) for 1 hour before being resuspended in LC-Load buffer at a concentration of 1 ug/ul with iRT peptides (Biognosys).
Samples were analyzed on an Orbitrap Lumos mass spectrometer (Thermo Fisher Scientific) equipped with an Easy-nLC 1200 (Thermo Fisher Scientific). Peptides were separated on an in-house packed 30 cm RP-HPLC column (Michrom BioResources, 75 μm i.d. x 30 cm; Magic C18 AQ 1.9 μm, 200 Å). Mobile phase A consisted of HPLC-grade water with 0.1% formic acid, and mobile phase B consisted of HPLC-grade ACN (80%) with HPLC-grade water and 0.1% (v/v) formic acid. Peptides were eluted at a flow rate of 250 nl/min using a non-linear gradient from 4% to 47% mobile phase B in 228 min. For data-independent acquisition (DIA), DIA-overlapping windows were used and a mass range of m/z 396-1005 was covered. The DIA isolation window size was set to 8 and 4 m/z, respectively, and a total of 75 or 152 DIA scan windows were recorded at a resolution of 30,000 with an AGC target value set to 1200%. HCD fragmentation was set to 30% normalized collision. Full MS were recorded at a resolution of 60,000 with an AGC target set to standard and the maximum injection time set to auto. DIA data were analyzed using Spectronaut v14 (Biognosys). MS1 values were used for the quantification process, peptide quantity was set to mean. Data were filtered using Qvalue sparse with a precursor and a protein Qvalue cut-off of 0.01 FDR. Interference correction and local cross-run normalization was performed. For PRM measurements, peptides were separated by reversed-phase chromatography on a 50 cm ES803 C18 column (Thermo Fisher Scientific) that was connected to a Easy-nLC 1200 (Thermo Fisher Scientific). Peptides were eluted at a constant flow rate of 200 nl/min with a 117 min non-linear gradient from 4–52% buffer B (80% ACN, 0.1% FA) and 25-50%B. Mass spectra were acquired in PRM mode on an Q Exactive HF-X Hybrid Quadrupole-Orbitrap MS system (Thermo Fisher Scientific). The MS1 mass range was 340–1,400 m/z at a resolution of 120,000. Spectra were acquired at 60,000 resolution (automatic gain control target value 2.0*10e5); Normalized HCD collision energy was set to 28%, maximum injection time to 118 ms. Monitored peptides were analyzed in Skyline v20 and results were uploaded to PanoramaWeb.
Incucyte live cell imaging
To measure cell proliferation in real-time, 2.5×10^3 LN-229 cells/well were plated in CellCarrier-96 Ultra Microplates (PerkinElmer, #6055302) 24 hours prior to the experiment, and transfected with BTG1, BTG2, and FLUC (-) MISSION® esiRNAs (Sigma-Aldrich, Euphoria Biotech, 40ng/well) using Lipofectamine RNAiMAX (Invitrogen, #13778075). Further details regarding siRNAs can be found in Supplementary Table 5 and Methods related to ‘siRNA knockdown and quantitative real-time PCR’. Real-time confluence of cell cultures (n=4 replicate wells/condition) was monitored by imaging every 2 hours for 7 days at 10x magnification with the ‘phase’ channel using the Incucyte live-cell analysis system S3 (Sartorius). Automatic image segmentation and analysis of the phase contrast images was performed by the Incucyte base analysis software (version 2020B).
Clonogenic survival assay
Adherent cells (LN-229: 50 cells; LN-308: 300 cells) were seeded in six replicates in 100 µL per well in 96-well plates and incubated overnight. On the following day, medium was replaced by fresh medium containing indicated final concentrations of Vortioxetine or DMSO. Glioblastoma-initiating cells (500 cells) were seeded in 75 µL medium and incubated overnight. Treatment was initiated by addition of 75 µL medium containing 2x concentrated Vortioxetine or DMSO to reach indicated final concentrations. DMSO concentration was kept at 0.5% for all treatments and controls. Following treatment addition, cells were cultured for 11 (LN-229) to 13 days (other cell lines) and clonogenic survival was estimated from a resazurin-based assay 90 using a Tecan M200 PRO plate reader (λEx = 560 nm / λEm = 590 nm).
Collagen-based spheroid invasion assay
Spheroid invasion assay was performed as described in (Kumar et al. 2015). Briefly, 2000 cells were seeded in six replicates into cell-repellent 96 U-bottom well plates (Greiner, #650979) and incubated for 48 hours to allow spheroid formation. Subsequently, 70 µL medium were removed, spheroids were overlaid with 70 µL 2.5% Collagen IV (Advanced Biomatrix, #5005-B) in 1xDMEM containing sodium bicarbonate (Sigma-Aldrich #S8761) and collagen was solidified in the incubator for 2 hours. Collagen-embedded spheroids were then overlaid with 100 µL chemoattractant (NIH-3T3-conditioned medium) containing 2x concentrated Vortioxetine/DMSO (0.5% final DMSO concentration across conditions) and incubated for 36 hours. Spheroids were stained with Hoechst and images were acquired on a MuviCyte imaging system (Perkin Elmer, #HH40000000) using a 4x objective. Images were contrast-enhanced and converted to binary using ImageJ/Fiji and quantified with the automated Spheroid Dissemination/Invasion counter software (aSDIcs), which quantifies the migration distance from the center of the spheroid for each detected cell nucleus 91.
In vivo drug testing
All animal experiments were done under the guidelines of the Swiss federal law on animal protection and were approved by the cantonal veterinary office (ZH98/2018). CD1 female nu/nu mice (Janvier, Le Genest-Saint-Isle, France) of 6 to 12 weeks of age were used in all experiments and 100’000 LN-229-derived- or 150’000 ZH-161-derived cells were implanted 92. Tumor-bearing mice were treated from day 5 – day 21 after tumor implantation with intraperitoneally (i.p.) administered Vortioxetine daily 10mg/kg, Paliperidone daily 5mg/kg, Apomorphine daily 5mg/kg, Aprepitant daily 20mg/kg, Brexpiprazole daily 1mg/kg, Chlorpromazine three time per week 10mg/kg, Temozolomide 50mg/kg for five consecutive days, CCNU 20mg/kg at day 7 and 14 after tumor implantation, or daily DMSO control. Magnetic resonance imaging (MRI) was performed with a 4.7 T imager (Bruker Biospin, Ettlingen, Germany) when the first mouse became symptomatic. Coronal T2-weighted images were acquired using Paravision 6.0 (Bruker BioSpin).
Acknowledgements
We thank the patients and their families for making this study possible. We thank core facilities at ETH Zurich and University of Zurich, including the Functional Genomics Center Zurich (FGCZ), the Flow Cytometry Core Facility (EPIC, FCCF), and central informatics services (ID). This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (Grant agreement No. 803063-SCIPER; BS), and from the Personalized Health and Related Technologies (PHRT) Strategic Focus Area of the ETH Domain (Project #2021-566; BS). We further acknowledge funding from the European Union/European Commission (MSCA Individual Fellowship; SL), the European Molecular Biology Organization (EMBO Long-term Post-doctoral Fellowship; SL), the Swiss National Science Foundation (grant no. 310030_204972; MAH), and financial support from the Department of Biology as well as the Institute of Molecular Systems Biology of the ETH Zurich. We thank the research and administrative staff who helped us throughout the study, including Yasmin Festl and Silas Krämer, and members of the Snijder lab for discussions.
Footnotes
↵* Shared-first authorship.
References
- 1.↵
- 2.↵
- 3.↵
- 4.↵
- 5.↵
- 6.↵
- 7.↵
- 8.↵
- 9.↵
- 10.
- 11.↵
- 12.↵
- 13.↵
- 14.↵
- 15.
- 16.↵
- 17.↵
- 18.
- 19.↵
- 20.
- 21.
- 22.↵
- 23.
- 24.
- 25.
- 26.
- 27.
- 28.↵
- 29.↵
- 30.
- 31.↵
- 32.↵
- 33.
- 34.↵
- 35.↵
- 36.↵
- 37.
- 38.↵
- 39.↵
- 40.
- 41.
- 42.↵
- 43.↵
- 44.↵
- 45.↵
- 46.
- 47.↵
- 48.↵
- 49.↵
- 50.
- 51.↵
- 52.↵
- 53.↵
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.↵
- 61.↵
- 62.↵
- 63.
- 64.↵
- 65.↵
- 66.↵
- 67.↵
- 68.↵
- 69.
- 70.
- 71.
- 72.
- 73.↵
- 74.↵
- 75.
- 76.↵
- 77.↵
- 78.
- 79.
- 80.
- 81.
- 82.↵
- 83.↵
- 84.↵
- 85.↵
- 86.↵
- 87.↵
- 88.↵
- 89.↵
- 90.↵
- 91.↵
- 92.↵