Patient sample processingEthics statement and patient cohortAdult patients with IDH-wildtype glioblastoma and CNS World Health Organization (WHO) grade 4 according to the 2021 WHO classification of CNS tumors treated either at the University Hospital of Zurich or the Cantonal Hospital St. Gallen provided informed consent to take part in the study with approval by the institutional review board (ethical approval number KEK-StV-Nr.19/08; BASEC numbers 2019-02027 and 2021-00652). There was no limit on tumor size for the human samples included in the study and no selection bias of the enrolled patients. Clinical characteristics of the prospective and retrospective patient cohorts, including clinical parameters, experiment inclusion, sex, age and genetics summary, can be found in Supplementary Table 1 and Supplementary Data 1. The prospective cohort consists of patients where fresh tissue was processed directly within 4 h after surgery (n = 27 patients for drug screening and an additional n = 17 patients for validation experiments). For PFS analysis of the prospective cohort, only patients with newly diagnosed glioblastoma who received radiotherapy and TMZ chemotherapy were included. The retrospective cohort (n = 18 patients) consists of patients for whom bio-banked tissue was available and who received maintenance TMZ, with OS documented as a clinical endpoint. Retrospective samples were selected to cover a broad spectrum of PFS outcomes and were further selected based on quality control measures, including cell viability, cell number and the amount of debris present in the sample.Patient sample dissociation for ex vivo drug screeningSurgically removed tissue samples were first washed with PBS and cut using single-use sterile scalpels. Subsequent dissociation was performed in reduced serum media (DMEM media, no. 41966, with 2% FBS no. 10270106, 1% pen–strep no. 15140122 and 25 mM HEPES no. 15630056, all from Gibco) supplemented with Collagenase IV (1 mg ml−1) and DNaseI (0.1 mg ml−1) using a gentle MACS Octo Dissociator (Miltenyi Biotec, 130-096-427) for maximally 40 min. Homogenates were filtered through a 70-µm cell strainer (Sigma-Aldrich, CLS431751) and washed once with PBS containing 2 mM 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), with an additional PBS wash. In case the cell pellet visibly contained a notable portion of red blood cells (RBCs), RBC lysis was performed with 1× RBC lysis buffer (BioLegend, 420301) at room temperature for 10 min before the PBS wash. Subsequently, cells were resuspended in reduced serum media, filtered once more through a 70-µm cell strainer and counted using a Countess II Automated Cell Counter (Invitrogen). In case sufficient cell numbers remained after cell seeding for ex vivo drug testing, cells were cryopreserved in 10% DMSO-containing cryopreservation media and/or cultured in DMEM media supplemented with 10% FBS, 1% pen−strep and 25 mM HEPES to obtain PDCs shortly maintained for a few passages as adherent cultures.Cell cultureThe adherent human glioblastoma cell lines LN-229 (American Type Culture Collection, CRL-2611, directly purchased from the vendor), LN-308 (obtained from the University Hospital of Zurich) and PDCs (patient IDs denoted with a ‘.C’) were cultured in standard serum media (DMEM media, no. 41966, with 10% FBS no. 10270106, 1% pen–strep no. 15140122 and 25 mM HEPES no. 15630056, all from Gibco). Adherent cell lines and PDCs were passaged using trypsin-EDTA (0.25%, Gibco, 25200056), with PDCs shortly maintained for a few passages after surgical dissociation. The spheroid human glioblastoma-initiating cell lines ZH-161 and ZH-562 were generated at the University Hospital of Zurich and cultured in Neurobasal (NB) medium (Gibco, 21103049) supplemented with B27 (Gibco, 17504044), 20 ng ml−1 b-FGF (PeproTech, AF-100-18B), 20 ng ml−1 EGF (PeproTech, AF-100-15) and 2 mM L-glutamine (Gibco, 25030081). Suspension spheroid cultures were passaged using Accutase (STEMCELL Technologies, 07920). Cell lines LN-308, ZH-161 and ZH-562 were authenticated at the Leibniz Institute DSMZ by short tandem repeat (STR) DNA analysis, whereas LN-229 was not authenticated as it was bought directly from the vendor. The LN-229 line is derived from a female patient, and LN-308, ZH-161 and ZH-562 are derived from male patients. LN-229 and LN-308 lines have methylated MGMT promoters. The LN-229 line is commercially available, and other glioblastoma cell lines/PDCs are obtainable from either the University Hospital of Zurich or the Snijder laboratory with the exception of the P022.C line that was not able to be expanded beyond five passages. All cell cultures were maintained at 37 °C, 5% CO2 in a humidified incubator.PCY (drug testing, immunocytochemistry, confocal microscopy and image analysis)The PCY method refers to high-content image-based ex vivo drug testing, including the following steps of cell seeding, drug testing, immunocytochemistry, confocal microscopy, image analysis and PCY score calculation for each tested drug32,35.Cell seeding and drug testingFreshly dissociated cells were seeded into CellCarrier 384 Ultra microplates (PerkinElmer, 6057300) typically within 4 h of surgery with 0.5–1.5 × 104 cells per well. For cultured glioblastoma cell lines and PDCs, trypsinized (adherent cultures) or accutase-treated (spheroid cultures) cells were seeded at 0.5–2.5 × 103 cells per well in 384-well plates. Before cell seeding, drugs were re-suspended as 5 mM stock solutions and dispensed into 384-well plates using an Echo 550 liquid handler (Labcyte) in a randomized plate layout to control for plate effects. Detailed information regarding drugs used in this study can be found in Supplementary Table 2. Different drug libraries included glioblastoma drugs (GSDs, n = 3 drugs), ONCDs (n = 65 drugs) and NADs (n = 67 drugs). The NAD library was based on purchasable drugs from the vendor (Sigma-Aldrich) of n = 119 CNS marketed drugs cited in Wager et al.66 and a curated list of n = 35 FDA-approved drugs for CNS indications between 2010 and 2018 after Wager et al.66 was published. All NADs were tested at 20 μM, and, for select NADs, a concentration range of 0.1–100 μM was tested (Extended Data Fig. 3a–d). GSDs were tested at the following concentrations: TMZ (first-line glioblastoma chemotherapy; 50, 100, 250 and 500 µM) and CCNU and carmustine (BCNU) (second-line glioblastoma chemotherapies; 10, 50, 100 and 250 µM). All ONCDs were tested at 10 μM concentrations. Drug plates included the following number of replicate wells per drug/concentration: GSD plate, drug, n = 3 wells, DMSO, n = 16 wells; NAD plate, drug, n = 4 wells, DMSO, n = 16–24 wells; ONCD plate, drug, n = 4 wells, DMSO, n = 16 wells. Cells were incubated for 48 h with drugs in reduced serum media at 37 °C, 5% CO2 before proceeding to cell fixation.ImmunocytochemistryCells were fixed with 4% paraformaldehyde (PFA) (Sigma-Aldrich, F8775) in PBS and blocked in 5% FBS and 0.1% Triton containing PBS. For characterization of cellular composition, cells were stained overnight at 4 °C in blocking solution with the following antibodies and dilutions: Alexa Fluor 488 anti-S100 beta (1:1,000, 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 (1:1,000, BioLegend, 422801, stock solution 10 mg ml−1). Due to the temporary manufacturer discontinuation of the Alexa Fluor 488 anti-S100 beta antibody, from patient sample P030 and onwards, samples were stained with either a self-conjugated Alexa Fluor 488 anti-S100 beta antibody, where Alexa Fluor 488 NHS Ester (Thermo Fisher Scientific, A20000) was conjugated to the anti-S100 beta antibody (Abcam, ab215989, clone EP1576Y), or the following antibody panel 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 (1:1,000, Abcam, ab274881, clone EP1576Y), PE anti-CD3 (1:300, BioLegend, 300441, clone UCHT1) and Alexa Fluor 647 anti-CD45 (1:300, BioLegend, 368538, clone 2D1).Other conjugated antibodies used included Alexa Fluor 647 anti-tubulin beta 3 (1:1,000, BioLegend, 657406, clone AA10); Alexa Fluor 488 anti-vimentin (1:500, BioLegend, 677809, clone O91D3); 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); and Alexa Fluor 594 anti-CD45 (1:300, BioLegend, 368520, clone 2D1). Other unconjugated antibodies used included anti-Connexin43 (1:500, Cell Signaling Technology, 83649T); anti-EGFR (1:300, Abcam, ab98133); anti-Nestin (1:150, BioLegend, 656802, clone 10C2); anti-S100 beta antibody (1:300, Abcam, ab215989, clone EP1576Y); and anti-Ki67 (1:300, Cell Signaling Technology, 9129S, clone D3B5). For unconjugated primary antibodies, the following secondary antibodies were used: donkey anti-sheep IgG (H + L) cross-adsorbed secondary antibody, Alexa Fluor 488 (Thermo Fisher Scientific, A11015); goat anti-mouse IgG (H + L) highly cross-adsorbed secondary antibody, Alexa Fluor Plus 555 (Thermo Fisher Scientific, A32727); and goat anti-rabbit IgG (H + L) highly cross-adsorbed secondary antibody, Alexa Fluor Plus 647 (Thermo Fisher Scientific, A32733). All secondary antibodies were used at 1:500 dilution.Confocal imaging and image analysisImaging of 384-well plates was performed with an Opera Phenix automated spinning-disk confocal microscope (PerkinElmer, HH14000000) at ×20 magnification for all assays with the exception of spheroid cell lines (ZH-161 and ZH-562) imaged at ×10 magnification. Select images were imaged at ×40 for visualization. Single cells were segmented based on their nuclei (DAPI channel) using open-source CellProfiler 2.2.0, and nuclear expansion was performed to assess cytoplasmic features, including marker expression. Downstream image analysis was performed with MATLAB R2019a–R2020a. Fractions of marker-positive cells for each sample and drug condition were based on local background-corrected intensity histograms across the whole drug plate. In patient samples, marker-positive populations were defined as follows: glioblastoma cells ((Nestin+ or S100B+) and CD45−), immune cells (CD45+ and S100B−Nestin−) and other marker-negative cells (S100B−Nestin−CD45−). Marker-positive fractions were averaged across each well/drug.PCY score calculationThe PCY score quantifies the drug-induced relative reduction of any marker-defined cell population by measuring the change of a defined target population upon drug treatment compared to DMSO vehicle control. In patient samples, the PCY score is calculated based on the fraction of ((Nestin+ or S100B+) and CD45− cells) out of all cells. In PDC lines, the score is based on (Nestin+) cells out of all cells. By all cells, we refer to any detected cell with a DAPI+ nucleus. PCY scores are averaged across technical replicates for each drug or control condition.$${\rm{PCY}}\; {\rm{score}}=1-\{[{{\rm{TP}}_{\rm{DRUG}}}]\div[{{\rm{TP}}_{\rm{DMSO}}}]\}$$where TPDRUG = fraction of the target population in a given DRUG condition of all cells and TPDMSO = fraction of the target population in the DMSO control condition of all cells.A positive PCY score of 1 represents the strongest possible ‘on-target’ response; a PCY score of 0 indicates no effect/equal cytotoxicity; and a negative PCY score indicates higher toxicity to other cell populations other than the defined target population. In cases where a target population is not defined, drug response and cell viability are measured as total cell number reduction in LN-229 and LN-308 lines and a reduction of 2D projected total spheroid area in ZH-161 and ZH-562 lines.Deep learning of apoptotic cell morphologiesTo generate a training dataset, cleaved CASP3+/− cells identified by IF and CellProfiler-based image analysis (n = 6 patient samples) were cropped as five-channel 50 × 50 pixel images around the nuclear centroid of each cell. In total, 6,072 single-cell image crops were manually curated and labeled as two classes (CASP3+/−) based on their cleaved CASP3 staining. A convolutional neural network (CNN) with a modified AlexNet architecture67 with the image input size set as 50 × 50 × 2 (two-channel bright-field (BF) and DAPI classifier) and the number of output classes set to 2 (CASP3+/−) was then trained on this manually curated image dataset (n = 6,072 single-cell images; split by a 8:2 ratio into training and test data, respectively). CNN training included use of the Adam optimizer, with a mini-batch size of 64 and a maximum number of 30 epochs. The initial learning rate was set to 0.01 with a piece-wise learning rate schedule and a drop factor of 0.1 every 10 epochs. Network performance on a manually curated test image dataset (n = 1,214 single-cell crops) is shown as a confusion matrix in Extended Data Fig. 1j. All DAPI+ nuclei detected in patient samples were retrospectively classified by this apoptotic classifier CNN based on the BF and DAPI channels to quantify apoptotic fractions across the prospective patient cohort, marker-based subpopulations and drug conditions. Cells were classified as apoptotic (CASP3+) based on a CNN confidence threshold of 87%, close to the true-positive rate of the classifier.Demonstration of PCY score robustness to apoptotic cellsWe performed ex vivo NAD (n = 67 drugs) screens in two patient samples (P048 and P049) by staining for cleaved CASP3. The drug response (Extended Data Fig. 2i–k) shows excellent reproducibility, both when comparing the original PCY scores with the PCY scores obtained after excluding CASP3+ cells by IF as well as when comparing the PCY scores after excluding CASP3+ cells defined either by IF or by the CNN apoptotic classifier. We also re-calculated the PCY scores by excluding the CNN-classified apoptotic cells measured across all 27 patient samples and 67 NADs and compared them to the original PCY scores reported in the manuscript (Extended Data Fig. 2k). The drug response correlation with or without the inclusion of apoptotic cells was 0.988, demonstrating that the PCY score is highly robust to the presence of apoptotic cells (Extended Data Fig. 2k) and can be expected to be equally robust to other forms of cell death.Targeted next-generation sequencing (Oncomine Comprehensive Assay)Formalin-fixed paraffin-embedded (FFPE) tissue blocks from patient-matched samples collected from the University Hospital of Zurich were used to determine genetic alterations. Tumor areas were marked on the hematoxylin and eosin (H&E) slide, and relative tumor content was estimated by a trained pathologist. One to three core cylinders (0.6-mm diameter) from the FFPE blocks (tumor areas) were used for DNA and RNA isolation. DNA was isolated with a Maxwell 16 FFPE Tissue LEV DNA Purification Kit (Promega, AS1130). DNA concentration was determined using a Qubit dsDNA HS Assay Kit. RNA was extracted with a Maxwell 16 FFPE Tissue LEV RNA Purification Kit (Promega, AS1260) after pre-treatment with DNase1 for 15 min at room temperature. Library preparation with 20 ng of DNA or RNA input was conducted using Oncomine Comprehensive Assay version 3. Adaptor/barcode ligation, purification and equilibration were automated with Tecan Liquid Handler (EVO-100). Next-generation sequencing (NGS) libraries were templated using Ion Chef and sequenced on an 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 version 3 Annotations version 1.4. For NGS data analysis, Torrent Suite software (Ion Reporter) 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 database68. Gene alterations described as (likely) benign were not included in the results. Non-pathogenic mutations harboring a minor allele frequency higher than 0.01 were not selected. The Default Fusion View parameter was selected. For the CNV confidence range, the default filter was used to detect gains and losses using a 5% confidence interval (CI) for minimum ploidy gain over the expected value and 95% CI for minimum ploidy loss under the expected value. CNV low-confidence range was defined for gain by copy number from 4 to 6 (minimum CNV CI 5%: 2.9) and loss from 0.5 to 1 (maximum CNV CI 95%: 2.43). High-confidence range was defined by gain up to 6 copy number (minimum CNV CI 5%: 4.54) and loss below 0.5 copy number (maximum CNV CI 95%: 1.37). The 5% and 95% CIs of all selected loss and gain are available in Supplementary Data 1. The minimum number of tiles required was 8. Results are reported as detected copy number.scRNA-seq and re-analysis of other published datasetsCryopreserved patient samples were thawed and used for scRNA-seq. Viability markers SYTOX Blue (1 μM, Thermo Fisher Scientific, S11348) and DRAQ5 (1 μM, BioLegend, 424101) were added to the cell suspension at least 15 min before sorting. Fluorescence-activated cell sorting (FACS) gates were set based on CD45 (Alexa Fluor 594 anti-CD45, 1:20, BioLegend, 368520, clone 2D1) and SYTOX Blue/DRAQ5 intensities to sort viable CD45+ and CD45− populations (Extended Data Fig. 1a) into DNA LoBind Eppendorf tubes (VWR, 525-0130) using the BD FACSAria Fusion Cell Sorter and FlowJo 10.4.2 software. CD45− and CD45+ cells were mixed at 2:1 to 10:1 ratios depending on availability to enrich for glioblastoma cells. Single-cell transcriptomes from four patient samples (P007, P011, P012 and P013), part of the prospective cohort, are referred to as ‘Lee et al.; this study’. For patient sample P024 that was used to measure the effect of vortioxetine drug treatment, cells sorted by FACS were incubated for 3 h with or without 20 µM vortioxetine before proceeding to library preparation. Libraries were generated using Chromium Next GEM Single Cell 3′ version 3.0 and version 3.1 kits (10x Genomics) and sequenced on a 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 were processed in November 2019 with the earlier version of 10x Genomics library prep kits and Cell Ranger analysis pipeline, whereas the later sample (P024) was processed in September 2021.Analysis of the cohort-matched in-house scRNA-seq datasetQuality control for the in-house dataset (Lee et al.) was performed by analyzing only high-quality cells with less than 10% of mitochondrial transcripts and genes that had at least a count of 2 in at least three cells. For the Lee et al. dataset, an expression threshold of log2(count+1) > 3 was applied to consider a gene expressed in a given cell. Uniform manifold approximation and projection (UMAP) clusters in Extended Data Fig. 1c are based on Leiden community detection, and cell types are assigned by marker expression. Glioblastoma clusters are numbered in descending order based on cluster-averaged expression of the Gene Ontology term ‘stem cell differentiation’ (GO:0048863).Re-analysis of other published scRNA-seq datasetsTo analyze additional glioblastoma patient cohorts by scRNA-seq, we used two published datasets: Neftel et al.4 and Yu et al.40. For Neftel et al., we removed cells with fewer than 29 detected genes and/or more than 15% of mitochondrial transcripts. For Yu et al. the data were already pre-filtered, but patient samples (7–9 and 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 two cells were included in the analysis. For the Neftel et al. and Yu et al. datasets, expression thresholds of log2(count+1) over 5 and 3, respectively, were applied to consider a gene expressed in a given cell. For all three scRNA-seq datasets, only patient samples with more than 50 positive cells for a given gene were considered in Fig. 1b and Extended Data Fig. 1d.Inferred CNA analysisCNAs were inferred using the ‘infercnv’ R package (version 1.18.0), using the same cell type definition in Fig. 1b and expression threshold as described above, sampling up to 70 cells per patient and cell type. ‘infercnv’ was run on the sampled cells with default settings with CD45+ immune cells across all patients set as the reference cell type. A cell was considered to have a detectable CNA if the mean ‘modified expression’ across all genes on each respective chromosome was either above a threshold of 1.1 for chromosome 7 (amplification) or below 0.9 for chromosome 10 (loss). Only patient samples that had detectable CNAs for their respective chromosomes in at least 5% of cells (combined across ‘Nestin+ or S100B+’ and ‘other’ cells) were included in the analysis presented in Fig. 1c.Cell-type-specific enrichment analysis of gene modules enriched in ‘other’ cellsTo determine putative cell types represented in Nestin−S100B−CD45− cells (‘other’) by scRNA-seq, we analyzed the log2(fold change) of ‘other’ enriched genes compared to glioblastoma cells. First, an aggregated average ‘metacell’ for each patient and subpopulation (either ‘other’ or glioblastoma cells) was created by summing the counts across each [patient-subpopulation] and dividing this by the corresponding number of cells. Next, considering only genes where the aggregate-averaged expression is above 1 in at least one ‘metacell’ type, we calculated the log2(fold change) of [‘other’ metacell] / [glioblastoma metacell] per gene and per patient. Manhattan distance-based clustering of the top 10 log2(fold change) of ‘other’ enriched genes per patient is visualized in Extended Data Fig. 1g. Dendrogram tree cutting of ‘other’ enriched genes yielded gene modules that were analyzed by WebCSEA69 to determine most likely cell types represented by the respective gene modules. The top seven most likely cell types representing each ‘other’ gene module ranked by the lowest combined P values are shown in Extended Data Fig. 1h.Neural specificity and patient specificity score analysisNeural specificity 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 Monocle70 and annotated the clusters as being either immune cells or neural cells based on known marker genes. DESingle71 analysis resulted in 11,571 neural-specific and 1,157 immune-specific genes (log2FC > 0.5). Using these cell-type-specific gene sets, we calculated an immune score and a neural score for each cell using singscore, and we classified every cell in the additional datasets as either neural or immune based on a linear combination of both scores. The ‘neural specificity score’ is defined as follows: [neural specificity = fraction of neural cells expressing gene – fraction of immune cells expressing gene] where expression of a given gene in a cell is defined as having any non-zero count. 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). For genes with low expression, this score will be close to 0, reflecting the fact that clear statements cannot be made about cell type specificity for these genes. To assess the variation of gene expression 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,:)].siRNA knockdown and quantitative real-time PCRAll siRNAs used in the study were part of the MISSION esiRNA library (Sigma-Aldrich, Euphoria Biotech; Supplementary Table 4) and ordered as custom gene arrays (esiOPEN and 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. siRNAs were transfected at 10 ng per well in 384-well plates (used for imaging and drug incubation) and 40 ng per well in 96-well plates (RNA extraction, quantitative real-time PCR (qRT–PCR)) with Lipofectamine RNAiMAX (Invitrogen, 13778075). For 384-well plates, both siRNAs and Lipofectamine were dispensed using a Labcyte Echo liquid handler in a randomized plate layout to control for plate effects when possible. For data presented in Figs. 3d and 5i and Extended Data Figs. 4d and 7j, LN-229 cells were incubated at 37 °C, 5% CO2 for 48 h after siRNA transfection before fixation, IF and RNA extraction. For Fig. 5k, after 48 h of siRNA transfection, LN-229 cells were incubated for an additional 24 h with either DMSO control or vortioxetine (10 µM) before fixing and subsequent analysis.siRNA knockdown efficiency and relative abundance for the genes BTG1, BTG2, JUN and MKI67 were measured by TaqMan Array 96-well plates (Applied Biosystems) using TaqMan Fast Advanced Master Mix (Thermo Fisher 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) and measured using a Qubit 4 fluorometer (Thermo Fisher Scientific). cDNA was synthesized with an iScript cDNA Synthesis Kit (Bio-Rad, 1708890). For each TaqMan biological replicate assay (n = 3 replicates), 25 ng 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 and GUSB) was subtracted from each [sample-target gene] Ct value to derive the deltaCt (dCt) value.COSTARCOSTAR is an interpretable molecular machine learning approach that uses 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 DTC41. 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’. ePTGs include all annotated drug bioactivities. STGs downstream 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 as 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 versus negative) 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 a cross-validation setting (Extended Data Fig. 5a). COSTAR performance was first evaluated on the training dataset, represented as a confusion matrix in Fig. 4d. Using this trained linear model, COSTAR was next used 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 Data 2). 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 Fig. 4h and Extended Data Fig. 5c. COSTAR predictions from this in silico screen were further experimentally validated ex vivo by PCY in glioblastoma patient samples (n = 4) on a set of untested drugs predicted as either COSTAR-HIT (n = 23) or COSTAR-NEG (n = 25).DRUG-seqHigh-throughput multiplexed RNA-seq was performed with the DRUG-seq method as described in Ye et al.43 with a few modifications. Oligonucleotides used for DRUG-seq are listed in Supplementary Table 5. Modifications to the published method are the following: (1) extraction of RNA before cDNA reverse transcription with the Zymo Direct-zol-96 RNA isolation kit (Zymo Research, R0256); (2) change of reverse transcription primers for compatibility with standard Illumina sequencing primers; (3) cDNA clean-up before library amplification performed with the DNA Clean & Concentrator-5 kit (Zymo Research, D4013); and (4) tagmentation performed with 2-ng input and sequencing library generated using the Nextera XT library prep kit (Illumina, FC-131-1024). In short, 1 × 104 LN-229 cells were plated in CellCarrier 96 Ultra microplates (PerkinElmer, 6055302) and incubated overnight in reduced serum media at 37 °C, 5% CO2 before drug treatment. A total of 20 drugs (Supplementary Table 2) were profiled across two different timepoints (6 h and 22 h; n = 4 replicates per drug/timepoint). These drugs included PCY-hit NADs spanning diverse drug classes (n = 11), PCY-hit ONCDs (n = 7), PCY-negative NADs (n = 2) and DMSO. Cells in drug-treated 96-well plates were lysed with TRIzol reagent (Thermo Fisher Scientific, 15596018), and then subsequent cDNA and library prep was performed as described above. Finally, 100-bp (80:20) paired-end reads were generated using Illumina’s NextSeq 2000 platform.Calcium assays on the FLIPR platformFor FLIPR calcium assays, LN-229 or P050.C cells were seeded on poly-d-lysine-coated ViewPlate-96 microplates (PerkinElmer, 6005182) in 100 µl of medium (LN-229: 70,000 cells per well; P050.C: 20,000 cells per well) 24 h before the experiment. Fluorescent Ca2+ signal was measured using the Calcium 6 assay kit (Molecular Devices, 5024048) by the FLIPR Tetra (Molecular Devices) using a 470–495-nm LED excitation module and a 515–575-nm emission filter. Calcium 6 dye stock solution was prepared in 10 ml of sterile-filtered nominal Ca2+ free (NCF) modified Krebs buffer containing 117 mM NaCl, 4.8 mM KCl, 1 mM MgCl2, 5 mM D-glucose and 10 mM HEPES (pH 7.4) stored as 500-µl aliquots 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, 50 µl of the diluted dye was applied per well followed by incubation at 37 °C for 2 h in the dark. For the assay setup outlined in Fig. 5e, cells were treated with their respective PCY-drug after a period of equilibration in 2 mM calcium-containing buffer. For fold change calculations presented in Fig. 5f and Extended Data Fig. 7b, normalized calcium levels for each drug were calculated by averaging calcium levels after drug treatment (400–600-s interval) divided by the basal level of calcium before drug administration (200–300-s interval). In the ER Ca2+ store release assay, stable baselines were established for 50 s before 50 µl of 2 µM (2×) thapsigargin (Sigma-Aldrich, T9033) or 40 µM (2×) drug solutions freshly prepared in NCF Krebs buffer were robotically dispensed. 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, 50 µl of 4 mM CaCl2 (2×) prepared in NCF Krebs buffer was dispensed onto the cells to re-establish a physiological 2 mM calcium concentration, and the fluorescence was monitored for 5 min. Next, 60 µM (3×) drug solutions freshly prepared in Krebs buffer containing 2 mM CaCl2 were robotically dispensed, and fluorescence was recorded for an additional 4 min. The raw data were extracted with ScreenWorks software version 3.2.0.14. The values represent average fluorescence level of the Calcium 6 dye measured over arbitrary selected and fixed timeframes.Calcium imaging using the Fura-2 calcium indicatorGlioblastoma cell lines (LN-229 and LN-308) and PDCs (P024.C, P040.C, P049.C and P050.C lines) were seeded in six-channel µ-Slide VI 0.4 ibiTreat (ibidi, 80606), with 30,000–100,000 cells per channel and up to three channels per slide. Seeded cells were cultured in these chamber slides 1–2 d before the experiment to achieve approximately 70–80% confluency. Before dye loading of the Fura-2 AM calcium indicator (Thermo Fisher Scientific, F1221), cells were washed two times with HEPES-buffered Krebs-Ringer Solution (referred to as Krebs buffer; Thermo Fisher Scientific, 67795.K2). Cell permeant Fura-2 dye resuspended in DMSO was incubated with cells (1 μM solution in Krebs buffer) for 15 min at 37 °C, 5% CO2 in a dark humidified incubator and washed three times with Krebs buffer before imaging. All subsequent calcium imaging and drug perfusion were performed in Krebs buffer.Live-cell calcium imaging was performed at ×20 magnification (S Fluor ×20 NA 0.75 objective) on a Nikon Ti2-E inverted microscope equipped with a Nikon DS-FI3 color camera (2,880 × 2,048 pixels, 2.4 μm × 2.4 μm), color BF camera, motorized fast emission filter wheels (Sutter Instrument) and a FURA dichroic mirror. FURA filterset specifications include: LED 1 (excitation window 1), 340/26; LED 2 (excitation window 2), 387/11; and an emission filter, 510/84. 2 × 2 binned images were acquired every 2 s throughout an imaging time of 10 min per experiment. CO2 levels and temperature were controlled by an Okolab box type incubation system. Vortioxetine (20 μM solution) was manually administered on the chamber slide. Timepoint of drug addition was, on average, between 125 s and 140 s after the start of imaging. Downstream image analysis was performed with ImageJ and R. In ImageJ, circular regions of interest (ROIs) were manually selected for each cell present in the first image frame of each experiment’s time series as well as five background ROIs to calculate the mean background intensity. For both the 340-nm and 380-nm channels, mean pixel intensities across each cell ROI and image frame were measured. Subsequently in R, mean background intensity was subtracted from each cell ROI before further downstream analysis. Cell ROIs with more than five image timeframes exhibiting a signal lower than background (lower one percentile of Fura-2 intensities across cells in the first 30 s of imaging) were excluded from the analysis. Timepoint of vortioxetine addition was determined either by outlier detection or by manual inspection between 120 s and 150 s after the start of imaging, and this single timeframe was assigned to N/A (not applicable) to exclude the possibility of imaging artifacts impeding the analysis. The raw Fura-2 calcium signal was defined as the ratio of 340/380 intensity. The mean change in calcium signal after vortioxtine treatment was defined as the baseline signal before drug treatment subtracted from the calcium signal after vortioxetine treatment, each averaged across a 120-s time window. Normalized Fura-2 calcium signal corresponds to the baseline signal subtracted from the raw signal on a cell ROI basis. The presence (Ψ) or absence (Ø) of oscillatory calcium signaling was determined by peak detection analysis. If a cell ROI had more than one or two peaks detected within its respective time span (baseline versus after vortioxetine drug treatment), the response type was assigned as oscillatory.ElectrophysiologyLN-229 and LN-308 glioblastoma cell lines were seeded at approximately 40% confluence in 35-mm Petri dishes (CLS430165, Corning). Whole-cell patch-clamp recordings were performed with a HEKA EPC10 USB amplifier using the following solutions: extracellular (in mM): 140 NaCl, 2 MgCl2, 2 CaCl2, 10 HEPES, 3 KCl, 10 D-glucose, pH 7.4; pipette (in mM): 4 NaCl, 120 K-gluconate, 10 HEPES, 10 EGTA, 3 Mg-ATP, 0.5 CaCl2, 1 MgCl2, pH 7.2 (liquid junction correction 12 mV). Patch pipettes (~10 MOhm) were pulled from borosilicate glass capillaries (Harvard Apparatus, 30-0038) using a two-step vertical pipette puller PC 100 (Narishige) and further fire-polished using a homemade microforge. Membrane voltage was measured during 10 s (current-clamp mode), and currents elicited upon changes in voltage (voltage-clamp mode) were assessed by keeping cells at −50 mV for 300 ms, followed by stepwise increments of +20 mV during 1,000 ms (−120 mV to +100 mV) and ending with −50 mV for 300 ms. Current-clamp and voltage-clamp protocols were executed automatically every minute during the experiment. Cells were kept at their respective membrane voltage (voltage clamp) in between protocols. For every cell, a 5-min control period was recorded after achieving whole cell followed by a 10-min recording with vortioxetine, 10 μM treatment. Average steady-state current and membrane voltage were calculated during 80% of recorded time. A linear mixed-effects model was fitted by: ‘Current density ~ Command voltage (Vcmd)* Condition (Cond.) + (1|Cell ID)’ to assess how command voltage and condition influence current density. Summary statistics are reported in Extended Data Fig. 7i.Incucyte live-cell imagingIn total, 2.5 × 103 LN-229 cells per well were plated in CellCarrier 96 Ultra microplates (PerkinElmer, 6055302) 24 h before the experiment and transfected with BTG1, BTG2 and FLUC (−) MISSION esiRNAs (Sigma-Aldrich, Euphoria Biotech, 40 ng per well) using Lipofectamine RNAiMAX (Invitrogen, 13778075). Real-time confluence of cell cultures (n = 4 replicate wells per condition) was monitored by imaging every 2 h for 7 d at ×10 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).Timecourse RNA-seq library preparation and sequencingLN-229 cells were seeded at 2 × 105 cells per well in six-well Nunc Cell-Culture Treated Multidishes (Thermo Fisher Scientific, 140675) and incubated overnight in reduced serum media at 37 °C, 5% CO2 before drug treatment. The following day, vortioxetine (AvaChem Scientific, 3380) was manually added to each well at a final concentration of 20 µM. At the start of the experiment, LN-229 cells that were not treated with vortioxetine were collected as the 0-h timepoint. After 3, 6, 9, 12 and 24 h following vortioxetine treatment, drug-containing media were removed, and cells were collected in TRIzol reagent (Thermo Fisher Scientific, 15596018). Total RNA was isolated using Direct-zol RNA MicroPrep Kit (Zymo Research, R2062), and RNA quality and quantity were determined with an Agilent 4200 TapeStation. Sample RNA integrity number (RIN) scores ranged from 5.9 to 10 (mean RIN, 9.33). RNA input was normalized to 300–400 ng, and RNA libraries were prepared using the Illumina TruSeq stranded mRNA library prep. Then, 100-bp 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 analysis, was performed with the R package ‘DESeq2’72.Timecourse proteomics and phosphoproteomicsCell preparation and vortioxetine treatment were performed as in the ‘Timecourse RNA-seq library preparation and sequencing’ subsection except that cell numbers were scaled to be seeded in T-150 culture flasks, and three timepoints were measured (0 h, 3 h and 9 h). Peptides were prepared using the PreOmics iST kit on the PreON (HSE AG) programmed to process eight samples in parallel. Cell pellets were resuspended in 50 µl of lysis buffer and denatured for 10 min at 95 °C, followed by 3 h of digestion with trypsin and Lys-C. Peptides were dried in a speed-vac (Thermo Fisher Scientific) for 1 h before being resuspended in LC-LOAD buffer at a concentration of 1 μg μl−1 with iRT peptides (Biognosys).Samples were analyzed on an Orbitrap Lumos mass spectrometer equipped with an Easy-nLC 1200 (both Thermo Fisher Scientific). Peptides were separated on an in-house packed 30-cm RP-HPLC column (Michrom Bioresources, 75 μm i.d. × 30 cm; Magic C18 AQ 1.9 μm, 200 Å). Mobile phase A consisted of HPLC-grade water with 0.1% formic acid (FA); mobile phase B consisted of HPLC-grade acetonitrile (ACN) (80%) with HPLC-grade water and 0.1% (v/v) FA. Peptides were eluted at a flow rate of 250 nl min−1 using a nonlinear 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 396–1,005 was covered. The DIA isolation window size was set to 8 m/z 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 1,200%. Higher-energy collisional dissociation (HCD) fragmentation was set to 30% normalized collision. Full mass spectra 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 version 14 (Biognosys). MS1 values were used for quantification, and peptide quantity was set to mean. Data were filtered using q value sparse with a precursor and a protein q value cutoff of 0.01 FDR. Interference correction and local cross-run normalization was performed. For PRM measurements, peptides were separated by reverse-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−1 with a 117-min nonlinear gradient from 4% to 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 × 105). Normalized HCD collision energy was set to 28% and maximum injection time to 118 ms. Monitored peptides were analyzed in Skyline version 20, and results were uploaded to PanoramaWeb.For phosphopeptide enrichment, protein lysate from LN-229 cells was prepared using a deoxycholate-based buffer. Five hundred micrograms of vortioxetine-treated cells at each timepoint (n = 3 replicates) were used as starting material. A tryptic digest was performed for 16 h. Samples were then purified on MACROSpin C18 columns (Harvard Apparatus). Phosphopeptides were specifically enriched using IMAC cartridges on the Bravo AssayMAP liquid handling platform (Agilent). Samples were dissolved in 160 μl of loading buffer (80% ACN, 0.1% trifluoroacetic acid (TFA)). Then, the AssayMAP phosphoenrichment protocol was performed with slight modifications. After purification, dried peptides were resuspended in LC buffer and subjected to DDA-MS on a Q Exactive H-FX mass spectrometer equipped with an Easy-nLC 1200 (both Thermo Fisher Scientific). Peptides were separated on an ES903 column (Thermo Fisher Scientific, 75 μm i.d. × 50 cm; particle size 2 μm). Mobile phase A consisted of HPLC-grade water with 0.1% FA; mobile phase B consisted of HPLC-grade ACN (80%) with HPLC-grade water and 0.1% (v/v) FA. Peptides were eluted at a flow rate of 250 nl min−1 using a nonlinear gradient from 3% to 56% mobile phase B in 115 min. MS1 spectra were acquired at a resolution of 60,000 with an AGC target value of 36 and a maximum injection time of 56 ms. The scan range was between 350 m/z and 1,650 m/z. A data-dependent top 12 method was used with a precursor isolation window of 1.3 m/z. MS/MS scans were acquired with normalized collision energy of 27 at a resolution of 15,000. AGC target was 15 with a maximum injection time of 22 ms. Dynamic exclusion was set to 30 s. Data analysis was performed using FragPipe (version 19.1) with the LFQ-phospho workflow73. Min site localization probability was set to 0.75 in IonQuant74. Statistical analysis was performed on the phosphoprotein-filtered combined protein output in FragPipe-Analyst.Clonogenic survival assayAdherent cells (LN-229: 50 cells; LN-308: 300 cells) were seeded in 96-well plates (n = 6 wells per condition; 100 µl of medium) and incubated overnight. On the following day, medium was replaced by fresh medium containing indicated final concentrations of vortioxetine or DMSO. Glioblastoma-initiating cells (ZH-161 and ZH-562; 500 cells) were seeded in 75 µl of medium and incubated overnight. Treatment was initiated by addition of 75 µl of medium containing 2× concentrated vortioxetine or DMSO to reach indicated final concentrations. DMSO concentration was kept at 0.5% for all treatments and controls. After treatment addition, cells were cultured for 11 d (LN-229) to 13 d (other cell lines), and clonogenic survival was estimated from a resazurin-based assay75 using a Tecan M200 PRO plate reader (λEx = 560 nm / λEm = 590 nm).Collagen-based spheroid invasion assaySpheroid invasion assay was performed as described (Kumar et al.76). In brief, 2,000 cells were seeded cell-repellent 96-well U-bottom plates (Greiner Bio-One, 650979, n = 6 wells per condition) and incubated for 48 h to allow spheroid formation. Subsequently, 70 µl of medium was removed, and spheroids were overlaid with 70 µl of 2.5% Collagenase IV (Advanced Biomatrix, 5005-B) in 1× DMEM containing sodium bicarbonate (Sigma-Aldrich, S8761), and collagen was solidified in the incubator for 2 h. Collagen-embedded spheroids were then overlaid with 100 µl of chemoattractant (NIH-3T3-conditioned medium) containing 2× concentrated vortioxetine/DMSO (0.5% final DMSO concentration across conditions) and incubated for 36 h. Spheroids were stained with Hoechst, and images were acquired on a MuviCyte imaging system (PerkinElmer, HH40000000) using a ×4 objective. Images were contrast enhanced and converted to binary using ImageJ/Fiji and quantified with automated Spheroid Dissemination/Invasion counter software (aSDIcs), which quantifies the migration distance from the center of the spheroid for each detected cell nucleus76.In vivo drug testingAll animal experiments were performed under the guidelines of Swiss federal law on animal protection and were approved by the cantonal veterinary office (ZH98/2018). CD1 female nu/nu mice (Janvier) of 6–12 weeks of age were used in all experiments, and 100,000 LN-229-derived or 150,000 ZH-161-derived cells were implanted77. Mice were euthanized when they exhibited neurological symptoms or a mouse grimace scale score of 2 (ref. 78). We confirm that these criteria were not exceeded. Mice were housed in groups of five mice per cage in the animal facility of LASC Zurich and kept in transparent plastic Eurostandard Type III cages measuring 425 × 266 × 155 mm. The cages contained autoclaved, dust-free sawdust bedding (80–90 g per cage) and one Nestlet (5 × 5 cm). The mice were fed a pelleted and extruded Kliba No. 3436 mouse diet (Provimi Kliba) ad libitum and had unrestricted access to sterilized drinking water. The room maintained a 12-h light/dark cycle with artificial light. The temperature was 21 ± 1 °C, and the relative humidity was 50 ± 5%.Test-naive mice were randomly assigned to drug treatment groups for experiments (in vivo drug treatment Trials I–V). Tumor-bearing mice were treated from day 5 to day 21 after tumor implantation with intraperitoneally administered vortioxetine daily 10 mg kg−1, citalopram daily 10 mg kg−1, paliperidone daily 5 mg kg−1, apomorphine daily 5 mg kg−1, aprepitant daily 20 mg kg−1, brexpiprazole daily 1 mg kg−1, chlorpromazine three times per week 10 mg kg−1, TMZ 50 mg kg−1 for five consecutive days, CCNU 20 mg kg−1 at day 7 and day 14 after tumor implantation or daily DMSO control. MRI was performed with a 4.7T imager (Bruker BioSpin) when the first mouse became symptomatic for in vivo Trials I–III or a 7T imager (Bruker BioSpin) at days 12, 25, 38 and 48 after tumor implantation for in vivo Trial IV. Coronal T2-weighted images were acquired using ParaVision 360 (Bruker BioSpin). Tumor regions were identified manually by two independent raters, and maximum perimeter was outlined and quantified using MIPAV (11.0.7).For immunohistochemistry analysis, mouse brains were embedded in Shandon Cryochrome (Thermo Fisher Scientific) and were cut horizontally by 8-μm steps until reaching the tumor. Tissue sections were stained for 1 s with 0.4% methylene blue and rinsed with deionized water (2 × 10 dips) to confirm tumors (when present) under the microscope. Sections were stored in dark dry boxes overnight before being stored at −80 °C. Sections were fixed with 4% PFA (Sigma-Aldrich, F8775) in PBS, blocked in 5% FBS and 0.1% Triton containing PBS and stained overnight at 4 °C in blocking solution with DAPI and the following antibodies and dilutions: Alexa Fluor 488 anti-vimentin (1:500, BioLegend, 677809, clone O91D3), anti-Ki67 (1:300, Cell Signaling Technology, 9129S, clone D3B5) and goat anti-rabbit IgG (H + L) highly cross-adsorbed secondary antibody, Alexa Fluor Plus 647 (1:500, Thermo Fisher Scientific, A32733). Imaging was performed by ×20 fluorescence imaging using the Pannoramic 250 slide scanner (3DHISTECH).Statistical analysisFor prospectively sampled patient material, no sample size determination was performed a priori as the effect size and variability of ex vivo drug response among patients were unknown before the study. Our sample sizes are similar to other published glioblastoma studies investigating the heterogeneity of patient samples and/or patient-derived explants4,38,60. For all other statistical analysis, their respective tests and significance values are reported in each corresponding figure panel and/or Methods. For linear correlations, Pearson correlation coefficients with two-tailed P values are annotated. When the Student’s t-test was used for comparisons between groups (for example, drug treatment versus control), data distribution was assumed to be normal, but this was not formally tested. The Wilcoxon test was also used as a non-parametric equivalent. For matched patient samples or cells, paired t-tests or Wilcoxon tests were used. Unless otherwise stated, multiple testing correction was performed using either the Holm method for comparisons of fewer than 20 data points or the FDR procedure for larger datasets.Reporting summaryFurther information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
