Glioblastoma remodelling of human neural circuits decreases survival – Nature

Patients, samples, consent and ethics approval

Each participant in the study was recruited from a prospective registry of adults aged 18–85 with newly diagnosed frontal, temporal and parietal IDH-WT high-grade gliomas with detailed language assessments and baseline MEG recordings. Inclusionary criteria included the following: native English speaking, aged 18–85 years, and no previous history of psychiatric illness, neurological illness, or drug or alcohol abuse. All human electrocorticography data were obtained during lexical retrieval language tasks from 14 adult awake patients undergoing intraoperative brain mapping for surgical resection. Tumours from eight patients were used for RNA-seq experiments. Site-directed tumour biopsies from 19 patients were used for immunofluorescence/immunohistochemistry analysis and 24 patients were used for immunocytochemistry and cell-based functional assays. Tumours from eight patients were used for mouse xenograft experiments. All of the participants provided written informed consent to participate in this study, which was approved by the University of California, San Francisco (UCSF) institutional review board (IRB) for human research (UCSF CC-171027, CHR 17-23215) and performed in accordance with the Declaration of Helsinki.

We began by studying short-range circuit dynamics in a subset of 14 patients with dominant hemisphere glioblastoma infiltrating speech production areas of the inferior frontal lobe using ECoG in the intraoperative setting (Extended Data Fig. 1a). We then focused molecular studies on patients with surgically treated IDH-WT glioblastoma, performed extraoperative language assessments and imaginary coherence as a long-range measure of functional connectivity using MEG (Extended Data Fig. 1a,b). This enabled us to import functional connectivity data into the operating room in which we performed site-specific tissue biopsies of human glioma from regions with differing measures of functional connectivity for in vivo and in vitro cell biology experiments including primary patient cultures (n = 19 patients) and multimodal tissue profiling, including microscopy, sequencing, proteomics and patient-derived tumour xenografting (Extended Data Fig. 1c). This layered approach—combining clinical variables, cognition assessments, human and animal models of network dynamics, in addition to cell biology—served as a platform to study the clinical implications of glioma–neuron interactions (Extended Data Fig. 1d and Supplementary Tables 1 and 2).

Human ECoG and data analyses

The hemisphere of language dominance was determined using baseline magnetic source imaging. In brief, the participants sat in a 275-channel whole-head CTF Omega 2000 system (CTF Systems) sampling at 1,200 Hz while they performed an auditory-verb generation task. The resulting time series were then reconstructed in source space with an adaptive spatial filter after registration with high-resolution MRI. Finally, changes in beta-band activity during verb generation were compared across hemispheres to generate an overall laterality index. All of the participants were left-dominant and underwent electrophysiological recording of the left hemisphere. We implemented an intraoperative testing paradigm that was previously established⁹. Noise in the operating room was minimized through rigorous enforcement of the following: (1) all personnel were requested to cease verbal communication; (2) telephones and alarms were muted; and (3) surgical suction and all other non-essential machinery were temporarily shut down. A 15 inch laptop computer (60 Hz refresh rate) running a custom MATLAB script integrated with PsychToolbox 3 (http://psychtoolbox.org/) was placed 30 cm away from each participant. The script initiated a picture-naming task that consisted of a single block of 48 unique stimuli, each depicting a common object or animal through coloured line drawings. Each stimulus was presented at the point of central fixation and occupied 75% of the display. After presentation of each stimulus, the participants were required to vocalize a single word that best described the item.

Intraoperative photographs with and without subdural electrodes present were used to localize each electrode contact combined with stereotactic techniques^9,53. Images were registered using landmarks from gyral anatomy and vascular arrangement to preoperative T1- and T2-weighted MRI scans. Tumour boundaries were localized on MRI scans and electrodes within 10 mm of necrotic tumour core tissue were identified as ‘tumour’ contacts. Electrodes overlying the hypointense core of the tumour extending from the contrast enhancing rim to the edge of FLAIR were considered to be tumour electrodes, and electrodes completely outside of any T1 post gadolinium or FLAIR signal were considered to be non-tumour or normal appearing by a trained co-author blinded to the electrophysiologic data². Glioma-infiltrated regions were defined on the basis of two criteria previously established in the literature⁹, including mass-like region of T2-weighted FLAIR sequences signal. Imaging was confirmed by gross inspection of the cortex confirming dilation and/or an abnormal vascular pattern. Previous research has shown that regions of non-enhancing disease consist of infiltrating tumour cells intermixed with neurons and normal glial cells^2,54. These labels were reviewed by the study principal investigator (S.L.H.-J.) and compared to labels derived during intraoperative stereotactic neuronavigation to reach a consensus (Brainlab).

Each participant received a training session 2 days before participation to ensure familiarity with the task. ECoG signals were acquired during a period after stopping the administration of anaesthetics (minimum drug wash out period of 20 min) and the patient was judged to be alert and awake after an extensive post-emergence wakefulness assessment to ensure adequate arousal⁵⁵. Intraoperative tasks consisted of naming pictorial representations of common objects and animals (picture naming) and naming common objects and animals through auditory descriptions (auditory naming)⁵⁶. Post-operative videos were reanalysed to ensure that all data were collected and correct responses only were included for analysis. Audio was sampled at 44.1 kHz from a dual-channel microphone placed 5 cm from the participant and electrophysiological signals were amplified (g.tec). Recordings were acquired at 4,800 Hz and downsampled to 1,200 Hz during the initial stages of processing. During offline analyses, audio and electrophysiological recordings were manually aligned, resampled and segmented into epochs (speech-locked). These epochs set time = 0 ms as speech onset and included ±2,000 ms for a total of 4,000 ms of signal per trial. Trials were discarded if (1) an incorrect response was given (including fillers and interjections) or (2) there was a greater than 2 s delay between stimulus presentation and response so as to maintain consistent trial dynamics and ensure that the neural signal indeed reflected the experimental manipulations. Channels with excessive noise artifacts were visually identified and removed if their kurtosis exceeded 5.0. After the rejection of artifactual channels, data were referenced to a common average, high-pass filtered at 0.1 Hz to remove slow-drift artifacts, and bandpass filtered between 70–110 Hz using a 300-Order FIR filter to focus the analyses on the high-gamma band range, which is strongly related to local mean population spiking rates. To extract the ERSPs, electrophysiological signals were first downsampled to 600 Hz, then high-pass filtered at 0.1 Hz to remove DC-offset and low-frequency drift, notch-filtered at 60 Hz and its harmonics to remove line noise, and bandpass-filtered between 70 and 170 Hz (that is, the high-gamma range) using a Hamming windowed sinc FIR filter. These signals were finally smoothed using a 100 ms Gaussian kernel, downsampled to 100 Hz and z-scored across each trial. Electrodes were subsequently rereferenced to the common average for each participant to facilitate group comparisons, and regions of interest were defined according to the Automated Anatomical Labelling Atlas (https://www.gin.cnrs.fr/en/tools/aal/). The location of grid implantation was solely directed by clinical indications. The accuracy of the final registration for each participant was independently confirmed using gyral and sulcal anatomy to triangulate the location of each electrode registered to the template surface and was then compared to intraoperative photographs of the actual cortex with the overlying grid(s)⁵⁷. The HGp was then calculated using the square of the Hilbert transform on the filtered data. The HGp was then averaged across the resting-state time series, yielding a single measure of neural responsivity for each electrode contact. The HGp was then averaged across patients during the task response period, yielding a single measure of neuronal responsivity for each channel. The HGp levels were then compared between tumour and normal appearing channels. Linear mixed-effects modelling was used to perform statistical comparisons with repeated measures using the nlme package in R (v.3.1-161; https://cran.r-project.org/web/packages/nlme/citation.html). The signal’s origin (that is, normal-appearing/glioma-infiltrated cortex) was modelled as a fixed effect and the participants were modelled as random effects. For continuous variables without repeated measures, t-tests were used. A threshold of P < 0.05 was used to denote statistical significance and corrections for multiple comparisons were made using the Bonferroni method.

To decode between low-frequency words (for example, rooster) and high-frequency words (for example, car), signals from normal-appearing and glioma-infiltrated electrodes were extracted from the anterior temporal lobe after participant-level registration to a common MNI atlas. Responses were time-locked to speech onset and the signal envelope was extracted using a Hilbert transform after applying a bandpass filter in the high-gamma range (70–170 Hz). Subsequently, an l2-regularized logistic regression classifier was trained (cost of 1) to distinguish neural responses during vocalization of low-frequency words (for example, rooster) from high-frequency words (for example, car). Model performance was determined by taking the accuracy on a held-out participant and averaging it across all folds (that is, leave-one-participant-out cross-validation) and statistical significance was determined by testing this accuracy against a binomial distribution. This process was conducted separately for normal-appearing and glioma-infiltrated cortex using an identical preprocessing, training and testing paradigm.

MEG recordings and data analysis

MEG recordings were performed according to an established protocol^30,31. In brief, the study participants had continuous resting state MEG recorded with a 275-channel whole-head CTF Omega 2000 system (CTF Systems) using a sampling rate of 1,200 Hz. During resting-state recordings, the participants were awake with their eyes closed. Surface landmarks were co-registered to structural magnetic resonance images to generate the head shape. Within the alpha frequency band, an artifact-free 1 min epoch was selected for further analysis if the patient’s head movement did not exceed 0.5 cm. This artifact-free, 1 min epoch was then analysed using the NUTMEG software suite (v.4; UCSF Biomagnetic Imaging Laboratory) to reconstruct whole-brain oscillatory activity from MEG sensors so as to construct functional connectivity (imaginary coherence (IC)) metrics^54,58,59. Spatially normalized structural magnetic resonance images were used to overlay a volume-of-interest projection (grid size = 8 mm; approximately 3,000 voxels per participant) such that each voxel contained the entire time series of activity for that location derived by all the MEG sensor recordings. The time series within each voxel was then bandpass-filtered for the alpha band (8–12 Hz) and reconstructed in source space using a minimum-variance adaptive spatial filtering technique^54,60. The alpha frequency band was selected because it was the most consistently identified peak in the power spectra from this sampling window in our patient series. Functional connectivity estimates were calculated using IC, a technique known to reduce overestimation biases in MEG data generated from common references, cross-talk and volume conduction^26,28.

Resting-state MEG was also used to measure intratumoural gamma activity. A spatial beamformer was applied to extract neural signals at the voxel level from manually defined regions of interest corresponding to FLAIR signal abnormality (that is, within the infiltrative margin of the tumour)⁶¹. These source-space signals were then downsampled to 300 Hz, notch filtered at 60 Hz to remove line noise and rereferenced to the common average. Spectral activity from 1 to 50 Hz was estimated at each voxel using Thomson’s multitaper method (pmtm in MATLAB R2021b) with 29 Slepian tapers. Next, gamma power from 30 to 50 Hz was computed at each voxel after subtracting the aperiodic component from each spectrum by fitting a Lorentzian function in semi-log space⁶². A point estimate of intratumoural gamma activity was subsequently computed by averaging the activity across all voxels for each participant and regressed against the corresponding number of manually counted intratumoural HFC nodes.

Functional connectivity map

The functional connectivity of an individual voxel was derived by the mean IC between the index voxel and the rest of the brain, referenced to its contralesional pair³⁰. It is possible that there are regions within gliomas with varying amounts of functional connectivity. Moreover, there are individual patients with more or less functional connectivity. We have addressed these differences in our experimental model. Intratumoural differences in functional connectivity were addressed by the following: in comparison to contralesional voxels, we used a two-tailed t-test to test the null hypothesis that the Z-transformed connectivity IC between the index voxel and non-tumour voxel is equal to the mean of the Z-transformed connectivity between all contralateral voxels and the same set of voxels. The resultant functional connectivity values were separated into tertiles: upper tertile (HFC) and lower tertile (LFC). Functional connectivity maps were created by projecting connectivity data onto each individual patient’s preoperative structural magnetic resonance images and imported into the operating room neuronavigation console. Stereotactic site-directed biopsies from HFC (upper tertile) and LFC (lower tertile) intratumoural regions were taken and x, y, z coordinates determined using Brainlab neuro-navigation. Thus, only the extremes of intratumoural connectivity (high and low connectivity, HFC and LFC, respectively) were analysed for these experiments. Rather than raw values, each functional connectivity measure represents a Z-transformed value and it therefore remains likely that the HFC distinction for one patient does not perfectly coincide with the HFC distinction in another patient’s tumour (intertumoural heterogeneity).

Measurement of tumour volume and calculation of volumetric extent of resection

Pre-operative and post-operative tumour volumes were quantified using BrainLab Smartbrush (v.2.6; Brainlab). Pre-operative MRI scans were obtained within 24 h before resection, and post-operative scans were all obtained within 72 h after resection. Total contrast-enhancing tumour volumes were measured at both pre-operative and post-operative timepoints. The total contrast-enhancing tumour volume was measured on T1-weighted post-contrast images, and the non- enhancing tumour volume was measured on T2 or FLAIR sequences. Manual segmentation was performed with region-of-interest analysis ‘painting’ inclusion regions based on fluid-attenuated inversion-recovery (FLAIR) sequences from pre- and post-operative MRI scans to quantify tumour volume. The extent of resection was calculated as follows: (pre-operative tumour volume − post-operative tumour volume)/pre-operative tumour volume × 100%. Manual segmentations were performed for which the tumour volumetric measurements were verified for accuracy after an initial training period. Volumetric measurements were performed blinded to patients’ clinical outcomes. All of the patients in the cohort had available preoperative and postoperative MRI scans for analysis. To ensure that post-operative FLAIR signal was not surgically induced oedema or ischaemia, FLAIR pre- and post-operative MRIs were carefully compared alongside DWI sequences before including each region in the volume segmentation⁴². HFC voxels with T1 post gadolinium contrast enhancing tumour were considered to be HFC-positive for survival analysis.

Language assessments

One to two days before tumour resection, patients underwent baseline language evaluation, which consisted of naming pictorial representations of common objects and animals (picture naming) and naming common objects and animals through auditory descriptions (auditory naming). Visual picture naming and auditory stimulus naming testing were used given their known significance and clinical correlation with outcomes in clinical patient population^63,64. The correct answers for these tasks (delivered on a laptop with a 15 inch monitor (60 Hz refresh rate) positioned two feet away from the seated patient in a quiet clinical setting) were matched on word frequency (that is, commonality within the English language) using SUBTLEX_WF scores provided by the Elixcon project and content category. Task stimuli were randomized and presented using PsychToolbox. The task order was randomly selected by the psychometrist for each participant. Slides were manually advanced by the psychometrist either immediately after the participant provided a response or after 6 s if no response was given. The tasks were scored on a scale from 0 to 4 by a trained clinical research coordinator who was initially blinded to all clinical data (including imaging studies). No participants had uncorrectable visual or hearing loss. Details of the administration and scoring of auditory and picture naming language tasks can be found in previous studies^27,55,65.

Isolation and culture of primary patient-derived glioblastoma cells

Tumour tissues with high (HFC) and low (LFC) functional connectivity sampled during surgery based on preoperative MEG were processed for quality control by a certified neuropathologist and were subsequently used to generate primary patient-derived cultures. Patient-matched samples were acquired from site-directed HFC and LFC intratumoural regions from the same patient. Intratumoural HFC and LFC tissues were dissociated both mechanically and enzymatically and then passed through a 40 µm filter to remove debris. The filtered cell suspension was then treated with ACK lysis buffer (Invitrogen) to remove red blood cells and subsequently cultured as free-floating neurospheres in a defined, serum-free medium designated tumour sphere culture medium, consisting of Dulbecco’s modified Eagle’s medium (DMEM-F12; Invitrogen), B27 (Invitrogen), N2 (Invitrogen), human-EGF (20 ng ml⁻¹; Peprotech), human-FGF (20 ng ml⁻¹; Peprotech). Normocin (InvivoGen) was also added to the cell culture medium in combination with penicillin–streptomycin (Invitrogen) to prevent mycoplasma, bacterial and fungal contaminations. Cell cultures were routinely tested for mycoplasma (PCR Mycoplasma Test Kit I/C, PromoCell) and no positive results were obtained (Extended Data Fig. 7d).

Bulk RNA-seq and analysis

RNA was isolated from HFC (n = 3) and LFC (n = 4) tumour samples using the RNeasy Plus Universal Mini Kit (QIAGEN) and RNA quality was confirmed using the Advanced Analytical Fragment Analyzer. RNA-seq libraries were generated using the TruSeq Stranded RNA Library Prep Kit v2 (RS-122- 2001, Illumina) and 100 bp paired-end reads were sequenced on the Illumina HiSeq 2500 system to at least 26 million reads per sample at the Functional Genomics Core Facility at UCSF. Quality control of FASTQ files was performed using FASTQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Reads were trimmed with Trimmomatic (v.0.32)⁶⁶ to remove leading and trailing bases with quality scores of less than 20 as well as any bases that did not have an average quality score of 20 within a sliding window of 4 bases. Any reads shorter than 72 bases after trimming were removed. Reads were subsequently mapped to the human reference genome GRCh38 (https://www.ncbi.nlm.nih.gov/assembly/GCF_000001405.39/)⁶⁷ using HISAT2⁶⁸ (v.2.1.0) with the default parameters. For differential expression analysis, we extracted exon-level count data from the mapped HISAT2 output using featureCounts⁶⁹. Differentially expression analysis was performed using DESeq2⁷⁰ using the apeglm parameter⁷¹ to accurately calculate log-transformed fold changes and setting a false-discovery rate of 0.05. Differentially expressed genes were identified as those with log-transformed fold changes of greater than 1 and an adjusted P value of less than 0.05. Unsupervised gene expression principal component analysis and volcano plots of IDH-WT glioblastoma (Extended Data Fig. 5b,c) revealed 144 differentially expressed genes between HFC and LFC tumour regions, including 40 genes involved in nervous system development (Supplementary Table 3).

Single-cell sequencing

Single-cell suspension generation

Fresh tumour samples were acquired from the operating room and transported to the laboratory space in PBS and on ice. Tumour tissue was minced with #10 scalpels (Integra LifeSciences) and then digested in papain (Worthington Biochemical, LK003178) for 45 min at 37 °C. Digested tumour tissue was then incubated in red blood cell lysis buffer (eBioscience, 00-4300-54) for 10 min at room temperature. Finally, the samples were sequentially filtered through 70 μm and 40 μm filters to generate a single-cell suspension.

Single-cell sequencing and analysis

Single-cell suspensions of three patient-matched HFC and LFC tumour tissues were generated as described above and processed for single-cell RNA-seq using the Chromium Next GEM Single Cell 3′ GEM, Library & Gel Bead Kit v3.1 on the 10x Chromium controller (10x Genomics) using the manufacturer’s recommended default protocol and settings, at a target cell recovery of 5,000 cells per sample. Although single-cell sequencing does not capture all cell types within the central nervous system microenvironment, the sequencing pipeline used in this study has been demonstrated to identify neurons and was therefore chosen for use in physiologically annotated fresh glioblastoma samples, compared with single-nucleus RNA-seq, which is commonly applied for frozen archived tissues^72,73. One hundred base pair paired-end reads were sequenced on the Illumina NovaSeq 6000 system at the Center for Advanced Technology at the University of California San Francisco, and the resulting FASTQ files were processed using the CellRanger analysis suite (v.3.0.2; https://github.com/10XGenomics/cellranger) for alignment to the hg38 reference genome, identification of empty droplets, and determination of the count threshold for further analysis. A cell quality filter of greater than 500 features but fewer than 10,000 features per cell, and less than 20% of read counts attributed to mitochondrial genes, was used. Single-cell UMI count data were preprocessed in Seurat (v.3.0.1)^74,75 using the sctransform workflow⁷⁶, with scaling based on the regression of UMI count and the percentage of reads attributed to mitochondrial genes per cell. Dimensionality reduction was performed using principal component analysis and then principal component loadings were corrected for batch effects using Harmony⁷⁷. Uniform manifold approximation and projection was performed on the reduced data with a minimum distance metric of 0.4 and Louvain clustering was performed using a resolution of 0.2. Marker selection was performed in Seurat using a minimum difference in the fraction of detection of 0.5 and a minimum log-transformed fold change of 0.5. We assessed the single-cell transcriptome from 6,666 HFC-region cells and 7,065 LFC-region cells (Supplementary Table 4).

Immunohistochemistry and immunofluorescence analysis

After rehydration, 5.0 μm paraffin-embedded sections were processed for antigen retrieval followed by blocking and primary antibody incubation overnight at 4 °C. The following primary antibodies were used: rabbit anti-synapsin 1 (1:1,000, EMD Millipore), mouse anti-PSD95 (1:100, UC Davis), mouse anti-nestin (1:500, Abcam), mouse anti-neurofilament (M+H; 1:1,000, Novus Biologicals), mouse anti-TSP-1 (1:20, Invitrogen), rabbit anti-TSP-1 (1:50, Abcam), rabbit anti-MET (1:100, Abcam) and rabbit anti-Ki-67 (1:100, Abcam). We used species-specific secondary antibodies: Alexa 488 goat anti-chicken IgG, Alexa 488 goat anti-rabbit IgG, Alexa 568 goat anti-rabbit IgG, Alexa 568 goat anti-mouse IgG, Alexa 647 goat anti-rabbit IgG, all used at 1:500 (Invitrogen). After DAPI nuclear counter staining (Vector Laboratories, 1:1,000), coverslips were mounted with Fluoromount-G mounting medium (SouthernBiotech) for immunofluorescence analysis. The number of synapsin-1 and PSD95 puncta was quantified using spots (with automatic intensity maximum spot detection thresholds and a spot diameter of 1.0 µm) detection function of Imaris. The ratio of pre- and postsynaptic puncta was calculated by dividing the total number of synapsin-1 or PSD95 puncta on neurofilament-positive neurons to the total number of cells stained with DAPI in 135 μm × 135 μm field areas for quantification. Alternatively, the sections were incubated in DAB horseradish peroxidase (Vector Laboratories) for chemical colorimetric detection after incubation in ImmPress anti-rabbit IgG (Novus Biologicals) and counterstained with Harris haematoxylin for immunohistochemistry analysis.

Glioma–mouse hippocampal neuron co-culture

Glioma cells were plated on poly-d-lysine and laminin-coated coverslips (Neuvitro) at a density of 10,000 cells per well in 24-well plates. Approximately 24 h later, 40,000 embryonic mouse hippocampal neurons (Gibco) were seeded on top of the glioma cells and maintained with serum-free Neurobasal medium supplemented with B27, gentamicin and GlutaMAX (Gibco). After 2 weeks of co-culture, cells were fixed with 4% paraformaldehyde (PFA) for 30 min at 4 °C and incubated in blocking solution (5% normal donkey and goat serum, 0.25% Triton X-100 in PBS) at room temperature for 1 h. Next, they were treated with primary antibodies diluted in the blocking solutions overnight at 4 °C. The following antibodies were used: rabbit anti-homer-1 (1:250, Pierce), mouse anti-synapsin-1 (1:200) and chicken anti-MAP2 (1:500, Abcam). The coverslips were then rinsed three times in PBS and incubated in secondary antibody solution (Alexa 488 goat anti-chicken IgG; Alexa 568 goat anti-mouse IgG, and Alexa 647 goat anti-rabbit IgG, all used at 1:500 (Invitrogen) in antibody diluent solution for 1 h at room temperature. The coverslips were rinsed three times in PBS and then mounted with VECTA antifade mounting medium with DAPI (Vector Laboratories).

Confocal imaging and quantification of synapsin-1 and homer-1 staining and colocalization analysis

Images were captured at 1,024 × 1,024 resolution using a ×10 objective on the Nikon C2 confocal microscope. The confocal microscope settings for the homer-1 Alexa488 and synapsin-1 Alexa647 channels were held constant across all of the samples that were used for the experiment. Collected images were then imported into Imaris software (Imaris v.9.2.1, Bitplane) and the threshold value for each channel was manually adjusted and the colocalized voxels of the synapsin-1 puncta with the homer-1 marker was detected by creating a colocalization channel using the built-in the colocalization module of the Imaris software. Furthermore, the colocalization events were quantified by running the built-in spot detection algorithm of Imaris in conjunction with the colocalization channel. Next, dendrites labelled by MAP2 was visualized in TRITC channel and reconstructed using the Filament tool of Imaris software; the number of colocalized puncta representing synapses were counted and presented as the number of synapsin-1- and homer-1-positive puncta per 10 μm of dendrite length. Areas of homer-1 immunolabelled synaptic puncta were reconstructed using Imaris software Surface tool on maximal-intensity projections. Surfaces were built using a surface area detail level of 0.1 µm, thresholding by absolute intensity and taking all voxel >1.0 into account. The area sizes of individual anti-homer1-immunostained puncta were analysed and the mean values were calculated.

Induced neuron organoid and glioma co-culture

Induced neuron organoids were generated from a WTC11 iPS cell clone integrated by human NGN2 transgene induction as described previously^38,78. In brief, iNeuron organoids were generated by the transgenic human iPS cell WTC11 line by NGN2 induction through addition of 2 μg ml⁻¹ doxycycline in the 1:1 mixture of Neurobasal and BrainPhys neural medium containing 1% B-27 supplement, 0.5% GlutaMAX, 0.2 μM compound E, 10 ng ml⁻¹ BDNF and 10 ng ml⁻¹ NT-3 for 10 days to induce neuronal differentiation. Next, neuron maturation was triggered by feeding the organoids with approximately 8-month-old organoid conditioned medium derived from astrocytes. Astrocytes were differentiated from the human iPS cell WTC11 line and cultured in a medium consisting of DMEM/F12 containing GlutaMAX, sodium bicarbonate, sodium pyruvate, N-2 supplement, B-27 supplement (Gibco), 2 μg ml⁻¹ heparin, 10  ng ml⁻¹ EGF and 10 ng ml⁻¹ FGF2. Neuron organoids were characterized as postmitotic and stained for MAP2 and βIII-tubulin to validate neuronal induction efficiency. After 14 days of neuronal differentiation, HFC/LFC glioma cells labelled with RFP were added to the neuron organoid culture at a ratio of 1:3. Before iNeuron induction, the transgenic human iPS cell line WTC11 was transduced with GFP lentivirus. A Zeiss Cell Observer spinning-disc confocal microscope (Carl Zeiss) fitted with a temperature- and carbon-dioxide-controlled chamber was used to record live interactions of glioma cells with neuron organoids. Organoids were imaged every 10 min for a 6 h period, starting at the time of co-culture initiation, using a 10× objective with 0.4 NA. To assess the effect of exogenous TSP-1 on the functional integration between glioma cells and neurons, human recombinant TSP-1 (R&D Systems) was applied at a dose of 5 µg ml⁻¹ to the LFC-neuron organoid co-culture. Live-cell image analyses were performed using ImageJ. In brief, a region of interest was drawn around each GFP-positive neuron organoid and the fluorescence intensity (integrated density) of the RFP-positive glioblastoma cells was measured in the outlined regions of interest for each of the indicated timepoints. At the end of two weeks, organoids from HFC and LFC co-cultures were embedded in OCT and sectioned at 10 μm thickness for homer-1 immunofluorescence staining. Determination of homer-1 expression was performed by analysing homer-1 puncta density of neuron-organoid-HFC and LFC co-cultures.

MEA recordings

Preparation of MEA plates

We prepared 24-well CytoView multi-electrode plates (Axion Biosystems) before the addition of cells by coating with poly-d-lysine (Thermo Fisher Scientific), laminin (Fisher Scientific) and fibronectin (Corning). In brief, 1 day before the establishment of cultures, a solution of 0.1 mg ml⁻¹ of poly-d-lysine was added to the MEA plates at a volume of 100 μl per well and incubated at room temperature for 2 h. After 2 h, poly-d-lysine was aspirated and the plates were washed three times with sterile water, and allowed to air dry in a biosafety cabinet and stored at 4 °C. The next day, the plates were coated with 100 μl of 5 µg ml⁻¹ laminin and 1 μg ml⁻¹ fibronectin and incubated for 2 h at 37 °C before cell seeding.

Preparation of cortical cultures

Primary cortical cultures were established from E18 CD1 mice (Charles River Laboratories). Timed-pregnant CD1 dams were killed by CO₂ euthanasia in accordance with UCSF Institutional Animal Care and Use Committee (IACUC). Dissection of complete cortex from E18 embryos was performed in ice-cold HBSS (Gibco) under a dissecting microscope (Zeiss). Dissected cortices were minced to 1 mm² pieces and enzymatically digested in 5 ml of 0.25% trypsin reconstituted from 2.5% trypsin (Corning) in calcium- and magnesium-free Hank’s Balanced Salt Solution (Worthington Biochemical Corporation) for 30 min at 37 °C. Then, 0.5 ml of 10 mg ml⁻¹ of DNase (Sigma-Aldrich) was added in the last 5 min of dissociation. Mechanical dissociation was then carried about by trituration using fire-polished glass Pasteur pipettes until tissue was homogeneously suspended with no visible sections/aggregates and subsequently filtered through a 40 μm cell strainer (Thermo Fisher Scientific). Cells were collected by centrifugation at 500g for 5 min and the resulting cell pellet was resuspended in fresh complete BrainPhys culture medium (1× BrainPhys culture medium (StemCell Technologies) supplemented with B27 (Invitrogen), N2 (Invitrogen) and penicillin–streptomycin antibiotics (Invitrogen). Then, 10 μl of cell suspension was mixed 1:1 with Trypan Blue, and the viable cell concentration was quantified using a haemocytometer. Further dilution was performed to bring viable cell concentration to 100,000 cells per 10 μl. Droplets of 10 μl were then added directly over the electrode field of each pretreated MEA well and stored in the cell culture incubator for 1 h to allow cell adhesion. The wells were then carefully flooded with 500 μl complete BrainPhys medium and the cultures were maintained and allowed to mature in a tissue culture incubator with semi-weekly half-volume medium changes.

Recordings of spontaneous neuronal activity and analysis

Spontaneous extracellular neuronal recordings were carried out using the Maestro Edge system with an integrated heating system and temperature controller (Axion Biosystems) in combination with the Axion 24-well CytoView MEA plates (each well housing a 4 × 4 16-channel electrode array that are 350 µm away from each other) and Axion Integrated Studio (AxIS) Navigator (v.3.5.2; Axion Biosystems). In brief, to record spontaneous neuronal activity, the Neural Real-Time module was used. The neuronal firing events/action potentials (herein referred to as the spike) was defined by applying an adaptive threshold crossing method, that sets the threshold for spike detection for each channel/electrode to 5 s.d. of the noise level⁷⁹; activity exceeding this threshold was counted as a spike. Unless otherwise stated, all analysis considers only active channels, defined as channels exhibiting ≥5 spikes per min. Raw data files were obtained by sampling the channels simultaneously with a gain of 1,000× and a sampling frequency of 12.5 kHz per channel using a band-pass filter (200–3,000 Hz). To detect single-electrode bursting activity, an interspike interval threshold was used, setting the minimum number of spikes at 5 and the maximum interspike interval at 100 ms. The network bursting activity (simultaneous bursts at multiple MEA electrodes) was analysed by Neural Metric Tool (v.1.2.3; Axion Biosystems). For this purpose, the Adaptive algorithm was selected using the following settings: minimum number of spikes = 50 and minimum electrodes = 35%. Quantification of network synchrony was computed through AxIS software by calculating the area under the normalized cross-correlogram (AUNCC) as described previously^80,81,82,83. AUNCC represents the area under interelectrode cross-correlation normalized to the autocorrelations, with higher values indicating greater synchronicity of the network. For additional neural data analysis, including mean firing rate of each electrode (the ratio of the total number of spikes per second and the total duration of recording (1,800 s)) and weighted mean firing rate (defined as the spike rate per well multiplied by the number of active electrodes in the associated well), raw data files were processed offline using the Statistics Compiler function in AxIS. Statistics Compiler output files were processed in Microsoft Excel (Microsoft) and with custom Python scripts to organize and extract individual parameter data for each well of each MEA plate and for data normalization. Raster plots illustrating spike histogram and network bursts were generated using Neural Metric Tool (Axion Biosystems).

Glioma–neuron co-culture and gabapentin treatment

Spontaneous neuronal activity from cortical cultures grown on MEA plates was recorded in 30 min sessions on days in vitro 1 (DIV1), DIV7 and DIV15. Bright-field images were captured at each of the above timepoints to assess the neuronal cell density and electrode coverage. Primary cortical neurons showed a constant maturation trend from DIV7 to DIV15, and the co-culture experiments were initiated when neurons showed a synchronous activity pattern network at DIV15. Baseline data were therefore recorded on DIV15 immediately before addition of glioma cells in the presence or absence of gabapentin. For glioma cell co-culture, a single-cell suspension from cultured neurospheres of primary patient-derived HFC and LFC were prepared and diluted to a viable cell concentration of 20,000 cells per 5 μl. Droplets of 5 μl were then plated on top of differentiating neurons in the MEA plate. After plating, glioma cells were allowed to adhere for approximately 1 h after which HFC cultures were exposed for next 24–48 h to a working concentration of 50 µM gabapentin^84,85 diluted in complete BrainPhys medium or an equivalent amount of vehicle (sterile water) as a control. Each condition was run on two wells (experimental replicates). Neurons from two different embryos were used as biological triplicates (n = 2). Presented data from MEA recordings reflects well-wide averages from active electrodes, with the number of wells per condition represented by n values.

Mice and housing conditions

All in vivo experiments were conducted in accordance with the protocols approved by the UCSF Institutional Animal Care and Use Committee (IACUC) and performed in accordance with institutional guidelines. Animals were maintained under pathogen-free conditions, in temperature- and humidity-controlled housing, with free access to food and water, under a 12 h–12 h light–dark cycle. For brain tumour xenograft experiments, the IACUC does not set a limit on maximal tumour volume but rather on indications of morbidity. These limits were not exceeded in any of the experiments as mice were euthanized if they exhibited signs of neurological morbidity or lost 15% or more of their body weight.

Orthotopic xenografting for neuronal circuit integration and mouse survival experiments

For all xenograft studies, NSG mice (NOD-SCID-IL2R gamma chain-deficient, The Jackson Laboratory) were used. Male and female mice were used equally. For immuno-electron microscopy experiments, a single-cell suspension from cultured neurospheres of HFC and LFC (n = 2 each) labelled with red fluorescent protein (RFP) were prepared in sterile DMEM immediately before the xenograft procedure. Mice (n = 8; 2 biological replicates per patient line) at postnatal day 28–30 were anaesthetized with 1–4% isoflurane and placed into a stereotactic apparatus. The cranium was exposed through a midline incision under aseptic conditions. Approximately 50,000 cells in 2 µl sterile PBS were stereotactically implanted into the CA1 region of the hippocampus through a 31-gauge burr hole, using a digital pump at infusion rate of 0.4 µl min⁻¹ and 31-gauge Hamilton syringe. Stereotactic coordinates used were as follows: 1.5 mm lateral to midline, 1.8 mm posterior to bregma, −1.4 mm deep to cranial surface. At the completion of infusion, the syringe needle was allowed to remain in place for a minimum of 2 min, then manually withdrawn at a rate of 0.875 mm min⁻¹ to minimize backflow of the injected cell suspension. For survival studies, morbidity criteria used were either: reduction of weight by 15% initial weight, or clinical signs such as hunched posture, lethargy or persistent decumbency. Kaplan–Meier survival analysis using log-rank testing was performed to determine statistical significance.

Quantification of tumour cell burden

Cell quantification was performed by a blinded investigator at 10–20× magnification using the Zeiss LSM800 scanning confocal microscope and Zen 2011 imaging software (Carl Zeiss). The area for quantification was selected for a 1-in-6 series of 40 μm coronal sections (240 µm apart from one another). Immunohistochemistry was performed on brain sections from HFC and LFC xenografts to stain for human nuclear antigen (HNA)-positive tumour cells. The tumour burden was evaluated using blinded rank-order analysis as previously reported⁸⁶. For each mouse, the section with the maximal amount of tumour burden was selected as defined by the number of HNA-positive cells. From this section, a tiled ×10 image of the centre of the section (3 × 3 tiles stitched together) was created, therefore generating a single image for each mouse that represents the maximal amount of tumour burden. Next, the images generated from each mouse were compared and ranked in order from the image with the least to maximum number of cells. Subsequently, scores were assigned and ranged from 1 to 24 with 1 and 24 representing images with the lowest and highest number of HNA-positive cells, respectively. After scoring, the experimenter was unblinded and the scores from each image were assigned to each of the respective HFC or LFC experimental conditions. Statistical differences between HFC and LFC scores were evaluated using two-tailed unpaired Mann–Whitney tests.

Sample preparation and image acquisition for electron microscopy

Twelve weeks after xenografting, mice were euthanized by transcardial perfusion with Karnovsky’s fixative: 2% glutaraldehyde (EMS, 16000) and 4% PFA (EMS, 15700) in 0.1 M sodium cacodylate (EMS, 12300), pH 7.4. Transmission electron microscopy (TEM) was performed in the tumour mass within the CA1 region of the hippocampus for all xenograft analysis. The samples were then post-fixed in 1% osmium tetroxide (EMS, 19100) for 1 h at 4 °C, washed three times with ultrafiltered water, then en bloc stained overnight at 4 °C. The samples were dehydrated in graded ethanol (50%, 75% and 95%) for 15 min each at 4 °C; the samples were then allowed to equilibrate to room temperature and were rinsed in 100% ethanol twice, followed by acetonitrile for 15 min. The samples were infiltrated with EMbed-812 resin (EMS, 14120) mixed 1:1 with acetonitrile for 2 h followed by 2:1 EMbed-812:acetonitrile overnight. The samples were then placed into EMbed-812 for 2 h, then placed into TAAB capsules filled with fresh resin, which were then placed into a 65 °C oven overnight. Sections were taken between 40 nm and 60 nm on a Leica Ultracut S (Leica) and mounted on 100-mesh Ni grids (EMS FCF100-Ni). For immunohistochemistry, microetching was done with 10% periodic acid and eluting of osmium with 10% sodium metaperiodate for 15 min at room temperature on parafilm. Grids were rinsed with water three times, followed by 0.5 M glycine quench, and then incubated in blocking solution (0.5% BSA, 0.5% ovalbumin in PBST) at room temperature for 20 min. Primary goat anti-RFP (1: 300, ABIN6254205) was diluted in the same blocking solution and incubated overnight at 4 °C. The next day, grids were rinsed in PBS three times, and incubated in secondary antibodies (1:10 10 nm gold-conjugated IgG, TED Pella, 15796) for 1 h at room temperature and rinsed with PBST followed by water. For each staining set, samples that did not contain any RFP-expressing cells were stained simultaneously to control for any non-specific binding. Grids were contrast stained for 30 s in 3.5% uranyl acetate in 50% acetone followed by staining in 0.2% lead citrate for 90 s. The samples were imaged using a JEOL JEM-1400 TEM at 120 kV and images were collected using a Gatan Orius digital camera.

Electron microscopy data analysis

Sections from the xenografted hippocampi of mice were imaged as above using TEM imaging. Here 101 sections of HFC xenografts across 4 mice and 104 sections of LFC xenografts across 3 mice were analysed. Electron microscopy images were taken at ×6,000 with a field of view of 15.75 μm². Glioma cells were counted and analysed after unequivocal identification of a cluster of immunogold particle labelling with 15 or more particles. The total number of synapses, including neuron-to-neuron and neuron-to-glioma synapses (identified by: (1) the presence of synaptic vesicle clusters; (2) visually apparent synaptic cleft; and (3) identification of clear postsynaptic density in the glioma cell) were counted.

EdU assay

The EdU-incorporation assay was performed using an EdU assay kit (Invitrogen) according to the manufacturer’s instructions. Patient-derived HFC/LFC glioma cells were seeded on poly-d-lysine- and laminin-coated coverslips at 10,000 cells per well of a 24-well plate. After 24 h of seeding, embryonic mouse hippocampal neurons were added to the glioma cells at 40,000 cells per well for the glioma-neuron co-culture group. After 72 h, glioma cells alone or in coculture with neurons were treated with 20 μM EdU overnight at 37 °C. Subsequently, the cells were fixed with 4% PFA and stained using the Click-iT EdU kit protocol. The proliferation index was then determined by quantifying the fraction of EdU-labelled cells/DAPI-labelled cells using confocal microscopy at ×10 magnification.

3D spheroid invasion assay

Glioma cell invasion was evaluated by performing an invasion assay using the Cultrex 3D Spheroid Cell Invasion Assay Kit (Trevigen) according to the manufacturer’s protocol. In brief, 3,000 cells were resuspended in 50 μl of spheroid formation matrix solution (prepared in culture medium) in a round-bottom 96-well plate. Spheroids were allowed to form for 72 h and images were taken at ×10 magnification before addition of invasion matrix (0 h). Working on ice, 50 μl of invasion matrix was then added to each well and the plate was incubated at 37 °C. After 1 h of gel formation, for the glioma cell + mouse conditioned medium (mCM) group, 100 μl of mouse hippocampal neuron supernatant was added to the wells to assess the effect of neuronal secreted factors on the invasiveness of the glioma cells. After 24 h of incubation at 37 °C, invasions were observed under a microscope and images were taken at ×10. Microtube lengths as well as the area of each spheroid measured at 0 h (pre-invasion) and 24 h (post-invasion) were analysed using ImageJ and the difference was used to calculate the total area of cell invasion.

SEM analysis

SEM was performed as described previously³⁹ to investigate the tumour microtubes in greater detail. Primary patient-derived HFC/LFC glioma cells were seeded on poly-d-lysine- and laminin-coated coverslips (Neuvitro) at 10,000 cells per well of a 24-well plate. For the glioma cell + mCM and glioma cell + mCM + GBP groups, mouse neuronal conditioned medium was added to the wells in the absence or presence of gabapentin (32 µM), respectively, to assess the effect of neuronal secreted factors and gabapentin modulation on tumour microtube formation of glioma cells^47,87,88. After culture for 1 week, the samples were fixed with 4% PFA for 30 min at 4 °C, washed with PBS and ultrafiltered water. After serial stepwise ethanol dehydration, coverslips were mounted on SEM stubs (TED Pella) using conductive adhesive tape (12 mm OD PELCO Tabs, TED Pella). The samples were then sputter-coated with a 2 nm layer of gold palladium. Tumour microtubes were observed under field emission SEM (Sigma 500; Carl Zeiss Microscopy) and micrographs were recorded at an accelerating voltage of 1.0 kV.

THBS1 shRNA knockdown

Primary patient-derived HFC glioma cells were seeded in 6-well plates at a density of 2 × 10⁵ cells per well. After overnight incubation, cells were infected with lentiviral particles expressing a shRNA targeting THBS1 or a control scramble construct expressing green fluorescent protein (GFP) according to the manufacturer’s protocol using polybrene transfection reagent and subsequently kept in a 5% CO₂ incubator at 37 °C for 24 h. After 24 h, the medium was replaced with fresh complete culture medium. Cells were checked 72 h after transduction for GFP expression to evaluate the efficiency of the transduction. The lentiviral shRNA constructs targeting THBS1 (5′-AGACATCTTCCAAGCATATAA-3′) and control scrambled shRNA (5′-CCTAAGGTTAAGTCGCCCTCG-3′) were designed and constructed by VectorBuilder.

Cell viability assay

The cell viability of primary patient-derived HFC cells after THBS1 knockdown was evaluated using the Live/Dead Viability/Cytotoxicity kit (Molecular Probes). One-week after transduction, HFC cells expressing scramble shRNA or shRNA targeting THBS1 were seeded on poly-d-lysine- and laminin-coated coverslips (Neuvitro) at 10,000 cells per well of a 24-well plate. After 1 week of culture, a 500 μl cell-staining solution at a final concentration of 2 µM calcein AM and 4 µM ethidium bromide (EthD-1) in DPBS was added to each well, and the plates were incubated for 45 min at room temperature in dark. As an indicator of cell viability, the calcein-AM is metabolically converted by intracellular esterase activity resulting in the green, fluorescent product, calcein. EthD-1 is excluded from live cells but is readily taken up by dead cells and stains the DNA emitting red fluorescence. Live and dead cells were imaged from four random fields per well and were visualized under a fluorescence microscope. The percentage of live cells was calculated as the number of live cells (in green) divided by the number of total cells (green + red) per image field.

In vitro cell proliferation assessment after THBS1 shRNA knockdown

After 1 week of transduction, primary patient-derived high connectivity glioma cells expressing scramble shRNA or shRNA against THBS1 were seeded on poly-d-lysine- and laminin-coated coverslips (Neuvitro) at 10,000 cells per well of a 24-well plate. Approximately 24 h later, 40,000 embryonic mouse hippocampal neurons (Gibco) were seeded on top of the glioma cells and maintained with serum-free Neurobasal medium supplemented with B27, gentamicin and GlutaMAX (Gibco). After 1 week of co-culture, cells were fixed with 4% PFA for 30 min at 4 °C and incubated in blocking solution (5% normal donkey and goat serum, 0.25% Triton X-100 in PBS) at room temperature for 1 h. Next, cells were treated with primary antibodies diluted in the blocking solutions overnight at 4 °C. The following antibodies were used: rabbit anti-Ki-67 (1:500, Abcam) and human nuclear antigen (HNA; mouse anti-human nuclei, 235-1; 1:100, Millipore). The coverslips were then rinsed three times in PBS and incubated in secondary antibody solution (Alexa 488 goat anti-rabbit IgG, and Alexa 647 goat anti-mouse IgG) all used at 1:500 (Invitrogen) in antibody diluent solution for 1 h at room temperature. The coverslips were rinsed three times in PBS and then mounted with VECTA antifade mounting medium with DAPI (Vector Laboratories). To calculate the proliferation index, the total number of HNA-positive cells co-labelled with Ki-67 was divided by the total number of human nuclei-labelled cells per image field visualized using confocal microscopy at ×40 magnification.

In vitro cell proliferation assessment after pharmacological inhibition of TSP-1 by gabapentin

The effect of TSP-1 inhibition by gabapentin on glioma cell proliferation was evaluated by Ki-67 immunofluorescence staining, as described above. In brief, primary patient-derived high connectivity glioma cells were seeded on poly-d-lysine- and laminin-coated coverslips (Neuvitro) at 10,000 cells per well of a 24-well plate. Approximately 24 h later, 40,000 embryonic mouse hippocampal neurons (Gibco) were seeded on top of the glioma cells and maintained with serum-free Neurobasal medium supplemented with B27, gentamicin and GlutaMAX (Gibco). The next day, cultures were treated either with vehicle (sterile water) or 32 µM gabapentin followed by daily half-medium switches of fresh 32 µM gabapentin until 1 week of coculture⁸⁸. Subsequently, cells were fixed and immunostained for Ki-67 and HNA labelling for proliferation assessment as described above.

Orthotopic xenografting for proliferation assessment and pharmacological inhibition of TSP-1 by gabapentin

A single-cell suspension from cultured neurospheres of primary patient-derived HFC and LFC were prepared in sterile HBSS immediately before the xenograft procedure by dissociation with TrypLE (Thermo Fisher Scientific). Mice (4–6 biological replicates per patient line) at postnatal day 28–30 were anaesthetized with 1–4% isoflurane and placed into a stereotactic apparatus. The cranium was exposed through midline incision under aseptic conditions. Approximately 150,000 cells in 3 µl sterile HBSS were stereotactically implanted into the premotor cortex (M2) through a 26-gauge burr hole, using a digital pump at infusion rate of 1.0 μl min⁻¹. Stereotactic coordinates used were as follows: 1.0 mm lateral to midline, 1.0 mm anterior to bregma, −1.0 mm deep to cortical surface. Four weeks after xenograft, HFC/LFC-bearing mice were treated with systemic administration of gabapentin (200 mg kg⁻¹; Sigma-Aldrich; formulated in saline) through intraperitoneal injection for 28 consecutive days. Controls were treated with an identical volume of the relevant vehicle. At 8 weeks after xenograft, mice were euthanized and coronal brain sections at 40 µm were obtained for immunohistochemistry.

Confocal imaging and quantification of tumour burden and cell proliferation

Cell quantification was performed by a blinded investigator at ×10–20 magnification using the Zeiss LSM800 scanning confocal microscope and Zen imaging software (Carl Zeiss). The area for quantification was selected for a 1-in-6 series of 40 μm coronal sections (240 µm apart from one another). Within 3 sections of maximal tumour burden, all HNA-positive (mouse anti-human nuclei, 235-1; 1:100, Millipore) cells were quantified to determine the tumour burden within the areas quantified. HNA-positive tumour cells were then assessed for double-labelling with Ki-67. To calculate the proliferation index (the percentage of proliferating tumour cells for each animal), the total number of HNA-positive cells co-labelled with Ki-67 across all areas quantified was divided by the total number of human nuclei-positive cells counted across all areas quantified. Differences in proliferation indices were calculated using unpaired, one-tailed Student’s t-tests.

ELISA

Peripheral blood samples from newly diagnosed patients with glioblastoma (n = 56) were collected and allowed to clot for 30 min at room temperature before centrifugation for 15 min at 1,000g. The serum was stored at −80 °C until analysis. The TSP-1 level was determined using the Quantikine immunosorbent assay kits according to the manufacturer’s instructions (R&D Systems). To confirm the functional protein-level knockdown of THBS1, the TSP-1 level was also measured in cell culture supernatants collected from scramble and THBS1-shRNA transduced HFC cells (collected one-week after infection) using the Picokine ELISA kit (Boster Biological Technology).

Statistics and reproducibility

Statistical tests were conducted using Prism (GraphPad) software unless otherwise indicated. The significance test of different groups was determined using Student’s t-tests and one-way ANOVA with Tukey’s post hoc tests. P ≤ 0.05 was considered to be statistically significant. Two-tailed unpaired Mann–Whitney tests were used to analyse the tumour cell burden of HFC and LFC hippocampal xenografts using the same significance denotation as above. Two-tailed log-rank analyses were used to analyse the statistical significance of Kaplan–Meier survival curves for human patients. All of the micrographs shown are representative of three independently conducted experiments, with similar results obtained. Statistical analyses for RNA-seq data are described above in the respective sections. The tumour volumes of patients with glioblastoma were calculated by manual segmentation with region-of-interest analysis ‘painting’ inclusion regions on the basis of FLAIR sequences. Volumetric measurements were made blinded to patients’ clinical outcomes. Manual segmentations were performed by co-authors A.E., A.A. and A.L. with tumour volumetrics verified for accuracy after an initial training period. Student’s t-tests and χ² tests were used to compare continuous and categorical variables between patient cohorts, respectively. Patient overall survival (OS) was defined as the time from the date of first surgery or original biopsy (if it occurred before surgery) until death or the last contact date. Alive patients were censored at the time of loss to follow-up or last follow-up date. Median follow-up was estimated using the reverse Kaplan–Meier method.

To identify clinically relevant survival risk groups in a multivariate setting, we used recursive partitioning analyses for survival data using the partDSA algorithm (v.0.9.14)^42,43,44. Survival trees use recursive partitioning to divide patients into different risk groups. The Brier score was the chosen loss function for splitting and pruning. Such methods are nonparametric and, therefore, do not require the proportional hazards assumption. All known prognostic variables were included in the trees, including age at diagnosis, sex, MGMT promoter methylation status, tumour location, chemotherapy, radiotherapy, the presence of functional connectivity within the tumour, pre- and post-operative tumour volume and the extent of resection. MGMT methylation status was included in both univariate Cox proportional-hazard modelling in addition to multivariate recursive partitioning survival analysis. The tree that minimized the fivefold cross-validated error as well as the most parsimonious tree within one standard error of the overall minimum error were selected for review. Leaves of the resulting trees defined the final risk groups from which the corresponding Kaplan–Meier curves were generated. Median OS times and hazard ratios were generated and compared between risk groups using the Kaplan–Meier method and the Cox proportional hazards model, respectively. The proportional hazards assumption was verified.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.