Informed consent was obtained from all patients prior to sample collection for diagnostic purposes. This study was reviewed and approved by the institutional review boards of Sainte Anne Hospital, Paris, France, and Laennec Hospital, Nantes, France, and performed per the Helsinki Protocol. Animal procedures were conducted in agreement with the European Convention for the Protection of Vertebrate Animals used for Experimental and other Scientific Purposes (ETS 123) and approved by the French Government (APAFIS#24400-2020022713064016.v2). 7–8 weeks-old male and female C57BL/6 mice (Janvier Labs) were housed in specific pathogen-free conditions.

Two different patient-derived glioblastoma stem-like cells from patient #1 (GSC1, mesenchymal, 68-year-old male) and patient #9 (GSC9, classical, 68-year-old female), were isolated and grown as spheres in mitogen-defined serum-free medium (NS34, DMEM/F12 with B27, G5, and N2 supplements, GlutaMAX and antibiotics; Life Technologies, Harford-Wright et al., 2017). Stealth non-silencing control duplexes (low-GC 12935200, Life Technologies) and small interfering RNA duplexes targeting human NRP1 (a mixture of HSS113022, HSS189594, HSS189595 in GSC1 and HSS113022 in GSC9, Life Technologies), ITGB3 (5′-GCU​CAU​CUG​GAA​ACU​CCU​CAU​CAC​CTT-3′), and GFP (#EHUEGFP, Sigma) were transfected using RNAiMAX lipofectamine (Life Technologies). Additionally, GSC9 was transduced with GFP and luciferase (pLNT-LucF/pFG12-eGFP mixture plasmid, a gift of Valérie Trichet, Nantes Universite, France).

The following primary antibodies were used for western blot (dilution 1:1000): GAPDH (SC-25778), ITGB1 (SC-365679), ITGB5 (SC-398214), ITGB7 (SC-515397), ITGB8 (SC-514150) from Santa Cruz Biotechnology, and ITGB3 (4702S, Cell Signaling Technologies). HRP-conjugated secondary antibodies (anti-rabbit, anti-mouse IgG1, IgG2a, and IgG2b) were purchased from Southern Biotech and used at 1:5000. APC-conjugated antibodies were used for flow cytometry: NRP1 (FAB3870A, R and D Systems, 1:20), ITGB1 (cat#559883, BD, 1:20), and ITGB3 (17-0619-42, Invitrogen, 1:20). Fasudil (Y₂₇₆₃₂, 50 nM) was from Sigma-Aldrich.

Selected single guide RNA (sgRNA: 5′-CTT​CAA​CCC​TCA​CTT​CGA​TT-3′) targeting human NRP1 was cloned into a lentiviral lentiCRISPRv2 backbone (GeCKO, ZhangLab), as described (Trillet et al., 2021). Puromycin-selected cells were single-cloned via NRP1 negative sorting (AriaIII, BD, Cytocell, SFR Francois Bonamy, Nantes). One bi-allelic KO clone was further selected after genomic DNA sequencing.

TCGA was explored via the Gliovis platform (http://gliovis.bioinfo.cnio.es/) (Bowman et al., 2017). RNAseq databases were used to interrogate clinical information associated with the expression of ITGB1, NRP1, and NRP2. The medium cutoff was applied, and all subtypes of primary GBM were included.

Surface staining was performed on living cells by incubating antibodies for 1 h at 4°C. Flow cytometry analysis was performed on FACSCalibur (Cytocell, SFR Francois Bonamy, Nantes) and processed using FlowJo software (BD, version X).

Cells were lysed in RIPA lysis buffer (25 mM Tris-HCl pH 7.4, 150 mM NaCl, 0.1% SDS, 0.5% Na-Deoxycholate, 1% NP-40, 1 mM EDTA, Pierce protease inhibitor) for 30 min on ice and cleared by centrifugation (10,000 g, 10 min, 4°C). 10 μg of post-nuclei supernatants were resolved by SDS-PAGE and transferred to nitrocellulose membranes. Signals were revealed using chemiluminescent HRP substrate (Merck) and visualized using Fusion imaging system (Vilber-Lourmart).

Equal amounts of RNA (RNeasy, Qiagen) were reverse-transcribed (Maxima Reverse Transcriptase, Life Technologies), and 50 ng of cDNA was amplified (PerfeCTA SYBR Green SuperMix Low ROX, QuantaBio). Data were analyzed by the 2-d dCt methods and normalized with the housekeeping genes ACTB and HPRT1. The following primers (human targets) were used: ACTB forward 5′-GGA​CTT​CGA​GCA​AGA​GAT​GG-3′; ACTB reverse 5′-AGC​ACT​GTG​TTG​GCG​TAC​AG-3′; HPRT1 forward 5′-TGA​CAC​TGG​CAA​AAC​AAT​GCA-3′; HPRT1 reverse 5′-GGT​CCT​TTT​CAC​CAG​CAA​GCT-3′; NRP1 forward 5′-ATC​ACG​TGC​AGC​TCA​AGT​GG-3′, NRP1 reverse 5′-TCA​TGC​AGT​GGG​CAG​AGT​TC-3’; ITGB1 forward 5′-CCT​ACT​TCT​GCA​CGA​TGT​GAT​G-3’; ITGB1 reverse 5′-CCT​TTG​CTA​CGG​TTG​GTT​ACA​TT-3; ITGB2 forward 5′-TTC​GGG​TCC​TTC​GTG​GAC​A-3’; ITGB2 reverse 5′-ACT​GGT​TGG​AGT​TGT​TGG​TC-3’; ITGB3 forward 5′-GTG​ACC​TGA​AGG​AGA​ATC​TGC-3’; ITGB3 reverse 5′-CCG​GAG​TGC​AAT​CCT​CTG​G-3’; ITGB4 forward 5′-CTC​CAC​CGA​GTC​AGC​CTT​C-3’; ITGB4 reverse 5′-CGG​GTA​GTC​CTG​TGT​CCT​GTA-3’; ITGB5 forward 5′-AAC​TCG​CGG​AGG​AGA​TGA​G-3’; ITGB5 reverse 5′-GGT​GCC​GTG​TAG​GAG​AAA​GG-3’; ITGB6 forward 5′-TCC​ATC​TGG​AGT​TGG​CGA​AAG-3’; ITGB6 reverse 5′-TCT​GTC​TGC​CTA​CAC​TGA​GAG-3’; ITGB7 forward 5′-CCA​TTC​AGC​TTT​CAC​CAT​GTG​C-3’; ITGB7 reverse 5′-ACC​TTC​AGG​CGA​GTC​CAG​ATT-3’; ITGB8 forward 5′-GTG​AAA​GTC​ATA​TCG​GAT​GGC​G-3’; ITGB8 reverse 5′-GCT​ATC​AAG​AGC​GAG​ATG​AGA​CG-3’.

8,000 GSCs were seeded in a 96-well Ultra-Low Attachment plate (Corning) in 100 μl NS34 + 0.16% Methylcellulose (4000 cP, Sigma Aldrich), centrifuged (100 g, 20 min), and incubated for 2 days. Spheroids were embedded in 1 mg/ml collagen hydrogel (rat-tail collagen I, Corning, 1 M NaOH). Upon polymerization, 100 μl of NS34 was added. Alternatively, cells were placed into growth factor-reduced matrigel (Corning). Images were acquired at the indicated times with an inverted microscope (Nikon Tie), equipped with a cmos sensor using a 4x/0.13 objective (Nikon). All images were analyzed and quantified using Fiji ImageJ software (version 2.1.0, https://imagej.nih.gov/ij/). The invasive index was calculated as the ratio of the spheroid areas day2/day0.

8,000 cells seeded in triplicate were treated as indicated for 3 days. Cell viability was measured using luminescent metabolic assay (CellTiterGlo, Promega) on a FluoStar Optima plate reader (BMG Biotech).

Experiments were performed as described in (Harford-Wright et al., 2017). GSCs (60 cells/μl) were seeded in triplicate in NS34, dissociated manually from day 1–3, and analyzed on day 4. The number of tumorspheres was manually counted (single-blinded) in five random fields of view (fov) per well.

Spheroids were collected, fixed (PBS-PFA 4%, 30 min), and permeabilized with PBS-Triton X-100 0.5%, BSA 3% (2 h). Alexa488-conjugated Phalloidin (1:300, Life Technologies) was added for 24 h. After PBS-Tween20 0.02% washes, spheroids were incubated for 24 h with TO-PRO-3 (1:1000, Life Technologies). Following washes, samples were cleared (2 days, 4°C, RapiClear 1.47, SunJin Lab). Images were acquired on a confocal microscope (Nikon A1rHD) using a 25x/1.05 silicon-immersion objective and reconstructed in 3D using the NIS-Element Software (Nikon).

7–8 weeks old female and male mice were euthanized and brains were harvested. Brains were cut in 8–10 fragments on ice and embedded in a 5% low gelling temperature agarose at 37°C (Sigma-Aldrich). 250 µm thick brain slices (vibratome, Leica VT1000S) were transferred onto culture inserts (Merck Millipore) in NS34 culture media. 30,000 GreenCellTracker-stained GSCs (1:1000, 1 h, 37°C, Life Technologies) were added together with isolectin GS-IB4 dye (1:60, Life Technologies) and incubated for 1 h more at 37°C. Slices were washed, fixed (PBS-PFA 4%, 20 min), and imaged with a confocal microscope (Nikon A1r). A depth of up to 60 µm was acquired (20x/0.75 objective, 4–6 areas per slice). The distance between GSCs (green) and blood vessels (red) was quantified using Fiji ImageJ software (version 1.53q, Schindelin et al., 2012). Images were Z projected at max intensity for 2D analysis. GSCs were segmented using the StarDist plugins (Schmidt et al., 2018) and the distance map was calculated outside the thresholded vessels using Li method.

Data are representative of at least three independent experiments unless otherwise stated. Statistical analyses were performed using GraphPad Prism9 using an unpaired t-test (Student’s t-test), one-way analysis of variance (ANOVA), or two-way ANOVA. For each statistical test, a p-value of <0.05 was considered significant.

To study the infiltrative behavior of patient-derived glioblastoma cells (GSCs), we applied a two-step process that allowed 1) to control for spheroid size and growth upon centrifugation of isolated patient cells, and, 2) to examine invasion in a serum-free defined matrix (Figure 1A). First, confocal analysis of spheroids that were cleared and stained with TO-PRO-3 to mark nuclei and Phalloidin to illuminate actin structures confirmed the integrity of the engineered spheroids (Figure 1B). Next, collagen was preferred over commonly used matrigel, in which GSC invasion was prevented (Figure 1C). We estimated the invasion index by the ratio of the invasive area (as delimited with the red line), normalized to the area of the core (as identified with the black circle) (Figure 1D). Collagen concentrations ranging from 0.5 to 1.5 mg/ml provided effective support for the invasion, as opposed to concentrations above 2 mg/ml that strongly hampered it (Figure 1E), most likely due to alterations of the bio-physical properties of the as prepared hydrogels (Antoine et al., 2014). At the chosen medium 1 mg/ml collagen concentration, invasion index augmented over time (from day 1–3), and was visible in two patient-derived lines (namely mesenchymal GSC1 and classical GSC9) with comparable intrinsic invasive properties (Figure 1F). Finally, we evaluated whether these experimental settings can be challenged with genetic and pharmacological manipulation of patient cells. As a proof-of-concept, GFP can be efficiently knocked-down in stably GFP-expressing spheroids (Figure 1G). Likewise, the use of Y₂₇₆₃₂, a ROCK inhibitor also known as fasudil, a well-established inhibitor of cell motility (Basile et al., 2007), reduced the invasion index (Figure 1H). Thus, we developed a reliable method to assess patient cell invasion in collagen scaffolds.

Several studies suggest that integrins, notably α_vβ₃ and those that encompass the β₁ subunit, regulate glioblastoma cell invasion (Martin et al., 2012). To identify ITGB1-related genes whose expression might impact the disease outcome, we interrogated The Cancer Genome Atlas (TCGA) through the Gliovis platform (Bowman et al., 2017) (Figure 2A). A total of 2,438 genes were found differentially expressed (i.e., adjusted p-value ≤ 0.05) between high and low ITGB1 expressed samples (median cutoff), among which 1,355 and 1,083 genes were positively and negatively associated, respectively (Figure 2A). A focus on the ITGB1-positive correlated genes highlighted three groups of 20 genes with a Pearson’s correlation r factor ranging between 0.7 and 0.8. Their GO functions were related to 1) matrix biology and remodeling (COL1A1, COL1A2, COL3A1, COL4A1, FN1, FNDC3B, LAMC1, LOXL2, LUM, and SERPINH1, please see the light-blue dots), 2) integrin sub-units (ITGA1, ITGA4, and ITGA5, please see the regular-blue dots), and 3) actin cytoskeleton and adhesion (CD93, LMAN1, TES, TMP4, please see the dark-blue dots) (Figure 2B). Of note, three miscellaneous genes namely IKBIP, SEC24D, and NRP1 (please see the red dots), emerged from this data mining. The glycoprotein neuropilin-1 (NRP1) was retained, as it was previously linked to integrin functions in the context of angiogenesis (Serini et al., 2008). Likewise, NRP1 has been involved in glioma cell migration processes, notably in 2D and chemotaxis models (Evans et al., 2011; Angom et al., 2020). To get further insights on the codependency between NRP1 and ITGB subunits in glioblastoma, TCGA was again queried (Figure 2C). The expression level of NRP1, and to a lesser extent of NRP2, strongly correlated with the one of ITGB1, ITGB3, and ITGB5 (Figure 2C). In agreement with previous studies (Angom et al., 2020), NRP1 expression was over-represented in glioblastoma patients, as opposed to non-tumor brain tissue (Figure 2D). The Kaplan-Meier survival analysis on TCGA database that includes 320 primary GBM patients suggests that high NRP1 expression (median cutoff) is of slight worsen prognosis (log-rank p-value 0.0204, Wilcoxon p-value 0.0072), with a noticeable difference in the earlier time points (Figure 2E). Paralleling the situation with endothelial cells (Serini et al., 2008), NRP1-dependent differentially expressed genes notably clustered in two biological functions related to integrin functions, namely “ECM receptor” and “focal adhesion” (Figure 2F). Alternate signatures also emerged and involved either dynamic adhesion and morphological remodeling processes (e.g., “protein digestion and absorption,” “proteoglycans in cancer”) or immune system and responses (e.g., “amoebiasis,” “complement and coagulation cascade,” “AGE-RAGE signaling pathway,” “IL17 signaling pathway,” “phagosome”) (Figure 2F). To next evaluate the relevance of the interconnection between ITGB and NRP1 in GSCs, NRP1 was knocked-down with transient RNA interference in two patient-derived cells (Figure 2G). This significantly reduced the level of ITGB3, at the RNA and protein level, while sparing other tested ITGB subunits, among which ITGB1, in GSC1 and GSC9 (Figures 2G,H). Of note, B2 in GSC1 and B2, B4, and B6 in GSC9 integrin subunits were barely expressed. Further flow cytometry analysis confirmed that NRP1 silencing caused a significant reduction in the surface staining of ITGB3, but not of the widely expressed ITGB1, in the two different patient-derived GSCs (Figure 2I). To reinforce our findings, a CRISPR-Cas9-based bi-allelic knockout of NRP1 was designed and selected in GSC1 (Figure 2J). NRP1 deletion resulted in a significant reduction of both RNA and protein surface staining of ITGB3, but not of ITGB1 (Figure 2K). Our data strengthened the concept that NRP1 is an integral part of the integrin network in GSCs.

We next explored the functional impact of NRP1 silencing on the invasive properties of patient-isolated GSCs. First, NRP1 knockdown did not impact neither GSC viability nor their ability to form spheres in floating non-adherent conditions (Figures 3A,B). Strikingly, the invasion in the collagen-based 3D matrix was significantly reduced in NRP1-silenced GSCs (Figure 3C). Likewise, NRP1 KO did not influence GSC viability and tumorsphere formation, yet resulted in the loss of invasion in 3D collagen scaffolds, therefore recapitulating the effect of the transient knockdown (Figures 3D–F). Because GSCs can invade the brain parenchyma by co-opting vessel routes on collagen tracks (Mancuso et al., 2006; Griveau et al., 2018; Rosińska and Gavard, 2021), we next deployed a model of organotypic brain slice culture. This system not only recapitulates the 3D brain architecture and stroma composition but also allows the tracking and visualization of fluorescent-labeled GSC behavior in interaction with isolectin-marked vasculature (Figure 3G). In such conditions, control GSC9 were mainly found in the proximity of vessels, largely overlapping (Figure 3H). Conversely, NRP1 knocked-down cells were stochastically dispersed on the brain slices, as the mean distance between GSCs and the closest blood vessels was significantly heightened (Figure 3H). Similarly, while WT GSC1 were in close contact with vessels, NRP1 KO cells positioned at distance, recapitulating the effects of the transient knockdown (Figure 3I). Our data thus highlighted the pivotal role of NRP1 in controlling the motility of GSCs, notably toward the host vasculature.

To further explore the correlation between NRP1 and ITGB3 in GSC invasion, ITGB3 was efficiently knocked-down using RNA interference (Figure 4A). It is noteworthy that NRP1 expression was left intact in ITGB3-silenced cells, suggesting a non-reciprocal regulatory mechanism (Figure 4A). Again, viability and sphere formation were not significantly impacted by the knockdown of ITGB3 (Figures 4B,C). Recapitulating NRP1 deficiency, the 3D collagen-based invasion was drastically reduced upon ITGB3 silencing (Figure 4D). In keeping with this, the mean distance of ITGB3-knockdown GSC9 from vessels increased (Figure 4E), like NRP1 silencing phenotype. Collectively, these results suggest that ITGB3 silencing mimics the NRP1 inhibition-provoked phenotype, as invasiveness was repressed and GSCs were secluded from the host vasculature.