Ethics statement

Written informed consent was obtained from all participants of this study. All procedures were performed in accordance with the Declaration of Helsinki and approved by the local ethics committees (University Hospital Essen, 19-8706-BO and 22-10564-BO; University Hospital Würzburg approval, 20230824 01).

Human biosampling

Clinical specimens were collected from patients newly diagnosed with IDH-wildtype glioblastoma, CNS WHO grade 4 (ref. 50), who had not undergone chemotherapy or radiotherapy. The specimens were obtained at the Department of Neurosurgery and Spine Surgery, University Hospital Essen. As control, tissue was collected from patients with nonmalignant intracranial disease (Supplementary Tables 1 and 2). At surgery, no patient suffered from acute infection or chronic inflammation. Calvarial bone chips derived during craniotomy from unplanned, intrasurgically required extensions of the burr hole or during necessary additional temporobasal decompression after craniotomy. Tumor tissue was obtained from contrast-enhanced, 5-aminolevulinic acid fluorescence, non-necrotic tumor areas by neuronavigation (Brainlab). Samples of tumor and paired bone were immediately stored in sterile Dulbecco’s Modified Eagle Medium (DMEM)/F12 (Gibco, 11320033), supplemented with antibiotics and antimycotics (2%; Gibco, 15240062). The standard collection of venous blood occurred at surgery or within 24 h. Standard dBM aspiration from posterior iliac crest was performed under general anesthesia before neurosurgery. Samples were immediately processed in the lab and registered at the Westdeutsche Biobank Essen (WBE; 22-WBE-137). Postsurgical CT scans were obtained within 24 h, and additional MRI scans of patients with glioblastoma within 72 h.

Clinical CXCR4 radiolabeling

PET–CT imaging data (University Hospital Wuerzburg)14 complemented data derived from presurgical 68Ga-labeled CXCR4 (Pentixafor) radiolabeling of patients with glioblastoma as part of clinical care at the University Hospital Essen (Supplementary Table 1). Intravenous (i.v.) administration of Pentixafor in Würzburg and Essen used activities of 1.94 ± 0.41 MBq kg−1 and 2.38 ± 0.39 MBq kg−1, respectively, followed by imaging 72 ± 14 min and 65 ± 19 min thereafter (mean ± s.d.). Integrated data (n = 19 histologically confirmed glioblastoma) underwent blinded consensus read by board-certified nuclear radiologists from both centers, using equal range settings. Cranial and calvarial enhancement was defined as focal uptake in the tumor-adjacent CB and absence of uptake in the contralateral reference point. Bridging tracer enhancement was classified as clearly distinguishable tracer transition between tumor and CB exceeding brain background uptake. Tracer uptake in the skin or in the venous sinuses was not assessed. As a control, patients not suffering from brain tumors (n = 6; Supplementary Table 2) received i.v. Pentixafor during clinical workup in Essen with an activity of 2.13 ± 0.25 MBq kg−1. Imaging was performed 81 ± 10 min thereafter on a Siemens Vision PET–CT scanner and CT–MRI fusion was conducted by board-certified nuclear medicine personnel using syngo.via (Siemens Healthineers) or Brainlab’s cranial navigation software (iPlanNet).

Tumor tissue processing

Within 30 min after resection, samples were minced and processed for derivation of primary cell cultures51. In parallel, single-cell suspensions were prepared18 by homogenizing tissue in Iscove’s Modified Dulbecco’s Medium (IMDM; Gibco, 12440053) with 0.11 DMC U ml−1 neutral protease (NP; Nordmark Biochemicals, S3030112) at 37 °C for ~30 min in a shaker incubator supported by intermittent resuspension. Cell suspension was filtered (35 µm cell strainer; Falcon, 352235) and washed twice with PBS (pH 7.4; Gibco, 14190169), supplemented with 0.04% BSA (Miltenyi Biotec, 130-091-376).

Bone sample processing

Bone chips were flushed with 0.11 DMC U ml−1 NP in IMDM for 10–15 min (at 37 °C) followed by PBS/0.04% BSA. Filtered cell suspensions (35 µm cell strainer) were centrifuged (10 min, 300g) and washed once in PBS/0.04% BSA. If available, excess bone tissue was flash frozen in liquid nitrogen and stored at −80 °C.

Blood sample and dBM processing

Blood and dBM samples arrived at room temperature in tubes containing EDTA or heparin for the isolation of PBMC or bone marrow mononuclear cells (BMMC) using Histopaque-1077 (Sigma-Aldrich, 10771) density gradient centrifugation, following the manufacturer’s protocol. Cells were washed twice in PBS/0.04% BSA.

Selection and preservation of immune cells

Single-cell suspensions from tumor tissue, bone and blood were enriched for vital CD45+ cells by the REAlease CD45 (TIL) MicroBead Kit (following the manufacturer’s protocol; Miltenyi Biotec, 130-121-563). Anti-CD45 antibodies were removed and cells were either used immediately or cryopreserved at −150 °C in 50% resuspension media (40% FBS in IMDM) and 50% freezing media (30% DMSO + 40% FBS in IMDM), according to CG00039 (10x Genomics). Derived samples were labeled sc-cohort 1 (Extended Data Fig. 3). Cells of sc-cohort 2 (Extended Data Fig. 9) underwent additional magnetic myeloid cell depletion by collecting the CD14− negative flow through (130-050-201).

scRNA-seq and analysis

Cell suspensions with >85% viable cells (trypan blue exclusion) were processed for scRNA-seq using Chromium Next GEM Single Cell 3′ Reagent Kit v3.1 and 5′ Reagent Kit v2 (10x Genomics, CG0000315 and CG0000331). Subsequent to quality control (2100 Bioanalyzer, Agilent), paired-end sequencing of pooled libraries was conducted on a NovaSeq 6000 System (Illumina). Reads were aligned to the hg38 human reference genome (2020) using Cell Ranger (v.7.0.1). The 5′ data, integrating V(D)J repertoire and gene expression, were processed with cellranger multi pipeline using 10x Genomics hg38 and V(D)J reference (7.0.0, GRCh38).

Analyses were performed in R (v4.2.0) on raw 3′ and filtered 5′ multi-output data. Using Seurat package (v.4.3.0)52, normalized cells (SCTransform) were filtered to remove cells with <500 or >7500 nFeature_RNA counts, or >15% mitochondrial genes and to identify doublets (DoubletFinder, v.2.0.3)53. The 3′ GEX (n = 21) and 5′ GEX/scVDJ (n = 8) Seurat objects (sc-cohort 1; Extended Data Fig. 3) were merged, cleaned of doublets and normalized regressing out mitochondrial percentage per cell and cell-cycle scores. Data integration used Harmony (v.0.1.1)54 by patient, followed by Seurat FindNeighbors (dims = 1:15) and FindClusters function (resolution = 0.6) with data visualization via RunUMAP (dims = 1:15; Fig. 2b,c). Cell-type annotation of integrated data was performed using SingleR (v.1.10.0)55 and marker-based identification via Seurat’s FindAllMarkers function and subsequent literature search. Expression of canonical marker gene sets was confirmed and visualized by gene set enrichment scores (AUCell score, v.1.18.1)56 (Extended Data Fig. 3c).

Myeloid cell compartment

Myeloid cell subset was refined by removing falsely clustered T cells (CD3D < 0.1), followed by normalization and data integration. Cell type annotation used Azimuth tool52 (v1.0.2) with GBMap dataset57. Only myeloid cells at annotation level 3 were kept, excluding cells expressing CD3, GFAP, OLIG1/OLIG2 or RBFOX3. The refined dataset was integrated by Harmony, followed by dimensional reduction using PHATE58 (v1.0.7) and cell-type identification via shared-nearest neighbor clustering.

Tumor-shared clonotypes in the CB

For integrating scVDJ information, TCRA/TCRB nucleotide sequences were assigned to T cells using Cell Ranger’s filtered contig annotation data (patients 4, 15 and 16; sc-cohort 1) and combineExpression function of scRepertoire (v.1.11.0)59. Differentially expressed genes (DEGs) were detected from tumor-shared clones ( ≥ 2 cells) versus nonexpanding singlets in the CB niche using Seurat’s FindMarkers() function with min.pct=0.2 (20% of cells). The top 10 DEGs (ranked by log2(FC)) were visualized using Seurat’s VlnPlot.

CD8+/− T cell compartment (sc-cohort 1)

T cells were subset removing falsely clustered myeloid cells (CD68 < 0.01), normalized and data integration reperformed. Using Seurat’s FindNeighbors (dims = 1:15), FindClusters (resolution = 0.5) and RunUMAP (dims = 1:15) functions, CD8+/CD4+ T cells, CD4+ Treg cells and MAIT cells were distinguished by cluster-based marker gene expression (Extended Data Fig. 3d,e). Cells lacking T cell genes (‘unknown’) or displaying high mitochondrial gene expression (‘low quality’) were excluded from analyses. CD8+ and CD4+ clusters were subsetted and remaining CD4- or CD8-expressing cells were removed (CD4 < 0.01/CD8A and CD8B < 0.01). A total of 6,550 cells remained unassigned. Their identity was determined by cluster-independent CD4 or CD8A/CD8B expression. Identified (CD4 < 1 × 10−15 and CD8A/CD8B > 1 × 10−15) CD8+ cells (n = 4,876) were added to the CD8+ T cell subspace for further analyses. Cells were normalized and integrated, followed by Seurat’s FindNeighbors (dims = 1:10), FindClusters (resolution = 0.5) and RunUMAP (dims = 1:15) commands for data visualization. Highly variable genes of the 11 distinct clusters were extracted via FindAllMarkers, and cellular identities were manually annotated (Extended Data Fig. 3f). Remaining MAIT cells, not belonging to CD8+ T cell subset, were re-assigned to the global T cell space (Extended Data Fig. 3d–f). Normalization and data integration revealed the final CD8+ T cell space (n = 18,973).

CB subspace assessment

The 3′ GEX CD8+ T cell subset of CB (n = 6,743 cells) was normalized (NormalizeData) using Seurat (v4.1.1), followed by FindVariableFeatures, ScaleData (default parameters) and RunPCA functions (npcs = 100). Following data integration by patient (Harmony), cells were ordered by differentiation trajectory using Python package Palantir24 (v1.0.1), visualized by RunUMAP (dims = 1:4). A numeric vector, predicting cellular status from least (1.0) to most (0.0) differentiated was generated from the RNA matrix by CytoTRACE25 (v.0.3.3). Pseudotime analysis was conducted using Monocle 3 (ref. 60; v.1.3.1). Following conversion into a CDS object using as.cell_data_set from SeuratWrappers, the cluster_cell and learn_graph functions from Monocle were applied. Location of naive CD8+ T cells was used to specify root node (order_cells function). Combined 3′ GEX CD8+ T data were used to compute cellular CytoTRACE scores of tumor and CB, visualized with ggplot2 (v.3.4.3; Extended Data Fig. 7a). VlnPlot2 (SeuratExtend v.0.6.0)61 was used to plot and compare (Wilcoxon test) CytoTRACE scores of CBe CD8+ T cells from both sources (Extended Data Fig. 7b).

Benchmarking to external signatures

Gene set enrichment scores of 19 curated CD8+ T cell gene signatures27 were computed by AUCell. Z scores across phenotypes and sources were calculated via CalcStats (SeuratExtend) and visualized as heatmap (Fig. 3h). Effector phenotypes were isolated and their activation:effector function signature27 AUCell score visualized (VlnPlot2, Seurat Extend).

Subanalysis of sc-cohort 2

Preprocessing of scData included removal of cells with <500 or >7500 nFeature_RNA counts, >15% mitochondrial genes and doublets before data integration. Normalization (SCTransform), including regression of mitochondrial read and cell cycle scores, and Harmony by patient was executed before FindNeighbors (dims = 1:15), FindClusters (resolution = 0.4) and RunUMAP (dims = 1:15) functions for data visualization (Extended Data Fig. 9b). SingleR and AUCell score of canonical T cell genes were employed to identify CD8+/CD4+ T cells, CD4+ Treg cells and MAIT cells (Extended Data Fig. 9c). Assignment of scVDJ information used Cell Ranger’s filtered contig annotation data and combineExpression function (scRepertoire). CD8+ T cells of sc-cohort 2 were subset and annotated by label transfer of sc-cohort 1 using singleCellNet62 (v.0.1.0; Extended Data Fig. 9d,e).

Tumor reactivity prediction

Gene count matrix was imported into R v4.1 and normalized using SCTransform on all genes (Seurat v.4). Normalized data was imported in Python with reactivity predicted by predicTCR32 model under xgboost (v1.7.4). Probability of reactivity was averaged for each clonotype, and threshold was determined using Fisher–Jenks natural break optimization. Clones with reactivity scores above threshold were designated as reactive and vice versa. For visualization, scVDJ data from sc-cohort 1 (patients 4, 15 and 16) and sc-cohort 2 (patients 21, 22 and 24) were integrated (n = 14,960). FindNeighbors (dims = 1:10), FindClusters (resolution = 0.5) and RunUMAP (dims = 1:10) functions were executed and data were visualized as UMAP.

T cell expansion

T cells were expanded from CD45+-enriched cells in T cell activation media (RPMI 1640 (Gibco, 72400021), human AB serum (10%; Sigma-Aldrich, H5667), sodium pyruvate (1 mM; Gibco, 11360039), β-mercaptoethanol (50 µM; Gibco, 21985023), antibiotic–antimycotic (1%), recombinant IL-2 (1000 U ml−1, 200-02), IL-15 (10 ng ml−1, 200-15) and IL-21 (10 ng ml−1, 200-21, all Peprotech)), similar to ref. 63. T cells expanded for 14–21 days in 96-well plates (Corning, 3596) with human T-activator CD3/CD28/CD137 Dynabeads (Gibco, 11163D) in a 1:5–10 bead-to-cell ratio. Before analyses, CD8+ T cells were enriched by magnetic separation (Miltenyi Biotec, 130-096-495), immediately used or stored at −150 °C.

ELISpot assays

Cellular IFNγ release (R&D Systems, EL285 and SEL285) was detected by incubating 10,000–20,000 bulk CD8+ T cells and 5,000–10,000 autologous tumor cells in 96-well plates (2:1, effector-to-target ratio). Autologous, short-term expanded tumor cells (passage 4–7) were prestimulated with IFNγ (1 µg ml−1; Peprotech, 300-02) for 48 h. T cells rested in reduced cytokine concentrations (20 U ml−1 IL-2, 1 ng ml−1 IL-15, 1 ng ml−1 IL-21) for at least 3 days and overnight in cytokine-free media. ELISpot assays were performed according to the manufacturer’s instructions after 24–48 h of co-incubation. MHCI/MHCII blockade was achieved by pre-incubating tumor cells with 5 µg ml−1 anti-HLA-DR (clone L243) and 5 µg ml−1 anti-HLA-A, anti-HLA-B and anti-HLA-C (clone W6/32) antibodies (BioLegend, 307648 and 311428) for 1 h. Background controls included wells with only CD8+ T cells or tumor cells. Spots were counted using ELISpot reader (AID iSpot, AID Autoimmun Diagnostika) and analyzed with Fiji Software (v1.0). MHC-dependent spots are defined as

$$\begin}\,}\,}(}8^}\; }+}\; }\; })\\-}\,}\,}(}8^}\; }+}\; }\; }+}\; })\end$$

Restimulation experiments

T cell activation was monitored by T cell clustering/aggregation during restimulation64. Expanded T cells rested in reduced cytokine conditions (see above) for at least 72 h. Restimulation cycles involved seeding 10,000 T cells in 96-well plates in activation media and CD3/CD28/CD137 Dynabeads (1:2, bead-to-cell ratio). Resilience assay evaluated CD8+ T cell fitness by counting successful restimulation cycles. Three 14-day-restimulation cycles, followed by 7 days of rest were performed in triplicates per patient and source. A restimulation cycle was successful if the mean cell count across all three wells exceeded the input of 10,000 cells per well. Resting T cells were cryopreserved in 80% FBS and 20% DMSO after expansion and restimulation.

Killing assay

Adapting protocols from refs. 65,66, we enriched tumor-reactive T cells by incubating 20,000 resting CD8+ T cells with 5,000 IFNγ stimulated autologous tumor cells on anti-CD28-coated 96-well plates (4 µg ml−1; BioLegend, 302934). Media consisted of one-fourth of maintenance media for primary tumor cells51 + three-fourths of T cell activation media. Cocultures were fed every other day and T cell outgrowth was incubated on fresh tumor cells weekly (up to 4 weeks). Derived T cells were used for the killing assay. Briefly, for patient 21, 20,000 tumor-reactive T cells were incubated at week 4 with 5,000 autologous tumor cells labeled with 1 µM CellTracker Red (Invitrogen, C34552) supplemented with caspase 3 substrate (NucView Biotium, 10402), according to manufacturer’s protocol. After 7 days, T cells were gently removed and live adherent tumor cells were detected (Nyone, Synentec) using 1 µg ml−1 Hoechst 33342 (Thermo Fisher Scientific, 62249). Celltracker signals and cellular morphology discriminated tumor cells and T cells, excluding Caspase-positive cells. MHCI blockade was achieved by pre-incubating tumor cells with 10 µg ml−1 anti-HLA-A, anti-HLA-B and anti-HLA-C (W6/32) antibodies for 1 h. During the assay, cells were fed once with fresh media and blocking antibodies at day 4.

Flow cytometry

Spectral flow cytometry-based immunoprofiles were detected using Cytek 25-Color Immunoprofiling Assay (Cytek Biosciences, R7-40002), with 18 cFluor reagents supplemented with seven antibodies from BioLegend (900004160) and ViaDye Red Fixable Viability Dye (Cytek), according to manufacturer’s protocol. Cryo-conserved single cells were thawed and washed twice before antibody labeling. Viability dye was used at 250 nM before blocking (Human TruStain FcX, BioLegend) and subsequent antibody labeling. Samples were measured on a Cytek Aurora flow cytometer in 5 l setup (16UV-16V-14B-10YG-8R), acquiring spectral profiles by SpectroFlo software (v3.0.3, Cytek). Unmixing was performed using the manufacturer’s recommended reference controls, with autofluorescence extraction enabled. Cell populations were quantified by recommended enhanced gating strategy (Cytek). Alternative gating to identify potential M-MDSCs was carried out via FlowJo (v10.9.0).

For phenotyping, indicated CD8+ cells from resilience assay were thawed, washed and incubated for 5 min with Fcγ receptor binding inhibitor (BD Pharmingen, 564220) before antibody labeling. The antibody cocktail consisted of BV421-CD95 (305623), BV711-CD8 (344733), BV510-CCR7 (353231), APC-CD4 (317415), FITC-CD161 (339905), PE-Cy7-CD3 (344815) and BV650-PD-1 (329949; all BioLegend), as well as BV786-CD45RA (563870) and PE-CD56 (555516) from BD Biosciences (all diluted at 1:20). Viability verified using 7AAD (Invitrogen, 00-6993-50). Cytometric profiling of S1PR1 on T cells was performed accordingly, by Fc-block and antibody panel (all diluted at 1:20)— BV421-CD95 (305623), BV711-CD8 (344733), BV510-CCR7 (353231), PE-CD4 (317410), PE-Cy7-CD3 (344815) and BV650-PD-1 (329949; all BioLegend), as well as BV786-CD45RA (BD Biosciences, 563870) and eFluor660-S1PR1 (50-3639-42) or respective Isotype Control (50-4714-82; both Thermo Fisher Scientific). Samples were incubated on ice in the dark for 30 min, washed and measured using FACS Celesta and FACS Diva software (v8.0.1.1, BD Biosciences), with FlowJo sub-analysis (v10.9.0). Gating strategies are depicted in the corresponding Extended Data Figs. 4–6.

Whole-mount staining and optical clearing

CB samples were fixed in 4% PFA (in PBS, pH 7.4) overnight at 4–8 °C and blocked (5% DMSO, 0.1% Tween 20, 1% BSA and 5 mM EDTA in PBS) for two days at room temperature. Immunofluorescence labeling was performed with PE/Dazzle594-CD45 (304052) and AlexaFluor647-CD34 (343508) BioLegend antibodies diluted 1:200 in blocking buffer for 5 days at room temperature. Samples were washed twice with 5% DMSO and 0.1% Tween 20 in PBS for 1 day at room temperature, respectively. Optical tissue clearing was performed by established methods15. Briefly, dehydration in increasing ethanol concentrations of 50%, 70% and 100% (room temperature, 1 day each) was followed by optical clearing in ethyl cinnamate (ECi; Sigma-Aldrich, 112372) at room temperature to achieve complete transparency.

Light-sheet fluorescence microscopy

ECi-cleared CB was imaged via light-sheet fluorescence microscopy, using a LaVision BioTec Ultramicroscope Blaze (Miltenyi Biotec and LaVision BioTec) with supercontinuum white light laser (460–800 nm), seven excitation and emission filters covering 450–865 nm, AndorNeo sCMOS Camera with pixel size of 6.5 × 6.5 µm2 , and ×1.1 (NA 0.1), ×4 (NA 0.35) and ×12 (NA 0.53) objectives with magnification changer ranging from ×0.66 to ×30. Cleared samples were immersed in ECi in a quartz cuvette and imaged using excitation (ex) and detection band-pass emission (em) filter settings—tissue autofluorescence, ex 500/20 nm, em 535/30 nm; CD45-PE-Dazzle594, ex 560/40 nm, em 650/50 nm; CD34-AlexaFluor647, ex 630/60 nm, em 680/30 nm. The z-step size was set to 5 or 10 µm based on the selected light-sheet NA. Depending on the objectives, optical zoom factor varied from ×4 to ×12, with a digital zoom factor of ×1. Data were processed with visualization tools from Imaris (Bitplane, v9.7.1).

Confocal laser scanning microscopy to assess TLS formation

CB samples were decalcified in 14% EDTA-free acid solution (pH 7.2) for 14 days at room temperature, washed with PBS, embedded in O.C.T. Compound (Sakura, 4583) and snap frozen. Tissue sections of 20 µm were generated on a CryoStar NX70 (Thermo Fisher Scientific) using Kawamoto’s film method (Section Lab, Cryofilm type 2C9) and stored at −20 °C. For immunofluorescence labeling, tissue sections were blocked (1% BSA, 0.1% Tween 20 and 0.1% DMSO in PBS) for 1 h at room temperature, and incubated with PE/Dazzle594 CD3 (1:100; BioLegend, 300450), AlexaFluor488 CD20 (1:100; Thermo Fisher Scientific, 53-0202-80) and DAPI (1:500; Carl Roth, 6335.1) in blocking buffer overnight (4–8 °C). Samples were washed thrice with washing buffer for 15 min at room temperature, and once with distilled water, covered with mounting medium (Agilent Dako, S3023) and imaged via high-resolution confocal laser scanning microscopy on a Leica TCS SP8 confocal laser scanning microscope equipped with acousto-optic tunable filters, an acousto-optical beam splitter, internal hybrid detectors (HyD SP) with use of an LMT200 high precision scanning stage. A Leica HC PL APO ×63/1.20W CORR objective combined with a digital zoom factor of 1.0 was used for imaging of sequential scans as follows: (1) CD20-AlexaFluor488, ex 488 nm (argon laser), em 500–550 nm; (2) CD3-PE/Dazzle594, ex 561 nm, em 600–650 nm and (3) Dapi, ex 405 nm, em 450–500 nm, with the last two being excited by a diode-pumped solid-state laser. The 3D reconstruction used Imaris software (v9.7.1, Bitplane) at maximum intensity projection.

Quantitative multiplex immunofluorescence imaging

CB samples were fixed with 4% methanol-free formaldehyde (Thermo Fisher Scientific) overnight with rotation at 4 °C. Decalcification (10% EDTA, pH 8; Sigma-Aldrich) for 14 days at room temperature with stirring was followed by dehydration (overnight) and paraffin embedding. Tissue sections of 10 µm were cut (pfm Slide 4004 M sledge microtome), deparaffinized, rehydrated and antigen-retrieved according to the manufacturer’s instructions (Agilent Technologies). Sections were blocked and permeabilized with TBS (0.1 M Tris, 0.15 M NaCl, pH 7.5) containing 0.05% Tween 20, 20% DMSO (Sigma-Aldrich) and 10% donkey serum (Jackson ImmunoResearch) for 15 min at room temperature. Antibodies (1:25) and DAPI were diluted in DAKO EnVision FLEX diluent (Agilent Technologies). Primary antibodies (CD45—Bio-Rad, MCA345G; CD146—R&D, AF932; CXCR4—Thermo Fisher Scientific, PA3-305) were applied overnight. Secondary antibodies (donkey anti-rabbit 488, anti-goat 555, anti-rat 594 (all Biotium)) were incubated for 5 h and DAPI (Thermo Fisher Scientific) was applied before mounting (Vector Laboratories, H-1400-10). Labeled sections were imaged on a Leica Stellaris 8 laser scanning confocal microscope equipped with 2x HyD-S, 2x HyD-X and one HyD-R detectors and two laser lines (405 and white-light laser) using ×20 multiple-immersion objective (NA 0.75, FWD 0.680 mm) at 400 Hz, 8-bit with 1024 × 1024 resolution.

Statistics and reproducibility

Statistical methods, sample size and replication for each experiment are indicated in the figure legends. Flow cytometry and ELISpot statistical analyses were performed using GraphPad Prism v9.5.1 or Microsoft Excel v16.79.2. Statistical analysis of survival data was executed in SPSS (v.29.0.2.0). For collection of tissue samples and clinical imaging data, no statistical method was used to predetermine sample size, but our sample sizes are similar to those reported in previous publications67,68,69. Tissue samples were collected consecutively. The sex of a patient was self-reported. No gender information was collected and sex was not considered in the study design. scRNA-seq data with low quality (see above) and patients not meeting inclusion criteria for survival analysis (Extended Data Fig. 10a) were excluded from the study. Experiments were not randomized. Except for PET data association with patient survival (Fig. 4m–o and Extended Data Fig. 10), data collection and analysis were not performed blind to the conditions of the experiments. In parametric statistical tests, data distribution was assumed to be normal but this was not formally tested.

Reporting summary

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