Meningeal solitary fibrous tumor cell states phenocopy cerebral vascular development and homeostasis

Meningeal solitary fibrous tumors (SFTs) are rare mesenchymal neoplasms that are associated with hematogenous metastasis, and the cell states and spatial transcriptomic architecture of SFTs are unknown. Here we use single-cell and spatial RNA sequencing to show SFTs are comprised of regionally distinct gene expression programs that resemble cerebral vascular development and homeostasis. Our results shed light on pathways underlying SFT biology in comparison to other central nervous system tumors and provide a framework for integrating single-cell and spatial transcriptomic data from human cancers and normal tissues.

cells with 30,934 cells from 6 meningioma samples 14 (Fig. 1g, Extended Data 2a-c, and Supplementary Table 2). Cell clusters were de ned using automated cell type classi cation, cell signature gene sets, cell cycle analysis, and differentially expressed cluster marker genes (Fig. 1h, Extended Data Fig. 2b, c). The distribution of cell states was analyzed across SFT and meningioma samples (Fig. 1i). SFT cell states were distinguished by expression of NAB2, STAT6, EGR1, VCAM1, NCAM2, and the endothelial cell marker CD34 (Fig. 1j). Meningioma cell states were distinguished by expression of SSTR2A, PDGFRB, and NOTCH3 (Fig. 1j). NOTCH3 drives meningioma tumorigenesis and is diffusely expressed in high-grade meningiomas that are resistant to radiation 14 . In contrast, immuno uorescence (IF) confocal microscopy showed NOTCH3 expression was restricted to a minority of cells in the perivascular niche in SFTs (Extended Data Fig. 3a, b). These data demonstrate that meningiomas and meningeal SFTs are distinguished by minimally overlapping tumor cell states despite their shared anatomic origin and imaging characteristics.
SFTs range from moderately to highly cellular neoplasms that are comprised of closely approximated, haphazardly arranged tumor cells with varying amounts of intervening collagenous stroma and numerous thin walled, ectatic, branching blood vessels 1 . To de ne the spatial transcriptomic architecture of meningeal SFTs, spatial RNA sequencing of 55µm regions from continuous tiled arrays across 6mm cores was performed on 8 SFT samples using an approach that integrates approximately 10 cells per capture area 15,16 . DNA methylation pro ling of SFTs (n = 8) and principal component (PC) analysis compared to DNA methylation pro ling of meningiomas (n = 213) 17 was used to validate SFT samples that were analyzed by spatial RNA sequencing (Extended Data Fig. 4). Tumor classi cation using DKFZ v12b6 18 showed the samples analyzed using spatial transcriptomics clustered in the SFT methylation class with con dence scores > 0.8 (n = 2) and > 0.99 (n = 6). Targeted next-generation DNA sequencing 19 of SFTs (n = 3) revealed NAB2-STAT6 gene fusions (n = 2) with breakpoints at exon 7 (NAB2) and exon 18 (STAT6), and multiple large-scale chromosome gains or losses 20,21 .
The spatial transcriptomic architecture of SFTs were analyzed across 2021 CNS World Health Organization (WHO) histological grades 1 (Fig. 2a and Extended Data Fig. 5a), or in patient-matched pairs of primary/recurrent or intracranial/metastatic samples ( Fig. 2b and Extended Data Fig. 5b). Spatial deconvolution of cell types from single-cell RNA sequencing (Fig. 1a) across WHO histological grades demonstrated regionally distinct SFT cell states and microenvironment cell types despite relatively uniform histological characteristics, including spatial heterogeneity in the distribution of macrophages (C6) or cell adhesion (C0, C2), ECM remodeling (C3), protein synthesis (C4), and mural (C7) SFT cells (Fig. 2a). Spatial NAB2 and STAT6 expression were relatively uniform across WHO histological grades, but cell stress (EGR1) and vascular (CD34, VWF, ACTA2) gene expression programs were heterogeneous (Extended Data Fig. 5a). Cell-cell communication analysis weighted for spatial transcriptome proximity across spatial RNA sequencing clusters showed VCAM1 interactions were heterogeneous and correlated with spatial expression of Integrin ligands that bind to VCAM1 (Fig. 2a). SFT cell states and microenvironment cell types also demonstrated regionally distinct expression in patient-matched pairs of primary/recurrent or intracranial/metastatic samples, which showed temporal or spatial evolution in protein synthesis (C4) and ECM remodeling (C5) SFT cells, or ECM remodeling SFT cells (C3) and macrophages (C6) (Fig. 2b). Spatial NAB2 and STAT6 expression were uniform across primary/recurrent and intracranial/metastatic samples, but cell stress (EGR1) and vascular (CD34, VWF, ACTA2) gene expression programs were heterogeneous (Extended Data Fig. 5b). Thus, meningeal SFTs demonstrate regionally distinct intratumor heterogeneity in cell states, gene expression programs, and cell-cell interactions across WHO histological grades and paired primary/recurrent or intracranial/metastatic samples.
These single-cell RNA sequencing (Fig. 1a-e, Extended Data Fig. 1, and Supplementary Table 1) and spatial RNA sequencing data (Fig. 2a, b Extended Data Fig. 5a, b) suggest the cell states and spatial transcriptomic architecture of meningeal SFTs resemble vascular cell types. To test this hypothesis, cell types from single-cell RNA sequencing of perinatal human brain vasculature (139,134 cells from gestational weeks 15, 17, 18, 20, 22, or 23) 22 or adult human brain vasculature (84,138 cells) 23 were deconvolved from SFT single-cell or spatial transcriptomes. SFT spatial transcriptomes (n = 23,682) were integrated and corrected for batch effects using Harmony and UMAP, and spatial transcriptome clusters were de ned using automated cell type classi cation, cell signature gene sets, cell cycle analysis, and differentially expressed cluster marker genes (Extended Data Fig. 6a-f and Supplementary Table 3).
Deconvolution of cell types underlying perinatal cerebral vascular development or adult cerebral vascular homeostasis revealed homology between perinatal venous cells, perinatal capillary cells, perinatal broblasts, adult venous cells, and adult pericytes in SFT single-cell and spatial transcriptomes, with lower homology to other cerebral vascular cell types (Fig. 2c). Thus, the cellular architecture of SFTs phenocopies endothelial and mural cell types involved in cerebral vascular development and homeostasis.
In summary, these data shed light on pathways underlying SFT biology in comparison to other intracranial meningeal tumors and provide a framework for integrating single-cell and spatial transcriptomic data from human cancers and normal tissues. We show SFTs are comprised of plastic cell states, regionally distinct gene expression programs, and cell-cell interactions that resemble cerebral vascular development and homeostasis. These data provide new insights into a rare, poorly understood cancer with a high rate of hematogenous metastasis that is unique in comparison to meningiomas and other tumors of the central nervous system.   Single-cell RNA sequencing and analysis

Declarations
Single cells were isolated from human SFT or meningioma samples, as previously described 17 . Singlecell suspensions were processed for single-cell RNA sequencing using the Chromium Single Cell 3' GEM, Library & Gel Bead Kit v3.1 (1000121, 10x Genomics) and a 10x Chromium or Chromium X controller, using the manufacturer recommended default protocol and settings for a target recovery of 5,000 cells per sample. Libraries were sequenced on an Illumina NovaSeq 6000, targeting >50,000 reads per cell, at the UCSF Center for Advanced Technology. Library demultiplexing, read alignment, identi cation of empty droplets, and UMI quanti cation were performed using CellRanger (https://github.com/10xGenomics/cellranger). Cells were ltered based on the number of unique genes, and single-cell UMI count data were preprocessed in R with the Seurat package (v4.3.0) 24,25 using the sctransform work ow. Dimensionality reduction was performed using PC analysis. UMAP and Louvain clustering were performed on the reduced data, followed by marker identi cation and differential gene expression.
Clusters were de ned using a combination of automated cell type classi cation 6 , cell signature gene sets 7 , cell cycle analysis, and differentially expressed cluster marker genes. The scType R package was used for automated cell type classi cation, with default parameters and an augmented list incorporating package-provided human 'Brain' marker genes speci c to each cell type 6 . Gene set enrichment analysis was performed on clusters using cell type signature gene sets from from MSigDB (https://www.gseamsigdb.org/gsea/msigdb) with the fgsea R package (Bioconductor v3.16). Cell cycle phases of individual cells were assigned with the 'CellCycleScoring' function in Seurat, using single-cell cell cycle marker genes 26 .
Meningeal SFT and meningioma single-cell samples were aligned to the GRCh38 human reference genome; ltered to cells with greater than 250 unique genes, less than 7,500 unique genes, and less than 25% of reads attributed to mitochondrial transcripts; scaled based on regression of UMI count and percentage of reads attributed to mitochondrial genes per cell; and corrected for batch effects using Deconvolution of SFT cell types from reference perinatal or adult vascular cell type single-cell RNA sequencing dataset was performed using SCDC (v0.0.0.9000) 27 . Single-cell transcriptomic data from the reference datasets were subsampled to 1000 cells per cluster, and the top differentially expressed genes were selected for each cell type. Using this expression set, SFT single cells were deconvolved to yield a matrix with predicted proportions of cell type for each cell, which were visualized using feature plots.
Gene enrichment analysis was performed using a list of 50 most differentially expressed candidate genes from previously published single-cell perinatal or adult vascular and mural cell types 22,23 . Average counts per cell were summarized, scored as mean, and visualized using feature plots.
Cell-cell communication networks were inferred and visualized using the CellChat R package (v1.5.0) 11 . Brie y, differentially expressed signaling genes were identi ed, noise was mitigated by calculating the ensemble average expression, intercellular communication probability was calculated by modeling ligand-receptor interactions using the law of mass action, and statistically signi cant communications were identi ed. The CellChat commands 'computeCommunProb', 'computeCommunProbPathway', and 'aggregateNet', were used for analysis, and 'netVisual_aggregate' was used for visualization.
Trajectory analyses was performed using monocle3 (v1.3.1) 10,28,29 for pseudotime, and velocyto (v0.17.16) 8 with scVelo (v0.2.5) 9 for RNA velocity. For pseudotime analysis, data were normalized followed by UMAP dimensionality reduction as described above. The 'cluster_cells' and 'learn_graph' monocle commands were used with default parameters and cells were ordered along pseudotime after manually selecting a root node (based on cluster, cell type, and cell cycle information). For RNA velocity analysis, velocyto was used to generate loom les with spliced and unspliced mRNA count information. scVelo was used to lter and normalize gene expression using criteria "min_shared_counts=3', and 'n_top_genes=2000' prior to computing RNA velocity and latent time. RNA velocity was visualized by projecting on to the UMAP generated using R and Seurat.
Spatial RNA sequencing and analysis Spatial transcriptomic pro ling was performed on FFPE sections using the 10x Genomics Visium Spatial assay (1000336, v1). 6 mm cores were mounted within capture areas on Visium glass slides, depara nized, stained with H&E, and imaged at the Gladstone Institutes Histology Core. Libraries were prepared according to manufacturer instructions at the Gladstone Institutes Genomics Core. Libraries were sequenced on an Illumina NovaSeq 6000 instrument at the UCSF Center for Advanced Technology.
Spaceranger generated ltered feature matrices were imported into a Seurat object (v4.3.0, arguments min.cells=3, min.features=100) using R (v4.2.1) and RStudio (v2022.07.2 Build 576). The individual count matrices were normalized based on nFeature_RNA count (subset=nFeature_RNA>1500 and nFeature_RNA<9500) with less than 10% of reads attributed to mitochondrial transcripts and integration Deconvolution of SFT cell types from reference SFT single-cell RNA sequencing was performed using SCDC (v0.0.0.9000) 27 . To do so, each spatial transcriptome was treated as a pseudobulked RNA sequencing dataset and leveraged against known cell types from reference single-cell RNA sequencing datasets comprised of 40,022 cells from 4 human SFT samples (Fig. 1a), perinatal human brain vasculature (139,134 cells from gestational weeks 15, 17, 18, 20, 22, or 23) 22 , or adult human brain vasculature (84,138 cells) 23 . Single-cell transcriptomic data were subsampled to 1000 cells per cluster, and the top differentially expressed genes were selected for each cell type. Using this expression set, spatial transcriptomes were deconvolved to yield a matrix with predicted proportions of cell type for each spatial transcriptome, which were visualized using spatial feature plots.
Gene enrichment analysis was performed using a list of 50 most differentially expressed candidate genes from previously published single-cell perinatal or adult vascular and mural cell types 22,23 . Average counts per spatial transcriptome were summarized, scored as a mean, and visualized using spatial feature plots.
The cell-cell communication network was inferred and visualized using the CellChat R package (v1.5.0) 11 similar to the method used for single cell RNA sequencing samples. Brie y, differentially expressed signaling genes were identi ed, noise was mitigated by calculating the ensemble average expression, intercellular communication probability was calculated by modeling ligand-receptor interactions using the law of mass action, and statistically signi cant communications were identi ed. The CellChat commands 'computeCommunProb', 'computeCommunProbPathway', and 'aggregateNet' were used for analysis, and 'netVisual_aggregate' was used for visualization. 'computeCommunProb' was run using spatial information from the Visium assay, including spatial dot coordinates and scale.factors for the ducials and low/high-resolution tissue images.

DNA methylation pro ling and analysis
Genomic DNA underwent bisul te conversion using the EZ DNA Methylation kit (D5004, Zymo Research), followed by ampli cation, fragmentation, and hybridization to In nium EPIC 850k Human DNA Methylation BeadChips (20020530, Illumina) according to manufacturer's instructions at the University of Southern California Molecular Genomics Core. Bioinformatic analysis was performed in R (v4.2.1). SFT or meningioma DNA methylation data were preprocessed using the mini pipeline 30 . In brief, probes were ltered and analyzed using normal-exponential out-of-band background correction, nonlinear dye bias correction, p-value with out-of-band array hybridization masking, and β value calculation (β=methylated/[methylated+unmethylated]). Principal component analysis was performed on the β methylation values from min pre-processing pipeline in R. Variable probes were identi ed from the rst three principal components (PCs). PCs greater than 4 contributed to less than 5% of β value variance. The top 2000 probes were selected for analysis by ranking the absolute gene loading score values within PCs and the tumors were projected along the rst two PCs.
Targeted next-generation DNA sequencing and analysis Targeted DNA sequencing was performed using the UCSF500 NGS panel, as previously described 19 . In brief, this capture-based next-generation DNA sequencing assay targets all coding exons of 479 cancerrelated genes, select introns, and upstream regulatory regions of 47 genes to enable detection of structural variants such as gene fusions and DNA segments at regular intervals along each chromosome

Statistics
All experiments were performed with independent biological replicates and repeated, and statistics were derived from biological replicates. Biological replicates are indicated in each gure panel or gure legend. No statistical methods were used to predetermine sample sizes, but sample sizes in this study are similar or larger to those reported in previous publications. Data distribution was assumed to be normal, but this was not formally tested. Investigators were blinded to conditions during clinical data collection and analysis. Bioinformatic analyses were performed blind to clinical features, outcomes, or molecular characteristics. The clinical samples used in this study were retrospective and nonrandomized with no intervention, and all samples were interrogated equally. Thus, controlling for covariates among clinical samples was not relevant. Cells and animals were randomized to experimental conditions. No clinical, molecular, or cellular data points were excluded from the analyses.

Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article. interactions and VCAM or NCAM interactions in SFTs using single-cell RNA sequencing cell-cell communication analysis. f, MR imaging, H&E staining, and IHC staining for STAT6 or SSTR2A in SFT versus meningioma. Scale bar, 100µm. g, h, Single-cell RNA sequencing UMAPs of 70,956 transcriptomes from SFT and meningioma samples shaded by tumor type of origin in g or by cell cluster in h. i, Cell cluster distribution across SFT and meningioma samples analyzed using single-cell RNA sequencing. j, Feature plots showing differentially expressed genes across UMAP clusters from SFT and meningioma samples analyzed using single-cell RNA sequencing.

Figure 2
Meningeal solitary brous tumor cell states are regionally distinct and phenocopy cerebral vascular development or homeostasis. a, b, MR imaging, H&E staining, spatial deconvolution of cell types from single-cell RNA sequencing, VCAM signaling networks weighted for spatial transcriptome proximity across spatial RNA sequencing clusters, and spatial expression of Integrin VCAM ligands across SFT CNS WHO histological grades in aor across patient-matched pairs of primary/recurrent and intracranial/metastatic samples in b. Scale bars, 10µm (top) and 100µm (bottom). c, Feature plots showing deconvolved perinatal or adult cerebral vascular cell types across single-cell RNA sequencing or spatial RNA sequencing UMAP clusters from SFT samples. Single-cells or spatial transcriptomes without concordance to cerebral vascular cell types are shown in grey.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.