Abstract

   Solitary fibrous tumors (SFTs) of the central nervous system (CNS) are
   rare, aggressive mesenchymal neoplasms with high recurrence and
   metastasis rates. Here, we perform a comprehensive multi-omics analysis
   of 189 cases of CNS SFTs integrating 94 whole genome sequencing, 88
   transcriptomics, 7 single-nucleus RNA sequencing and 3 spatial
   transcriptome sequencing. We find that receptor tyrosine kinase
   mutations are significantly more prevalent besides the widespread
   NAB2-STAT6 fusion and correlate with tumor grade. Transcriptomic
   analysis reveals four molecular subtypes—classical, neural-like,
   inflamed and migratory—each associated with distinct clinical and
   biological characteristics. Single-nucleus RNA sequencing identifies
   five tumor cell states, with the SFT_classical state serving as a
   precursor to other states influenced by hypoxia and inflammation.
   Notably, we identify FER kinase as a key therapeutic target, with FER
   inhibition significantly reducing tumor cell proliferation, migration
   and invasion. These findings provide important insights into CNS SFT
   biology and suggest potential therapeutic strategies for this
   challenging tumor type.

   Subject terms: CNS cancer, Targeted therapies, Cancer genomics,
   High-throughput screening, Transcriptomics
     __________________________________________________________________

   Solitary fibrous tumors (SFTs) of the central nervous system (CNS) are
   uncommon mesenchymal neoplasms with limited treatment options. Here,
   the authors perform a multi-omics analysis of 189 cases of CNS SFTs to
   identify potential clinical biomarkers and therapeutic targets.

Introduction

   Solitary fibrous tumor (SFT) is a rare mesenchymal spindle cell
   neoplasm with unpredictable clinical and biological behavior.
   Molecularly defined by the NAB2-STAT6 gene fusion^[58]1,[59]2, SFT
   accounts for <2% of all soft tissue sarcomas. Although SFT can arise in
   various anatomical locations including the central nervous system
   (CNS), CNS SFT was not distinctly documented until 1996 by Carneiro et
   al.^[60]3.

   CNS SFTs represent <1% of all primary CNS tumors but are frequently
   misdiagnosed as meningiomas due to their dural attachment and similar
   radiographic appearance^[61]4. CNS SFTs demonstrate more aggressive
   behavior than meningiomas, with recurrence rates reaching 57% even
   after complete resection, compared to only 5% for
   meningiomas^[62]5,[63]6. These tumors also exhibit significant
   metastatic potential, spreading to extracranial sites including the
   liver, lungs, bone, and peritoneum^[64]7. Malignant transformation and
   dedifferentiation further compromise prognosis and can rapidly lead to
   patient mortality^[65]8,[66]9.

   Surgical resection remains the primary treatment for CNS SFTs, with
   gross total resection serving as the most important prognostic
   factor^[67]4. Achieving complete resection often proves technically
   challenging due to the tumors’ deep location and aggressive nature.
   While adjuvant radiotherapy may enhance local control, its effects on
   metastasis prevention and overall survival remain uncertain^[68]10.
   Traditional chemotherapy shows limited efficacy against SFT, and
   standardized systemic treatment protocols are lacking^[69]11.

   The latest World Health Organization (WHO) classification distinguishes
   CNS SFTs from peripheral SFTs based on their distinct
   pathophysiological features and worse prognosis^[70]11,[71]12. Despite
   recent progress in understanding their clinical and molecular
   characteristics, the mechanisms driving SFT recurrence, malignant
   transformation, dedifferentiation, and distant metastasis remain
   incompletely understood^[72]13,[73]14. Additionally, systematic
   research on potential therapeutic targets—including signaling pathway
   aberrations associated with aggressive phenotypes and dynamic changes
   in the immune microenvironment—remains insufficient.

   To address these knowledge gaps, we conducted a multi-omics analysis of
   189 histologically confirmed CNS SFT cases and 8 peripheral SFT samples
   from three neurosurgical centers. We employed single-nucleus
   transcriptomic sequencing, spatial transcriptomic sequencing,
   whole-genome sequencing, and transcriptomic sequencing to
   comprehensively characterize the molecular landscape of CNS SFT.
   Through primary SFT cell culture, high-throughput small molecule drug
   screening, and lentivirus-mediated gene intervention experiments, we
   identified potential clinical biomarkers and therapeutic targets,
   providing insights to enhance the diagnosis, treatment, and prognostic
   management of CNS SFTs (Fig. [74]1A).

Fig. 1. Landscape of high-frequency driver mutations and associated
functional pathways in SFT.

   [75]Fig. 1
   [76]Open in a new tab

   A Overview of the experiment design and the number of samples for WGS,
   bulk RNA seq, single nucleus RNA seq, spatial transcriptome seq and
   primary cell culture. Created in BioRender. Wang, T. (2025)
   [77]https://BioRender.com/8urwveb. B Landscape of high-frequency
   candidate driver mutations in SFT. C KEGG pathway enrichment analysis
   of genes with non-synonymous mutations identified in all WGS samples.
   Term Enrichment is performed by R package clusterProfiler. P-value was
   calculated by two-sided Kolmogorov–Smirnov test and adjusted by FDR. D,
   E Representative results and quantification of p-ERK, p-S6, and p-AKT
   immunohistochemistry (IHC) staining in different grades of SFTs. D
   Representative images of p-ERK, p-S6, and p-AKT IHC staining in
   different grades of SFTs. E Quantification of the relative intensity
   scores of p-ERK, p-S6, and p-AKT staining across different SFT grades.
   For p-ERK: Grade 1 (n = 12), Grade 2 (n = 16), Grade 3 (n = 9); for
   p-S6: Grade 1 (n = 13), Grade 2 (n = 15), Grade 3 (n = 9); for p-AKT:
   Grade 1 (n = 9), Grade 2 (n = 12), Grade 3 (n = 11). Data are presented
   as mean ± SEM. *p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001.
   (Student’s t-test, two-sided). Source data are provided as a Source
   Data file.

Results

Whole genome sequencing of CNS SFTs

   To gain a comprehensive understanding of the genetic underpinnings of
   CNS SFTs, we performed whole genome sequencing (WGS) on 94 CNS SFT
   samples. Post-processing steps, including trimming and alignment,
   yielded an estimated average sequencing depth of ~32×. In agreement
   with previous studies^[78]1,[79]2, NAB2-STAT6 DNA fusions were detected
   in 87.23% (82/94) of the samples (Fig. [80]1B, Supplementary
   Data [81]1). Following the established framework by Robinson D. R. et
   al.^[82]1, we systematically categorized the fusion transcripts based
   on their protein domain in both NAB2 and STAT6 genes. This
   classification was primarily based on the key protein domains contained
   within each of these genes: NAB2 comprises a N-terminal EGR1-binding
   domain (EBD), a NAB conserved domain (NCD2), and a C-terminal
   transcriptional repressor domain (RD); whereas STAT6 contains a
   coiled-coil domain 1 (CCD1), a DNA-binding domain (DBD), a Src homology
   2 (SH2) domain, and a C-terminal transcriptional activation domain
   (TAD). This domain-based analysis revealed statistically significant
   differences in the prevalence of specific fusion subtypes, notably the
   RD-CCD1 variant (fusion between the RD of NAB2 and the CCD1 of STAT6),
   across WHO-defined diagnostic categories (Supplemental Fig. [83]1A). In
   addition, the fusion frequency of RD-SH2 (fusion between the RD of NAB2
   and the SH2 of STAT6) tends to be higher in high-grade tumors than in
   low-grade ones. Furthermore, mutations in TP53 (6%) and the TERT
   promoter (Not observed) were found to be relatively rare. The
   epigenetic regulators EP300 and ASXL1 were the most frequently mutated
   genes in CNS SFTs, albeit with mutation rates not exceeding 10%
   (Fig. [84]1B). Interestingly, mutations in these genes are also common
   in certain hematologic malignancies, such as acute myeloid leukemia
   (AML)^[85]15,[86]16.

   Although no other dominant mutations were identified, we observed a
   notable enrichment of mutations in receptor tyrosine kinases (RTKs)
   involved in cell migration and signaling pathways (Fig. [87]1B, C).
   Immunohistochemical analysis revealed a significant increase in
   phosphorylated extracellular signal-regulated kinase (p-ERK),
   phospho-S6 ribosomal protein (p-S6) and phosphorylated protein kinase B
   (p-AKT) levels in higher grade tumors (Fig. [88]1D, E), reflecting the
   increasing proportion of RTK mutations across different tumor grades
   (Grade 1: 36.84%, Grade 2: 56.82%, Grade 3: 64.52%). These results
   suggest a potential link between RTK pathway activation and CNS SFT
   progression, which is reflected by both protein expression and genomic
   alterations. Similar to this, as the tumor grade increases, the overall
   mutation burden rises, accompanied by shifts in mutation patterns
   (Supplemental Fig. [89]1B, [90]C).

Transcriptomic analysis and molecular subtypes in CNS SFTs

   To further elucidate the transcriptional landscape of SFTs, we
   conducted RNA sequencing (RNA-seq) on 96 rRNA-depleted SFT tissue
   samples, including 84 CNS SFTs and 8 peripheral SFTs after removing
   four low-quality samples (Supplemental Data [91]1). The NAB2-STAT6
   fusion was identified in 93.48% (86/92) of the samples using
   FusionCatcher (Supplemental Data [92]1). Fusion domain analysis of our
   RNA sequencing dataset confirmed malignancy-associated fusion patterns,
   with the RD-SH2 variant showing a significant correlation with
   high-grade histopathological features (Supplemental Fig. [93]1D).
   Integrated analysis of genomic and transcriptomic results revealed that
   RD-CCD1 is associated with lower-grade CNS SFTs. (Supplemental
   Fig. [94]1E).

   Transcriptomic comparisons between CNS and peripheral SFTs revealed the
   expression of several neuron-specific genes (e.g. NRCAM^[95]17,
   ZIC2^[96]18, AQP4^[97]19) in CNS SFTs (Fig. [98]2A, Supplemental
   Fig. [99]1F). Immunohistochemical analysis confirmed the abundant
   expression of the neuronal marker Synaptotagmin^[100]20 in CNS SFTs, in
   contrast to the more generalized expression of the extracellular matrix
   protein Decorin in both CNS and peripheral SFTs (Fig. [101]2B–D).
   Single-cell sequencing data further corroborated that these neuronal
   markers were directly expressed by the tumor cells (Supplemental
   Fig. [102]2A).

Fig. 2. Transcriptomic characterization and molecular subtypes of CNS SFTs.

   [103]Fig. 2
   [104]Open in a new tab

   A Volcano plot comparing gene expression between CNS and peripheral SFT
   tumors. The names of the genes marked in red are genes that are highly
   expressed in the nervous system. Differential analysis was performed
   using the R package DESeq2, and the p-value was calculated using the
   two-sided Wald test and corrected with FDR. B–D Representative IHC
   staining results for Synaptotagmin and Decorin in CNS SFT (n = 6) and
   peripheral SFT (n = 11) (B) and their quantitative results (C, D). Data
   are presented as mean ± SEM. ****, p < 0.0001 (Student’s t-test,
   two-sided). E GO enrichment analysis of differentially expressed genes
   between WHO grade 3 and WHO grade 1 SFT. F GO enrichment analysis of
   differentially expressed genes between CNS SFT and meningioma (a merged
   dataset of [105]GSE136661 and [106]GSE189672). Term Enrichment is
   performed by R package clusterProfiler. P-value was calculated by
   two-sided Kolmogorov–Smirnov test and adjusted by FDR. G Sankey plot
   illustrating the relationship between WHO grade and molecular subtypes.
   H Heatmap showing significantly activated signaling pathways across
   four molecular subtypes. I–K The maximum percentage of Ki67% (I), the
   log10-transformed tumor volume (J) and the frequency of cystic necrosis
   according to radiographic result (K) in different molecular subtypes of
   CNS SFT. Neural-like (n = 13), Classical (n = 37), Inflamed (n = 13),
   Migratory (n = 17). Data are presented as mean ± SEM. *, p < 0.05; **,
   p < 0.01 (For (I) and (J) Student’s t-test, two-sided; for (K)
   Chi-square test, two-sided). Source data are provided as a Source Data
   file.

   Analysis of transcriptomic profiles among different CNS SFT grades
   revealed relatively minor heterogeneity, with WHO grade 2 tumors
   exhibiting overlapping characteristics with both grade 1 and grade 3
   tumors (Supplemental Fig. [107]1G, [108]H). Consistent with the WGS
   data, signaling pathways related to cell surface tyrosine kinase
   activity were significantly enriched in WHO grade 3 tumors compared to
   WHO grade 1 (Fig. [109]2E). Since CNS SFTs often have similar imaging
   features to meningiomas^[110]21, we also compared the transcriptomes of
   CNS SFTs and meningiomas. Interestingly, angiogenesis signaling and
   cell surface receptors, particularly insulin-like growth factor-related
   receptors, were more enriched in CNS SFT samples (Fig. [111]2F).

   To delineate molecular subtypes of CNS SFTs, we employed non-negative
   matrix factorization (NMF) clustering, which classified the samples
   into four distinct subtypes (Fig. [112]2G, Supplemental Fig. [113]1I).
   Similar to the differences in mutation burden among samples of
   different WHO grades, there are also variations in mutation burden
   among the subtypes defined in our transcriptomic classification.
   However, the types of mutations present in each subtype differ, which
   may be related to the characteristics of the respective samples
   (Supplemental Fig. [114]1J, [115]K). The most prevalent subtype, termed
   the “Classical” subtype, was characterized by traditional SFT features.
   Gene Set Variation Analysis (GSVA) identified a “Neural-like” subtype
   enriched in cell differentiation and neurotransmitter secretion
   pathways (Fig. [116]2H, Supplemental Fig. [117]1L). Another subtype,
   termed “Inflamed”, exhibited high activation of the NF-κB pathway,
   while the “Migratory” subtype showed enrichment in pathways related to
   cytoskeletal remodeling (Fig. [118]2H, Supplemental Fig. 1L). Clinical
   correlation revealed that the Neural-like subtype predominantly
   comprised WHO grade 1 and 2, which exhibited lower proportion of
   Ki67-positive cells and generally smaller volumes (Fig. [119]2I, J).
   The Classical subtype, on the other hand, displayed the highest
   incidence of cystic necrosis (Fig. [120]2K).

Diverse cell states in SFT tumors revealed by single-nucleus RNA sequencing

   To explore the heterogeneity of SFT tumor cells and the tumor
   microenvironment, we performed single-nucleus RNA sequencing
   (snRNA-seq) on seven SFT tissues of different grades and obtained
   58,847 high-quality nuclei. We integrated samples across a broad
   spectrum of characteristics and classifications without applying
   selection biases to ensure comprehensive representation. While cell
   type proportions varied somewhat among samples, all specimens contained
   diverse cell populations (Supplemental Fig. [121]2B). The tumor cells
   were classified into five distinct states, with the “SFT_Classical”
   state being the most prevalent across grades, which expressed several
   SFT and mesenchymal cell markers such as CD34^[122]22, CD44^[123]23,
   ALDH1L1^[124]24, and ENG (CD105)^[125]25 (Fig. [126]3A–C, Supplemental
   Fig. [127]2C, D). Regulon analysis revealed an enrichment of
   transcription factors involved in limb and craniofacial transcriptional
   regulation during embryonic development and epithelial-mesenchymal
   interactions, including PRRX2^[128]26, HOXB6^[129]27, ALX4^[130]28 and
   GLI3^[131]29 (Fig. [132]3D). RNA velocity and pseudotime trajectory
   analyses suggested that cells in the SFT_Classical state cells might
   serve as precursors for other SFT cell states (Fig. [133]3E, F).

Fig. 3. Cellular states of CNS SFTs by snRNA-seq.

   [134]Fig. 3
   [135]Open in a new tab

   A Uniform Manifold Approximation and Projection (UMAP) diagram of the
   different SFT states and cell types of the tumor microenvironment
   identified by snRNA-seq. SFT_Classical, SFT_Inflamed, SFT_Neural,
   SFT_Migratory, SFT_Angiogenic: SFT cells with different cellular
   states; Endo: Endothelial cell; Peri: Pericyte; Macro: Macrophage;
   Fibro: Fibroblast. B Proportion of different SFT cellular states in
   seven samples subjected to snRNA-seq sequencing (n = 7). Data are
   presented as mean ± SEM. *, p < 0.05; **, p < 0.01. (Student’s t-test,
   two-sided). C Projection of CD34, CD44, ALDH1A1 and ENG expression on
   the UMAP plot of the major cell types identified in SFT snRNA-seq. D
   Heatmap of SCENIC binary regulon activities in different SFT states. E
   RNA velocity plot embedded in UMAP space showing the putative
   near-transcriptional state of the different CNS SFT cellular states. F
   Pseudotime trajectory analysis of CNS SFT cell states. G Gene set
   variation analysis (GSVA) of gene sets or pathways enriched in
   different SFT cellular states. H Projection of the average expression
   of 200 genes from the HALLMARK_HYPOXIA gene set onto the UMAP plot of
   SFT snRNA-seq. I Proportion of different cell types deconvoluted from
   the bulk transcriptomic data (n = 88). Only cell types with a
   proportion >5% in the sample are shown. J Proportion of SFT-specific
   cell types in the four molecular subtypes (total sample n = 88).
   *p < 0.05; **p < 0.01; ***p < 0.001. (two-sided Wilcoxon Rank Sum and
   Signed Rank tests). Boxplots display the median (center line),
   interquartile range (box), and whiskers extending to the smallest and
   largest values within 1.5 × IQR from the lower (Q1) and upper (Q3)
   quartiles, respectively. Source data are provided as a Source Data
   file.

   Pathway enrichment analysis revealed distinct biological
   characteristics among the remaining four cell states (Fig. [136]3G):
   the “SFT_Inflamed” state exhibited activation of TNF signaling and
   apoptotic pathways; the “SFT_Neural-like” state was characterized by
   expression of neuronal markers; the “SFT_Migratory” and
   “SFT_Angiogenic” states, predominantly observed in grade 3 tumors, were
   enriched in hypoxia-related pathways (Fig. [137]3H), with the
   SFT_Migratory state also showing upregulation of TGF-β signaling and
   cell migration pathways^[138]30.

   To further explore the proportions of different SFT cell states across
   various tumor molecular subtypes, we employed BayesPrism^[139]31 for
   the deconvolution of bulk RNA-seq results. The analysis revealed that,
   consistent with snRNA-seq findings (Supplemental Fig. [140]2D), the
   SFT_Classical state was the most prevalent across all molecular
   subtypes (Fig. [141]3I, J). The SFT_Neural-like state was more abundant
   within the Neural-like subtype, while SFT_Inflamed cells were least
   represented in this subtype. As expected, the SFT_Angiogenic and
   SFT_Inflamed states were primarily observed in the Inflamed and
   Migratory subtypes (Fig. [142]3I, J).

   Spatial transcriptomics was employed to map the spatial distribution of
   the different SFT cell states (Supplemental Fig. [143]3A). The results
   were largely corroborated the findings from single-cell sequencing,
   with the SFT_Classical state being the most abundant. The
   SFT_Neural-like state was found almost exclusively in WHO grade 1
   tumors (Fig. [144]4A). Spatial analysis further revealed that the cells
   of the SFT_Inflamed state were concentrated in regions with vascular
   leakage or inflammatory cell infiltration. The SFT_Migratory state was
   detected in moderately hypoxic areas, while the SFT_Angiogenic state
   cells were confined to regions with extreme hypoxia (Fig. [145]4B,
   Supplemental Fig. [146]3B, [147]C). These results suggest that
   different SFT cell states may evolve from the SFT_Classical state in
   response to different tumor microenvironments, particularly under
   different inflammatory and hypoxic conditions.

Fig. 4. Spatial heterogeneity and cell-cell interactions in the CNS SFTs.

   [148]Fig. 4
   [149]Open in a new tab

   A Proportion of distinct cell types in three spatial samples,
   categorized by WHO grade 1-3. B Hematoxylin and eosin (H/E) staining of
   representative regions of different SFT grades and the spatial
   distribution of SFT cellular states in the corresponding regions as
   well as the average expression levels of genes from the
   HALLMARK_HYPOXIA gene set. A total of six samples were included: one
   grade I, two grade II, and one grade III. Staining was performed in
   three independent replicates. C Proportion of cell types from the tumor
   microenvironment in the four molecular subtypes (total sample n = 88).
   *p < 0.05; **p < 0.01; ***p < 0.001. (two-sided Wilcoxon Rank Sum and
   Signed Rank tests). Boxplots display the median (center line),
   interquartile range (box), and whiskers extending to the smallest and
   largest values within 1.5 × IQR from the lower (Q1) and upper (Q3)
   quartiles, respectively. D The immune infiltration in different
   molecular subtypes by EPIC (left) and MCP-counter (right). Two-sided
   Wilcox test was used to calculate the differences among four subtypes.
   E Cell-cell communication between different cell types in CNS SFTs
   deduced from snRNA-seq. Source data are provided as a Source Data file.

Tumor microenvironment composition in CNS SFTs

   Deconvolution analysis of bulk RNA-seq data revealed differences in the
   composition of the tumor microenvironment across four SFT molecular
   subtypes (Fig. [150]4C). Unlike the complex microenvironment observed
   in malignancies such as glioblastoma^[151]32, the SFT microenvironment
   was relatively simple, consisting of endothelial cells, pericytes,
   macrophages, fibroblasts, and a sparse population of T cells
   (Figs. [152]3A, [153]4C). Immunohistochemical analysis confirmed a
   higher proportion of T cells compared to B cells or neutrophils within
   the SFT microenvironment (Supplemental Fig. [154]3D, [155]E). Of note,
   the Inflamed subtype had the highest proportion of T cells, whereas the
   Migratory subtype exhibited the greatest abundance of macrophages
   (Fig. [156]4C). We assessed immune cell infiltration levels across
   different subtypes using EPIC and MCP_counter algorithms for immune
   deconvolution^[157]33. This analysis demonstrated comparable immune
   cell infiltration levels in the “Inflamed” and “Classical” subtypes,
   with both showing significantly higher infiltration than the
   “Neural-like” subtype according to both algorithms. The “Inflamed”
   subtype specifically exhibited greater prominence of CD4 T cells and
   cytotoxic T cells, suggesting their potential functional importance in
   this subtype’s immune landscape (Fig. [158]4D, Supplemental
   Fig. [159]3F). Single-cell sequencing further validated the
   identification of multiple receptor-tyrosine kinase (RTK)
   ligand-receptor interactions, including VEGFA-KDR, PDGFB-PDGFRB,
   IGF1/2-IGF1R, HBEGF-EGFR and FGF2-FGFR1 (Fig. [160]4E, Supplemental
   Fig. [161]4A, [162]B). Among these, VEGF and PDGF were mainly produced
   by endothelial cells and SFT_angiogenic state cells, pericytes provided
   PDGFB, macrophages primarily supplied the EGFR ligand HBEGF, while
   IGF-IGFR1 formed an autocrine loop within SFT tumor cells
   (Fig. [163]4E). Consistent with this phenomenon, large peripheral SFT
   tumors secreting large amounts of IGF can lead to paraneoplastic
   syndromes, a phenomenon that has not yet been reported in CNS SFT
   tumors, possibly due to the limited intracranial space restricting
   tumor size^[164]34,[165]35. In addition, fibroblasts and SFT_Migratory
   cells secreted large amounts of TGF-β, which may induce
   epithelial-mesenchymal transition (EMT), cell migration and tumor
   metastasis.

FER kinase as a therapeutic target in CNS SFTs

   To investigate the therapeutic vulnerabilities of CNS SFT cells, we
   undertook the establishment of primary CNS SFT cell cultures, as there
   are no commercially available cell lines for this rare tumor type.
   Recognizing the transcriptional similarity between CNS SFTs and
   mesenchymal stem cells (MSCs), we employed a modified MSC culture
   medium and were able to successfully propagate CNS SFT cells in 13 out
   of 18 attempts (Supplemental Data [166]1). However, these cells
   exhibited limited proliferative capacity, with most cultures ceasing
   proliferation after about 7 passages and exhibiting marked senescence
   (Supplemental Fig. [167]5A).

   Given this limitation, we focused on examining the drug sensitivity of
   CNS SFT cells at early passage with a library of over 4000 compounds
   comprising FDA-approved drugs and investigational agents in Phase
   II/III clinical trials (Supplemental Data [168]2). Our screening
   revealed a notable sensitivity of primary CNS SFT cells to tyrosine
   kinase inhibitors (Fig. [169]5A, B). To elucidate the underlying kinase
   targets, we selected Ceritinib, an ALK inhibitor, and Pyrotinib, an
   EGFR family inhibitor, both of which demonstrated potent inhibitory
   effects on SFT cells. We also included Tucatinib, another EGFR family
   inhibitor, that lacked significant activity against SFT cells, to
   assess specificity. Remarkably, kinase inhibition profiling of the
   three selected compounds across 76 tyrosine kinases revealed that all
   effective inhibitors exhibited activity against the FER kinase, whereas
   Tucatinib did not (Fig. [170]5C, Supplemental Fig. [171]5B–[172]D).
   Further, the known FER kinase inhibitors E260 and DS21360717 showed
   comparable or even superior efficacy in inhibiting CNS SFT cell
   viability and proliferation than Ceritinib and Pyrotinib
   (Fig. [173]5D–G, Supplemental Fig. [174]5E, [175]F). Compared to
   Tucatinib, E260, DS21360717, pyrotinib, and ceritinib all effectively
   inhibited the invasion and migration of primary SFT cells
   (Fig. [176]5H–K, Supplemental Fig. [177]5G, [178]H).

Fig. 5. Identification of FER kinase as a therapeutic target in SFT.

   [179]Fig. 5
   [180]Open in a new tab

   A Viability of primary SFT cells (CBSFT6940 and CBSFT6596) treated with
   FDA-approved and experimental compounds. B Proportion of compounds with
   >90% inhibition among all screened vs. tyrosine kinase inhibitors
   (two-sided Chi-square test). C Heatmap of the inhibitory effect of
   ceritinib, pyrotinib and tucatinib on 76 tyrosine kinases at a
   concentration of 1 μM. D Dose–response curves of primary SFT cells
   treated with ceritinib, pyrotinib, tucatinib, E260, and DS21360717
   (biological replicates n = 3). E IC[50] of ceritinib, pyrotinib,
   tucatinib, E260, and DS21360717 on primary SFT cells (n = 13, except
   for Tucatinib n = 6). F Representative EdU staining image of primary
   SFT cell CBSFT3965 treated with DMSO, tucatinib, pyrotinib, ceritinib,
   E26,0 and DS21360717. G Quantification results of EdU positive cell
   ratio from primary SFT cells showing reduced proliferation upon
   treatment (n = 4). H−K Representative Crystal Violet staining images
   and quantifications for invasion (H, I) and migration (J, K) assays in
   primary SFT cells treated with DMSO, tucatinib, pyrotinib, ceritinib,
   E260, and DS21360717 (n = 5). L Western blot showing FER knockdown in
   CBSFT1248 and CBSFT6596 (4 cells were validated). M Growth curve of
   FER-silenced CBSFT1248 vs. control (biological replicates n = 3). N
   Normalized growth of shControl, shFER1 and shFER2 in primary SFT cells
   (n = 9). O, P Representative EdU staining and quantification in SFT
   cells expressing control or FER shRNAs (n = 9). Q−T Representative
   Crystal Violet staining images and quantifications for invasion (Q, S)
   and migration (R, T) assays in SFT cells expressing control or FER
   shRNAs (n = 9). U Gene Set Enrichment Analysis showing pathways
   affected by FER silencing. V−X Western blots of phosphorylated ERK (V),
   AKT (W) and FAK (X) upon FER silencing (4 cells were validated). Data
   are presented as mean ± SD. *p < 0.05; **p < 0.01; ***p < 0.001;
   ****p < 0.0001. Tests: two-sided Student’s t-test, Chi-square test,
   one-way or two-way ANOVA. Source data are provided as a Source Data
   file.

   FER kinase, a member of the FPS/FES family of non-receptor tyrosine
   kinases, has been implicated in the regulation of receptor tyrosine
   kinase (RTK) signaling, cell adhesion and migration^[181]36,[182]37.
   Our transcriptomic analysis revealed that FER, but not its family
   members FES or FRK, is highly expressed in CNS SFT cells (Supplemental
   Fig. [183]5I, [184]J). To investigate the functional role of FER in
   these cells, we employed RNA interference to silence FER expression
   (Fig. [185]5L, Supplemental Fig. [186]5K, L). Knockdown of FER
   expression led to a significant reduction in cell viability,
   proliferation, migration, and invasion in primary CNS SFT cultures
   (Fig. [187]5M–T, Supplemental Fig. [188]5M–P).

   Comparative transcriptomic analysis of CNS SFT cells with FER
   expression knocking down (shFER) versus control cells (shCon) by RNA
   sequencing revealed downregulation of signaling pathways associated
   with cell cycle progression, tumor metastasis, and Rho kinase activity,
   as well as upregulation of signaling pathways involved in cell
   differentiation and p53 signaling (Fig. [189]5U). These findings were
   corroborated by Western blot analysis, which demonstrated reduced
   levels of phosphorylated ERK (p-ERK), AKT (p-AKT), and FAK (p-FAK) in
   FER-silenced cells (Fig. [190]5V–X, Supplemental Fig. [191]5Q, R).

Discussion

   This study presents an integrated multi-omics analysis of CNS SFTs,
   revealing molecular heterogeneity and key driving mechanisms that offer
   insights into this rare neoplasm’s biology.

   Our findings confirm NAB2-STAT6 fusion as a core feature in CNS SFTs,
   with RC-CCD1 or RD-SH2 fusion subtypes correlating with low-grade or
   high-grade histopathology, respectively^[192]8. This pattern mirrors
   peripheral SFTs, though the specific mechanisms by which different
   fusion transcripts drive tumorigenesis require further
   investigation^[193]38. Transcriptomic analysis further classified CNS
   SFTs into four molecular subtypes: Classical, characterized by cystic
   necrosis and representing the predominant subtype; Neural-like,
   enriched for neural differentiation pathways and more common in
   low-grade tumors, exhibiting smaller size and low proliferative
   activity; Inflamed, characterized by NF-κB pathway activation and high
   T-cell infiltration; and Migratory, interacting with macrophages
   through cytoskeletal remodeling pathways, potentially driving
   aggressive behavior.

   Single-nucleus RNA sequencing not only validated the bulk RNA-seq
   subtyping results but also revealed five distinct cell states. Among
   these, the Classical state may represent a progenitor-like cell, with
   its gene expression profile displaying MSC-like characteristics,
   potentially explaining the prominent mesenchymal differentiation and
   histological heterogeneity, as well as treatment resistance and
   recurrence often observed in SFT histopathology^[194]9,[195]39,[196]40.
   The Neural-like state was concentrated in low-grade regions, the
   Inflamed state was associated with areas of vascular leakage, and the
   Migratory state was localized to the hypoxic microenvironment. This
   consistency between molecular subtyping and the spatial distribution of
   cell states indicates that CNS SFTs develop heterogeneity by
   dynamically adapting to the microenvironment, offering a
   multi-dimensional perspective on their biological behavior. It’s also
   worth noting that this study has several limitations. First, to ensure
   the quality of omics data, the CNS SFT samples included in this study
   were collected from three hospitals within the past 2 years. Given the
   relatively long recurrence-free survival (RFS) and overall survival
   (OS) of this tumor^[197]8,[198]41,[199]42, most patients in the current
   cohort have not yet reached recurrence endpoints, which limits our
   ability to definitively establish the dynamic associations between
   molecular features and disease progression. Second, the sample size of
   some subtypes, such as the Migratory subtype, was relatively small,
   potentially impacting the statistical power of subgroup analyses.
   Future research should involve multi-center collaborations to build a
   large-sample cohort with long-term follow-up, combined with
   longitudinal multi-omics analysis, to validate the prognostic value of
   the molecular subtyping and explore dynamic monitoring biomarkers.

   While the tumor microenvironment of CNS SFTs is relatively simple
   compared to other CNS malignancies, significant heterogeneity exists
   among subtypes. The enrichment of T-cell infiltration in the Inflamed
   subtype suggests the potential efficacy of immune checkpoint inhibitors
   in this subtype, while the high infiltration of macrophages in the
   Migratory subtype implies their potential role in driving tumor
   progression by secreting pro-invasive factors such as TGF-β, the
   potential link of which to EMT warrants further exploration. Notably,
   the ligand-receptor interactions of RTK pathways related to
   angiogenesis, such as VEGFA-KDR and PDGFB-PDGFRB, not only highlight
   the central role of angiogenesis in SFT progression but also exhibit an
   activation pattern consistent with the high similarity observed
   recently in single-cell sequencing between SFT cell states and brain
   vascular development pathways^[200]43, suggesting that the tumor may
   achieve immune evasion or invasion by mimicking the normal vascular
   microenvironment. Furthermore, the discovery of an IGF-IGF1R autocrine
   loop within SFT cells, although not associated with paraneoplastic
   syndrome in this cohort, requires validation for its potential
   pro-tumorigenic role in larger samples. These dynamic interaction
   networks provide insights into the microenvironment-dependent
   progression mechanisms of CNS SFTs and guide the development of
   targeted therapeutic strategies.

   Our integrated analysis of mutation profiles, molecular subtypes, tumor
   heterogeneity, immune microenvironment, and drug sensitivity in CNS
   SFTs consistently emphasizes the central role of tyrosine kinases and
   their downstream MAPK and PI3K-AKT signaling cascades in the
   pathophysiology of CNS SFTs. This indicates that aberrant activation of
   these pathways is a key mechanism driving aggressive behavior, a
   finding that synergizes with recent reports of mTOR pathway mutations
   being significantly associated with prognosis in peripheral SFT
   patients. Notably, we established FER kinase as a core driver of CNS
   SFTs and demonstrated, using the largest primary SFT cell culture
   cohort to date, that FER kinase synchronously regulates tumor cell
   proliferation, migration, and invasion by integrating the
   RTK-ERK/AKT/FAK signaling axis. This finding provides a clear direction
   for targeted therapy, especially considering the limitations of
   traditional treatments (such as surgery/radiotherapy) in complex
   neuroanatomical settings.

   While previous studies identified EGR1 as an oncogenic driver in SFTs
   (suppressed by NAB2 and dependent on tyrosine kinase signaling)^[201]6,
   directly targeting transcription factors presents significant
   challenge^[202]44. FER kinase on the other hand, as a targetable
   upstream regulatory node, exhibited significant anti-tumor activity
   when inhibited in vitro. This promising avenue for targeted therapy
   development holds potential to improve prognosis by providing more
   effective and safer treatment options for patients with this refractory
   tumor.

Methods

Human tissue samples collection and preparations

   All participants provided written informed consent. The sampling
   procedures and experimental protocols in this study were approved by
   the Institutional Review Boards (IRB) of Beijing Tiantan Hospital,
   Capital Medical University (KY 2024-278-01); Wannan Medical College
   Yijishan Hospital (2024, Ethics review research No.15); and Sanbo
   Hospital of Capital Medical University (SBNK-YJ-2024-019-01). Sex and
   gender were not considered in this study. Participants received no
   compensation.

   A total of 197 histologically confirmed, formalin-fixed and
   paraffin-embedded (FFPE) solitary fibrous tumor (SFT) specimens were
   collected from three neurosurgical centers: Beijing Tiantan Hospital of
   Capital Medical University (122 cases of CNS SFTs), Sanbo Hospital of
   Capital Medical University (34 cases of CNS SFTs), and the First
   Affiliated Hospital of Wannan Medical College (41 cases of SFTs) (see
   Supplementary Data [203]1 for further details). The cohort was
   assembled based on the following inclusion criteria: (i)
   histopathologic confirmation of SFT in all specimens, (ii) availability
   of comprehensive clinicopathologic data for all patients, and (iii)
   absence of prior chemotherapy, radiotherapy, or targeted therapy before
   the initial surgical intervention. Clinical data, including patient
   gender, age at diagnosis, WHO grade, tumor size, anatomic location,
   Ki67 labeling index, presence of necrosis, mitotic rate, and patient
   outcomes, were retrieved from medical records and follow-up visits.
   Full clinicopathologic details are provided in Supplemental
   Data [204]1.

Whole genome sequencing

   Whole genome sequencing (WGS) was performed by GENESEEQ Technology Inc
   (Nanjing, China). DNA was extracted from FFPE tumor samples, and DNA
   concentration and purity were assessed using the Qubit 3.0 Fluorometer.
   Library preparation was conducted using Qiagen reagents. This included
   random fragmentation of genomic DNA, end repair, A-tailing and adapter
   ligation, followed by polymerase chain reaction (PCR) amplification to
   generate the final library. Sequencing was performed on the DNBSEQ-T7
   platform.

Whole genome sequencing (WGS) process and mutation calling

   Somatic and germline single-nucleotide variations (SNVs) and small
   insertions/deletions (indels) were identified using Sentieon software
   (Release 202308) and the TNhaplotyper2 and TNfilter pipelines. Analysis
   was conducted on WGS data using several reference files and databases,
   including the Panel of Normals (PoN) from GATK4^[205]45, dbSNP156
   (hg38, GATK4), COSMIC^[206]46 (version 90) and gnomAD^[207]47 (version
   4.0). Variants were filtered based on a variant allele frequency (VAF)
   threshold of 0.05 and quality score criteria in accordance with GATK4
   best practices. Variants that passed these filters were annotated using
   ANNOVAR^[208]48 (version 3.3.9) with reference to the UCSC refGene
   (hg38) database. For somatic mutations, only exonic variants were
   considered, whereas for germline mutations, only non-synonymous SNVs
   were selected for further analysis.

   Copy number variations (CNVs) were detected using CNVnator^[209]49
   (version 0.4.1) in conjunction with the HMF pipeline. Functional copy
   number variations (fCNVs) were further refined by integration with
   paired RNA-seq data using the methods described previously^[210]50.
   Briefly, samples without copy number variations were used to establish
   baseline levels of gene expression by calculating the mean and standard
   deviation of expression. Z-scores and 95% confidence intervals were
   then calculated for genes in samples with significant amplifications or
   deletions to identify statistically significant changes in expression
   levels. A subset of samples exhibiting mild RTK amplifications (copy
   number 3–4) was subjected to fluorescence in situ hybridization (FISH)
   for further validation (Supplemental Data [211]1).

Fusion domain annotation

   To characterize the functional implications of fusion transcripts, we
   adapted the domain-centric classification framework established by
   Robinson et al.^[212]1 For each NAB2-STAT6 fusion event, we mapped the
   genomic breakpoints to RefSeq gene annotations (GRCh38) and compared
   their positions relative to the canonical domain architecture defined
   in Robinson et al.

Bulk RNA sequencing

   Total RNA extraction and RNA sequencing was performed by Genewiz
   (Suzhou, China). Basically, total RNA was extracted from FFPE tumor
   samples, and RNA concentration and quality were determined using a
   NanoDrop spectrophotometer. For library construction, 300–500 ng of RNA
   per sample was subjected to specific hybridization with single-stranded
   DNA probes targeting ribosomal RNA (rRNA), followed by rRNA digestion
   and removal of DNA probes. The remaining RNA was purified using RNA
   clean beads. The RNA was then fragmented with divalent cations and
   reverse transcribed with random primers. During second-strand
   synthesis, deoxyuridine triphosphate (dUTP) was incorporated. End
   repair and 3’-adenylation were performed, followed by ligation of
   sequencing adapters to both ends of the double-stranded complementary
   DNA (cDNA). The adapter-ligated fragments were purified and
   size-selected using DNA clean beads. PCR amplification was performed
   with P5 and P7 primers to generate the final libraries, which were
   pooled and sequenced on the Illumina NovaSeq 6000 platform (PE150)
   according to the manufacturer’s protocol. An average of 20 million
   reads per sample was obtained.

Quantification of bulk RNA-seq

   Adapter trimming, quality filtering and per-read quality trimming were
   performed using TrimGalore (v0.6.2). Clean paired-end reads were
   subsequently aligned to the human reference genome (hg38, UCSC) with
   HISAT2^[213]51 (v2.2.1) using the RF parameters --sensitive and
   --rna-strandness. Gene expression matrices were generated with
   StringTie^[214]52 (v2.2.1) using the gencode.v45.basic GTF file as
   reference annotation. Batch effects were assessed and corrected using
   the RunHarmony algorithm of the Seurat package^[215]53 (v5.1.0).
   Differential expression analyzes and downstream transcriptomic analyzes
   were conducted on the normalized data.

Identification of NAB2::STAT6 fusion transcripts

   Structural variations (SVs) in whole-genome sequencing (WGS) data were
   identified using the Docker version of Parliament2^[216]54 (version
   0.1.11-0-gb492db6d) and annotated with AnnotSV^[217]55 (v3.4.2). The
   presence of NAB2::STAT6 fusion transcripts in RNA-seq data was detected
   using FusionCatcher^[218]56 (v1.33)
   ([219]https://github.com/ndaniel/fusioncatcher).

Batch effect correction

   To address technical variability across sequencing batches, we
   evaluated multiple batch correction methods, including ComBat
   (parametric adjustment) and Harmony (non-linear integration). Raw count
   matrices were first normalized using DESeq2’s variance-stabilizing
   transformation (VST). Principal Component Analysis (PCA) revealed
   pronounced batch effects in uncorrected data. After benchmarking,
   Harmony demonstrated superior performance in preserving biological
   variation while minimizing batch-associated clustering, and thus
   subsequent analyses utilized Harmony-adjusted expression matrices.

Molecular subtype clustering and enrichment of bulk RNA-seq

   Molecular subtypes of SFT were delineated using non-negative matrix
   factorization (NMF) via the NMF package^[220]57 (v0.27). Genes were
   ranked based on their median absolute deviation (MAD), and clustering
   was performed for the top 2000 genes in increments of 100 genes. NMF
   was executed across ranks (k) from 3 to 6, with 20 independent runs for
   each rank. The optimal number of clusters was determined by selecting
   the rank with the highest cophenetic coefficient, and the consistency
   of clustering was validated by leave-one-out cross-validation.

   Principal component analysis (PCA) was performed using the FactoMineR
   package^[221]58 (v2.8). Gene Set Variation Analysis (GSVA) was
   performed with the Top 2000 Gene Matrix using the GSVA package^[222]59
   (v1.48.3). The statistical significance of differences in normalized
   enrichment scores (NES) between clusters was assessed using the
   Wilcoxon test, with p-values adjusted using the Bonferroni method. A
   significance threshold of 0.05 was applied.

   Differentially expressed genes (DEGs) between clusters were identified
   using the edgeR package^[223]60 (v4.0.16) and the limma package^[224]61
   (v3.58.1), with a threshold of p < 0.05 and |log[2]FoldChange | > 0.58.

Differential expression analysis between SFT and normal brain samples

   Bulk RNA-seq expression matrices from normal brain regions, including
   BA24, BA9, caudate, cortex, hippocampus, hypothalamus, nucleus
   accumbens, and putamen, were obtained from the GTEx project^[225]62.
   After standard normalization, batch effects between the GTEx data and
   our SFT cohort were corrected using the RunHarmony algorithm (Seurat
   v5.1.0). Differential expression analyses comparing SFT and normal
   brain sample groups were then performed on the harmonized data using
   edgeR (v4.0.16) and limma (v3.58.1), with significance defined as
   adjusted p-value < 0.05. To focus on the most robust genes, we retained
   the top 20 % of DEGs ranked by absolute log[2] fold-change and applied
   a minimum effect-size cutoff of |log[2]FC | ≥ 1.6. Functional
   enrichment analyzes were conducted using the ClusterProfiler
   package^[226]63 (v4.8.2).

Differential expression analysis between SFT and meningioma bulk RNA-seq data

   Bulk RNA-seq count matrices for two meningioma cohorts (Meningioma1:
   [227]GSE136661^[228]64, Meningioma2: [229]GSE189672^[230]65) and our
   solitary fibrous tumor (SFT) samples were processed together. After
   normalization, batch effects were corrected with RunHarmony (Seurat
   v5.1.0). We then performed differential expression analyses on the
   harmonized data. Genes differentially expressed between predefined
   sample clusters (SFT vs. meningioma) were identified using edgeR
   (v4.0.16) and limma (v3.58.1), applying an adjusted p-value cutoff of
   <0.05. To focus on the most robust changes, we retained only the top 20
   % of DEGs ranked by absolute log[2] fold-change (|log[2]FC | ≥ 3.3).
   Functional enrichment—including Gene Set Enrichment Analysis (GSEA)—was
   carried out with ClusterProfiler (v4.8.2).

Inference of cellular compositions from bulk RNA-seq data

   Single-cell RNA-seq data were integrated with bulk RNA-seq data to
   investigate cellular composition and gene expression profiles.
   Overlapping genes between the datasets were identified and selected for
   further analysis. The expression matrix was filtered by removing
   non-coding genes to focus exclusively on protein-coding genes annotated
   in the hg38 genome assembly. Bayesian inference of cell type
   proportions was conducted using BayesPrism package^[231]31 (v2.2.2) to
   infer the contribution of single cell populations to the overall
   changes in gene expression observed in the bulk RNA-seq data.

Single-nucleus RNA sequencing

   Single nucleus RNA sequencing (snRNA-seq) was performed by Singleron
   Biotechnologies Ltd (Nanjing, China). Fresh tissue was stored in
   sCelLive™ Tissue Preservation Solution (Singleron) on ice immediately
   after surgical resection within 30 min. Samples were washed three times
   with Hanks’ Balanced Salt Solution (HBSS), minced and digested with
   3 mL of sCelLive™ Tissue Dissociation Solution (Singleron) using the
   Singleron PythoN™ Tissue Dissociation System at 37 °C for 15 min. The
   resulting cell suspension was filtered through a 40 µm sterile
   strainer. Red blood cells were lysed with GEXScope® Red Cell Lysis
   Buffer (RCLB, Singleron) at a ratio of 1:2 by volume. The mixture was
   incubated for 5−8 min at room temperature, centrifuged at 300 × g at
   4 °C for 5 min, washed with PBS and resuspended in PBS to a final
   concentration of 2 × 10^5 cells/ml. The single-cell suspension was
   loaded onto microwell chips using the Singleron Matrix® Single Cell
   Processing System, followed by capture of mRNA onto barcoding beads,
   reverse transcription, and cDNA synthesis. PCR amplification was then
   performed and the amplified cDNA was fragmented and ligated with
   sequencing adapters. The snRNA-seq libraries were prepared using the
   GEXSCOPE® Single Cell RNA Library Kits (Singleron) and sequenced using
   the Illumina NovaSeq 6000 platform.

Dimensionality reduction, clustering, and cell type annotation

   For data integration, we utilized scVI from the scvi-tools package,
   following established protocols
   ([232]https://docs.scvi-tools.org/en/stable/tutorials/index.html).
   Clustering of the integrated data was performed using the scanpy
   package. The resulting H5ad files were converted into Seurat objects
   using Sceasy ([233]https://github.com/cellgeni/sceasy). Data were
   analyzed in RStudio (v4.2.25) using Seurat (v4.1), and cluster-specific
   markers were identified using the FindAllMarkers function, in
   accordance with previously established workflows^[234]66. Signal
   enrichment for different cell states was analyzed across gene sets from
   the Molecular Signatures Database (MSigDB) using the
   Over-Representation Analysis (ORA) module available in the Python
   implementation of decoupleR.

RNA velocity analysis

   RNA velocity analysis was conducted using scVelo with default
   parameters, following the software instructions
   ([235]https://smorabit.github.io/tutorials/8_velocyto/). The input
   files for scVelo were generated using the Python version of velocyto.
   Pseudotime trajectory analysis was performed using Monocle 3 with
   default parameters
   ([236]https://cole-trapnell-lab.github.io/monocle3/docs/getting_started
   /). The regulatory network analysis was carried out using the Docker
   version of pySCENIC^[237]67 (v0.12.0)
   ([238]https://hub.docker.com/r/aertslab/pyscenic).

   Differential expression between cell types was determined using
   FindAllMarker based on area under the curve (AUC) scores. For
   visualization, we used ComplexHeatmap and ClusterProfiler, consistent
   with previous methods. Conversion between H5ad and Seurat objects was
   performed using Sceasy, as previously described.

Immune infiltration profiling

   As in a previous study^[239]33, tumor microenvironment composition was
   inferred using two complementary deconvolution tools: EPIC^[240]68 and
   MCP-counter^[241]69 using default parameters.

Spatial transcriptome sequencing (Visium HD)

   Spatial transcriptome analysis was conducted using the Visium HD
   Spatial Gene Expression platform from 10x Genomics, following the
   manufacturer’s protocol. This work was carried out by Yimei Tongde
   Technology Ltd (Beijing, China). The FFPE tissue sections were placed
   within the 6.5 × 6.5 mm capture regions on the Visium slide. After
   drying overnight, the sections were incubated at 60 °C for 2 h, stained
   with hematoxylin and eosin (H&E), and covered with 85% glycerol. The
   slides were imaged with an Olympus VS120 Slide Scanner at 40×
   magnification. The glycerol was then removed and the tissue sections
   were decrosslinked with TE buffer (pH 9.0). Human transcriptome probes
   targeting ~18,000 genes were ligated to mRNA for capture and library
   construction following the Visium Spatial Gene Expression for FFPE
   protocol. Sequencing was performed on an Illumina NovaSeq SP flow cell
   with paired-end 50 bp reads. Fastq files were generated using the
   bcl2fastq pipeline (v2.20.0.4).

Analysis of visium HD spatial gene expression

   Raw FASTQ files were processed with Cell Ranger using default
   parameters and the hg38 reference genome. Spatial transcriptomic
   analyses were conducted using Seurat package (v5.1.0), including
   normalization and feature identification with the NormalizeData,
   FindVariableFeatures, and ScaleData functions. To reduce dataset size,
   20% of the 8 × 8 μm-binned spots were sampled using leverage score
   sampling. Dimensionality reductions and clustering results from the
   sampled data were projected onto the full dataset using PCA and UMAP.
   To analyze the localization and expression of cell type mixtures, we
   performed robust cell type decomposition (RCTD) analysis with the
   spacexr package^[242]70 (v2.2.1) using default parameters, with
   scRNA-seq as guidance. To determine the hypoxia condition of
   differential cell types, the hypoxia scores were calculated using the
   HALLMARK_HYPOXIA gene set from MSigDB via single-sample gene set
   enrichment analysis (ssGSEA) implemented in GSVA (v1.48.3) package. To
   measure intercellular communication taking into account spatial
   coordinates of cell types, we performed cell-cell communication
   analysis utilizing statistical analysis method of CellPhoneDB^[243]71
   (v3.0) under database cellphonedb-data v5.0.

Histology and immunohistochemistry

   Formalin-fixed SFT tissue samples were embedded in paraffin, cut into 4
   μm thick sections and stained with the H&E staining kit (Solarbio,
   G1120). All H&E-stained slides were reclassified according to the fifth
   edition of the WHO classification for CNS SFTs. Immunohistochemistry
   was performed with the following primary antibodies: STAT6
   (Proteintech, 51073-1-AP), Ki-67 (Origen, TA801577), p-ERK (Cell
   Signaling Technology, [244]P28482), p-S6 (Cell Signaling Technology,
   88441), EGFR (Cell Signaling Technology, 4267), Decorin (Proteintech,
   14511-1-AP), Synaptotagmin-1 (Proteintech, 14511-1-AP), CD4 (Cell
   Signaling Technology, 25229), CD8a (Cell Signaling Technology, 85336),
   CD19 (Cell Signaling Technology, 3574) and CD66b (Genetex, GTX19779). A
   two-step universal reagent kit (ZSGB-BIO, PV-8000-1) and a DAB kit
   (ZSGB-BIO, ZLI9018) were used for detection. Images were acquired using
   a pathology microscope, and histopathologic analysis was performed with
   QuPath digital pathology software. The H-score was calculated as
   follows:

   H-score = 0×(%negative cells)+1×(%weakly positive
   cells) + 2 × (%moderately positive cells) + 3 × (%strongly positive
   cells), with a total score range of 0–300.

Primary SFT cell cultivation

   Fresh SFT tumor tissue was immediately placed in Hibernate A medium
   (Thermo Scientific, A1247501) under sterile conditions and processed
   within 6 h after resection. Blood components were removed by washing
   with cold, sterile PBS, and the electrocauterized tissue was trimmed.
   Tumor tissue was minced into uniform pieces (~1 mm^3) using microscopic
   scissors. The tissue was digested for 16 min with Accutase (Biolegend,
   423201) in RWD tissue dissociator (RWD Life Technology Ltd, DSC-410,
   Shenzhen, China). The digestion process was terminated with PBS, and
   the suspension was centrifuged at 500 × g for 5 min. The supernatant
   was removed, and the red blood cells were lysed with a red blood cell
   lysis buffer for 5 min. After another centrifugation step, the cells
   were resuspended in DMEM/F12 medium (Gibco, C11330500BT), supplemented
   with 10% fetal bovine serum (Gibco, 16000044), 20 ng/mL EGF
   (Novoprotein, C029-500 μg), 20 ng/mL bFGF (ORIGENE, TP750002), and 1%
   penicillin/streptomycin (Solarbio, P8420/S8290). The cells were seeded
   in culture dishes coated with Matrigel (CORNING, 354248) and incubated
   at 37 °C in a 5% CO[2]. The medium was changed after 24 h and
   spindle-shaped adherent cells were observed within a week. The medium
   was changed every 2 days and the cells were passaged twice a week with
   trypsin/EDTA.

Cell viability assay

   For the small molecule drug library screening, two compound libraries
   were utilized: the Selleck 1466 FDA-approved small molecule library and
   the TargetMol 2566 small molecule library, which contains both
   FDA-approved drugs and investigational agents in Phase II/III clinical
   trials, all at a stock concentration of 10 mM in DMSO or water. These
   compounds were added to 384-well plates to identify inhibitors of
   primary CNS SFT cell viability at a final concentration of 10 μM. Drug
   response data were analyzed according to methods described in previous
   studies^[245]72. For the cell viability assay, 2000 cells per well were
   seeded in 96-well plates and incubated at 37 °C for 3 h. After
   incubation, the cells were treated with various small-molecule
   compounds for 72 h. Cell proliferation was assessed using the
   CellTiter-Glo Luminescent Cell Viability Assay (Promega, G9243). The
   small molecule compounds used included E260 (HY-112097), DS21360717
   (HY-128576), ceritinib dihydrochloride (HY-15656A) and tucatinib
   (HY-16069) from MedChemExpress, and pyrotinib dimaleate ([246]T12594)
   from Shanghai TargetMol Bio-Technology. Relative viability for each
   treatment group was calculated as percent change relative to the
   vehicle control, and dose-response curves were fitted using the log
   (inhibitor) vs. normalized response model in GraphPad Prism (v7.0, La
   Jolla, USA).

Kinase inhibitor profiling

   The inhibitory efficacy of ceritinib, pyrotinib dimaleate, and
   tucatinib on kinase activity was evaluated by ICE Bioscience Inc. All
   compounds were tested at a concentration of 1 μM. Inhibitory activity
   against various kinases was assessed using both the HTRF kinase assay
   and the ADP-Glo kinase assay. The reaction control (1% DMSO) served as
   a reference for 0% inhibition, while the background control (10 µM
   positive control) was set as 100% inhibition. The percent inhibition
   for each compound was calculated using the following formula: %
   Inhibition = 100%-(compound-positive control)/(negative
   control-positive control)*100%.

Lentiviral-based shRNA knockdown and transfection

   Lentivirus-mediated transfection was used to knock down FER (Ensembl:
   ENSG00000151422) expression. Lentiviral particles were produced by
   co-transfection of HEK293T cells with a mixture of lentiviral
   expression plasmids, psPAX2 and pMD2.G, at a ratio of 5:3:2. The shRNA
   sequence targeting FER was cloned into the pLKO.1-Puro vector
   (Addgene). The shRNA targeting sequences were:

   shFER1: CAGAACAACTTAGTAGGATAA

   shFER2: CCACCTCCAGTAGTAAATTAT

   Primary SFT cell lines and HCT116 cells were transduced with 5 μL
   lentivirus per well in 6-well plates using 5 μg/mL polybrene (Beyotime
   Biotechnology, C0351) and 10 μg/mL protamine sulfate to increase
   transfection efficiency. Fresh culture medium was added 24 h
   post-transfection. After 48 h, the medium was replaced with complete
   medium containing 0.625 μg/mL puromycin for sthe election of
   successfully transfected cells. Selection was maintained for 2 days.
   The cells were then divided into the negative control group (Sh-con),
   transduced with control shRNA, and the FER knockdown group (Sh-FER).
   Knockdown efficiency was determined by quantitative PCR (qPCR) and
   Western blot analysis.

Quantitative polymerase chain reaction (qPCR) analysis

   Total RNA was extracted from cultured SFT and HCT116 cells using the
   FastPure Cell/Tissue Total RNA Isolation Kit V2 (Vazyme, RC112-01)
   according to the manufacturer’s instructions. The concentration and
   purity of the RNA were assessed using a NanoDrop spectrophotometer.
   Subsequently, 1 μg of total RNA was reverse transcribed into
   complementary DNA (cDNA) using the HiScript III All-in-one RT SuperMix
   (Vazyme, R333). Quantitative PCR (qPCR) was performed using the Taq Pro
   Universal SYBR qPCR Master Mix (Vazyme, Q712) on a CFX96 Real-Time PCR
   Detection System (Bio-Rad). Relative transcript levels were calculated
   using the comparative Cq method (2^-ΔΔCq), with GAPDH serving as the
   reference gene. The primer sequences used were as follows:

   HuGAPDH-F: 5’-AATCCCATCACCATCTTCCA-3’

   HuGAPDH-R: 5’-TGGACTCCACGACGTACTCA-3’

   HuFER-F: 5’-CAGCACCGGAAGCTCTTAAT-3’

   HuFER-R: 5’-CCCCTAAGCTGAAGGTCTCC-3’.

Western blotting

   Cell lysates were prepared with RIPA lysis buffer (Beyotime, P0013B)
   supplemented with protease inhibitors (Bimake, [247]B14001) and
   phosphatase inhibitors (Merck Millipore, 524625) and quantified using a
   BCA Protein Assay Kit (Thermo Scientific, 23227). Protein samples were
   separated via 7.5% sodium dodecyl sulfate-polyacrylamide gel
   electrophoresis (SDS-PAGE) and transferred to a polyvinylidene
   difluoride (PVDF) membrane (Merck Millipore, ISEQ00010). The membranes
   were blocked for 2 h at room temperature and then incubated overnight
   at 4 °C with the following primary antibodies: FER (Proteintech,
   25287-1-AP), ERK (Cell Signaling Technology, #[248]P27361), p-ERK (Cell
   Signaling Technology, #[249]P28482), AKT (Cell Signaling Technology,
   #[250]P31751), p-AKT (Cell Signaling Technology, #[251]P31749), FAK
   (Proteintech, 12636-1-AP), p-FAK (Cell Signaling Technology, #3283),
   and α-tubulin (Proteintech, 11224-1-AP). Membranes were washed four
   times with TBST (10 min each time) and incubated with either
   peroxidase-conjugated rabbit anti-mouse IgG (Proteintech) or mouse
   anti-rabbit IgG H&L (Proteintech). Protein bands were visualized using
   chemiluminescence reagents and grayscale analysis was performed using
   ImageJ software. α-Tubulin was used as a loading control.

EdU incorporation assay

   Cell proliferation was further evaluated using the BeyoClick EdU-647
   Cell Proliferation Detection Kit (Beyotime Biotechnology, C0081S). A
   total of 2,000 cells per well were seeded in 96-well plates and
   cultured for 72 h. Subsequently, the medium was replaced by a complete
   medium containing EdU and the cells were incubated for 2 h. After
   incubation, the cells were fixed with methanol for 15 min, washed with
   BSA and permeabilized with Triton X-100 for 15 min. The Click-iT
   reaction cocktail was added and the cells were incubated for 30 min in
   the dark. Hoechst 33342 was used to stain the cell nuclei (30 min
   incubation in the dark). Images was captured using fluorescence
   microscopy and the EdU-positive rate was calculated by dividing the
   number of EdU-positive cells by the total number of nuclei based on
   three random fields of view.

Transwell migration and invasion assays

   Transwell chambers (Corning, 3422) were used to assess the migration
   and invasion capabilities of primary SFT cells. For the migration
   assay, 30,000 cells were seeded in the upper chamber and 600 μL of
   medium supplemented with 10% FBS was added to the lower chamber. For
   the invasion assay, the upper chamber was pre-coated with Matrigel
   (Matrigel: serum-free medium = 1:4, Corning, 356234). After 48 h of
   incubation, the cells were fixed with methanol, stained with crystal
   violet, and washed. Microscopic images were captured, and the number of
   cells that had migrated or invaded through the membrane was quantified.
   Three random low-magnification image fields were selected, and the
   total number of cells at the bottom of the upper chamber was counted.

Statistics and reproducibility

   Statistical analyzes were performed with R (v4.3.1) and GraphPad
   (v7.0). For box plots, the median is represented by the center line and
   the upper and lower quartiles by the box boundaries. The whiskers
   indicate the 1.5× the interquartile range, and the individual points
   represent raw values. Unless otherwise stated, Student’s t-tests were
   performed.

   Detailed statistical information was listed in the respective figure
   legends or source data. Asterisks are used to indicate the statistical
   significance (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001);
   n.s. means statistically non-significant (p > 0.05).

   No statistical methods were employed to predetermine sample sizes. No
   data points were excluded from analyses. Group allocations were
   randomized. Investigators were not blinded during outcome assessments.

Reporting summary

   Further information on research design is available in the [252]Nature
   Portfolio Reporting Summary linked to this article.

Supplementary information

   [253]Supplementary Information File^ (1.9MB, pdf)
   [254]41467_2025_63039_MOESM2_ESM.pdf^ (77.3KB, pdf)

   Description of Additional Supplementary Files
   [255]Reporting Summary^ (5.2MB, pdf)
   [256]Transparent Peer review file^ (5.2MB, pdf)
   [257]Supplementary data^ (402.2KB, xlsx)

Source data

   [258]Source data^ (6.1MB, xlsx)

Acknowledgements