Abstract

   Diffuse‐type tenosynovial giant cell tumor (D‐TGCT) and localized‐type
   tenosynovial giant cell tumor (L‐TGCT) share common genomic aberrations
   and histopathological features, but the former has a more aggressive
   nature and a higher recurrence rate, leading to worse prognoses for
   patients. In this study, single‐cell RNA sequencing (scRNA‐seq) on
   human D‐TGCT and L‐TGCT lesions is conducted to discover
   transcriptional differences. A unique cluster of tumor cells in D‐TGCT
   is identified that regulated differentiation of CD34 ^+ fibroblasts
   into MMP3^+ fibroblasts or APOE ^+ fibroblasts via COL6A3 − (ITGAV +
   ITGB8) interaction. The APOE ^+ fibroblasts further activated IL‐1B ^+
   CCL20 ^+ macrophages through the CXCL12/CXCR4 axis. IL‐1B ^+ CCL20 ^+
   macrophages and MMP3 ^+ fibroblasts participated in local aggression of
   D‐TGCT. Two effective biomarkers, ROR1 and PRKD1 are also identified
   and validated, to predict disease recurrence. This study not only
   clarified the underlying mechanisms of aggressive behavior in D‐TGCT
   but also provided a theoretical basis and potential targets for
   intervention into and treatment of this disease.

   Keywords: local aggression, recurrence, single‐cell RNA sequencing,
   tenosynovial giant cell tumors
     __________________________________________________________________

   MP3 tumor cells, a specific subpopulation of tumor cells in D‐TGCT,
   regulated the differentiation of CD34 ^+ Fbs into MMP3 ^+ Fbs and APOE
   ^+ Fbs through COL6A3 − (ITGAV + ITGB8) interaction. APOE ^+ Fbs
   activated IL‐1B ^+ CCL20 ^+ Mφs through the CXCL12/CXCR4 axis. IL‐1B ^+
   CCL20 ^+ Mφs and MMP3 ^+ Fbs participated in the local invasion of
   D‐TGCT.

   graphic file with name ADVS-12-2415835-g003.jpg

1. Introduction

   Tenosynovial giant cell tumor (TGCT), also known as pigmented
   villonodular synovitis, is a rare proliferative mesenchymal neoplasm
   that affects bones, synovium, tendon sheaths, and bursae of the
   joints.^[ [54]^1 ^] Two distinct subtypes, L‐TGCT and D‐TGCT, are
   defined by different radiological, biological, and clinical features.^[
   [55]^2 , [56]^3 ^] L‐TGCT typically manifests as a well‐circumscribed,
   non‐destructive nodule, whereas D‐TGCT presents as uncontrolled
   synovial proliferation with locally aggressive and invasive behavior,
   causing severe joint pain and recurrent articular hemorrhage.
   Extra‐articular infiltration and malignant transformation have been
   incidentally reported,^[ [57]^4 , [58]^5 ^] forcing amputation of limbs
   and leaving patients disabled.^[ [59]^6 , [60]^7 ^]

   Currently, no recognized clinical guideline for the treatment of TGCT
   exists, but the mainstay treatment option is still surgical resection,
   whether open or arthroscopic, to relieve pain and lower the risk of
   joint destruction.^[ [61]^8 ^] For L‐TGCT, which often occurs in the
   hands and feet, surgical excision with clear margins is an effective
   treatment; the average recurrence rate after these procedures is <6%.^[
   [62]^9 ^] However, D‐TGCT has a much higher recurrence rate, almost
   44%, after surgical resection, which seriously affects patients’
   quality of life (QoL)^[ [63]^10 , [64]^11 ^] and necessitates
   additional interventions such as radiotherapy^[ [65]^12 , [66]^13 ^]
   and drug therapy.^[ [67]^14 ^] The long‐term risks of vascular
   necrosis, joint fibrosis, and stiffness caused by radiation therapy, as
   well as other sequelae, merit further research still.^[ [68]^15 ^] In
   addition, drug research mainly focuses on colony‐stimulating factor 1
   receptor (CSF1R) inhibitors, such as pexidartinib, the first drug to
   receive US Food and Drug Administration (FDA) approval for treating
   D‐TGCT.^[ [69]^14 ^] However, the poor effectiveness and side effects
   of these medications, such as liver and kidney toxicity, remain
   unresolved challenges.^[ [70]^16 ^] Therefore, comprehensive treatment
   of D‐TGCT is difficult to achieve.

   To date, the pathogenesis and molecular features of TGCT are still
   unclear and debated, with conflicting theories proposing it as either
   an inflammatory disease or a neoplastic disease. This lack of clarity
   restricts the development of systemic treatment strategies for TGCT,
   especially molecular targeted therapy and drug research. Previous
   studies involving pathological and flow cytometric (FCM) analyses have
   demonstrated that both subtypes of TGCT are characterized by disruption
   and rearrangement of the 3′‐end of CSF^[ [71]^17 , [72]^18 , [73]^19 ^]
   and that both are mainly composed of mononuclear macrophages,
   fibroblasts, and lymphocytes.^[ [74]^20 , [75]^21 ^] However, the
   clinical characteristics, postoperative‐recurrence rates, and prognoses
   of the two subtypes are quite different. Why D‐TGCT is more aggressive
   and has a higher recurrence rate is not fully understood. Given these
   limitations, revealing the mechanisms that drive such aggression and
   recurrence has become crucial for developing novel effective treatments
   and early‐intervention strategies for D‐TGCT.

   Interactions among tumor cells, immune cells, and stromal cells in the
   tumor microenvironment (TME) could influence tumor progression and
   immune status.^[ [76]^22 , [77]^23 , [78]^24 , [79]^25 ^] However, the
   differences in TME between the two TGCT subtypes, the classification
   and functional annotation of each cell subpopulation, and potential
   biomarkers for predicting recurrence have not been explored. The advent
   of scRNA‐seq provides an optimal means of shedding light on the
   pathological profile and molecular features of TGCT. In this study, we
   examined 10 D‐TGCT samples and 7 L‐TGCT samples using scRNA‐seq to
   investigate intratumoral heterogeneity and analyze the TME as well as
   intercellular interactions to explain the discrepancy in manifestations
   between D‐TGCT and L‐TGCT. Since osteoarthritis (OA) is a degenerative
   disease of articular cartilage, characterized by relatively mild
   synovial inflammation compared to other arthritis such as rheumatoid
   arthritis (RA), we also sequenced 3 OA samples as control in our study.
   Using in silico analysis, we identified and validated two effective
   biomarkers to predict disease recurrence. Our findings yielded an
   in‐depth single‐cell transcriptomic atlas of TGCT and can offer a
   theoretical basis of and potential targets for intervention into and
   treatment of this disease.

2. Results

2.1. Single‐Cell Profiling of D‐TGCT and L‐TGCT

   To fully decipher global differences in TMEs between D‐TGCT and L‐TGCT,
   we performed scRNA‐seq on 17 samples (nine primary D‐TGCT,
   one recurrent D‐TGCT, six primary L‐TGCT, and one recurrent L‐TGCT) as
   a discovery cohort using the 10x Genomics Chromium platform (10x
   Genomics, Inc., Pleasanton, CA, USA; Figure [80] 1A). Following
   standard data processing and quality control (QC) procedures, we
   obtained 123970 TGCT cells, which were then clustered into
   32 subpopulations via uniform manifold approximation and projection
   (UMAP; Figure [81]S1A, Supporting Information). The uniform
   distribution of these 32 cell populations across each sample suggested
   successful correction for the batch effect (Figure [82]S1B, Supporting
   Information). Then, we annotated these populations into 10 major cell
   types based on expression levels of canonical marker genes
   (Figure [83]1B; Figure [84]S1C, Supporting Information), including
   macrophages (Mφs), proliferating macrophages (pro‐Mφs), osteoclasts
   (OCs), type 2 conventional dendritic cells (cDC2s), mast cells (MCs),
   B lymphocytes (B‐lyms), T lymphocytes (T‐lyms), endothelial cells
   (ECs), fibroblasts (Fbs), and smooth‐muscle cells (SMCs). All annotated
   cell types were distributed consistently between D‐TGCT and L‐TGCT
   (Figure [85]1C), as well as between recurrent and primary TGCT
   (Figure [86]1D). Similarly, we clustered and annotated five major cell
   types in OA synovium (Figure [87]S1D,E, Supporting Information): Mφs,
   T‐lyms, MCs, Fbs, and ECs. Compared with OA synovium, immune cells were
   the predominant cellular components in TGCT, suggesting a significant
   pathological role of these cells in the disease.

Figure 1.

   Figure 1
   [88]Open in a new tab

   Single‐cell transcriptome atlas of diffused‐type tenosynovial giant
   cell tumor (D‐TGCT) and localized‐type tenosynovial giant cell tumor
   (L‐TGCT). A) Workflow of the current study. B) UMAP projection of
   123970 cells from 10 D‐TGCT and 7 L‐TGCT samples, which were annotated
   into 10 main cell types. Each dot represents a single cell and is
   colored according to its cell population. C) UMAP projection of all
   TGCT cells grouped by different subtypes. D) UMAP projection of all
   TGCT cells grouped by disease status. E) Histogram showing the ratio of
   each cell type across the 10 D‐TGCT and 7 L‐TGCT samples. F) Histogram
   showing the ratio of each cell type between D‐TGCT and L‐TGCT samples.

   Notably, scRNA‐seq analysis identified macrophages and fibroblasts as
   the predominant cellular components (Figure [89]1E,F), exhibiting
   differential abundance between D‐TGCT and L‐TGCT (Figure [90]S1F,
   Supporting Information). Mφs have previously been reported as the
   responsive cell type in TGCT.^[ [91]^17 , [92]^26 ^] Therefore, we
   first focused on these two cell types and their heterogeneity between
   D‐TGCT and L‐TGCT.

2.2. IL‐1B ^+ CCL20 ^+ Macrophages in D‐TGCT Showed Increased Invasive
Capability, Potentially Modulated by Fibroblasts

   Macrophagic lineages from primary TGCT, including Mφs, pro‐Mφs, and
   OCs, were clustered into seven subpopulations (Figure [93] 2A) and
   subsequently annotated as IL‐1B ^+ CCL20 ^+ Mφs, EGR1 ^+ Mφs, MARCO ^+
   Mφs, FMNL2 ^+ Mφs, CCL18 ^+ Mφs, pro‐Mφs, and OCs based on their
   differentially expressed genes (DEGs; Figure [94]S2A, Supporting
   Information). We then conducted Gene Ontology (GO)
   functional‐enrichment analysis of these DEGs to infer their cellular
   functions. GO analysis revealed that upregulated DEGs in IL‐1B ^+ CCL20
   ^+ Mφs were enriched in pathways pertinent to inflammation activation,
   M1 polarization, and invasion (Figure [95]2B). In particular,
   “regulation of apoptotic signaling pathway,” “response to tumor
   necrosis factor,” “chemokine‐mediated signaling pathway,” and “negative
   regulation of cell junction assembly” were enriched in IL‐1B ^+ CCL20
   ^+ Mφs. Similarly, we designated the remaining Mφ subpopulations based
   on their respective enriched GO terms: EGR1 ^+ Mφs were termed another
   immune‐activated subset due to enrichment of “cellular response to
   tumor necrosis factor,” “cellular response to heat,” and “cellular
   response to chemical stress”; MARCO ^+ Mφs were termed phagocytic type
   due to enrichment of “humoral immune response,” “apoptotic‐cell
   clearance,” and terms related to phagocytosis; FMNL2 ^+ Mφs were
   subsets related to angiogenesis; and CCL18 ^+ Mφs were subsets related
   to inherent immunity (Figure [96]2B). We further validated the
   functional enrichment analysis results by calculating phenotypic gene
   set scores for each Mφ subset. Consistently, the expression levels of
   M1, pro‐inflammatory, and Mφ invasion signatures of IL‐1B ^+ CCL20 ^+
   Mφs were higher than those of the other Mφs (Figure [97]S2B, Supporting
   Information). Based on these findings, we assumed that IL‐1B ^+ CCL20
   ^+ Mφs might contribute to the aggressive behavior of D‐TGCT.

Figure 2.

   Figure 2
   [98]Open in a new tab

   Single‐cell transcriptome atlas of macrophages. A) UMAP projection of
   seven subpopulations generated from unsupervised clustering of
   macrophagic lineages. B) GO enrichment analysis of DEGs of each Mφ
   subpopulation, suggesting that different biological functions were
   present in different Mφ subpopulations. C) Pie charts showing the
   proportion of each Mφ subpopulation in primary TGCT. D) Histogram
   showing the number of DEGs of each Mφ subpopulation between D‐TGCT and
   L‐TGCT. E) Bar plot showing the enriched GO pathways in IL‐1B ^+ CCL20
   ^+ Mφs from D‐TGCT compared with those from L‐TGCT. F) Circle plot
   showing the cell–cell communication between IL‐1B ^+ CCL20 ^+ Mφs and
   other major immune cell populations. Line thickness indicates relative
   interaction intensity. Red and blue represent higher interaction
   intensities observed in D‐TGCT and L‐TGCT, respectively.

   Notably, IL‐1B ^+ CCL20 ^+ Mφs constituted the majority subpopulation
   of Mφs, but they showed minimal variation between D‐TGCT and L‐TGCT
   with no statistically significant differences (Figure [99]2C; Figure
   [100]S2C, Supporting Information). We also performed diffusion map
   dimensionality reduction to interrogate differentiation trajectories of
   macrophage subpopulations across two types of TGCT. As mentioned above,
   macrophages in TGCT originated from circulating monocytes^[ [101]^17
   ^]. We accordingly calculated the monocyte signature score for each
   macrophage subpopulation to identify a monocyte‐like subset and found
   out MARCO ^+ Mφs exhibited the highest score (Figure [102]S2D,
   Supporting Information). By utilizing it as the starting point, we
   constructed the differentiation trajectories of macrophage
   subpopulations (Figure [103]S2E,F, Supporting Information). RNA
   velocity showed a similar developmental trajectory, starting with MARCO
   ^+ Mφs and ending with FMNL2^+ Mφs (Figure [104]S2G, Supporting
   Information). Consistent with the results of cellular proportion
   analysis, the developmental trajectories of all macrophage subsets
   exhibited no differences between D‐TGCT and L‐TGCT (Figure [105]S2H,
   Supporting Information), prompting us to further investigate whether
   they exhibited discrepancies in functionality.

   To identify the functional variances of Mφ subpopulations between
   D‐TGCT and L‐TGCT, we calculated the number of DEGs within each Mφ
   lineage across both types of TGCT. Of these lineages, IL‐1B ^+ CCL20 ^+
   Mφs had the highest DEG count, suggesting the most pronounced
   functional diversity (Figure [106]2D). We next conducted GO functional
   enrichment analysis of DEGs of IL‐1B ^+ CCL20 ^+ Mφs from D‐TGCT
   compared with those from L‐TGCT. We found that the upregulated genes in
   D‐TGCT were enriched in “cellular response to tumor necrosis factor,”
   “macrophage activation,” “chemokine‐mediated signaling pathway,” “cell
   junction disassembly,” and “cellular component disassembly”
   (Figure [107]2E), indicating that IL‐1B ^+ CCL20 ^+ Mφs from D‐TGCT had
   a more apparently invasive phenotype than those from L‐TGCT. To further
   investigate potential upstream cell populations contributing to the
   altered functions of IL‐1B ^+ CCL20 ^+ Mφs across both types of TGCT,
   we conducted cell–cell interaction (CCI) analysis using the R software
   package CellChat. Extensive cellular communication was observed in
   TGCT, and it was stronger between Fbs and IL‐1B ^+ CCL20 ^+ Mφs in
   D‐TGCT than in L‐TGCT (Figure [108]2F). This suggested that the
   functional variance of IL‐1B ^+ CCL20 ^+ Mφs in D‐TGCT might be
   influenced by Fbs.

2.3. Higher Proportions of MMP3 ^+ Fb and APOE ^+ Fb are Associated with the
Aggressive Behavior of D‐TGCT

   To gain further insight into the functional heterogeneity of Fbs, we
   re‐clustered them into six subpopulations (Figure [109]S3A, Supporting
   Information) and annotated them by DEG expression level (Figure
   [110]S3B, Supporting Information). Specifically, within these six
   subpopulations, we identified a subgroup expressing the molecular
   characteristics of TGCT tumor cells. Previous studies have shown that
   tumor cells in TGCT derive from Fbs.^[ [111]^27 , [112]^28 ^] A recent
   study has indicated the potential of GFPT2 as a marker for tumor cells
   in TGCT and its association with activation of the Hippo signaling
   pathway.^[ [113]^29 ^] Consistent with those findings, we discovered
   that the above‐described Fb subpopulation indeed expressed higher
   levels of GFPT2 (Figure [114]S3D, Supporting Information) and scored
   higher on activation of the Hippo pathway (Figure [115]S3E, Supporting
   Information) than other Fb subpopulations. In the above‐mentioned cell
   type annotations, we compared differences in proportions of the six
   identified Fb subpopulations between the two types of TGCT and
   discovered that the proportions of MMP3 ^+ Fbs and APOE ^+ Fbs were
   significantly higher in D‐TGCT than in L‐TGCT (Figure [116] 3A,B),
   which we further validated via multiplex immunohistochemical (mIHC)
   staining (Figure [117]3C,D); these findings suggested that these Fbs
   played key roles in the development of D‐TGCT. MMP3 is a secreted
   protein reported to be involved in cell invasiveness and cancer
   progression.^[ [118]^30 ^] GO functional enrichment analysis of MMP3 ^+
   Fbs revealed functions of matrix remodeling and matrix degradation
   (Figure [119]5E), suggesting these Fbs’ potential role in local joint
   destruction. A previous study identified an Fb subpopulation in the
   synovial lining that was closely related to active rheumatoid arthritis
   (RA).^[ [120]^31 ^] Notably, MMP3 ^+ Fbs in our study also highly
   expressed marker genes associated with an Fb subpopulation closely
   related to active RA (Figure [121]S3F, Supporting Information). Using
   these marker genes as a gene set of Fb activation, we found that MMP3
   ^+ Fbs scored significantly higher than the other subpopulations
   (Figure [122]S3G, Supporting Information), indicating their activated
   state and correlation with aggression of D‐TGCT.

Figure 3.

   Figure 3
   [123]Open in a new tab

   Higher proportions of MMP3 ^+ fibroblasts and APOE ^+ fibroblasts were
   associated with locally aggressive behavior of D‐TGCT. A) Pie charts
   showing the percentages of different Fb subpopulations between the two
   types of TGCT. B) Box plots showing the differences in the proportions
   of MMP3 ^+ Fbs and APOE ^+ Fbs between the two TGCT subtypes.
   Statistical significance was inferred by two‐sample t test. C)
   Representative mIHC staining of D‐TGCT (n = 10) and L‐TGCT (n = 10)
   lesions. Scale bar, 200 µm. Histogram on the right shows the
   percentages of MMP3 ^+ Fbs in D‐TGCT and L‐TGCT. D) Representative mIHC
   staining of D‐TGCT (n = 10) and L‐TGCT (n = 10) lesions. Scale bar,
   200 µm. Histogram on the right shows the percentages of APOE ^+ Fbs in
   D‐TGCT and L‐TGCT. E) GO enrichment analysis of MMP3 ^+ Fb marker
   genes. F) GO enrichment analysis of APOE ^+ Fb marker genes. G) Circle
   plot showing the cell–cell communication between APOE ^+ Fbs and other
   immune cells. Line thickness indicates the relative interaction
   intensity between D‐TGCT and L‐TGCT. H) Dot plot showing the
   ligand–receptor interaction of the chemokine pathway between APOE ^+ Fb
   and Mφ subpopulations.

Figure 5.

   Figure 5
   [124]Open in a new tab

   Identification of invasion‐related subpopulations of tumor cells
   (meta‐program 3) and the ability of ROR1 and PRKD1 to predict
   recurrence. A) Heatmap showing PCC among 94 sample‐level programs,
   calculated based on their top 50 marker genes. Programs were clustered
   into three MPs with similar functions (framed by black dashed lines).
   B) GO and KEGG enrichment analysis of the three identified MPs. C) Pie
   charts showing the proportions of MPs between D‐TGCT and L‐TGCT. D)
   Correlative network showing CCI based on scRNA‐seq and array data.
   Nodes represent different tumor cell MPs or other cell subpopulations.
   Line thickness indicates the strength of correlation. E) Representative
   mIHC images of CD34 ^+ Fbs treated with CM of synoviocytes from D‐TGCT
   in the absence (top) and presence (bottom) of integrin αvβ8 receptor
   Ab. Scale bar, 50 µm. F) Identification of PRKD1 as the marker gene for
   MP3 tumor cells. G) Representative IHC staining images of PRKD1 in
   D‐TGCT, L‐TGCT, and OA synovium. Scale bar, 20 µm. Histogram on the
   right shows the percentages of PRKD1 ^+ cells in D‐TGCT, L‐TGCT, and OA
   synovium. Statistical significance was inferred by two‐sample t test.
   ****P < 0.0001, ns = not significant. H) Representative IHC staining
   images of ROR1 in non‐recurrent (n = 17) and recurrent (n = 16) D‐TGCT.
   Scale bar, 20 µm. Histogram on the right shows the percentages of ROR1
   ^+ cells in non‐recurrent and recurrent D‐TGCT. Statistical
   significance was inferred by two‐sample t test. ***P < 0.001. I)
   Representative IHC staining images of PRKD1 in non‐recurrent (n = 17)
   and recurrent (n = 16) D‐TGCT. Scale bar, 20 µm. Histogram on the right
   shows the percentages of PRKD1 ^+ cells in non‐recurrent and recurrent
   D‐TGCT. Statistical significance was inferred by two‐sample t test.
   *P < 0.05. J) Receiver operating characteristic (ROC) curve evaluating
   the performance of ROR1 and PRKD1 in predicting disease recurrence. K)
   Determination of the threshold for ROR1 and PRKD1 positivity rates
   using the Youden index.

   For APOE ^+ Fbs, enriched GO terms were related to functions of
   regulation in myeloid differentiation and chemokine‐mediated signaling
   pathway (Figure [125]3F). CCI analysis also suggested that APOE ^+ Fbs
   had the strongest interactions with IL‐1B ^+ CCL20 ^+ Mφs
   (Figure [126]3G), the major Mφ subpopulation related to the invasive
   ability of D‐TGCT reported in the previous subsection of this article.
   Specifically, of all ligand–receptor pairs in the chemokine and
   cytokine pathways, the CXCL12/CXCR4 axis had the highest communication
   probability between APOE ^+ Fbs and IL‐1B ^+ CCL20 ^+ Mφs
   (Figure [127]3H). Such communication has been reported to mediate the
   transformation of circulating monocytes into perivascular Mφs in the
   brain, causing neuroinflammation.^[ [128]^32 ^] Taken together, the
   above‐mentioned analyses suggested that APOE ^+ Fbs in D‐TGCT might
   activate IL‐1B ^+ CCL20 ^+ Mφs through the CXCL12/CXCR4 axis and that,
   together with MMP3 ^+ Fbs, they might contribute to the aggression of
   D‐TGCT.

2.4. Tumor Cells Promoted Differentiation of CD34 ^+ Fibroblasts into MMP3 ^+
Fibroblasts and APOE ^+ Fibroblasts

   To further investigate why we observed higher proportions of MMP3 ^+
   Fbs and APOE ^+ Fbs in D‐TGCT, we drew a diffusion map of our scRNA‐seq
   cohort and reconstructed the differentiation trajectory of Fb subsets.
   The results showed that CD34 ^+ Fbs had the smallest diffusion mapping
   value (dptval) and were located at the beginning of the trajectory,
   whereas APOE ^+ Fbs and MMP3 ^+ Fbs were developed at the end of the
   trajectory (Figure [129] 4A,B). In addition, RNA velocity showed a
   similar developmental trajectory, starting with CD34 ^+ Fbs and ending
   with APOE ^+ Fbs and MMP3 ^+ Fbs (Figure [130]4C). We found a higher
   proportion of CD34 ^+ Fbs, which have been found to be fibroblast
   progenitors,^[ [131]^33 ^] in L‐TGCT than in D‐TGCT (Figure [132]3A;
   Figure [133]S3C,D, Supporting Information). These results suggested a
   possible underlying mechanism in D‐TGCT regulating the differentiation
   of CD34 ^+ Fbs into APOE ^+ Fbs and MMP3 ^+ Fbs.

Figure 4.

   Figure 4
   [134]Open in a new tab

   Identification of tumor cells and their specific markers. A) Diffusion
   maps visualizing the differentiation trajectories of CD34 ^+ Fbs, POSTN
   ^+ Fbs, APOE ^+ Fbs, and MMP3 ^+ Fbs. Cells were colored by inferred
   diffusion pseudotime (left) and subpopulation (right). B) CD34 ^+ Fbs,
   POSTN ^+ Fbs, APOE ^+ Fbs, and MMP3 ^+ Fbs ordered by diffusion map
   pseudotime. C) RNA velocities of CD34 ^+ Fbs, POSTN ^+ Fbs, APOE ^+
   Fbs, and MMP3 ^+ Fbs visualized as a streamline plot (left) and a
   partition‐based graph abstraction (PAGA) plot (right) in a UMAP‐based
   embedding. D) Correlative network showing CCI based on scRNA‐seq and
   array data. Nodes represent different cell subpopulations. Line
   thickness indicates the strength of correlation. E) Heatmap showing the
   CNV profile of each Fb subpopulation. ECs, B cells, and T cells were
   selected as reference cells. Blue and red colors represent lost and
   amplified chromosomes, respectively. F) Box plot showing the CNV scores
   of different Fb subpopulations. G) Identification of ROR1 as the marker
   gene for TGCT tumor cells. H) Representative IHC staining images of
   ROR1 in D‐TGCT, L‐TGCT, and OA synovium. Scale bar, 20 µm. Histogram on
   the right shows the percentages of ROR1 ^+ cells in D‐TGCT, L‐TGCT, and
   OA synovium. Statistical significance was inferred by two‐sample
   t test. ****P < 0.0001, ***P < 0.001. Scale bar, 200 µm.

   Subsequently, we applied a computational model previously used in
   cancer research to further characterize this hypothetical regulatory
   mechanism.^[ [135]^34 ^] Using a public‐microarray dataset of TGCT,^[
   [136]^35 ^] researchers can identify genes that exhibit strong
   correlations with high abundance of a particular cell population, such
   as MMP3 ^+ Fbs or APOE ^+ Fbs. We speculated that these genes might
   potentially regulate changes in cell abundance of the specific cell
   population. When matching these highly correlated genes back to our
   scRNA‐seq data, we could thus infer cell types that highly expressed
   those genes and denote them as potential upstream regulators. We
   thereby established a correlative CCI network and identified strong
   interactions between tumor cells and Fbs that were CD34 ^+, MMP3 ^+, or
   APOE ^+ (Figure [137]4D; and Table [138]S4, Supporting Information).
   These analyses suggested that tumor cells might promote differentiation
   of CD34 ^+ Fbs into APOE ^+ Fbs and MMP3 ^+ Fbs in D‐TGCT, turning
   “mild” fibroblast clusters into “aggressive” ones.

2.5. Tumor Cells in TGCT Exhibited Higher Copy Number Variation and were
Characterized by Expression of ROR1

   To detect the functional characteristics of tumor cells and reveal the
   specific mechanisms by which they regulated the development of
   aggressive Fbs, we first established a procedure to identify tumor
   cells and their specific markers. Chromosomal abnormalities, most
   commonly translocation of chromosome 1p11–13, have recently been
   characterized as the major pathogenic mechanisms in TGCT.^[ [139]^17 ^]
   Therefore, we suspected that quantifying chromosomal instability, such
   as copy number variation (CNV) events, in Fb subsets could effectively
   identify neoplastic cells. To that end, we initially performed CNV
   inference on all Fb subpopulations using inferCNV (Broad Institute,
   Cambridge, MA, USA), with ECs, T cells, and B cells selected as
   references (Figure [140]4E). Subsequently, the CNV score of each subset