Single-cell landscape of human adipose tissue reveals THBS1 as a potential driver of fibro-inflammation in metabolic disease

Abstract

The cellular complexity of adipose tissue plays a critical role in the pathogenesis of metabolic disease; however, mechanisms underlying adipose dysfunction remain poorly understood. Here, we constructed a single-nucleus and bulk transcriptomic atlas of human visceral adipose tissues from an ethnically homogeneous cohort with defined metabolic states: lean with normal glucose tolerance, obesity, and obesity with type 2 diabetes. Single-nucleus RNA sequencing (snRNA-seq) revealed a fibro-inflammatory subpopulation of adipose stem cells and adipocytes characterized by high THBS1 expression, which were enriched in individuals with obesity and obesity with type 2 diabetes compared to lean controls. Through complementary in vitro (human primary ASCs, THP-1 macrophages, and Matrigel-captured mature adipocytes), ex vivo (mouse and human visceral adipose tissue explants), and in vivo (gain-of-function and antibody-mediated neutralization) experiments, we demonstrated that THBS1 promoted adipose fibrosis and inflammation via macrophage-mediated TGFβ activation, impairing adipogenesis and adipocyte function. Inhibition of THBS1 attenuated fibro-inflammatory features in metabolically dysfunctional adipose tissues. Moreover, THBS1 levels in plasma and adipocytes correlated with metabolic traits, including BMI and HbA1c. Collectively, our study provides a cellular atlas of human visceral adipose tissues across metabolic disease states and identifies THBS1 as a key pathogenic driver and potential therapeutic target.

Fibro-inflammatory remodeling of visceral fat accompanies obesity and type 2 diabetes, but its drivers are poorly defined. Here, the authors show, using single-nucleus transcriptomics of human visceral adipose tissue, that THBS1 promotes fibrosis and inflammation via macrophage TGFβ activation.

Introduction

Adipose tissue is essential for metabolic homeostasis, functioning both as an energy reservoir and a modulator of systemic inflammation^1–3. Due to its cellular heterogeneity, which includes adipocytes, adipose-derived stem cells (ASCs), immune cells, and endothelial cells, understanding the pathological processes underlying adipose tissue dysfunction requires in-depth investigation at single-cell resolution.

Visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT) exhibit distinct metabolic properties. VAT drains into the portal circulation, exposing the liver to elevated free fatty acids and inflammatory mediators, and exhibits more pronounced inflammatory and fibrotic remodeling than SAT under obesity-related metabolic disease conditions. Clinical studies have consistently shown that VAT accumulation, rather than total adiposity, is more strongly associated with insulin resistance, type 2 diabetes, and cardiovascular disease^4,5.

Recent advances in single-cell transcriptomic profiling have provided valuable insights into the cellular and molecular mechanisms underlying adipose tissue dysfunction^6–8. Notably, ref. ⁹ and ref. ¹⁰ generated comprehensive single-cell atlases of adipose tissue across individuals with varying degrees of obesity, while ref. ¹¹ and ref. ¹² conducted longitudinal analyses of human adipose tissue before and after bariatric surgery, revealing dynamic transcriptomic and epigenetic remodeling. Although these studies have identified potential marker genes and cell subpopulations associated with adipose dysfunction, their functional roles remain incompletely understood. This gap hinders the translation of these findings into actionable therapeutic strategies for metabolic diseases. Moreover, these studies included heterogeneous cohorts without clear stratification by metabolic disease stage, limiting their ability to delineate stage-specific cellular signatures and robust disease-associated markers.

In the present study, we addressed these limitations by establishing an ethnically homogeneous human cohort with clearly defined metabolic disease states, namely lean, obesity, and obese with type 2 diabetes, to enable comprehensive comparisons across disease stages. Using bulk and single-nucleus transcriptomic analyses of visceral adipose tissue, we constructed a detailed cellular and molecular atlas and identified key pathways driving adipose fibro-inflammation. Specifically, we observed upregulation of thrombospondin 1 (THBS1) in ASCs and adipocytes as a hallmark of fibro-inflammatory remodeling. In an independent validation human cohort, we found that THBS1 promotes fibro-inflammation and impairs adipogenesis through both direct and paracrine mechanisms, particularly via adipocyte–macrophage crosstalk. Furthermore, we explored the potential of THBS1 as a candidate blood biomarker and a therapeutic target to reverse adipose fibro-inflammation, using a chemical inhibitor and neutralizing antibody. Full validation of THBS1 as a diagnostic biomarker will require assessment in larger and more diverse clinical cohorts. Collectively, our findings highlight THBS1 as a key mediator of adipose tissue dysfunction and a promising target for therapeutic intervention in metabolic disease.

Results

Single-nucleus transcriptomics reveal distinct cellular states in human visceral adipose tissues across metabolic disease status

To characterize the cellular and molecular phenotypes of adipose tissue across stages of metabolic disease, we analyzed visceral adipose tissue surgical samples from 21 participants stratified by metabolic status: lean (Lean; body mass index [BMI] <23 kg/m², hemoglobin A1c [HbA1c] <6.5%), obesity (Ob; BMI ≥30 kg/m², HbA1c < 6.5%), and obese with type 2 diabetes (Ob-DM; BMI ≥30 kg/m², HbA1c ≥6.5%) (Fig. 1A and Supplementary Fig. 1A for clinical traits). Principal component analysis of bulk RNA-seq data (n = 20, with one sample excluded because of QC failure) revealed distinct transcriptional profiles associated with metabolic status (Fig. 1B). Gene set enrichment analysis (GSEA) showed that the Lean group exhibited enrichment in gene sets related to oxidative phosphorylation, adipogenesis, and fatty acid metabolism, while the Ob and Ob-DM groups were enriched for gene sets involved in inflammatory response, hypoxia, and TGFβ signaling (Fig. 1C).

Fig. 1. Single-nucleus transcriptomics reveal distinct cellular states in human visceral adipose tissues across metabolic disease status.

A Study design. 21 individuals (13 female and 8 male) were categorized into lean, obesity (Ob), and obesity with diabetes (Ob-DM) groups. B, C Bulk RNA-seq of visceral adipose tissue. B Principal component analysis. C Gene set enrichment analysis (fgsea; Benjamini–Hochberg-adjusted P < 0.05). D–H Single-nucleus transcriptomics of visceral adipose tissue. D Uniform manifold approximation and projection of 86,000 nuclei. E Violin plot with cell type-specific markers. F Cell type proportion according to metabolic disease status (n = 8 lean, 5 Ob, 8 Ob-DM individuals). The first quartile, median and third quartile are shown, with whiskers extending 1.5 × the interquartile range (IQR). G Heatmap showing donor score matrix and corresponding metadata associations found without using any batch correction. P values were calculated using univariate linear model F-tests with Benjamini–Hochberg FDR correction. H Heatmap showing Benjamini–Hochberg-adjusted P values for gene set enrichment analysis in each cell type for factor 4. The enrichment is divided into “up” (gene sets with positive normalized enrichment score (NES), red) and “down” (gene sets with negative NES, blue). Bar plot (top) shows the number of significantly enriched gene sets per cell type. *P < 0.05, **P < 0.01, and ***P < 0.001. Panel A was created in BioRender. Jung, J. (2026) https://BioRender.com/eeqajwq.

To identify cell type-specific transcriptional alterations driven by metabolic status, we performed snRNA-seq on visceral adipose tissue from the same 21 individuals. A total of 86,000 high-quality nuclei passed quality control criteria. Based on canonical marker gene expression, we identified eight distinct cell types: adipocytes, ASCs, mesothelial cells, endothelial cells, lymphatic endothelial cells, smooth muscle cells, myeloid cells, and T/NK cells (Fig. 1D–F and Supplementary Fig. 1B–E). Notably, the proportions of adipocytes and ASCs decreased, while the proportion of myeloid cells increased in individuals with obesity.

Next, to examine inter-individual transcriptional variation associated with metabolic status across cell types, we applied single-cell interpretable tensor decomposition (scITD)¹³, which identified five distinct transcriptional programs in an unbiased manner. Program 2 was associated with metabolic state, while Program 1 exhibited sex-specific variation (Fig. 1G). Genes contributing to Program 2 in ASCs and adipocytes were enriched in pathways related to metabolic processes and immune response regulation, consistent with findings from the bulk RNA-seq analysis (Fig. 1H). Together, these results suggest that transcriptional reprogramming and alterations in cellular composition, particularly in ASCs, adipocytes, and myeloid cells, are key features of adipose tissue remodeling associated with metabolic disease.

Fine dissection of the immune compartment identifies proinflammatory SLC11A1⁺ macrophages associated with metabolic disease status

Immune cells were subdivided into four major populations: dendritic cells, two macrophage subsets, and T/NK lymphocytes (Fig. 2A and Supplementary Fig. 2A). Notably, reciprocal changes in macrophage subpopulations were observed across metabolic states (Fig. 2B and Supplementary Fig. 2B). SLC11A1-expressing macrophages (Mac.SLC11A1) expanded from Lean to Ob to Ob-DM groups, whereas LYVE1-expressing macrophages (Mac.LYVE1) gradually declined. Genes enriched in Mac.SLC11A1 were associated with inflammatory pathways, including TNF production and NF-κB signaling, while those enriched in Mac.LYVE1 were linked to vascular-related functions, such as endothelial cell chemotaxis and blood vessel morphogenesis (Fig. 2C). Mac.SLC11A1 exhibited elevated expressions of lipid-associated macrophage (LAM) markers, including APOE, TREM2, GPNMB, and PLIN2, as well as M1-associated markers, including TNF and IL1B. Mac.LYVE1 showed higher expression of tissue-resident macrophage and M2-associated markers such as LYVE1, FOLR2, and MRC1, indicating distinct functional states between the two populations (Supplementary Fig. 2C).

Fig. 2. Fine dissection of the immune compartment identifies proinflammatory SLC11A1⁺ macrophages associated with metabolic disease status.

A–F Single-nucleus transcriptomics of immune cells in visceral adipose tissue. A UMAP of immune cells; color indicates cell type. B Cell type proportions by metabolic disease status (n = 8 lean, 5 Ob, 6 Ob-DM). Box plots show first quartile, median, and third quartile with whiskers extending 1.5 × IQR. Two-sided Wilcoxon rank-sum test with Benjamini–Hochberg correction; adjusted P < 0.05. C Gene Ontology (Biological Process) enrichment for Mac.SLC11A1 and Mac.LYVE1 marker genes using topGO with Fisher’s exact test and the elim algorithm; P values not adjusted for multiple comparisons. D Heatmap of donor score matrix and metadata associations. P values calculated using univariate linear model F-tests or logistic regression with χ² tests (discrete variables); Benjamini–Hochberg FDR correction applied. E Gene loading scores for factor 4; genes with Benjamini–Hochberg-adjusted P < 0.05 shown. Significance determined by univariate linear regression with FDR correction. F PROGENy pathway activity heatmap across conditions and cell types. G–I Flow cytometry of visceral adipose tissue macrophages (CD68⁺CD45⁺) from lean, Ob, and Ob-DM groups. G Proportion of SLC11A1⁺ macrophages; n = 5 lean, 6 Ob, 6 Ob-DM; mean ± SEM; each data point represents one individual. One-way ANOVA with Tukey’s multiple-comparison test (lean vs Ob-DM P = 0.005, Ob vs Ob-DM P = 0.003). H, I Geometric mean fluorescence intensity of CD11c (H) and CD163 (I) in SLC11A1⁺ vs SLC11A1⁻ macrophages; n = 4 Ob-DM individuals; mean ± SEM; paired two-tailed t-test. J Representative immunofluorescence of human VAT from an Ob-DM individual. Left: SLC11A1 (green)/CD68 (red)/DAPI. Right: SLC11A1 (green)/IL6 (red)/DAPI. Scale bar, 100 μm. K Representative IHC of SLC11A1 and CD163 counterstained with Mayer’s hematoxylin. Scale bar, 100 μm. J, K Representative of ≥4 independent experiments. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

To further explore transcriptional dynamics within the immune compartment, we performed a separate immune-focused scITD analysis, which identified Program 4 as being associated with metabolic disease status (Fig. 2D). The genes contributing to this program were predominantly expressed in macrophages and included diabetes-associated genes such as GREB1 and LGMN (Fig. 2E). Pathway activity analysis revealed a gradual increase in proinflammatory signaling, including NF-κB, TNFα, and hypoxia pathways, in Mac.SLC11A1 cells across metabolic disease states (Fig. 2F).

To assess the physiological relevance of these findings, we mapped our macrophage subpopulations to known human adipose tissue macrophages (hMac1-3) using publicly available single-cell data from ref. ⁹. The proportion of hMac1 decreased with increasing BMI, while hMac2 and hMac3 were enriched in individuals with BMI between 35 and 50 (Supplementary Fig. 2D). The proinflammatory Mac.SLC11A1 subpopulation was predominantly enriched within the inflammatory hMac2 and hMac3 subsets (Supplementary Fig. 2E), supporting the enrichment of SLC11A1⁺ macrophages in metabolically unhealthy adipose tissue.

In an independent validation cohort, we further characterized the Mac.SLC11A1 subpopulation to confirm their inflammatory phenotype and tissue localization. Flow cytometric analysis revealed a significant increase in SLC11A1⁺ macrophages from Lean to Ob-DM individuals (Fig. 2G). SLC11A1 expression positively correlated with the expression of canonical M1 macrophage markers CD11c and CD80 (Fig. 2H and Supplementary Fig. 2F), while showing mutually exclusive expression with the M2 marker CD163 (Fig. 2I). Histological analysis corroborated these findings, demonstrating that SLC11A1⁺ macrophages localized to crown-like structures, co-expressed CD68 and IL6, and were distinct from CD163⁺ M2 macrophages (Fig. 2J, K). Furthermore, the proportion of SLC11A1⁺ macrophages showed significant positive correlations with adipose tissue fibrosis (FAT histological score), HbA1c, fasting glucose, and circulating IL6 (Supplementary Fig. 2G), linking this subset to both local fibro-inflammatory remodeling and systemic metabolic dysregulation. Taken together, these findings suggest that SLC11A1⁺ macrophages represent a proinflammatory subset enriched in adipose tissue of obesity-related clinical conditions and associated with adipose tissue fibrosis, glycemic control, and systemic inflammation.

Endothelial and mesothelial cells exhibit distinct inflammatory signatures in metabolic disease

The stromal compartment comprised endothelial cells (EC.1 (SEMA3G), EC.2 (RGCC), EC.3 (TLL1)), pericytes (PRKG1), and two mesothelial subtypes (classical (cMeso; CD55) and mesenchymal (meMeso; CDON)) characterized by their specific marker gene expression patterns (Supplementary Fig. 3). Significant compositional changes were observed in EC.3, cMeso, and meMeso across disease states (Supplementary Fig. 3B, C, G, H). Pathway activity analysis revealed subtype-specific activation of inflammatory and angiogenic pathways. In particular, EC.3 and cMeso exhibited increased activity of hypoxia, NF-κB, and TNFα signaling pathways across metabolic disease states (Supplementary Fig. 3E, J). The upregulation of these proinflammatory pathways in cMeso cells from Ob-DM is consistent with previous reports¹⁰ and suggests a potential role of cMeso cells in sustaining adipose tissue inflammation associated with metabolic disease.

THBS1⁺ adipose stem cells represent a distinct fibro-inflammatory subpopulation

To characterize the cellular heterogeneity within ASCs, we performed unsupervised subclustering of ASC populations. This analysis resolved three major ASC subsets, which were annotated according to their top differentially expressed marker genes as ASC.PI16, ASC.PPARG, and ASC.THBS1 (Fig. 3A–C). ASC.PI16 cells expressed canonical progenitor markers including PI16, CD55 and DPP4, consistent with the multipotent stem-like ASC phenotype previously described in adipose tissue¹⁴. ASC.PPARG cells showed elevated expression of adipogenic commitment genes, including PPARG, PIK3R1, and IGFBP3, consistent with a preadipocyte-like state primed for adipogenic differentiation. By contrast, ASC.THBS1 cells were characterized not only by THBS1 expression but also by a fibro-inflammatory and extracellular matrix-remodeling signature including THBS1, IL1R1¹⁵, HAS1¹⁶, and FGFR1¹⁷, COL1A1 and other collagen genes (Fig. 3B, C). Systematic comparison with published fibro-inflammatory ASC populations further showed that ASC.THBS1 shares core features with anti-adipogenic progenitor¹⁰, mesothelial-like stem cells¹⁴, CD9⁺ ASCs¹⁸, and SDC1⁺ fibrotic ASCs¹⁹, particularly through fibro-inflammatory genes such as COL8A1, CCL2, and TIMP1 (Supplementary Fig. 4A). Reference mapping of recently defined adipose progenitor cell (APC) clusters²⁰ to our ASC reference revealed that stressed (APC3) and profibrotic (APC5, ADAM12-high/POSTN-high) clusters were predominantly classified as ASC.THBS1 (57.5% and 53.3%, respectively) compared to other clusters (χ² test, P < 0.001; Supplementary Fig. 4B). Flow cytometry validation confirmed elevated SDC1 and CD9 protein expression in THBS1⁺ ASCs (P < 0.001; Supplementary Fig. 4C), validating the fibro-inflammatory phenotype at the protein level.

Fig. 3. THBS1⁺ adipose stem cells represent a distinct fibro-inflammatory subpopulation.

A–F Single-nucleus transcriptomics of ASCs in visceral adipose tissue. A UMAP of ASCs; color indicates cell type. B Heatmap of scaled marker gene expression across three ASC populations; genes hierarchically clustered with distinct gene clusters (C1–C3) indicated. C Gene ontology enrichment for gene clusters in (B) (topGO, Fisher’s exact test with elim algorithm; C1, n = 134; C2, n = 111; C3, n = 166 genes; background, 35,105 genes; P values not adjusted for multiple comparisons); color intensity indicates statistical significance. D Pseudotime trajectory analysis of ASC populations. E PROGENy pathway activity heatmap across ASC subpopulations. F ASC subtype proportions. G Flow cytometric quantification of THBS1⁺ ASCs (CD45⁻CD31⁻CD140⁺) in lean (n = 3), Ob (n = 3), and Ob-DM (n = 6) groups; mean ± SD; each data point represents one individual. One-way ANOVA (F₂,₉ = 30.048, P = 0.0001) with Tukey’s multiple-comparison test (lean vs Ob P = 0.0015, lean vs Ob-DM P = 0.0001). H, I Fibrosis marker expression in THBS1⁺ vs THBS1⁻ ASCs. Representative histograms of H ACTA2 and I COL6A1; black/red lines indicate THBS1⁻/THBS1⁺ ASCs from an Ob-DM patient. Geometric mean fluorescence intensity shown as mean ± SD (n = 5 Ob-DM individuals; paired two-tailed t-test within each individual). J Representative fluorescence image of adipogenic differentiation in human primary ASCs; BODIPY (green), PDGFRA (red), DAPI (blue). Scale bar, 10 μm. K BODIPY fluorescence intensity in primary ASCs treated with vehicle or rhTHBS1; mean ± SEM (n = 5 lean, 6 Ob-DM; paired two-tailed t-test within each individual). L, M Time-course analysis in ASC52Telo treated with vehicle or rhTHBS1 (20 nM). L BODIPY geometric mean fluorescence intensity (arb. units) on days 5, 10, 15, 20. M Proportion of ICAM1⁺PPARG⁺ adipogenic (left) and ICAM1⁺PPARG⁻ fibrotic (right) cells on days 5 and 10. Data in (L, M): mean ± SEM; n = 6 independent experiments (n = 5 for day-15 rhTHBS1 in (L) and day-10 vehicle in (M)); unpaired two-tailed t-test. N Representative immunofluorescence of COL6A1 (green) and ACTA2 (red) in ASC52Telo. Scale bar, 50 μm. O Proportion of COL6A1⁺ (left) and CD9⁺ (right) ASCs; mean ± SEM (n = 4 independent experiments; unpaired two-tailed t-test). *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Trajectory analysis inferred two putative differentiation paths: Path 1 leading toward an adipogenic state and Path 2 toward a fibro-inflammatory state (Fig. 3D). We further performed RNA velocity analysis using scVelo to independently assess trajectory directionality. RNA velocity vectors emanated from the ASC.PI16-enriched region toward both ASC.PPARG- and ASC.THBS1-enriched regions, supporting the bifurcating trajectory structure inferred by Palantir (Supplementary Fig. 4D). The terminal state of Path 1, predominantly composed of committed ASC.PPARG cells, was characterized by increased expression of PPARG and ADIPOQ. In contrast, Path 2, enriched in fibro-inflammatory ASC.THBS1 cells, exhibited activation of proinflammatory and fibrotic pathways (Fig. 3E). The proportion of ASC.THBS1 cells increased from Lean to Ob to Ob-DM, a trend validated by flow cytometry in an independent cohort of 12 individuals (Fig. 3F, G). In addition, a specific subtype of ASCs (hASPC6) identified from a previously published single-cell atlas⁹, which was enriched in the visceral adipose tissue of individuals with BMI 40–50, mapped to the terminal fibro-inflammatory state of Path 2 (Supplementary Fig. 4E).

To functionally characterize THBS1⁺ ASCs, we isolated these cells from the visceral adipose tissue of an independent validation cohort. Consistent with the findings from snRNA-seq analysis, THBS1⁺ ASCs exhibited significantly higher expression of fibrotic markers ACTA2 and COL6A1 compared to THBS1^– ASCs (Fig. 3H, I). Furthermore, treatment of primary ASCs with recombinant human THBS1 (rhTHBS1) upon differentiation inhibited healthy adipogenesis, as evidenced by a reduced formation of mature adipocytes and decreased PPARγ expression in rhTHBS1-treated ASCs (Fig. 3J, K and Supplementary Fig. 4F). For further validation of the inhibitory effect on adipogenesis of THBS1, we performed a time-course adipogenic differentiation assay using the immortalized human ASC cell line, ASC52Telo, with the presence or absence of recombinant THBS1. THBS1-treated cells showed perturbed adipogenic differentiation with reduced BODIPY from differentiation day 10 and thereafter (Fig. 3L). THBS1 treatment also significantly reduced the ICAM1⁺PPARγ⁺ committed preadipocyte population at day 10, while increasing the ICAM1⁺PPARγ⁻ non-adipogenic/fibrotic population (Fig. 3M), indicating that THBS1 impairs adipocyte differentiation at the early commitment stage by shifting the balance from adipogenic toward non-adipogenic cell fates. In parallel, recombinant THBS1 treatment of the ASC cell line dose-dependently increased the proportion of CD9⁺ and COL6A1⁺ ASCs by FACS and upregulated COL6A1 and ACTA2 protein expression by immunocytochemistry (Fig. 3N, O and Supplementary Fig. 4G–I), providing direct protein-level evidence that THBS1 alone is sufficient to drive ECM production and myofibroblast-like transition in adipose progenitor cells. Collectively, these results demonstrate that THBS1⁺ ASCs represent a distinct fibro-inflammatory subtype and that THBS1 directly drives a coordinated dual program, suppressing adipogenic commitment while promoting fibrotic conversion, thereby contributing to adipose tissue dysfunction in metabolic disease.

THBS1 expression in adipocytes is associated with fibrosis, inflammation, and systemic metabolic phenotypes

Given the enrichment of fibro-inflammatory ASC subpopulations under metabolically unhealthy conditions, we next sought to characterize the adipocyte subpopulation structure. We identified three adipocyte subsets: Adipo.PIK3R1, Adipo.PLIN4, and Adipo.THBS1 (Fig. 4A). These adipocyte clusters were reproducibly identified across condition-stratified clustering (Supplementary Fig. 5A, B). Adipo.PIK3R1 represented a metabolically healthy adipocyte subset, characterized by the expression of genes involved in adipogenesis, insulin signaling, and lipid metabolism (IGF1, LPL, PIK3R1, PPARG, and ADIPOQ) (Fig. 4C, D and Supplementary Fig. 5C). Adipo.PLIN4 was characterized by increased PLIN4 expression and enrichment of genes associated with fatty acid biosynthesis and homeostasis (Fig. 4C, D). Notably, Adipo.THBS1, similar to ASC.THBS1, exhibited strong expression of genes associated with fibro-inflammation, TGFβ signaling, and thrombosis (THBS1, SVEP1, and STAT3) (Fig. 4C). The proportion of Adipo.THBS1 cells increased across metabolic disease states, mirroring the trend observed in ASC.THBS1 cells (Fig. 4B).

Fig. 4. THBS1 expression in adipocytes is associated with fibrosis, inflammation, and systemic metabolic phenotypes.

A UMAP of adipocyte clusters: Adipo.PIK3R1, Adipo.PLIN4, and Adipo.THBS1. B Adipocyte subtype proportions across lean, Ob, and Ob-DM conditions. C Heatmap of scaled marker gene expression across adipocyte populations; hierarchically clustered with gene clusters (C1–C3) indicated. D Gene ontology enrichment for clusters in (C) (topGO, Fisher’s exact test with elim algorithm; P values not adjusted for multiple comparisons); color intensity indicates significance. E Representative immunocytochemistry of THBS1⁺ adipocytes; THBS1 (green), DAPI (blue). Scale bar, 100 μm. F THBS1 concentration in primary adipocyte lysates from lean (n = 11), Ob (n = 9), and Ob-DM (n = 11) individuals; mean ± SEM; one-way ANOVA (F₂,₂₈ = 22.48, P < 0.0001) with Tukey’s test (P = 0.012, lean vs Ob; P < 0.0001, lean vs Ob-DM; P = 0.008, Ob vs Ob-DM). G Correlation of THBS1 with BMI and HbA1c (n = 13; two-sided Spearman: BMI ρ = 0.571, P = 0.044; HbA1c ρ = 0.637, P = 0.022). H Immunocytochemistry and quantification of COL6A1, IL6, and 4-HNE in freshly isolated human adipocytes from lean, Ob, and Ob-DM groups. Violin plots show per-cell fluorescence intensity (arb. units); COL6A1 (n = 28/22/15 adipocytes, 5/4/4 donors), IL6 (n = 19/36/90 adipocytes, 6/7/10 donors), 4-HNE (n = 13/16/26 adipocytes, 4/5/6 donors); unpaired two-tailed t-test. Scale bars, 50  μm. I Schematic of rhTHBS1 treatment in Matrigel-captured mouse mature adipocytes (vehicle or 20 nM, 6 days). J qPCR of treated adipocytes; mean ± SEM (n = 5 per group for Adipoq and Pparg; n = 4 for Col1a1 and Ccl2; unpaired two-tailed t-test). K Immunocytochemistry of COL6A1 (green) and TNF (red); quantification of COL6A1⁺ and TNF⁺ ratios; mean ± SEM (n = 7; unpaired two-tailed t-test). Scale bar, 50 μm. L Schematic of lentiviral shThbs1 knockdown in HFD mouse mature adipocytes using the MAAC method. M 2-NBDG glucose uptake in shThbs1 vs scramble adipocytes; mean ± SEM (n = 10; unpaired two-tailed t-test). Scale bar, 50 μm. N pAKT/AKT and GLUT4/GAPDH by western blot; mean ± SD (pAKT, n = 6/5; GLUT4, n = 5/4; unpaired two-tailed t-test). *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Panel (I) was created in BioRender. Jung, J. (2026) https://BioRender.com/4j2r0p1. Panel (L) was created in BioRender. Jung, J. (2026) https://BioRender.com/56pe5ch.

We validated these findings using human primary adipocyte fractionated from human VAT samples obtained from an independent cohort. THBS1 protein levels in adipocytes increased from Lean to Ob to Ob-DM groups (Fig. 4E, F). Furthermore, THBS1 expression in adipocytes was positively correlated with clinical metabolic indicators, including BMI (Spearman’s ρ = 0.58, P = 0.044) and HbA1c (ρ = 0.64, P = 0.022) (Fig. 4G). In addition, markers of fibrosis, inflammation, and cellular stress (COL6A1, IL6, and 4-HNE) were significantly upregulated in THBS1-expressing adipocytes and showed increasing trends across metabolic disease stages (Fig. 4H and Supplementary Fig. 5D). Hyperglycemic and hyperinsulinemic stimulation of primary adipocytes robustly induced Thbs1 expression (Supplementary Fig. 5E).

To assess the functional role of THBS1 in adipocytes, we treated Matrigel-embedded mouse primary adipocytes with rhTHBS1. This treatment inhibited the expression of key adipogenic and functional markers (Pparg and Adipoq) and significantly upregulated fibro-inflammatory related genes (Col1a1 and Ccl2) (Fig. 4J and Supplementary Fig. 5F). Consistently, rhTHBS1 reduced adiponectin secretion into the conditioned media (Supplementary Fig. 5G) and increased the proportion of adipocytes positive for COL6A1 and TNF by immunocytochemistry (Fig. 4K), corroborating that THBS1 directly impairs adipocyte endocrine function while driving fibro-inflammatory remodeling at both transcriptional and protein levels. We further performed reciprocal gain- and loss-of-function experiments to assess the role of adipocyte THBS1 in glycemic regulation. Recombinant THBS1 reduced 2-NBDG glucose uptake in primary adipocytes from standard chow diet-fed mice (Supplementary Fig. 5H, I), whereas shRNA-mediated THBS1 knockdown in adipocytes from 14-week high-fat diet (HFD)-fed mice, in which THBS1 was markedly upregulated (Supplementary Fig. 5J), significantly increased 2-NBDG uptake (Fig. 4M) and elevated the phospho-AKT/total AKT ratio and GLUT4 expression (Fig. 4N and Supplementary Fig. 5K, L). Furthermore, mRNA expression of Glut4 and Fabp4 were significantly increased by Thbs1 knockdown in adipocytes (Supplementary Fig. 5M), indicating enhanced insulin signaling. These reciprocal findings demonstrate that THBS1 directly impairs AKT-mediated insulin signaling and suppresses adipocyte glucose uptake and lipid transport. Collectively, these results suggest that THBS1 expression in adipocytes contributes to the development of adipose tissue fibrosis, inflammation, and functional impairment, thereby implicating adipocyte THBS1 signaling in systemic metabolic dysregulation.

Bulk RNA-seq of visceral adipose tissues identifies metabolic disease-associated transcriptional programs

To complement the snRNA-seq data and assess disease-related transcriptional changes, we performed bulk RNA-seq of visceral adipose tissue from 20 individuals matched to the snRNA-seq cohort. Differential expression analysis identified 1778 and 2520 differentially expressed genes (DEGs) in the Ob and Ob-DM groups, respectively, compared with the Lean group. Notably, marker genes in our snRNA-seq data—including THBS1, SLC11A1, FN1, IL6, TGFB1, PPARG, and ADIPOQ—were consistently recapitulated in the bulk RNA-seq data across disease stages (Fig. 5A, B).

Fig. 5. Bulk RNA-seq of visceral adipose tissues identifies metabolic disease-associated transcriptional programs.

A, B Volcano plots of DEGs in obesity with diabetes (Ob-DM) and obesity (Ob) groups compared to those in the lean group. Major markers identified in snRNA-seq, including THBS1, SLC11A1, TGFB1, FN1, IL6, PPARG, and ADIPOQ are highlighted. Significant genes were filtered based on log₂ fold change >0.58 or <−0.58 and adjusted P < 0.05 (Benjamini–Hochberg). C Venn diagrams illustrating DEGs between Ob and Ob-DM groups relative to lean groups. Genes are categorized into six groups: g1 (546 genes upregulated in Ob-DM only), g2 (1042 genes upregulated in both Ob and Ob-DM), g3 (194 genes upregulated in Ob only), g4 (419 genes downregulated in Ob-DM only), g5 (513 genes downregulated in both Ob and Ob-DM), and g6 (29 genes downregulated in Ob only). D, E KEGG pathway enrichment analysis comparing DEGs from g2 and g1 groups. Selected KEGG terms were shown based on biological relevance and statistical significance (adjusted P < 0.05) as bubble plots, where bubble size represents the number of genes involved in each pathway and color intensity indicates the odds ratio. The red dashed line indicates the statistical significance threshold. F Heatmap of module eigen gene expression by WGCNA for each group. G Heatmap depicting Pearson correlations between gene modules and clinical features. H, I KEGG pathway enrichment analysis of module 1 (H) and 6 (I), represented as bubble plots. The size of each bubble corresponds to the number of genes involved in the pathway, and the color intensity highlights the odds ratio red dotted line indicates points P < 0.05. J Protein–protein interaction (PPI) network of primary interactors of THBS1 within module 1. Interactions were obtained from STRING (v11.5) using a string score cutoff of 700, and the network was visualized using the iGraph and ggplot2 R packages. Nodes represent individual proteins, and edge thickness reflects interaction strength. The interaction strength is color-coded, with darker orange indicating stronger interactions. *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001.

We categorized the DEGs into six gene sets (g1–g6) based on the overlap and specificity from pairwise comparisons of Lean vs. Ob and Lean vs. Ob-DM. Among these, 1555 DEGs were shared (g2: 1042 upregulated; g5: 513 downregulated), while 965 were unique to Ob-DM (g1: 546 upregulated; g4: 419 downregulated) (Fig. 5C). The shared upregulated DEGs (g2) were enriched for pathways such as phagosome formation, extracellular matrix (ECM) organization, cell adhesion, and TGFβ signaling (Fig. 5D). Ob-DM–specific upregulated DEGs (g1) showed greater enrichment in inflammatory and fibrotic pathways, including TNF, NF-κB, Hippo signaling, and hypoxia (Fig. 5E). The shared downregulated DEGs (g5) were associated with PPAR and insulin signaling, while Ob-DM–specific downregulated genes (g4) were enriched for fatty acid metabolism and AMPK signaling (Supplementary Fig. 6A, B). Together, these findings support a fibro-inflammatory transcriptional program in both Ob and Ob-DM groups, with pronounced impairment of lipid metabolism and energy-sensing pathways in the Ob-DM group.

To identify co-regulated transcriptional programs, we applied weighted gene co-expression network analysis (WGCNA)²¹, which identified seven functional gene modules based on eigengene correlation (Fig. 5F). Modules 1–3 displayed stepwise upregulation from Lean to Ob-DM and showed positive correlation with HbA1c, BMI, and LDL levels. Module 4 was upregulated in both Ob and Ob-DM and was positively correlated with HbA1c, BMI, and HOMA-IR. In contrast, Module 6 was enriched in Lean and negatively correlated with HbA1c and BMI (Fig. 5G). Module 1, which includes THBS1, TGFB1, and multiple collagen genes, was enriched for TGFβ, Hippo, insulin resistance, ECM-receptor interaction, and hypoxia pathways (Fig. 5H and Supplementary Fig. 6C), a pattern consistent with the pathway enrichment observed in ASC.THBS1 cells. Conversely, Module 6, specific to the Lean group, featured enrichment in the TCA cycle, fatty acid metabolism, and PPAR signaling, mirroring the transcriptomic profile of ASC.PPARG cells (Fig. 5I and Supplementary Fig. 6D). Detailed information on the genes included in each module and their enriched pathways is provided in Supplementary Fig. 6E–H and Supplementary Data 1 and 2.

To further explore the regulatory landscape, we performed a protein–protein interaction analysis of Module 1 using the STRING database. This analysis revealed a THBS1-centered network that included fibrosis-related genes such as ITGB1, COL1A1, FN1, and TGFBs, as well as canonical THBS1 receptors CD36 and CD47, supporting a central role for THBS1 in the fibro-inflammatory axis of adipose tissue (Fig. 5J). Together, these results validate and extend the snRNA-seq findings by uncovering coordinated fibro-inflammatory and metabolic gene programs in adipose tissue.

THBS1-dependent intercellular communication drives fibro-inflammatory changes in adipose tissues

To investigate intercellular communication differentially regulated according to metabolic disease status in adipose tissue, we performed ligand–receptor interaction analysis across cell types using CellChat²². THBS1⁺ cells (ASC.THBS1 and Adipo.THBS1) were the most prominent contributors to altered cell-to-cell communication between Ob-DM and Ob (Fig. 6A, B). Grouping 77 differentially regulated ligand–receptor pairs into 38 signaling pathways, we identified 25 pathways that were more enriched in Ob-DM than in Ob. Notably, THBS- and fibrosis-related pathways (e.g., collagen and FN1) were among the most strongly upregulated (Fig. 6C).

Fig. 6. THBS1-dependent intercellular communication drives fibro-inflammatory changes in adipose tissues.

A Heatmap of differential cell–cell interaction strengths between Ob and Ob-DM groups. B Scatter plots of incoming and outgoing interaction strengths for Ob and Ob-DM; point size indicates the number of significant interactions. C Bar plot of relative information flow of signaling molecules; bar length indicates contribution; red indicates highly active signals. D Schematic of functional assays: THP-1 macrophages treated with rhTHBS1 ± LSKL; conditioned media (CM) applied to ASC52Telo ± SB431542 to assess TGFβ-mediated fibrotic conversion. E TGFβ concentration in CM from THP-1 macrophages treated with vehicle or rhTHBS1 (5, 10, 20 nM) ± LSKL (25, 50, 100 nM); mean ± SEM (n = 4 per condition; one-way ANOVA with Šidák’s test). F Representative immunofluorescence of COL6A1 (green) and ACTA2 (red) in ASC52Telo treated with macrophage CM. Scale bar, 50 μm. G Quantification of ACTA2 and COL6A1 fluorescence intensity; mean ± SEM (n = 5/7/5/7/9 independent samples; one-way ANOVA with Šidák’s test for selected pairs). H Schematic of transwell co-culture: THP-1 macrophages in upper insert treated with vehicle, rhTHBS1 (20 nM), rhTHBS1 + LSKL (100 nM), or rhTHBS1 + SB431542 (100 nM); lower compartment contains ASC52Telo (I) or mouse mature adipocytes (J). I Representative immunofluorescence of COL6A1 (green) and ACTA2 (red) in cocultured ASC52Telo. Scale bar, 50 μm. Quantification of COL6A1 and ACTA2 fluorescence intensity; mean ± SEM (n = 4/4/5/4; one-way ANOVA with Šidák’s test). J Representative immunofluorescence of COL6A1 (green) and TNF (red) in cocultured mouse adipocytes. Scale bar, 50 μm. COL6A1⁺ adipocyte ratio; mean ± SEM (n = 11/13/3/4; one-way ANOVA with Šidák’s test). *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Panel (D) was created in BioRender. Jung, J. (2026) https://BioRender.com/i0djmnw. Panel (H) was created in BioRender. Jung, J. (2026) https://BioRender.com/zwvqp7k.

Given the prominent changes in intercellular communication involving THBS1⁺ cells, we next performed a targeted ligand–receptor interaction analysis using MultiNicheNet²³. After accounting for interpatient heterogeneity, we found that THBS1-mediated signaling from both Adipo.THBS1 and ASC.THBS1 was significantly increased in Ob-DM, targeting a broad range of recipient cell subpopulations, including ASCs, adipocytes, immune cells, and endothelial cells (Supplementary Fig. 7A, B). These findings indicate a central role for THBS1⁺ cell subtypes in mediating crosstalk with multiple cell types, contributing to adipose tissue dysfunction and metabolic disease.

To explore the spatial organization of THBS1 signaling, we analyzed publicly available spatial transcriptomic data from adipose tissue samples of lean (BMI, 20 kg/m²) and obese (BMI, 36 kg/m²) individuals²⁴. In obese adipose tissues, a significantly larger area expressed THBS1, its receptors (LRP5, ITGB1, and CD47), and fibrosis-related genes (COL1A1, ACTA2, and FN1) (Supplementary Fig. 7C). Importantly, within obese adipose tissues, THBS1⁺ spatial spots were more frequently adjacent to spots expressing THBS1 receptors and fibrosis-related genes, compared to THBS1^- spots. These results suggest that enhanced localized THBS1-mediated intercellular communication contributes to fibrosis by increasing spatial proximity between THBS1, its receptors, and fibrosis-related pathways in metabolically unhealthy adipose tissue.

THBS1 promotes adipose tissue fibrosis through macrophage-mediated TGFβ activation

THBS1 is a glycoprotein that promotes tissue remodeling, inflammation, and angiogenesis by interacting with multiple binding partners^25,26. To investigate its pathological role, mouse visceral adipose tissues were cultured ex vivo in the presence of rhTHBS1 for 4 days. This treatment increased the expression of fibrotic markers (Col1a1, Col6a3, and Tgfb1) and proinflammatory cytokine (Il6), along with decreased expression of adipocyte markers (Fabp4 and Pparg; Supplementary Fig. 8A, B).

THBS1 has been proposed to convert latent TGFβ to its active form, thereby promoting tissue fibrosis^26,27. Given that TGFB1 was highly expressed in myeloid and T/NK cells (Supplementary Fig. 8C), we hypothesized that THBS1 derived from ASCs and adipocytes may trigger fibro-inflammation via TGFβ activation in immune cells. To test this, we treated differentiated THP-1 macrophage cells with rhTHBS1, which induced TGFβ secretion in a dose-dependent manner, whereas co-treatment with LSKL reduced TGFβ secretion (Fig. 6D, E). Conditioned media from rhTHBS1-treated macrophages were subsequently treated with the human ASC cell line. This conditioned media treatment upregulated the expression of fibro-inflammatory markers such as COL6A1 and ACTA2 in human ASCs (Fig. 6F, G). Co-treatment with the TGFβRI inhibitor SB431542 abolished this conditioned media-induced fibrotic response in ASCs, confirming that macrophage-derived paracrine signaling drives ASC fibrotic activation in a TGFβ-dependent manner (Fig. 6F, G and Supplementary Fig. 8D, E). To further validate this paracrine interaction in a bidirectional setting, we established ASC–THP-1 and adipocyte–THP-1 transwell co-culture systems (Fig. 6H). Pharmacological inhibition of either THBS1–TGFβ interaction (LSKL) or TGFβRI signaling (SB431542) attenuated THBS1-induced fibrotic marker expression in both ASC–THP-1 co-cultures and Matrigel-captured mature adipocytes, whereas TNF expression was not attenuated by either inhibitor (Fig. 6I, J and Supplementary Fig. 8F), suggesting TGFβ as a key downstream mediator of THBS1-driven fibrotic remodeling. The selective preservation of TNF expression despite TGFβ blockade suggests that THBS1 may engage additional TGFβ-independent inflammatory pathways, possibly through CD36 or CD47 signaling, which warrants further investigation.

We next examined the effect of THBS1 in macrophages in different polarized states. In pre-polarized M0 state of THP-1 cells, THBS1 dose-dependently increased the proportion of CD11c⁺SLC11A1⁺ pro-inflammatory macrophages (Supplementary Fig. 9A). In M1-polarized macrophages by LPS and IFNγ, baseline expression of pro-inflammatory cytokines (TNF, IL6, IL1B, and CXCL10) was already saturated by the polarizing stimulus, whereas SLC11A1 and TIMP1 were further elevated by THBS1 treatment (Supplementary Fig. 9B). In M2-polarized macrophages by IL-4 and IL-13, THBS1 treatment did not significantly alter canonical M2 identity markers (MRC1, CD163, IL10) but induced a fibrotic transcriptional program (TGFB1, TGFB2, PDGFB, TIMP1, SPP1, and LGALS3) (Supplementary Fig. 9C). These results indicate that THBS1 not only triggers TGFβ release from macrophages but also reprograms them toward a profibrotic state in a polarization-state-dependent manner.

THBS1 as a biomarker for adipose tissue dysfunction and metabolic disease

To investigate the physiological relevance of THBS1, we analyzed tissue-specific expression profiles using publicly available data from the Genotype-Tissue Expression (GTEx) database²⁸. Among various human tissues, THBS1 showed particularly high expression in visceral adipose tissue (Fig. 7A). Paired analysis of matched VAT and SAT samples further confirmed that THBS1 expression was significantly higher in VAT than SAT in both males (n = 233, P = 1.6 × 10⁻²⁹) and females (n = 96, P = 3.6 × 10⁻¹⁶) (Supplementary Fig. 10A). Furthermore, an expression-based phenome-wide association study (ePheWAS) using data from BXD mice revealed significant associations between Thbs1 expression in adipose tissue and obesity-related traits, such as fat mass and body fat proportion (Supplementary Fig. 10B). We assessed the disease relevance of ASC subtypes by integrating UK Biobank GWAS with our snRNA-seq data. ASC.THBS1 was significantly associated with four metabolic traits, including BMI, total cholesterol, LDL cholesterol, and hypertension (FDR <0.05), and showed a consistent trend for HbA1c (Fig. 7B).

Fig. 7. THBS1 as a biomarker for adipose tissue dysfunction and metabolic disease.

A Tissue-specific THBS1 expression from GTEx v8 (n = 17,382 samples, 948 donors, 54 tissues); violin plots show full log2(TPM + 1) distribution as provided by GTEx v8. B Bar plot of significant associations between ASC clusters and GWAS risk scores (BMI, cholesterol, HbA1c, hypertension, LDL) computed by scDRS; one-sided Monte Carlo association P value after Benjamini–Hochberg correction across all tested trait–cluster pairs. C Plasma THBS1 in lean (n = 9), Ob (n = 7), and Ob-DM (n = 9); mean ± SEM; one-way ANOVA with Tukey’s test (lean vs Ob P = 0.004, lean vs Ob-DM P = 0.0001). D Two-sided Spearman correlation of plasma THBS1 with BMI and HbA1c (n = 13). E Schematic of ex vivo human VAT culture ± αTHBS1 mAb (5 days). F KEGG pathway enrichment of downregulated genes in αTHBS1 mAb-treated VAT (EnrichR, Fisher’s exact test); bubble size indicates gene count, color intensity indicates odds ratio (P < 0.05). G Schematic of in vivo THBS1 overexpression in mouse eWAT via jetPEI-mediated injection of pcDNA3-EGFP or pcDNA3-THBS1-His. H Picrosirius Red staining of eWAT at 2 and 4 weeks; FAT score quantification; mean ± SEM (n = 5; paired two-tailed t-test). Scale bar, 100 μm. I, J Immunofluorescence of eWAT at 4 weeks: reporter (green), COL6A1 (I, red) or SLC11A1 (J, red), DAPI (blue). Representative of n = 5 animals. Scale bar, 100 μm. K Schematic of bilateral αTHBS1 mAb injection into iWAT of HFD mice (14–18/20 weeks). L Picrosirius Red staining and FAT score of iWAT; mean ± SEM (n = 7 at 4 weeks, n = 6 at 6 weeks; paired two-tailed t-test). Scale bar, 100 μm. M Heatmap of YAP/TAZ (WWTR1, THBS1, MYC, SOX6, AXL, TEAD1) and adipogenic (PPARG, CD36, FABP4, LPL) gene expression across ASC populations. N THBS1 and WWTR1 immunocytochemistry in lean and Ob-DM adipocytes; WWTR1 fluorescence intensity; mean ± SEM (n = 18 lean, 17 Ob-DM adipocytes, 4 donors each; unpaired two-tailed t-test, P = 0.040). Scale bar, 50 μm. O ChIP schematic with THBS1 promoter coordinates (chr15:39,579,678-39,580,199). P ChIP-qPCR for WWTR1 and TEAD1; mean ± SEM (n = 3; paired two-tailed t-test vs IgG). Q Thbs1, Acta2, and Pparg expression after GA-017 ± verteporfin treatment; mean ± SEM; one-way ANOVA with Šidák’s test (n = 6/5/5, 3/4/4, 4/3/3 for naïve/GA-017/GA-017 + verteporfin). *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Panel (E) was created in BioRender. Jung, J. (2026) https://BioRender.com/ykzfdnu. Panel (G) was created in BioRender. Jung, J. (2026) https://BioRender.com/5k21kx6. Panel (K) was created in BioRender. Jung, J. (2026) https://BioRender.com/nz72alo. Panel (O) was created in BioRender. Jung, J. (2026) https://BioRender.com/y6869v3.

To assess the clinical utility of THBS1 as a biomarker, we measured its circulating protein levels in an independent cohort. Plasma THBS1 levels increased from Lean to Ob to Ob-DM groups (Fig. 7C). Furthermore, circulating THBS1 levels significantly correlated with metabolic traits such as BMI and HbA1c levels (Fig. 7D). These results support THBS1 as a mechanistically informative and clinically relevant biomarker linking adipose tissue dysfunction to systemic metabolic dysregulation.

THBS1 inhibition ameliorates fibro-inflammatory remodeling in metabolically dysfunctional adipose tissue

To evaluate the therapeutic potential of targeting THBS1, we examined transcriptomic changes in adipose tissues following THBS1 modulation. In mice, ex vivo culture of visceral adipose tissue with LSKL, a peptide inhibitor of THBS1^29,30, for four days reversed the THBS1-induced upregulation of key fibrosis markers, including Col1a1, Col6a3, and Acta2 (Supplementary Fig. 10C, D). To further evaluate the implication of THBS1 inhibition in insulin sensitivity, we cultured mouse epididymal white adipose tissue (WAT) explants with anti-THBS1 neutralizing antibody (αTHBS1) or LSKL peptide for 7 days. Both αTHBS1 and LSKL treatments significantly increased the phospho-AKT/total AKT ratio and adiponectin compared to vehicle, while total GLUT4 levels were unchanged (Supplementary Fig. 10E, F).

To assess the translational relevance of THBS1 inhibition in humans, we treated visceral adipose tissue samples from individuals with Ob-DM with a neutralizing anti-THBS1 monoclonal antibody (Fig. 7E). Following 5 days of ex vivo culture, we performed bulk RNA-seq and identified transcriptional changes indicative of attenuated fibro-inflammation. Specifically, THBS1 inhibition significantly reduced the expression of genes involved in fibrosis and ECM remodeling, including FN1, COL5A3, ICAM2, and ADAMTS5. Additionally, the expressions of TGFBI and TGFB2 were downregulated. Pro-inflammatory cytokines such as TNF and CCL8 were also decreased upon THBS1 inhibition (Supplementary Fig. 10G). KEGG pathway enrichment analysis of downregulated DEGs revealed suppression of key fibro-inflammatory pathways, including cell adhesion, ECM-receptor interaction, and TGFβ signaling (Fig. 7F). These findings suggest that THBS1 may serve as a therapeutic target for ameliorating fibro-inflammatory changes in adipose tissue.

In vivo gain- and loss-of-function models establish THBS1 as a local driver of adipose fibro-inflammation

To directly test the causal contribution of THBS1 to adipose fibro-inflammation in vivo, we performed complementary gain-of-function and antibody-mediated neutralization experiments. In the gain-of-function model, a pcDNA3-THBS1-His expression vector was delivered into the right epididymal fat pad (THBS1 O/E), while a pcDNA3-eGFP control vector was delivered into the left fat pad (GFP O/E) of the same animal, enabling paired intra-animal comparison while controlling for systemic effects (Fig. 7G). Picrosirius Red staining at 2 and 4 weeks post-injection demonstrated significantly increased collagen deposition in THBS1-overexpressing epididymal WAT compared to GFP controls, with progressive accumulation over time (Fig. 7H, n = 5 mice per timepoint). Dual immunofluorescence further revealed dense peri-adipocyte COL6A1 deposition (Fig. 7I and Supplementary Fig. 10H) and accumulation of pro-inflammatory SLC11A1⁺ macrophages (Fig. 7J) surrounding transgene-expressing adipocytes, while contralateral GFP-expressing epididymal WAT remained largely COL6A1-negative and showed minimal SLC11A1⁺ macrophage infiltration. These results demonstrate that local THBS1 overexpression is sufficient to induce fibrotic ECM deposition and recruit a pro-inflammatory macrophage phenotype in adipose tissue in vivo.

For the loss-of-function model, an anti-THBS1 neutralizing monoclonal antibody (αTHBS1 mAb) was delivered by bilateral intra-inguinal WAT injection in HFD-fed C57BL/6J mice (10 µg/injection, 3×/week), with normal IgG injected into the contralateral (IgG) inguinal WAT pad of the same animal (Fig. 7K). At both the 4-week (n = 7) and 6-week (n = 6) timepoints, αTHBS1 mAb-treated inguinal WAT pads showed significantly reduced FAT scores compared to contralateral IgG-injected controls (Fig. 7L), demonstrating that local THBS1 neutralization ameliorates fibrotic remodeling in adipose tissue.

Hippo signaling regulates transcriptional activity of THBS1 in adipocytes

To identify upstream regulators that control THBS1 expression and potentially modulate its fibro-inflammatory function, we applied decoupleR³¹ to infer transcription factor (TF) activity in ASCs. Compared with ASC.PPARG cells, YAP/TAZ and inflammatory TFs such as RELA and NFKB were upregulated in ASC.THBS1, whereas adipogenic regulators including SREBF1 and MECOM were downregulated (Supplementary Fig. 10I). Consistent with this, YAP/TAZ target genes were highly expressed in ASC.THBS1 cells (Fig. 7M). In primary adipocytes, THBS1 and WWTR1 were co-localized, and WWTR1 nuclear levels were elevated in adipocytes from Ob-DM individuals compared to Lean adipocytes (Fig. 7N). Chromatin immunoprecipitation followed by PCR (ChIP-PCR) confirmed direct binding of both WWTR1 and TEAD1 to the THBS1 promoter (Fig. 7O, P).

To validate the functional relevance of YAP/TAZ, we treated primary mouse adipocytes with GA-017, a LATS1/2 inhibitor that activates YAP/TAZ, and with verteporfin, a TEAD1 inhibitor. GA-017 treatment increased the expression of Thbs1 and Acta2, whereas co-treatment with verteporfin reversed these effects and restored Pparg expression (Fig. 7Q). These results indicate that Hippo signaling is involved in the transcriptional regulation of THBS1 in adipocytes and may contribute to fibro-inflammation and adipose tissue dysfunction.

Discussion

Despite technical challenges inherent in analyzing adipose tissue, previous single-cell studies have significantly advanced our understanding of adipose biology. While these efforts have characterized the distribution and composition of diverse adipose cell populations, key questions remain regarding the primary regulators of fibro-inflammation and their functional validation in human adipose tissue. In this study, we established an ethnically homogeneous human cohort with well-defined metabolic disease states: lean individuals with normal glucose tolerance, individuals with obesity, and those with obesity and type 2 diabetes. Among various fat depots, we focused on visceral adipose tissue, given its strong link to systemic metabolic dysregulation. Consistent with this, paired GTEx analysis confirmed that THBS1 is expressed at significantly higher levels in VAT than in SAT in both sexes (Supplementary Fig. 10A), supporting the notion that THBS1-driven fibro-inflammation is preferentially active in VAT. Whether this axis operates equivalently in SAT warrants further cross-depot mechanistic studies.

Through comprehensive functional validation, including human primary cell culture, ex vivo culture of adipose tissue explants, and Matrigel-embedded adipocyte assays, we uncovered key regulatory mechanisms underlying visceral adipose tissue dysfunction. We identified fibro-inflammatory subpopulations of ASCs and adipocytes characterized by THBS1 expression, which were enriched from lean to obesity to individuals with obesity and diabetes. These THBS1⁺ ASCs and adipocytes share phenotypic features with previously described fibrotic and anti-adipogenic progenitors, including SDC1⁺/COL8A1⁺ fibrotic ASCs described in our previous study¹⁹, CD9⁺ and CD55^low ASCs¹⁸, anti-adipogenic progenitors¹⁰, mesothelial-like stem cells¹⁴, and stressed and profibrotic progenitor, APC3 and APC5²⁰, further supporting their identity as key drivers of adipose tissue fibro-inflammation. Notably, hyperglycemic and hyperinsulinemic stimulation of primary adipocytes robustly induced Thbs1 expression (Supplementary Fig. 5E), supporting the possibility that Adipo.THBS1 may represent a metabolically-induced activation state rather than a stable lineage-defined subtype. However, distinguishing an inducible activation state from a distinct lineage will require further investigation.

We propose that THBS1 is an important contributor of adipose tissue fibrosis, inflammation, and metabolic dysfunction. THBS1 is a secreted glycoprotein involved in ECM remodeling and blood clotting and interacts with multiple receptors and ligands. In particular, THBS1 activates latent TGFβ via competitive interaction with latent TGFβ binding protein (LTBP), promoting fibrotic responses^25–27. It also binds ECM components such as β1 integrin and engages CD47 through its C-terminal domain to regulate ECM remodeling and inflammation³². Several previous studies have implicated THBS1 in obesity and metabolic disease, including increased adipose THBS1 expression in individuals with obesity³³, protection from HFD-induced insulin resistance in Thbs1-deficient mice³⁴, adipose macrophage infiltration via THBS1³⁵, and profibrotic THBS1⁺ fibro-adipogenic progenitors identified by scRNA-seq³⁶. Our study extends these findings by providing single-nucleus profiling of THBS1⁺ subpopulations in VAT across human metabolic disease states and mechanistic insights into the THBS1–TGFβ axis via THBS1⁺ ASC and adipocytes–macrophage intercellular interactions. Supporting this role, our reciprocal in vivo gain-of-function and antibody-mediated neutralization experiments demonstrated that local THBS1 perturbation modulates fibro-inflammatory remodeling in mouse VAT. Future studies employing adipocyte- or ASC-specific conditional Thbs1 knockout models will be helpful for dissecting the cell-autonomous contributions of THBS1 in vivo.

We validated THBS1 as a clinically relevant biomarker and potential therapeutic target through in vitro and ex vivo functional experiments. Using adipose tissue explants and primary adipocyte cultures from over 20 individuals, we showed that THBS1 promotes fibrosis and impairs adipogenesis. Importantly, inhibition of THBS1 using either an anti-THBS1 monoclonal antibody or the LSKL peptide consistently attenuated fibro-inflammatory changes in adipose tissue from individuals with obesity and diabetes. In an independent cohort of over 60 individuals, THBS1 levels in both adipocytes and plasma increased from Lean to Ob to Ob-DM, correlating positively with BMI and HbA1c. These findings indicate promising associations of plasma THBS1 with metabolic parameters, although establishing its clinical utility as a diagnostic biomarker will require validation in larger, independent cohorts with formal diagnostic performance assessment, including ROC analyses.

Understanding the upstream regulatory mechanisms of THBS1 is essential for therapeutic development. We propose a model wherein adipocyte hypertrophy, a hallmark of overnutrition and metabolic stress, leads to inactivation of the Hippo signaling pathway, with reduced MST1/LATS1-mediated phosphorylation of YAP/TAZ permitting their nuclear translocation and TEAD1-mediated transcriptional activation of THBS1. YAP/TAZ are known to regulate organ size, mechanosensing, and tissue homeostasis³⁷. While YAP/TAZ have been implicated in adipogenesis, adipocyte cell death, and adipose plasticity, their role in human adipose dysfunction remains unclear^38,39. Wang et al. demonstrated that YAP/TAZ and their target genes are activated in adipocytes during obesity in both mice and humans³⁹; Lee et al. further showed that adipocyte hypertrophy mechanistically drives YAP/TAZ nuclear translocation via increased cell stiffness and F-actin rearrangement⁴⁰; and Shen et al. identified the Hippo–YAP/TAZ–TEAD axis as a key molecular switch in obesity-induced adipose tissue fibrosis⁴¹. Our integrative approach demonstrated that YAP/TAZ-TEAD1 directly binds to the THBS1 promoter and regulates its transcription. While we present adipocyte hypertrophy as a putative upstream contributor rather than a proven causal trigger in our system, alternative upstream signals, including ECM stiffening and fibrotic remodeling, both known modulators of Hippo signaling, may also activate this axis. These results suggest that Hippo pathway inhibition could serve as an upstream strategy to mitigate THBS1-driven adipose pathology. To further translate these findings, we are actively screening compound libraries, including natural products, for additional inhibitors targeting THBS1 and the YAP/TAZ-TEAD1 axis.

Additionally, we identified two macrophage subtypes with opposing distributions across metabolic states: proinflammatory SLC11A1⁺ macrophages^42–44, which expanded from Lean to Ob to Ob-DM, and LYVE1⁺ macrophages, which declined across the same spectrum. While SLC11A1 regulates intracellular iron homeostasis and has several pleiotropic effects on macrophage activation⁴⁵, its role in metabolic diseases has been underexplored. The functional role of SLC11A1 in macrophages, especially in fibro-inflammatory remodeling, warrants further mechanistic investigation. Our findings align with the expansion of hMac2 and hMac3 subsets observed in ref. ⁹, which harbor similar proinflammatory signatures. We validated the enrichment of SLC11A1⁺ macrophages in obesity-associated diabetes and found their strong association with M1 polarization markers (CD11c and CD80), proinflammatory cytokines, and crown-like structures, implicating them in visceral adipose tissue inflammation. In contrast, LYVE1⁺ macrophages, which are associated with vascular homeostasis and anti-inflammatory functions^46,47, were correspondingly depleted, suggesting a loss of tissue-protective immunity across metabolic disease states.

This study has several limitations. First, we employed snRNA-seq rather than scRNA-seq as the primary single-cell profiling strategy. This approach enabled the capture of fragile adipocytes and simultaneous analysis of multiple human samples, but it may underrepresent cytoplasmic or low-abundance transcripts, including inflammatory cytokine transcripts. To address this limitation, we performed bulk RNA-seq on matched tissue samples, which validated the snRNA-seq results and additionally identified key inflammatory genes, including IL1B, IL6, and TGFB1, that were not detected in the snRNA-seq data. Second, our study cohort consisted exclusively of Korean individuals with a relatively homogeneous genetic background. While this homogeneity helps reduce inter-individual variability, it may limit the generalizability of our findings to other ethnic groups. To partially overcome this, we projected our cell type-specific findings onto publicly available single-cell datasets, including that of Emont et al.⁹, and observed consistent patterns across populations, supporting the broader applicability of our results. Third, although our cohort size is comparable to those of previously published human adipose snRNA-seq/scRNA-seq studies and was designed to provide adequate statistical power for condition-based comparisons, the overall sample size remains relatively small and insufficient for robust sex-stratified analyses. Although both female and male donors were included (13 females and 8 males), the cohort yields only two to five individuals per stratum, precluding reliable sex-specific inference. We are therefore currently planning to evaluate the generalizability of THBS1 as a biomarker in larger cohorts balanced across sex and metabolic disease status. Fourth, circulating THBS1 can originate from multiple tissues, the relative contribution and clinical significance of adipose-derived versus systemic THBS1 will require further investigation in future studies. Fifth, for multiplexed snRNA-seq channels, donors within each channel belonged to the same metabolic condition, preventing complete separation of pool-specific technical effects from condition-associated biological effects. Although Souporcell-based genotype demultiplexing enabled donor-level cell assignment, it does not remove potential technical effects shared among donors processed in the same channel. Therefore, residual pool- or channel-specific effects may influence the interpretation of inter-individual transcriptional variation.

This study provides a high-resolution single-nucleus transcriptomic atlas of human visceral adipose tissue across distinct metabolic states. Through the integration of snRNA-seq, bulk RNA-seq, and in vitro functional assays, we identify THBS1 as a candidate mediator associated with adipose tissue fibrosis, inflammation, and metabolic dysfunction. Our findings uncover two key mechanisms contributing to adipose tissue dysfunction: THBS1-mediated intercellular communication and YAP/TAZ-TEAD1-dependent transcriptional regulation of THBS1 expression. Together, these results offer mechanistic insights into adipose tissue remodeling in obesity-associated metabolic disease and highlight THBS1 as a promising biomarker and therapeutic target. This work lays the foundation for the development of new interventions aimed at mitigating adipose tissue inflammation and fibrosis in the treatment of metabolic disorders.

Methods

Human adipose tissue collection

Visceral adipose tissues were collected under the approval of the Seoul National University Bundang Hospital Institutional Review Board (B-1203-147-006 and B-1801-445-301). Written informed consent was obtained from all participants. Visceral adipose tissues were collected from patients undergoing abdominal surgery: living kidney donors (Lean group, BMI <23, n = 8) and individuals undergoing laparoscopic sleeve gastrectomy (Obesity and Obesity with diabetes groups, BMI ≥30, n = 13). The obtained tissue samples were processed into stromal vascular fractions (SVF) and adipocyte fractions or frozen in liquid nitrogen and stored at −80 °C for future use. For histological analysis, small tissue sections were fixed in 4% paraformaldehyde for 24 h, followed by standard tissue processing protocols to construct formalin-fixed paraffin-embedded blocks. These blocks were then sectioned into 5-µm-thick slices, which were subsequently used for histological and immunohistochemical analyses. Peripheral blood plasma was separated via centrifugation at 3000 rpm for 15 min at room temperature (RT), and separated plasma samples were then collected and stored at – 80 °C.

Animal studies

All animal experimental procedures were conducted in accordance with protocols approved by the Institutional Animal Care and Use Committee of Seoul National University Bundang Hospital (IACUC No. BA-2406-393-001-02 and BA-2601-431-005). Male C57BL/6J mice were purchased from Jackson Laboratory (Bar Harbor, ME, USA) and maintained under specific pathogen-free conditions on a 12-h light/dark cycle at 23 ± 1 °C and 60% relative humidity, with food and water provided ad libitum. For diet-induced obesity experiments, mice were placed on a 60% high-fat diet (HFD; D12492, Research Diets, Inc., New Brunswick, NJ, USA) starting at 6 weeks of age and maintained on HFD until sacrifice. For experiments using lean controls or ex vivo/adipocyte-captured culture, mice were maintained on a standard chow diet (Purina Laboratory Rodent Chow, cat. no. 10000102251; Cargill Agri Purina, Seongnam, Korea) provided ad libitum by the facility. Mice were euthanized by CO₂ inhalation at the indicated age, and adipose tissues were collected for downstream analyses.

Anti-THBS1 neutralizing antibody injection

To assess the in vivo requirement of THBS1 in obesity-induced adipose fibro-inflammation, HFD-fed male C57BL/6J mice were subjected to bilateral intra-iWAT antibody injection. After 8 weeks of HFD feeding (at 14 weeks of age), αTHBS1 neutralizing monoclonal antibody (10 µg/injection in 50 µL PBS, Genetex GTX21823) was injected into one inguinal white adipose tissue (iWAT) pad, while an equal amount of normal IgG (50 µL PBS, Cell Signaling Technology, 68860L) was injected into the contralateral pad of the same animal, three times per week. HFD feeding and bilateral antibody injections were maintained concurrently, and mice were sacrificed after either 4 or 6 weeks of injection (at 18 and 20 weeks of age, respectively) for histological and molecular analyses.

In vivo transfection of epididymal white adipose tissue (eWAT)

For in vivo gene delivery into eWAT, all surgical procedures were performed under aseptic conditions. The pcDNA3 mammalian expression vector (originally from Invitrogen, now Thermo Fisher Scientific, Waltham, MA, USA) was used as the backbone for both control and overexpression constructs. The full-length human THBS1 coding sequence with a C-terminal 6×His tag was synthesized and cloned into pcDNA3 to generate pcDNA3-THBS1-His. The EGFP coding sequence was sub-cloned from pEGFP-C1 into pcDNA3 to generate the pcDNA3-EGFP control construct. Custom gene synthesis, cloning, and large-scale plasmid preparation were performed by Cosmo Genetech (Seoul, Republic of Korea). Plasmid DNA was further processed for endotoxin removal using EndoZero II Spin Columns (C1060-25, Zymo Research, Irvine, CA, USA) according to the manufacturer’s instructions and resuspended in nuclease-free water at a final concentration of 1 µg/µL prior to use.

Twelve-week-old male C57BL/6J mice maintained on a normal chow diet were anesthetized by inhalation of 1–2% isoflurane. The abdominal fur was removed with an electric clipper, and the surgical site was disinfected with povidone-iodine (Betadine). A small midline incision was made through the skin and peritoneum using iris scissors, and bilateral eWAT pads were gently exteriorized onto sterile saline-soaked gauze. Plasmid DNA was delivered using in vivo-jetPEI (Polyplus Transfection, Illkirch, France) according to the manufacturer’s instructions. DNA–jetPEI complexes were prepared at an N/P ratio of 7 in 5% glucose and incubated at room temperature for 15 min before injection. A total of 10 µg of plasmid DNA in 50 µL was injected into each fat pad, distributed across three to four sites per pad using a 1 mL syringe fitted with a 25-gauge needle to accommodate the limited fat pad volume. After injection, eWAT pads were carefully repositioned into the peritoneal cavity, and the peritoneum and skin were closed separately with three to five simple interrupted sutures per layer. The suture site was disinfected with povidone-iodine. Gentamicin (5 mg/kg, subcutaneous) was administered immediately after suture closure to prevent post-operative infection, and carprofen (5 mg/kg, subcutaneous) was administered for post-operative analgesia. Mice were placed under an infrared warming lamp until full recovery from anesthesia before being returned to their home cages and were sacrificed at 2 or 4 weeks post-surgery for downstream histological and molecular analyses.

Sample preparation and nuclei isolation for snRNA-seq

Snap-frozen visceral adipose tissues were obtained and immediately stored at -80 °C until use. The tissues were thawed in warm media (Dulbecco’s modified eagle medium [DMEM]; 11965-092; Gibco, MA, USA; 10% FBS; 16000-044; Gibco) at 37 °C for 10 min and then washed twice with phosphate-buffered saline (PBS; 10010-031; Thermo Fisher Scientific, MA, USA) to remove debris and residual red blood cells. The tissues were then minced using a sterile razor blade (1000548452179; Dorco, Seoul, Korea) into small pieces of ~2–4 mm in size. Nuclei were extracted as previously described⁴⁸, with minor modifications. Minced tissues were incubated in TST buffer (10 mM Tris-HCl, pH 8.0, 0.25% SDS, 1 mM EDTA, 150 mM NaCl, and 1x protease inhibitor cocktail; 78430; Thermo Fisher Scientific) for 20 min on ice with gentle shaking. The lysate was filtered through a 100 µm cell strainer (732-2759; VWR, PA, USA) to remove large debris. Subsequently, the filtrate was passed through sequential strainers with pore sizes of 70 µm (93070; SPL Life Sciences, Gyeonggi-do, Korea) and 10 µm (43-50010-50; pluriSelect, Leipzig, Germany) to isolate nuclei. The resulting nuclear suspension was centrifuged at 500×g for 5 min at 4 °C to pellet the nuclei. For single-nucleus RNA-sequencing, nuclei isolated from three individual donors were pooled in equal proportions (~10,000 nuclei per donor) and loaded into a single 10X Genomics Chromium channel for GEM formation.

snRNA-seq library preparation and sequencing

The isolated nuclei were resuspended in nuclei buffer (PN-1000268; 10x Genomics, CA, USA). The concentration of the nuclei was determined using a hemocytometer (HSU-0650030; Paul Marienfeld, Germany) and was adjusted to an appropriate concentration for further processing. Immediately after isolation, the nuclei were processed using the 3’ Single Cell Gene Expression Kit (v3.1; 10x Genomics), following the manufacturer’s instructions. Sequencing of snRNA-seq libraries was performed by Genome Insight Inc. (Korea).

Bioinformatic analysis

snRNA-seq data preprocessing

Raw reads from snRNA-seq data were mapped to the human reference genome GRCh38 using CellRanger (v7.0.1)⁴⁹. A total of 11 10x Genomics channels were processed, comprising 21 unique donors. Nuclei were processed either individually or as multiplexed samples. Four channels contained single donors, whereas the remaining multiplexed channels contained nuclei pooled from up to three donors belonging to the same metabolic condition. The complete donor composition for each channel is provided in Supplementary Fig. 1A. For the Ob group, Ob1 and Ob2 were processed in both AD6 and AD8 because AD8 was generated as a technical re-processing replicate of the initially low-yield AD6 pool. These repeated measurements were merged by biological donor identity after demultiplexing and were not counted as independent biological donors. A gene-by-cell count matrix was generated with default parameters. Ambient RNA contamination and low-quality cells were removed using the remove-background function of CellBender (v0.3.2)⁵⁰. Additional cell-specific quality control metrics were calculated using the perCellQCMetrics function in the scuttle (v1.4.0) R package⁵¹. Cells with <2.5 log₁₀-scaled unique molecular identifiers or >10% mitochondrial gene content were excluded.

Cell-specific size factors were computed using the quickCluster and computeSumFactors functions in the scran (v1.22.1) R package⁵². Normalized data were obtained by using the logNormCounts function in the scuttle R package. Highly variable genes (HVGs) were identified using the decomposeVar function in scran, with a false discovery rate <0.05 and biological variability >0.1. Principal component analysis was performed using the RunPCA function in Seurat on HVGs expression data that were scaled using the ScaleData function. Batch effects across patients were corrected using the RunHarmony function in the harmony (v0.1.1) R package⁵³.

For cell clustering, the FindClusters function in the Seurat (v4.3.0) R package⁵⁴ was applied using the first 15 principal components at a resolution of 0.8. The UMAP visualization was generated with the first 20 principal components using the RunUMAP function. Putative doublets were identified using the scDblFinder function in the scDblFinder (v1.8.0) R package⁵⁵ with the sample parameter. Doublet predictions were validated through manual inspection of clusters exhibiting co-expression of lineage-specific markers from different cell types and elevated UMI counts. For multiplexed snRNA-seq channels, donor identities were assigned using Souporcell (v2.5)⁵⁶, with the expected number of genotype clusters set according to the number of donors included in each channel. To assign Souporcell genotype clusters to known donors, we performed genotype matching with donor-specific whole-exome sequencing (WES) data. WES reads were aligned using BWA-MEM2 (v2.2.1), followed by duplicate marking, base quality recalibration using ExAC r1.0 known sites, and variant calling using GATK (v4.4.0.0). Variants were filtered to retain high-quality exonic SNPs using SelectVariants for SNP-only filtering, QD ≥10, and BEDTools (v2.30.0) for exonic region filtering. Genotype concordance between Souporcell clusters and WES samples was calculated across common variants using vcfR (v1.14.0). Clusters were assigned to donors based on maximum genotype similarity, using a threshold of ≥0.7 similarity and ≥100 common variants. Ambiguous nuclei, including predicted doublets, unassigned nuclei, and low-confidence genotype matches, were excluded. For Ob1 and Ob2, which were processed in both AD6 and AD8 as technical re-processing replicates, nuclei were merged by biological donor identity, and each donor was counted once in downstream donor-level analyses.

Transcriptional variation analysis using scITD

To capture inter-individual transcriptional variation in a cell type-aware manner, we applied single-cell interpretable tensor decomposition (scITD). This method decomposes gene expression variation across donors and cell types into coordinated latent transcriptional programs, referred to as “Programs” in the original scITD framework. These programs are conceptually analogous to components in exploratory factor analysis: they represent axes of co-variation across the dataset. Each program is characterized by: (1) donor scores indicating how strongly each individual exhibits the corresponding pattern and (2) cell type-specific gene loadings indicating the genes contributing to that pattern within each cell type. We performed two separate scITD analyses. First, we performed a tissue-wide analysis including all major cell types, including adipocytes, ASCs, endothelial cells, myeloid cells, and T/NK cells, to capture global patterns of inter-individual variation across the adipose tissue compartment (Fig. 1G). Second, we performed an immune-focused analysis restricted to immune cell populations to resolve immune-specific transcriptional variation at higher cellular resolution (Fig. 2D). Because these analyses were performed independently using different input cell populations, the program numbers are analysis-specific and should not be interpreted as corresponding directly between Figs. 1G and 2D.

Condition-stratified clustering for adipocytes

Adipocyte populations showed strong condition-dependent transcriptional variation, particularly for THBS1, which was minimally expressed in Lean samples but markedly upregulated in Ob-DM samples. To evaluate whether disease-associated adipocyte states could be recovered reproducibly while minimizing overcorrection of condition-associated biological signals, we performed condition-stratified adipocyte clustering as a sensitivity analysis. Lean and obesity-associated (Ob/Ob-DM) samples were clustered separately using the standard Seurat workflow, including FindNeighbors and FindClusters at a resolution of 0.8. Cluster annotations were assigned based on marker gene expression and compared across conditions. For integrated visualization and cross-condition analyses, Harmony batch correction was applied to the combined adipocyte dataset. Clustering robustness was further assessed by repeating clustering across resolution parameters from 0.7 to 0.9 and quantifying cluster purity, defined as the proportion of cells assigned to the dominant reference annotation within each cluster.

Cell–cell interaction

To analyze intercellular communication patterns and identify ligand–receptor interactions between cell types, we used the CellChat (v1.0.0) R package²². The analysis identified significantly overexpressed ligand–receptor genes within each cell type compared to other cell types and computed communication probabilities between different cell types based on these ligand–receptor pairs. Signaling pathway information flow was compared between the Ob and Ob-DM groups using the rankNet function. Interpatient heterogeneity was addressed using multinichenetr (v2.0.1) R package²³, with particular focus on adipocytes and ASCs as sender cells, using the following parameters: min_cells = 5, min_sample_prop = 0.25, fraction_cutoff = 0.25, and logfoldchange_threshold = 0.30.

Trajectory analysis

We performed trajectory analysis on ASC using the Palantir (v1.0.0) Python package⁵⁷. First, we identified marker genes for each ASC subtype using the FindAllMarkers function in Seurat and selected genes with log2 fold change >0.25 and adjusted p value <0.05 to create a subtype-specific gene expression matrix. For ASC trajectory visualization, we applied principal component analysis to this matrix and corrected batch effects across individuals using HarmonyPy (v0.0.4)⁵³ on the first 400 principal components (PCs). We then constructed a k-nearest neighbor graph (k = 30) using 20 diffusion components derived from the batch-corrected PCs, and computed t-distributed stochastic neighbor embedding plots (perplexity = 500). Trajectory analysis was performed using the run_palantir function with 500 waypoints, designating PI16⁺ cells as starting points for adipose lineage trajectory.

Differential expression and gene ontology analysis

For macrophage subtype characterization, we identified DEGs between Mac.SLC11A1 and Mac.LYVE1 using the FindMarkers function in Seurat, with MAST test and patient correction. Significant DEGs were those with adjusted P value <0.05 and log₂ fold change >0.25. Gene ontology enrichment analysis was performed using the runTest function in the topGO (v2.38.1) R package⁵⁸.

Transcription factor inference

To infer transcription factor activities in adipose stem cells, we applied the run_ulm function in deCoupleR (v2.9.1) R package³¹. ASC.THBS1-specific upregulated transcription factors were identified using FindMarkers with MAST test, patient correction, a log fold-change threshold of 0, and adjusted P values <0.05.

Spatial neighbor analysis

We investigated the spatial relationships between THBS1+ spots and cells expressing various markers (SLC11A1, LYVE1, COL1A1, CD47, ITGB1, ACTA2, LRP5, and FN1) using 10x Visium spatial transcriptomics in two tissue sections (sample numbers: S44 and S49)²⁴. We categorized spots into four groups based on gene expression: spots with no expression (“NO”), THBS1-only, marker-only, or co-expression (“BOTH”), and visualized their spatial distributions. We then implemented a k-nearest neighbor (k = 20) approach using the FNN (v1.1.3.2) R package⁵⁹ to quantify each spot’s neighborhood composition, specifically by counting the marker-positive spots within the 20 nearest neighbors. Differences in neighborhood marker-positive spot counts between THBS1⁺ and THBS1^- spots were assessed using Wilcoxon rank-sum tests to determine whether THBS1⁺ cells exhibit distinct spatial relationships with other cell populations in the tissue microenvironment.

Transcriptional variation analysis

To investigate coordinated multicellular patterns of transcriptional variation across multiple cell types and donors, we performed tensor decomposition using the scITD (v1.0.4) R package⁶⁰. For both main adipose tissue and immune cell analyses, we required a minimum of three cells per donor for each cell type to ensure robust pseudobulk expression profiles. Pseudobulk expression matrices were generated by summing UMI counts across cells for each donor-cell type combination.

Data were normalized using the trimmed mean of M-values (TMM) method, then normalized by dividing by adjusted library size, multiplying by 10000, and log-transforming with a pseudocount of 1. Variable genes were selected based on normalized variance calculations, with the top overdispersed genes retained (threshold p-value < 0.1 for adipose analysis; top 1,000 genes for immune analysis). Expression values were centered and unit-scaled across donors for each gene within cell types, then rescaled by normalized variance (scale_var =TRUE) to allow genes with greater biological variation to contribute proportionally to decomposition.

Tucker decomposition with hybrid rotation was applied using the run_tucker_ica function to extract donor factors (5 for adipose and 6 for immune) and gene factors (10 for adipose and 12 for immune). Metadata associations between donor scores and clinical variables (condition, sex) were tested using the get_meta_associations function with linear regression F-tests for continuous variables or logistic regression with χ² tests for categorical variables, with FDR correction applied.

Single-cell disease relevance score analysis

To investigate the association between adipose stem cell populations and various disease/metabolic traits, we used the scDRS (v1.0.3) Python package⁶¹ to integrate the GWAS and single-cell RNA-seq data. We applied the compute_score function to our quality-controlled ASC data in h5ad format using UK Biobank GWAS summary statistics⁶² processed using MAGMA (10 kb window) with the top 1000 genes ranked by MAGMA z-scores and their corresponding z-scores. We specified the parameters for human species alignment, raw count data processing, and normalized control score generation. For downstream analysis, we used the perform-downstream function with group analysis parameters set to condition and cell type to examine disease associations across different ASC subpopulations.

Bulk RNA-seq preprocessing and analysis

For the read sequences resulting from total RNA-seq, adapter sequences (Nextera adapters) and low-quality reads were removed using Trimmomatic software (v0.39)⁶³. The resulting reads were aligned to the human reference genome GRCh38 using the HISAT2 aligner (v2.2.1)⁶⁴, and PCR or optical duplicate reads were filtered out using Picard (v2.18.29)⁶⁵. The remaining reads were counted for gene features (GTF file of GRCh38) using HTSeq (v0.11.2)⁶⁶.

To identify DEGs across Lean, Obesity, and Obesity with diabetes (Ob-DM) groups, we used raw FPKM expression values without additional log-transformation or normalization. Genes with FPKM >1 in more than half of the samples (n > 10 out of 20 total) and expressed in at least one sample per group were retained for downstream analyses.

For pairwise differential expression analyses between groups (Ob vs. Lean and Ob-DM vs. Lean), we applied the limma package⁶⁷ using a linear model approach. A design matrix was constructed based on the three clinical groups, and empirical Bayes moderation was performed to estimate fold changes and p values. Genes with adjusted P < 0.05 and absolute log2 fold change >0.58 were considered significantly differentially expressed. Volcano plots were generated to visualize the DEG distribution, and selected marker genes were labeled.

To further characterize group-specific gene expression patterns, DEGs from the two pairwise comparisons (Ob vs. Lean and Ob-DM vs. Lean) were classified into six subsets based on directionality and overlap. Upregulated genes were categorized as Ob-DM-specific (g1), common to both Ob and Ob-DM (g2), or Ob-specific (g3), while downregulated genes included Ob-DM-specific (g4), commonly downregulated (g5), and Ob-specific (g6). These subsets were visualized using Venn diagrams, and each group of DEGs (g1–g6) was subjected to pathway enrichment analysis using EnrichR⁶⁸. Enrichment was performed against the KEGG pathway database to identify functional pathways associated with condition-specific transcriptional changes. Significant pathways were defined by p values <0.05 and visualized as dot plots.

Gene set enrichment analysis

To evaluate coordinated changes in pathway activity across conditions, gene set enrichment analysis was performed using the fgsea R package⁶⁹. Gene-wise log2(FPKM + 1)-transformed expression values were averaged for each group (Lean, Obesity, Obesity with diabetes) and converted to ranked statistics. Hallmark gene sets were retrieved from the MSigDB database (version 2024.1) using the msigdbr R package⁷⁰. The fgsea algorithm was applied to each ranked gene list using all hallmark gene sets (H), with parameters set to minSize = 15 and maxSize = 500. Normalized enrichment scores (NES) were extracted, and a selected panel of pathways was visualized using heatmaps. Significant enrichment was defined as an adjusted p value <0.05.

Weighted gene co-expression network analysis (WGCNA)

To identify co-expressed gene modules associated with clinical traits, we applied WGCNA²¹ to the log2-transformed FPKM matrix, retaining only samples with matched clinical metadata. The top 25% most variable genes were selected as input. The soft-thresholding power (β) was determined by evaluating scale-free topology fit (R² > 0.8) across powers ranging from 1 to 20. Modules were identified using hierarchical clustering with dynamic tree cutting (minimum module size = 70), yielding 20 initial modules. Modules were subsequently merged into seven final modules based on eigengene correlation and functional similarity. Modules lacking meaningful enrichment (i.e., fewer than five significantly enriched pathways with adjusted P < 0.05) or corresponding to the gray (unassigned) category were grouped into a unified unassigned module (module 7) and excluded from downstream interpretation. Module–trait associations were evaluated using Pearson correlation with continuous clinical traits, including HbA1c, BMI, LDL, and HOMA-IR. Functionally relevant modules were annotated using KEGG pathway enrichment analysis via EnrichR⁶⁸. While all KEGG terms were tested, only selected pathways with adjusted P < 0.05 and known biological relevance were visualized as bubble plots based on odds ratios.

Protein–protein interaction network and hub gene identification

Genes from module 1 were mapped to STRING v11.5⁷¹ to construct a protein–protein interaction network. First-degree neighbors of THBS1 were extracted using the iGraph R package⁷², and only edges with combined STRING scores greater than 0.9 were retained. Louvain community detection was applied to identify densely connected submodules, and the network was visualized using Fruchterman–Reingold and Kamada–Kawai layouts, with edge thickness representing interaction strength and node colors denoting community clusters.

GTEx paired visceral and subcutaneous adipose tissue analysis

Bulk RNA-seq gene expression data from the Genotype-Tissue Expression project (GTEx Analysis v11) were obtained from the GTEx Portal (https://www.gtexportal.org). To compare THBS1 expression between visceral (Adipose–Visceral (Omentum); VAT) and subcutaneous (Adipose–Subcutaneous; SAT) adipose tissue, we restricted the analysis to donors from whom both depots were sampled, identified by matching subject IDs (SUBJID) across the two tissue annotations. Gene-level transcripts per million (TPM) values were log2-transformed (log2(TPM + 1)) and compared between paired VAT and SAT samples using the paired Wilcoxon signed-rank test, stratified by donor sex. The final paired cohort comprised 233 male and 96 female donors with both VAT and SAT samples available.

Ex vivo adipose tissue culture

Visceral adipose tissues (human visceral adipose tissue and mouse epididymal white adipose tissue) were carefully dissected into ~5 mm pieces under sterile conditions and thoroughly washed with sterile PBS to remove blood and debris. Explants were maintained as free-floating cultures in 24-well plates containing DMEM/F-12 supplemented with 10% fetal bovine serum at 37 °C in a humidified 5% CO₂ incubator, where explants remained suspended in the medium by their inherent buoyancy. Cultures were maintained for up to 7 days, with culture medium and treatment reagents replenished every other day. Treatment conditions and durations are described in the corresponding sections below.

THBS1 inhibition in ex vivo adipose tissue explants

To assess the transcriptional and metabolic consequences of THBS1 inhibition, ex vivo cultured adipose tissue explants (prepared as described in the Ex vivo adipose tissue culture section) were treated with either the LSKL peptide (100 nM; HY-P0299, MedChemExpress, Monmouth Junction, NJ, USA), a competitive inhibitor of THBS1-mediated TGF-β activation, or an αTHBS1 monoclonal antibody (clone A6-1, GTX21823, GeneTex, Irvine, CA, USA; 1:100 dilution), with saline as the vehicle control. The treatment medium was replenished every 2 days, with treatment duration and downstream readout depending on the experimental endpoint as described below.

Transcriptional phenotyping: Mouse adipose tissue explants were treated for 4 days and harvested for qPCR analyses. Human visceral adipose tissue explants were treated with the αTHBS1 monoclonal antibody for 5 days and harvested for bulk RNA-sequencing.

Metabolic phenotyping (insulin signaling): Mouse epididymal white adipose tissue (eWAT) explants were treated for 7 days. On day 6, explants were subjected to 24 h starvation in glucose-free DMEM containing 2% FBS at 37 °C to suppress baseline insulin signaling. On day 7, starved explants were stimulated for 30 min with 100 nM insulin in growth medium (DMEM/F-12, 10% FBS), and immediately rinsed in ice-cold PBS to halt signaling. Tissues were then harvested for protein extraction and Western blot analysis as described in the Western blot analysis section.

Fractionation of SVFs and mature adipocytes

The obtained adipose tissue from human or mouse visceral fat were immediately minced into small pieces using sterile razor blades under sterile conditions. The minced tissues were digested with type I collagenase (17100017; Gibco) at a final concentration of 1 mg/mL in serum-free DMEM (11965-092; Gibco) for 20 min in a 37 °C warm water bath with gentle shaking to facilitate the breakdown of the extracellular matrix. The digestion reaction was neutralized by adding an equal volume of digestion medium (serum-containing DMEM). The mixture was centrifuged at 500×g for 5 min at RT to fractionate the adipocytes and SVFs. Floating adipocytes were collected from the top layer, whereas the remaining SVF settled at the bottom layer. The adipocytes and the remaining adipose tissue layer were carefully harvested and filtered through 100-µm cell strainers (732-2759, VWR) to remove any undigested tissue fragments. Floating adipocytes were washed twice with PBS and collected for subsequent immunostaining or Matrigel-captured adipocyte culture. The SVF pellet, which contained a mixture of preadipocytes, endothelial cells, immune cells, and fibroblasts, was washed with PBS to remove any residual collagenase and debris. To remove contaminating red blood cells, the pellet was treated with red blood cell lysis buffer (A10492-01; Invitrogen, MA, USA) for 5 min at RT. The SVF was washed twice with PBS, resuspended, and used for subsequent immunostaining, culture or adipocyte differentiation.

Matrigel-captured mature adipocyte culture

Fully mature, unilocular adipocytes were isolated from collagenase-digested mouse epididymal white adipose tissue or human visceral adipose tissue as described in the Fractionation of SVFs and mature adipocytes section. To isolate and culture mature adipocytes, 100 µL of the floating mature adipocyte layer was carefully mixed with 200 µL of Matrigel (354277; Corning, NY, USA). The adipocyte-Matrigel mixture was seeded onto a 48-well plate (30048; SPL Life Sciences) and allowed to solidify at 37 °C for 30 min in a cell culture incubator. After the Matrigel had solidified, the adipocytes were cultured in DMEM/F-12 medium (11330-032; Gibco) supplemented with 10% FBS (16000-044; Gibco) and 1% penicillin–streptomycin (15140-122; Gibco) at 37 °C in a 5% CO₂ incubator. The medium was replaced every other day, and the culture was maintained up to day 7 to support adipocyte viability and homeostatic function.

Lentiviral shRNA production and transduction of mature adipocytes

Lentiviral shRNA expression plasmids targeting mouse Thbs1 (clones #1 and #3, recommended by the manufacturer) and a non-targeting scramble control, all encoding a TagBFP2 reporter, were obtained from VectorBuilder (custom orders). Lentiviruses were produced in HEK293T cells by polyethylenimine (PEI)-mediated co-transfection. Briefly, HEK293T cells (5 × 10⁶) were seeded in 100 mm dishes overnight, then transfected with transfer vector, psPAX2 and pMD2.G at a 5:4:1 ratio (total 20 µg DNA per dish). Six hours post-transfection, the medium was replaced with fresh Opti-MEM (Gibco). Lentiviral supernatants were collected at 48 and 72 h post-transfection, clarified by centrifugation, and concentrated by ultracentrifugation at 50,000 × g. Viral pellets were resuspended in sterile saline. Functional viral titer was determined by transducing ASC52Telo cells with serial dilutions and quantifying TagBFP2⁺ cells by flow cytometry.

For mature adipocyte transduction, freshly isolated floating adipocytes from 14-week HFD-fed C57BL/6 J mice were aggregated and adhered onto a transwell insert using the membrane-anchored mature adipocyte culture (MAAC) approach⁷³. Concentrated lentivirus was added to the surrounding medium at a multiplicity of infection (MOI) over 10 and incubated overnight at 37 °C. The following day, transduced adipocyte aggregates were transferred and embedded in Matrigel as described in the Matrigel-captured mature adipocyte culture section and maintained for 7 days to allow knockdown stabilization before functional assays (2-NBDG uptake, Western blot). Transduction efficiency was confirmed by TagBFP2 fluorescence imaging.

Recombinant THBS1 challenge of mature adipocytes

For functional perturbation experiments, Matrigel-captured mature adipocytes were pre-equilibrated for 24 h post-isolation, then challenged with recombinant human THBS1 (rhTHBS1; 0, 5, 10, 20 nM, Sigma-Aldrich, ECM002) or vehicle (saline) in DMEM/F-12 supplemented with 10% FBS for 6 days, with media replenishment every other day. At the end of treatment, adipocytes were either harvested for downstream RNA or protein analysis (after Matrigel digestion, see below) or fixed for immunocytochemistry as described.

Adipocyte recovery from Matrigel

For cell-based or molecular analyses requiring intact adipocyte recovery, Matrigel-captured adipocytes were dissociated using a sequential collagenase and dispase digestion protocol. Briefly, Matrigel-embedded cultures were incubated with type I collagenase (final 1 mg/mL; C0130, Sigma-Aldrich) and dispase (final 1 mg/mL, diluted 1:1 v/v with Matrigel from a 2 mg/mL stock; 17105-041, Gibco) in serum-free DMEM at 37 °C for 30 min with gentle agitation, followed by 10 min on ice to facilitate complete adipocyte release. Floating adipocytes were collected from the top layer after centrifugation at 150 × g for 5 min at RT and washed twice with PBS to remove residual Matrigel debris. Recovered adipocytes were then processed for insulin-stimulated glucose uptake, immunocytochemistry, or mRNA expression analysis by qRT-PCR (see the detailed information of each step below).

Human primary SVF and ASC52Telo culture and adipogenic differentiation

Fractionated primary SVF cells were cultured and expanded in growth medium (DMEM supplemented with 1% penicillin–streptomycin and 10% FBS) for 1 week prior to adipocyte differentiation. A total of 1.5 × 10⁵ cells were seeded into a 12-well plate, and differentiation was initiated the day after seeding by replacing the growth medium with differentiation medium. Human primary SVF cells were differentiated into adipocytes using the StemPro Adipogenesis Differentiation Kit (A1007001, Gibco), with replenishment of the differentiation medium every other day for 10 days. After differentiation stimulation, the cells were maintained in growth medium until day 20, then harvested for subsequent experiments.

Human telomerase-immortalized adipose-derived stem cells (ASC52Telo; SCRC-4000, ATCC) were maintained in MesenPRO medium (Gibco) supplemented with 1% penicillin–streptomycin at 37 °C in 5% CO₂. For adipogenic differentiation, confluent ASC52Telo cells were induced with the StemPro Adipogenesis Differentiation Kit (A1007001, Gibco) for 14 days, with media replenishment every 2 days. After differentiation, cells were maintained in growth medium (DMEM/F-12, 10% FBS, 1% penicillin–streptomycin) for an additional 7 days before downstream analysis. For THBS1 challenge experiments, rhTHBS1 (0, 5, 10, or 20 nM) was added concurrently with adipogenic induction and maintained throughout the differentiation course. For time-course assays, samples were harvested at Days 5, 10, 15, and 20 for FACS analysis (BODIPY for lipid accumulation, ICAM1 and PPARγ for adipogenic commitment) as described in the Flow cytometry section.

Lipid droplets were stained using BODIPY (D39221, Invitrogen) for 5 min, after which the cells were washed and observed using an Axio Observer 7 fluorescence microscope (Zeiss).

THP-1 macrophage differentiation and conditioned media preparation

Human THP-1 monocytes (ATCC TIB-202) were maintained in RPMI-1640 (Gibco) supplemented with 10% FBS and 1% penicillin–streptomycin at 37 °C in 5% CO₂. To induce macrophage differentiation, THP-1 monocytes (1 × 10⁶ cells/well) were treated with phorbol 12-myristate 13-acetate (PMA; 25 nM, Sigma-Aldrich, P8139) for 48 h, followed by 24 h of rest in PMA-free medium. For polarization, M0 macrophages were further treated with LPS (100 ng/mL, MedChemExpress, HY-D1056) plus IFN-γ (20 ng/mL, MedChemExpress, HY-P78456) for M1 polarization, or IL-4 (20 ng/mL, MedChemExpress, HY-P70764) plus IL-13 (20 ng/mL, MedChemExpress, HY-P72795) for M2 polarization, for 48 h. Polarized (M1 or M2) or unpolarized (M0) macrophages were then challenged with rhTHBS1 (0, 5, 10, or 20 nM) or vehicle for 48 h in serum-free DMEM. For experiments addressing TGFβ activation by THBS1, THP-1 macrophages were co-treated with rhTHBS1 (20 nM) and LSKL peptide (25, 50, or 100 nM), which competitively blocks THBS1-mediated activation of latent TGFβ. Conditioned media (CM) were collected, centrifuged at 500×g for 5 min at RT to remove cell debris, and either used immediately or stored at −80 °C until further use. THBS1 concentration in 10% FBS-supplemented culture media was 105 ± 4.3 ng/mL (~0.23 nM; n = 8 wells) which are substantially lower than the concentrations we tested. Conditioned media from vehicle or rhTHBS1-treated THP-1 cells ranged from 0.53 to 0.61 nM, also substantially lower than the concentrations we tested.

Conditioned media treatment of mature adipocytes and ASCs

Macrophage CM was applied to (i) Matrigel-captured mature adipocytes pre-equilibrated for 24 h post-isolation or (ii) ASC52Telo cells maintained in subconfluent monolayer culture. CM treatment was performed for 6 days with replenishment every 2 days. For TGFβ receptor signaling inhibition, CM was supplemented with SB431542 (25, 50, or 100 nM; HY-10431, MedChemExpress), a selective TGFβRI kinase inhibitor, for the duration of CM treatment. Samples were harvested at the end of treatment for qPCR or immunocytochemistry as described.

Transwell co-culture of THP-1 macrophages with mature adipocytes or ASC52Telo cells

To assess bidirectional paracrine interactions between macrophages and adipose-derived cells, a transwell co-culture system using 0.4 µm pore-size inserts was established to permit soluble factor exchange while preventing direct cell contact. To eliminate any potential carry-over of PMA from the differentiation step, THP-1 macrophages and target cells were prepared separately in two independent plates and combined at the time of co-culture initiation, defined as the end of the THP-1 resting period.

THP-1 + Matrigel-captured mature adipocytes (12-well format): THP-1 monocytes (7 × 10⁵ cells/well) were seeded in 12-well plates and differentiated into M0 macrophages with PMA (25 nM, 48 h) followed by 24 h of rest in PMA-free medium. In parallel, mature adipocytes were freshly isolated by sacrificing mice 1 day prior to the end of the THP-1 resting period and Matrigel-captured on transwell inserts (0.4-µm pore size, 12-well compatible; 37012, SPL, Pocheon, Republic of Korea) as described in the Matrigel-captured mature adipocyte culture section, allowing one day of pre-equilibration to match the readiness of the two compartments. Co-culture was initiated by placing the adipocyte-bearing inserts into the upper chamber of the THP-1–containing wells.

THP-1 + ASC52Telo (24-well format, for immunocytochemistry): ASC52Telo cells (2 × 10⁴ cells/well) were seeded onto 9 mm glass coverslips placed in 24-well plates one day prior to the end of the THP-1 resting period. Cells were maintained in their standard growth medium without adipogenic induction, as the readout focused on fibrotic—not adipogenic—responses. In a separate 24-well plate, THP-1 monocytes (3.5 × 10⁵ cells/insert) were seeded into 24-well-compatible transwell inserts (0.4 µm pore size; 37024, SPL, Pocheon, Republic of Korea) and differentiated into M0 macrophages using the same PMA protocol described above. Co-culture was initiated by transferring the THP-1–bearing inserts onto the ASC52Telo–containing wells.

In both systems, unpolarized M0 macrophages were used at the start of co-culture to allow phenotype to be shaped by the cocultured adipose target cells. Treatments—including recombinant human THBS1 (rhTHBS1; 0, 5, 10, or 20 nM), LSKL peptide (25, 50, or 100 nM; HY-P0299, MedChemExpress), or the TGFβRI kinase inhibitor SB431542 (25, 50, or 100 nM; HY-10431, MedChemExpress)—were added at matched concentrations to both upper and lower chambers to maintain a media-sharing paracrine environment. Co-cultures were maintained for 7 days at 37 °C in a humidified 5% CO₂ incubator, with media in both chambers replenished every 2 days. At the experimental endpoint, ASC52Telo coverslips were processed for immunocytochemistry, and Matrigel-captured adipocytes were recovered as described in the Adipocyte recovery from Matrigel section for downstream immunocytochemistry, qPCR, or Western blot analyses.

2-NBDG glucose uptake assay in Matrigel-captured adipocytes

Insulin-stimulated glucose uptake was assessed using the fluorescent glucose analog 2-NBDG (D-(+)-2-deoxy-2-(N-(7-nitrobenz-2-oxa-1,3-diazol-4-yl)amino)-glucose; N1819, Sigma-Aldrich, St. Louis, MO, USA). Matrigel-captured mature adipocytes (after rhTHBS1 challenge or shThbs1 lentiviral transduction) were subjected to overnight starvation in low-serum, glucose-free DMEM (2% FBS, glucose-free; Gibco) at 37 °C. The following day, Matrigel was removed by sequential collagenase and dispase digestion as described in the “Adipocyte recovery from Matrigel” section, and recovered floating adipocytes were stimulated with insulin (100 nM) and 2-NBDG (10 µM) simultaneously in glucose-free medium for 30 min at 37 °C. After incubation, adipocytes were washed twice with ice-cold PBS to remove unincorporated 2-NBDG, and the floating adipocyte fraction was collected for fluorescence imaging on an Axio Observer 7 fluorescence microscope (Zeiss). 2-NBDG mean fluorescence intensity was quantified per adipocyte using region-of-interest (ROI)-based analysis in ImageJ. Per-field 2-NBDG intensities were averaged within each biological replicate and compared by unpaired two-tailed Student’s t-test.

Flow cytometry

Immune cells and adipose-derived stem cells from the SVF of human visceral adipose tissues were isolated. The adipose tissues were processed to isolate the SVFs as described above. Isolated cells were fixed with 4% paraformaldehyde (PC2031-100-00; Biosesang, Korea) in PBS for 30 min at RT, washed twice with PBS, and centrifuged at 500 × g for 5 min at RT. The cells were permeabilized with 0.1% Triton X-100 (T8787; Sigma-Aldrich, MO, USA) in PBS for 10 min at RT to allow for antibody penetration. The cells were then blocked with 3% bovine serum albumin (BSA; 821006; Millipore, MA, USA) in PBS for 1 h at RT to prevent nonspecific antibody binding. The blocking solution was removed via centrifugation, and the cells were resuspended in PBS. The cells were incubated in a rocker overnight at 4 °C with the following primary antibodies diluted 1:100 in blocking buffer: CD68 (333828; BioLegend, CA, USA), SLC11A1 (ANT-201; Alomone Labs, Israel), CD80 (305220; BioLegend), CD11c (25-0114-82; Thermo Fisher Scientific), CD163 (ab182422; Abcam, UK), CD31 (14-0311-82; Thermo Fisher Scientific), CD140a (14-1401-82; Thermo Fisher Scientific), THBS1 (NBP1-52410; Novus), COL6A1 (NB120-6588; Novus, MO, USA), anti-ACTA2 (NBP2-32808; Novus), CD9 PerCP Cy5.5 conjugated (BD, 561326) and unconjugated syndecan-1 (R&D Sys, MAB2780). For phenotypic characterization of ASC52Telo cells during adipogenic differentiation, three additional staining panels were used: (i) FITC-conjugated anti-CD9 (clone HI9a, 312103; BioLegend; 1:100) was applied as a single-color stain in a separate tube; and (ii) a co-staining panel comprising BODIPY 493/503 (D39221, Invitrogen) for neutral lipid detection, APC-conjugated anti-ICAM1/CD54 (clone HA58, 353111; BioLegend; 1:100), and unconjugated rabbit polyclonal anti-PPARG (bs-4590R; Bioss; 1:100), the latter detected with mouse anti-rabbit IgG-CFL594 (sc-516250; Santa Cruz Biotechnology; 1:200) as the secondary antibody. The following day, the cells were washed thrice with 3% BSA in PBS via centrifugation at 500 × g for 5 min at RT. After washing, cells were resuspended in PBS and analyzed using a BD Symphony flow cytometer (BD Biosciences, NJ, USA), which was configured with fluorescence compensation for spectral overlap. Raw flow cytometry data were collected and processed using the FlowJo software (FlowJo, LLC). The gating strategy was designed based on forward and side scatter profiles to exclude debris and doublets, followed by marker-based identification of immune subsets (CD45⁺, CD68⁺) and ASCs (CD140a⁺/CD31⁻) using appropriate fluorescence compensation.

Histological assessment

Adipose tissue samples (human VAT, mouse eWAT and iWAT) were fixed in 4% paraformaldehyde, embedded in paraffin, and sectioned at 5-μm thickness onto glass slides. Slides were dried on a 37 °C warm plate, melted in a 60 °C oven to remove residual paraffin, deparaffinized in three changes of 100% xylene, and rehydrated through a graded ethanol series (100% × 2, 90%, 70%, 50%, 30%, and distilled water; 10 min each).

Picrosirius Red staining

Following rehydration, sections were stained for collagen using a Picrosirius Red Stain Kit (Polysciences, Warrington, PA, USA) according to the manufacturer’s instructions, with the following modifications. Sections were pre-treated with phosphomolybdic acid (Solution A) for 5 min, washed, and stained with Picrosirius Red (Solution B) for 2 h at RT. After staining, sections were differentiated in acidified HCl (Solution C) for 5 min, briefly rinsed twice with 0.5% acetic acid in PBS, dehydrated through a graded ethanol series, cleared in xylene, and mounted with DPX mountant (06522; Sigma-Aldrich). Stained slides were imaged on a bright-field light microscope. Adipose tissue fibrosis was assessed semiquantitatively using the Fibrosis of Adipose Tissue (FAT) score, which integrates perilobular and pericellular fibrosis grading, as previously described by ref. ⁷⁴.

Immunohistochemistry

Following rehydration, heat-induced epitope retrieval was performed in citrate buffer (pH 6.0) using a steam cooker for 15 min, followed by cooling to room temperature in lukewarm water. Sections were outlined with a hydrophobic barrier pen (PAP pen) and treated with 1% H₂O₂ in 5% methanol to quench endogenous peroxidase activity and eliminate red blood cell autofluorescence. After PBS washes, sections were blocked with 5% bovine calf serum (BCS) in PBS for 2 h at RT.

Sections were then incubated overnight at 4 °C with primary antibodies diluted 1:100 in blocking buffer. Primary antibodies used included: anti-SLC11A1 (rabbit polyclonal, ANT-201; Alomone Labs, Jerusalem, Israel), anti-CD68 (mouse IgG1 chimeric monoclonal, clone EPR20545, ab283316; Abcam, Cambridge, UK), anti-CD163 (mouse monoclonal, clone GHI/61, MA5-17716; Invitrogen, Carlsbad, CA, USA), anti-IL6 (mouse monoclonal, SAB3300071, Sigma-Aldrich), anti-GFP (mouse monoclonal, clone 4B10B2, MA5-15349; Invitrogen), anti-6×His tag (mouse monoclonal, clone HIS.H8, MA1-21315; Invitrogen), anti-THBS1 (goat polyclonal, NBP1-52410 Novus), and anti-PLIN1 (rabbit polyclonal, 27716-1-AP, ProteinTech). Antibody combinations and corresponding panels are described in figure legends.

Two detection modalities were used:

• (i)Fluorescence detection: After three PBS washes, sections were incubated for 1 h at RT with appropriate species-matched fluorescent secondary antibodies diluted 1:200 in PBS. Sections were mounted with VECTASHIELD Mounting Medium with DAPI (H-1200; Vector Laboratories, Burlingame, CA, USA) for nuclear counterstaining. Imaging was performed on an Axio Observer 7 fluorescence microscope (Zeiss).
• (ii)Chromogenic (DAB) detection—performed for SLC11A1 only: After three PBS washes, sections were incubated with HRP-conjugated anti-rabbit IgG secondary antibody (#7074; Cell Signaling Technology, Danvers, MA, USA) for 1 h at RT, washed, and developed with the DAB Substrate Kit (SK-4100; Vector Laboratories, Newark, CA, USA) according to the manufacturer’s instructions. Sections were counterstained with Mayer’s Modified Hematoxylin (ab220365; Abcam) for nuclear staining, dehydrated through a graded ethanol series, cleared in xylene, and mounted with DPX mountant (06522; Sigma-Aldrich). Slides were imaged on an Axio Observer 7 microscope under bright-field illumination.

Immunocytochemistry of primary ASC culture

Differentiated human primary ASCs were cultured on glass coverslips in 24-well plates. The cells were fixed with 4% paraformaldehyde in PBS for 10 min at RT and washed thrice with PBS with gentle agitation to remove the residual fixative. The fixed cells were permeabilized with 0.1% Triton X-100 in PBS for 5 min at RT and blocked with 3% BSA in PBS for 1 h at RT. The cells were then incubated overnight at 4 °C with the following primary antibodies diluted 1:100 in blocking buffer: anti-COL6A1 (17023-1-AP; Proteintech, Rosemont, IL, USA), and anti-ACTA2 (NBP2-32808; Novus). The following day, the cells were washed thrice with 3% BSA in PBS to remove any unbound primary antibodies. The cells were then incubated for 30 min at RT with the following secondary antibodies diluted 1:200 in PBS: mouse anti-rabbit IgG–CFL488 (sc-516248; Santa Cruz Biotechnology, Dallas, TX, USA) for COL6A1 detection, and m-IgGκ BP–CFL594 (sc-516178; Santa Cruz Biotechnology) for ACTA2 detection. Subsequently, the cells were washed thrice with 3% BSA in PBS and stained with 4’,6-diamidino-2-phenylindole (DAPI; 62248, Thermo Fisher Scientific) diluted 1:1000 in PBS for 5 min at RT to visualize the nuclei.

Immunostaining of floating adipocytes and adipose tissues

Floating mature adipocytes were prepared from two distinct experimental contexts: (i) freshly isolated human visceral adipocytes from Lean, obesity, and donors with obesity and diabetes and (ii) Matrigel-captured mouse mature adipocytes following recombinant THBS1 (rhTHBS1) challenge.

In both contexts, adipocytes were fixed with 4% paraformaldehyde in PBS for 30 min at RT and washed thrice with PBS using a 1 mL syringe (1CC-26G; KoVax, Israel) to gently aspirate wash buffer without disrupting the floating layer. Cells were permeabilized with 0.05% Tween 20 (P2287; Sigma-Aldrich) in PBS for 10 min at RT, washed thrice with PBS, and blocked with 3% BSA in PBS for 1 h at RT. Cells were then incubated overnight at 4 °C with the following primary antibodies diluted 1:50 in 3% BSA in PBS: anti-THBS1 (NBP1-52410; Novus Biologicals), anti-COL6A1 (NB120-6588; Novus Biologicals; for human samples), anti-COL6A1 (17023-1-AP; Proteintech; for mouse samples), anti-IL6 (SAB3300071 Sigma-Aldrich), anti-4-HNE (ab48506; Abcam), and anti-TNF (mouse monoclonal,60291-1-Ig, ProteinTech). Antibody combinations used per panel are described in the corresponding figure legends.

After three PBS washes, samples were incubated for 30 min at RT with species-matched secondary antibodies diluted 1:100 in PBS: mouse anti-rabbit IgG–CFL488 (sc-516248; Santa Cruz Biotechnology, Dallas, TX, USA) for COL6A1 and 4-HNE detection, m-IgGκ BP–CFL594 (sc-516178; Santa Cruz Biotechnology) for IL6 and TNF detection, and mouse anti-goat IgG–CFL680 (sc-516245; Santa Cruz Biotechnology) for THBS1 detection. Nuclei were counterstained with DAPI (62248; Thermo Fisher Scientific) diluted 1:1000 in PBS for 5 min at RT.

Because mature adipocytes float in suspension due to their lipid content and resist conventional coverslip mounting, stained adipocyte suspensions were imaged directly on disposable cell counting slides (Countess™ Cell Counting Chamber Slides; Invitrogen). Briefly, 10 µL of the stained adipocyte suspension was slowly loaded into each chamber to avoid disrupting cell integrity, and slides were imaged in their entirety. Images were acquired on a Zeiss LSM710 confocal microscope or an Axio Observer 7 fluorescence microscope (Zeiss) using appropriate fluorescence filter sets for CFL488 (excitation, 495 nm; emission, 519 nm), CFL594 (excitation, 590 nm; emission, 617 nm), AF647 (excitation, 650 nm; emission, 668 nm), and DAPI (excitation, 358 nm; emission, 461 nm), and processed using ZEN software (Zeiss).

Whole-mount immunostaining of adipose tissue explants

For whole-mount tissue-level immunostaining, freshly isolated mouse eWAT was cut into small pieces (~5 mm³) and cultured ex vivo as described in the Ex vivo adipose tissue culture section. Tissue explants were treated with recombinant human THBS1 (rhTHBS1; 5 nM) for 4 days, with medium changed every 2 days, to induce THBS1-driven downstream responses prior to fixation.

After the challenge, tissue pieces were fixed with 4% paraformaldehyde in PBS overnight at 4 °C, followed by three PBS washes. Tissues were permeabilized with 0.5% Triton X-100 (T8787; Sigma-Aldrich) in PBS for 15 min at RT, washed once with PBS, and blocked with 5% fetal bovine serum (FBS) in PBS for 1 h at RT.

Tissues were then incubated overnight at 4 °C with the following primary antibodies diluted 1:100 in blocking buffer: anti-Col1a1 (rabbit polyclonal, ab21286; Abcam, Cambridge, UK), anti-TNF (mouse monoclonal, 60291-1-Ig, ProteinTech), and anti-Perilipin (goat polyclonal, NB100-60554; Novus Biologicals, Centennial, CO, USA). After three PBS washes, tissues were incubated for 1 h at RT with species-matched secondary antibodies diluted 1:200 in PBS: mouse anti-rabbit IgG–CFL488 (sc-516248; Santa Cruz Biotechnology, Dallas, TX, USA) for Col1a1, m-IgGκ BP–CFL594 (sc-516178; Santa Cruz Biotechnology) for TNF, and mouse anti-goat IgG–CFL680 (sc-516245; Santa Cruz Biotechnology) for Perilipin. Nuclei were counterstained with DAPI (62248; Thermo Fisher Scientific) diluted 1:1000 in PBS for 5 min at RT.

Stained tissue explants were embedded in 1% low-melting-point agarose (16520050; Thermo Fisher Scientific) on glass slides for whole-mount imaging, which preserves three-dimensional tissue architecture without compression. Imaging was performed on an Axio Observer 7 fluorescence microscope (Zeiss, Oberkochen, Germany) using appropriate fluorescence filter sets for CFL488 (excitation, 495 nm; emission, 519 nm), CFL594 (excitation, 590 nm; emission, 617 nm), CFL680 (excitation, 683 nm; emission, 700 nm), and DAPI (excitation, 358 nm; emission, 461 nm). For figure presentation, fluorescence channels were pseudocolored using ZEN software: Col1a1 (CFL594) in red, TNF (CFL488) in cyan, and Perilipin (CFL680) in green. Images were processed using ZEN software (Zeiss).

Western blot analysis

Adipocyte protein lysates (recovered from Matrigel as described in the Adipocyte recovery from Matrigel section) were prepared in RIPA buffer, denatured in Laemmli sample buffer at 95 °C for 5 min, and resolved on SDS–PAGE gels cast in-house using Tris (T1503; Sigma-Aldrich), 10% ammonium persulfate (APS; A3678; Sigma-Aldrich), SDS (L3771; Sigma-Aldrich), and 30% acrylamide/bis-acrylamide solution (1610154; Bio-Rad), run on a Bio-Rad Mini-PROTEAN Tetra electrophoresis system. Proteins were transferred onto PVDF membranes (1620177; Bio-Rad) using a Bio-Rad wet transfer system at 10 V overnight in transfer buffer containing 10% methanol to enable detection of high-molecular-weight THBS1 (~450 kDa trimer, ~140 kDa monomer).

Membranes were blocked in 3% bovine serum albumin (BSA; A7906; Sigma-Aldrich) in TBST for 1 h at RT and incubated overnight at 4 °C with the following primary antibodies (1:1000 unless specified): anti-phospho-AKT (Ser473) (rabbit polyclonal, #9271; Cell Signaling Technology, Danvers, MA, USA), anti-AKT pan (rabbit polyclonal, #9272; Cell Signaling Technology), anti-THBS1 (1:500; mouse monoclonal, clone A6-1, GTX21823; GeneTex, Irvine, CA, USA), anti-β-tubulin (mouse monoclonal, clone BT7R, MA5-16308; Thermo Fisher Scientific, Waltham, MA, USA), anti-GAPDH (rabbit monoclonal, clone D16H11, #5174; Cell Signaling Technology), anti-GLUT4 (mouse, 2213 s; Cell Signaling Technology), and anti-Adiponectin (rabbit polyclonal, NB100-65810; Novus Biologicals, Centennial, CO, USA).

After three TBST washes, membranes were incubated with HRP-conjugated species-matched secondary antibodies (1:5000) for 1 h at RT, washed, and developed using enhanced chemiluminescence substrate (Cytiva). Chemiluminescent signals were captured on a ChemiDoc MP imaging system (Bio-Rad), and densitometric quantification was performed using ImageJ (NIH, Bethesda, MD, USA), with target protein levels normalized to β-tubulin or GAPDH loading controls.

RNA extraction, cDNA synthesis, and quantitative PCR

Total RNA was extracted from differentiated adipocytes (in vitro-differentiated human primary ASCs harvested at day 14 of adipogenic induction), from Matrigel-captured mature adipocytes (recovered from Matrigel as described in the Adipocyte recovery from Matrigel section), and from mouse eWAT tissue explant cultures. For tissue explant samples, mouse eWAT explants were minced into small pieces with sterile surgical scissors directly in TRIzol Reagent and disrupted by vigorous vortexing prior to phase separation. Total RNA was extracted using TRIzol Reagent (15596018; Ambion, Thermo Fisher Scientific, Waltham, MA, USA) following the manufacturer’s protocol. RNA concentration and purity were assessed using a NanoDrop spectrophotometer (ND-1000; Thermo Fisher Scientific).

Complementary DNA (cDNA) was synthesized from 2 µg of total RNA using the High-Capacity cDNA Reverse Transcription Kit (K1642; Thermo Fisher Scientific) on a thermal cycler programmed for 25 °C × 5 min, 42 °C × 60 min, and inactivation at 85 °C × 5 min.

Quantitative PCR was performed on a QuantStudio™ 7 Flex Real-Time PCR System (Thermo Fisher Scientific) using TOPreal™ qPCR 2× PreMIX (SYBR Green with low ROX) (RT5015; Enzynomics, Daejeon, Republic of Korea). Reactions were prepared in 20 µL volumes with the following amplification protocol: initial denaturation at 95 °C for 10 min, followed by 40 cycles of 95 °C × 15 s (denaturation), 60 °C × 20 s (annealing), and 72 °C × 40 s (extension). Primer sequences are listed in Supplementary D3. Gene expression was normalized to ACTB and calculated using the ΔΔCt method.

ChIP and ChIP-PCR

To validate the direct upstream regulators of THBS1 expression, ChIP and PCR were performed as previously described⁷⁵, with minor modifications. Human visceral adipose tissues (1 g) were used for ChIP. They were thawed on ice and minced in TST buffer (10 mM Tris-HCl, pH 8.0, 0.25% SDS, 1 mM EDTA, 150 mM NaCl, and 1x protease inhibitor cocktail) using a razor. After mincing, the samples were incubated on ice for 20 min to allow for complete lysis. Nuclei from adipose tissues were purified with a 100 µm pore size cell strainer. Purified nuclei were sonicated using Ultrasonic dismembrator (VCX-130; Invitrogen) to shear the chromatin into fragments of ~500 bp. Sonication was performed at 70 W with 1 min “on” and 2 min “off” cycles, for a total of 15 min. After sonication, the lysate was clarified via centrifugation at 16,000×g for 10 min at 4 °C, and the supernatant containing the chromatin fragments was collected and transferred to new tubes for immunoprecipitation.

For immunoprecipitation, chromatin was incubated with antibodies against the transcription factors TEAD1 (PA5-37075; Invitrogen) and WWTR1 (NB110-58359; Novus) to enrich the specific chromatin-protein complexes of interest. Normal rabbit IgG (02-6102; Invitrogen) was used as a negative control. The antibody-chromatin mixture was incubated overnight at 4 °C with gentle shaking to allow sufficient binding between the antibodies and their respective targets. Following incubation, 2 μL of protein A/G magnetic beads (88802; Pierce Biotechnology, MA, USA) was added to the chromatin-antibody complex and incubated for 2 h at 4 °C with gentle shaking to capture the antibody-bound chromatin complexes. The beads were washed sequentially with three buffers to reduce nonspecific binding: low-salt buffer (20 mM Tris-HCl, pH 8.0, 150 mM NaCl, 0.1% SDS, and 1% Triton X-100), high-salt buffer (20 mM Tris-HCl, pH 8.0, 500 mM NaCl, 0.1% SDS, and 1% Triton X-100), and LiCl wash buffer (10 mM Tris-HCl pH 8.0, 250 mM LiCl, 0.5% NP-40, and 0.5% sodium deoxycholate). The beads were washed once more with TE buffer (10 mM Tris-HCl, pH 8.0, and 1 mM EDTA) to remove the residual washing buffer. The protein-DNA cross-links were reversed by adding 5 M NaCl to the immunoprecipitated complexes and incubating them at 65 °C for 4 h. To ensure the complete reversal of cross-linking, the samples were vortexed every 30 min. The chromatin was digested by adding 200 µg/mL proteinase K (P2308-10MG; Sigma-Aldrich) and incubating the sample at 45 °C for 4 h, after which proteinase K was inactivated via incubation at 65 °C for 30 min. DNA was purified using a PCR Purification Kit (113-102; Gene All) following the manufacturer’s instructions. DNA concentration was determined using a NanoDrop spectrophotometer (ND-1000; Thermo Fisher Scientific), and its quality was assessed by running a portion of the sample on a 1% agarose gel. qPCR was performed as previously described to analyze the promoter regions of human THBS1 (chr15:39,579,678-39,580,199), which overlaps with a TEAD1/WWTR1 binding peak identified in ChIP-seq datasets from human adipocytes (ChIP-Atlas dataset SRX7807127⁷⁶;). The relative enrichment of THBS1 promoter DNA was calculated using the ΔCt method, with IgG isotype control samples serving as the baseline reference for normalization,.

Enzyme-linked immunosorbent assay (ELISA)

THBS1 levels were quantified in two sample types: (i) patient plasma collected from human VAT donors and (ii) whole-cell lysates of in vitro-differentiated human adipocytes, using the Human THBS1 ELISA Kit (BMS2100; Invitrogen). For TGFβ analysis, conditioned media collected on day 4 of rhTHBS1 challenge from THP-1-derived macrophages (prepared as described in the “Conditioned media preparation” section) were assayed using the TGFβ ELISA Kit (BMS249-4; Invitrogen). Adiponectin levels in conditioned media collected on day 4 of rhTHBS1 challenge from Matrigel-captured mature mouse adipocytes were quantified using the Mouse Adiponectin ELISA Kit (KMP0041; Invitrogen). All ELISAs were performed according to the manufacturer’s protocol, including standards, sample incubation, washing, and detection steps. The concentrations of THBS1, TGFβ, and adiponectin were determined by comparing the optical density values of the samples to the standard curves generated from known concentrations of recombinant proteins. Protein concentrations were normalized to the total protein content of the samples, which was determined using a BCA assay kit (23228; Thermo Fisher Scientific) according to the manufacturer’s instructions.

Statistical analysis

Statistical analyses were performed using GraphPad Prism (v8.4.3). Two-group comparisons used unpaired two-tailed Student’s t-tests. Three-group and multi-group comparisons used one-way ANOVA followed by Tukey’s (all pairwise; Figs. 2G, 3G, 4F, and 7C) or Šidák’s (selected pairwise; Figs. 6E, G, I, J and 7Q) multiple-comparison test. Paired comparisons within the same individual or sample (Figs. 2H, I, 3H, I, and 7P) used paired two-tailed t-tests. For 2-NBDG (Fig. 4M), per-field intensities were averaged within each sample and compared by an unpaired two-tailed t-test. Data were mean ± SEM unless otherwise stated; error-bar type is specified in each legend. P < 0.05 was considered significant.

Reporting summary

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

Author contributions

J.J. and E.S.P. contributed equally. J.J., E.S.P., J.H.M, J.K.K., and S.H.C. conceptualized the study. Y.S.P., H.J.K., J-K.K., and T.J.O. provided adipose tissue samples and clinical data. N.K.L., J-J.L., U.J., and Q.Z.H. prepared adipose tissue samples. J.J. made single-nuclei RNA-sequencing and bulk RNA-sequencing libraries and performed wet experiments. J.J., E.S.P., and J.H.M. analyzed the data. J.H.M., J.K.K., and S.H.C. supervised the study. J.J., E.S.P., J.H.M., J.K.K., and S.H.C. wrote the manuscript. All authors contributed to editing and reviewing the manuscript.

Peer review

Peer review information

Nature Communications thanks Tae-Hwa Chun, Tuo Deng and the other anonymous reviewer for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the Korea Health Industry Development Institute (RS-2024-00403679 and RS-2024-00404132) and the National Research Foundation of Korea (RS-2025-00555731) to J.H.M.; the National Research Foundation of Korea (RS-2024-00451940 and RS-2025-00559738) to E.S.P.; the National Research Foundation of Korea (RS-2024-00426031, 2021M3H9A1030158, and RS-2023-NR077247) to J.K.K.; and the National Research Foundation of Korea (2018R1A5A2024425 and RS-2023-00218616) to S.H.C.

Data availability

Raw and processed single-nucleus RNA-seq and bulk RNA-seq data analysed in this study are publicly available in the NCBI Gene Expression Omnibus under accession number GSE342773, with raw reads deposited in the Sequence Read Archive. The same datasets have also been deposited in the Korea BioData Station (K-BDS) under accession number KAP241889. Source Data files containing all quantitative data underlying the figures are provided with this paper. Source data are provided with this paper.

Code availability

No custom code was generated in this study. All analyses were performed using publicly available software packages, as described in the Methods.

Competing interests

Sung Hee Choi, Joon Ho Moon, Jisu Jung, Jong Kyoung Kim, and Eun Seo Park are named as inventors on a pending patent application related to the findings reported in this manuscript (Republic of Korea application no. 10-2025-0130870). The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77107-w.