White adipose tissue (WAT) is a uniquely plastic organ that can expand or shrink in response to caloric supply and demand. The ability to function across pronounced variations in tissue mass is governed by a plethora of resident and recruited cell types1. Disturbed WAT remodeling leads to changes in the cell composition of the tissue, which in turn increases the risk of developing insulin resistance, type 2 diabetes, and other cardiometabolic complications2. Defining the cellular landscape and microarchitecture of human WAT in health and disease is therefore of considerable clinical relevance.

To determine cell composition, single-cell technologies have been applied to WAT obtained from different depots. Together, these studies have identified novel specialized cells in adipose tissue such as: (i) adipogenic precursor cells with anti-adipogenic effects3,4,5, (ii) lipid-associated macrophages (LAMs) with central roles in metabolic health6 and (iii) adipocyte subtypes with distinct sensitivities to insulin7 or thermogenic effects8. However, a caveat with most single-cell studies is that issues related to data production restrict sequencing depths and the number of samples that can be processed. Consequently, most published reports in the adipose field are based on small cohort sizes and even the largest ones have, so far, included fewer than 15 individuals9,10. This, together with qualitative differences between technical platforms and bioinformatic approaches, limits the generalizability of the observed findings.

To address this, we performed a comprehensive meta-analysis of newly generated and publicly available data where we integrated single-cell (scSeq) and single-nucleus (snSeq) RNA sequencing results. Based on this, we created a cellular meta-map which we used to deconvolve spatial and bulk transcriptomic data from women and men spanning over a very broad range in age, BMI, and metabolic states. By mining this rich resource, we provide a nomenclature of cells residing in human WAT, their localization and how they relate to metabolic health.


Human WAT contains four major cell classes

To define the cellular landscape of human WAT, we retrieved scSeq, snSeq and spatial transcriptomic (STx) results from ten published reports comprising samples from subcutaneous, visceral, and perivascular WAT6,7,8,9,10,11,12,13,14,15. We combined this with unpublished data from four additional cohorts (Massier et al. #1–4), resulting in a total of 17 datasets across studies and depots (Supplementary Table 1). As displayed in Fig. 1a, together these included 401,320 quality-filtered cells/nuclei (hereafter referred to as objects) obtained from 103 samples of 83 donors spanning over a broad range in age (22–77 years) and body mass index (BMI, 17–55 kg/m2).

Fig. 1: A meta-map to define human WAT composition.
figure 1

a For each included cohort, number (n) and gender of subjects, age (years) and body mass index (BMI) ranges (min–max) as well as number of objects (cells/nuclei), method (single-nucleus Seq [snSeq], single-cell Seq [scSeq] or spatial transcriptomics [STx]) and Jaccard index are displayed. Massier et al. #1–4 refer to data generated for the present meta-analysis and gray bars (n/a) indicate that no information was obtained. Boxes for age and BMI represent a range (min–max), boxplots are presented as interquartile range plus median and Tukey whiskers. Summarizing statistics are displayed in the right panels as mean ± S.D. b Network displaying nodes (subclusters from each study) and edges (marker gene overlap). Data were distributed into four major classes and named based on prominent marker genes. Node sizes are reflecting cluster proportions. The displayed network does not include results from Hildreth et al.11, as data from this study overlapped poorly with the others. c, d Cell class proportions comparing c) methods (snSeq vs. scSeq) and d depots (subcutaneous [sc] WAT vs. omental [om] vs. perivascular [pv]). Note that adipocytes are only available by snSeq. Data are shown as mean ± S.D. Statistical differences were calculated by two-sided Mann–Whitney U test between sc (n = 10) and om (n = 4) WAT or scSeq (n = 6) and snSeq (n = 8). Because of fewer cases, no statistics were calculated for adipocytes and pvWAT. e K-nearest-neighbor batch-effect test (kBET) and adjusted Rand index (ARI) for raw or integrated data using the indicated methods displayed according to method, depot, and cohort. BBKNN batch balanced k-nearest neighbors, scVI single-cell variational inference, rPCA reciprocal principal component analysis. Source data are provided as a Source data file.

To create a meta-map, we first defined cell clusters by processing all datasets individually using Seurat v4.116 (Supplementary Fig. 1a and Supplementary Data 1). Based on these results, we calculated a Jaccard index comparing cell cluster marker genes between studies. We found that all data, except the one from Hildreth et al.11, largely overlapped (Fig. 1a). This demonstrates that: (i) most cell types were present across the different cohorts/depots and (ii) snSeq, scSeq, and STx on average capture similar trends. Because STx is a spot-based approach that requires specific tools for data deconvolution17, we analyzed the Bäckdahl et al.7 study separately. For the remaining snSeq/scSeq datasets, we performed a high-level topological analysis, which revealed that clusters identified in the individual studies separated into four major cell classes: adipocytes, fibroblast and adipogenic progenitors (FAPs), vascular and immune cells (Fig. 1b). To validate our findings, we applied CellTypist18 to the immune cells which confirmed our annotations for this class (Supplementary Fig. 1b).

Our comprehensive classification allowed us to estimate the cellular composition of WAT (Fig. 1c, d). For this, we first used snSeq data, as this method, in contrast to scSeq, captures fat cells. After quantifying the number of adipocytes in each dataset, we next compared the proportions of the other cell classes between methods. We found that FAPs constituted the largest class of ~40% of the total cell population according to both snSeq and scSeq. This was followed by adipocytes and immune cells, which were present in a 1:1 ratio and together constituted ~40% of the cells. For adipocytes, this is in line with what has previously been reported in the literature19. About 15% of the objects were vascular cells, but the recovery of this cell class was strongly influenced by methods where the proportion was approximately three times higher with snSeq compared to scSeq. This could be due to pre-selection bias introduced by enzymatic digestion as has been shown before for endothelial cells in mouse kidney20. Comparisons across depots showed that there were no significant differences in cell class abundance. As expected, scSeq studies where cells had been enriched for either FAPs (CD45)13 or immune cells (CD45+)14 prior to sequencing displayed markedly higher proportions of these cell types compared to the rest of the datasets, supporting the validity of our classification (Supplementary Fig. 1a). Of note, for perivascular WAT, snSeq data were only available from one cohort implying that the generalizability of the results from this depot needs to be further validated.

Cell class-based analyses successfully integrates data across studies

As much of the biological variability overlapped between cohorts, we performed data integration across studies. In contrast to the label-centric analysis presented in Fig. 1b, this approach facilitates the identification of rare cell types which risk being omitted in more sparsely sampled datasets. Many tools are available for integrative analyses and all of them introduce different biases where a trade-off between overfitting vs. insufficient integration must be considered21,22. In a first step, we applied single-cell annotation using variational inference (scANVI) as prior knowledge of cell type annotations can improve integration results23. In line with the findings in Fig. 1a, b, data from Hildreth et al. could not be integrated with scANVI and this dataset was therefore excluded from all subsequent analyses (Supplementary Fig. 1c). After removing Hildreth et al. the scANVI results mirrored the high-level topological analysis shown in Fig. 1b (Supplementary Fig. 1d). We therefore split the data into these four classes and tested different integrative frameworks. We included reciprocal principal component analysis (rPCA)16, batch balanced k-nearest neighbors (BBKNN)24, Harmony25 and single-cell variational inference (scVI)26, as all of them have benchmarked well in prior systematic comparisons27,28. To evaluate these methods across the main confounders (techniques, depots and cohorts), we calculated the adjusted Rand index (ARI)29, k-nearest-neighbor batch-effect test (kBET)30 and Local Inverse Simpson Index (LISI)25,31. In comparison to using all objects as input, we obtained enhanced integration results when splitting the data into immune cells, vascular cells, FAPs, and adipocytes. Our results showed that (i) Harmony performed well for FAP integration, (ii) scVI and rPCA provided the best results across all scores for methods, cohorts, and depots, and (iii) kBET showed overall the lowest rejection rates for scVI (Fig. 1e, Supplementary Fig. 2a, b). This in combination with a previous report suggesting that scVI works best for complex data27, prompted us to apply this method to all four cell classes separately. Given that the included datasets varied in object number, we capped them to contain similar object counts to reduce bias and enhance integration (Supplementary Fig. 2c, d). Based on this, we included an optimal number of objects for each cell class which resulted in cell clusters represented by most studies and methods (Supplementary Fig. 2d, e). With these criteria, we created WAT annotation models which are publicly available and can be applied in future studies (see Methods).

Immune cells with different origins and activation states are present in human WAT

The immune cell class integrated well across techniques, studies, and depots (Fig. 1e, Supplementary Fig. 2b). As displayed in Supplementary Figures 3a, b, we annotated two major groups containing distinct subpopulations of: (i) T, natural killer (NK), and NKT cells (28.8%) and (ii) monocytes, macrophages, and dendritic cells (67.3%) as well as three minor populations including mast (2.94%), B (0.87%), and plasma B cells (0.16%).

The large number of objects allowed us to further dissect these groups separately to provide a high-resolution map of the different cell populations. In the major lymphocyte group, we identified 11 different clusters which were denoted lyC0-10, in order of abundance from high to low (Fig. 2a, b, Supplementary Table 2 and Supplementary Data 2). These included subtypes of (i) Th1-polarized (lyC0), tissue-resident memory (TRM, lyC01), naive/early differentiated (lyC04), and naive/regulatory (TREG, lyC08) CD4+ T cells (ii) CD8+ T cells including early- (lyC02) and late-differentiated (lyC03) cells, (iii) NKT cells (lyC05), as well as (iv) CD16+ (lyC06) and CD16 (lyC09) NK cells. Multiple cell classes, including TREG and TRM, could be further subdivided based on their differentiation states (Fig. 2a, lower panel). We confirmed these data by flow cytometry (Supplementary Fig. 3c) using a panel of antibodies identified in the transcriptomic analyses (see Methods). Comparisons across depots revealed that all identified T and NK cell populations were present in subcutaneous, omental, and perivascular WAT albeit with slightly different proportions (Fig. 2b). For example, Th1 CD4+ T cells (lyC0) were more abundant in omental while late-differentiated CD8+ T cells (lyC03) enriched in subcutaneous WAT (Fig. 2b). These observations were confirmed by deconvolution of bulk transcriptomic data in two independent cohorts32,33 (Fig. 2c).

Fig. 2: Analyses of the WAT immune cell panorama reveal novel subtypes.
figure 2

a Nomenclature (upper panel) and expression patterns of selected marker genes (lower panel) for T, NKT, and NK cells (lyC0-10). b Proportions (%) of T, NKT, and NK cells (lyC0-10) in subcutaneous (sc), omental (om), and perivascular (pv) WAT depots. c The enrichment of lyC0 in omental and lyC03 in subcutaneous WAT was supported by deconvolution of bulk transcriptomic data from Arner et al.32 (left panel) and Krieg et al.33 (right panel). p values were calculated by two-sided Wilcoxon signed-rank test. d Same visualization as a, but for monocytes and macrophages. e Selected marker gene expression profiles for omental-enriched myC08 and myC12. f Same visualizations as b, but for monocytes and macrophages. g Same as c, but for myC07, myC08, and myC12. DC dendritic cells, diff differentiated, LAM lipid-associated macrophages, MMe metabolic-regulated macrophages, Mo monocytes, Mox redox-regulatory metabolic macrophages, NK natural killer cells, NKT natural killer T cells, Th helper T cells, TREG regulatory T cells, TRM tissue-resident memory T cells. Source data are provided as a Source data file.

We next classified the myeloid group. This resulted in 16 different clusters (myC0-15) including macrophages, non-classical (myC05) and classical (myC14) monocytes, as well as class 2 dendritic cells (myC03) (Fig. 2d, Supplementary Table 2 and Supplementary Data 2). In-depth analyses of macrophages revealed several M2-like subpopulations (myC0-01, myC04, myC07-09, myC11-13) out of which two (myC08 and −12) have not been previously reported in WAT (Fig. 2d, e). We also identified mixed M1/M2-like (myC06), lipid-associated (LAM, myC02), metabolic-regulated (MMe, myC10) and redox-regulatory metabolic (Mox, myC15) macrophages (Fig. 2d). Similar to lymphoid cells, our flow-cytometry-based follow-up studies confirmed the presence of several of the identified myeloid subpopulations in subcutaneous WAT (Supplementary Fig. 3d). Analyses over the three depots revealed that LAMs (myC02) were present in both subcutaneous and omental WAT but were virtually absent in the perivascular depot, while Mox (myC15) were only present in the subcutaneous region (Fig. 2f). The omental region was enriched for different types of M2-like cells (myC07-08 and myC12), observations which were confirmed by deconvolution of bulk transcriptomic data (Fig. 2g). In addition to these larger groups, we also analyzed B and mast cells (Supplementary Fig. 3e, f, Supplementary Data 2). B cells separated into activated (bC01, characterized by upregulation of genes related to regulation of T cell-mediated cytotoxicity) and non-activated states (bC0). While mast cells could not be clearly subdivided, there were trends in the data indicating that both MCT (tryptase-positive) and MCTC (positive for chymase and carboxypeptidase) cells were present.

WAT contains vascular cells with angio- and adipogenic expression profiles

For vascular cells, our data integrated well between techniques, studies, and depots (Fig. 1e, Supplementary Fig. 2b). We identified 12 distinct cell populations (vC0-11), which were broadly split into four major groups including several blood (vC0-03, vC05, vC08-11), and lymphatic (vC06) endothelial cells as well as vascular smooth muscle cells (vC07) and pericytes (vC04) (Fig. 3a, Supplementary Table 2 and Supplementary Data 2). Based on nomenclatures from several scSeq atlases, blood endothelial cells were further classified into the classical capillary (vC0), venous (vC01 and vC11), mixed capillary/venous (vC03), and arterial (vC02) subpopulations (Fig. 3b). Three subtypes (vC08-10) separated from the major endothelial cell populations. Top marker genes for these revealed that vC08 displayed a considerable overlap with the herein-identified myC08 (Jaccard index: 0.33), indicating that they may represent cells with similar function or intermediate cell states. Two of the shared marker genes were TIMP1 and ITLN1 which encode secreted proteins inhibiting neovascularization34,35, suggesting that these cells may exert anti-angiogenic effects (Fig. 3c). vC09 was enriched for genes previously described in “early endothelial progenitor cells” (e.g., TYROBP, FCER1G), a hematopoietic cell type that promotes angiogenesis via paracrine mechanisms (Fig. 3d)36,37. Thus, human WAT may contain vascular cells that either stimulate or inhibit vascularization. One vascular subtype (vC05) expressed multiple marker genes for committed preadipocytes (e.g., CXCL14, APOD, and CFD) and was the only vascular population that expressed PDGFRA (Fig. 3e). In mice, all peri-aortic adventitial fibroblasts are PDGFRA+ and give rise to perivascular adipocytes15. It is therefore possible that vC05 represents an intermediate cell state between endothelial and adipocyte precursor cells. Admittedly, this notion needs further functional studies. Comparisons between depots revealed that all cell types, including vC05, were present in the three regions. However, lymphatic (vC06) and TIMP1-expressing (vC08) endothelial cells were enriched in omental and perivascular WAT (Fig. 3f), observations which were confirmed by deconvolution of bulk transcriptomic data (Fig. 3g).

Fig. 3: The vascular cell class contains mixed and intermediate cell states.
figure 3

a, b Nomenclature and visualizations of selected marker genes for vascular cells (vC0-11), including a UMAPs and b violin plots. ce Multiple UMAPs of marker genes for three subtypes of blood ECs (vC08, vC09, and vC05, respectively). f The proportions (%) of different vascular subtypes in subcutaneous (sc), omental (om), and perivascular (pv) WAT depots. g The proportion of vC06 and vC08 in sc and om depots was supported by deconvolution of bulk transcriptomic data from Arner et al.32. (upper panel) and Krieg et al.33 (lower panel). p values were calculated by two-sided Wilcoxon signed-rank test. EC endothelial cells, VSMC vascular smooth muscle cells. Source data are provided as a Source data file.

Subcutaneous FAPs include subpopulations distinguished by degree of commitment

Although FAPs include fibroblasts and stem cells at different stages of commitment, an established nomenclature for this class is still lacking. This, in combination with their pronounced heterogeneity and depot-specific contribution to WAT expansion (at least in mice38), prompted us to analyze this cell class depot-by-depot. The subcutaneous FAPs separated into 17 clusters (sfC0-16) distributed into four smaller (sfC03, −11, −13, and −16) and one very large (86.5% of all FAP objects) group (Fig. 4a and Supplementary Data 2). The former included mesothelial-like cells (sfC11), CD74+ stromal cells (sfC13) reported to have antifibrotic properties39, and late committed preadipocytes (sfC16). One cluster (sfC03) enriched for cell adhesion molecules and was the only FAP that did not express CD34, PDGFRA, or PDGFRB (Fig. 4b).

Fig. 4: FAPs display different levels of commitment in human WAT.
figure 4

a Nomenclature and proportions for subcutaneous FAPs (sfC0-16) including a UMAP with selected marker genes (left panel) and a stacked bar chart displaying the proportion (%) of different subtypes in subcutaneous white adipose tissue (WAT) (right panel). b, c Selected FAP marker gene expression profiles displayed in UMAPs. d Pseudo-time trajectory analysis initiated from the CD55/PI16-enriched cell cluster (sfC02). Two main trajectories were discovered: route 1 (upper) and route 2 (lower). eg CD55+ positive human adipose-derived stem cells were e analyzed by flow-cytometry and f imaged before/after adipogenic induction in vitro. Nuclei are stained by Hoechst (blue) and lipid droplets by BODIPY (green). Experiment was repeated three times with similar results. Scale bar is 20 μm. g deconvolution of bulk RNAseq data from these cells shows how the expression of marker genes for FAP subtypes in panel a vary during adipogenesis (colors are matched in a and g). h Heat map displaying similarities (Jaccard index) of gene expression profiles between inguinal WAT from mice (P1.1 to P4) and subcutaneous FAPs (sfC0-16). Note that human cells displaying overlap with mouse FAPs are only found in route 1. Explanatory legend is visualized in j. i Same as a, but for omental FAPs (ofC0-14). j Same as in h; but for comparisons of subcutaneous and omental human FAPs. k Flow cytometric analysis of CD55+, APOD+, CD74+, and EZR+ FAPs from stromal vascular cells (CD45, CD31 and CD34+) of subcutaneous and omental WAT, respectively. Percentages represent the frequency of all gated live single cells in the representative sample. APC adipose precursor cell, CPA committed preadipocytes, FAP fibroblast and adipogenic progenitor cells, MSL mesothelial-like cells. Source data are provided as a Source data file.

In contrast to the smaller clusters, the larger group included multiple extracellular matrix- and CD55/PI16-expressing cells (Fig. 4a, c, Supplementary Fig. 4a). The latter marks a universal fibroblast population that can differentiate into more specialized cells40, including adipocytes13. Therefore, we used them as a starting point in pseudo-time analyses which identified two distinct trajectories (routes 1–2) branching into opposite directions at the first step (Fig. 4d). To test if these trajectories were recapitulated during de novo adipocyte formation, we deconvolved bulk RNAseq data from human CD55+ cells undergoing adipogenesis in vitro41 (Fig. 4e, f). Our results show that the CD55/PI16 gene cluster (sfC02) was highly expressed prior to differentiation and was markedly downregulated upon adipogenic induction (Fig. 4g). This was followed by a transient upregulation of route 1-localized clusters (sfC04-05 and sfC07-08), which thus constitute transient cell states. Route 1 ended in sfC0, a cell population enriched for genes that were upregulated at late stages of differentiation (e.g., APOD and CFD) and most likely represent committed preadipocytes (Fig. 4a, d, g). In contrast to route 1, the cell populations in route 2 (sfC01, sfC06, sfC09-10, sfC12, and sfC14-15) could not be detected during in vitro adipogenesis. These results were recapitulated in three additional cell models41,42,43 (Supplementary Fig. 4b), suggesting that under standard in vitro conditions, CD55+ adipocyte precursor cells cannot recapitulate the full FAP heterogeneity observed in subcutaneous WAT.

To improve our FAP annotations and determine inter-species similarities, we systematically compared our findings with scSeq data obtained from mouse inguinal WAT (which is frequently used as a proxy for human subcutaneous WAT)3,5,44. We found that mouse early adipocyte progenitors (P1-1 and P1-2) overlapped with CD55/PI16-expressing adipocyte precursors (sfC02), while more committed preadipocytes (P2-1 and P2-2) resembled sfC0 (Fig. 4h). The gene expression profile of the P3 population, which has been defined as FAPs with anti-adipogenic properties (AREG), showed overlap with one of the transient cell clusters along route 1 (sfC04). There were no overlaps with any of the route 2-localized FAPs indicating that clusters along this trajectory are not overlapping with the Lin-/SCA+ cells pre-selected in the studies by Schwalie et al.3 and Dong et al.5. However, DPP4, which is an established marker of pro-adipogenic FAPs in mouse inguinal WAT44, projected onto route 2 and was enriched in sfC06 and −10 (Supplementary Fig. 4c). This demonstrates that specific FAPs in this trajectory overlap with murine PDGFRB+ cells with adipogenic capacity.

Omental and subcutaneous FAP signatures largely overlap

The omental FAPs separated into 15 clusters (ofC0-14) (Fig. 4i, Supplementary Data 2). To annotate these, we compared their signatures to subcutaneous FAPs. We found that several cell populations shared overlapping marker genes between depots (Fig. 4j, Supplementary Fig. 4c-d). Thus, CD55/PI16-expressing adipose precursors (sfC02) matched ofC07, and APOD/CFD-expressing committed preadipocytes (sfC0) corresponded to ofC02. All FAPs in route 2 showed strong similarities to three omental clusters (ofC05, ofC09, and ofC14). CD74 was highly enriched in sfC13 and ofC10 (Supplementary Fig. 4d), but the overall Jaccard index between these two cell types was modest. The mesothelial-like cell signature in subcutaneous WAT (sfC11) was recapitulated in multiple omental FAPs (ofC03, ofC06, and ofC11-12). Of note, DPP4 was enriched in these omental clusters, but not in the corresponding subcutaneous cells, indicating that this gene marks different subsets of FAPs in the two depots (Supplementary Fig. 4c), a notion previously suggested in mice44. Flow cytometric analysis of subtype-specific surface markers in the stromal vascular fraction of WAT biopsies from subcutaneous and omental WAT confirmed the presence of sf02/ofC07 (CD55+), sfC0/ofC02 (APOD+), sfC13/ofC10 (CD74+) and mesothelial-like cells (EZR+) in both depots (Fig. 4k, Supplementary Fig. 4d). We did not find any expression profiles similar to the fibro-inflammatory progenitors previously described in mouse epigonadal WAT4 (Supplementary Fig. 4e). Data from the perivascular depot separated into eight clusters (pf0-pf07) (Supplementary Fig. 4f). While pfC0 resembled ofC07 (adipose precursors) and ofC12, pfC03 was linked to ofC02 (committed preadipocytes). The remaining clusters did not display strong similarities to any of the other FAP subtypes identified in subcutaneous WAT.

Adipocytes display inconsistent heterogeneity between studies

In contrast to scSeq, snSeq allows for analyses of mature adipocytes. In the present study, we analyzed snSeq datasets from subcutaneous (n = 4), omental (n = 2), and perivascular (n = 1) fat depots. As with the other cell classes, we jointly analyzed objects classified as fat cells first. However, this resulted in poor ARI, kBET, and LISI indices (Fig. 1e, Supplementary Fig. 2b), suggesting that complete data integration was not possible to achieve between studies and depots (Supplementary Fig. 5a). As a next step, we analyzed the depots separately. This, however, did not improve our integration as most clusters separated according to studies rather than biological similarities between datasets (Supplementary Fig. 5b).

A possible reason for incomplete data integration is that there are few common features between studies, even in the low-dimensional space21. To test if data harmonization was influenced by a limited degree of overlap, we next analyzed the studies individually. We focused on data produced by Emont et al.9 and our own results (Massier et al. #1) generated in subcutaneous WAT as they contain the largest number of subjects and objects. We selected the top 50 marker genes for all clusters and found that adipocytes displayed lower fold-changes compared to the other cell classes (Fig. 5a, left panel). This indicates that the degree of cell heterogeneity is less pronounced in adipocytes. However, there may still be consistencies between the studies. We, therefore, transferred cell type classifications between datasets. In comparison with the other cell types where marker genes displayed clear enrichments and overlap between studies, we found a low overlap and a limited set of reproducible adipocyte marker genes across studies (Fig. 5a, right panel). These included a cluster of genes encoding proteins involved in lipid metabolism e.g., ABCD2, ACACB, CD36, DGAT2, GPAM, HACD2, and LPL (Fig. 5b).

Fig. 5: Adipocyte snSeq data display inconsistent marker gene overlaps in WAT.
figure 5

a Single-nucleus sequencing (snSeq) data from Emont et al. (#1) and Massier et al. (#1) were analyzed according to top marker genes for adipocytes, FAPs, vascular, and immune cells, respectively. In comparison with other cell classes, adipocyte marker genes displayed lower fold-changes (left panel) and a limited overlap (right panel). b Representative examples of adipocyte marker genes in subcutaneous WAT displaying overlap between the indicated studies. c Dendrograms of snSeq, bulk RNAseq of isolated mature adipocytes (from the FANTOM5 atlas46,81 and Harms et al.45) and spatial transcriptomics7 (STx) (upper panel). Comparisons of scSeq, snSeq, bulk RNAseq, and STx data of subcutaneous white adipose tissue from the same individual (lower panel). d Heatmap of adipocyte marker genes with a >50-fold enrichment in adipocytes vs. other tissues included in the FANTOM 5 atlas. Results are shown for each study as well as for the combined snSeq data after integration. e Same as in d, but for genes with discordant expression (Δz-score>5) comparing STx and snSeq data. Source data are provided as a Source data file.

Because of the modest heterogeneity and low reproducibility across studies of adipocytes, we benchmarked the snSeq data and our recently published STx-based results7 against rRNA-depleted45 and cap-trapped41 bulk RNAseq of isolated human mature fat cells. This revealed that the STx data were positioned close in space to the bulk results, while all snSeq datasets clustered together with considerably fewer similarities to the other samples (Fig. 5c, upper panel). To test the reproducibility of these results, we generated snSeq, STx, and bulk RNAseq from the same individual and repeated the analyses. This confirmed that STx correlated better with bulk RNAseq data than snSeq (Fig. 5c, lower panel). We characterized this further by using the FANTOM5 expression atlas46, where we identified 218 transcripts that were enriched at least 50-fold in adipocytes compared to other cell types (Supplementary Fig. 5c). These genes included multiple established adipocyte markers and we analyzed their expression levels in data from the three platforms (Fig. 5d). We observed marked variations in the expression between methods. By filtering based on genes that displayed a high discordance (>5 |Δ z-scores|) between snSeq and STx, we found that several STx adipocyte subpopulation marker genes were less expressed or not detected by snSeq, but highly abundant in bulk RNA and STx data (Fig. 5e, Supplementary Fig. 5d). These included marker genes of previously described7 adipocyte subtypes (Fig. 5e). Conversely, PPARG, WDPCP, and PDE3B were higher in snSeq. These types of platform-specific biases may contribute to the low reproducibility and heterogeneity scores as has been shown previously for adipocytes47. In subsequent analyses, we, therefore, pooled the adipocyte snSeq data into one class.

Multiple FAP-relayed signals target M2-like macrophage subpopulations

Having defined cell types present in WAT through our meta-map, we inferred their communication routes using CellChat48. We first summarized the expression of ligands vs. receptors across the different cell types and next identified specific cell-cell interactions via ligand-receptor patterns in the subcutaneous and omental depots (Fig. 6a, b, Supplementary Fig. 6a). Our analysis suggested that FAPs were mainly relaying information, which in turn was primarily received by M2 macrophages. Clustering of pathways by functionality allowed us to link these FAP-myeloid communication routes to for example complement, chemerin, and IL16 signaling (Supplementary Fig. 6b). Another striking finding was that Mox (myC15), which are only present in subcutaneous WAT and characterized by exceptionally high expression levels of various pro-inflammatory cyto- and chemokines (Fig. 6c), both relayed and received CCL, CXCL and TNF signals. Thus, these cells received CCL5 input from different CD8+ T cells (in particular lyC02-03) via CCR1, and signaled to vascular cells (vC01 and vC11) via ligands (CCL2, CXCL2, CXCL8) recognized by ACKR1 (Fig. 6d). For TNF, Mox signaled to adipocytes and myeloid cells via TNFRSF1A, and to CD8+ T cells/NK cells (lyC03, −06, −09) and endothelial cells (vC01-03) via TNFRSF1B (Fig. 6d). The validity of these in silico analyses was supported by the observation that, the KIT receptor was only present in mast cells49 while periostin was solely relayed from pericytes50 (Supplementary Fig. 6c).

Fig. 6: FAPs send multiple signals which are received by M2-like macrophages.
figure 6

a Incoming and outgoing interaction strengths for FAPs/adipocyte, vascular, myeloid, and lymphoid cells in subcutaneous (upper panels) and omental (lower panels) WAT. Selected cell types are indicated. b Subcutaneous (upper) and omental (lower) cell-cell communication predictions for clusters identified in Supplementary Fig. 6a. Lines indicate interactions between cell types where the strength is proportional to the line width and the color defines the sending subpopulation. c Violin plots for selected marker genes enriched in the Mox cluster (myC15). d Contribution of each ligand-receptor (L-R) pair to the overall cell-cell communication strength. Blue bars are strongly contributed by Mox (myC15). e Predicted cell-cell interactions for the indicated L-R pairs. Line widths and colors indicate signaling strengths and sending subpopulations, respectively. Cell types are in numerical order as shown in the left-most panel. Note that the myeloid and lymphoid clusters ends with mast cells and B cells, respectively. Source data are provided as a Source data file.

Spatial distribution of FAPs suggest distinct adipogenic niches

To add a spatial dimension to our meta-map, we re-analyzed previously generated STx data of subcutaneous human WAT from ten individuals7 using six deconvolution tools: DestVI, stereoscope, Tangram, RCTD, cell2location and SPOTlight. Although all six frameworks provided similar results (Supplementary Fig. 7a), we opted to use cell2location as it has been shown to be a robust deconvolution tool17. In contrast to FAPs, lymphocytes, and adipocytes, we found that vascular and myeloid cells were concentrated in specific areas of the tissue (Fig. 7a). For the latter, these regions included classical monocytes (myC14) and M1-like macrophages (LAMs [myC02], Mmes [myC10] and Mox [myC15]) but were devoid of M2-like macrophages (Fig. 7a). We confirmed clustering of LAMs in specific areas of WAT by immunofluorescence (Fig. 7b), thereby confirming previous findings in mice6. To complement these studies, we systematically identified within-spot colocalization patterns. For this, we correlated deconvolution scores for each cell type across all spots and identified three clusters (Fig. 7c, Supplementary fig. 7b). These included endothelial cells as well as two groups of FAPs. Further analyses showed that one FAP subtype (sfC12) was enriched in areas close to endothelial cells (Fig. 7d). This contrasted with the other FAPs, which displayed strongly negative relationship to vascular cells. In addition, another FAP subtype (sfC08) was found close to LAMs (myC02) (Fig. 7d). These associations were present in multiple individuals (Supplementary Fig. 7c) and were confirmed by immunofluorescence (Fig. 7e). Altogether, this suggests that FAPs have specific tissue distributions, possibly to form different types of adipogenic niches.

Fig. 7: WAT contains niches populated by specific sets of cells.
figure 7

a Representative sections from two subjects displaying areas densely populated by myeloid cells (left panels). The indicated regions are magnified where the hematoxylin & eosin stain is shown in the middle and the Visium slide myeloid score is shown. Deconvolution scores for myeloid subpopulations in the inlay regions are shown in the right panels. Boxplots are presented as interquartile range plus median and Tukey whiskers; scale bar is 100 µm. b Representative immunostaining of human subcutaneous white adipose tissue incubated with antibodies targeting LAM marker proteins TREM2 and CD9, respectively. Nuclei were stained with Hoechst. The experiment was repeated three times with similar results. Scale bar is 100 μm. c Pair-wise correlation heatmap displaying within-spot associations between cellular subpopulations. Full heatmap is shown in Supplementary Fig. 7b. d Representative sections displaying the distributions of selected subpopulations of FAPs (sfC08 and −12), myeloid (myC02), and vascular cells (vC01). Scale bar is 500 µm. e Representative immunostaining of human subcutaneous white adipose tissue incubated with antibodies targeting the sfC12 marker protein SLIT2 as well as the endothelial protein CD31. Nuclei were stained with Hoechst. The experiment was repeated seven times with similar results. Scale bar is 50 μm in the merged panel and 10 μm in the inlay. Source data are provided as a Source data file.

Two clusters of cells are reciprocally associated with metabolic health

A major weakness of most transcriptional studies with single-cell resolution is that the clinical relevance of the identified cell types is difficult to determine. To address this limitation, we used single-cell marker genes to deconvolve subcutaneous WAT bulk transcriptomic data from eight independent studies comprising a total of 864 individuals. As displayed in Fig. 8a and further detailed in Supplementary Table 3, together these cohorts included both adult men and women with a broad range in age, BMI, waist-to-hip ratio, circulating triglycerides, HDL-cholesterol, and leptin levels as well as insulin sensitivity estimated by HOMAIR. We also retrieved data on fat cell volume and lipolysis (basal and isoprenaline-stimulated), which are two measures linked to metabolic health51. By clustering the correlations between cell types and the mentioned parameters, we identified major trends in the data. Thus, while cluster A included six cell types that associated positively with a metabolically beneficial profile, cluster B contained 15 cell types that correlated negatively with the same parameters (Fig. 8b). Cluster C included 37 cell types which displayed weaker links to metabolic states. Of note, age was not associated with any specific cell type (Fig. 8b).

Fig. 8: Deconvolution of transcriptomic data reveals cluster-specific clinical associations.
figure 8

a Bulk transcriptomic data from eight cohorts were retrieved, the distribution in age, BMI, and HOMA-IR are shown in the left panel. Summary statistics are detailed in the right panels. b Heatmap displaying the association between individual cell types (denoted by numbers and color according to the classification in Figs. 24) with: anthropometric measures, HOMA-IR, circulating levels of HDL-cholesterol, triglycerides, and leptin (all in the fasted state), fat cell volume as well as adipocyte lipolysis (basal, isoprenaline-stimulated and isoprena-stimulated/basal). Three main clusters (A–C) were identified where cluster A and B are magnified in the right panel. c Representative Forest plots displaying the associations between individual measures and cell types. Data are shown as correlations with 95% confidence intervals for each study and summarized using both common and random effects models. For all displayed data, p values were <0.0001. d Stability of clusters A and B were determined in the two indicated cohorts where WAT bulk transcriptomes were generated before and two years following bariatric surgery. e Effects of weight loss induced by bariatric surgery in two cohorts. Panels display deconvolution scores for the indicated cell subpopulations. p values were calculated by two-sided paired sample t test (n = 15; Petrus et al.54 and n = 37; Kerr et al.53) and boxplots are presented as interquartile range plus median and Tukey whiskers with individual, paired data points. Source data are provided as a Source data file.

In more detailed analyses, we found that cluster B was overrepresented by immune cells, e.g., LAMs (myC02), Mmes (myC10), DC2 (myC03), six out of nine M2-like macrophages (myC0-01, myC07-08, myC11 and myC13), early differentiated CD8+ (lyC02) and CD16+ NK cells (lyC06) (Fig. 8b, c). These results are in line with previous data demonstrating that specific immune cell subtypes are enriched in states of insulin resistance and obesity1. In contrast to cluster B, cluster A enriched for FAPs and vascular cells (Fig. 8b, c). The latter included blood endothelial capillary cells (vC0) and could be due to capillary rarefaction, a phenomenon characterized by reduced capillary beds previously reported in WAT from people with obesity52. For FAPs, we observed that CD55/PI16-expressing adipose precursors (sfC02) were the only cells present in cluster B while some intermediate/late states (sfC04, −14, and −16) were found in cluster A. This suggests that adipogenesis may be impacted by metabolic health at several different levels. To reveal associations between adipocytes and clinical measures, we used STx marker genes as this platform reflects the transcriptional profiles of adipocytes more closely than snSeq. In concordance with previous results obtained in a small cohort7, we found that AdipoPLIN correlated negatively with BMI, insulin resistance and circulating leptin levels while AdipoLEP was positively associated with all these measures (Supplementary Fig. 8a). In contrast, AdipoSAA displayed weak correlations with all investigated parameters.

Finally, to test if clusters A and B were impacted by weight changes induced by bariatric surgery, we retrieved WAT transcriptomic data from two of the studies where subjects were followed two and five years post-operatively (n = 52)53,54. Our results revealed that in both cohorts, clusters A and B were stable and normalized by weight loss (Fig. 8d, Supplementary Fig. 8b). One exception was blood endothelial capillary cells (vC0), which remained unaltered by weight loss (Fig. 8e). These data53 also allowed us to assess effects of weight regain comparing follow-ups at two and five years. As displayed in Supplementary Figure 8c, subdividing subjects into tertiles based on long-term weight regain or stability showed that except for vC0, several cell subtypes followed the changes in body weight. Altogether, these observations suggest that the cellular landscape of WAT is dynamic and mirrors alterations in fat mass.


Our meta-analysis integrates existing and newly generated single-cell data with bulk sequencing of in vitro adipogenesis and intact WAT from large clinical cohorts. By overlapping this with spatial transcriptomics and data from human and murine single-cell resources generated in different organs, we provide a meta-map of cell types and their spatial organization in human WAT. Altogether, this allowed us to define >60 distinct cell types including immune cells with diverse activation states, intermediate vascular cell types with hybrid transcriptional profiles, and FAPs displaying distinct tissue localization and different levels of adipogenic commitment.

To create a cellular nomenclature across adipose depots, we included data from subcutaneous, omental, and perivascular WAT in our meta-analysis. Based on depot comparisons, we found that although proportions differed, most cell subpopulations were present in all three regions. This was also true for FAPs, even though they have been suggested to contribute to depot-specific differences in tissue growth. For example, both adipose precursors and committed preadipocytes were found in all three sites. However, within subcutaneous WAT, the FAPs displayed distinct localizations where some were found close to vessels and others were adjacent to specific macrophages. The latter is of interest given that our ligand-receptor analyses suggested that FAPs relayed multiple signals to myeloid cells. Apart from these quantitative and microarchitectural aspects, we also observed that a limited number of cell types were unique to either region, including Mox and a few M2-like macrophage subtypes. Although we corroborated these results by deconvolving bulk transcriptomic data from paired samples of subcutaneous and omental WAT, they need to be validated in additional cohorts and the function of these cells needs to be determined. In addition, the spatial analyses presented herein were only performed in subcutaneous WAT and should be followed-up in other depots.

In our meta-analysis, adipocytes comprised ~20% of the WAT cell population and displayed a transcriptional fingerprint that was distinct from FAPs, immune and vascular cells. However, in contrast to these cell classes, adipocyte data were exclusively generated using snSeq and the results displayed less pronounced and reproducible heterogeneity between studies. A possible reason for this lack of consistency is that snSeq preferentially detects nascent and long transcripts47. This is further supported by our benchmarking of technical platforms where snSeq data, in comparison to STx results, associated poorly with the transcriptional signature of isolated fat cells. In fact, numerous adipocyte subtype marker genes identified by STx (e.g., LEP, PLIN4, SAA1, RBP4)7 were not, or very weakly, detected by snSeq. We therefore conclude that combined analyses using different technical platforms are required to confidently identify adipocyte subtypes.

Although we have combined studies to obtain data at the single-cell level from approximately 100 samples, we have not determined the influence of age, anthropometric measures, and disease states on these results. Instead, we created a cartography of cells present in WAT and used this framework to deconvolve bulk transcriptomic results from over 860 samples. This expands previous efforts55,56 and allowed us to link the identified cell populations to multiple clinical and WAT parameters. More specifically, we show that CD55/PI16-expressing adipose precursors as well as a large set of immune cells, including LAMs and Mmes, were enriched in individuals with markers of a pernicious metabolic phenotype, i.e., subjects with large fat cell volume, high waist-to-hip ratio, high HOMA-IR and impaired lipid mobilization. Conversely, a group of intermediate FAPs and capillary endothelial cells associate negatively with the same parameters. Additional analyses in cohorts before and after bariatric surgery, revealed that multiple cell clusters enriched in people with obesity and insulin resistance are normalized upon weight loss. Together, these data extend previous results on WAT expansion by confidently identifying specific cell types linked to increased inflammation as well as attenuated adipogenesis and vascularization. Nevertheless, a limitation with the present work is that we did not investigate longitudinal data following other types of interventions including, life-style-related changes, and/or pharmacological treatments.

Taken together, in this meta-analysis of 17 datasets and >800 bulk transcriptional profiles from eight clinical studies, we have comprehensively defined the cellular composition of human WAT in health and metabolic disease. Thus, by jointly analyzing data from multiple types of studies, we have created a framework that is easily accessible and includes additional tools for WAT analyses such as new models in CellTypist. We, therefore, provide a rich resource to facilitate future studies of specific WAT-resident cell types in relation to aspects not investigated herein, such as ethnic differences and the impact of therapeutic interventions.


Clinical studies

Inclusion and Ethics

This study was performed in agreement with the Declaration of Helsinki. Studies of cohorts presented for the first time herein (Massier et al. #1–4), were approved by the regional ethics boards in Stockholm (clinical trials identifiers: NCT01785134 and NCT01727245) and Leipzig (approval numbers: 159-12-21052012 and 004/21-ek) and explained in detail to each participant who gave informed written consent. For retrospective analyses of published data, the studies have been approved by the respective ethical boards where informed written consent was obtained from all participants. For data detailed in Fig. 8d, e, primary outcomes for both clinical trials (NCT01785134 (DEOSH) and NCT01727245 (NEFA)) have been described at and previously published7,53,57,58,59, and the studies have been completed.

Sample collection and preparation

Samples of cohort 1 and 3 were collected in Stockholm (Sweden) and processed by T.W. in Zürich (Switzerland). Adipocyte nuclei were isolated following a modified nuclear isolation protocol60. In total, 50 mg of fresh or frozen WAT was first minced into 1–3 mm pieces and then homogenized on ice in 0.1% CHAPS in CST buffer supplemented with 0.2 U/μl RNAase inhibitor (RI) using a Dounce homogenizer. After homogenization, samples were left on ice for five minutes following which PBS supplemented with BSA and 0.2U/μl RI was added to obtain a final concentration of 1% BSA. The lysates were filtered through 40-μm cell strainers and centrifuged at 500 × g for five minutes at 4 °C. The nuclei pellets were resuspended with 1% BSA in PBS supplemented with 0.2 U/μl RI and centrifuged again at 50 × g for five minutes at 4 °C. This step was repeated once more. After the final resuspension, nuclei were filtered through 20-μm cell strainers and loaded directly on a 10X Chip G. 10X-libraries were prepared with the Chromium Single-Cell v3.1 reagent kit following the manufacturer’s protocol (10X Genomics). Suspensions containing around 1200 nuclei per μl were loaded on Chip G followed by reverse transcription to obtain cDNA, which subsequently was amplified and used for library construction. After preparation, the libraries were sequenced on a NovaSeq 6000 platform (Illumina). For data analysis, the human genome assembly GRCh38.p13 was used. Mapping was performed using 10x Genomics Cell Ranger (v6.0.2). CellBender (v0.2.0)61 was used on ‘raw_feature_bc_matrix’ to remove empty droplets and ambient RNA; scDblFinder (v1.5.11)62 was applied to exclude potential doublets. Downstream analyses were performed as all other included studies (see below).

Samples of cohort 2 and 4 were collected and processed in Leipzig (Germany) by P.A.N.N. Paired samples of omental and subcutaneous WAT were obtained from female patients with obesity undergoing different bariatric surgery procedures (Roux-en-Y Gastric Bypass and sleeve gastrectomy). After collection, samples were washed using PBS, placed on ice until the end of the surgery procedure, snap-frozen in liquid nitrogen and stored at −80 °C for later use. Sequencing analyses were performed on isolated nuclei as detailed above. Single-nuclei RNAseq. libraries were generated using the Chromium Single Cell 3′ v3 assay (10× Genomics) and sequenced with NovaSeq 6000 S4 flow cell platform (Illumina). Raw reads were aligned to the human genome (hg38) and cells were called using 10x Genomics Cell Ranger (v.6.0.1).

Collection of publicly available datasets

Peer-reviewed WAT datasets containing either snSeq, scSeq or STx with publicly available results published until 31.03.2022 were included in the present meta-analysis (Supplementary Table 1). If not present in the published articles, the corresponding authors were contacted by email to obtain information regarding sample numbers and gender as well as ranges of age and BMI.

Re-analysis of publicly available datasets

Publicly available datasets were re-analyzed with Seurat v4.1.016 in R v4.1.263 (pipeline available via GitHub []). Mitochondrial and hemoglobin genes, as well as further confounding transcripts, including MALAT2 and NEAT1, were removed prior to analysis. All data were normalized using sctransform64 and corrected for subject effects using Harmony v0.1.025 before performing independent component analysis65. Clusters were determined using FindNeighbors and FindClusters in Seurat after generating Uniform Manifold Approximation and Projection (UMAP) data projections using RunUMAP66. Spatial data were analyzed as described recently7.

Cluster classification and annotation

Cluster classification for individual cohorts was assisted by a supervised network analysis where each cell cluster was represented as a node. Nodes were connected by edges by calculating overlap percentages of positively enriched marker genes (FDR-adjusted p value <0.05). Minimum requirements for edge connections were >15% genes overlapped in one of the two nodes and >5% in both nodes. Based on this, Jaccard similarity scores were calculated, and edges were created based on the five highest values. The network was built in R v4.1.2. using igraph v1.2.11 and visualized in Cytoscape 3.7.1 using Rcy3 v2.14.167. Integrated cluster annotations were performed manually using multiple reference datasets (Supplementary Table 2).

Data integration and bechmarking

We evaluated the following integration tools: rPCA provided with Seurat16, Harmony25, BBKNN as well as scVI. In addition to identification/clustering of prominent marker genes in the integrated data, benchmarking included calculations of ARI coefficients, LISI scores, kBET acceptance rates (1- rejection rate) for integration across methods, depots and cohorts. ARI was calculated using the adj.rand.index function in pdfCluster v1.0-3 by supplying factors of either depots, methods or cohorts in addition to the identified clusters. Average LISI scores were estimated using lisi v1.0 compute_lisi command adding embeddings of the UMAPs along the meta data. kBET scores were calculated using kBET v0.99.6 and the respective integrated reductions (e.g., Harmony or scVI) in the Seurat object. Data from different depots or methods within the same study were treated as independent cohorts, which was used as batch variable to integrate over, thereby correcting for differences in methods and sequencing platforms. Of note, all cohorts except for Acosta et al12. (<0.5% of analyzed cells), were sequenced in 3’ direction, thereby facilitating integration (Supplementary Table 1). Guided by our testing, we opted to integrate data individually for adipocytes, FAPs, immune and vascular cells. Based on these results, scVI integration using the 2000 most variable features (using VariableFeatures in Seurat) was applied in the final analysis. Seurat objects were transcribed into anndata objects using the sceasy v0.0.6 function convertFormat and scVI was run using default settings in R with reticulate v1.24 (setup_anndata, SCVI, train, get_latent_representation). Subsequent subcluster analyses were performed based on different depots (omental, perivascular, and subcutaneous WAT) or lineages (e.g., myeloid and lymphoid cells). When comparing similarity between individual data sets using Jaccard index, marker genes were selected based on a log2 fold-change > 0.5, and adjusted p value <0.05.


Deconvolution of bulk transcriptomic data of human WAT was performed using BisqueRNA v1.0.568 using the marker gene approach with a minimum gene count of six. To validate depot differences, two cohorts32,33 with available data from omental and subcutaneous WAT were used. We also retrieved sequencing as well as clinical phenotype data from six additional published datasets53,54,69,70,71,72 (Supplementary Table 3). Deconvolved data were compared in a meta-approach using Hmisc v4.6–0 to calculate Spearman correlation and meta v5.2–0 to calculate and visualize summarized results using both common and random models73.

Flow-cytometry staining and analysis of data

The procedures for preparing cells for flow-cytometry have been described in detail elsewhere74. In brief, stromal vascular fractions (SVF) were thawed and stained with different antibody combinations prior to analysis with a flow cytometer. Washing steps were performed with wash buffer (PBS supplemented with 0.5% BSA [#A4503, Sigma-Aldrich] and 2 mM EDTA [#E7889, Sigma-Aldrich]) and the cells were centrifuged at 200 × g for 10 minutes. To remove red blood cells, samples were incubated in red blood cell lysis buffer (15.5 mM NH4Cl, 0.57 mM K2HPO4, 0.01 mM EDTA × 2 H2O in PBS) for 6 minutes and subsequently washed with wash buffer. Samples were divided into aliquots containing approximately one million cells and subsequently stained with either (i) a lymphocyte antibody cocktail, (ii) an antibody panel for myeloid cells and fibroblasts combined, or (iii) a separate antibody panel for FAPs. The antibody-fluorochrome conjugates used are listed in Supplementary Table 4. Staining with the lymphocyte antibody cocktail was performed by resuspending the cells first in stain buffer (1% FBS, 2 mM EDTA in PBS) with CCR7-APC-Cy7 and incubating for ten minutes at 37 °C. Subsequently, the surface marker antibodies were added to the cell suspension and the incubation was carried out at room temperature for 20 minutes. The cells were washed once with stain buffer, centrifuged at 400 × g for five minutes, and resuspended into 1× fixation/permeabilization solution (#00-5223-56 and #00-5123-43, eBioscience). After incubating the cell suspensions for 30 minutes at room temperature in the dark, they were washed with permeabilization buffer (#00-8333-56, eBioscience), centrifuged at 400 × g for 5 minutes, and resuspended with an intracellular staining cocktail in 1× permeabilization buffer for 30 minutes at room temperature in the dark. Lastly, the cells were washed with permeabilization buffer and fixed with 1% PFA (#22023-20 ml, Biotium) for 15 minutes prior to analysis with a BD Symphony analyzer equipped with 355, 405, 488, 561 and 640 nm lasers and DIVA software (BD Biosciences). Fixable live/dead Aqua stain (#L34957, Invitrogen) was included in the surface stain cocktail and used for dead cell exclusion. Gating was performed as outlined in Supplementary Figure 9a. The cells stained with the myeloid panel were incubated in the antibody cocktail for 30 minutes at 4 °C in the dark, washed once with wash buffer, resuspended into flow buffer (0.1% BSA, 2 mM EDTA in PBS), and analyzed immediately with the flow cytometer with a previously set compensation and gating setup. Fixable live/dead yellow dye (#L34968, Invitrogen) was used to exclude dead cells. The gating setup preceding the UMAP analysis is outlined in Supplementary Figure 9b. Cells stained with the FAPs panel were stained in a similar fashion as with the myeloid panel. 7-AAD (#559925, BD Biosciences) was used as live/dead exclusion dye and fluorescence minus one (FMO) controls were used for the gating (Supplementary Fig. 9c). The results were analyzed with FlowJo Software v10.7.1 and v10.8.0 (BD Biosciences). Dimensionality reduction for analyzing the myeloid cells was performed using the UMAP FlowJo plugin v3.1. For this analysis, 1267 myeloid cells from each individual were exported into a new file, barcoded and concatenated. FlowJo Phenograph plugin v3 was applied for unsupervised clustering, with the optimal k-nearest neighbors implemented automatically.


Lipid droplet staining before and after adipogenesis

For imaging of CD55+ cells before and after adipogenesis, cells were fixed in 4% PFA for 15 minutes at room temperature and washed twice with PBS. Lipid droplets and nuclei were stained with PBS containing BODIPY 493/503 (1:2500, ThermoFisher) and Hoechst 33342 (1:5000, #ab228551, Abcam) for 15 minutes. Cells were then washed four times with PBS and images were acquired using CREST V3 confocal system (Crest Optics) mounted on an inverted Nikon Ti2 microscope equipped with a Prime BSIexpress sCMOS camera (pixel size 6.5 μm) from Photometrics. A Nikon 20x/0.75 air objective was used to acquire images.

Immunostaining of LAMs (myC02)

For immunofluorescence, WAT samples were fixed in cold 4% paraformaldehyde (PFA) for 24 hours, embedded in paraffin and then sliced into 6 μm thick sections. Antigens were retrieved by heating up the sections for 20 minutes in 10 mM citrate buffer pH 6.0 (tri-sodium citrate in distilled water) using a microwave. The samples were subsequently washed three times with PBS containing 0.3% Triton X-100 and blocked for one hour at room temperature in PBS containing 0.1% BSA, 0.1% Triton X-100, 50 mM glycine, and 0.05% Tween. Primary antibodies (anti-TREM2 [1:100, #13483-1-AP, Proteintech] and anti-CD9 [1:500, #60232-1-Ig, Proteintech]) were diluted in 0.1% BSA, 0.1% Triton-X-100, 10 mM glycine and 0.05% Tween in PBS and incubated with the sections overnight at 4 °C. This was followed by three wash steps with PBS containing 0.3% Triton X-100 and incubations for 10 minutes at room temperature. After this, slides were incubated with secondary antibodies (donkey α-rabbit conjugated with Alexa Fluor 594 [1:200, #A-21207, Thermofisher] and donkey α-mouse conjugated with Alexa Flour 488 [1:200, # A-21202, Thermofisher]) diluted in 0.1% Tween in PBS for an hour at room temperature. Following three wash steps with PBS containing 0.3% Triton-X-100, 100 µL Sudan Black B in 70% ethanol was added to each section and the samples were incubated for two minutes at room temperature. The slides were thereafter rinsed in PBS and incubated with Hoechst 33342 (1:10000) diluted in PBS for ten minutes to stain nuclei. Prior to mounting in DAKO Fluorescence mounting media (S302380-2, Agilent Technologies), the samples were washed with PBS and swirled in distilled water.

Immunostaining of FAPs (sfC12) and endothelial cells

Subcutaneous WAT blocks embedded in optimal cutting temperature compound were sliced into 16 μm thick sections, which were then fixed in 4% PFA for five minutes at room temperature and washed twice with PBS. After this, glycine was added (100 mM final concentration) for ten minutes and slides were blocked for one hour in PBS containing 1% BSA, 0.3% Triton X-100, and 10% normal donkey serum. Subsequently, the slides were incubated overnight in 4 °C with primary antibodies (anti-SLIT2 [1:200, #20217-1-AP, Proteintech] and anti-CD31 1:100, #M082329-2, Agilent Technologies]) in incubation buffer containing 5% normal donkey serum, 1% BSA and 0.3%Triton X-100 in PBS. Slides were washed three times with PBS containing 0.3%Triton X-100 for five minutes. They were thereafter incubated with secondary antibodies (donkey α-rabbit conjugated with Alexa Fluor 594 [1:200, #A-21207, Thermofisher] and donkey α-mouse conjugated with Alexa Flour 488 [1:200, #A-21202, Thermofisher]) in incubation buffer for 1 hour at room temperature. Additional washing steps were performed with PBS supplemented with 0.3%Triton X-100. Then, Hoechst 33342 (1:10,000) diluted in PBS was added for 10 minutes to stain nuclei. Prior to mounting in DAKO Fluorescence mounting media, the samples were washed with PBS and swirled in distilled water.

For immunostaining of both myC02 and sfC12/endothelial cells, images were acquired using the NIS Elements software, a CSU-X1 spinning disk confocal (Yokogawa) mounted on an inverted TiE microscope (Nikon) equipped with a ×1.2 magnification lens and a Kinetix back-illuminated sCMOS camera (pixel size 6.5 μm QE95%) (Photometrics). A Nikon ×20/0.75 air objective was used to acquire images.

Bulk RNA sequencing of in vitro adipogenesis

To annotate FAP clusters, we retrieved bulk sequencing data of human subcutaneous adipocyte precursor cells undergoing adipogenesis from four model systems: (i) adipose-derived stem cells41, (ii) primary SVF-derived cells from subcutaneous WAT41, (iii) Simpson-Golabi-Behmel (SGBS) syndrome cells42, and (iv) human multipotent adipose-derived stem (hMADS) cells43. Scores were generated for each FAP cluster at each timepoint of adipogenesis by calculating expression fold-changes of top 30 marker genes over the background of the respective bulk datasets.

Comparing adipocyte-enriched transcripts between platforms

To identify adipocyte-enriched genes unrelated to single-cell approaches, we retrieved bulk RNAseq data of human isolated subcutaneous adipocytes from the FANTOM5 database46. This was compared to pseudobulk data, which was retrieved using AverageExpression in Seurat, of both individual snSeq and STx studies, as well as the integrated snSeq data. We calculated transcript fold-changes in these samples compared to the complete dataset. In total, 218 genes with >50-fold enrichment in fat cells compared to all other samples were considered to be adipocyte-enriched. We compared expression levels of these genes between snSeq and STx platforms after z-score normalizing each dataset. As an extra control, we included an additional bulk RNAseq dataset of human isolated subcutaneous adipocytes45. By filtering for genes with a |Δz-score| >5, we identified genes with the strongest differences between snSeq and STx.

Spatial deconvolution

We performed cell2location v0.175 for deconvolution of STx by using a subset of annotated scRNA-seq data as input. We subsampled the single-cell data set for each cell type according to following criteria: (i) If a cell type had ≤1500 cells, select all cells; (ii) if a cell type had >1500 cells, randomly select 1500 cells. We compared cell2location with five additional spatial deconvolution methods: stereoscope v0.2.076, SPOTlight v0.1.077, RCTD v2.0.078, Tangram v1.0.279, DestVI v0.16.280 to validate the robustness of deconvolution results. Default parameter settings were used for deconvolution analysis.

We used spot-wise Pearson correlation between the estimated cell type abundances to quantify cell type colocalization pattern. The Pearson correlations were computed across all spots for each pair of cell types and each subject. High positive correlation indicated that two cell types exhibited similar spatial distributions, while negative correlation suggested distinct spatial distributions between two cell types.

Cell-cell communication analysis

Cell-cell communication analysis was performed using CellChat v1.4.048 based on the curated ligand-receptor interaction database (CellChatDB). In brief, subcutaneous and omental normalized gene expression matrices were provided as input to CellChat, respectively. The total numbers of interactions and interaction strengths were computed by the computeCommunProb function, and the communication probabilities for each cell signaling pathway were calculated by computeCommunProbPathway function.


As detailed under each subheading above, statistics were performed in R v4.1.2 or GraphPad Prism v9. Data distributions were tested by Kolmogorov-Smirnov and Shapiro-Wilk tests and parametric vs. non-parametric tests were used accordingly. Spearman’s rank correlation was used to assess relationship of two continuous variables. For analyses requiring family-wise error rate corrections, p value <0.05 after correcting for false discovery rate using Benjamini-Hochberg was considered significant.

Reporting summary

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