Study design
Single-nucleus RNA sequencing (snRNA-seq) was done in subcutaneous abdominal AT from 25 people with obesity before and after marked WL (paired samples) and in 26 healthy lean controls. Two lean samples were removed as described below. Obese case and control groups were well matched for age, *** and ethnicity. Spatial transcriptomics was done in equivalent groups (n = 4 per condition), as were histological cross-validation studies (n = 4–5 per condition). All molecular phenotyping was done after overnight fasting. The WL interval was a minimum of 5 months (median 7, range 5–18 months). Median percentage WL was 22% (range 13–33%). Primary snRNA-seq data were integrated with previously published snRNA-seq data in whole subcutaneous AT from nine obese and four lean people to increase obese and lean cell numbers and improve cell annotation11. Participant characteristics are provided in Extended Data Table 1.
Sample collection
The AT samples were obtained intra-operatively from morbidly obese individuals (BMI > 35 kg m−2) undergoing laparoscopic bariatric surgery (gastric byp*** or gastric sleeve) and healthy controls (BMI < 26 kg m−2) undergoing non-bariatric laparoscopic abdominal surgery51. Subcutaneous AT was collected from abdominal surgical incision sites. Follow-up subcutaneous AT samples were collected from the peri-umbilical region using needle biopsy more than 5 months after WL intervention. Whole AT samples were snap frozen at collection and stored at −80 °C for future use. Participants were unrelated, between 20 and 70 years of age, from a multiethnic background and free from systemic illnesses not related to obesity. People with treated type 2 diabetes were excluded because of the potential effects of hypoglycaemic medications on AT metabolism. Metabolic characteristics were collected at baseline and follow-up. All participants gave informed consent. The study complies with all relevant ethical regulations and was approved by the London – City Road and Hampstead Research Ethics Committee, United Kingdom (reference 13/LO/0477). Human tissue validation also used samples from the Imperial College Healthcare Tissue Bank, approved by Wales REC3 to release human material for research (reference 17/WA/0161).
Nucleus isolation
The AT nuclei from individual participants were pooled for library preparation and sequencing to increase efficiency, cohort diversity and study power. Pooled samples were separated by condition to avoid cross-over (4–5 samples per pool; a total of 6 pools per group). Sample pools for each experimental group were processed through to sequencing in lean–obese–WL trios to minimize between-group batch effects. For each participant sample, nucleus extraction was done using a modified version of a previously described protocol52. In brief, frozen human AT (about 100 mg) was cut into pieces of less than 0.2 cm and homogenized with 1 ml ice-cold lysis buffer (Tris-HCl 10 mM (Invitrogen, 15567-027), NaCl 10 mM (Invitrogen, AM9760G), MgCl2 3 mM (Invitrogen, AM9530G), 0.1% NP40 (BioBasic, NDB0385), 0.2 U µl−1 RNase inhibitor (Roche, 03335402001)) in a gl*** dounce homogenizer (Merck, T2690/P0485/P1110, 15 strokes, loose then tight pestles) on ice. After homogenization, samples were transferred through a 100 µM cell strainer (Greiner Bio-One, 542000) into a prechilled tube using ART wide-bore tips (Thermo Scientific, 2079 G). The filtered homogenate was then transferred to 1.5 ml low DNA-bind tubes (Sarstedt, 72.706.700) and centrifuged at 500g and 4 °C for 5 min. After lipid/supernatant removal, the nuclei pellet was resuspended in 1 ml wash buffer (PBS with 0.5% BSA (Invitrogen, AM2616) and 0.2 U µl−1 RNase inhibitor), transferred to new 1.5 ml low DNA-bind tubes and recentrifuged at 500g and 4 °C for 5 min. After repeat lipid/supernatant removal, the nuclei pellet was resuspended in 300 µl wash buffer containing DAPI (Thermo Scientific, 62248) at 0.1 µg ml−1 to stain nuclei, and filtered through a 35 µM cell strainer into a fluorescence-activated cell sorting (FACS) tube (Falcon, 352235) on ice. At this point, the isolated nuclei from 4–5 samples from the same experimental group were pooled before sorting by flow cytometry.
FACS was used to clean up residual debris and lipid from isolated nuclei and to remove doublets. Pooled nuclei were sorted on a BD FACS Aria SORP. The sheath tank was bleach cleaned before each run and nuclease-free PBS (1×) (Invitrogen, AM9625) was used as sheath fluid. A 405 nm laser was used to excite DAPI, and emission was collected using a 450/50 nm bandp*** filter. Single nuclei were selected by gating on the first DAPI-positive band on the DAPI versus forward scatter (FSC) plot and then subsequently gating on side scatter (SSC) versus FSC and FSC A versus FSC H to ensure better debris and doublet removal. All sorts were performed using an 85 μm nozzle. The sorted nuclei were collected into a BSA- and RNase inhibitor-rich collection buffer (70 µl of PBS with 1.375% BSA and 2.15 U µl−1 RNase inhibitor) in low DNA-bind tubes kept at 4 °C. After sorting, nuclei were centrifuged at 500g for 5 min at 4 °C to pellet. Supernatant was removed to leave about 40 µl, which was used to resuspend pellets with a wide-bore pipette tip.
Single-nucleus library preparation and next-generation sequencing
Pooled single-nucleus suspensions were used to generate barcoded single-nucleus libraries for next-generation sequencing. For each pool, 5,000–10,000 nuclei were co-encapsulated with 10x barcoded gel beads to generate gel beads in emulsion (GEMs) using a 10x Chromium Controller and a 10x Genomics Single Cell 3′ v.3.1 kit, according to the manufacturer’s instructions. After GEM-RT and clean-up, the quantity and fragment size distribution of amplified cDNAs derived from barcoded single-cell RNAs were ***essed using an Agilent 2100 Bioblockyzer High Sensitivity DNA ***ay. From this cDNA, snRNA-seq libraries were constructed and sequenced (Illumina NextSeq2000) in three batches, containing equal numbers of obese, lean and control library pools, to minimize between-group batch effects. Each unique library was sequenced to a minimum depth of more than 20,000 paired-end reads per nucleus (read 1, 28 base pairs (bp) and read 2, 90 bp, with unique dual 10-bp indexes). Raw sequencing data were demultiplexed and blockysed using CellRanger v.5.0.1 and bcl2fastq v.2.20.0. Libraries were demultiplexed using CellRanger mkfastq based on the sample indices (allowing one mismatch), and the CellRanger count pipeline was used to perform alignment against human genome GRCh38 (using STAR), filtering and counting unique molecular identifiers (UMIs) (including introns).
Single-nucleus quality control
For each pooled library, raw count matrices from CellRanger were processed using CellBender53 (–epochs 150-200, –learning-rate 0.0001-0.00005) to remove ambient RNA molecules and random barcode swapping, and filter inferred cells. The number of expected cells was based on CellRanger estimations. Filtered count matrices were processed separately using Seurat54 and SeuratObject. Low-quality cells with low read or gene counts (less than 1,000 UMIs or less than 400 genes), low complexity (log10(genes per UMI) < 0.85) and high mitochondrial or ribosomal fractions (greater than 5%) were removed from each pooled dataset. Clean libraries were normalized and transformed (sctransform v.2 regularization55) to stabilize count variances. Potential doublet nuclei were detected using three approaches: expression-based DoubletFinder56, using doublet estimates from genotyping to set the expectation; genotype-based, Vireo57 (details below); and iterative clustering and detection of clusters with high expression or genotype-based doublet fractions. Assigned doublets, ambiguous cells and doublet clusters were then removed and singlet-only datasets were retransformed. Participant-level annotation information from genotyping was then added to generate high-quality cell datasets.
Participant annotation from genotype information
Genotype information present in the RNA sequencing reads was aligned to existing genome-wide genotyping to attribute specific cells to specific participants in each sample pool. Participant-level genotype data were generated from whole blood using Illumina Infinium OmniExpress-24 v.1.2 bead chips. Directly genotyped single-nucleotide polymorphisms (SNPs) with call rates of less than 90%, minor allele frequency of less than 0.01, Hardy–Weinberg equilibrium P < 1 × 10−6, SNPs on *** chromosomes and duplicated SNPs were removed. After quality control, 649,007 SNPs were taken forward for imputation. SHAPEIT58 (v.2.r900) was used to infer haplotypes, and imputation was done in IMPUTE2 (v.2.3.2)59 using a 1,000 genomes reference panel phase 3 (all ancestries). Each chromosome was divided into 5-megabase chunks for imputation and merged at the end. A random seed was supplied automatically. An effective population size (Ne) reflecting genetic diversity was 20,000, as recommended when using a multi-population reference panel. After imputation, genotype data were available for 81,656,368 SNPs.
Cell-level SNP data were generated for each pooled sample using cellsnp-lite60 (using the combined imputed SNP list as the reference). Cell-level SNP data were then intersected with participant-specific genotype references in Vireo57 to identify variants that segregated the samples, and we used these variants to demultiplex participant specific cells, ambiguous cells and doublets. A range of cellsnp-lite MAF settings were tested and MAF > 0.05 was selected to maximize singlet recovery. Participant-level cell annotations were then incorporated into pre-cleaned high-quality cell datasets.
Integration
High-quality, doublet-removed cell libraries containing participant-level annotations were then integrated to a unifying atlas. Two samples, one with very high lymphocyte counts and one with very few cells, were removed at this stage, leaving 24 samples in the lean group. A further 13 whole subcutaneous AT samples from obese and lean people in a previously published dataset11 were also incorporated in the integration phase to increase cohort diversity and improve cell annotation. Of note, only samples meeting the following criteria were selected: whole tissue; nucleus only; subcutaneous depot; and BMI < 26 or BMI > 30 kg m–2. Previously published samples were individually reprocessed from raw counts using thresholds equivalent to our own datasets.
To integrate our dataset with the previously published dataset11, we updated the gene IDs from the latter to match the same Ensembl release. Both datasets were then normalized to 10,000 counts per nuclei before proceeding with downstream blockysis. To minimize any sample-driven effect for cell-type identification, we took a three-step approach. First, we regressed out the effects of number of original counts, as well as the percentage of mitochondrial and ribosomal genes. Then we calculated the PCA space on the highly variable genes, detected by Scanpy61, followed by correction of the PCA space with Harmonypy62 using samples as batches. Finally, we used BBKNN63 with samples as a batch to identify neighbourhoods.
Analysis overview
Cell type and state annotation was done in the combined (our own and that from ref. 11) integrated dataset. Primary exploratory blockyses were performed in our own dataset, which was processed in experimental group trios (lean–obese–WL) to minimize batch effects and comprised paired obese–WL samples and age-, ***- and ethnicity-matched lean controls. Differential neighbourhood abundance and expression blockyses between groups (in which biological, technical and batch covariates could be adjusted for) were repeated using the combined dataset to verify reproducibility.
Cell annotation
We identified the main cell types with unbiased clustering, using a low-resolution (0.15) Leiden algorithm, and each cell type was annotated according to known markers. To identify cell states, we isolated the barcodes for each of the main cell type identities, except for mast and lymphatic endothelial cells, owing to low numbers. Each cell type was then reintegrated and reclustered twice, as described above. First, we used a high-resolution Leiden (1.2 or higher) to identify barcodes that contained a mixed signature, with markers of different lineages. These barcodes were flagged as ‘un***igned’ and were excluded from any downstream blockysis. Then, we removed these barcodes and proceeded with the second round of reintegration and clustering. Resolution varied across cell types (0.65 or higher), with myeloid cells requiring the highest Leiden to identify rare, known cell types (2.25). Clusters that were similar to each other and had no unique identifiable features between them were merged. Cell states were annotated based on a mix between unbiased and known markers. To identify unbiased markers, we used Scanpy’s rank_gene_groups function to perform a Wilcoxon test.
Compositional blockyses
To blockyse changes in cellular composition, we used a neighbourhood graph-based approach in Milo64. We performed comparisons of lean–obese and WL–obese groups, adjusting for biological covariates in the lean–obese blockyses. Neighbours were recalculated with BBKNN using samples as a batch, restricted to the comparison groups (lean–obese and WL–obese). To blockyse global shifts, we used Milo on all cell types together and within each cell type to blockyse shifts in cell state composition. Only neighbourhoods containing at least 90% of a single cell type or state were considered neighbourhoods, and those with a spatial FDR < 0.1 were considered significant. Fasting insulin adjusted for BMI abundance blockyses were carried out in steady state lean and obese samples, using lean–obese neighbourhoods, adjusting for biological covariates.
Metabolic blockyses
The metabolic profiles of different cells were inferred using flux-based blockysis modelling in COMPASS65. For this, we created an expression matrix for every main cell type, consisting of the mean expression of each gene per sample. These matrices were then used to run COMPASS. Statistical blockysis to compare conditions was performed with a Wilcoxon test for every reaction, using their COMPASS score. COMPASS plots consisted of both positive and negative reactions grouped by their defined subsystem.
Differential expression blockyses
Differential expression blockyses were carried out between obese cases and controls, and between obese–WL pairs, in Nebula66 using negative binomial mixed-effect models to correct for subject- and cell-level correlation structure. In all comparisons, further thresholding was applied (mitochondrial fraction less than 1% and ribosomal fraction less than 1%) to minimize false discovery, and fractions of mitochondrial and ribosomal counts were incorporated as technical covariates; in obese–lean comparisons, age, *** and ethnicity were included as covariates; in obese–WL comparisons of paired samples, biological covariates were not included. Statistical significance was inferred at P < 0.05 Bonferroni corrected for obese–WL pairwise comparisons (where power was higher) and FDR < 0.01 for lean–obese comparisons. Cell type and state differences were examined using Scanpy’s rank_gene_groups function to perform a Wilcoxon test, as were spatial differences in gene expression within cell types between conditions. Amphiregulin (AREG), which is known to be secreted67, was added to the curated secretory protein list from the Human Protein Atlas41 for comparisons in stressed and basal cells.
Inference of regulatory networks
To infer regulon activity, we used the Python implementation of the SCENIC68 pipeline (pySCENIC). The expression matrix used consisted of nuclei from all 3 conditions, downsampled to the same number of nuclei (20,000 each). Genes that were expressed in all nuclei, or in less than 5% of nuclei for any given cell state, as well as mitochondrial, ribosomal, haemoglobin, non-coding, antisense, contig and microRNA genes, were also removed from the blockysis. For TF binding sites, we used the Encode 2019/06/21 ChIP-seq hg38 refseq-r80 10 kilobases up and down database. Only regulons with a minimum of five target genes were considered. Analyses in adipocytes were restricted to all TF genes and genes in dysregulated metabolic pathways from COMPASS. Differential networks between cell states and within cell types between conditions were identified by comparing cell-level network scores between groups (non-parametric Wilcoxon rank-sum test). Significance was inferred at P < 0.05 (Bonferroni corrected). Within a cell state, fold changes were scaled for visualization.
Cell–cell communication
We used CellChat69 to infer intercellular communication, based on known receptor–ligand interactions. For the purpose of this blockysis, to compare the differences between each condition, cellular communication was inferred for each condition separately. Each condition was down-sampled to 20,000 barcodes to avoid any confounding effects arising from higher cell numbers in obese and lean groups, and cell types with very low numbers were removed because these cell types often have higher mean gene expression owing to low cluster background. To blockyse the differential communication between two conditions, we used the rankNet function in CellChat to obtain overall signalling differences, as well as pairwise comparison with each cell type as a sender and as a receiver. To blockyse communication at the cell state level, we performed a condition-agnostic blockysis to maintain cell states with low numbers of nuclei. For intra- and inter-niche communication blockyses, because of the lack of most ligand–receptor pairs in the Xenium gene panel, we imputed spatial data using ENVI70. This was done for each condition separately, training on the single-nucleus data for each condition. We did this step ten times and averaged the results in a final imputed expression matrix because of the stochastic nature of imputation. Imputed genes with low expression (below the mean across all genes, the gene-level quality control) and those with below the mean for that gene (cell-level quality control) were removed.
Metabolic and senescence scores
Gene list scoring was done in Scanpy using the score_genes function, with the normalized ln expression and a control size of 50. Senescence signatures were obtained from MSigDB71,72. Housekeeping genes were obtained from the 20 most stable human transcripts in the Housekeeping Transcription Atlas73, supplemented with commonly used housekeeping genes (RRN18S, ACTB, GAPDH, PGK1, PPIA, RPL13A, RPLP0, ARBP, B2M, YWHAZ, SDHA, TFRC, GUSB, HMBS, HPRT1 and TBP). The BCAA score was performed using the genes ***ociated with the respective pathways on COMPASS.
Pathway blockyses
Pathway blockyses of differentially expressed genes were done in ClusterProfiler74 using the Over Representation Analysis and MSigDB71,72 datasets (H, C2 and C5) as inputs. All genes present in the comparison datasets were used as background. Significant pathway enrichment was inferred at FDR < 0.01.
Tissue processing for spatial transcriptomics and histology
Frozen stored AT samples (stored at −80 °C) were directly thawed in a 4% paraformaldehyde solution and kept at 4 °C for 24 h. Samples were then transferred to a 70% ethanol solution and stored until paraffin embedding. Ethanol-dehydrated samples were cleared with xylene, infiltrated with molten wax using the Sakura Tissue Tek VIP6 vacuum infiltration processor and embedded in paraffin using the Sakura Tissue Tek TEC5 embedding system.
Spatial transcriptomic preparation
Slide preparation
Formalin-fixed paraffin-embedded (FFPE) blocks were stored at 4 °C. Xenium slides stored at −20 °C were equilibrated to room temperature for 30 min before sectioning. The FFPE blocks were rehydrated in an ice bath with distilled water for 10–30 min and sectioned at 5 µm thickness. Sections were floated in a 42 °C water bath and slides containing tissue sections were incubated at 42 °C for 3 h and then dried overnight at room temperature in a desiccator. Slides were kept at 4 °C in a desiccator until use. All histology was done in RNase-free conditions using sterilized equipment.
Technical pilot
A technical pilot was done on a single frozen stored AT sample separated into three sections for fixation at 24 h, 48 h and 72 h to evaluate the effects on tissue integrity (H&E) and transcript recovery using the 10x Xenium Human Multi-Tissue and Cancer Panel (P/N 1000626), with two slides and one tissue section for each fixation time/slide (Institute of Developmental and Regenerative Medicine (IDRM), Oxford).
Panel design
A 10x Xenium Human Multi-Tissue and Cancer Panel (P/N 1000626) supplemented by 100 custom genes was selected to annotate prominent cell types, states and effector pathways identified in single-nucleus datasets.
Xenium in situ transcriptomics
The FFPE tissues were blockysed on a 10x Xenium Analyser following 10x Genomics Xenium in situ gene expression protocols CG000580, CG000582 and CG000584. In brief, 5-µm FFPE tissue sections on Xenium slides were deparaffinized and permeabilized to make the mRNA accessible. Gene panel probes were hybridized for 20 h overnight followed by washing, ligation of probe ends to targeted RNAs, generating circular DNA probes with high specificity. Rolling circle amplification was used to generate hundreds of copies of gene-specific barcodes for each RNA-binding event, resulting in a strong signal-to-noise ratio. Background fluorescence was quenched chemically to mitigate tissue auto-fluorescence. Tissues sections were stained with DAPI nuclear stain and Xenium slides were loaded onto the Xenium instrument for imaging and then decoding of image data to transcripts. Secondary blockysis to segment cells and ***ign transcripts was performed on-instrument (Xenium Analyser v.1.7.1.0). Xenium Explorer was used to evaluate the initial data quality and visualize morphology images, transcript localization at subcellular resolution, segmentation and data clustering.
Post-Xenium processing
After Xenium in situ transcriptomics, slides were kept in PBS and stored at 4 °C for up to 24 h. For immunofluorescence staining, slides were washed three times in PBS for 5 min and then incubated in CF 594 wheat germ agglutinin (1:200; Biotium, 29023-1) for 20 min. Slices were then rewashed three times with PBS, and tissue stained with DAPI (1:5,000; Thermo Scientific, 62248) for 10 min at room temperature. Finally, sections were rewashed as before and then mounted using antifade medium Vectashield (Vector Laboratories, H-1000). Full slide scans for the immunofluorescence channels were performed at 20× magnification using a ZEISS Axio Scan.Z1 slide scanner.
Spatial data blockysis
Xenium data were blockysed by three different methods, depending on the purpose of the blockysis. Regardless of the type of blockysis, only transcripts with a quality value higher than 35 were considered.
To plot transcript and score densities, regardless of cell type we took a segmentation-free approach creating 50-µm bins using the transcript coordinates provided by Xenium. Only bins that contained more than ten transcripts were kept for downstream blockysis.
For cell-type identification, we took the nucleus segmentation from Xenium and ***igned only transcripts within 2 µm of each nucleus (selected to maximize recovery of transcripts but minimize the capture of known cross-contaminating marker transcripts from adjacent cells, designated nucleus segmentation). The resulting matrices were then imported into Scanpy for blockysis. Here, only nuclei with more than 40 transcripts were kept for downstream blockysis. Clustering was performed similarly to the single-nucleus data, with Harmonypy62 and BBKNN63 used to correct batch effects in the PCA and neighbourhoods, respectively. However, here gene expression was scaled using Scanpy’s61 scale function to give more weight to low-expression genes. A low-resolution Leiden algorithm was then used to identify the main cell types, and cell states were identified by reintegrating and reclustering each of these cell types individually. Clusters were labelled to match the single-nucleus reference. Ambiguous clusters were labelled ‘un***igned’. To delineate rarer LAM subtypes in the spatial dataset we used CellTypist for label transfer75, creating a model trained on the single-nucleus LAM subtypes and applying a ‘best match’ prediction on the MYE2 LAM spatial cluster.
To correlate genes with adipocyte size, we performed a semi-manual segmentation using ImageJ, designated boundary segmentation. WGA staining, performed after the Xenium run, was aligned to the Xenium data using the DAPI channel as a guide and utilized for segmentation. To avoid any issues arising for multiple adipocytes being merged in the segmentation, we manually closed some gaps where the WGA staining was not strong enough to be detected by the binary threshold of ImageJ. We then used the Analyse particles function of ImageJ to detect each object and measure the area and centroid coordinates. Furthermore, we created a separate table with coordinates for each pixel contained in each object. To ***ign transcripts to the ImageJ objects, and to remove any noise derived from other cell types, we first removed any transcript that was ***igned to non-adipocytes during the nuclei segmentation. We then created a distance tree between the remaining transcript coordinates and the pixel coordinates obtained for every ImageJ object. This was achieved using the KDTree function from Scipy’s spatial module. Adipocyte transcripts that were found on the cell boundary were ***igned to the closest adipocyte(s) (any adipocyte within 2 µm of the nearest segmented pixel). Only objects with an area greater than 1,000 µm2 and less than 25,000 µm2 were considered as adipocytes for this blockysis. As larger objects were found to have higher probability of capturing more transcripts, gene expression was normalized to the total number of counts per cell. Clustering was done as described above, using a high resolution to identify and then remove fine clusters containing contaminating transcripts from other cell types. A Spearman correlation was done to investigate which genes correlated with adipocyte area.
Finally, to cluster cells in spatial niches, we made use of Scipy’s KDTree function to create a distance tree between every cell in each sample. We then created a neighbourhood matrix by counting, for each cell, the number of proximate cells (within 300 µm) at a cell state level. Because adipocyte sizes increased in obesity, cells in lean samples had roughly twice the number of neighbouring cells that cells in obese samples did. To prevent this from biasing the niche clustering, the neighbourhood matrix was normalized such that each cell was represented by the percentage of neighbouring cells in each cell state. To cluster cells into niches, we created an anndata object of the neighbourhood matrix for use in Scanpy and corrected for batch effects with Harmony and BBKNN before Leiden clustering. Very similar clusters, driven by small fluctuations, were merged into the AD niche.
Tissue immunohistochemistry
The FFPE blocks were sectioned at 5 µm thickness for immunohistochemistry and immunofluorescence. Sections were deparaffinized and hydrated, and then heat-mediated antigen retrieval was done in an EDTA-based pH 9.0 solution. Endogenous peroxidase was quenched with 3% hydrogen peroxide. Sections were incubated with rabbit monoclonal to p21 Waf1/Cip1 (1:50 dilution; Cell Signalling, 2947, clone 12D1), followed by anti-rabbit IgG conjugated with polymeric horseradish peroxidase linker (25 μg ml−1; Leica Bond Polymer Refine Detection, DS9800). DAB was used as the chromogen and the sections were then counterstained with haematoxylin and mounted with DPX. Immunohistochemistry was performed on a Leica BOND RX. To evaluate p21-positive cells, full virtual slide scans were loaded into QuPath 0.5.1 (ref. 76) and the positive cell detection module was used to count the total haematoxylin and DAB-positive nuclei in two slices per sample. The fraction of p21-positive cells relative to the total cell number was then calculated for each slice, and the mean was used for between-group blockyses.
Tissue immunofluorescence
Tissue sections of 5 µm were deparaffinized by submerging three separate times in Histoclear (National Diagnostics, HS-200) for 5 min and then rehydrated by submerging in a series of graded alcohol solutions of decreasing concentrations for 5 min each. After rehydration, antigen retrieval was done by heating the samples in 10 mM sodium citrate buffer, pH 6 (Abcam, ab64236) for 5 min in a decloaking chamber (Biocare Medical, DC2012-220V). The sections were then permeabilized in 0.2% Triton X (Sigma-Aldrich, X100-500mL) in PBS for 10 min and subsequently blocked in 1× ACE blocking solution (Bio-Rad, BUF029) for 30 min. After blocking, sections were incubated in primary antibody solutions diluted in 0.5× block ACE at 4 °C overnight: anti-NAMPT (1:200, Affinity Biosciences, DF6059); anti-TREM2 (clone D8I4C, 1:400, Cell Signalling, 91068); or anti-TLR2 (clone TL2.1, 1:400, Invitrogen, 14-9922-82). After primary antibody removal, the tissue was washed in PBS and then incubated with secondary antibody, goat anti-rabbit Alexa Fluor 488 (1:200, Invitrogen, A11034), donkey anti-rabbit Alexa Fluor Plus 488 (1:250, Invitrogen, A32790) or goat anti-mouse Alexa Fluor Plus 647 (1:250, Invitrogen, A32728) in 0.5× block ACE for 45 min at room temperature. For NAMPT, sections were incubated with DyLight 594 Lycopersicon Esculentum Lectin (1:250, Invitrogen, L32471) for 20 min (room temperature), rewashed with PBS and then stained with a DAPI solution (1:5,000, Thermo Scientific, 62248) for 10 min at room temperature. For TREM2/TRL2 at CLS, only DAPI was used. Finally, sections were washed and mounted using antifade medium Vectashield (Vector Laboratories, H-1000). For each sample, representative images were taken at 40× magnification (NAMPT) or 20× (CLS) using a Leica SP8 DLS confocal microscope. Image blockysis was done in QuPath 0.5.1 (ref. 76). To quantify the NAMPT:lectin ratio, the positive pixel area of the NAMPT and lectin channels was measured in two z-stack maximum projection images per sample using the pixel cl***ifier module. Measurement precision was evaluated between two images per sample (to confirm low within-sample variability) and the mean sample intensity was used for between-group blockysis.
Macrophage isolation and HPG uptake
We used a modified SCENITH-based approach to evaluate human macrophage metabolic pathways ex vivo77,78. Fresh subcutaneous AT was cut into approximately 2-mm pieces with 30 ml HBSS (Gibco, 14175-053) in a 50 ml tube, washed and collected using a 100 µM cell strainer. Tissue was digested for 20 min at 37 °C with 3 mg ml−1 collagenase II (Sigma C6885) in methionine-free RPMI (Sigma, R7513), 65 mg l−1 l-cystine dihydrochloride (Sigma, C6727), 1× GlutaMAX (Gibco, 35050061), 10% dialysed fetal bovine serum (FBS, Gibco, A3382001). Digested tissue was filtered through a 100 μm strainer and digestion was terminated by adding methionine-free RPMI containing 10% FBS, followed by centrifugation (300g at 4 °C for 7 min). After resuspension in methionine-free RPMI (65 mg l−1 cystine, 10% FBS, 1× glutamax), cells were plated (160 µl) into wells on a 96-well V-bottomed plate. Cells were methionine starved for a further 15 min (total starvation of 45 min including digestion and isolation) before treatment with inhibitors or control media (40 µl) for 15 min. The four treatments were medium, 2-deoxy-d-glucose (2-DG; 100 mM final concentration; Sigma, D8375), oligomycin (2 µM final concentration; Sigma, 495455) and 2-DG plus oligomycin (100 mM and 2 µM final concentration, respectively). Homopropargylglycine (HPG; Cayman Chemical, 11785) was then added to wells at a final concentration of 500 µM and incubated for 30 min to initiate cell HPG uptake. An additional well received cells and media but no HPG and no treatment (click chemistry negative control). After HPG uptake, cells were stained with zombie aqua live/dead stain (1:500 in PBS; BioLegend, 423101) for 20 min at room temperature in the dark, washed with PBS and then fixed with 2% PFA for 15 min.
Click chemistry, staining and FACS blockysis
Fixed cells were permeabilized (0.1% saponin and 1% BSA in PBS) for 15 min, washed with click buffer (100 mM Tris-HCl, pH 7.4; Invitrogen, 1556-027) and incubated with Fc receptor blocker (25 µg ml−1 in PBS; Fc1, BD Biosciences, 564765) for 10 min. Cells were rewashed and incubated in 100 µl of click reaction mix in the dark at room temperature for 30 min. Click reaction mix was made sequentially, adding CuSO4 (final concentration, 0.5 mM; Sigma, 209198), THPTA (final concentration, 2 mM; Antibodies.com, A270328), sodium ascorbate (final concentration, 10 mM; Sigma, A7631) and then AZDye 555 (final concentration, 25 µM; Vector Laboratories, CCT1479) to click buffer (final concentration, 100 mM Tris-HCl).
After click chemistry exposure, cells were washed using FACS buffer (PBS, 1% BSA, 5 mM EDTA, 25 mM HEPES) and stained with antibody mix (FACS buffer, anti-CD45 FITC (1:20; H130; BioLegend, 304006), anti-FOLR2 APC (1:20; 94b/FOLR2; BioLegend, 391705), anti-CD9 APC-fire (1:20; H19α; BioLegend, 312114), Fc block reagent (25 µg ml−1)) at 4 °C in the dark for 30 min. After rewashing, cells were filtered (35 µM cap strainer) for FACS blockysis.
Spectral flow cytometry was done on a Sony ID7000 in standardization mode. The ID7000 software was used to calculate distinct spectral signatures for each fluorochrome based on single stained controls. Fluorochrome signatures, together with autofluorescence signatures identified in unstained aliquots of each sample using the AF finder software feature, were used to unmix the signals in fully stained samples with the built-in WLSM algorithm. Unmixed signals were used for gating (Extended Data Fig. 2i and Supplementary Fig. 1) and blockysis of median fluorescence intensity in FlowJo.
In vitro stress studies
Immortalized human adipose-derived stromal cells (Bmi-1/hTERT, iHASC) were acquired from Applied Biological Materials (T0540). For differentiation experiment cells, iHASC were seeded in six-well plates. Differentiation was induced at confluence using growth medium (DMEM/F-12 (Gibco, D8437), 10% FBS (Gibco, F7524), 2 ng ml−1 rhbFGF (Z101455), 1% gentamicin (G255)) supplemented with 10 µg ml−1 insulin (Actrapic, Novo Nordisk), 500 µM 3-isobutyl-1-methylxanthine (Sigma, I5879), 1 µM dexamethasone (Sigma, D4902) and 2 µM rosiglitazone (Sigma, R2408) for 15 days. Etoposide (Sigma-Aldrich, E1383) was used to induce the DNA damage stress response79. From day 1 to day 5 of differentiation, cells were treated with DMSO (Fisher-Scientific, BP231100) (control) or etoposide 5 µM or 10 µM. Medium was refreshed every 3 days. For stress-marker experiments, undifferentiated cells were seeded in 96-well plates and treated with DMSO control or etoposide (5 µM and 10 µM) at 80% confluence.
O-Red-oil (ORO) staining was performed as previously described51. In brief, cells were fixed with formalin, washed with sterile water, treated with 60% isopropanol and stained with ORO solution (Sigma, O0625) and DAPI (1:5,000). After washing, stained cells were imaged on an Evos m7000 (Thermo Scientific) capturing a minimum of 100 fields at 20× magnification per well. Marker quantification was done in Qupath; nuclear segmentation was done using the cell-detection module in the DAPI channel. Mean ORO intensity was quantified in a 15 µm radius to each nucleus. Positive cells were called empirically at a threshold greater than 32.2, 8-bit depth. The proportion of ORO-positive cells to the total number of nuclei was calculated.
For stress-marker quantification, after etoposide and media treatment, 96-well plates were fixed in 10% formalin for 10 min and then washed with PBS. The following primary antibodies were used for staining: anti-STAT3 (clone 124H6, 1:500; Cell Signalling, 9139S) and anti-JUN (clone 60A8, 1:500; Cell Signalling, 9165S). Otherwise, staining procedures used the same steps, reagents and concentrations as for tissue immunofluorescence. After staining, wells were kept in PBS and imaged using a high-throughput fluorescent microscope IN Cell Analyzer 2500HS (Cytiva, objectives 20× for JUN and 40× for STAT3). Positive cells were determined using IN Carta image blockysis software (v.1.14), based on the nuclear fluorescence intensity for the target protein (empirical positive threshold for JUN, greater than 396.9, and STAT3, greater than 505.3, 16-bit depth). Data were expressed as the percentage of positive cells (JUN or STAT3) of the total number of nuclei.
Statistics and reproducibility
Unless otherwise stated, significance was inferred at P < 0.05 for single-variable tests and FDR < 0.05 for multiple-hypothesis tests. For spatial datasets, where representative images are provided, all blockyses were repeated in n = 4 samples per group. For histological verification, where representative images are shown, all blockyses were repeated in n = 4–5 samples per group.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
