Adipose tissue retains an epigenetic memory of obesity after weight loss

Data reporting

No statistical methods were used to predetermine sample size. The experiments were not randomized, and the investigators were not blinded to allocation during experiments and outcome assessment.

Clinical sample acquisition

Human AT biopsies were obtained from three independent studies: MTSS, LTSS and NEFA.

MTSS

The MTSS samples comprised samples from omental visceral AT biopsies obtained in the context of a two-step BaS treatment, which included a sleeve gastrectomy as the first step (T0) and laparoscopic RYGB as the second step (T1)¹⁶. Individuals with syndromal, monogenic, early-onset obesity or individuals with other known concurrent diseases, including acute infections or malignant diseases, were not included in the study. Individuals were not required to adhere to any specific diet before or after surgery but received individual dietary recommendations during regular visits in the obesity management centre. Insulin resistance was determined using a hyperinsulinaemic–euglycaemic clamp technique or the homeostatic model assessment for insulin resistance (HOMA-IR). Only biopsies from individuals that (1) lost 25% or more of BMI between T0 and T1 (Extended Data Table 1), (2) had undergone surgery at the Municipal Hospital Karlsruhe or Municipal Hospital Dresden-Neustadt, (3) were not diagnosed with diabetes, and (4) did not receive any glucose-lowering medication were used for snRNA-seq in this study. AT samples were collected during elective laparoscopic abdominal surgery as previously described⁶³, snap-frozen in liquid nitrogen and stored at −80 °C. Body composition and metabolic parameters were measured as previously described⁶⁴. Samples of healthy individuals who were not obese were collected during routine elective surgeries such as herniotomies, explorative laparoscopies and cholecystectomies at the same hospitals. The study was approved by the Ethics Committee of the University of Leipzig under approval number 159-12–21052012 and was performed in agreement with the Declaration of Helsinki.

LTSS

The human study samples comprised samples from omental visceral and subcutaneous abdominal AT, collected in the context of a two-step BaS treatment. Following an initial sleeve gastrectomy (T0), a laparoscopic RYGB was made in the second step (T1)¹⁶. Individuals with syndromal, early-onset obesity or individuals with other known concurrent diseases, including acute infections or malignant diseases, were not included in the study. Individuals did not adhere to any specific diet before or after surgery but received individual healthy diet recommendations during regular visits in the obesity management centre. Insulin resistance was determined using HOMA-IR. Only individuals that (1) lost 25% or more of BMI between T0 and T1 (Extended Data Table 1), (2) had undergone surgery at the Leipzig University Hospital, (3) were not diagnosed with diabetes and (4) did not receive any glucose-lowering medication were included. AT samples were collected during elective laparoscopic abdominal surgery as previously described⁶³, snap-frozen in liquid nitrogen and stored at −80 °C. Body composition and metabolic parameters were measured as previously described⁶⁴. Samples from healthy donors that were not obese were collected during routine elective surgeries (herniotomies, explorative laparoscopies, cholecystectomies) at the same hospital. The study was approved by the Ethics Committee of the University of Leipzig under approval number 159-12–21052012 and performed in agreement with the Declaration of Helsinki.

NEFA study

The NEFA study (NCT01727245) comprises samples from subcutaneous abdominal AT from individuals before and after RYGB surgery, as well as healthy controls who had never been obese^8,65. For this, biopsies were obtained under local anaesthesia before (T0) and 2 yr post-surgery (T1). Only samples from individuals that (1) lost more than 25% BMI between T0 and T1, (2) were not diagnosed with diabetes at T0 and T1 and (3) did not take glucose-lowering medication were included in the present study (Extended Data Table 1). Samples from control subjects were obtained from individuals that were BMI- and age-matched to RYGB patients at T1 as reported previously⁸. AT samples were handled as reported before⁶⁵, snap-frozen in liquid nitrogen and stored at −80 °C. The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the Karolinska Institute, Stockholm (approval number 2011/1002-31/1).

Mice

All mice were kept on a 12-h/12-h light/dark cycle at 20–60% (23 °C) humidity in individually ventilated cages, in groups of between two and five mice, in a pathogen-free animal facility in the SLA building at ETH Zurich. The health of mice was monitored closely, and any mouse exhibiting persistent clinical signs of ill health or distress was excluded from this study. The 16- and 29-week-old male C57BL/6J diet-induced obesity mice (catalogue no. 380050) and diet-induced obesity control mice (catalogue no. 380056) were obtained from The Jackson Laboratory and were kept on the respective diets for another 2 weeks until tissue harvest or diet switch. Different mice were used for insulin tolerance tests and glucose tolerance tests. AdipoERCre⁶⁶ and NuTRAP⁶⁷ mice were maintained on a C57BL/N background. Homozygous NuTRAP and AdipoERCre mice were bred to generate AdipoERCre x NuTRAP mice. AdipoERCre x NuTRAP mice were kept on HFD or chow diet for 12 or 25 weeks before tissue harvest or diet switch. The HFD used contained 60% (kcal%) fat (diet no. 2127, Provimi Kliba); the low-fat chow diet used contained 10% (kcal%) fat (diet no. 2125, Provimi Kliba). During the WL period both experimental groups received chow diet (diet no. 3437, Provimi Kliba). All animal experiments were approved by the Cantonal Veterinary Office, Zurich.

Tamoxifen application

The 4–5-week-old AdipoERCre x NuTRAP mice were gavaged two times with 1 mg of tamoxifen dissolved in corn oil. Tamoxifen was washed out for 2 weeks before starting HFD.

Physiological measurements

Glucose tolerance test

Mice were fasted for 6 h during dark phase before administration of 1 g of glucose per kg body weight by intraperitoneal injection. Blood was collected from the tail vein at 0, 15, 30, 60, 90 and 120 min and blood glucose concentrations were measured using an Accu-Check Aviva glucometer.

Insulin tolerance test

Mice were fasted for 6 h during dark phase before administration of 1 U per kg body weight of human insulin (insulin Actrapid HM, Novo Nordisk) by intraperitoneal injection. Blood was collected from the tail vein at 0, 15, 30, 60, 90 and 120 min and blood glucose concentrations were measured using a Accu-Check Aviva glucometer.

In vivo indirect calorimetry

Measurements were obtained from one 8-cage and one 16-cage Promethion Core Behavioral System that were in the same room. Mice were habituated to the system for 36 h before measurements were started.

Live body composition

Mice were fasted for 6 h during dark phase. Live mouse body composition was measured with a magnetic resonance imaging technique (EchoMRI 130, Echo Medical Systems). Fat and lean mass were analysed using EchoMRI 14 software.

Fasting insulin

EDTA plasma was isolated from fasted blood samples (fasting 6 h). Insulin was measured with Ultra Sensitive Mouse Insulin ELISA Kit (Crystal Chem, catalogue no. 90080).

Postprandial insulin

EDTA plasma (50 µl) was thawed on ice and used in a custom U-PLEX assay (Meso Scale Discovery) according to the manufacturer’s instructions. A Mesoscale SI 2400 was used to read the plate.

Postprandial leptin

EDTA plasma (50 µl) was thawed on ice and used in a custom U-PLEX assay (Meso Scale Discovery) according to the manufacturer’s instructions. A Mesoscale SI 2400 was used to read the plate.

Liver triglycerides

First, 50 mg of frozen liver was homogenized in 1 ml of isopropanol, lysed for 1 h at 4 °C and centrifuged for 10 min at 2,000g at 4 °C. The supernatant was transferred into a new tube and stored at −80 °C until use. Triglyceride levels were measured by mixing 200 µl of reagent R (Monlab, catalogue no. SR-41031) and 5 µl of sample or Cfas calibrator dilutions (Roche, catalogue no. 10759350; lot no. 41009301), then incubating for 10 min while shaking at room temperature and measuring optical density at 505 nm (OD₅₀₅) with a plate reader (BioTek Gen5 Microplate Reader).

Cell culture experiments

AT digestion

AT was minced and digested at 37 °C while shaking in collagenase buffer (25 mM NaHCO₃, 12 mM KH₂PO₄, 1.3 mM MgSO₄, 4.8 mM KCl, 120 mM NaCl, 1.2 mM CaCl₂, 5 mM glucose, 2.5% BSA; pH 7.4) using 2 mg of collagenase type II (Sigma-Aldrich, catalogue no. C6885-1G) per 0.25 g of tissue. After 30 min tissues were resuspended, and for ingAT digestion continued for 15 min whereas epiAT was processed immediately. An equal volume of growth medium (DMEM (Gibco, catalogue no. 31966021), 10% FBS (Gibco, catalogue no. 10500-064, Lot no. 2378399H), 1% penicillin-streptomycin (Gibco, catalogue no. 15140-122)) was added and digested tissue was centrifuged for 4 min at 300g, and the floating fraction was transferred into a new Falcon tube and kept at 37 °C. The SVF was resuspended in 5 ml of erythrocyte lysis buffer (154 mM NH₄Cl, 10 mM NaHCO₃, 0.1 mM EDTA, 1% penicillin-streptomycin), incubated at room temperature for 5 min, filtered through a 40 µM mesh filter and centrifuged for 5 min, 300g. The SVF was resuspended in growth medium and counted.

SVF differentiation

A total of 10,000 cells were plated into one well of a collagen-coated (Sigma-Aldrich, catalogue no. C3867) 96-well plate and kept in culture until they reached confluency, with media change every 48 h. At 2 d post-confluence, medium was changed to induction medium (DMEM, 10% FBS, 1% penicillin-streptomycin, 10 nM insulin (Sigma-Aldrich, catalogue no. I9278), 0.5 mM 3-isobutyl-1-methylxanthin (Sigma-Aldrich, catalogue no. I7018-1G), 1 µM dexamethasone (Sigma-Aldrich, catalogue no. D4902), 1 µM rosiglitazone (Adipogen, catalogue no. AG-CR1-3570-M010)). After 48 h medium was changed to maintenance medium (DMEM, 10% FBS, 1% penicillin-streptomycin, 10 nM insulin). Medium was changed every 48 h for 8 d.

AdipoRed assay

The SVF was cultured as described and controls were either kept in growth medium or only maintenance medium without induction. On day 8 after induction, cells were washed twice in PBS, and AdipoRed (Lonza, catalogue no. LZ-PT-7009) reagent was used according to the manufacturer’s instructions and read with a plate reader (BioTek Gen5 Microplate Reader).

Primary adipocyte culture

Primary floating adipocytes were cultured under membranes according to Harms et al.⁶⁸. Packed adipocytes (30 µl) were seeded onto one membrane and kept in inverted culture for 48 h in maintenance medium (DMEM-F12 (Gibco, catalogue no. 31330095), 10% FBS, 1% penicillin-streptomycin, 10 nM insulin). After 48 h of maintenance, adipocytes were washed and serum and glucose starved overnight in KREBBS-Ringer buffer (120 mM NaCl, 4.7 mM KCl, 1.2 mM KH₂PO₄, 1.2 mM MgSO₄, 2.5 mM CaCl₂, 25 mM HEPES (Lonza, catalogue no. BEBP17-737E), pH 7.4) and 2.5% fat-free BSA (Sigma-Aldrich, catalogue no. A6003).

Glucose uptake

Glucose uptake from primary adipocytes was measured using the Glucose Uptake-Glo Assay Kit (Promega, catalogue no. J1341) according to the manufacturer’s instructions. Adipocytes were preincubated with 5 nM insulin for 15 min before 2-deoxy-d-glucose was added at 1 mM final concentration. Protein concentration was measured using a Pierce 660 nm Protein Assay Kit (Thermo Fisher, catalogue no. 22662) and the Ionic Detergent Compatibility Reagent (Thermo Fisher, catalogue no. 22663). Both assays were read with a plate reader (BioTek Gen5 Microplate Reader).

C16 uptake

Starved adipocytes were incubated with 5 nM BODIPY-palmitate (Thermo Fisher, catalogue no. D3821) in the presence of 10 nM insulin for 1 h. Subsequently, adipocytes were washed twice and lysed in 200 µl of RIPA buffer. Then, 100 µl of lysate was used to measure BODIPY signal. Diluted lysate was used to measure protein concentration using a DC Protein Assay Kit II (Bio-Rad Laboratories, catalogue no. 5000112) for normalization. Both assays were read with a plate reader (BioTek Gen5 Microplate Reader).

Histology

Tissues were collected, fixed in 4% PBS-buffered formalin for 72 h at 4 °C and stored in PBS at 4 °C. Following paraffin embedding, tissues were sent to the pathology service centre at Instituto Murciano de Investigación Biosanitaria Virgen de la Arrixaca for sectioning, trichrome staining, haematoxylin and eosin staining, and imaging. Tissues from two independent experiments were sent for sectioning.

Adipocyte size quantification

Images of ingAT and epiAT were taken with 3DHISTECH Slide Viewer 2 and then analysed with Adiposoft⁶⁹ using Fiji ImageJ⁷⁰. Five to ten images were taken of each section belonging to a biological replicate (n = 4).

Sample processing and library preparation

Isolation of nuclei from mouse tissue

Nuclei were isolated from snap-frozen epiAT in ice-cold Nuclei Extraction Buffer (Miltenyi, catalogue no. 130-128-024) supplemented with 0.2 U µl^−1 recombinant RNase Inhibitor (Takara, catalogue no. 2313) and 1× cOmplete EDTA-free Protease Inhibitor (Roche, catalogue no. 5056489001) using the gentleMACS Octo Dissociator (Miltenyi, catalogue no. 130-096-427), using C-tubes (Miltenyi, catalogue no. 130-093-237). Nuclei were subsequently filtered through a 50 µm cell strainer (Sysmex, catalogue no. 04-0042-2317) and washed two times in PBS-BSA (1% w/v) containing 0.2 U µl^−1 RNase inhibitor. For snRNA-seq, five mice were pooled per condition.

Isolation of nuclei from human tissue

Nuclei were isolated from snap-frozen human AT (10–50 mg) in ice-cold Nuclei Extraction Buffer (Miltenyi, catalogue no. 130-128-024) supplemented with 1 U µl^−1 recombinant RNase Inhibitor (Takara, catalogue no. 2313), 1× cOmplete EDTA-free Protease Inhibitor (Roche, catalogue no. 5056489001) and 10 mM sodium butyrate using the gentleMACS Octo Dissociator (Miltenyi, catalogue no. 130-096-427), using C-tubes (Miltenyi, catalogue no. 130-093-237).

The nuclei suspension was filtered through a 50 µm strainer, supplemented with PBS-BSA (1% w/v) containing 1× protease inhibitor and RNase inhibitor and centrifuged at 4 °C, at 500g for 10 min. The nuclei pellet was resuspended in 1 ml of PBS-BSA (1%, w/v) supplemented with RNase inhibitor (0.5 U µl^−1) and 1× protease inhibitor and was transferred into a new 1.5 ml tube.

snRNA-seq of AT

Nuclei were counted using a haemocytometer and Trypan blue, concentration was adjusted to approximately 1,000 nuclei per µl and they were loaded onto a G-chip (10x Genomics, catalogue no. PN-1000127). Single-cell gene expression libraries were prepared using the Chromium Next GEM Single Cell 3′ v3.1 kit (10x Genomics) according to the manufacturer’s instructions. To accommodate for low RNA content, two cycles were added to the complementary DNA amplification PCR. Libraries were pooled equimolecularly and sequenced in PE150 (paired-end 150) mode on a NovaSeq 6000 with about 40,000 reads per nucleus at Novogene or using a NovaSeqX at the Functional Genomics Center, Zurich.

Paired TRAP–seq, CUT&Tag and ATAC–seq

Paired TRAP–seq, CUT&Tag and ATAC–seq protocols were developed on the basis of published protocols^{67,71,72,73,74}.

Ribosome and nuclei isolation

Nuclei and ribosomes were isolated from snap-frozen epiAT from AdipoERCre x NuTRAP mice in ice-cold Nuclei Extraction Buffer (Miltenyi, catalogue no. 130-128-024) supplemented with 0.2 U µl^−1 recombinant RNase Inhibitor (Takara, catalogue no. 2313), 1× cOmplete EDTA-free Protease Inhibitor (Roche, catalogue no. 5056489001) and 10 mM sodium butyrate using the gentleMACS Octo Dissociator (Miltenyi, catalogue no. 130-096-427), using C-tubes (Miltenyi, catalogue no. 130-093-237). The nuclei suspension was filtered through a 50 µm strainer and centrifuged at 4 °C, 500g for 5 min. The supernatant was transferred into a new tube and supplemented with 2 mM dithiothreitol, 100 µg ml^−1 cycloheximide (Sigma-Aldrich, catalogue no. 01810) and 1 mg ml^−1 sodium heparin (Sigma-Aldrich, catalogue no. H3149-10KU) and kept on ice. The nuclei pellet was resuspended in 1 ml of PBS-BSA (1%, w/v) supplemented with 0.2 U µl^−1 RNase inhibitor, 1× cOmplete EDTA-free Protease Inhibitor and 10 mM sodium butyrate and transferred into a new 1.5 ml tube. Nuclei were centrifuged and subsequently bound to Dynabeads MyOne Streptavidin C1 beads (Thermo Fisher, catalogue no. 65002) for 30 min at 4 °C followed by three washes with PBS-BSA (1% w/v).

TRAP–seq

Per sample, 25 µl of GFP-Trap Magnetic Agarose Beads (ChromoTEK, catalogue no. gtma-20) were washed in 2 ml of polysome lysis buffer (50 mM TRIS-HCl pH 7.5, 100 mM NaCl, 12 mM MgCl₂, 1% Igepal CA-630 (Sigma-Aldrich, catalogue no. I8896), 1× protease inhibitor). The supernatant was mixed with the beads and incubated at 4 °C on a rotator for 1–2 h. Subsequently, tubes were put on a magnetic stand and the supernatant was removed. The beads were washed three times with polysome lysis buffer supplemented with 2 mM dithiothreitol (Sigma-Aldrich, catalogue no. D0632-10G), 100 µg ml^−1 cycloheximide (Sigma, catalogue no. D0632-10G) and 1 mg ml^−1 sodium heparin (VWR, catalogue no. ACRO411210010) and resuspended in 1 ml Trizol (Thermo Fisher, catalogue no. 15596). Trizol preserved samples were kept at −80 °C until RNA isolation. RNA was isolated by adding 200 µl of chloroform (Sigma-Aldrich, catalogue no. 288306) to samples, followed by shaking and centrifugation at 4 °C, 12,000g for 15 min. The aqueous phase was transferred into a new tube and RNA was isolated and DNase treated with the RNA Clean and Concentrator-5 kit (Zymo Research, catalogue no. R1016), following the manufacturer’s instructions.

RNA libraries were prepared by performing reverse transcription and template switching using Maxima H Minus reverse transcriptase (Thermo Fisher, catalogue no. EP0753), a template switch oligo and an oligodT primer to generate full-length cDNA. cDNA was amplified using the KAPA Hotstart 2x ReadyMix (Roche Diagnostics, catalogue no. 7958935001). Then, 1–3 ng of cDNA was tagmentated using 1.3 µg of Tn5 and amplified using KAPA HiFi plus dNTPs (Roche Diagnostics, catalogue no. 07958846001) and the following PCR settings: 72 °C 5 min, 98 °C 30 s, 10 cycles of 98 °C for 10 s, 63 °C for 30 s, 72 °C for 1 min, hold at 4 °C. Libraries were quantified using the KAPA library quantification kit (Roche Diagnostics, catalogue no. 079602), and sequenced in PE150 mode on a NovaSeq 6000 at Novogene.

CUT&Tag

CUT&Tag was performed as previously described with minor adjustments^74,75. All buffers were supplemented with 1 x cOmplete EDTA-free Protease Inhibitor and 10 mM sodium butyrate. Briefly, nuclei bound to beads were aliquoted into 96-well LoBind plates (Eppendorf, catalogue no. 0030129547) and incubated with primary antibodies—anti-H3K4me3 (abcam, catalogue no. ab8580), anti-H3K27me3 (Cell Signaling Technology, catalogue no. C36B11), anti-H3K27ac (abcam, catalogue no. ab4729), anti-H3K4me1 (abcam, catalogue no. ab8895)—overnight at 4 °C. With the plate on a magnet, the primary antibody solution was removed, and the beads were resuspended in secondary antibody solution (guinea pig anti-rabbit IgG (antibodies-online, catalogue no. ABIN101961)) and incubated at room temperature. pA-Tn5 was bound to antibodies, and transposition was performed at 37 °C and stopped using TAPS-Wash solution. Nuclei were lysed and pA-Tn5 decrosslinked using SDS-release solution. PCR was performed using KAPA HiFi plus dNTPs (Roche Diagnostics, catalogue no. 07958846001) with the following PCR settings: 72 °C 5 min, 98 °C 30 s, 15 cycles of 98 °C 10 s, 63 °C 30 s, and 72 °C final extension for 1 min, hold at 4 °C.

ATAC–seq

Beads with nuclei were resuspended in ATAC–seq solution (10 mM TAPS pH 8.5, 5 mM MgCl₂, 10% DMF (Sigma-Aldrich, catalogue no. D4551), 0.2 µg µl^−1 transposase (Tn5)) and incubated at 37 °C for 30 min. Thereafter, 100 µl of DNA binding buffer (Zymo Research, catalogue no. D4003-1) was added and samples were stored at −20 °C. Then, DNA was extracted using Zymo DNA Clean and Concentrator-5 (Zymo Research, catalogue no. D4004). Library amplification was performed using KAPA HiFi plus dNTPs (Roche Diagnostics, catalogue no. 07958846001) and the following PCR settings: 72 °C 5 min, 98 °C 30 s, 10 cycles of 98 °C 10 s, 63 °C 30 s, 72 °C 1 min, hold at 4 °C.

Both ATAC–seq and CUT&Tag libraries were cleaned using SPRI beads, eluted in nuclease-free water and pooled equimolecularly after library quantification using the KAPA library quantification kit (Roche Diagnostics, catalogue no. 079602). Libraries were sequenced in PE150 mode on a NovaSeq 6000 at Novogene.

Sequencing data processing

snRNA-seq data processing and analysis

Data integration and differential expression analysis for mouse snRNA-seq

The 10x Genomics Cell Ranger v.6.1.2 pipeline was used for demultiplexing, read alignment to reference genome mm10-2020A (10x Genomics), barcode processing and unique molecular identifier (UMI) counting with Include introns argument set to ‘True’. The R package Seurat v.4.1.0 (ref. ⁷⁶) was used to process, integrate and analyse datasets. scDblFinder⁷⁷ was used to identify and remove doublets. Nuclei with unique feature counts less than 500 or greater than 3,000 and UMI counts greater than 40,000 were discarded during quality control (Extended Data Fig. 11a). Highly expressed genes such as mitochondrial genes, pseudogenes and Malat1 were excluded from the count matrix before normalization. SoupX⁷⁸ was used to estimate potential ambient RNA contamination in all samples, but no sample required any correction. Samples were normalized using sctransform and integrated using the CCA (canonical correlation analysis) method built into Seurat. Filtered, normalized and integrated nuclei data were clustered by using the Louvain algorithm with a resolution of 0.4 using the first 30 principal components. Cluster markers were identified on the basis of differential gene expression analysis (Wilcoxon rank-sum test with |log₂FC| > 0.25 and adjusted P < 0.05). Clusters were then annotated on the basis of known markers from literature^{34,36,37,46,79,80}. Additionally, our manual cluster annotation was confirmed by reference mapping against a reference male mouse epiAT³⁴ dataset (Extended Data Fig. 11b,c). Differential expression analysis (Wilcoxon rank-sum test with |log₂FC| > 0.5 and adjusted P < 0.01) per cell type between different conditions was done using the FindMarkers function from Seurat. Differential expression analysis hits were intersected with a list of epigenetic modifier genes (see the Source Data to Extended Data Fig. 8) to investigate their expression dynamics. For visualization of snRNA-seq data we used the R package SCpubr v.1 (ref. ⁸¹).

Data integration and differential expression analysis for human snRNA-seq

The 10x Genomics Cell Ranger v.7.2.0 pipeline was used for demultiplexing, read alignment to reference genome GRCh38-2020-A (10x Genomics), barcode processing and UMI counting, with force cells set to 10,000. The R package Seurat v.4.1.0 (ref. ⁷⁶) was used to process, integrate and analyse datasets. scDblFinder⁷⁷ was used to identify and remove doublets. Nuclei with unique feature counts <300 or >4,000 (LTSS) / 6,000 (NEFA), UMI counts >15,000 (LTSS) / 25,000 (NEFA) and mitochondrial gene counts greater than 5% were discarded during quality control (Extended Data Fig. 12). SoupX⁷⁸ was used to estimate and correct for potential ambient RNA contamination in all samples. Samples were normalized using sctransform and integrated using the CCA method built into Seurat. Filtered, normalized and integrated nuclei data were clustered by using Louvain algorithm using the first 30 principal components. For each study, the cluster resolution was determined using the R package clustree⁸². Cluster markers were identified on the basis of differential gene expression analysis (Wilcoxon rank-sum test with |log₂FC| > 0.25 and adjusted P < 0.01). Clusters were then annotated on the basis of known markers from literature^{34,35,36,37,83}. Additionally, our manual cluster annotation was confirmed by reference mapping against reference human white AT atlas³⁴ (Extended Data Figs. 2 and 3). For each AT depot, adipocytes from two studies were integrated together using the first 20 principal components following the steps as mentioned above. Differential expression analysis (Wilcoxon rank-sum test with |log₂FC| > 0.5 and adjusted P < 0.01) per cell type between different conditions was done using the FindMarkers function from Seurat. Differential expression analysis hits were validated using MAST and likelihood-ratio tests using the FindMarkers function from Seurat. For visualization of snRNA-seq data, we used the R package SCpubr v.1 (ref. ⁸¹).

SNP-based demultiplexing of human snRNA-seq datasets

To perform SNP calling and demultiplexing on the pooled samples, cellsnp-lite⁸⁴ was first used to call SNPs on a cell level using the 1000 Genomes-based reference variant call file for hg38 at a resolution of 7.4 million SNPs. SNPs with less than 20 counts and a minor allele frequency of less than 10% were filtered out, as per the developer recommendations. Finally, the tool vireo⁸⁵ was used to demultiplex the pooled data using the cellsnp-lite-derived genotype information.

For each donor, we analysed tissue composition and removed nuclei belonging to donors in the case in which no nuclei were assigned as adipocytes (one case in NEFA) or more than 50% or nuclei were assigned as B cells (one case in MTSS; lean donor) after correspondence with surgeons.

Transcriptional retention

DEGs from obese and WL cells from mouse and human were overlayed, respectively. A DEG was considered restored if it was no longer deregulated in WL cells when compared with controls. If not restored, we considered a DEG part of a transcriptional memory. Clusters identified as similar cell types (for example, three clusters of endothelial cells) were merged for DEG quantification but not differential expression analysis itself. For human snRNA-seq, only cell types for which we obtained at least 30 cells per donor were considered for the retention analysis. T cells were not included in differential expression analysis or transcriptional retention analysis. For integrated human adipocyte differential expression analysis quantification, non-coding transcripts were excluded.

TRAP–seq

Quality control of the raw reads was performed using FastQC v.0.11.9. Raw reads were trimmed using TrimGalore v.0.6.6 (https://github.com/FelixKrueger/TrimGalore). Filtered reads were aligned against the reference mouse genome assembly mm10 using HISAT2 v.2.2.1. Raw gene counts were quantified using the featureCounts⁸⁶ program of subread v.2.0.1. Differential expression analysis was performed using the R package EdgeR⁸⁷, with |log₂FC| ≥ 1 and nominal P < 0.01 as cut-offs.

CUT&Tag and ATAC–seq data processing and analysis

Quality control of CUT&Tag and ATAC–seq data and generation of bedgraph files was performed as described previously⁷⁵. Peaks were called from CUT&Tag sequencing and ATAC–seq libraries on individual bedgraph files using SEACR⁸⁸ v.1.3 in stringent mode with a peak calling threshold of 0.01. Peaks overlapping with mouse blacklist regions⁸⁹ were filtered out. Called peaks were annotated using the R package ChIPSeeker⁹⁰. Peak fold enrichment against genomic features was calculated using the formula: Σ(base pair (bp) overlap) × genome_size/[Σ(bp hPTM peak) × Σ(bp genomic feature)]. Genomic features tracks were downloaded from ENCODE using the R package annotatr⁹¹. Visual quality control of bam files was performed with Seqmonk⁹². Called peaks were combined to generate a union peak list and quantified using the R package chromVAR⁹³ v.1.16, generating a raw peak count matrix.

MOFA

MOFA^50,94 was run to identify the driving variation source across all conditions using all data modalities. For each modality, the top 3,000 variable features (genes or peaks) between all samples were selected using the R package DESeq2 (ref. ⁹⁵) and used as input to train the MOFA model. The trained MOFA model represented data variability in terms of five latent factors, which were further explored and visualized.

Generation of enhancer tracks of adipocytes

Adipocyte chromatin states were identified using ChromHMM v.1.22 (ref. ⁹⁶) in concatenated mode with binned bam files (200-bp bins) from each condition combining all hPTMs and ATAC–seq. After final model selection⁷⁵ with eight chromatin states and emission parameter calculation of hPTMs and ATAC–seq, chromatin state fold enrichment was performed against genomic features and ENCODE candidate cis-regulatory elements. Enhancer states were selected on the basis of genomic localization and hPTM enrichment. Subsequently, an enhancer track was generated per condition and merged for differential analysis.

Differential analysis of hPTMs and ATAC–seq

Promoters

Promoters were defined using the getPromoters function from ChIPSeeker with TxDb.Mmusculus.UCSC.mm10.knownGene as input and setting the TSSRegion to c(-2000, 2000). Peaks overlapping with promoters were extracted using the annotatePeak function from ChIPseeker⁹⁰ by selecting peaks annotated as promoters. For differential analysis, our raw peak count matrix was filtered for these promoter regions and counts were aggregated at gene level. Differential analysis of the same hPTM between two conditions was performed using the R package EdgeR⁸⁷ with nominal P < 0.01 and |log₂FC| > 1 as cut-offs.

Enhancers

ChromHMM was used to identify regions in the genome that were marked by H3K4me1, H3K27ac and open (ATAC–seq) but not enriched for H3K4me3 and that were not promoters (Extended Data Fig. 9b–e). States 6 and 5 were selected as enhancer regions on the basis of their genomic locations (distal enhancer elements) (Extended Data Fig. 9b–e).

Our raw peak count matrix was filtered for enhancer regions defined by chromHMM, and peaks around the TSS (±2,000 bp) were discarded. Linkage of putative enhancers to genes was done using the R package ChIPSeeker by selecting the closest gene (TSS or gene body) within 20,000 bp distance. Putative enhancers farther away than 20,000 from a TSS or gene body were not linked to any gene and were discarded from downstream GSEA.

For each hPTM, the raw filtered peak matrices were log-normalized using the R package EdgeR and Pearson’s correlation coefficient was computed using the cor function from the R package stats v.3.6.2.

Differential analysis of the same hPTM between two conditions was performed using the R package EdgeR with nominal FDR < 0.05 and |log₂FC| > 1 as cut-offs.

PCA

Raw gene and promoter/enhancer-specific peak count matrices were log-normalized using the R package EdgeR. PCA of the normalized count matrices was performed using the prcomp function of R package stats v.3.6.2.

GSEA

GSEA was performed using the R package enrichR^97,98,99. For generation of heatmaps summarizing GSEA across cell types, significantly enriched terms were selected using the adjusted P value (<0.01) and the combined.score (enrichment score) was scaled and visualized.

Visualization

R v.4.2, GraphPad Prism v.9.5.1 and Seqmonk v.1.48.1 were used to generate plots and Affinity Designer and Publisher were used to adjust plots for clarity (for example, colour schemes).

Statistical analysis of physiological parameters from mice

GraphPad Prism v.9.5.1 was used to analyse physiological data from mice. Each dataset of physiological parameters was tested for normality using the Shapiro–Wilk test. On the basis of the results, parametric or non-parametric tests were used to compare experimental with age-matched control groups. Tests are indicated in figure legends and the Source Data.

Reporting summary

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