Abstract

   Adipose tissue is a central player in energy balance and glucose
   homeostasis, expanding in the face of caloric overload in order to
   store energy safely. If caloric overload continues unabated, however,
   adipose tissue becomes dysfunctional, leading to systemic metabolic
   compromise in the form of insulin resistance and type 2 diabetes.
   Changes in adipose tissue during the development of metabolic disease
   are varied and complex, made all the more so by the heterogeneity of
   cell types within the tissue. Here we present detailed comparisons of
   atlases of murine WAT in the setting of diet-induced obesity, as well
   as after weight loss induced by either vertical sleeve gastrectomy
   (VSG) or treatment with the GLP-1 receptor agonist semaglutide. We
   focus on identifying populations of cells that return to a lean-like
   phenotype versus those that persist from the obese state, and examine
   pathways regulated in these cell types across conditions. These data
   provide a resource for the study of the cell type changes in WAT during
   weight loss, and paint a clearer picture of the differences between
   adipose tissue from lean animals that have never been obese, versus
   those that have.

   Keywords: Obesity, Bariatric surgery, Semaglutide, Single cell, Adipose
   tissue

Highlights

     * •
       Atlas of almost 50,000 cells from obese mouse inguinal and
       epididymal adipose tissue after VSG or semaglutide.
     * •
       VSG leads to an inflammatory phenotype in the peritoneal
       compartment compared to sham.
     * •
       Semaglutide induced weight loss leads to remodeling in adipocytes
       and other cell types.
     * •
       Some adipocyte genes revert to a chow-like state after weight loss,
       while others remain locked in the obese pattern.

1. Introduction

   Even modest weight loss can result in marked improvements in metabolic
   function and cardiovascular health [[39]1]. An entire industry worth
   billions of dollars has emerged centered on helping people to lose
   weight, encompassing a wide range of strategies such as diet, exercise,
   drugs, and surgery. Weight loss operations, such as Roux en Y gastric
   bypass (RYGB) and the now dominant vertical sleeve gastrectomy (VSG),
   are some of the most effective ways to lose weight and maintain weight
   loss [[40]2]. Recently, GLP-1 agonists like semaglutide, originally
   developed for the treatment of type 2 diabetes, have been shown to be
   nearly as effective for weight loss as bariatric surgery even in
   non-diabetic individuals [[41]3]. This has led to an unprecedented
   amount of interest in their use around the world.

   White adipose tissue (WAT) remodels during the development of obesity,
   a process that involves a profound increase in immune cell infiltration
   leading to a more inflammatory profile. Less, however, is known about
   WAT remodeling that occurs during weight loss. Some studies have shown
   reduction in pro-inflammatory immune cell content of adipose tissue
   following significant weight loss in humans [[42][4], [43][5],
   [44][6]], while others have suggested that immune cell populations do
   not change following weight loss despite the return to a more
   metabolically healthy phenotype [[45][7], [46][8], [47][9], [48][10]].

   Beyond the immune compartment, relatively little is known about changes
   in the transcriptional programs in other adipose cell types after
   weight loss. Adipocytes, of course, become smaller and secrete more
   adiponectin, and this is associated with improvements in local and
   systemic metabolism [[49]11].

   Single nuclear RNA-sequencing (sNuc-seq) enables the analysis of all
   adipose tissue cell types, including large, fragile adipocytes which
   are refractory to traditional single cell techniques [[50][12],
   [51][13], [52][14]]. Here, we use sNuc-seq to examine changes in mouse
   adipose tissue after weight loss via two distinct modalities, VSG and
   semaglutide. Our data indicate that weight loss by any method improves
   metabolic homeostasis despite method-dependent differences in immune
   cell composition. We also show that while many of the adipocyte
   transcriptional pathways that are characteristic of the obese state
   revert to a pattern seen in lean animals, many other genes and pathways
   remain “locked” in the obese pattern despite metabolic improvement.

2. Results

2.1. VSG and semaglutide cause weight loss in obese mice

   Male C57Bl/6 mice were fed a high fat diet (60% fat; HFD) for 8 weeks
   prior to either VSG or sham surgery. After surgery, animals were
   maintained on HFD, and a subset of the Sham cohort was treated with
   semaglutide (0.04 mg/kg subcutaneously) (Sema) starting on post-op Day
   5. Animals from each of these three groups were sacrificed at two
   postoperative timepoints, and success of the VSG was determined
   post-mortem by application of an unbiased scoring system (see Methods).
   Epididymal (EPI) and inguinal (ING) fat pads were harvested
   ([53]Figure 1A). As expected, both the VSG and Sema groups lost weight
   ([54]Figure 1B), consistent with a reduction in food intake
   ([55]Extended Data Figure 1a). Notably, at 10 days post-surgery
   (Timepoint 1) both groups exhibited an equal amount of weight loss,
   after which the VSG group started to regain weight. Although the VSG
   group re-attained their pre-surgery weight by Timepoint 2, they were
   still significantly leaner than sham-operated mice, which continued to
   gain weight throughout the experiment. Regardless, the VSG and Sema
   groups both demonstrated significantly reduced fasting glucose compared
   to Sham controls at both timepoints, with fasting insulin trending
   downward ([56]Figure 1C).

Figure 1.

   [57]Figure 1
   [58]Open in a new tab

   Two models of weight loss. A) Experimental schematic. Male mice
   underwent sham surgery, vertical sleeve gastrectomy, or sham surgery
   followed by semaglutide treatment. At two post-operative timepoints,
   inguinal and epididymal fat were isolated. Fat from three animals per
   cohort/timepoint was subjected to single nuclear RNA sequencing. B-C)
   Body weight and fasting glucose and insulin levels in animals from A
   (Sham T1 n = 7, VSG T1 n = 13, Sema T1 n = 6, Sham T2 n = 6, VSG T2
   n = 12, Sema T2 n = 8). D) UMAP plot of cells split by timepoint and
   condition. Data is randomly subsampled to the same number of cells per
   timepoint, condition, and depot (n = 2256 cells).

2.2. Weight loss induces changes in WAT cellular composition

   We selected three weight-matched animals with low surgical complication
   scores at each combined timepoint and condition for sNuc-seq analysis,
   and sequenced both the EPI and ING depots. After quality control,
   including stringent doublet removal, we identified 47,180
   high-confidence cells. All expected cell types were identified,
   including adipocytes, adipose stromal and progenitor cells (ASPCs),
   mesothelial cells, vascular cells, and immune cells ([59]Extended Data
   Table 1 and [60]Extended Data Figures 1, 2). At Timepoint 1, the
   overall cell type proportions were not strongly altered by either
   weight loss treatment, although examination of uniform manifold
   approximation and projection (UMAP) plots reveal that the distribution
   within cell types, particularly adipocytes, was affected
   ([61]Figure 1D, [62]Extended Data Figure 1). By Timepoint 2, the
   overall proportions of many cell types had changed. Both weight loss
   groups showed an increase in adipocyte and decrease in macrophage and
   monocyte proportions when compared to Sham mice ([63]Figure 1D,
   [64]Extended Data Figure 2).

2.3. Vascular cells are largely unchanged by weight loss

   Vascular cell type proportion changes very little with weight gain in
   mice [[65]12,[66]13]. Similarly, we noted overall stability in vascular
   cell type proportion with weight loss ([67]Extended Data Figure 3). The
   VSG group in particular exhibited no differences compared to Sham mice
   at either timepoint, while the Sema group demonstrated a small increase
   in arteriole population mEndo^Dkk2 at Timepoint 2 in both EPI and ING,
   with a corresponding decrease in venule mEndoV cells. There was also an
   increase in the proportion of lymphatic endothelial cells (LECs) in the
   ING of Sema animals, and an increase in the proportion of endothelial
   stalk mEndo^Bmp4 in EPI of Sema animals. These changes may indicate
   increased remodeling of WAT in Sema vs. VSG.

2.4. Depot influences the ASPC response to weight loss

   ASPC populations were also largely stable after weight loss, with only
   small changes at Timepoint 1 in both VSG and Sema groups ([68]Extended
   Data Figure 4). At Timepoint 2, VSG populations were largely similar to
   those seen in Sham, while the Sema group exhibited an increase in
   mASPC^Pde11a and mASPC^Cd36 in EPI, with a corresponding decrease in
   mASPC^Aldh1a3 ([69]Figure 2A,B). This suggests there may be an increase
   in adipogenesis in EPI, as mASPC^Pde11a and mASPC^Cd36 resemble
   previously described populations of committed and differentiating
   preadipocytes, while mASPC^Aldh1a3 expresses markers such as Dpp4 and
   Pi16, resembling a previously described population of non-committed
   mesenchymal stem cells [[70]15]. In contrast, the populations of ASPCs
   in ING are reversed, with semaglutide causing an increase in
   mASPC^Aldh1a3 cells and a decrease in the more committed preadipocyte
   populations ([71]Figure 2C,D). This is reminiscent of previous
   observations that weight gain leads to adipocyte hyperplasia in EPI but
   hypertrophy in ING [[72]16], although it is notable that here we see
   this in response to acute weight loss.

Figure 2.

   [73]Figure 2
   [74]Open in a new tab

   ASPC and mesothelial compartments show increased remodeling with
   Sema-induced weight loss. UMAP plot of EPI (A) and ING (C) ASPCs at
   Timepoint 2, split by condition. Data is randomly subsampled to contain
   the same number of cells (n = 494 for EPI cells, n = 461 for ING cells)
   in each panel. Proportion of cells in EPI (B) and ING (D) ASPC
   subclusters for each sample at Timepoint 2. E) UMAP plot of mesothelial
   cells at Timepoint 2, split by condition. Data is randomly subsampled
   to contain the same number of cells (n = 336) in each panel. F)
   Proportion of cells in mesothelial subclusters at Timepoint 2. G) UpSet
   plot of differential genes from mesothelial pseudobulk analysis. H) Top
   pathways for VSG or Sema upregulated genes from mesothelial pseudobulk
   analysis.

2.5. Weight loss effects on the mesothelium vary by modality

   The mesothelium comprises a layer of specialized stromal cells
   encapsulating the EPI fat [[75]17]. Weight loss, regardless of
   modality, induces prominent differences in mesothelium populations,
   although the specifics vary greatly. While populations mMes^Fn1 and
   mMes^Cst3 are seen in VSG animals, mMes^Pla2r1 and mMes^Lama2 cells are
   enriched in Sema mice ([76]Figure 2E,F, and [77]Extended Data
   Figure 5). To further explore these differences, we performed
   pseudobulk analysis, in which the counts for all cells of a specific
   type are summed together, and the resulting data is analyzed using
   “standard” bulk RNA-seq data tools and methods. This type of analysis
   has been shown to outperform other methods of gene expression analysis
   across conditions in single cell data [[78]18]. Differential expression
   and gene enrichment pathway analysis on the mesothelial pseudobulk data
   revealed that genes enriched in VSG were linked to genes and pathways
   like “negative regulation of alpha-beta T cell activation”, while those
   enriched in the semaglutide mesothelium were related to tissue
   remodeling genes and related pathways such as “phospholipid transport”
   and “muscle tissue morphogenesis” ([79]Figure 2G,H, [80]Supplemental
   Table 1); the significance of these pathways to mesothelial biology
   remain unclear.

2.6. Modality-dependent effects of weight loss on immune cells in WAT

   Despite significant improvements in glycemia and body weight in both
   the VSG and Sema groups at Timepoint 1, we noted very few differences
   in immune cell proportions in these groups compared to the Sham group
   ([81]Extended Data Figure 6). At Timepoint 2, however, differences
   became more apparent, most notably in the macrophage and lymphocyte
   populations. Consistent with prior flow-sorting data [[82]8], we saw
   increased B and T lymphocytes in EPI after VSG, however, in the Sema
   group the increase in such lymphocytes was restricted to ING
   ([83]Figure 3A,B). Even more pronounced was the difference in
   macrophage populations between the groups, with the mMac^Trem2
   population enriched in the Sham and Sema groups, and mMac^Cxcl13
   enriched in the VSG group. When we interrogated cell–cell interactions
   across conditions using CellChat [[84]19], inflammatory pathways such
   as the CCL signaling pathway, while present in both the Sham and Sema
   groups, were enriched in magnitude and affected more cell types in the
   VSG cohort ([85]Figure 3C).

Figure 3.

   [86]Figure 3
   [87]Open in a new tab

   Immune cell changes depend upon weight loss modality. A) UMAP plots of
   immune cells split by condition and depot at Timepoint 2. Data are
   randomly subsampled to contain the same number of cells (n = 457) in
   each panel. B) Proportion of cells in immune subclusters for each
   sample at Timepoint 2. C) CCL signaling pathway network activity across
   clusters at Timepoint 2 determined using CellChat. D) Heatmap of
   selected differentially expressed genes from pseudobulk analysis of
   macrophage gene expression plotted on macrophage pseudobulk calculated
   per sample. Analysis and plots were made using macrophages at Timepoint
   2. E) PCA plot of pseudobulk analysis of EPI macrophages at Timepoint
   2.

   These results, which mirror the pattern seen in the mesothelium,
   suggest that there is increased inflammation in VSG EPI macrophages
   compared to Sema, and even relative to the obese Sham group. We
   examined the macrophage populations further by performing pseudobulk
   analysis of the major macrophage clusters (mMac^Cxcl13, mMac^Iigp1,
   mMac^Engase, and mMac^Trem2) which again revealed strong differences in
   gene expression between the VSG macrophages compared to both the Sham
   and Sema macrophages ([88]Figure 3D,E and [89]Figure 4A,B). Pathway
   analysis of significantly regulated genes found that inflammatory
   pathways such as “leukocyte chemotaxis” were increased in VSG-enriched
   genes, well those such as “regulation of lipid storage” were
   upregulated in the shared Sham and Sema genes, fitting with the
   previously described function of Trem2+ macrophages [[90]20]
   ([91]Figure 4C,D, [92]Supplemental Table 1).

Figure 4.

   [93]Figure 4
   [94]Open in a new tab

   Pro-inflammatory changes to macrophages after VSG. UpSet plots
   depicting the intersection of up-regulated (A) and down-regulated (B)
   genes between VSG and other conditions in pseudobulk analysis of
   macrophages at Timepoint 2. C) Selected GO pathways from pathway
   enrichment analysis on differentially regulated genes in EPI. D)
   Heatmap of genes from highlighted pathways on pseudobulk of EPI
   Timepoint 2 macrophages.

   We hypothesized that the increased inflammation of the EPI depot after
   VSG might be a consequence of the nature of the VSG operation, in which
   part of the stomach is excised. To test this hypothesis, we performed
   sNuc-seq on Timepoint 2 EPI samples from VSG animals with high surgical
   complication scores and compared to Timepoint 2 immune cells from the
   other groups. We noted that some pro-inflammatory immune populations
   (e.g., mBcell^Irf4, mMac^Cxcl13, and mMac^Iigp1) are strongly
   correlated with the surgical complication score, while other
   populations (e.g., mTcell^Il23r/Ctla4, which includes anti-inflammatory
   Treg cells) are decreased proportionally ([95]Extended Data Figure 7).
   These data suggest that the success of the VSG operation is linked to
   at least some of the increased inflammation seen in the EPI of these
   animals, but may not be the only factor.

2.7. Adipocyte subpopulation abundance is highly dependent on weight

   Several studies have now identified subpopulations of white adipocytes
   in humans and mice [[96][12], [97][13], [98][14],[99]21]. Our
   clustering analysis reveals six subpopulations that express distinct
   marker genes ([100]Figure 5A, [101]Extended Data Figure 8a, b). In
   contrast to the modest changes in non-adipocyte cell subpopulations at
   Timepoint 1, there was a notable shift in some of the adipocyte
   subpopulations, with increased mAd^Angpt1 and mAd^Erbb4 and reduced
   mAd^Fasn in both the VSG and Sema groups ([102]Extended Data
   Figure 8c). Of note, in this study we see increased resolution of
   subgroups associated with lean healthy adipocytes and reduced
   resolution of subgroups associated with obese, stressed adipocytes,
   likely because of the difference in the proportion of obese animals in
   this study. This can be seen in reference mapping to our previous
   dataset, where the majority of our new clusters map to previous
   chow-enriched clusters mAd1, mAd2, and mAd3, while obeseity related
   adipocytes only form one cluster that largely maps to our previous
   HFD-enriched clusters mAd4, mAd5, and mAd6 ([103]Extended Data
   Figure 8d).

Figure 5.

   [104]Figure 5
   [105]Open in a new tab

   Adipocyte gene expression is primarily influenced by weight loss. A)
   UMAP plots of adipocytes split by condition and depot at Timepoint 2.
   Data are randomly subsampled to contain the same number of cells
   (n = 746) in each panel. B) Proportion of cells in adipocyte
   subclusters for each sample at Timepoint 2. C) PCA plot of pseudobulk
   analysis of ING (top) or EPI (bottom) adipocytes. D) Heatmap of
   selected differentially expressed genes from pseudobulk analysis of
   adipocyte gene expression.

   Integration of adipocytes from our current and prior studies, as well
   as data from prior work by the Mandrup and Seale groups, show
   reasonable concordance. Adipocytes from Sham animals cluster with those
   enriched in HFD in our previous dataset and from Sárvári et al.
   [[106]13]. Additionally, mAd^Cidea adipocytes, which express a number
   of thermogenic markers in specifically found in the ING Sema samples
   cluster with beige adipocytes induced by cold exposure from Holman
   et al. [[107]21], consistent with previous reports that GLP-1 agonism
   activates thermogenesis [[108]22] ([109]Extended Data Figure 9).

   At Timepoint 2, the differences in adipocyte subclusters were even more
   pronounced. Levels of mAd^Fasn remain high in both the EPI and ING of
   sham animals, and begin to increase in VSG animals, which have regained
   weight. Conversely, Sema animals show increased levels of mAd^Angpt1 in
   both depots, reflecting the lack of weight gain in that cohort
   ([110]Figure 5A,B). To further explore the gene expression differences
   across conditions, we performed pseudobulk analysis on the EPI and ING
   adipocytes. In both cases the largest axis of difference in gene
   expression (PC1) corresponds with body weight. Interestingly, PC2 in
   ING appears to separate VSG gene expression from the other two
   conditions, a difference not seen in EPI samples ([111]Figure 5C,D).
   Pathway analysis of the enriched genes in each condition found genes
   associated with thermogenesis specifically in Sema ING adipocytes,
   while the pathways shared between Sema and VSG included terms related
   to remodeling like stem cell population maintenance ([112]Extended Data
   Figure 10a-c, [113]Supplemental Table 1).

2.8. Persistence of obesity-associated gene expression programs in adipocytes
following weight loss

   We next asked whether the adipocytes of an animal that was once obese
   and has subsequently lost weight differ from the adipocytes of animals
   that had never been fat. To assess this, we first generated new
   sNuc-seq profiles from the ING depot of chow-fed male mice that
   experimentally matched those that we generated for the cohort in this
   study. For the purposes of this comparison, we focused on the Sema
   group, because they exhibited more sustained weight loss, and better
   metabolic outcomes, than VSG mice. We proposed three potential models
   for gene regulation. First, some genes (which we term
   “weight-dependent”) might be purely correlated with weight; such genes
   would change (up or down) in Sham vs Chow, but would then revert to the
   Chow-like profile in Sema. Another pattern (designated “weight memory”)
   would include genes that change in Sham vs. Chow, but which do not
   revert to the chow-like state in the Sema group. Finally, we envisioned
   a set of genes that would not change between Chow and Sham, but which
   respond uniquely to Sema treatment (“sema-specific”) ([114]Figure 6A).
   We performed pseudobulk analysis on ING adipocytes from all three
   groups and were able to identify genes that fell into all three of
   these patterns ([115]Figure 6B,C). “Weight-dependent” genes included
   Agmo, which we previously found as associated with lean adipocytes and
   insulin sensitivity [[116]12], while those specifically associated with
   HFD included the inflammatory transcription factor Irf8
   ([117]Figure 6D). Pathway analysis on weight loss upregulated
   “weight-dependent” genes identified pathways related to metabolism like
   “hydrogenated halocarbon metabolic process”, while the HFD specific
   pathways included “fatty acid beta oxidation” ([118]Extended Data
   Figure 10d-f, [119]Supplemental Table 1). Among the “weight memory”
   genes, we noted genes such as Nfib and Adam19, and associated pathways
   which included “lipid storage” and “fat cell differentiation”. Finally,
   some genes were altered uniquely by Sema, including Rgcc and Sgpp1,
   with associated pathways including “carbohydrate biosynthetic process”.

Figure 6.

   [120]Figure 6
   [121]Open in a new tab

   Adipocyte gene expression patterns with Sema-induced weight loss. A)
   Cartoon depicting three proposed models of gene regulation. The number
   next to each group is the number of significantly changed genes
   following that pattern. B, C) UpSet plots depicting the intersection of
   differentially regulated genes between conditions in pseudobulk
   analysis of inguinal adipocytes. Plots are grouped to look at genes
   that are positively (B) or negatively (C) regulated by weight loss. D)
   Heatmap of selected genes from ING adipocyte pseudobulk showing
   differential expression.

3. Discussion

   Bariatric surgery is the gold standard method of weight loss in humans,
   as it provides significant and lasting metabolic benefits that commence
   shortly after the operation. There are several weight loss operations
   in current use, but VSG is the most popular by virtue of its relative
   simplicity and reduced rate of medical complication [[122]23]. More
   recently, GLP-1 receptor agonists like semaglutide and its derivatives
   have been demonstrated to produce a similar degree of weight loss and
   metabolic benefit as surgery [[123]3]. In this study, we were able to
   identify changes in the cellular composition and gene expression
   profile of white adipose tissue in male mice due to weight loss;
   moreover, were able to distinguish between effects that are due to
   weight loss per se from those due to the specific modality (e.g.,
   surgery vs. drug).

   There is an enormous body of literature that indicates that macrophages
   and other pro-inflammatory immune cells accumulate in WAT in the
   setting of overnutrition. These cells secrete a variety of chemokines,
   cytokines, and other inflammatory mediators that induce insulin
   resistance and subsequent dysglycemia. In this light, one interesting
   finding of this study is that metabolic homeostasis is restored by
   acute weight loss by Timepoint 1, prior to any significant changes in
   the immune compartment. This is true of both VSG and Sema-treated mice.
   At Timepoint 2, there is sustained metabolic improvement in the VSG and
   Sema groups, although the former had regained much of the lost weight,
   and their glucose homeostasis was correspondingly worse than that seen
   in the Sema animals. By Timepoint 2, the cellular immune profiles of
   the VSG and Sema groups diverged significantly, with some reduction in
   pro-inflammatory cells in the latter group and an increase in the
   former. The idea that weight loss and metabolic improvement can be
   uncoupled from the continued presence of pro-inflammatory cells is not
   entirely new. Hasty and colleagues reported a similar phenomenon in a
   diet-dependent weight cycling model in mice [[124]7], flow-sorting
   studies have shown retention of multiple immune populations after
   weight loss in mice [[125]8,[126]9] and a few human studies have shown
   that 5–10% weight loss improves metabolism without reducing adipose
   inflammation [[127]24,[128]25]. More profound weight loss in humans
   following bariatric surgery did lead to changes in immune cell
   composition [[129]4], although the temporal relationship of the
   cellular changes to metabolic improvement was not assessed.

   We noted a significant increase in pro-inflammatory immune cells in the
   VSG cohort at Timepoint 2, which was particularly evident in the EPI
   depot. This effect appeared to be independent of weight loss per se,
   because it occurred when weight was being regained in the VSG mice, and
   because it was not observed in the Sema group. This phenomenon was
   mirrored in the mesothelial cells, which show VSG-specific changes
   indicative of exposure to a more inflammatory milieu. We speculated
   this could be a result of “leakage” from the transected stomach, even
   in those animals with the lowest complication scores as determined
   during post-mortem examination. To assess this, we performed sNuc-seq
   on a few samples from mice that were originally rejected from the study
   because of their high complication scores; those data show that the
   cellular changes we see in the “best” VSG mice are exaggerated in the
   ones with the “worst” scores. This suggests that some of the
   VSG-dependent changes in immune cells may result from complications
   specific to the surgery itself. Our data are not entirely consistent
   with a prior report using Q-PCR to demonstrate reduction of
   inflammatory macrophage and T cell markers in the early post-operative
   period of animals receiving VSG [[130]26].

   Because mice are a widely used model of metabolic disease,
   understanding the dynamics of different weight loss modalities in mice
   is valuable for interpretation of many studies, however it is also
   relevant how translatable the findings in mice are to humans. In our
   cohort, the VSG mice lost weight post-surgery, but their weight
   rebounded to levels similar to what was seen at baseline by the end of
   this study, resulting in significantly less weight gain than the sham
   group but more than seen in the Sema group ([131]Figure 1B). Notably,
   despite the weight regain seen in the VSG mice, their serum glucose and
   insulin are still comparable to the Sema animals, which did not
   experience a similar regain, at both timepoints post-surgery
   ([132]Figure 1C). This weight rebound is similar to what is seen in
   other studies of rodent bariatric surgery [[133]27] and also resembles
   the rebound in weight seen in humans post-bariatric surgery, although
   the rebound in humans is often less pronounced than what is seen in
   mice [[134]28]. It is also unknown if the increase in adipose tissue
   inflammation seen post VSG in mice is a phenomenon also observed after
   human bariatric surgery. Profiles of circulating immune cells in humans
   post-bariatric surgery have suggested that an increase in some
   populations of circulating immune cells is seen in humans at timepoints
   immediately following bariatric surgeries [[135]29], although over time
   these populations decrease to levels lower than baseline [[136]30].
   These results suggest that if we were to follow mice to a later
   timepoint, their VSG-related inflammation might similarly resolve.

   Within the adipocyte population, weight loss by either VSG or Sema
   caused similar changes in cellular composition. This was exemplified by
   a reduction in “stress-associated” adipocyte subpopulations that
   express genes associated with insulin resistance and metabolic
   dysfunction. We also noted an interesting increase in thermogenic
   adipocytes after Sema, which was not seen in the VSG group. Several
   studies in mice have indicated that increased beige adipogenesis and
   adipose thermogenesis can occur after GLP1R agonism [[137]22,[138]31].
   In addition, there are limited data from humans suggesting that GLP1R
   agonists may increase brown and/or beige fat activity
   [[139]32,[140]33]. More interesting, however, were gene expression
   patterns that indicated the presence of “weight-dependent”, “weight
   memory”, and “sema-specific” pathways. “Weight-dependent” genes are the
   easiest to understand, as they reflect the current nutritional
   conditions that the adipocytes are experiencing. Expression of these
   genes tracks with, and in some cases may cause, improvements in insulin
   sensitivity and glycemia. Similarly, the “sema-specific” category of
   genes likely reflects idiosyncratic responses to semaglutide treatment,
   likely due to indirect actions of Sema elsewhere in the body, as
   adipocytes express little if any GLP1R. The “weight memory” genes are
   the most interesting, however, as they indicate that there must be
   mechanisms that “lock” some genes into an on or off conformation in
   obesity. Some of this might reflect ongoing needs that persist after
   weight loss. For example, adipogenesis is required to make new
   adipocytes in obesity, and adipogenesis continues after weight loss as
   healthy new adipocytes replace some of the dying, old adipocytes that
   were stressed beyond repair during the period of overnutrition.
   Alternatively, or in addition, it is known that animals that were once
   obese and then lose weight are prone to accelerated weight gain when
   reexposed to a high fat diet [[141]7,[142]34,[143]35]; these locked
   “weight memory” genes might contribute to this phenomenon.

   Finally, while in many respects the effect of weight loss on adipose
   tissue was similar between both weight loss modalities, there were some
   indications that semaglutide specifically causes adipose tissue
   remodeling. The change in proportion of ASPC populations after
   semaglutide treatment indicated that there was increased
   differentiation in EPI, and the pathways enriched in Sema mesothelium
   were related to tissue remodeling. While it is difficult to say if this
   Sema-specificity is due to specific GLP-1-related effects or if it is
   due to the weight regain in the VSG animals, the upregulation of these
   pathways suggest that weight loss leads to a remodeling of adipose
   tissue that is more complex than simply the reduction of adipocyte
   size.

   Limitations of this study include discrepancies in timing and weight
   gain between the VSG and Sema cohorts. We chose to perform the sham
   surgery on the Sema animals 10 days later than the VSG and Sham,
   knowing that Sema treatment would cause more rapid weight loss than the
   VSG and wanting to match the age of the animals at sacrifice. This
   resulted in a 10 day difference between surgery and time point 2 for
   Sema vs. VSG (25 vs. 35 days). Because both weight loss regimens
   reduced body weight significantly compared to Sham, and because both
   regimens improved glucose and insulin levels at both timepoints tested,
   we believe the 10 day difference is negligible, but we acknowledge that
   timing may affect our data. We also note that the VSG group
   demonstrated weight regain by timepoint 2, while the Sema group did
   not. It is notable that despite these differences, some results are
   clearly a result of weight loss modality itself, as for example the VSG
   animals exhibit an increase in a distinct inflammatory macrophage
   population as early as 10 days post-surgery that increases even further
   at 35 days post-surgery which is not seen in either the heavier Sham or
   lighter Sema animals ([144]Figure 3, [145]Figure 4 and [146]Extended
   Data Figure 6). Because of these discrepancies, we focused our analyses
   primarily on Sham vs. VSG or Sham vs. Sema, avoiding direct comparison
   of the two weight loss methods. Other limitations include the
   relatively short time frame studied. It should be noted, however, that
   even within this restricted time period, we observed significant
   metabolic benefits in both the VSG and Sema groups, suggesting that any
   long-term remodeling not captured here is incidental to, or the result
   of, the return to metabolic health, and not the cause thereof. We also
   made comparisons to samples from chow fed mice that were not part of
   the original weight loss cohort. These chow samples were processed in
   the same way as the weight loss samples, and there was excellent
   concordance between gene expression profiles in these samples (i.e.,
   genes expected to be expressed at low levels in lean mice were low in
   both chow and Sema samples), suggesting that these provide a good
   comparator. Finally, we only studied male mice. Obese female mice have
   less adipose inflammation than males, and differences have been
   reported in their response to weight loss [[147]36].

   In conclusion, we find that the cellular composition of WAT exhibits
   several changes in response to bariatric surgery and GLP1 agonism. Some
   of these changes are specifically weight loss-dependent, i.e., they are
   seen in both surgical and pharmacological weight loss. Other changes
   are specific to the modality used. An important example of the latter
   are the changes in pro-inflammatory immune cells, which go up in VSG
   and down in Sema. Importantly, these changes do not correlate well with
   the beneficial effects of weight loss on metabolic function, which are
   seen with both modalities.

4. Methods

4.1. Animal models of weight loss

   C57BL/6J DIO male mice, 12–13 weeks of age, were purchased from Jackson
   Labs and maintained on 60% HFD (Research Diets Inc, Cat#D12492) for an
   additional 2–4 weeks prior to surgery. Body composition was measured
   using EchoMRI (Echo Medical Systems) before surgery and at the end of
   study. Sham and VSG surgeries were performed as described previously
   [[148]37,[149]38]. Briefly, mice were fasted overnight prior to
   surgery. Animals were anesthetized using isoflurane, and a small
   laparotomy incision was made in the abdominal wall. The lateral 80% of
   the stomach along the greater curvature was excised in VSG animals by
   using an ETS 35-mm staple gun (Ethicon Endo-Surgery). Sham surgery was
   performed by the application of gentle pressure on the stomach with
   blunt forceps for 15 s. All mice received one dose of Buprinex
   (0.1 mg/kg) and Carprofen (5 mg/kg) immediately after surgery. All mice
   received Carprofen (5 mg/kg) for 3 days after surgery. Animals were
   placed on DietGel Boost (ClearH[2]O; Portland ME) for 3 days after
   surgery. They were placed back on the pre-operative solid diet (60%
   HFD) on day 4 post-surgery. Body weight and food intake as well as
   overall health were monitored daily for the first 7 days after surgery
   and once weekly until the end of the studies. Semaglutide (0.04 mg/kg)
   was administered to some Sham-operated mice subcutaneously once daily
   starting at day 5 post-surgery. Animals were sacrificed at two
   timepoints post-surgery, at 10 days for all groups (Timepoint 1), and
   at 25 days for semaglutide treated animals and at 35 days for sham and
   VSG animals (Timepoint 2). All mice were sacrificed on the same
   calendar days with the order of sacrifice counter-balanced across all
   the experimental groups. Because we knew that weight loss would occur
   faster with Sema, we chose to dose them for 25 days to more carefully
   match the weight loss between the Sema and VSG-treated mice, therefore
   their surgeries were performed 10 days later than the others so at
   sacrifice all animals would be the same age. All mice were fasted for
   4 h prior to tissue collection. Terminal fasting (4 h) glucose and
   insulin levels were measured at the end of the study for each cohort.

4.2. Calculation of VSG score

   The presence of contained leaks is checked and evaluated at sacrifice
   in a blind manner by three experienced surgeons. Each surgeon assigns a
   score to each particular sleeve and then an average score (out of three
   values) is calculated.

   Scores of 1 are assigned when the gastric sleeve is intact and similar
   in size to that at the time of surgery, with no adhesions present.
   Scores of 2–3 are assigned when there is a small, localized leak. Any
   adhesions are minor. Score 4 is assigned when the sleeve is slightly
   enlarged but has a single leak of a bigger size, with more moderate
   adhesions to the surrounding organs. Scores of 5–7 are assigned when
   there is a large solitary leak to a very large leak with multiple
   chambers or multiple big leaks. Adhesions range from well-established
   to very severe. Mice with scores of 4–7 are excluded from the initial
   analysis primarily due to their pathomorphological changes and
   inflammatory profile.

4.3. Single nuclear sequencing

   Adipose sNuc-seq was performed as previously described [[150]39].
   Briefly, for each 10× run, 8-12 adipose samples were separately
   homogenized in TST buffer using a gentleMACS Dissociator (Miltenyi
   Biotec). Lysate was filtered through 40 μm and 20 μm nylon filters
   (CellTreat), centrifuging and resuspending in fresh buffer each time.
   After the second wash, each sample was incubated with NucBlue
   (Thermofisher scientific) as well as an individual hashtag antibody
   (BioLegend) for 45 min. After incubation, nuclei were washed and flow
   sorted together into RT Reagent B (10× Genomics) using a Beckman
   Coulter MoFlo Astrios EQ with a 70 μm nozzle to remove poor quality
   nuclei and count the same number of nuclei per sample. After sorting,
   samples were immediately loaded on the 10× Chromium controller (10×
   Genomics) according to the manufacturer's protocol. Single Cell 3′ v3.1
   chemistry was used to process all samples and cDNA and gene expression
   libraries were generated according to the manufacturer's instructions
   (10× Genomics). Gene expression libraries were multiplexed and
   sequenced on the Nextseq 500 (Illumina).

4.4. Single cell data processing and analysis

   Raw sequencing reads were processed into digital expression matrices
   (DGE) using CellRanger (10× Genomics, version 6.1.2) and using GENCODE
   annotation GRCm38. DemuxEM [[151]40] (version 0.1.6) was used to sort
   cells into individual samples based on their antibody hashtags. Ambient
   RNA was removed from the processed DGEs using cellbender [[152]41]
   (version 0.2.0), and doublet scores were calculated using both
   scDblFinder [[153]42] (version 1.2.0) and scds [[154]43] (version
   1.4.0). Cells were removed as doublets if they were both determined to
   be a doublet using scDblFinder and if they had a scds hybrid score
   >1.5. Cells with <800 UMIs were removed from the dataset, as were cells
   with >10% of mitochondrial reads. Genes found in fewer than 2 cells
   were similarly removed. The data were then normalized using SCTransform
   and integrated using RPCA in Seurat (version 4.1.1), integrating by 10×
   run so as not to integrate over differences seen between conditions. To
   subcluster, cells were subset into broad cell types and reintegrated
   and clusters were recalculated. Subclusters with high doublet scores
   and/or high percentages of mitochondrial reads were removed and the
   data was again reintegrated and clusters were recalculated.

4.5. Reference mapping and annotation

   Reference mapping was performed between the reported dataset and the
   mouse dataset from Emont et al. (2022) using Seurat multimodal
   reference mapping [[155]44]. To run, the object was then split by 10×
   run and mapped onto the sNuc-seq data from the matching Emont et al.
   (2022) mouse all-cell or subset object using the RNA assay and PCA
   reduction. To annotate cell types, calculated clusters were evaluated
   for similarity to previously annotated clusters. If clusters mapped
   exactly, they were annotated as the same cell type. If not, they were
   evaluated for distinct marker genes and annotated as a new cluster. All
   clusters without distinct names were renamed to be defined by a top
   marker gene.

4.6. Pseudobulk and pathway analysis

   To calculate pseudobulk counts, objects for individual cells (i.e.
   adipocytes, macrophages, etc) were split by sample and randomly subset
   to contain the same number of cells per sample. Differential expression
   analysis (DEA) was run on pseudobulk samples using edgeR (version
   3.30.3). To reduce [[156]45] the variation introduced by the random
   subsampling, subsampling and DEA were performed 10,000 times,
   significantly regulated genes used on plots and for pathway analysis
   were defined as those with log2 fold change greater than 0.5 and FDR
   less than 0.1 in at least 9000 iterations. To plot, heatmaps were
   generated using pseudobulk sums of all cells in each sample, and edgeR
   was used to calculate cpms from the pesudobulk counts. PCA plots were
   generated using a random subsampled pseudobulk. Heatmaps were plotted
   using pheatmap (version 1.0.12) and UpSet plots were generated using
   UpSetR [[157]46] (version 1.4.0). Pathway analysis was performed with
   clusterProfiler [[158]47] (version 4.12.0) using GO biological
   processes pathways.

4.7. Prediction of cell–cell interactions

   To predict cell–cell interactions, data were subset to exclude
   “epithelial”, “schwann/glial” and “stromal” clusters and split by
   timepoint and condition. Each resulting dataset was analyzed using
   CellChat [[159]19], filtering the output to exclude interacting
   clusters with fewer than 5 cells.

4.8. Integration across datasets

   To compare adipocyte datasets, adipocytes from Emont et al., 2022
   [[160]12], Sárvári et al. [[161]13], and Holman et al. [[162]21] were
   combined. To integrate, data from this paper were split by 10× run,
   data from Emont 2022 were split by animal, and data from Sárvári and
   Holman were split by sample. Data were integrated using CCA, with the
   data from this paper as reference. UMAP plots and clusters were
   calculated using 30 PCs and a resolution of 0.5.

CRediT authorship contribution statement

   Margo P. Emont: Writing – review & editing, Writing – original draft,
   Visualization, Investigation, Funding acquisition, Formal analysis,
   Data curation, Conceptualization. Adam L. Essene: Methodology,
   Investigation. Anton Gulko: Methodology, Investigation. Nadejda
   Bozadjieva-Kramer: Methodology, Investigation. Christopher Jacobs:
   Formal analysis, Data curation. Soumya Nagesh: Formal analysis. Randy
   J. Seeley: Writing – review & editing, Supervision, Methodology,
   Conceptualization. Linus T. Tsai: Writing – review & editing,
   Supervision, Methodology, Conceptualization. Evan D. Rosen: Writing –
   review & editing, Writing – original draft, Supervision, Funding
   acquisition, Conceptualization.

Declaration of competing interest

   The authors declare the following financial interests/personal
   relationships which may be considered as potential competing interests:

   Evan D. Rosen reports financial support was provided by National
   Institutes of Health. Evan D. Rosen reports a relationship with Source
   Bio that includes: consulting or advisory. Randy Seeley reports a
   relationship with Novo Nordisk Inc that includes: consulting or
   advisory and funding grants. Randy Seeley reports a relationship with
   Eli Lilly and Company that includes: consulting or advisory and funding
   grants. Randy Seeley reports a relationship with Fractyl Health, Inc.
   that includes: funding grants. Randy Seeley reports a relationship with
   AstraZeneca Pharmaceuticals LP that includes: funding grants. Randy
   Seeley reports a relationship with Congruence that includes: consulting
   or advisory and funding grants. Randy Seeley reports a relationship
   with Amgen Inc that includes: consulting or advisory and funding
   grants. Randy Seeley reports a relationship with Glyscend Inc that
   includes: funding grants. Randy Seeley reports a relationship with
   BullFrog AI Holdings, Inc. that includes: equity or stocks and funding
   grants. Randy Seeley reports a relationship with Cinrx Pharma, LLC that
   includes: consulting or advisory. Randy Seeley reports a relationship
   with Structure that includes: consulting or advisory. Randy Seeley
   reports a relationship with Crinetics Pharmaceuticals Inc that
   includes: consulting or advisory. Randy Seeley reports a relationship
   with Rewind that includes: equity or stocks. If there are other
   authors, they declare that they have no known competing financial
   interests or personal relationships that could have appeared to
   influence the work reported in this paper.

Acknowledgements