A genetic map of the mouse dorsal vagal complex and its role in obesity

Abstract

The brainstem dorsal vagal complex (DVC) is known to regulate energy balance and the target of appetite suppressing hormones, such as glucagon-like peptide-1 (GLP-1). Here we provide a comprehensive genetic map of the DVC and identify neuronal populations that control feeding. Combining bulk and single-nucleus gene expression and chromatin profiling of DVC cells, we reveal 25 neuronal populations with unique transcriptional and chromatin accessibility landscapes and peptide receptor expression profiles. GLP-1 receptor (GLP-1R) agonist administration induces gene expression alterations specific to two distinct sets of Glp1r neurons – one population in the area postrema (AP) and one in the nucleus of the solitary tract (NTS) that also expresses calcitonin receptor (Calcr). Transcripts and regions of accessible chromatin near obesity-associated genetic variants are enriched in AP and NTS neurons that express Glp1r and/or Calcr, and activation of several of these neuronal populations decreases feeding in rodents. Thus, DVC neuronal populations associated with obesity predisposition suppress feeding and may represent targets for therapy of obesity.

Introduction

Despite the increasing incidence and prevalence of obesity and obesity-related complications across the globe, effective treatments remain scarce^1,2. A variety of data suggest the importance of central nervous system (CNS) pathways for the control of body weight and in susceptibility to and treatment of obesity³. Not only do CNS pathways mediate the effects of a variety of monogenic obesity alleles⁴, but most weight-lowering pharmacotherapeutics act in the brain⁵. Furthermore, body mass index (BMI; a surrogate for obesity)-associated genetic variants identified in genome-wide association studies (GWAS) map predominantly to CNS-expressed genes^6,7.

Within the CNS, the brainstem dorsal vagal complex (DVC) – composed of the area postrema (AP), the nucleus of the solitary tract (NTS), and the dorsal motor nucleus of the vagus nerve (DMV) – plays important roles in integrating peripheral satiety signals and regulating energy balance⁸. Indeed, the AP (a circumventricular organ that lacks a blood-brain barrier and thus directly senses blood-borne signals) and adjacent NTS express a variety of receptors that suppress feeding, including GLP-1 receptor (Glp1r)^9,10, GDNF family receptor α-like (Gfral)¹¹, calcitonin receptor (Calcr), amylin receptor (a Calcr/receptor activity modifying protein (Ramp) heterodimer)¹², and gastric inhibitory polypeptide receptor (Gipr)¹³. The AP and NTS contribute to the anorexic effects of amylin and GLP-1R agonists (GLP-1RAs)^14–16, respectively, and both the AP and NTS are involved in salmon calcitonin (sCT)-induced food suppression^14,17,18. Furthermore, Calcr neurons in the NTS contribute to the physiologic control of feeding and body weight¹⁷. Hence, neurons in the DVC represent likely targets for therapeutic intervention in obesity. It will be important to agnostically and comprehensively understand the identities and heterogeneity of DVC cell populations and to understand potential DVC controllers of food intake and body weight to define potential targets for therapy of obesity.

Large-scale assessment of cellular heterogeneity and the unbiased identification of cell populations require single-cell profiling techniques, such as single-nucleus RNA-sequencing (snRNA-seq) and single-nucleus assay for transposase-accessible chromatin by sequencing (snATAC-seq)¹⁹. These techniques not only permit comprehensive mapping of cell populations by their genetic signatures, but can also reveal cell population-specific responses to pharmacological treatments²⁰. When integrated with GWAS data, these cellular atlases can also be leveraged to agnostically nominate candidate cell populations and regulatory networks mediating susceptibility to complex traits and disease (including obesity)^7,21,22.

Here, we set out to identify specific DVC cell populations that control energy balance. Towards that goal, we combined data from bulk RNA-seq with snRNA-seq and snATAC-seq from the mouse AP and surrounding DVC, thereby constructing transcriptional and chromatin accessibility atlases of DVC cell populations. In addition, we identified DVC cell population-specific responses to anorectic GLP-1RA administration and utilized BMI GWAS data to identify DVC cell populations that likely contribute to obesity predisposition. We demonstrate that two of these neuronal populations (one expressing the amylin receptor complex and another expressing Calcr and Glp1r) can control food intake, consistent with their potential utility as targets for therapeutic intervention in obesity.

Results

RNA-seq of areas activated by GLP-1 receptor agonist administration

The GLP-1RA liraglutide represents one of the most effective currently-available pharmacological anti-obesity treatment⁵; with the more long-acting and potent GLP-1RA semaglutide having completed phase III clinical trials and been submitted for regulatory approval²³. Glutamatergic Glp1r neurons likely mediate GLP-1RA effects on feeding and body weight²⁴, and GLP-1RAs engage cells in the hypothalamus (including in the arcuate (ARH), dorsomedial (DMH) and paraventricular (PVH) nuclei), the lateral septal nucleus (LS), the AP (tyrosine hydroxylase (TH) neurons), and the NTS²⁵. However, we lack a more complete understanding about cell population-specific responses to GLP-1RAs in general and potential differences in the responses to individual GLP-1RAs.

To understand the transcriptional response to GLP-1RA administration, we randomized diet-induced obese (DIO) male mice into four groups; 1) semaglutide-administered, 2) liraglutide-administered, 3) ad libitum fed vehicle-administered, and 4) weight-matched controls (to discriminate between weight loss effects induced by GLP-1RA administration and food restriction; Fig. 1a). Semaglutide was administered at a lower dose than liraglutide to obtain similar effects on weight loss. Mice were dosed once daily for seven days and, as expected, the semaglutide- and liraglutide-administered groups lost weight over the course of treatment (mean ± SEM 10.34 ± 0.59 g and 8.7 ± 0.46 g, respectively) relative to baseline, corresponding to a 18% and 15.5% reduction in body weight, respectively (Fig. 1b). The amount of food given to the weight-matched group was lower than in the GLP-1RA-administered groups, suggesting a reduction in energy expenditure in the weight-matched group compared to the GLP-1RA-administered groups, consistent with previous observations²⁵ (Fig. 1c).

Fig. 1: Overview of transcriptional changes in response to GLP-1 receptor agonist administration.

a Overview of bulk RNA-seq study design. Diet-induced obese mice were randomized into four groups; 1) semaglutide-administered (SQ, day 0, 0.02 mg/kg; day 1, 0.04 mg/kg; day 2–6, 0.1 mg/kg; n=15), 2) liraglutide-administered (SQ, day 0, 0.2 mg/kg; day 1, 0.4 mg/kg; day 2–6, 1.0 mg/kg; n=15), 3) ad libitum fed vehicle-administrated (n=15), and 4) weight-matched controls (n=15) that were dosed once daily for seven days. Six brain areas were isolated with LCM, dissociated, and subjected to RNA-seq. b Daily body weight and c food intake following GLP-1RA administration (n=15 mice/group). Mean ± SEM. *P<0.05, **P<0.01, ***P<0.001 vs. vehicle and ^#P<0.05, ^##P<0.01, ^###P<0.001 semaglutide and liraglutide vs. weight-matched, linear mixed effects model, Bonferroni-adjusted least-squares means two-tailed t-test. d PCA plot of samples colored by brain area. e Number of differentially expressed genes compared to weight-matched animals. Benjamini-Hochberg-adjusted DESeq2 P<0.05. f Log₂ fold-changes of differentially expressed genes for semaglutideand liraglutide-administered animals vs. weight-matched animals. g AP volcano plot for semaglutide-administered vs. weight-matched animals. Dark red indicates log₂ fold-change>0.5 and Benjamini-Hochberg-adjusted DESeq2 P<0.05, top 10 significant genes are labelled.

HFD, high-fat diet; LCM, laser capture microscopy; LS, lateral septum; PVH, paraventricular nucleus of the hypothalamus; DMH, dorsomedial hypothalamic nucleus; ARH, arcuate nucleus of the hypothalamus; AP, area postrema; NTS, nucleus of the solitary tract; SQ, subcutaneous; GLP-1RA, GLP-1 receptor agonist; PCA, principal component analysis; PC, principal component; RNA-seq, RNA sequencing.

Data to reconstruct panels d-g can be found in Supplementary Data 1 and/or through the NCBI Gene Expression Omnibus.

Following treatment, we employed laser capture microscopy (LCM) to dissect relevant brain areas and identify global patterns of differential gene expression using bulk RNA-seq. Following data normalization and quality control filtering, we performed principal component analysis (PCA), revealing that samples clustered by brain area and developmental compartment (Fig. 1d). Of the brain areas examined, the AP exhibited the largest number of differentially expressed genes between the GLP-1RA-administered and weight-matched control mice (Benjamini-Hochberg (BH)-adjusted P<0.05; Fig. 1e, Supplementary Data 1). Apart from only one and two differentially expressed genes in the LS and DMH, respectively, we found no significant differences between the semaglutide- and liraglutide-induced transcriptional changes, suggesting that the two GLP-1RAs have similar transcriptional consequences in the six brain areas investigated (Fig. 1f, Supplementary Data 1); we thus focused subsequent analyses on semaglutide-administered animals. Among the most upregulated genes in the AP from semaglutide-administered animals, many encoded neurotrophins and neuropeptides, including growth hormone releasing hormone (Ghrh), VGF nerve growth factor inducible (Vgf), CART prepropeptide (Cartpt), and prodynorphin (Pdyn), suggesting that GLP-1RA administration alters neuronal function and neurotransmission in the AP (Fig. 1g).

Transcriptional single-cell atlas of the dorsal vagal complex

Directing our attention to the AP and adjacent areas, we investigated whether distinct cell populations could represent potential targets of GLP-1RAs and other therapeutics of obesity. To that end, we subjected DVC tissue from AP-centered dissections from semaglutide-administered, ad libitum fed vehicle-administered, and weight-matched control DIO mice (replicating the previous in vivo set-up) to snRNA-seq (Fig. 2a, Extended Data Fig. 1). The quality-controlled dataset contained a total of 72,128 cells consisting of 49,392 neurons and 22,736 glial cells that distributed across eight distinct cell types (Fig. 2b–c). Marker genes for all cell populations were identified using the CELLEX tool⁷, computing a combined expression specificity measure (ES_μ; range [0,1]; low to high cell population specificity) across four individual metrics (Supplementary Data 2). One glial cell population could not easily be assigned a label based upon known marker genes and turned out to be similar to tanycytes when compared to a published ARH-median eminence (ME) single-cell atlas²² (henceforth tanycyte-like cells²⁶). Top marker genes of these tanycyte-like cells included Wilms tumor 1 homolog (Wt1), Wnt inhibitory factor 1 (Wif1), solute carrier family 22 member 3 (Slc22a3), and cell adhesion molecule-related/downregulated by oncogenes (Cdon; Supplementary Data 2); all of which were highly specific to the AP but not restricted to the ventricles when cross-referenced with in situ hybridization (ISH) data from Allen Mouse Brain Atlas²⁷ (Extended Data Fig. 2). We conclude that the tanycyte-like cells represent a unique glial cell population in the AP although the exact identity of these cells warrants further exploration.

Fig. 2: Transcriptional atlas of DVC cell populations.

a Overview of single-nucleus RNA-seq study design. Diet-induced obese mice were randomized into three groups; 1) semaglutide-administered (n=7), 2) ad libitum fed vehicle-administered (n=6), and 3) weight-matched controls (n=7) that were dosed once a day for seven days as described for the bulk RNA-seq in vivo study. The AP, NTS, and DMV were dissociated and subjected to single-nucleus RNA-seq. b UMAP plot of 72,128 cells colored by cell type. c Non-neuronal marker genes. Top, number of cells per cluster. Bottom, violin plot showing the normalized transcript counts for marker genes. A black label indicates that the gene is also a marker of the cell type in the hypothalamus, a red label indicates lack of specificity in the hypothalamus. d UMAP plot of 49,392 neurons colored by population. e Enrichment of AP and NTS marker genes. Left, dots indicate the significance level of overlap between AP and NTS marker genes and cell population marker genes (ES_μ>0). Bonferroni-adjusted one-tailed Fisher’s exact test. Right, the most likely DVC origin of the cell populations. f Neuronal marker genes. From top to bottom, number of cells per cluster, violin plots of the normalized transcript count for marker genes, the most likely DVC origin of the neuronal populations. Chat, Slc32a1, Slc17a6, Th, Ddc, and Dbh were used as markers for neurotransmitter subtypes.

DVC, dorsal vagal complex; AP, area postrema; NTS; nucleus of the solitary tract; DMV, dorsal motor nucleus of the vagus nerve; UMAP, uniform manifold approximation and projection; OPCs, oligodendrocyte precursor cells; RNA-seq, RNA sequencing; VLMCs, vascular leptomeningeal cells.

Data to reconstruct panels b-f can be found in Supplementary Data 2–3 and/or through the NCBI Gene Expression Omnibus.

Reclustering all neurons using a measure of cluster cohesiveness to determine the optimal number of clusters, we identified 25 neuronal populations (Fig. 2d, Supplementary Data 2). Because our dissections included cells from the AP, NTS, and DMV, we set out to determine the most plausible origin of the different cell populations. Using the LCM-dissected bulk RNA-seq data, we identified neuronal populations enriching for genes that were differentially expressed between the AP and NTS. These data, along with manual curation using the Allen Mouse Brain Atlas²⁷, allowed us to assign five populations of neurons to the AP, 15 to the NTS, and five to the DMV (Fig. 2e, Extended Data Fig. 3, Supplementary Data 3).

The DVC contains GABAergic, glutamatergic, and catecholaminergic (dopaminergic and noradrenergic) neurons^10,24,28. Examining the expression of the vesicular GABA transporter (Slc32a1) and the vesicular glutamate transporter 2 (Slc17a6), we detected an abundance of glutamatergic (Glu1–15) over GABAergic neurons (GABA1–7; Fig. 2f). Assessing the expression of Th, dopamine decarboxylase (Ddc), and dopamine beta hydroxylase (Dbh) as markers for dopaminergic (Th⁺/Ddc⁺/Dbh⁻) and noradrenergic (Th⁺/Ddc⁺/Dbh⁺) neurons, respectively, we further identified several populations of catecholaminergic neurons, all of which were also glutamatergic; one population expressed dopaminergic markers (Glu7^NTS), and four expressed noradrenergic markers (Glu2^NTS, Glu4^AP, Glu5^NTS, and Glu11^NTS). Additionally, using the marker choline O-acetyltransferase (Chat), we identified three cholinergic neuronal populations (Chat1–3) of which two also expressed dopaminergic markers (Chat1^DMV and Chat3^DMV; Fig. 2f). Together these results demonstrate a transcriptionally diverse organization of non-neuronal and neuronal cells in the AP, NTS, and DMV.

Expression of appetite-suppressing genes across cell populations

To understand DVC cell populations that potentially are involved in energy balance control, we examined the expression of Glp1r along with other anorexic receptors and peptides across DVC cell populations. Despite the reportedly high expression of Glp1r in the AP^9,10, we were initially unable to identify Glp1r populations. Remapping the genome annotation to include reads falling within downstream alternative polyadenylation (poly(A)) signals and recomputing the transcript counts for Glp1r, we identified three neuronal populations with high expression specificities of Glp1r; GABA7^AP (ES_μ=0.72), Glu4^AP (ES_μ=0.99), and Glu11^NTS neurons (ES_μ=0.64) (Fig. 3a–b, Supplementary Data 2).

Fig. 3: Neuronal expression of receptors and peptides involved in body weight control.

a Overview of Glp1r gene reannotation. The transcript counts for Glp1r were computed following manual re-annotation to include alternative polyadenylation signals. b Cell population expression specificities of genes previously linked to body weight control. Top, CELLEX expression specificity values (ES_μ) of selected genes. Bottom, the most likely DVC origin of the neuronal populations. c Representative image showing combined IHC of GLP-1R (green) and ISH of Gfral (magenta; n=3). d Representative image showing db-ISH of Glp1r (blue) and Calcr (red; n=3). e Representative image showing combined IHC of GLP-1R (green) and ISH of Casr (magenta; n=3). f Representative image showing db-ISH of Grpr (blue) and Glp1r (red; n=4). g Representative image showing db-ISH of Ramp3 (blue) and Calcr (red; n=3). h Representative image showing db-ISH of Olfr78 (blue) and Calcr (red; n=4). Scale bar for panels c-h, 100μm.

DVC, dorsal vagal complex; poly(A), polyadenylation; pol II, RNA polymerase II; AP, area postrema; NTS, nucleus of the solitary tract; DMV, dorsal motor nucleus of the vagus nerve; IHC, immunohistochemistry; ISH, in situ hybridization; db, double; ES_μ, expression specificity.

Data to reconstruct panel b can be found in Supplementary Data 2.

We observed an overlap between neuronal populations expressing Glp1r and two other receptors known to be involved in DVC body weight control, namely Gfral and Calcr (Fig. 3b, Supplementary Data 2). Glu4^AP (Glp1r), as well as non-Glp1r Glu5^NTS and Glu13^NTS neurons, expressed Gfral (ES_μ=0.99, ES_μ=0.96, and ES_μ=0.73, respectively). We confirmed with immunohistochemistry (IHC) and ISH in mice and non-human primates that a subset of cells in the AP but not in the NTS co-expressed Glp1r and Gfral (Fig. 3c, Extended Data Fig. 4a). Calcr was expressed by Glu11^NTS (Glp1r) as well as Glu2^NTS and Glu10^AP neurons (ES_μ=0.91, ES_μ=0.56, and ES_μ=0.99, respectively), and applying double (db)-ISH in mice and non-human primates, we validated that Calcr co-localized with the majority of Glp1r cells in the NTS; in contrast, few Glp1r cells co-expressed Calcr in the AP (Fig. 3d, Extended Data Fig. 4b). The top marker genes for Glu4^AP (Glp1r/Gfral) and Glu11^NTS (Glp1r/Calcr) neurons were calcium-sensing receptor (Casr) and gastrin releasing peptide receptor (Grpr), respectively (Fig. 2f). The co-expression of Casr and Glp1r exclusively in the AP matched our db-ISH data in mice and non-human primates; as did the co-expression of Grpr with Glp1r and Calcr exclusively in the NTS (Fig. 3e–f, Extended Data Fig. 4c–d). Two other anorexic genes were expressed in Glu4^AP (Glp1r/Gfral) and Glu11^NTS (Glp1r/Calcr) neurons, namely brain derived neurotrophic factor (Bdnf; ES_μ=0.94 and ES_μ=0.89, respectively) and melanocortin 4 receptor (Mc4r; ES_μ=0.43 and ES_μ=0.94, respectively; Fig. 3b). Applying db-ISH, we validated the co-localization of Casr and Bdnf in the AP and the co-localization of Grpr with Bdnf and Mc4r in the NTS, though we were unable to detect any Mc4r expression in the AP of control mice (Extended Data Fig. 4e–h). Finally, Glu4^AP (Glp1r/Gfral) and Glu11^NTS (Glp1r/Calcr) neurons specifically expressed the three components to synthesize noradrenaline, namely Th (ES_μ=0.49 and ES_μ=0.96, respectively), Ddc (ES_μ=0.97 and ES_μ=0.63, respectively), and Dbh (ES_μ=0.83 and ES_μ=0.90, respectively); findings which we for Th and Ddc replicated with IHC experiments in Glp1r-Cre and Calcr-Cre mice (Extended Data Fig. 4i–l). Together these results identify noradrenergic Glp1r neurons in the AP and NTS of mice and non-human primates that express receptors capable of sensing distinct, but overlapping, anorectic signals.

We also detected appetite-regulating receptors in cell populations that did not express Glp1r and Gfral. Glu10^AP (Calcr) neurons expressed Ramp3 (ES_μ=0.99; Fig. 3b, Supplementary Data 2). The top marker of this population was olfactory receptor 78 (Olfr78; Fig. 2f), and the co-expression of Calcr with Ramp3 and Olfr78 exclusively in the AP aligned with our IHC and ISH data in mice and non-human primates (Fig. 3g–h, Extended Data Fig. 5a). This further corroborates that Glu10^AP (Calcr/Ramp3) neurons represent an amylin-sensing neuronal population. In addition, GABA4^AP and GABA5^AP neurons expressed gastrin inhibitory polypeptide receptor (Gipr; ES_μ=0.79 and ES_μ=0.98, respectively; Fig. 3b). Collagen and calcium-binding EGF domain-containing protein 1 (Ccbe1) was the top marker of GABA4^AP neurons and also specifically expressed in GABA5^AP neurons (ES_μ=0.97), while paired box 5 (Pax5) was the top marker of GABA5^AP neurons (Fig. 2f). Applying db-ISH, we validated that a subset of Ccbe1 and the majority of Pax5 cells co-localized with Gipr in the AP (Extended Data Fig. 5b–c). We also verified that Gipr and Glp1r were expressed in non-overlapping AP cells (Extended Data Fig. 5d). These results demonstrate that the partially overlapping expression of receptors for anorexic peptides are distributed across several distinct neuronal cell populations in the DVC of mice and non-human primates.

Furthermore, the NTS is known to comprise a small population of preproglucagon (Gcg)-expressing neurons encoding a precursor peptide for GLP-1²⁹. Indeed, we detected high expression of Gcg in Glu14^NTS neurons (ES_μ=0.95; Fig. 3b, Supplementary Data 2). In line with previous observations, this population also expressed Lepr (ES_μ=0.92; Fig. 3b)³⁰. The top marker gene of Glu14^NTS neurons was ankyrin repeat and SOCS box-containing 4 (Asb4, which is known to be highly-expressed and regulated in Lepr-expressing neurons³¹), which was consistent with our db-ISH data (Fig. 2f, Extended Data Fig. 5e). Finally, we were able to map additional DVC cell populations with previously-defined roles in metabolism onto our atlas (Extended Data Fig. 6).

Transcriptional programs induced by GLP-1 receptor agonists

To better understand the transcriptional response to GLP-1RA administration in AP and neighboring DVC cells, we applied weighted gene co-expression network analysis (WGCNA; a computational approach to identify sets of co-regulated genes associated with treatment outcomes) on the bulk RNA-seq data from the AP. In total, we identified 23 gene modules (M1-M23), four of which were associated with semaglutide administration (module M1 positively and modules M18–20 negatively; Bonferroni-adjusted P<2.0×10⁻⁴; Fig. 4a, Supplementary Data 4).

Fig. 4: Transcriptional changes induced by GLP-1RA administration in glutamatergic Glp1r neurons.

a Modules correlated with GLP-1RA administration. The bulk RNA-seq data was clustered into modules of co-regulated genes. Median, first and third quartiles, and whiskers with minimum and maximum 1.5 interquartile ranges are shown. P<0.05 are specified, semaglutide (n=14 mice) vs. weight-matched (n=13 mice), logistic regression with Bonferroni-adjusted likelihood ratio test. b Cell population enrichment of module genes. Top, dot size indicates the significance level of overlap between module genes and cell population marker genes (ES_μ>0). Bonferroni-adjusted one-tailed Fisher’s exact test. Bottom, the most likely DVC origin of the cell populations. c Five top-most enriched Gene Ontology terms for module M1. Bonferroni-adjusted g:Profiler P-value. d Top 30 genes for module M1. Dot size indicates log₂ fold-change between semaglutide and weight-matched animals. All genes were upregulated in the AP following semaglutide administration (Benjamini-Hochberg-adjusted DESeq2 P<0.05). e Differential expression of top 10 M1 genes in Glu4^AP neurons. Mean ± SEM. P<0.05 are specified, semaglutide (784 neurons, n=7 mice) vs. weight-matched (734 neurons, n=7 mice), Bonferroni-adjusted pseudo-bulk DESeq2 P-value. f Representative images showing FOS immunoreactivity (purple) and GFP immunoreactivity (green) in Glp1r-Cre;GFP mice treated with Exendin-4 (IP, 150 μg/kg; n=3) or vehicle (n=3). Scale bar, 150μm.

DVC, dorsal vagal complex; AP, area postrema; NTS; nucleus of the solitary tract; DMV, dorsal motor nucleus of the vagus nerve; GLP-1RA, GLP-1 receptor agonist; GFP, green fluorescent protein; IP, intraperitoneal.

Data to reconstruct panel d can be found in Supplementary Data 1, Supplementary Data 4, and through the NCBI Gene Expression Omnibus.

To assess whether these GLP-1RA-associated modules represented cell population-specific processes, we tested whether module genes selectively overlapped with cell population marker genes. The 581 genes in module M1 were broadly expressed in glutamatergic neurons and most strongly overlapped with Glu4^AP (Glp1r/Gfral) and Glu11^NTS (Glp1r/Calcr) neuron marker genes (Bonferroni-adjusted P<7.2×10⁻²⁰), suggesting that these populations, but not GABA7^AP (Glp1r) neurons, respond directly to GLP-1RA administration (Fig. 4b). The 67 genes constituting module M18 overlapped with marker genes of ependymal cells, tanycyte-like cells, and vascular and leptomeningeal cells (VLMCs), the 206 M19 genes were broadly expressed in glial cells, and the 143 M20 genes overlapped with marker genes of astrocytes and tanycyte-like cells (Bonferroni-adjusted P<0.05; Fig. 4b). These results demonstrate that GLP-1RA administration upregulates distinct biological processes in glutamatergic Glp1r neurons while decreasing transcriptional processes in glial cells.

To understand the transcriptional response to GLP-1RA administration in greater depth, we examined the enrichment of Gene Ontology (GO) biological processes and molecular functions in each module. We found that module M1 was most strongly associated with Modulation of Chemical Synaptic Transmission (Bonferroni-adjusted P=5.3×10⁻⁷; Fig. 4c), whereas M18 was most strongly associated with Calcium-Dependent Protein Binding, M19 with Rhythmic Process, and M20 Negative Regulation of Neuron Differentiation (Bonferroni-adjusted P<0.05, Extended Data Fig. 7). Among the top genes in module M1 were Bdnf, Mc4r, Vgf, Cartpt, prohormone convertase 1 (Pcsk1), and Th (Fig. 4d, Supplementary Data 4). Applying db-ISH in mice treated with Exendin-4 (Ex4; another GLP-1RA), we verified that Bdnf and Mc4r (which was undetectable in control mice) were upregulated in the AP upon GLP-1RA administration (Extended Data Fig. 4e, Extended Data Fig. 4g, Extended Data Fig. 8). Finally, the top genes of module M1 were indeed upregulated upon semaglutide administration in Glu4^AP (Glp1r/Gfral) neurons; four of these significantly, namely Vgf, peptidylglycine alpha-amidating monooxygenase (Pam), protein tyrosine phosphatase receptor type N (Ptprn), and Bdnf (Bonferroni-adjusted P<0.05; Fig. 4e). Together, these data suggest that GLP-1RA administration induces upregulation of neurotrophic and neuroendocrine genes in glutamatergic Glu4^AP (Glp1r/Gfral) neurons.

To further verify that Glp1r neurons in the AP and NTS are activated by GLP-1RA administration, we evaluated the expression of the activity-dependent immediate early gene FOS in Glp1r-Cre and Calcr-Cre mice before and after administration of Ex4. While GLP-1RA administration induced FOS immunoreactivity in Glp1r AP and NTS cells and a few Calcr NTS, no Calcr AP cells were positive for FOS upon Ex4 administration (Fig. 4f, Extended Data Fig. 9a). As expected, an induction of FOS immunoreactivity in Calcr AP and NTS cells were seen upon administration of sCT (Extended Data Fig. 9b). These results confirm selective activation of Glp1r AP and NTS neurons by GLP-1RAs.

Chromatin accessibility single-cell atlas of the dorsal vagal complex

To further characterize the regulatory landscape of DVC cells, we utilized snATAC-seq to profile chromatin accessibility in semaglutide-administered, ad libitum fed vehicle-administered, and weight-matched control animals (from the same in vivo study as used for the snRNA-seq data; Fig. 5a). The DNA fragment sizes of our snATAC-seq data displayed a classical decreasing distribution with peaks around 100 base pairs (bp), corresponding to nucleosome free regions, and around 200 and 400 bp, corresponding to regions bound by monoor dinucleosomes, respectively (Fig. 5b). We retained 22,545 quality-filtered cells and identified neurons and the previously described eight glial cell types in similar proportions as in the gene expression atlas (Spearman’s rho=0.92, P=0.001; Fig. 5c). Likewise, all neuronal populations, except three populations (GABA7^AP, Glu14^NTS, and Glu15^NTS neurons), were corroborated by snATAC-seq in comparable proportions (Spearman’s rho=0.91, P=3.7×10⁻⁶; Fig. 5d). Thus, our single-nucleus chromatin accessibility data aligned well with our single-nucleus gene expression data.

Fig. 5: Chromatin accessibility atlas of DVC cell populations.

a Overview of single-nucleus ATAC-seq study design. Diet-induced obese mice were randomized into three groups; 1) semaglutide-administered (n=5), 2) ad libitum fed vehicle-administered (n=5), and 3) weight-matched controls (n=5) that were dosed once daily for seven days as described for the bulk RNA-seq in vivo study. The AP, NTS, and DMV were dissociated and subjected to single-nucleus ATAC-seq. b Fragment size distribution for single-nucleus ATAC-seq data. c UMAP plot of the 22,545 cells colored by cell type. d UMAP plot of the 11,651 neurons colored by population. e Motif enrichment in non-neuronal cells. From top to bottom, dendrogram showing the cell type similarity in motif accessibility, number of cells per cluster, most enriched motifs colored by the expression specificity values (ES_μ) of the corresponding transcription factor genes. Enriched motifs were identified using logistic regression with Bonferroni-adjusted likelihood ratio test. f Motif enrichment in neuronal populations. From top to bottom, dendrogram showing the neuronal population similarity in motif accessibility, number of cells per cluster, most enriched motifs colored by the ES_μ values of the corresponding transcription factor genes, the most likely DVC origin of the neuronal populations. g Differential motif accessibility in Glu4^AP neurons for GLP-1RA-induced transcription factors. Mean ± SEM. **P=0.007, semaglutide (182 neurons, n=5 mice) vs. weight-matched (239 neurons, n=5 mice), pseudo-bulk logistic regression with Bonferroni-adjusted likelihood ratio test.

DVC, dorsal vagal complex; AP, area postrema; NTS; nucleus of the solitary tract; DMV, dorsal motor nucleus of the vagus nerve; ATAC, assay for transposase-accessible chromatin using sequencing; RNA-seq, RNA-sequencing; bp, base pairs; UMAP, uniform manifold approximation and projection; OPCs, oligodendrocyte precursor cells; VLMCs, vascular leptomeningeal cells; GLP-1RA, GLP-1 receptor agonist

Data to reconstruct panels b-f can be found in Supplementary Data 5 and/or through the NCBI Gene Expression Omnibus.

We next investigated whether we could identify relevant cell population-specific transcription factor binding sites (henceforth motifs). Across the glial cell types, we identified a total of 174 enriched motifs (Bonferroni-adjusted P<0.05; average of 45 motifs per cell population; Fig. 5e, Supplementary Data 5). There was a significant correlation between the motif enrichment and the expression specificity of the corresponding transcription factor (median Spearman’s rho=0.21; P=1.1×10⁻²³), suggesting our data allows for identification of relevant cell population-specific accessible motifs. Across the neuronal populations, we identified 315 enriched motifs (Bonferroni-adjusted P<0.05; average of 45 enriched motifs per neuronal population; Fig. 5f, Supplementary Data 5). We again observed a correlation between motif enrichment and transcription factor expression specificity (median Spearman correlation=0.11, Fisher’s method P=7.6×10⁻²⁶). There was a high similarity between neurons of the same neurotransmitter class at the motif level (dendrogram in Fig. 5f). Among the most enriched motifs in Glu4^AP (Glp1r/Gfral) and Glu11^NTS (Glp1r/Calcr) neurons were PHOX2A and PHOX2B, which are involved in the transcription of Dbh, consistent with the noradrenergic profile of these neurons³².

Finally, we assessed whether GLP-1RA administration induced changes at the chromatin level. Since 16 genes in the semaglutide-associated module M1 encoded transcription factors, we asked whether the corresponding motifs were enriched in Glu4^AP (Glp1r/Gfral) neurons from semaglutide-administered animals. Fifteen out of 16 motifs displayed enriched accessibility upon GLP-1RA administration of which eight remained significant after multiple testing correction (Bonferroni-adjusted P<0.05; Fig. 5g). Four of these motifs (FOS, FOSL1, FOSL2, and JUNB) can be bound by stimulus-induced early response gene transcription factors³³. These results reinforce the above observations that Glu4^AP (Glp1r/Gfral) neurons are activated by GLP-1RA administration.

Assessing relevance in humans using genetic data

To assess the extent to which DVC cell populations are likely to impact genetic susceptibility to obesity, we integrated our single-cell atlases with GWAS data for BMI. Specifically, we asked whether human orthologs of genes marking cell populations non-randomly co-localized with genetic variants associated with BMI. We subsequently validated these findings by testing whether cell population-specific motifs also preferentially co-localized with BMI-associated genetic variants (Fig. 6a).

Fig. 6: Glp1r and Calcr neurons enrich for genetic variants associated with BMI.

a Overview of BMI GWAS integration. The expression specificity profile (top) or the motif enrichment (bottom) for each cell population was integrated with BMI GWAS data to compute the genetic enrichment in each cell population. b BMI GWAS integration with gene expression and c motif accessibility. Top, genetic enrichment. Bonferroni-adjusted CELLECT P-value. Dashed line indicates significance threshold. Bottom, the most likely DVC origin of the neuronal populations. d Representative image showing tdTomato immunoreactivity (red) in Calcr-Cre;tdTomato reporter rats (n=5). e Representative image showing mCherry immunoreactivity (red) in Calcr-Cre injected with hM3Dq-mCherry in the AP (n=4). Scale bar for panels d-e, 200μm. f Food intake in Calcr-Cre rats injected with hM3Dq-mCherry in the AP following treatment with saline (n=4) or CNO (IP, 1 mg/kg; n=4). g Food intake in Calcr-Cre rats following treatment with saline (n=10) or CNO (IP, 1 mg/kg; n=10). Mean ± SEM are shown. P<0.05 are specified, linear mixed effects model, Bonferroni-adjusted two least squares means two-tailed t-test.

BMI, body mass index; GWAS, genome-wide association study; OPCs, oligodendrocyte precursor cells; VLMCs, vascular leptomeningeal cells; DVC, dorsal vagal complex; AP, area postrema; NTS; nucleus of the solitary tract; DMV, dorsal motor nucleus of the vagus nerve; CNO, Clozapine-N-oxide; IP, intraperitoneal.

Applying CELLECT⁷ to address these questions, we identified four cell populations, exclusively glutamatergic neurons, comprising marker genes co-localizing with BMI-associated genetic variants (Bonferroni-adjusted P<0.05; Fig. 6b). Glu10^AP (Calcr/Ramp3) neurons displayed the strongest enrichment, implicating a role for amylin-sensing AP neurons in body weight control. Glu9^NTS neurons (top marker gene, Eya1) displayed the second strongest enrichment. Finally, we also detected enrichments for Glu4^AP (Glp1r/Gfral) and Glu11^NTS (Calcr/Glp1r) neurons. These results suggest that Calcr and/or Glp1r-expressing AP and NTS neurons may contribute to the genetic control of body weight.

Integrating the BMI GWAS with the chromatin accessibility data supported the above results (Spearman’s rho=0.8, P=1.3×10⁻⁶); five exclusively glutamatergic neuronal populations comprised accessible motifs near BMI-associated genetic variants (Bonferroni-adjusted P<0.05; Fig. 6c). Again, Glu10^AP (Calcr/Ramp3) neurons displayed the strongest enrichment, thus establishing that both the gene expression and chromatin accessibility profiles of these neurons position them as a likely regulator of susceptibility to obesity. While Glu4^AP (Glp1r/Gfral) and Glu11^NTS (Glp1r/Calcr) neurons were nominally enriched, these results did not withstand adjustment for multiple testing (non-adjusted P=1.7×10⁻³ and P=4.7×10⁻³, respectively).

We conducted a meta-analysis on the results from the integration of BMI GWAS with gene expression and chromatin accessibility data. Top-ranking cell populations mediating genetic susceptibility to obesity were 1) Glu10^AP (Calcr/Ramp3), 2) Glu9^NTS (Eya1), 3) Glu4^AP (Glp1r/Gfral), and 4) Glu11^NTS (Calcr/Glp1r) neurons (Bonferroni-adjusted P<3.0×10⁻³; Supplementary Data 6). Together these results imply a physiological role of AP and NTS glutamatergic neurons, especially those that express Calcr and/or Glp1r, in the control of food intake and body weight, consistent with the capabilities of pharmacological GLP-1RA and CALCR agonist administration to reduce body weight. These results are also consistent with previous findings¹⁷ and our observation that activation of Calcr NTS neurons suppresses food intake and body weight (Extended Data Fig. 10).

We next tested the predicted role of Glu10^AP (Calcr/Ramp3) neurons in the control of food intake. Because stereotaxic manipulation of AP neurons in mice is difficult, we used CRISPR–Cas9 genome engineering to generate a Calcr-Cre knock-in rat model (Fig. 6d, Supplementary Fig. 1).

We injected AAV-hM3Dq-mCherry, which expresses the activating hM3Dq designer receptor exclusively activated by designer drugs (DREADD) in a Cre-inducible manner, into the AP of Calcr-Cre rats and confirmed that the DREADD expression was specific to the AP (Fig. 6e, Supplementary Fig. 2). We then examined food intake over 24 hours in these rats following the injection of saline or clozapine-N-oxide (CNO, which activates hM3Dq), revealing that the DREADD-mediated activation of Calcr AP neurons in rats durably decreased food intake, consistent with a role for Glu10^AP (Calcr/Ramp3) neurons in the control of feeding (Fig. 6f). Importantly, CNO injection in Calcr-Cre rats without DREADD expression did not alter feeding (Fig. 6g). The observed anorexic effects following activation of Calcr AP and NTS neurons reinforce their potential utility as targets for therapeutic intervention in obesity.

Discussion

Here we investigated the contribution of DVC cell populations in the pharmacological and physiological control of energy balance. Applying snRNA- and snATAC-seq, we identified eight non-neuronal and 25 neuronal populations with distinct gene expression and motif accessibility landscapes. Combining our single-cell atlases with bulk RNA-seq, we show that GLP-1RA-induced alterations in gene expression and chromatin accessibility are specific to AP neurons that express Glp1r and Gfral and NTS neurons that express Glp1r and Calcr. These neuronal populations, along with AP neurons that express Calcr and Ramp3, exhibit marker genes and accessible motifs co-localizing with BMI-associated genetic variants. Furthermore, using a novel Calcr-Cre rat model, we show that activation of Calcr AP neurons suppresses food intake. Together, these findings implicate Glp1r/Gfral AP neurons, Glp1r/Calcr NTS neurons, and Calcr/Ramp3 AP neurons in the genetic predisposition to obesity and suggest that these cell populations represent relevant targets for the therapy of obesity.

Several of the 33 cell populations identified here align well with previous observations. First, we report an overrepresentation of glutamatergic (Glu1–15) neurons in the AP and NTS²⁴. Second, we identified a partial overlap between Glp1r and Gfral (Glu4) neurons exclusively in the AP³⁴. Third, we report a noradrenergic phenotype of Glp1r (Glu4 and Glu11) neurons in the AP and NTS^10,28. Fourth, we detected a population of Calcr (Glu10) neurons co-expressing Ramp3 in the AP³⁵. Finally, we identified distinct glial cell types with previously characterized marker genes. Among these populations, VLMCs, ependymal, and endothelial cells were more strongly enriched in the AP in the DVC consistent with increased vascularization of circumventricular organs³⁶. We extend these findings by showing that: 1) there exists a subset of neurons (Glu11) containing Glp1r and Calcr in the NTS; 2) that Gipr is expressed in GABAergic (GABA5) neurons in the AP; and 3) that Casr, Grpr, Olfr78, and Pax5 mark Glp1r (Glu4) AP, Glp1r/Calcr (Glu11) NTS, Calcr (Glu10) AP, and Gipr (GABA5) AP neurons, respectively. During the revision phase of this paper, a single-nucleus gene expression atlas focused on nausea-promoting cells of the AP was published³⁷. The AP and nausea-focused study supports a number of findings reported here, the two most important being that distinct glutamatergic neuronal populations in the AP express Glp1r, Gfral, and Casr (Glu4) while others express Calcr, Ramp3, and Olfr78 (Glu10), and that Gipr is expressed by GABAergic (GABA5) AP neurons. In addition to Eya1 (Glu9) NTS neurons that represent candidates for the control of energy balance, other cell populations identified in our atlas with relevance to metabolism include Gfral (Glu5) NTS neurons, Lepr/Gcg (Glu14) NTS neurons that suppress feeding^30,38, Mc4r (Chat3) DMV neurons that are involved in insulin regulation³⁹, Hsd11b2 (Glu2) NTS neurons that drive sodium appetite⁴⁰, and Lepr/Gal (Glu3) NTS neurons that modulate respiration⁴¹ (Fig. 7). These consistencies in DVC cell population identities with previous studies and our non-human primate db-ISH data further position our transcriptional and chromatin accessibility single-cell atlases as valuable resources for future studies directed towards the DVC.

Fig. 7: Overview of DVC neuronal populations and their role in metabolism.

Single-nucleus RNA-seq profiling of 49,392 neurons in the DVC identified a total of 25 neuronal populations distributed across the AP (n=5), NTS (n=15), and DMV (n=5). Integrating the single-cell data with GWAS data for BMI showed that Glu4^AP, Glu9^NTS, Glu10^AP, and Glu11^NTS neurons expressed marker genes with human orthologs that non-randomly co-localized with BMI-associated genetic variants. Glu10^AP, Glu11^NTS, and Glu14^NTS[30,38] neurons suppress feeding when activated by DREADDs. Glu4^AP neurons are activated by GLP-1RAs, Glu10^AP neurons by sCT, and Glu11^NTS neurons by GLP-1RAs and sCT. Chat3^DMV neurons regulate circulating insulin³⁹, Glu2^NTS neurons drive sodium appetite⁴⁰, and Glu3^NTS neurons modulate breathing via leptin-mediated mechanisms⁴¹. We note that functions for other DVC populations remain to be determined. Within each of the three DVC nuclei, the spatial placement of the neuronal populations is arbitrary. Magnified neuronal shapes denote neuronal populations with genetic enrichments, potential or proven utility as targets for therapeutic intervention in obesity, or roles in central regulation of feeding or other metabolic processes. Shapes and colors indicate neurotransmitter type (circle, glutamatergic; square, GABAergic; diamond, cholinergic).

DVC, dorsal vagal complex; BMI, body mass index; AP, area postrema; NTS, nucleus of the solitary tract; DMV, dorsal motor nucleus of the vagus nerve; GWAS, genome-wide association study; BMI, body mass index; DREADD, designer receptors exclusively activated by designer drugs; GLP-1RA, GLP-1 receptor agonist; sCT, salmon calcitonin.

Among the six brain areas transcriptionally characterized in our work, GLP-1RA administration induced the strongest transcriptional changes in the AP. We did not identify major differences between semaglutide and liraglutide administration, suggesting that the superior weight-lowering effects of semaglutide is mediated through brain areas or mechanisms not covered by our analyses. Genes upregulated following semaglutide administration (captured in module M1) were specifically expressed in glutamatergic Glp1r/Gfral (Glu4) AP and Glp1r/Calcr (Glu11) NTS neurons, supporting a role for glutamatergic neurons in GLP-1RA-induced weight loss²⁴ and aligning with observations that NTS intraparenchymal delivery of GLP-1RA suppress food intake^15,16. These genes were not specific to Gcg (Glu14) neurons in the NTS, consistent with previous observations showing that GLP-1RAs have anorexic effects independently of Gcg NTS neurons, and that these neurons have limited interactions with AP neurons and vice versa^30,42. Top genes of the GLP-1RA-induced response included Bdnf, Mc4r, and Pcsk1. Several of these genes have established roles in body weight control: 1) protein-altering mutations and deletions in BDNF, MC4R, and PCSK1 can cause early-onset severe obesity in humans^43,44, 2) DVC intraparenchymal infusion of BDNF and MC4R agonists both reduce food intake and body weight^45,46, and 3) body weight suppressive effects of GLP-1RA administration is attenuated with DVC intraparenchymal infusion of a MC4R antagonist suggesting that GLP-1RA stimulation depends on MC4R activation by endogenous ligands such as α-melanocyte-stimulating hormone (α-MSH)⁴⁷.

A key strength of our approach is that we can implicate cell populations with relevance to human obesity by matching hundreds of genetic loci associated with BMI to the gene regulatory signatures of DVC cell populations. Our data-driven approach associated four neuronal populations with genetic obesity predisposition (Glu4, Glu9, Glu10, and Glu11), three of which carry pharmacological relevance by expressing Glp1r and/or Calcr. These findings indicate their long-term weight-regulatory relevance, and indeed, knock-out of Calcr in the AP and NTS blunts the anorexic effects of amylin¹⁸, while both Glp1r NTS knock-down and intraparenchymal delivery of Exendin-9 (a GLP-1R antagonist) increases food intake^48,49. Furthermore, knock-out of Calcr in the NTS attenuates the anorexic effects of sCT, while Calcr neurons in the AP may also contribute to the sCT response as AP lesions blunt both amylinand sCT-mediated suppression of food intake¹⁴. Our findings further establish the importance of Calcr (Glu10 and Glu11) neurons in the DVC and show that activation of both Calcr AP and Calcr NTS neurons decrease food intake. In contrast to Calcr NTS neurons, the vast majority of Calcr AP neurons did not express Glp1r, however, additional experiments are needed to fully rule out the possibility that Glp1r neurons contribute to the Calcr AP-mediated feeding suppression. We additionally report a previously uncharacterized population of Eya1 (Glu9) NTS neurons to be associated with susceptibility to obesity. The exact role of these neurons in body weight control will have to be addressed in future studies.

Nausea, an unpleasant feeling of visceral malaise, represents an important therapy-limiting side effect of previous drug candidates targeting the AP⁵⁰. Hence a key question remains whether any of the likely body weight-regulatory DVC neuronal populations induce nausea. Notably, Calcr AP and Calcr NTS neurons do not to promote conditioned taste aversion in mice (a response that can be indicative of nausea)^17,37. In contrast, activation of Glp1r AP and Gfral AP neurons induce conditioned taste aversion³⁷. Due to the co-localization of Calcr and Glp1r in the NTS and the non-aversive nature of Calcr NTS neurons¹⁷, it seems less likely that Glp1r NTS neurons are aversive. The findings in rodents are partly contradicted by phase III clinical data in type 2 diabetes patients, showing that semaglutide-induced weight loss can only to a minor extent by explained by nausea or vomiting⁵¹. Thus, whereas at least activation of rodent Glp1r AP neurons may lead to nausea, the feeding response driven by activation of Calcr AP neurons, the cell population most strongly enriched for genes co-localizing with BMI GWAS variants, is less likely to be caused by nausea. However, the present study does not fully eliminate aversion as a possible mechanism of reduced feeding.

Despite its strengths, our study comes with a number of limitations. First, there were slight differences between the two in vivo studies performed here; semaglutide-administered mice weighed approximately one gram less than their weight-matched controls in the single-nucleus RNAand ATAC-seq-focused in vivo study, whereby the difference in food intake was not as significant as in the bulk RNA-seq-focused study. Second, due to a typically relatively low messenger RNA-capture rate in the droplet-based single-cell techniques, we cannot rule out that additional relevant genes are expressed in the cell populations characterized here. For instance, although α-MSH is a proopiomelanocortin (POMC)-derived peptide reported to be expressed in the ARH and NTS⁵², we were not able to identify Pomc NTS neurons in our DVC atlas. Third, although our results imply a role for Glp1r AP neurons as key mediators of GLP-1RA-induced weight loss, other studies have shown that the AP may not be required for these anorectic effects^37,53,54. This apparent contradiction may be explained by the expression of Glp1r in the NTS and elsewhere in the brain and by the access of long-acting GLP-1RAs to several sites contributing to the efficacy of GLP-1RA administration²⁵. Finally, the genetic enrichment analysis relies on the assumption that gene expression levels and chromatin accessibility are relatively conserved between the mouse and human species.

In conclusion, our results suggest that Glp1r and Calcr neurons in the AP and NTS represent therapeutic targets that express genes and contain accessible DNA sequence motifs with likely roles in predisposition to obesity in humans, and our single-cell transcriptomics and chromatin accessibility atlases constitute comprehensive molecular resources for further exploration of DVC cells in obesity and beyond.

Methods

Bulk and single-nucleus RNAand ATAC-seq

Mice

All in vivo experiments were conducted in accordance with internationally accepted principles for the care and use of laboratory animals approved by the Danish Ethical Committee for Animal Research.

For the bulk RNA-seq experiment, seven weeks old C57BL/6 male mice were obtained from JanVier Labs (Saint-Berthevin Cedex, France). Upon arrival to the animal unit, mice were group-housed 10 per cage with ad libitum access to 60% HFD (#12492 Research Diets) and tap water. Animals were fed the 60% HFD for at least 24 weeks prior to study start to ensure DIO. The animal room environment was controlled (temperature 22 ± 2°C; relative humidity 50 ± 10%; 12-hour light/dark cycle).

For the snRNA-seq experiment, 20 weeks old DIO (16 weeks of HFD) male C57BL/6 mice were obtained from Charles River (France). Upon arrival to the animal unit, mice were single-housed with ad libitum access to 60% HFD (#12492 Research Diets) and tap water for six weeks prior to study start. The animals were housed as described above.

Randomization and dosing

For the bulk RNA-seq experiment, two weeks before study start animals were single-housed and their daily food intake was recorded for the last six days prior to study start. On the day before study start animals were block-randomized based on body weight into the following groups; 1) semaglutide-administered (n=15), 2) liraglutide-administered (n=15), 3) vehicle-administered and ad libitum fed (n=15), and 4) vehicle-administered and weight-matched to the semaglutide group (n=15). For each group, animals were divided into two subgroups where dosing was initiated with one day’s interval. Animals were dosed once daily, subcutaneous, for seven days approximately three hours before lights off. Dosing of the animals was gradually up-titrated with daily increments to avoid initial discomfort and potential dehydration: 0.02, 0.04, 0.1 mg/kg for semaglutide and 0.2, 0.4, 1.0 mg/kg for liraglutide. Animals were terminated by decapitation under CO₂/O₂ anaesthesia the day after the last dose. All animals were terminated in three to four hours after lights on in the morning the day after the last dose, to ensure minimal variation in gene expression caused by circadian rhythm.

The above experimental set-up was replicated for the snRNAand snATAC-seq experiments unless otherwise described. On the day before study started animals were block-randomized based on body weight into the following groups; 1) semaglutide-administered (n=9; snRNA-seq, n=4; snATAC-seq, n=2; snRNAand snATAC-seq, n=3), 2) vehicle-administered and ad libitum fed (n=9; snRNA-seq, n=4; snATAC-seq, n=3; snRNAand snATAC-seq, n=2), and 3) weight-matched to the semaglutide group (n=8; snRNA-seq, n=3; snATAC-seq, n=1; snRNAand snATAC-seq, n=4). Please refer to Reporting Summary for additional information on the in vivo studies and related experiments.

Generation of bulk RNA-seq data

After decapitation, brains were stored at −80°C until further processing. The frozen brains were slightly thawed to −20°C and divided into three parts (forebrain, midbrain, and hindbrain) by two dorsoventrally cuts at the optic chiasm and just rostral to the cerebellum. Each part of the brain was cut on a cryostat (CM3050S, Leica, Germany) into 12 mm thick coronal sections, and serial sections were collected on Arcturus PEN membrane slides (Applied Biosystems, Foster City, USA). Slides were stored at −80°C until laser capture microdissection (LCM). Sectioning was optimized for each region to ensure a sufficient RNA (approximately 30 ng) for the sequencing process; LS, Bregma 0.74 to 0.14 mm; PVH, Bregma −0.58 to −1.22 mm; ARH, Bregma −1.46 to −2.18mm; DMH, Bregma −1.58 to −2.18 mm; NTS, Bregma −7.20 mm to end of NTS; AP, Bregma −7.20 mm to end of AP. Following LCM, RNA extraction was performed using the PicoPure^™ RNA Isolation Kit (Applied Biosystems, Foster City, USA) as recommended by manufacturer, including an Ambion^™ DNase treatment (Invitrogen, Carlsbad, USA).

Generation of AP single-nucleus suspensions

After decapitation, the brains were excised, and the AP was directly removed from the brainstem by dissection and stored at −80°C. Nuclei were purified with reagents from the Nuclei Pure Prep Nuclei Isolation Kit (Sigma-Aldrich, NUC-201). Lysis buffer (LB) and 1.8 M sucrose cushion solution (SC) was prepared as recommended by the manufacture, with the addition of 80 U Protector RNase Inhibitor (Sigma-Aldrich, 3335399001)/ml LB. All reagents and samples were kept cold on wet-ice throughout the process. Samples were thawed on ice for 1 min and homogenized in 10–20 μl LB by pipetting up and down till the tissue appeared thoroughly degraded. Then LB to 50 μl was added, and the homogenate was gently laid on top of two pre-vetted filters (a pluriStrainer Mini 40 μm on top of a pluriStrainer Mini 20 μm (pluriSelect)). The filters were rinsed with 2 x 200 μl LB and next 2 x 500 μl SC. The top filter was discarded and the sample plus the remaining filter was spun for 30 s at 290 g. The filtered sample was mixed by pipetting and carefully laid on top of 500 μl SC and spun for 10 min at 10,000 g at 4°C. The top layer was removed, leaving ~500 μl sample with the pellet. PBS-BSA (PBS with 1% BSA, sterile filtered) was added to a total of 2 ml, the sample was mixed by pipetting, and spun for 5 min at 500 g at 4°C. The supernatant was carefully removed, leaving ~100 μl sample with the pellet, and the PBS-BSA wash was repeated. Following the final spin, ~100 μl sample was left after removal of supernatant, and the pelleted nuclei were resuspended in this remaining sample buffer. The nuclei concentration was measured, and 9000 nuclei were used for snRNA-seq and snATAC-seq, respectively.

Bulk RNA-seq library preparation

From the extracted RNA, stranded messenger RNA libraries were produced using the TruSeq Stranded mRNA Kit (Illumina, San Diego, USA), followed by sequencing of single-end, 75 base pair reads on a NextSeq 500 platform (Illumina, San Diego, USA) to a depth of approximately 1.5×10⁷ reads per sample.

Single-nucleus RNAand ATAC-seq library preparation

Libraries for snRNA-seq and snATAC-seq were prepared using the 10x Genomics Chromium single-cell 3’ reagent kits (version 2) and Chromium single-cell ATAC reagent kits as recommended by the manufacturer. Libraries were sequenced on an Illumina NovaSeq^™ 6000 and a NextSeq 500^™ sequencing system to a depth of approximately 2.3×10⁸ and 7×10⁷ reads per sample for the snRNA-seq and snATAC-seq libraries, respectively.

In vivo study data statistical analyses

Statistical analyses on mouse in vivo study data was performed using a linear mixed effects model with food intake or body weight as dependent variable, treatment group and time point as independent variables with interaction effects, and mouse as random effect. Group contrasts were tested using least-squares means two-tailed t-test, and P-values were adjusted for multiple testing using the Bonferroni method.

Bulk and single-nucleus RNA-seq data analysis

Bulk RNA-seq raw data processing

Base calling and demultiplexing was performed with Illumina bcl2fastq version 2.17.1.14. The sequencing data was aligned to the Mus musculus (mm10) genome using STAR version 2.5.2a with default parameters. STAR was also used for feature counting. Quantification and identification of differentially expressed genes, reported in Supplementary Data 1, were carried out using DEseq2⁵⁵, and P-values were adjusted for multiple testing using the BH method.

Reannotation of the Glp1r genomic coordinates

The PolyA_DB database⁵⁶ version 3.2 was used to identify genomic coordinates for alternative polyadenylation sites in the Glp1r and Ghsr genes. UCSC liftOver⁵⁷ was used to convert the coordinates from the genome build mm9 to mm10. The transcription end coordinates for Glp1r and Ghsr were reannotated to be the most downstream polyadenylation sites.

Single-nucleus RNA-seq raw data processing and quality control

Raw sequencing reads were processed using the 10x Genomics Cell Ranger version 3.0 pipeline and aligned to the Mus musculus (mm10) genome with default parameters. For each sample, a unique molecular identifiers (UMIs) count matrix was generated using both exonic and intronic reads. Normalization, alignment, and dimensionality reduction were performed using Seurat⁵⁸ version 3.1.1. Cells with a mitochondrial RNA content >5% were discarded (n=2,003) and doublets were removed (n=2,815) using DoubletFinder⁵⁹ version 2.0.0. Counts for each cell were normalized by the total gene expression of that cell, multiplied by 10,000, and log-transformed. To remove variance attributed to batch for downstream visualization and clustering, samples were integrated using the FindIntegrationAnchors and IntegrateData Seurat functions, which employs canonical correlation (CCA) analysis followed by mutual nearest neighborhood detection to align cells across samples. The integrated dataset was centered and scaled, PCA was carried out, and the top 30 principal components were used as input for UMAP dimensionality reduction. The above quality controls reduced our snRNA-seq dataset from 82,413 to 77,605 cells.

Single-nucleus RNA-seq cell population identification

Cell clustering was carried out on the integrated dataset using the two Seurat functions FindNeighbors and FindClusters using Louvain community detection. Major cell populations were separated by an initial round of clustering and classified based on the expression of known marker genes (resolution=0.3, chosen based on best visual separation of non-neuronal cell clusters). Mapping of microglia and tanycyte-like cells between the present and the published ARH-ME dataset²² was performed by projecting the ARH-ME dataset PCA structure onto the present dataset followed by mutual nearest neighborhood detection using the two Seurat functions FindTransferAnchors and TransferData (confidence score>0.5). For each cell, the silhouette index, a measure quantifying the similarity of a given cell to its assigned cluster compared to other clusters, was computed, and subsequently, in order to remove cells that could not confidently be assigned to a single cluster, cells with a negative silhouette index were discarded (n=1,822).

After these quality control steps, it was evident that a few cells labelled as astrocytes, oligodendrocytes and microglia were most similar to other cell populations than their assigned ones in UMAP space. To remove likely missed doublets, all genes across all cells were clustered into modules of co-expressed genes using WGCNA⁶⁰. Cells that were not annotated as oligodendrocytes or OPCs and loaded high (module eigengene>5% quantile of oligodendrocytes) on the oligodendrocyte-specific module were removed. Likewise, cells that were not astrocytes, VLMCs, tanycyte-like, endothelial, or ependymal cells and loaded high on the astrocyte-specific module were removed as were non-microglial cells that loaded high on the microglia-specific module. During this filtering step 758 cells were discarded.

Finally, to identify robust neuronal populations we applied the following two-step procedure. First, neurons (n=52,289) were clustered at 10 different low resolutions (0.01–0.1), the run with the highest average silhouette index was chosen (resolution=0.1; 11 clusters). Cells that could not be confidently assigned to a major neuronal class were discarded (removing cells with negative silhouette indices). Second, neurons (mean silhouette index, 0.31; median silhouette index, 0.30) were subclustered in a more fine-grained manner by testing 100 different resolutions varying from 0.01 to 1. Each run was scored by the silhouette index, and the run exhibiting the highest average silhouette index was chosen for all downstream analyses (resolution=0.44; 25 clusters; mean silhouette index, 0.34; median silhouette index, 0.36). This final filtering step removed 2,897 neuronal cells, leaving us with 49,392 neuronal and 22,736 non-neuronal cells for the downstream analyses (total=72,128). The mean and median number of transcripts recovered per cell were 2,093 and 1,354, respectively, and the mean and median number of unique genes recovered were 1,112 and 876, respectively.

Expression specificity

Cell population marker genes were identified with the tool CELLEX⁷ version 1.1.1 (retrieved from https://github.com/perslab/CELLEX) using default parameters. For glial cell types, the gene with the highest CELLEX score that was also a marker of the corresponding glial cell type in the hypothalamus^22,61,62 was depicted. For neuronal cell populations, the gene with the highest CELLEX score that was expressed in >10% of cells of the corresponding cell population was reported.

Enrichment of AP and NTS markers

Differentially expressed genes between the bulk RNA-seq AP and NTS samples, reported in Supplementary Data 3, were identified using DESeq2⁵⁵, and P-values were adjusted for multiple testing using the BH method. AP and NTS markers were ranked by P-values, and cell population marker genes (ES_μ>0) were tested for the enrichment of the top 1000 marker genes for the AP and NTS, respectively, using a one-tailed Fisher’s exact test. P-values were adjusted for multiple testing using the Bonferroni method (number of cell populations x number of brain areas).

Weighted gene co-expression network analysis

WGCNA⁶⁰ was run using the R implementation. The bulk RNA-seq data was subject to variance stabilizing transformation normalization using DESeq2⁵⁵, and genes with a variance of zero were removed. Using the biweight midcorrelation, a similarity matrix was computed from which a signed network was constructed using a soft thresholding power of five, maximizing the scale free topology R² fit. Genes were clustered hierarchically based on the average topological overlap measure, and modules of co-expressed genes were identified using the cutreeDynamic function with the parameters minClusterSize set to 30, deepSplit set to three, and pamStage parameter set to false. Finally, the Pearson correlation between the module eigengenes was computed, and modules with a correlation above 0.75 were merged.

Gene ontology analysis

Modules were tested for enrichment of GO terms (categories; Molecular Function, Biological Process) using the gProfiler⁶³ R implementation. Module genes were ordered by their kME (module membership) values, and the geneset enrichment analysis was carried out with the parameters ordered_query set to true, max_set_size limited to 500, and hier_filtering set to strong.

Module and treatment associations

The association between the module eigengene and treatment was tested using logistic regression and reported in Supplementary Data 4. A logistic model with the module eigengene as dependent and treatment group as independent variable (degrees of freedom (df)=2) was constructed for each module and compared to the null model (df=1) using a likelihood ratio test. P-values were adjusted for multiple testing using the Bonferroni method (adjusting for the number of modules).

Cell population enrichment for module genes

Cell population marker genes (ES_μ>0) were tested for enrichment of module genes using a one-tailed Fisher’s exact test. P-values were adjusted for multiple testing using the Bonferroni method (adjusting for the number of cell populations x number of modules).

Pseudo-bulk differential gene expression analysis

A pseudo-bulk gene expression matrix was generated by summing the transcripts counts for all cells with the same cell population and sample combination. DESeq2⁵⁵ was applied on the pseudo-bulk gene expression matrix, and P-values were adjusted for multiple testing using the Bonferroni method (adjusting for the number of genes).

Single-nucleus ATAC-seq data analysis

Single-nucleus ATAC-seq raw data processing and quality control

Raw sequencing reads were processed using the Cell Ranger ATAC version 1.1.0 pipeline and aligned to the Mus musculus (mm10) genome with default parameters. The aligned reads were further processed using SnapATAC⁶⁴ version 1.0.0. Fragments that were uniquely mapped (mapping quality score>30) and with proper length (50 bp<length<1000 bp) were kept. Based on visual inspection, high-quality cells were detected based on the number of unique fragments and fragments in promoter ratio (selecting cells with 1000<unique fragments<100,000; 0.15<fragments in promoter ratio<0.6). Using these criteria, 29,743 cells were called for downstream analysis.

The genome was segmented into 5 kb bins, and the chromatin accessibility profiles were represented as a binary matrix (1 denoting accessible chromatin; 0 denoting inaccessible chromatin). Bins overlapping blacklisted regions (retrieved from http://mitra.stanford.edu/kundaje/akundaje/release/blacklists/), mapping to unwanted (mitochondrial-, sex-, or unknown) chromosomes or exceeding a high coverage (95% quantile), thus likely representing invariable genomic regions, were removed. The binary cell-by-bin matrix was converted to a cell-by-cell distance matrix (Jaccard metric) and was normalized using the regression-based normOVE method implemented in SnapATAC. PCA was performed, and principal components 2–40 (determined based on visual inspection of the knee plot and pairwise PCA plots) were subjected to Leiden clustering using the resolution resulting in the highest average silhouette index (resolution: 0.1,0.2, …,1). These steps resulted in 31 clusters (resolution=0.9). Cells belonging to the same cluster were aggregated, and accessible peaks were called clusterwise using MACS2⁶⁵ version 2.2.6, generating a peak count matrix (253,507 peaks x 29,743 cells). These clusters were solely used to call peaks as described in the next section.

Single-nucleus ATAC-seq cell population identification

A Seurat object was initialized from the peak count matrix. Using Signac⁵⁸ version 0.1.5, the peak count matrix was normalized using term-frequency inverse-document-frequency normalization, and the 95% most common features were identified and used as input for latent semantic indexing (LSI) dimensionality reduction. The first 30 singular values were used as input for UMAP dimensionality reduction. A gene activity matrix (henceforth pseudo-snRNA-seq matrix) used for snATAC-seq cell population labelling was computed from the chromatin accessibility within and 2000 bp upstream of protein coding genes and stored in the present Seurat object. The gene activity counts for each cell were normalized by its total gene activity, multiplied by the median gene activity across all cells, and log-transformed. To annotate cells, CCA and mutual nearest neighborhood detection were performed using the Seurat function FindTransferAnchors with the snRNA-seq and pseudo-snRNA-seq datasets as inputs. TransferData was used to project the labels from the snRNA-seq to the snATAC-seq dataset. Cells with prediction scores below 0.5 were discarded (n=28). As the last filtering step, cells with a negative silhouette coefficient computed from the top 30 peak count matrix singular values were removed (n=478).

The following steps were applied to classify neuronal populations. First, major neuronal populations were identified by projecting labels from the neuronal snRNA-seq to the snATAC-seq dataset as described above. Neurons that could not be confidently classified into one of the 11 neuronal populations (prediction score<0.5) were discarded (n=507). Second, as for the snRNA-seq data processing, neurons with a negative silhouette index were removed (n=3,729). Third, neuronal population labels were projected from the snRNA-seq dataset, and populations with a prediction score below 0.5 were removed (n=456). These filtering steps reduced our snATAC-seq dataset to 11,651 neuronal and 10,894 non-neuronal cells for downstream analysis (total=22,545). The mean and median number of accessible regions recovered per cell were 7,028 and 4,401, respectively.

Motif analysis

ChromVar⁶⁶ was used to compute Z-scores for motif enrichment (one Z-score per motif per cell). For each cell population, a logistic regression model was constructed with Z-score as independent variable and cell population as dependent variable (df=2) and compared to the null model (df=1) using a likelihood ratio test. Enriched motifs, reported in Supplementary Data 5, were identified by averaging ChromVar Z-scores for a given motif across cells in a given population, retaining only motifs with an average Z-score >1, and then adjusting P-values for multiple testing using the Bonferroni method (adjusting for the number of cell populations x number of motifs).

The following steps were performed to correlate motif enrichment with transcription factor gene expression levels. First, for each cell population, we ranked all transcription factors (up to 334 transcription factors were expressed in our snRNA-seq data) based on their cell population gene expression specificity (mean rank of the four expression specificity metrics used in CELLEX). Second, for each transcription factor and motif pair, the transcription factor rank was correlated (Spearman) with the cell population motif enrichment Z-score across all cell populations. A meta P-value of the correlation was calculated using Fisher’s method.

Pseudo-bulk differential chromatin accessibility analysis

A pseudo-bulk chromatin accessibility matrix was generated by summing the peak counts for all cells with the same cell population and sample combination. ChromVar⁶⁶ was applied on the pseudo-bulk chromatin accessibility matrix to compute Z-scores for motif enrichment (one Z-score per motif per cell population per sample). To assess differential chromatin accessibility within a cell population, a logistic regression model was constructed with Z-score as independent variable and treatment as dependent variable (df=2) and compared to the null model (df=1) using a likelihood ratio test. P-values were adjusted for multiple testing using the Bonferroni method (n-motifs).

Genetic enrichment analysis

CELLECT analysis

Genetic enrichment analysis was carried out by running CELLECT⁷ version 1.0.0 (retrieved from https://github.com/perslab/CELLECT) on BMI GWAS summary statistics from the UK Biobank comprising > 457,000 individuals^67,68 and the DVC cell population gene expression or chromatin accessibility profiles. For the gene expression atlas, CELLECT was run on the CELLEX matrix using default parameters (window size=100 kb). For the chromatin accessibility atlas, the genetic enrichment analysis was carried out using the following steps. Peak coordinates were mapped from the mouse genome build mm10 to the human genome build hg19 using liftOver⁵⁷ with the minMatch parameter set to 0.1 (recommend parameter value for cross-species mapping). Eighty-three percent of the peaks could be mapped to the human genome. For each cell population, all peaks containing ≥1 enriched motif (Z>1 and Bonferroni-adjusted P<0.05) were ranked based on their maximum Z-score and rank-normalized to values between ]0;1]. Remaining peaks were assigned value 0. The resulting cell population-by-peak matrix was used as input for CELLECT using a window size of 1 kb as used in reference⁶⁹. For both genetic enrichment analyses P-values were corrected for multiple testing using the Bonferroni method (n-cell populations).

Meta-analysis

One-tailed meta-analysis, reported in Supplementary Data 6, was performed with the metafor R package using the rma.uni function using the fixed effects method with CELLECT coefficients as input values and CELLECT precision as weights. The meta-analysis P-values were corrected for multiple testing using the Bonferroni method (adjusting for the number of cell populations).

Immunohistochemistry and RNA in situ hybridization

All procedures were conducted in accordance with national regulations in Denmark, which are fully compliant with internationally accepted principles for the care and use of laboratory animals, and with animal experimental licenses granted by the Danish Ministry of Justice. Six eight weeks old C57BL/6 male mice were dosed IV with Ex4 (200 μg/kg; n=2) or vehicle (n=4) for 2 hours, then sacrificed and perfusion-fixed with formalin, and the brains were removed and embedded in paraffin after post-fixing in formalin overnight. Immersion-fixed brain samples collected at necropsy from two cynomolgus monkeys (one male, seven years old; one female, five years old) were purchased from Charles River and were embedded in paraffin upon receipt. Samples were cut (4 μm), and sections with AP areas represented (2–5 sections for each animal) were used for automated RNAscope (ACD, Bio-Techne) duplex ISH/ISH with chromogenic substrates or ISH/IHC fluorescence protocols on the Ventana Discovery Ultra platform using RNAscope^® VS automated workflows⁷⁰, using the following probes and antibody. Mouse probes (ACDBio, Bio-Techne): Ramp3 (497139), Glp1r (418859), Casr (423459), Gfral (439149), Olfr78 (436609), Grpr (317879), Asb4 (435099), Gipr (455789), Ccbe1 (485659), Bdnf-C2 (316039-C2), Mc4r-C2 (319189-C2), Glp1r-C2 (418859-C2), Lepr-C2 (471179-C2), Gcg-C2 (400609-C2), Gipr-C2 (319129-C2), Pax5-C2 (319189-C2), Calcr-C2 (494079-C2). Primate probes (ACDBio, Bio-Techne): Mfa-Ramp3-C2 (486679-C2), Mmu-Glp1r-C2 (449299-C2), Mfa-Casr (481659), Hs-Gfral (483049), Hs-Calcr (483049). Negative control probes (ACDBio, Bio-Techne): Dabp (312039), Dapb-C2 (312039-C2). Rodent GLP-1R (monoclonal rabbit) antibody: ab218532, Abcam (1:200). Fluorescent slides were scanned on an Olympus VS120 scanner, and slides developed using chromogens were scanned on a Hamamatsu Nanozoomer XR scanner.

Cre-transgenic mice studies

Mice

Mice were bred in the colony in the Unit for Laboratory Animal Medicine at the University of Michigan (USA); these mice and procedures were approved by the University of Michigan Committee on the Use and Care of Animals and in accordance with Association for the Assessment and Approval of Laboratory Animal Care and National Institutes of Health guidelines. Eight-12 weeks old C57BL/6 male mice were purchased from the Jackson Laboratories (Bar Harbor, USA). Mice were provided with standard chow diet (Purina Lab Diet 5001; except as noted below) and water ad libitum in temperature-controlled rooms on a 12-hour light-dark cycle. Calcr-Cre mice⁷¹ were propagated by intercrossing homozygous mice of the same genotype and used for the DREADD-experiments. Glp1r-Cre or Calcr-Cre line crossed with Cre-inducible GFP reporter mice were used for immunohistochemistry.

Viral Reagents and Stereotaxic Injections.

AAV8-hSyn-DIO-hM3Dq-mCherry was prepared by the University of North Carolina Vector Core (Chapel Hill, USA). For injection, following the induction of isoflurane anesthesia and placement in a stereotaxic frame, the skulls of adult mice were exposed. For NTS injection, the obex was set as the reference point for injection. After the reference was determined, a guide cannula with a pipette injector was lowered into the approximate NTS coordinates, which was A/P, −0.2; M/L, ±0.2; D/V, −0.2 from the obex, and 100 nl of virus was injected with using a picospritzer at a rate of 5–30 nl/min with pulses. Five minutes following injection, to allow for adequate dispersal and absorption of the virus, the injector was removed from the animal; the incision site was closed and glued. The mice received prophylactic analgesics before and after surgery. To assess the effect of CNO on short-term chow food intake, food intake was monitored over four hours in the dark cycle in wildtype mice injected with hM3Dq-mCherry in the NTS after administration with saline (n=6) or CNO (n=6). To assess the effect of CNO on long-term chow food intake, food intake was monitored over 24 hours in wildtype mice injected with hM3Dq-mCherry in the NTS after administration with saline (n=6) or CNO (n=6).

Phenotypic assessment

DREADD-expressing mice and their controls that were either at least two months old or two months post-surgery were treated with CNO (4936, Tocris, IP, 1 mg/kg) at the onset of dark cycle. For long-term standard chow or HFD (Research Diets D12492, 60% from fat) food intake and body weight assessment, DREADD-expressing mice (chow, n=7; HFD, n=7) and their controls (chow, n=6; HFD, n=5) were given saline for two to three days prior to injecting CNO twice per day (approximately 5:30 PM and 8:30 AM) for two days, followed by saline injections for another one to two days to assess recovery from the treatment. For short-term HFD food intake assessment, food intake was monitored over four hours in the dark cycle in DREADD-expressing mice after administration of saline (n=7) or CNO (n=7)

Immunohistochemistry

Mice were anesthetized in isoflurane and perfused with PBS followed by 10% buffered formalin. Brains were removed, placed in 10% buffered formalin overnight, and dehydrated in 30% sucrose for one week. With use of a freezing microtome (Leica, Buffalo Grove, USA), brains were cut into 30-μm sections. Sections were treated sequentially with 1% hydrogen peroxide/0.5% sodium hydroxide, 0.3% glycine, 0.03% sodium dodecyl sulfate, and blocking solution (PBS with 0.1% triton, 3% normal donkey serum). Three mice were used for the assessment of TH and DDC expression. For Ex4 FOS studies, mice were treated with saline (n=3, IP) or Ex4 (n=3, 6355, Tocris, IP, 150 μg/kg) two hours prior to perfusion. For sCT FOS studies, mice were treated with sCT (n=3, 4033011, Bachem, IP, 150 μg/kg) two hours prior to perfusion. The perfused sections were incubated overnight at room temperature in rabbit anti-FOS (FOS, #2250, Cell Signaling Technology, 1:1000) and exposed the next day with fluorescent secondary antibody (Molecular Probes, 1:200) to visualize proteins. Immunofluorescent staining was performed using primary antibodies (GFP, GFP1020, Aves Laboratories, 1:1000; dsRed, 632496, Takara, 1:1000; TH, NB300–109, Novus Biologicals, 1:1000; DDC, 101661-AP, Proteintech, 1:1000). Antibodies were reacted with species-specific Alexa Fluor-488, -568 or -647 conjugated secondary antibodies (Invitrogen, Thermo Fisher, 1:200). Images were collected on an Olympus (Center Valley, PA) BX53F microscope. Images were pseudo-colored using Photoshop software (Adobe) or Image J (NIH).

In vivo study data statistical analyses

Statistical analyses on mouse in vivo study data was performed by constructing a linear mixed effects model with food intake or body weight as dependent variable, genotype and time point as independent variables with interaction effects, and mouse as random effect. Group contrasts were tested using least-squares means two-tailed t-test, and P-values were adjusted for multiple testing using the Bonferroni method.

Cre-transgenic rat studies

Rats

Rats were bred in the Unit for Laboratory Animal Medicine at the University of Michigan; these rats and the procedures performed were approved by the University of Michigan Committee on the Use and Care of Animals and in accordance with Association for the Assessment and Approval of Laboratory Animal Care and National Institutes of Health guidelines. Wildtype Sprague-Dawley rats were obtained from Charles River Laboratories.

Calcr-Cre rats were produced by CRISPR/Cas9-mediated gene editing in collaboration with the Molecular Genetics Core of the Michigan Diabetes Research Center (MDRC, Ann Arbor, MI, USA). Briefly, we designed standard gRNAs homologous to opposite-stranded sequences beginning 24 and 73 bp downstream of the end of the coding sequences (gRNA sequences: ctcagtggatcacaatgttg and tgggatcacttgaaacgca). Synthetic sgRNAs containing these sequences (Synthego) were co-injected into fertilized oocytes together with Cas9 protein and an editing template containing the 200 bp 5’and 3’-homology arms surrounding sequences for a self-cleaving P2A peptide plus coding sequences for a nuclear-localized Cre recombinase in place of the CALCR STOP codon plus the subsequent 93 bp of genomic sequences (Supplemental Fig. 3a). The embryos were implanted into pseudopregnant females, and all pups were genotyped for the presence of Cre. Cre-containing pups were subjected to long-range PCR to determine insertion of the desired sequences in the correct genomic locus (Supplementary Fig. 3b); these genomic fragments were also subjected to Sanger sequencing to ensure the lack of adventitious mutations. Eight out of 78 were positive founders.

Applying immunohistochemistry, we later confirmed that AP Cre reporter activity was specific to Calcr cells (Supplementary Fig. 3c). Calcr-Cre rats were bred to ROSA26^{em1(CAG-tdTomato)1} reporter rats (purchased from Sigma Advanced Genetic Engineering Labs (SAGE, St. Louis, MO, USA)), and offspring were genotyped and confirmed by DNA sequencing, as above. Rats were housed in a 12-hour light-dark cycle at 21 °C constant temperature. All rats were housed in cages with Bed-O’Cobs bedding with ad libitum access to standard chow and Lixit water, except during the experiment.

Genotyping and PCR analysis

Calcr-Cre rats were genotyped using the following primers for CalcRCre: 5’ Cre-tatcaactcgcgccctggaag 3’ and 5’ CalcR 3’ HA-tatttgggtctgcctggtgac 3’, expected size 748bp. ROSA26^{em1(CAG-tdTomato)1} reporter rats were genotyped for Td-Tomato using the following primers: CAG-3F-GCAACGTGCTGGTTATTGTG and Tdtomato-5R-TGATGACCTCCTCGCCCTTG, expected band size 550bp.

Viral Reagents and Stereotaxic Injections.

Prior to surgery, six-12 months old Calcr-Cre female rats (n=6; 250–400 g) were handled for 3 min for three days. Rats were anesthetized with isoflurane (induced in a chamber at 3% and maintained under anesthesia at 2–2.5% delivered via face mask). The oxygen flow was delivered at 0.8 l/min. Rats received buprenorphine hydrochloride (0.03 mg/kg; SC), Buprenex, and Carprofen (5 mg/kg; SC) injections prior to brain stereotaxic surgery (Model 942 Kopf with digital display console, Tujunga, CA). Rats were placed on the stereotaxic frame with their head inclined to about a 90° angle. To reach the AP, an incision was made on the skin and three muscles on a rostral-caudal axis direction. An incision was made from left-right (horizontally) to reveal the fourth ventricle; the area postrema and the obex are found caudal of the fourth ventricle. Rats received an AAV8-hSyn-DIO-hM3Dq-mCherry⁷² virus injection in the AP (Z=-0.3 mm) of a volume of 500 nl using a Hamilton syringe (5 ul syringe 800 series Hamilton syringe, 33-gauge small hub RN needle). Rats continued to receive Buprenex (twice a day; three days total) and Carprofen (once a day; three days total). They received post-operative care for seven-10 days after their sutures were removed. Twenty-one days were allowed before experimentation for the virus to be expressed.

Phenotypic assessment

Rats were moved to an experimental room and placed in a new cage set-up (Pure-O’Cel bedding and water bottles). They received three days of handling and habituation in their new room and cage. They received CNO (4936, Tocris, 1 mg/kg; IP) or 0.9% Sodium Chloride Injection (1 mg/kg; IP) 30 min prior to the onset of the dark cycle. Additionally, food measures were collected at time-point 0 (onset of dark-cycle), 2, 4, 6, 16, 24 hours. After a week, conditions were counterbalanced amongst rats. Animals that received CNO on the first trial received saline on the second trial and vice-versa. Following these experiments, rats were perfused under anesthesia and processed for the detection of mCherry (as below). Animals lacking AP mCherry-IR or with mCherry-IR outside of the AP were excluded from further analysis (two animals were excluded, leaving four animals in the final analysis). To assess the effect of CNO on food intake, three-four months old Calcr-Cre female rats (n=10) received CNO (4936, Tocris, 1 mg/kg; IP) or 0.9% Sodium Chloride Injection (1 mg/kg; IP) 30 min prior to the onset of the dark cycle, and food measures were collected as above. Likewise, after a week, conditions were counterbalanced amongst rats, and animals that received CNO on the first trial received saline on the second trial and vice-versa.

Immunohistochemistry

Rats were placed under CO₂ and then transcardially perfused with phosphate-buffer saline (PBS) and 10% buffered formalin. Brains were extracted and placed in 10% buffered formalin overnight and cryoprotected in 30% sucrose. The brains were coronally sectioned into 30-μm sections using a microtome (Leica, Buffalo Grove, IL). Sections were washed in PBS and placed in PBS with 0.1% triton, 3% normal donkey serum blocking solution. Tissue was incubated overnight at room temperature in 1:1000 dsRed (Rabbit, 632392, Takara Bio Clontech). The tissue was incubated in Alexa Fluor-568 secondary antibody.

Rat in vivo study statistical analyses

Statistical analyses were carried out as described above for the mouse in vivo studies, using food intake as the dependent variable, treatment (CNO or saline) and time point as independent variables with interaction effects, and sample as random effect. Group contrasts were tested using least-squares means, and P-values were adjusted for multiple testing using the Bonferroni method.

Extended Data

Extended Data Fig. 1. Body weight and food intake for the single-nucleus RNA- and ATAC-seq in vivo study.

a Daily body weight and b food intake in semaglutide-administered (n=9), ad libitum fed vehicle-administered (n=9) and weight-matched control mice (n=8). Mean ± SEM are shown. *P<0.05, **P<0.01, ***P<0.001 vs. vehicle and ^#P<0.05, ^##P<0.01, ^###P<0.001 semaglutide vs. weight-matched, linear mixed effects model, Bonferroni-adjusted least-squares means two-tailed t-test.

Extended Data Fig. 2. Expression of tanycyte-like cell marker genes.

a Expression of marker genes for tanycyte-like cells. b Wt1 (n=1 mouse), c Wif1 (n=2 mice), d Slc22a3 (n=1 mouse), and e Cdon (n=1 mouse) in situ hybridization of sagittal brain sections (Allen Mouse Brain Atlas²⁷). Scale bar for panel b is representative for panels c-e, 100 μm. AP, area postrema; NTS, nucleus of the solitary tract; DMV, dorsal motor nucleus of the vagus nerve; OPCs, oligodendrocyte precursor cells; VLMCs, vascular leptomeningeal cells. Data to reconstruct panel a can be found in Supplementary Data 2.

Extended Data Fig. 3. Expression of DVC neuronal marker genes.

a Expression of marker genes for different neuronal populations. From top to bottom, dendrogram illustrating the similarity of the neuronal populations computed based on their gene expression levels, heatmap depicting the gene expression specificity values (ES_μ) of the neuronal marker genes, the most likely DVC origin of the neuronal populations. b-z In situ hybridization of coronal brain sections (Allen Mouse Brain Atlas²⁷). N=1 mouse for all hybridizations except panels m and o (n=2 mice) and panels p (n=3 mice). Scale bar for panel b is representative for panels c-z, 100 μm. DVC, dorsal vagal complex; AP, area postrema; NTS, nucleus of the solitary tract; DMV, dorsal motor nucleus of the vagus nerve. Data to reconstruct panel a can be found in Supplementary Data 2.

Extended Data Fig. 4. Expression of appetite-supressing receptors in mice and non-human primates.

a Representative image showing db-ISH of Casr (blue) and Glp1r (red) in non-human primates (n=2). Scale bar, 250 μm. b Representative image showing db-ISH of Gfral (blue) and Glp1r (red) in non-human primates (n=2). Scale bar, 100 μm. c Representative image showing db-ISH of Calcr (blue) and Glp1r (red) in non-human primates (n=2). Scale bar, 250 μm. d Representative image showing db-ISH of Grpr (blue) and Calcr (red) in mice (n=4). Scale bar, 100 μm. e Representative image showing db-ISH of Casr (blue) and Bdnf (red) in mice (n=4). Scale bar, 100 μm. f Representative image showing db-ISH of Grpr (blue) and Bdnf (red) in mice (n=4). Scale bar, 100 μm. g Representative image showing db-ISH of Casr (blue) and Mc4r (red) in mice (n=4). Scale bar, 100 μm. h Representative image showing db-ISH of Grpr (blue) and Mc4r (red) in mice (n=4). Scale bar, 100 μm. i-l Representative images showing TH (purple) or DDC immunoreactivity (purple) and GFP immunoreactivity (green) in Glp1r-Cre;GFP or Calcr-Cre;GFP mice (n=3). Scale bar, 150 μm. db-ISH; double in situ hybridization; AP, area postrema; NTS, nucleus of the solitary tract; GFP, green fluorescent protein.

Extended Data Fig. 5. Expression of appetite-supressing receptors and peptides in mice and non-human primates.

a Representative image showing db-ISH of Calcr (blue) and Ramp3 (red) in non-human primates (n=2). Scale bar, 100 μm. b Representative image showing db-ISH of Ccbe1 (blue) and Gipr (red) in mice (n=4). Scale bar, 100 μm. c Representative image showing db-ISH of Pax5 (blue) and Gipr (red) in mice (n=4). Scale bar, 100 μm. d Representative image showing db-ISH of Gipr (blue) and Glp1r (red) in mice (n=4). Scale bar, 100 μm. e Representative image showing db-ISH of Asb4 (blue) and Gcg (red) in mice (n=4). Scale bar, 100 μm. db-ISH; double in situ hybridization; AP, area postrema; NTS, nucleus of the solitary tract.

Extended Data Fig. 6. Additional DVC neuronal populations with previously-defined functions.

Expression of genes defining DVC neurons previously implicated in metabolic control. Top, expression specificity (ES_μ) of selected genes. Bottom, the most likely DVC origin of the neuronal populations. Mc4r cholinergic (Chat3) DMV neurons regulate circulating insulin³⁹, Hsd11b2 and Nr3c2 glutamatergic (Glu2) NTS neurons drive sodium appetite⁴⁰, and Lepr and Gal glutamatergic (Glu3) NTS neurons module breathing⁴¹. DVC, dorsal vagal complex; AP, area postrema; NTS, nucleus of the solitary tract; DMV, dorsal motor nucleus of the vagus nerve. Data to reconstruct figure can be found in Supplementary Data 2.

Extended Data Fig. 7. GLP-1RA-downregulated modules.

a Most enriched Gene Ontology terms for modules M18–20. Bonferroni-adjusted g:Profiler P-value. b Top 10 genes for modules M18–20. Data to reconstruct figure can be found in Supplementary Data 1, Supplementary Data 4, and through the NCBI Gene Expression Omnibus.

Extended Data Fig. 8. Expression of Bdnf and Mc4r following GLP-1RA administration.

a Representative image showing db-ISH of Casr (blue) and Mc4r (red) in mice administered with Exendin-4 (IV, 200 μg/kg; n=2). Scale bar, 100 μm. b Representative image showing db-ISH of Casr (blue) and Bdnf (red) in mice administered with Exendin-4 (IV, 200 μg/kg; n=2). Scale bar, 100 μm. db-ISH; double in situ hybridization; AP, area postrema; NTS, nucleus of the solitary tract; IV, intravenous.

Extended Data Fig. 9. FOS immunoreactivity in Calcr AP and NTS cells following GLP-1RA or salmon calcitonin administration.

a Representative image showing FOS immunoreactivity (purple) and GFP immunoreactivity (green) in Calcr-Cre;GFP mice administered with Exendin-4 (IP, 150 μg/kg; n=3) or vehicle (n=3). Scale bar, 150 μm. b Representative image showing FOS immunoreactivity (purple) and GFP immunoreactivity (green) in Calcr-Cre;GFP mice administered with sCT (IP, 150 μg/kg; n=3) or vehicle (n=3). Scale bar, 150 μm. AP, area postrema; NTS, nucleus of the solitary tract; GFP, green fluorescent protein; GLP-1RA, GLP-1 receptor agonist; sCT, salmon calcitonin; IP, intraperitoneal.

Extended Data Fig. 10. Activation of Calcr NTS neurons suppresses feeding.

a Representative image showing mCherry immunoreactivity (pseudo-colored green) and FOS immunoreactivity (purple) after CNO treatment in Calcr-Cre mice injected with hM3Dq-mCherry in the NTS (n=7). Scale bar, 150 μm. b Long-term chow food intake and c body weight in control (n=6) and Calcr-Cre mice injected with hM3Dq-mCherry in the NTS (n=7) measured during 1 day of saline, 2 days of CNO (IP, 1 mg/kg) followed by 1 day of saline treatment. d Short-term HFD food intake in Calcr-Cre mice injected with hM3Dq-mCherry in the NTS and treated with saline (n=7) or CNO (n=7; IP, 1 mg/kg). e Long-term HFD food intake and f body weight in control (n=5) or Calcr-Cre mice injected with hM3Dq-mCherry in the NTS (n=7) measured during 3 days of saline, 2 days of CNO (IP, 1 mg/kg) followed by 2 days of saline treatment. g Short-term chow food intake in control mice injected with hM3Dq-mCherry in the NTS and treated with saline (n=6) or CNO (n=6; IP, 1 mg/kg). Mean ± SEM are shown. P<0.05 are specified, linear mixed effects model, Bonferroni-adjusted least squares means two-tailed t-test. h Long-term chow food intake in control mice injected with hM3Dq-mCherry in the NTS and treated with saline (n=6) or CNO (n=6; IP, 1 mg/kg. Mean ± SEM are shown. Linear model, least squares means two-tailed t-test. AP, area postrema; NTS, nucleus of the solitary tract; HFD, high-fat diet; CNO, Clozapine-N-oxide; IP, intraperitoneal.

Supplementary Material

supplementary data 1

Supplementary Data 1 Differentially expressed genes in bulk RNA-seq data.

supplementary data 2

Supplementary Data 2 CELLEX scores for cell populations identified in the single-nucleus RNA-seq data.

supplementary data 3

Supplementary Data 3 AP and NTS marker genes.

supplementary data 4

Supplementary Data 4 Gene modules and their association with treatment.

supplementary data 5

Supplementary Data 5 Motif enrichment for cell populations identified in the single-nucleus ATAC-seq data.

supplementary data 6

Supplementary Data 6 Meta-analysis of CELLECT results.

Data availability

All genetic data generated in this study (bulk RNA-seq, snRNA-seq and snATAC-seq) are available in GEO under SuperSeries accession number GSE166649. All other data are available from the authors upon reasonable request.