Cross-species analysis defines the conservation of anatomically segregated VMH neuron populations

The ventromedial hypothalamic nucleus (VMH) controls diverse behaviors and physiologic functions, suggesting the existence of multiple VMH neural subtypes with distinct functions. Combing translating ribosome affinity purification with RNA-sequencing (TRAP-seq) data with single-nucleus RNA-sequencing (snRNA-seq) data, we identified 24 mouse VMH neuron clusters. Further analysis, including snRNA-seq data from macaque tissue, defined a more tractable VMH parceling scheme consisting of six major genetically and anatomically differentiated VMH neuron classes with good cross-species conservation. In addition to two major ventrolateral classes, we identified three distinct classes of dorsomedial VMH neurons. Consistent with previously suggested unique roles for leptin receptor (Lepr)-expressing VMH neurons, Lepr expression marked a single dorsomedial class. We also identified a class of glutamatergic VMH neurons that resides in the tuberal region, anterolateral to the neuroanatomical core of the VMH. This atlas of conserved VMH neuron populations provides an unbiased starting point for the analysis of VMH circuitry and function.


Introduction
The ventromedial hypothalamic nucleus (VMH, which primarily contains glutamatergic neurons) plays important roles in a variety of metabolic responses and in the control of behaviors relevant to panic, reproduction, and aggression. The VMH contains several anatomic subdivisions, including the dorsomedial and central VMH (VMH DM and VMH C , respectively, which control autonomic outputs and behavioral responses to emergencies; Lindberg et al., 2013;Vander Tuig et al., 1982), and the ventrolateral VMH (VMH VL ; known for roles in sexual and social behaviors; Hashikawa et al., 2017;Krause and Ingraham, 2017). The predominantly GABAergic tuberal region of the hypothalamus lies anterolateral to the core of the VMH.
Each VMH subdivision mediates a variety of outputs and thus presumably contains multiple functionally distinct cell types. For example, activating adult Nr5a1-expressing VMH neurons (which includes most cells in the VMH DM and VMH C ) promotes panic-related behaviors, augments hepatic glucose output to increase blood glucose, and elevates energy expenditure (Meek et al., 2016, 201;Flak et al., 2020;Kunwar et al., 2015). In contrast, activating the subset of VMH DM cells that expresses leptin receptor (Lepr, which encodes the receptor for the adipose-derived, energy balance-controlling hormone, leptin; Chen et al., 1996;Tartaglia et al., 1995) promotes energy expenditure without altering these other parameters (Sabatini et al., 2021;Meek et al., 2013;Meek et al., 2016). Hence, each VMH subregion may contain multiple discrete neuron populations that mediate unique functions.
To date, most analyses of VMH function have utilized Nr5a1 or candidate markers that do not necessarily align with functionally and/or transcriptionally unique VMH cell types (Bingham et al., 2006). Thus, to understand VMH-controlled responses, we must use unbiased methods to define discrete subpopulations of VMH neurons, along with markers that permit their selective manipulation. Single-cell approaches (such as single-nucleus RNA-sequencing [snRNA-seq]) can identify neuronal populations in an unbiased manner and have previously suggested parceling schemes for neurons in many brain areas, including the VMH (Kim et al., 2019;Campbell et al., 2017;Habib et al., 2017). Many such analyses define large numbers of highly interrelated cell populations of unclear functional significance and conservation, however (Kim et al., 2019;Chen et al., 2017;Lam et al., 2017). Determining functions for dozens of cell populations that lie in the same anatomic region and which possess overlapping gene expression profiles (i.e., that do not contain unique marker genes) would represent a daunting task.       In the present study we use translating ribosome affinity purification with RNA-sequencing (TRAPseq) in mice together with snRNA-seq of mouse and macaque VMH neurons to define transcriptionally unique, anatomically discrete, conserved, and genetically tractable classes of VMH neurons. These include a distinct Lepr-expressing VMH neuron class, along with a set of glutamatergic VMH neurons that resides in the tuberal region. These findings define a starting point for the comprehensive analysis of VMH circuitry and function.

Results
Combining snRNA-seq with Nr5a1-directed TRAP-seq defines mouse VMH neuron populations To define neuronal populations within the mouse VMH in an unbiased manner, we microdissected the VMH of mice and subjected 10 individual tissue samples to snRNA-seq using the 10Â Genomics platform ( Figure 1A), collecting a total of 42,040 nuclei that passed quality control ( Many adult VMH neurons express Nr5a1 (which encodes the transcription factor, SF1; Cheung et al., 2013) and/or Fezf1 (Kurrasch et al., 2007), whose detection was restricted to a confined cluster of neurons in UMAP space ( Figure 1D). Although essentially all VMH neurons express Nr5a1 during development, only a subset of VMH cells express Nr5a1 and/or Fezf1 in adult animals (Cheung et al., 2013;Kurrasch et al., 2007). Furthermore, the inherent noise in snRNA-seq data risks false positives and negatives when using only one or two genes for cell-type identification. To ensure that we identified all VMH cell groups for our analysis, we performed TRAP-seq using Nr5a1-Cre;Rosa26 eGFP-L10a mice, in which the early developmental expression of Nr5a1-Cre promotes the permanent expression of tagged ribosomes across the VMH ( Figure 1E). TRAP-seq identified 4492 transcripts significantly enriched in cells marked by the developmental expression of Nr5a1, including Nr5a1 and Fezf1 ( Figure 1F, Supplementary file 1). Applying this broader VMH-enriched transcriptome to our snRNA-seq clusters revealed six populations of neurons (clusters 6-11; corresponding to the populations with highest Nr5a1 and Fezf1 expression) that contain VMH neurons ( Figure 1G,H). Importantly, Nr5a1-negative cells (clusters 6 and 7) were identified as VMH by their enrichment of Nr5a1-Cre TRAP-seq genes ( Figure 1H), suggesting they were developmentally labeled by Nr5a1-Cre; these cells presumably reside in the VMH VL (which expresses Nr5a1 during development, but not in adulthood).
To compare TRAP-seq and snRNA-seq results, we performed 'pseudo-TRAP' on pseudobulk samples of our snRNA-seq data, aggregated by cell type (i.e., VMH clusters vs. non-VMH clusters; see Supplementary file 2 for enrichment results). While many genes were enriched in both datasets (1977), more were specific to one method (Figures 1I, 2354 genes specific to TRAP and 2828 specific to pseudo-TRAP). Notably, enrichment was largely a function of expression level: the genes that were enriched in both datasets were highly expressed in both, while the genes that were specific to TRAP or pseudo-TRAP were highly expressed in their enriched dataset, but expressed at lower levels in the other ( Figure 1J). This suggests that the two methods may illuminate partially distinct aspects of the transcriptome-TRAP-seq for ribosome-associated genes and snRNA-seq for nuclear-enriched and nascent transcripts-and that their combined use provides a more comprehensive view of cellular state. Finally, many of the genes enriched in both datasets were limited to one or a few populations, highlighting the heterogeneity of gene expression across VMH cell types, even for prominent VMH marker genes ( Figure 1K).
To understand the landscape of mouse VMH neuron populations in more detail, we subjected the VMH neurons in clusters 6-11 ( Figure 1H) to further clustering, which identified 24 transcriptionally defined neuronal populations (Figure 2A (Kim et al., 2019) analysis represented a notable difference between our analyses, however. In our reanalysis of their data, we found that Nfib-marked cells were present in their samples but were filtered out before the final VMH clustering (Figure 2-figure supplement 1). A previous scRNA-seq study of the neighboring ARC also mapped a neuron population marked by Nfib expression to the VMH, however (Campbell et al., 2017), and these cells correlated to our Nfib-marked VMH neuron populations ( A simplified parceling scheme defines anatomically distinct VMH neuron populations Hierarchical clustering and marker gene analysis for our 24 mouse VMH neuron clusters using CEL-LEX (Timshel et al., 2020) revealed that many cell groups were highly related to other VMH neuron clusters ( Figure 2D,E; Supplementary file 3). Furthermore, many of these populations share marker gene expression to an extent that renders it impossible to specifically manipulate single populations given current approaches that use a single gene for cell-type manipulation (e.g., Cre-based mouse models) ( Figure 2D,E). To identify classes of genetically distinguishable cells, we cut the hierarchical tree at different levels and measured the maximum pairwise expression correlation to highlight the level at which few pairs of clusters exhibited highly correlated transcriptomes. We found that six classes represented the largest number of classes that retained minimal correlated expression ( Figure 2F). To avoid confusion, we refer to these as VMH neuron classes, while referring to the cell groups of which each class is composed as clusters or subpopulations. The mean silhouette width (a measure of clustering robustness) for the various tree cuts also supported the use of six classes (Figure 2-figure supplement 3), and these six classes corresponded to the cluster designations from the broader neuron dataset ( Figure 2G). Hence, a parceling scheme for mouse VMH neurons that contains six classes, each composed of highly similar subpopulations, captures the transcriptional patterns of the VMH. Importantly, this approach identified numerous specific marker genes (e.g., Dlk1, Esr1, Nfib, Foxp2, Fezf1, and Lepr) for each class of mouse VMH neurons ( Figure 2D,H, Supplementary file 4), which should facilitate their manipulation and study. Consistent with the distinct nature of these six VMH neuron classes and the utility of this parceling scheme, each class demonstrated a circumscribed anatomic distribution. As previously reported Persson-Augner et al., 2014), Dlk1-and Esr1-expressing neurons (VMH Dlk1 and VMH Esr1 , respectively) map to the VMH VL . While Dlk1 is unique to a single class of VMH neurons, it is also expressed in neighboring hypothalamic neurons (Figure 2-figure supplement 4), complicating its utility for manipulating this population without using an intersectional approach. In contrast to  these ventrolateral populations, Lepr-expressing cells lie within the core of the VMH DM (Elmquist et al., 1998) and the expression of marker genes for the Nfib-marked clusters (VMH Nfib ) resides in the most dorsomedial compartment of the VMH DM (Figure 2-figure supplement 5). Fezf1-marked populations (VMH Fezf1 ) include cells with similarity to both ventrolateral VMH Dlk1 and dorsomedial VMH Lepr neurons, and many VMH Fezf1 neurons lie in VMH C , in the transition between the dorsomedial and ventrolateral zones of the VMH. Markers for the Foxp2-expressing populations (VMH Foxp2 ) reside anterior and lateral to the core of the VMH, in the so-called tuberal region (Figure 2-figure supplement 6). Thus, the major VMH classes identified by our analysis each map to specific and distinct anatomic locations, consistent with their unique genetic signatures.

Lepr-directed TRAP-seq analysis of VMH Lepr cells
Given that Lepr-expressing VMH neurons mediate only a subset of VMH DM functions, we were intrigued by the finding that unbiased snRNA-seq identified a distinct class of VMH DM cells marked by Lepr (VMH Lepr neurons), consistent with a specialized role for Lepr-expressing VMH cells (Minokoshi et al., 1999;Toda et al., 2013;Noble et al., 2014;Gavini et al., 2016). The finding that Lepr-expressing VMH neurons map onto a single VMH neuron class also suggests a uniformity of function for these cells, which contrasts with the situation in other brain regions (such as the neighboring ARC, where multiple cell types with opposing functions [e.g., Agrp and Pomc cells] express Lepr; Campbell et al., 2017). We utilized TRAP-seq to assess the extent to which Leprexpressing VMH neurons correspond to the VMH Lepr clusters and to compare the genetic program of Lepr-expressing VMH neurons with gene expression in snRNA-seq-defined VMH neuron classes and subpopulations.
We microdissected the VMH of Lepr Slc17a6 -L10a mice and performed TRAP-seq, identifying 3580 transcripts that were enriched in the TRAP material relative to the supernatant ( Figure 3C, Supplementary file 5). Importantly, Lepr itself was not enriched, suggesting that we successfully purified ribosome-associated mRNA from Lepr-expressing VMH cells away from that derived from other Lepr-expressing populations ( Figure 3B). We found that most non-VMH genes enriched in our conventional (Lepr Cre -only) TRAP-Seq (including Agrp, Nts, and Ghrh) were not enriched in this analysis ( Figure 3-figure supplement 1). Pomc and Prlh remained somewhat enriched, however, suggesting that Lepr-expressing DMH Prlh cells are glutamatergic (Dodd et al., 2014), and consistent with the finding that some ARC Pomc cells express Slc17a6 (Jones et al., 2019).
As expected, we observed a high degree of concordance between snRNA-seq-defined gene expression in neurons of the VMH Lepr cluster with gene expression in Lepr-expressing VMH neurons by TRAP-seq. We identified 3576 genes enriched in VMH-centered Lepr Slc17a6 TRAP-seq material ( Figure 3C), 1174 of which were also enriched in pseudo-TRAP analysis of the neurons assigned to the VMH Lepr population (Supplementary file 6). Among the top enriched genes shared by these two methods were Gpr149, Rai14, and Tnfrsf8 ( Figure 3F), which exhibit a similar VMH DM -centered expression pattern as Lepr ( Figure 3G-I). As with the Nr5a1-Cre TRAP-seq, enrichment was largely a function of expression level: the genes that were enriched in both datasets were highly expressed in both, while the genes that were specific to TRAP or pseudo-TRAP were highly expressed in their enriched dataset, but more lowly expressed in the other. Notably, gene ontology (GO) analysis of the common genes revealed many terms related to synaptic function, while genes unique to TRAPseq were enriched for ribosomal and mitochondrial function (data not shown).
We mapped enriched genes from Lepr Slc17a6 TRAP-seq to the gene expression profiles of our snRNA-seq-defined VMH neuron populations ( Figure 3F), revealing the bias of Lepr Slc17a6 TRAP-seq gene expression toward VMH Lepr neurons and the exclusion of markers from other VMH Lepr cells from Lepr Slc17a6 TRAP-seq-enriched genes. Thus, this analysis demonstrates that Lepr-expressing VMH neurons map specifically to VMH Lepr cell clusters, suggesting that they represent a transcriptionally and functionally unique set of neurons.

Conservation of VMH neuronal populations across species
While many previous studies of VMH neuron function have suggested that this brain region contains neurons that could represent therapeutic targets to aid people with obesity, diabetes, and other diseases, most of these studies have been performed in mice (Hashikawa et al., 2017;Flak et al., 2020;Meek et al., 2013;Kim et al., 2012). We know little about the cross-species conservation of VMH cell populations, however. To assess the potential conservation of VMH neuron populations across species, we microdissected macaque (Macaca mulatta) VMH and performed snRNA-seq using the same techniques as for mouse VMH (Figure 4-figure supplement 1). A subset of macaque neurons expressed NR5A1 and/or FEZF1 and exhibited similar gene expression profiles to our mouse Nr5a1-Cre TRAP-seq enriched genes (Figure 4-figure supplement 2), suggesting that these cells represent the macaque VMH and indicating that the mouse and macaque VMH share similar global gene expression signatures.
Graph-based clustering of the macaque VMH neurons yielded seven populations with unique marker genes (with the partial exception of two related ESR1-expressing cell types) ( Figure 4B-E). Most of these macaque populations, including populations marked by LEPR, FOXP2, NFIB, and ESR1 (VMH LEPR , VMH FOXP2 , VMH NFIB , and VMH ESR1 , respectively), have presumptive orthologs in the mouse ( Figure 4D, see Supplementary file 7 for a complete list of markers). While populations marked by DLK1 and FEZF1 were absent from this analysis, the macaque VMH contained a population marked by QRFPR expression (VMH QRFPR ).
To determine whether the macaque populations were orthologous to those from the mouse, we first compared expression of orthologous genes among mouse and macaque cell clusters. We found that all macaque populations had clear correlates in the mouse, which mapped according to their marker genes (as expected) ( Figure 4F, Figure 5-figure supplement 1D). Notably, the macaque VMH QRFPR population correlated with clusters from both VMH Dlk1 and VMH Fezf1 classes in the mouse, suggesting that VMH QRFPR contains orthologs of the mouse VMH Dlk1 and VMH Fezf1 classes. Projecting the mouse or macaque cluster labels onto the other species using Seurat anchors (see Materials and methods for more details) yielded similar results ( Figure 5-figure supplement 1). All major classes from mouse and macaque projected with high confidence onto the equivalent major classes of the other species, and the macaque VMH QRFPR population was represented by both mouse VMH Dlk1 and VMH Fezf1 cells ( Figure 5-figure supplement 1A).
To generate an atlas of conserved mouse and macaque VMH populations, we integrated the mouse and macaque data and clustered the merged dataset using the Seurat canonical correlation analysis (CCA) framework ( Figure 5A-B, see Materials and methods for more details). This analysis revealed populations of VMH neurons that each contained mouse and macaque cells in roughly equal proportions ( Figure 5E). As predicted, neurons from the macaque VMH LEPR , VMH FOXP2 , VMH NFIB , and VMH ESR1 classes mapped directly with murine VMH Lepr , VMH Foxp2 , VMH Nfib , and VMH Esr1 classes, respectively ( Figure 5G). We examined the potential co-expression of conserved marker genes for VMH LEPR cells by in situ hybridization (ISH) for ACVR1C in the macaque hypothalamus ( Figure 5-figure supplement 2), confirming ACVR1C is co-expressed with LEPR in the macaque VMH DM . We also examined the potential co-expression of SLC17A8 (which marks VMH NFIB ) with LEPR. SLC17A8 identified a population of cells that lay at the medial extreme of the macaque VMH DM , corresponding to the most dorsomedial aspect of the rodent VMH DM , as for murine VMH Nfib cells; these cells did not colocalize with LEPR-expressing cells ( This analysis also identified macaque populations that mapped with the murine VMH Dlk1 and VMH Fezf1 populations; these derived mainly from the macaque VMH QRFPR population ( Figure 5G). Notably, FEZF1 and DLK1 were poorly enriched in the macaque VMH QRFPR population ( Figure 5D), while mouse Qrfpr expression was biased toward the VMH Esr1 populations ( Figure 5D). Despite differences in some marker genes, however, the orthologous macaque and mouse VMH neuron populations share dozens of other genes across species ( Figure 5H, Supplementary file 8, Figure 5figure supplement 1B), suggesting that the mouse and macaque cells in each group represent similar cell types. Gene ontology (GO) analysis of common marker genes (Source data 2, Source data 1) revealed that most of these mediate core neuronal functions, such as ion channel activity ( Figure 5-figure supplement 1C). Thus, while some marker genes vary across species, the mouse VMH populations have close orthologs in the macaque VMH based upon their gene expression profiles.

Discussion
We combined TRAP-and snRNA-seq analysis of the VMH to identify 24 mouse neuronal populations with complex interrelations. These 24 populations represented six distinct cell classes that demonstrated unique anatomic distribution patterns. The main VMH classes were highly conserved between the mouse and the macaque in terms of gene expression profiles and anatomic distribution within the VMH. This atlas of conserved VMH neuron populations provides an unbiased and tractable starting point for the analysis of VMH circuitry and function.
Having many populations of VMH neurons with highly related gene expression profiles complicates the functional analysis of VMH cell types suggesting the importance of simplifying the map of these heterogeneous populations to permit their manipulation and study. By using hierarchical clustering, we were able to identify six maximally unrelated classes of VMH neurons with distinct gene expression signatures. These six discrete transcriptionally defined VMH neuronal classes demonstrated distinct anatomic distribution patterns (three located in the VMH DM , two in the VMH VL , and one in the tuberal region), revealing a transcriptional basis for the previously suggested functional architecture of the VMH.
Our VMH cell populations were similar to those previously described by Kim et al., although they more finely split the VMH VL populations than did we (presumably because their dissection bias toward the VMH VL yielded more VMH VL neurons) (Kim et al., 2019). We were able to more clearly distinguish among VMH DM cell types in our VMH DM -focused analysis, however, including by dividing the VMH DM neurons that they identified into two major classes (VMH Lepr and VMH Fezf1 ), as well as identifying a third VMH DM population (VMH Nfib ) absent from their analysis. Interestingly, the VMH Nfib population was quite distinct from the more closely related VMH Lepr and VMH Fezf1 VMH DM cell types, suggesting a potentially divergent function for this most dorsomedial of the VMH DM populations.
Our mouse VMH cell types mapped clearly onto specific populations of macaque VMH neurons, revealing the utility of the mouse as a surrogate for the primate in terms of VMH cell types and, presumably, function. While the macaque single VMH QRFRP population mapped to two mouse classes (VMH Fezf1 and VMH Dlk1 ) by orthologous gene expression, these mouse classes are closely related to each other (transcriptionally and anatomically) and the macaque cells from VMH QRFRP segregate to distinct cell clusters defined by Dlk1 and Fezf1 in the mouse when we integrated and reclustered the mouse and macaque cell data. Also, while DLK1 is not specific to the macaque VMH QRFPR cells, the otherwise similar gene expression profiles of mouse and macaque cells that map to VMH Dlk1 or VMH Fezf1 populations suggest the conserved nature of these cell types across species.
While not all Lepr-expressing cell types track in the brain across species (e.g., preproglucagon [Gcg]-containing NTS neurons in the mouse express Lepr, rat NTS Gcg cells do not; Huo et al., 2008), mouse VMH Lepr neurons map directly with macaque VMH LEPR neurons by all of the measures that we examined. The finding that Lepr/LEPR expression marks a unique and conserved cell type in rodent and primate is consistent with the notion that this class of VMH neuron mediates a discrete component of VMH function, as suggested by previous work in the mouse demonstrating roles for Lepr-expressing VMH neurons in the control of energy balance, but not glucose production or panic-like behaviors (Meek et al., 2013;Meek et al., 2016;Sabatini et al., 2021).
While the tuberal region contains more GABAergic than glutamatergic neurons, this region projects to similar target areas as does the core VMH and contains substantial numbers of neurons marked by Nr5a1-Cre activity. Hence, although VMH Foxp2 cells lie in the tuberal region, their glutamatergic nature, their marking by Nr5a1-Cre activity, and the finding that they are transcriptional most similar to other VMH populations mark VMH Foxp2 cells as VMH neurons. While few data exist to suggest the physiologic roles played by these cells, it will be interesting to manipulate VMH Foxp2 neurons to determine their function.
The identification of distinct transcriptionally defined VMH cell populations provides the opportunity to develop new tools that can be used to understand the nature and function of VMH DM cell types and their roles in metabolic control. The finding that the major VMH cell classes found in the mouse are present in the macaque supports the use of the mouse to study the metabolic functions of the VMH as a means to identify potential therapeutic targets for human disease. It will be important to use these findings to dissect functions for subtypes of VMH cells, which may represent targets for the therapy of diseases including obesity and diabetes.

Materials and methods Animals
Mice were bred in the Unit for Laboratory Animal Medicine at the University of Michigan. These mice 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 (AAALAC) and National Institutes of Health (NIH) guidelines. Mice were provided with ad libitum access to food (Purina Lab Diet 5001) and water in temperature-controlled (25˚C) rooms on a 12 hr light-dark cycle with daily health status checks.
Nonhuman primate tissue was obtained from the Tissue Distribution Program at ONPRC. Animal care is in accordance with the recommendations described in the Guide for the Care and Use of Laboratory Animals of the NIH and animal facilities at the Oregon National Primate Research Center (ONPRC) are accredited by AAALAC.
Tissue prep, cDNA amplification, and library construction for 10Â snRNA-seq Mice were euthanized using isoflurane and decapitated, the brain was subsequently removed from the skull and sectioned into 1-mm-thick coronal slices using a brain matrix. The VMH was dissected out and flash frozen in liquid N 2 . Nuclei were isolated as previously described (Habib et al., 2017) with modifications as follows. On the day of the experiment, frozen VMH (from 2 to 3 mice) was homogenized in Lysis Buffer (EZ Prep Nuclei Kit, Sigma) with Protector RNAase Inhibitor (Sigma) and filtered through a 30 mm MACS strainer (Myltenti). Strained samples were centrifuged at 500 rcf Â 5 min and pelleted nuclei were resuspended in wash buffer (10 mM Tris Buffer pH 8.0, 5 mM KCl, 12.5 mM MgCl 2 , 1% BSA with RNAse inhibitor). Nuclei were strained again and recentrifuged at 500 rcf Â 5 min. Washed nuclei were resuspended in wash buffer with propidium iodide (Sigma) and stained nuclei underwent FACS sorting on a MoFlo Astrios Cell Sorter. Sorted nuclei were centrifuged at 100 rcf Â 6 min and resuspended in wash buffer to obtain a concentration of 750-1200 nuclei/mL. RT mix was added to target~10,000 nuclei recovered and loaded onto the 10Â Chromium Controller chip. The Chromium Single Cell 3' Library and Gel Bead Kit v3, Chromium Chip B Single Cell kit, and Chromium i7 Multiplex Kit were used for subsequent RT, cDNA amplification, and library preparation as instructed by the manufacturer. Libraries were sequenced on an Illumina HiSeq 4000 or NovaSeq 6000 (pair-ended with read lengths of 150 nt) to a depth of at least 50,000 reads/cell.

snRNA-seq data analysis
Count tables were generated from the FASTQ files using cellranger and analyzed in R 3.6.3 using the Seurat three framework. Genes expressed in at least four cells in each sample and were not gene models (starting with 'Gm') or located on the mitochondrial genome were retained. Cells with at least 500 detected genes were retained. Doublets were detected using Scrublet (Wolock et al., 2019). For each 10Â run, the expected number of doublets was predicted using a linear model given 10Â data of the detected doublet rate and the number of cells. Then, each cell was given a doublet score with Scrublet and the n cells (corresponding to the expected number of doublets) with the top scores were removed.
The data was then normalized using scran (Lun et al., 2016) and centered and scaled for each dataset independently and genes that were called variable by both Seurat FindVariableFeatures and sctransform (Hafemeister and Satija, 2019) were input to principal component analysis (PCA). The top principal components (PCs) were retained at the 'elbow' of the scree plot (normally 15-30, depending on the dataset) and then used for dimension reduction using UMAP and clustering using the Seurat FindNeighbors and FindClusters functions. Both were optimized for maximizing cluster consistency by clustering over a variety of conditions: first, varying the number of neighbors from 10 to the square root of the number of cells while holding the resolution parameter in FindClusters at one and finding the clustering that maximized the mean silhouette score; then, this number of neighbors was held fixed while varying the resolution parameter in FindClusters from 0.2 upward in steps of 0.2 until a maximal mean silhouette score was found. Clusters were then hierarchically ordered based on their Euclidean distance in PC space and ordered based on their position in the tree.
Marker genes were found using the Seurat function FindAllMarkers for each sample with resulting p-values combined using the logitp function from the metap package or using CELLEX 1.0.0. Cluster names were chosen based on genes found in this unbiased marker gene search and known marker genes.
From the all-cell data, cell types were predicted using gene set enrichment analysis from the marker genes and a manually curated set of genes known to mark specific CNS cell types. From this, clusters that were highly enriched for markers from two (or more) distinct cell types were labeled as 'doublets' and those with no enrichment were labeled as 'junk', the remaining clusters were labeled based on their lone CNS cell type.
To predict VMH neurons from all neurons, we first found genes significantly enriched in the bead fraction in Nr5a1-Cre TRAP-seq (see below for details) and expressed above one count per million. The scaled count matrix containing these genes was then used as input to PCA. The magnitude of the first PC (loading) was then used to generate a VMH similarity score and the clusters that had a high Nr5a1-Cre TRAP loading were glutamatergic (express Slc17a6 and not Gad1 or Slc32a1 above the mean value), and expressed either Nr5a1 or Fezf1 above the mean value were included as presumptive VMH.

TRAP-seq analysis
Mice (Lepr Cre ;ROSA26 EGFP-L10a or Lepr Cre ;Slc17a6 FlpO ;ROSA26 EGFP-L10a ) were euthanized and decapitated, the brain was subsequently removed from the skull and sectioned into 1-mm-thick coronal slices using a brain matrix. The VMH or hypothalamus was dissected and homogenized in lysis buffer. Between 15 and 20 mice were used for TRAP experiments. GFP-tagged ribosomes were immunoprecipitated and RNA isolated as previously described (Allison et al., 2018). RNA was subject to ribodepletion and the resultant mRNA was fragmented and copied into first strand cDNA. The products were purified and enriched by PCR to create the final cDNA library. Samples were sequenced on a 50-cycle single end run on a HiSeq 2500 (Illumina) according to manufacturer's protocols.
FASTQ files were filtered using fastq_quality_filter from fastx_toolkit to remove reads with a phred score <20. Then reads were mapped using STAR with a custom genome containing the Ensembl reference and sequences and annotation for Cre and EGFP:L10a (and Flpo in the RCFL:: eGFP-L10a dataset). Count tables were generated using the STAR -quantMode GeneCounts flag.
Count tables were analyzed in R 3.6.3 and were subject to quality control to ensure read adequate library size (20-30 million reads), enrichment of positive control genes (e.g., EGFP:L10a and/ or Cre, Nr5a1), and appropriate sample similarity in both hierarchical clustering of Euclidean distance and TSNE/UMAP space (e.g., bead samples are more similar to one another than to any sup sample). All samples passed quality control. Enriched genes were determined using DESeq2 including an effect of sample pair in the model to account for pairing of the bead-sup samples (~pair + cells).

Integration with published data
Count tables from Kim et al., 2019 were downloaded from https://data.mendeley.com/datasets/ ypx3sw2f7c/3 and count tables from Campbell et al., 2017 were downloaded from https://www. ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE93374. Note: Only the 10Â data from Kim et al. was used. The data was then preprocessed and clustered in the same way as above, though some samples were removed from the Kim et al. dataset for low mean read depth that confounded clustering (Samples 2018_0802, 2018_0803, 2018_0812_1, 2018_0812_2, and 2018_0812_3). For the Kim et al. dataset, cells were clustered as above (FindNeighbors then FindClusters), and neuron clusters were predicted using WGCNA to identify correlated gene expression modules. The modules contained dozens of genes that mapped clearly onto a small set of clusters in UMAP space and-based on known marker genes-corresponded to the most prevalent cell types in the brain (e.g., neurons, astrocytes, microglia, oligodendrocytes, etc.). For the Campbell et al. data, neurons were labeled in the metadata from the authors, so neuronal barcodes were simply selected based on their annotation. For both neuronal datasets, VMH neurons were predicted using the same procedure as above: clusters that were glutamatergic, Fezf1/Nr5a1-expressing, and similar to Nr5a1-Cre TRAP-seq. This corroborated the clusters called VMH in both datasets by the original authors, with the exception of the Nfib populations in the Kim et al. dataset that was not called VMH and therefore not assigned a cluster name; we refer to these cells as (Missing) in our integrated dataset.
To find shared populations across datasets, we took two approaches. First, variable genes were found for both datasets using the Seurat FindVariableFeatures function. Then, the pairwise Pearson's correlation was found for the mean scaled expression in each cluster in each dataset for the set of genes called variable in both datasets. Additionally, we used the Seurat FindTransferAnchors and IntegrateData functions to generate a merged dataset that was then subjected to PCA, UMAP reduction, and clustering in the same way as above. These new clusters containing cells from both datasets were then manually named using marker genes from the original datasets (e.g., Dlk1, Esr1,Satb2,Lepr,Nfib,Foxp2,etc.).

Immunostaining
Animals were perfused with phosphate buffered saline (PBS) for 5 min followed by an additional 5 min of 10% formalin. Brains were then removed and post-fixed in 10% formalin for 24 hr at room temperature, before being moved to 30% sucrose for 24 hr for a minimum of 24 hr at room temperature. Brains were then sectioned as 30-mm-thick free-floating sections and stained. Sections were treated with blocking solution (PBS with 0.1% triton, 3% normal donkey serum; Fisher Scientific) for at least 1 hr. The sections were incubated overnight at room temperature in blocking solution containing primary antibodies. The following day, sections were washed and incubated with fluorescent secondary antibodies with species-specific Alexa Fluor-488 or -568 (Invitrogen, 1:250) to visualize proteins. Primary antibodies used include: GFP (1:1000, #1020, Aves Laboratories) and NFIA (1:500, #PA5-35936, Invitrogen). Images were collected on an Olympus BX51 microscope. Images were background-subtracted and enhanced by shrinking the range of brightness and contrast in ImageJ.

Macaque snRNA-seq
Whole Rhesus macaque brains were obtained from the Tissue Distribution Program at ONPRC. Brains were centered within a chilled brain matrix (ASI Instruments, catalog # MBM-2000C) and 2 mm slices were obtained from rostral to caudal. Slices containing the hypothalamus were placed in saline and the PVH, ARC, VMH, and DMH were punched out and samples were placed in pre-chilled tubes on dry ice. Samples were stored at À80˚C until shipment on dry ice. Nuclei were isolated from frozen macaque tissue as described above for mice.
The FASTQ files were mapped to the macaque genome (Mmul_10) using cellranger and count matrix files were analyzed in R using Seurat three as above, with the exception that gene models and genes mapping to the macaque mitochondrial genome were not removed. Macaque neurons were predicted using orthologs of mouse cell-type marker genes and macaque VMH neurons were identified using macaque orthologs of Nr5a1-Cre TRAP-seq enriched genes. Orthologs were identified using Ensembl and only 1:1 orthologs were retained.

Species integration
The mouse and macaque datasets were integrated in a similar way to integrating the Kim et al. and Campbell et al. mouse datasets. First, the macaque genes were renamed to their mouse orthologs and only 1:1 orthologs and genes expressed in both species were retained. Importantly, because our dataset was biased toward the dorsomedial VMH and the Kim et al. dataset was biased toward the ventrolateral VMH, we also included four randomly chosen samples from the Kim et al. data to get a more representative picture of shared VMH populations across species. The data was then preprocessed, normalized, and scaled in the same way as previously. The mouse and macaque data was then integrated using the Seurat FindTransferAnchors and IntegrateData functions and marker genes were found that were common across species by running FindAllMarkers for each species separately and then using the metap logitp function to find genes that are significantly enriched.

Macaque ISH
Whole Rhesus macaque brains were obtained from the Tissue Distribution Program at ONPRC. Hypothalamic blocks fixed with 4% paraformaldehyde were incubated in glycerol prior to freezing with isopentane. Tissue was sectioned at 25 mm using a freezing stage sliding microtome and freefloating sections were stored in glycerol cryoprotectant at À20˚C. Tissue was mounted on slides prior to ISH, which was performed using ACD Bio RNAScope reagents (Multiplex Fluorescent Detection Kit v2, 323100) for Acvr1c (ACD 591481), Slc17a8 (ACD 543821-c2), and Lepr (ACD 406371-C3). Negative and positive control probes were included in all runs. Slides were imaged on an Olympus VS110 Slidescanner and processed using Visiopharm software.

Statistical analysis
All data is displayed as mean ± SEM. All plotting and statistical analysis was performed using R 3.6.3. Specific statistical tests are listed in the figure legends.

Resource availability
All mouse strains will be made available upon reasonable request.