Spatially resolved molecular sex differences at single-cell resolution in the adult human ventromedial and arcuate hypothalamus

SUMMARY

The hypothalamus contains multiple regions, including the ventromedial hypothalamus (VMH) and arcuate (ARC), which are responsible for sex-differentiated functions such as endocrine signaling, metabolism, and reproductive behaviors. While molecular, anatomic, and sex-differentiated features of the rodent hypothalamus are well established, much less is known about these regions in humans. Here, we provide a spatially resolved single-cell atlas of sex-differentially expressed (sex-DE) genes in the human ARC and VMH. We identify neuronal populations governing hypothalamus-specific functions, define their spatial distributions, and show enrichment of sex-DE in retinoid metabolism- and retinoid receptor-regulated genes. Within the ARC and VMH, we find correlated autosomal expression differences localized to ESR1/TAC3-expressing and corticotropin-releasing hormone receptor 2 (CRHR2)-expressing neurons and extensive sex-DE genes linked to sex-biased disorders, including autism, depression, and schizophrenia. Our molecular mapping of disease associations to hypothalamic cell types with established roles in sex-divergent physiology and behavior provides insights into the mechanistic bases of sex bias in neurodevelopmental and neuropsychiatric disorders.

In brief

The ventromedial hypothalamus (VMH) and arcuate (ARC) nuclei of the hypothalamus play critical roles in regulating metabolism and behavior. Here, Mulvey et al. create a focused spatial transcriptomic atlas of sex-differential gene expression in these two regions and localize their molecular sex differences to specific neuronal populations.

Graphical Abstract

INTRODUCTION

The hypothalamus (HYP) is a specialized brain structure that modulates behavioral and physiological drives essential to survival—ranging from social behavior ¹ to metabolism ² (Figure 1A). This functional breadth is mediated by different HYP regions (“nuclei”) with distinct circuitry, neurotransmitters, and neuropeptides. The tuberal portion of the HYP contains the ventromedial hypothalamus (VMH) and arcuate (ARC), which show sex-differentiated development ² and sex divergence at the levels of behavior,^3–5 physiology, ^6–8 cell types,^{9, 10} and gene expression^{8, 11} in model species.

Figure 1. Study overview and identification of VMH and ARC domains.

(A) Sagittal view of the human HYP highlighting the arcuate (ARC) and ventromedial hypothalamus (VMH) and main functions.

(B) Tissue block (Br1225) sectioned for Visium and Xenium, indicating the ARC, VMH, and landmark white matter tracts (fornix [Fx] and optic tract [OT]). Boxes denote sampling areas for Visium (blue) and Xenium (red), which have different-sized capture areas.

(C) Study design overview. We first utilized the Visium spatial gene expression platform, which uses ∼5,000 expression spots with unique spatial barcodes that correspond to an associated H&E image for a single capture area. We then performed in situ sequencing on the Xenium platform to acquire cellular resolution data for 266 pre-defined “brain panel” genes and 100 custom genes based on domain marker, spatial, and sex-differential patterns from Visium.

(D) BayesSpace clustering of Visium data shown for v1225A_M at k = 15, with 3 OT/white matter clusters grouped together and 2 GABAergic neuron clusters grouped together.

(E–G) Same sample as (D), illustrating the Visium (v)VMH and vARC domains, which, respectively, express known markers NR5A1 (F) and POMC (G). Approximate boundaries of the vVMH and vARC domains are labeled in the same colors as those in (E).

(H) Correlation between marker statistics for vVMH/vARC and mouse cell subclasses assigned to these domains in a recent whole-brain mouse atlasing dataset.¹²

(I) GSEA of vVMH- and vARC-enriched genes for ontology/function terms aggregated in MSigDB. NES, normalized enrichment score (a GSEA output metric). Sample identifiers are given as platform (v or x), donor number (4 digits), and replicate tissue section (A–D), followed by an underscore and donor sex.

The rodent VMH has molecularly distinct anatomic subdivisions ^3,13 linked to regulation of metabolism, appetite, and parenting behavior. The VMH modulates both adrenal ¹⁴ and peripheral ^15,16 nervous catecholamine release and contains a neuronal population expressing corticotropin-releasing hormone receptor 2 (Crhr2),¹⁷ suggesting critical roles in regulating the autonomic nervous system and stress-related behaviors. ^18,19 The VMH’s roles in metabolism ²⁰ include glucose-mediated modulation of brain-derived neurotrophic factor (BDNF) expression, ⁷ potentially in a sex-differential fashion. ²¹ The VMH is necessary for several sex-differentiated mouse behaviors, including mating, maternal aggression, and male fighting. ^1,22–24 Both of these functions are partially mediated by ventral lateral subdivision (VMHvl) neurons coexpressing estrogen receptors (Esr1) and progesterone receptors (Pgr)^3,22,23 and sexually differentiated during developmental androgen exposure. ¹⁰

The ARC is adjacent to the VMH and serves two primary functions, the first of which is metabolic regulation. ARC neurons detect circulating nutrients and nutritional status hormones (e.g., leptin)—in part via primary cilia²⁵ —and release neuropeptides including ghrelin, growth hormone-releasing hormone (GHRH), ²⁶ neuropeptide Y (NPY), and agouti-related peptide (AgRP) ²⁷ throughout the HYP and pituitary to modulate metabolism and behavior. Second, Kiss1 (kisspeptin)-Esr1-coexpressing ARC neurons regulate cyclic release of gonadotropin-releasing hormone (GnRH) from the HYP, resulting in pulsatile pituitary release of luteinizing hormone (LH) and follicle-stimulating hormone (FSH). ^9,28,29 Kiss1- and metabolism-regulating populations of the ARC bidirectionally influence one another to alter metabolism/growth according to reproductive status in a sex-differentiated manner. ^6,30

While the ARC and VMH have been extensively studied in rodents, little is known about their function in the human brain despite their fundamental importance to survival and their potential roles in neuropsychiatric disorders. At the molecular scale, in vitro studies of human ARC-like neurons have identified enrichment in genomic regulators associated with major depressive disorder (MDD) or bipolar disorder. ³¹ Furthermore, diagnostic features of several neuropsychiatric disorders implicate pathways in which the VMH/ARC participate: examples include eating disorders, which by definition involve appetitive/metabolic (dys)regulation; MDD, which can involve increased or decreased appetite; and anxiety/post-traumatic stress disorders, which may stem from deficits in fear extinction (in part a VMH-regulated process ³²). Notably, all of these disorders have a sex-differential prevalence. ^33–37 Likewise, sex differences in molecular and behavioral phenotypes relevant to mental health are documented in many rodent HYP studies.^{1, 3,4,22,23,32,38} In sum, observations ranging from animal models to clinical phenotypes suggest that the HYP may mediate sex differences in molecular risk for—and presenting features of—neuropsychiatric/neurodevelopmental disorders (NDDs), highlighting the importance of investigating molecular sex differences in the human VMH and ARC.

In-depth molecular profiling of individual subregions of the human HYP has been limited. Previous analyses of bulk RNA sequencing (RNA-seq) data from adult human tissue by UKBEC ³⁹ and GTEx ^40,41 revealed substantial sex differences in the HYP, though these could not be attributed to specific regions. ^42,43 To date, single-cell ⁴⁴ and spatial transcriptomic ⁴⁵ atlases from adult human HYP tissue have included only male samples or were underpowered for sex-differential analyses, respectively. To our knowledge, transcriptional sex differences within molecularly defined HYP cell types have been the focus of only one report, with limited information on tissue anatomy and sample handling. ⁴⁶ Thus, significant questions remain regarding the regional enrichment and cell-type specificity of molecular sex differences in the human VMH and ARC.

Here, we utilized the 10× Genomics Visium and Xenium platforms to profile spatial gene expression in the human postmortem VMH and ARC from 8 adult control donors without psychiatric diagnoses, evenly split by sex. A preliminary analysis of transcriptome-wide Visium data was used to prioritize genes for follow-up at single-cell resolution with a custom gene panel using Xenium. Using Visium, we illustrate that transcriptome-wide sex-differential expression patterns in the VMH and ARC are well correlated. Subsequent single-cell-resolution measurements with Xenium confirmed regional sex-DE observations and pinpointed the cell types responsible for observed sex-DE in each region. Moreover, nominally significant Visium sex-DE observations were resolved into significant, cell-type-specific sex-DE using Xenium; for example, MC4R showed only nominally significant sex-DE in the entire VMH but became significant at the cell-type level in CRHR2-expressing neurons of the VMH. Further, we identified potential transcription factors (TFs), including gonadal hormone- and retinoid-binding TFs and TFs associated with neuropsychiatric disorders, whose putative regulatory targets are enriched for sex-DE genes (sex-DEGs). Across the cell types residing in the VMH or ARC, we identify 122 genes that are both sex-DE and linked to neuropsychiatric disorders with sex differences in prevalence or clinical features. Finally, we provide evidence that VMH and ARC cell types are enriched for sex-DEGs associated with disorders that show sex differences in prevalence rates, including autism, depression, and schizophrenia (SCZ). We made this state-of-the-art multisample, multimodal spatial atlas of molecular features and sex differences in the adult human VMH and ARC available via interactive web resources at https://research.libd.org/spatial_HYP/.

RESULTS

Molecular identification of the human VMH and ARC using transcriptome-wide spatial transcriptomics

To understand sex-differentiated gene expression at cellular and spatial resolution in the human VMH and ARC, we performed Visium spatial gene expression profiling followed by Xenium in situ sequencing (10× Genomics) on adjacent sections of postmortem human HYP tissue from male and female adult neurotypical donors (N = 4 per sex; BMI range: for males, 23.1–28.7, and for females, 24–31.6; Figures 1A–1C). We used Visium to measure transcriptome-wide gene expression across the VMH and ARC, facilitating data-driven selection of genes for expression profiling at cellular resolution using Xenium (Figure 1C). To locate the VMH and ARC, we performed single-molecule fluorescence in situ hybridization (smFISH) for NR5A1 and NPY, marker genes for the VMH and ARC, respectively. We used NR5A1 expression to identify VMH-containing tissue sections, as it is expressed in very few areas of the brain, and the rodent ortholog Nr5a1 is a canonical VMH marker. When the VMH and ARC were reached, tissue sections were first collected for Visium (10 total samples, 1–2 sections per donor) (Table S1); sections from these same tissue blocks were later obtained for Xenium experiments. After quality control (Figures S1 and S2), the filtered Visium dataset included ∼45,000 spots and ∼30,000 genes.

To identify data-driven spatial domains representing the VMH and ARC from Visium, we performed unsupervised spatial clustering. Specifically, we used 1,816 spatially variable genes (SVGs; Table S2)⁴⁷ as features for dimensionality reduction and batch correction with harmony ⁴⁸ (Data S1), followed by unsupervised clustering at k = 15, 20, or 31 (STAR Methods). We present results using k = 15 (Figure 1D), as this value is consistent with the number of anatomical regions annotated in the Allen Brain Atlas in our samples. At k = 15, we discerned 2 oval-shaped VMH clusters and two ARC clusters ventral to them (Figure 1E). The VMH clusters appeared somewhat concentric in many samples, at first suggesting organization similar to the excitatory core and thin, cell-poor inhibitory shell found in the rodent VMH.^{3, 49} However, these VMH clusters were significantly enriched in glutamate transporter SLC17A6, precluding us from annotating a possible inhibitory shell at Visium resolution. Both clusters also expressed signature VMH marker genes FEZF1 and NR5A1, along with ANKRD34B (Figures 1F, S3, and S4); NR5A1 was sparsely expressed compared to FEZF1, ANKRD34B, and SLC17A6. ARC clusters were enriched in KISS1—a regulator of GnRH—as well as neuropeptide-encoding genes GHRH, AGRP, NPY, and POMC (Figures 1G and S4; Tables S3A–S3C). We henceforth refer to these Visium (v) clusters as vVMH.1, vVMH.2, vARC.1, and vARC.2 and, when considered together, as “vVMH” and “vARC.”

We examined our two vVMH clusters to determine whether they contained spatially restricted cell types. Pairwise differential expression between vVMH clusters demonstrated that vVMH.1 was more enriched for canonical marker genes, while vVMH.2 was more enriched for glial markers (e.g., GFAP and MBP; 0 < log fold changes [logFCs] <0.75; Figure S4; Table S4). This was also the case for vARC, suggesting that the vVMH and vARC clusters reflect local differences in glial content rather than spatially exclusive neuronal populations per se. In neither the vVMH nor the vARC did higher k values yield further molecularly distinct (neuronal) clusters inside the ARC or VMH domains (Figures S4, S5, and S6). Therefore, we present the below results at k = 15, with vVMH.1 and vVMH.2 joined into a single vVMH cluster, and vARC.1/vARC.2 joined likewise. Nonetheless, supplemental information from our downstream analyses (markers, sex-DE, and gene set enrichments) provides results for each Visium cluster at k = 15, 20, or 31 (including individual VMH and ARC clusters) alongside the results from the joined vVMH and joined vARC domains we focus on below.

To examine how well the vVMH and vARC domain marker genes corresponded to those of the adult mouse VMH and ARC, we compared our human data to a recently published single-cell atlas of the adult mouse brain. ¹² vVMH, vARC, and optic tract (OT) clusters shared gene marker signatures with mouse VMH, ARC, and white matter cell types, respectively (Figure 1H; Table S5). We obtained similar results when comparing marker gene expression against adult mouse in situ hybridization data from the Allen Brain Atlas⁵⁰ (Table S5).

We next performed gene set enrichment analysis (GSEA) for the vVMH and vARC using MSigDB-indexed gene sets (STAR Methods) (Figure 1I; full results of this analysis at k = 15 [with the vVMH and vARC as well as their individual clusters], 20, and 31 are available in Data S2). As expected, the vARC was enriched for inhibitory neuron terms, the vVMH for excitatory neuron terms, and both for terms related to neuropeptides. vVMH-enriched genes were overrepresented for kisspeptin signaling pathways, which is consistent with previous literature demonstrating VMH responses to kisspeptinergic neuron activity ⁵¹ and with our observed expression of kisspeptin receptor (KISS1R) in the vVMH (Tables S3A–S3C). Further, “regulation of insulin secretion” genes were enriched among highly expressed genes in the vVMH, consistent with its glucoregulatory function in rodents. ^52–54 Interestingly, vARC-enriched genes were overrepresented for terms relating to primary cilia and their organizing body, the axoneme. We confirmed this by performing GSEA with a curated set of human primary cilia genes⁵⁵ and found ARC-specific enrichment for primary-cilium-localized genes (false discovery rate [FDR] < 5 × 10⁻⁷ ; Tables S6I and S6J). Taken together, these results show that we successfully identified the ARC and VMH in our human Visium samples and that the signals, circuits, and cellular biology in these regions are broadly consistent with expectations from the rodent literature.

Correlated molecular sex differences in VMH and ARC domains

To assess transcriptome-wide sex-DE, we first verified that the recorded donor sex and expression of sex-specific genes (chromosome Y [chrY] or XIST) were concordant for each donor. We then analyzed the vARC and vVMH to identify genes with statistical (Table S7) or visual evidence of sex-DE and prioritized these for cell-resolution follow-up using the Xenium platform. Pseudobulk DE analyses revealed 11 significant vVMH (Figure 2A) sex-DEGs (FDR < 0.05), all of which were allosomal (10 on chrY and XIST) (Figure S7). Orthologs to sex-DEGs previously identified in the mouse VMH (Cckar, Rprm, and Pdyn) ⁸ did not achieve nominal significance in the human vVMH, though we did note nominally significant male-biased expression of an obesity-associated ⁵⁶ melanocortin receptor, MC4R (Figures 2B and 2C), which offers an example case for how genes were selected for Xenium. We included MC4R for follow-up on Xenium based on (1) prior evidence for MC4R involvement in VMH-mediated metabolic regulation, ^57,58 (2) nominally significant sex-DE in the vVMH, and (3) appreciable vVMH expression across the cohort.

Figure 2. Sex-differential expression at the domain level in the human VMH and ARC.

(A) Volcano plot illustrating autosomal sex-differential expression within vVMH. The black line indicates nominal p < 0.05; the red line indicates FDR < 0.05.

(B) Spotplots from one male (top; v6197A_M) and one female (bottom; v6588A_F) Visium sample illustrating the nominally significant male upregulation of MC4R in vVMH (thicker domain border colored in purple).

(C) Pseudobulk domain-level expression (log2 counts per million [CPM]) of MC4R in all vVMH samples, stratified by sex. Each point is one pseudobulked sample. Boxplots indicate the median (center lines) and 25^th /75^th percentiles (bottom/top box edges), with error bars spanning samples within ±1.5× the interquartile range.

(D) Volcano plot illustrating autosomal sex-DE within vARC. Lines are the same as those described for (A).

(E) Spotplots from the same male (top) and female (bottom) Visium samples as in (B), illustrating female-preferential expression of RDH10 in ARC (thicker domain border colored in green).

(F) Pseudobulk domain-level expression (log2 CPM) of RDH10 in all vARC samples, stratified by sex. Boxplot extents are the same as those described for (C). Sample identifiers are given as platform (v or x), donor number (4 digits), and replicate tissue section (A–D), followed by an underscore and donor sex.

Meanwhile, we identified 26 significant vARC (Figure 2D) sex-DEGs (FDR < 0.05, Table S7), 13 of which were autosomal. These included female-biased expression of RDH10, which encodes an enzyme responsible for converting retinoids into active ligands for retinoid receptors (retRs) (Figures 2E and 2F). The distribution of p values for both analyses showed an overrepresentation of p values close to zero, suggesting a sizable number of “true” differentially expressed genes in both domains, higher than the 11 or 26 we report here (Figures S8A–S8D). We also performed sex-DE analyses for individual vVMH and vARC clusters, revealing the same general pattern, with 36 sex-DEGs (13 allosomal) in vARC.1, 17 (12) in vARC.2, 15 (12) in vVMH.1, and 9 (9) in vVMH.2 (Table S7B).

We observed a strong concordance between autosomal sex-DE in the vARC and vVMH, with greater FCs in the vARC. The effects of sex were correlated (t statistic Spearman’s rho = 0.542; logFC Spearman’s rho = 0.585; Figures S8A and S8B), and the correlation increased when we restricted to genes nominally DE in either of the two domains (t statistic rho = 0.586; logFC rho = 0.683). Among nominal autosomal sex-DEGs in Visium, 211 were found in both domains, of which 210 had concordant effects. The one exception was histamine H1 receptor (HRH1), which showed greater vARC expression in females (logFC = 0.77) and greater vVMH expression in males (logFC = 0.49). Analysis of the p value distribution of vVMH conditional on the vARC (and vice versa) likewise suggested a substantial overlap among the sex-DEGs in the two domains (Figures S8E and S8F).

Cell-level expression of sex-DEGs in VMH and ARC with imaging-based spatial transcriptomics

Following an analysis of preliminary Visium data, a unique panel of 100 genes was designed for the Xenium platform (Table S8). This panel included vVMH and vARC marker genes, 36 putative sex-DEGs identified with Visium, sex hormone receptors (the androgen receptor, AR; nuclear estrogen receptor ESR1; and a G-protein-coupled estrogen receptor, GPER1), and genes with spatial or sex patterns from rodent HYP literature (STAR Methods). We additionally performed SVG analyses of the vVMH or vARC in isolation to identify expression patterns suggestive of fine-grained cellular organization in these domains. We identified 277 SVGs in the vVMH and 1279 SVGs in the vARC (Table S9); among genes in the latter set were 11 markers of 6 ARC clusters previously reported in the developing human HYP, including TAC3, GHRH, and DIRAS3. ⁵⁹ All 11 of these ARC cluster markers were significant SVGs in ≥6 tissue sections, with mean ranks skewed toward the top (Wilcoxon test, p < 1 × 10⁻⁷), and 8/11 were in the top 150 vARC SVGs overall. We thus concluded that genes with strong SVG signals and consistent patterns within a single Visium domain were likely indicative of discrete neuron types not resolvable as distinct clusters using Visium; therefore, several of these SVGs were included in the Xenium panel (Figures S9–S17). Our 100 custom-selected genes were supplemented by a commercially available “human brain” panel of 266 genes spanning markers for broad cell types largely derived from human cortex data. Since our goal was to study sex differences, we balanced each slide to contain sections from both sexes and each processing batch to include an equal number of male and female donors. This ensures that any unwanted variation associated with the processing batch is independent of sex differences.

From all 8 donors, tissue sections adjacent to those profiled with Visium were collected for Xenium (N = 7 male and N = 6 female sections). After quality control (Figure S18), data were clustered (Figures 3A and S19), yielding 33 clusters total, including 5 Xenium ARC (xARC) clusters and 3 Xenium VMH (xVMH) clusters identifiable by marker genes and spatial distributions (Figures S19, S20A, and S20B; Table S3G). All three xVMH clusters shared enrichment for a number of VMH marker genes (SLC17A6, NR5A1, FEZF1, NRGN, ANKRD34B, ADCYAP1, and HS3ST4; Figure S20B). Comparison across these three clusters revealed that xVMH.1 was highly enriched for CRHR2, known to be expressed in the rodent VMH. ^60,61 Interestingly, xVMH.2 was comparatively enriched for the oxytocin receptor (OXTR) but also for astrocyte (GJA1 and SOX9) and microglia (P2RY13) markers (Figure S20C; Table S3I); as such, we refer to xVMH.2 as “VMH-glia mixed.” xVMH.3 was enriched for LAMP5; as LAMP5 is associated with inhibitory neurons in other brain regions (in contrast to the VMH, which is composed of excitatory neurons), we performed smFISH to confirm LAMP5 colocalization with excitatory neuron marker SLC17A6 and VMH marker NR5A1 (Figures S21 and S22). A fourth, HCRT-expressing cluster showed striking restriction to the VMH perimeter; we grouped this “lateral border” population as a fourth xVMH cluster in downstream analyses based on its surprising spatial pattern (Figure S23).

Figure 3. Xenium clustering, domain identification, and sex-DE analyses at the domain level.

(A) Xenium cell-type clustering from the same donor (sample x1225B_M) as shown in Figure 1D, with clusters combined into Visium-analogous groups for visualization. The legend notes the number of cell-type clusters in the visualized group.

(B) Xenium (x)VMH and xARC anatomical boundaries as inferred using k-nearest neighbors analyses to quantify the number of spatial neighbor cells from xVMH or xARC clusters.

(C) Volcano plot illustrating autosomal sex-DE within xVMH (all cells from all clusters and within xVMH). The black line indicates nominal p < 0.05; the red line indicates FDR < 0.05.

(D) Volcano plot illustrating autosomal sex-DE within xARC (all cells from all clusters and within xARC). Lines are the same as those used in (C).

(E) Expression spotplots of example male (top) and female (bottom) Xenium samples x6197B_M and x6588C_F, illustrating male-preferential expression of MC4R in VMH (thicker domain border colored in purple). The boxplot inset (as described in the legend of Figure 2C) shows pseudobulk domain-level expression (log2 counts per million [CPM]) of MC4R from each xVMH sample, stratified by sex.

(F) Xenium samples from the same representative donors as in (E), illustrating female-preferential expression of RDH10 in ARC (thicker domain border colored in green). The boxplot inset (as described in the legend of Figure 2C) shows pseudobulk domain-level expression (log2 CPM) of RDH10 in each xARC sample, stratified by sex. Sample identifiers are given as platform (v or x), donor number (4 digits), and replicate tissue section (A–D), followed by an underscore and donor sex.

Of our 5 xARC clusters (Figure S24), 3 were readily identifiable based on known marker genes from the literature: xARC.1 as AGRP-expressing neurons, xARC.3 as a TRH-expressing subpopulation of GHRH neurons, and xARC.4 as TAC3- and ESR1-expressing neurons, possibly representing KISS1-expressing KNDy (kisspeptin-neurokinin-dynorphin) neurons ^62,63 (Figure S20). For the two remaining clusters, xARC.5 was comparatively enriched in oligodendrocyte markers (OLIG1, OLIG2, and PDGFRA) along with POMC and thus was designated “POMC-oligo mixed,” while xARC.2 was enriched for SLC17A7 and glial markers and designated “SLC17A7-glia mixed” (Figure S20C; Table S3H). Regarding relative enrichments of glial genes when comparing among xARC or xVMH clusters, we repeated our analysis using only nuclear-expressed transcripts and obtained similar results (Tables S3J and S3K), suggesting close interactions between glial processes and neurons in select clusters.

We then sought to examine whether Xenium sex-DE patterns were consistent with Visium at the domain level. To achieve this, we first defined ARC and VMH domains analogous to those from Visium by developing a smoothing algorithm (STAR Methods) to obtain anatomically contiguous domains corresponding to the xARC and xVMH (Figure 3B). Consistent with Visium, Xenium domains for the xVMH and xARC were respectively enriched for expected marker genes NR5A1 and POMC, with qualitatively strong anatomical agreement within donors and across platforms (Figure S25). With the xARC and xVMH defined, we pseudobulked across all cells within the xVMH or xARC and performed sex-DE analyses, covaried by Xenium run (Figures 3C and 3D), to assess cell-type-agnostic sex-DE (analogous to vVMH/vARC sex-DE analyses). While no genes achieved significance at an FDR of <0.05 in Xenium domains, trends identified in Visium were replicated, such as male upregulation of MC4R in the VMH (Figure 3E) and female upregulation of RDH10 in the ARC (Figure 3F). Systematic examination of sex-DE across platforms revealed correlated autosomal effects of sex (t statistics; ARC, 277 genes, Spearman’s rho = 0.63, p < 2.2 × 10⁻¹⁶; VMH, 245 genes, Spearman’s rho = 0.41, p < 5.5 × 10⁻¹¹; Figure S8G). The correlation increased when only considering the 36 genes earmarked for sex-DE follow-up on Xenium (ARC, 33 genes, rho = 0.80, p < 5.8 × 10⁻⁷; VMH, 27 genes, rho = 0.51, p < 0.01; Figure S8H). These findings suggest that sex effects at the domain level were consistent across platforms, regardless of DE significance.

Additionally, we examined domain-level expression of estrogen, progesterone, and androgen receptors using both platforms. As anticipated from rodent and human data, ⁶⁴ ESR2 expression was weaker relative to ESR1 in the VMH and ARC (Figure S26A). Visium expression of progesterone receptor (PGR) was moderate in both the vVMH and vARC, consistent with its protein expression pattern in the human HYP. ⁶⁵ PGR was also nominally sex-DE (male elevated) in the vARC (logFC = 0.94, p < 0.001) and trended similarly in the vVMH (logFC = 1.23, p = 0.053; Table S7). PGR (Visium only), ESR1, and AR were detected in a greater fraction of ARC spots/cells than in the VMH. Contrary to expectations from rodent literature showing enrichment of ESR1 in the VMHvl, ^{2–4,8,10,22,23} ESR1 expression was sparse and spatially uniform in the human xVMH. smFISH on Visium-adjacent tissue sections reproduced this finding (Figure S27A; see data and code availability for full-tissue ESR1 smFISH microscopy). The dense expression of ESR1 in the human ARC compared to the VMH further suggests that our observations of more marked sex-DE in the ARC domain are driven by hormonally mediated (activational) signaling.

Single-cell analysis reveals especially marked sex effects in ESR1-expressing neurons of ARC

Next, to determine whether the VMH and ARC showed cell-type-specific sex differential expression, we examined all Xenium-measured genes for sex-DE within individual cell types in the xVMH or xARC domains (Figure S28; Tables S10A and S10B). In the xVMH, we found the most significant (FDR < 0.05) sex-DEGs in excitatory CRHR2-expressing cells, involving 39 genes, with another six DEGs in the xVMH-glia-mixed cluster (Figures 4A and S29). Among these, TAC1 was female upregulated in the xVMH-glia-mixed cluster with strong trends (FDR < 0.1) toward female upregulation in CRHR2 and LAMP5 clusters—consistent with sex-DE of Tac1 observed in the mouse VMH ^8,66 —and CYP26A1 trended toward female upregulation in CRHR2 and xVMH-glia-mixed clusters (Figure 4B). Male xVMH CRHR2 neurons showed significantly (FDR < 0.1) greater expression of MC4R (Figure 4B), suggesting that sex-DE in xVMH CRHR2 neurons was driving our above observations of a trend toward domain-level sex-DE in the vVMH and xVMH. Likewise, MYT1L, a TF mutated in a syndromic form of autism spectrum disorder (ASD) involving obesity, ⁶⁷ was selectively male upregulated in the CRHR2 population (Figure 4B).

Figure 4. Sex-DE within Xenium-resolved cell types of the human xVMH and xARC.

(A) Volcano plots of sex-DE for selected xVMH cell clusters. The black lines indicate nominal p < 0.05 while red lines indicate FDR < 0.05.

(B) Top candidate sex-DE genes from vVMH measured using Xenium. The boxplots (as described in the legend of Figure 2C) show intra-xVMH gene expression for a cluster (color legend is shared with A), pseudobulked by sample and stratified by sex. Cell types trending toward sex-DE (FDR < 0.1) are colorized in this plot.

(C) Volcano plots of sex-DE for selected xARC cell clusters, highlighting that sex-DE is most prominent in terms of significance and effect size in TAC3-ESR1 cells.

(D) Top candidate sex-DEGs identified in vARC measured using Xenium. The boxplots (as described in the legend of Figure 2C) show intra-xARC gene expression for a cluster (color legend is shared with C), pseudobulked by sample and stratified by sex. Cell types trending toward sex-DE (FDR < 0.1) are colorized in this plot.

(E) Visualization of individual xVMH/xARC clusters wherein trending sex-DE was most marked. The top row illustrates individual xVMH clusters in shared x-y space from one Xenium sample (x6588C_F); the bottom row illustrates individual xARC clusters in the same manner. As with other spatial plots, down is ventral, and left is medial.

(F) Confocal image showing expression of ESR1, TAC3, and KISS1 in an adjacent tissue section by smFISH in a region of ARC expressing all 3 genes (scale bar: 25 μm). Arrows indicate cells expressing all three genes of interest. The arrowhead indicates an example cell with only TAC3 and ESR1 expression. Sample identifiers are given as platform (v or x), donor number (4 digits), and replicate tissue section (A–D), followed by an underscore and donor sex.

Sex-DE (FDR < 0.05) in the xARC was most extensive in TAC3-ESR1 neurons (67 genes, Figure 4C) and detected in all xARC populations except the POMC-oligo-mixed cluster (Figure S29). As with xVMH cell types, we found male upregulation of MYT1L to be cell-type specific in the xARC, only observed in TAC3-ESR1 cells (Figure 4D). TAC3-ESR1 neurons also showed specific and marked male upregulation of CDH13, a receptor for insulin-sensitizing peptide adiponectin ⁶⁸ implicated in obesity and type 2 diabetes ^69–71 (Figure 4D). RetR RXRG, a male-upregulated DEG in the vARC, showed less cell-type-selective sex-DE, with significant male upregulation in three xARC clusters and a trend toward the same in TAC3-ESR1 neurons (Figure 4D). Consistent with domain-level findings, retinoid metabolism genes were female upregulated in the xARC, with RDH10 showing DE in four populations (Tables S10A and S10B) and vARC marker CYP26A1 showing significant or trending DE in two xARC clusters (Figure 4D). No significant DEGs showed opposite directions of effect among or across xARC and xVMH cell types. However, in agreement with vARC and vVMH findings, HRH1 was significantly upregulated in two xARC clusters from females while trending toward male upregulation (p < 0.07) in xVMH CRHR2 cells.

We next examined sex hormone receptor expression patterns at the finer resolution of individual cell clusters (Figure S26B). The greatest proportion of cells with a ≥1 count of ESR1 included xARC TAC3-ESR1, xARC AGRP, and xVMH CRHR2—all populations notable for having a number of significant sex-DEGs—in addition to cells of microglia cluster 1 and oligodendrocyte progenitors (OPCs) residing in either the xVMH or xARC. AR was generally found in a similar proportion of cells for each cluster-domain pair, with the notable exception of xARC GAL-TRHGHRH, where it was detected in ∼2-fold more cells than ESR1.

To confirm that our sex-DE findings were not confounded by ARC/VMH cell-type intermixing at domain boundaries, we repeated the sex-DE analyses by reducing both the xVMH and xARC domain boundaries inward by 4%, which gave well-correlated t statistics relative to the original domain boundaries for all 5 xARC clusters and the 3 main xVMH clusters (all Spearman’s rho > 0.99) (Tables S10F–S10H). Similarly, we ruled out effects of cell segmentation by repeating the analyses using only nuclear transcript counts, which also yielded consistent results (Tables S10D, S10E, and S10H).

As several instances of sex-DE in the xARC or xVMH were cell-type specific, we next examined whether these cell types displayed spatial organization that would suggest a specific anatomic subdomain where sex effects were most pronounced. xVMH clusters showed no spatial organization, with each occupying the entirety of their domain (Figure 4E). The xARC, on the other hand, showed exquisite spatial isolation of TAC3-ESR1 cells in particular. Consistent with Visium SVG results (Figures S14 and S15), GHRH and TAC3 expression was mutually exclusive within the xARC (Figure S30). Given both spatial and sex-DE features of our TAC3-ESR1 cluster, we sought to more definitively annotate these cells. Although KISS1 was not included in our Xenium panel, smFISH showed that ∼14% of ESR1+/TAC3+ neurons in the ARC also express KISS1, suggesting that KNDy neurons are included within this population (Figures 4F and S27). Additionally, orthologs to secondary markers of mouse KNDy neurons ^62,63 were also enriched in our TAC3-ESR1 population (ECEL1 and NR4A2; Figure S20; Table S3G). These findings highlight a previously unrecognized ARC subdivision in which TAC3-ESR1 neurons (both KISS1-positive and -negative), and thus much of sex-DE in the adult ARC, are restricted. Altogether, we replicated and refined several Visium findings at higher spatial resolution and identified specific xARC and xVMH cell types responsible for sex-DE signals. Moreover, we identified cell populations under the strongest transcriptional influence of sex in each domain and illustrated that genes—including those linked to known VMH/ARC functions such as metabolism and obesity (e.g., MC4R and CDH13)—are influenced by sex in the human VMH and ARC.

Male-upregulated genes in the VMH and ARC are enriched for NDDs

We next sought to leverage our transcriptome-wide data to identify potential transcriptional regulators of adult vVMH- and vARC-defining expression patterns. To achieve this, we performed GSEA on one-vs.-all marker analysis test statistics in each Visium domain using aggregated databases of putative TF target genes across mammalian tissues and cell types (Table S6; Data S3; see STAR Methods). Results are also provided for individual vVMH/vARC domains at k = 15 and for all domains at k = 20 or 31 in the supplemental information of this resource (Table S11; Data S2, S3, S4, and S5). Gene targets of the glucocorticoid receptor (NR3C1) and of a putative retinoic acid receptor (NR2F2) were enriched in the vVMH. ⁷² Similarly, targets of the traditional retRs RXRA and RARA were enriched in the vVMH and vARC, respectively (Figure 5A). Finally, several results suggested sex differences in the vVMH and vARC, including AR target enrichment in the VMH and AR interactor (i.e., co-TFs) ⁷³ target enrichment in ARC. Furthermore, genes upregulated or downregulated by the perturbation of EGR1, a TF that sex-differentially modifies chromatin in the hippocampus, ⁷⁴ were respectively enriched in the vARC or vVMH. These findings collectively highlight previously underappreciated signaling pathways in the human VMH and ARC.

Figure 5. Sex-DE patterns of the vARC/vVMH are enriched in ASD/NDD risk genes, including disease-associated TFs.

(A) GSEA enrichment of vVMH/vARC markers, weighted by one-vs.-all marker t statistics in neuropsychiatric disorder- and BMI-associated genes. *FDR < 0.05; ⋅uncorrected p < 0.05.

(B) GSEA enrichment of genome-wide vVMH/vARC sex-DE for TF targets/TF interactors, with points sized based on normalized enrichment score (NES). For GEO TF perturbation signatures identified in paired perturbation-control replicates, a point representing each replicate’s signature is plotted for a given sex-DE test.

(C) GSEA enrichment of genome-wide vVMH/vARC sex-DE, weighted by signed DE t statistics, in neuropsychiatric disorder- and BMI-associated genes. Tests of enrichment for patterns of male and female upregulation are shown, along with a test considering sex-DE in both directions. *FDR < 0.05; ⋅uncorrected p < 0.05.

(D) GSEA enrichment of Visium sex-DE for ontological and functional terms indexed in MSigDB, with points sized based on NES. The legend is shared with (B).

(E) ASD-associated TF, MYT1L, whose targets are overrepresented in genes with male-preferential vARC and vVMH expression (B), is also male upregulated in xVMH, predominantly in the CRHR2 population. Cell-level expression is plotted for two samples per sex (sex is denoted in the sample name) for xVMH clusters CRHR2 and LAMP5.

(F) A hypothesized mechanism that may connect our observations of sex-DE genes, the shared TFs regulating them, and sex-DE enrichment for ASD genes. MYT1L (itself an ASD risk gene) is a male-upregulated TF in our data (left side); likewise, male-upregulated genes in VMH are overrepresented in MYT1L regulatory targets, including ASD risk genes (as discussed in the results; right side), suggesting that at the level of domains or cell populations, sex-DE of TFs such as MYT1L may also drive sex-DE of several further disease risk genes. AD, Alzheimer’s disease; ADHD, attention-deficit hyperactivity disorder; ASD, autism spectrum disorder; AUD, alcohol use disorder; BIP, bipolar disorder; BMI, body mass index (quantitative trait); Cannabis, cannabis use disorder; GAD, generalized anxiety disorder; MDD, major depressive disorder; NDD, neurodevelopmental disorder; OCD, obsessive-compulsive disorder; OUD, opioid use disorder; PTSD, posttraumatic stress disorder; SCZ, schizophrenia; TS, Tourette’s syndrome.

We then performed a similar GSEA examining sex-DEGs and the same TF-target gene sets, again filtering results to TFs expressed in the vVMH or vARC (see STAR Methods; Figure 5B; Data S4). Sex hormone receptor targets were enriched for DE in both the ARC and VMH. While ESR1 target sets were strongly enriched in the female vVMH and vARC, the male vVMH was weakly enriched for one target set and for ESR1 protein interactors. The female vVMH showed stronger enrichment for inferred AR targets than did the male vVMH, presumably representing female expression of genes repressed by AR activity. Finally, GSEA findings indicated that the AR targets enrichmed among vARC sex-DEGs were specifically those upregulated by hormone-independent AR activator ⁷⁵ and chromatin modifier ⁷⁶ PKN1 (Data S4).

Given that the prevalence rates for many NDDs and neuropsychiatric disorders are sex biased and that the HYP plays established functional roles in behaviors relevant to these disorders, we also hypothesized that expression patterns that define or are sex differentiated in the VMH and ARC may be enriched in NDD and neuropsychiatric disorder gene sets. Toward this end, we first performed summary LDSC applied to significantly expressed genes analysis ⁷⁷ to test whether genomic intervals surrounding marker genes or sex-DEGs were enriched for heritability of neuropsychiatric and metabolic diseases/traits based on genome-wide association studies (GWASs). These analyses only identified significant heritability enrichment (FDR < 0.05) for marker genes of the vVMH (in intelligence, educational attainment, and BMI; Table S11E) and for sex-DEGs of Visium domain GABA.1 in bipolar disorder (Table S7F). However, this approach accounts for neither long-distance gene regulation nor disease-associated genes identified in studies of rare or de novo genetic variation. Therefore, we collected gene sets from (non-proximity-based) analyses in GWAS publications, rare-variant studies (where applicable), transcriptome-wide association studies (TWASs), ^78–82 and the database DisGeNET, ⁸³ which includes curated disease-gene links (Table S6). These gene sets were combined into a single gene set for each of 18 disorders. We then performed GSEA to identify whether genes associated with traits/disease via functional genomics demonstrated marker-like expression or sex-DE in the vARC or vVMH. The vARC did not show an enrichment of genes associated with neuropsychiatric disorders, while the vVMH showed an enrichment of genes associated with several sex-biased disorders, including ASDs, NDDs, and SCZ (Table S11). A similar GSEA on sex-DE patterns in the vVMH or vARC revealed that male-upregulated genes from both the vVMH and vARC were overrepresented in genes associated with risk for ASDs and Tourette’s syndrome (TS), which are disorders more prevalent in males (Figure 5C; Table S7E). While our Visium analysis only identified three ASD-associated genes with nominal sex-DE in the vARC, TAC3-ESR1 cells in the xARC showed sex-DE (FDR < 0.1) for 11 such genes (AR, CDH13, DCC, GABRA5, GABRQ, GLRA2, MYO16, MYT1L, NR4A2, PCLO, and THBS1; Table S12). MYT1L, a gene mutated in a syndromic form of ASD involving obesity, ⁶⁷ was significantly male upregulated selectively in TAC3-ESR1 neurons, providing a previously unreported cell-type context for this risk gene. These findings exemplify how gene panel prioritization using Visium improves our ability to detect sex-DEGs in anatomically restricted cell types while retaining transcriptome-wide information to link such genes to broader pathways.

We then asked whether vVMH or vARC sex-DE corresponds to broader aspects of HYP biology that may be relevant to the enriched neuropsychiatric disorders, or to gonadal hormone signaling, by performing similar GSEAs of sex-DE patterns with pathway and ontology gene sets (Data S5). Male-biased vVMH genes were enriched for gene sets associated with aggression, which is consistent with behavioral functions of the male rodent VMH in regulating aggressive behavior. ^22,24 On the other hand, male-biased vARC genes were enriched for targets of miR30a, a regulator of puberty onset downregulated by perinatal estrogen exposure in female rodents ⁸⁴ (Figure 5D). Progesterone-responsive genes were enriched among female-biased genes of the vVMH (Figure 5D), despite trends toward greater male PGR expression in both the vVMH and vARC.

Last, we examined individual sex-DEGs on Xenium at an FDR of <0.1, focusing on 8 neuropsychiatric disorders/NDDs with prominent sex differences in prevalence and/or clinical features (attention-deficit hyperactivity disorder [ADHD], anorexia nervosa, ASD, MDD, TS, SCZ, generalized anxiety disorder, and panic disorder). In total, we obtained 122 disease-associated genes with sex-DE in ≥1 cluster within the ARC and 35 in ≥1 cluster within the VMH, with 29 genes in common between the two domains. These clusters include glial cells, interneurons, and other broad cell types that may interact differently with VMH- and ARC-specific neuronal populations (Figure S28). Strikingly, there were 70 ARC/22 VMH sex-DEGs associated with MDD, 68 ARC/16 VMH genes associated with SCZ, and 31 ARC/12 VMH genes associated with ASD (Table S12). MDD-, SCZ-, and ASD-associated genes were also frequent among sex-DEGs of domain-specific clusters, with 25 genes in ARC neurons and 22 in VMH neurons for MDD, 31 ARC/16 VMH for SCZ, and 13 ARC/12 VMH for ASD. While our Xenium data are not a genome-wide assessment, they establish that genes associated with several sex-biased disorders are also sex-DE in individual cell types of the ARC and/or the VMH, including their defining neuronal populations.

Our TF-target analysis also suggested sex-DE regulation by multiple TF-coding genes associated with ASDs and NDDs. The female vVMH and vARC were enriched for targets of ZNF142, which is associated with a rare, female-predominant neurodevelopmental syndrome, ⁸⁵ and had divergent target patterns between sexes. Genes upregulated by disruption of ZNF142 were enriched among female-biased vVMH and vARC genes, while those downregulated by disruption of ZNF142 were enriched among male-biased vVMH DEGs. As these targets are ascertained from TF perturbation experiments, this would suggest that ZNF142 activity in the VMH is ordinarily greater in males than in females. On the other hand, the male vVMH was enriched for targets of two syndromic ASD genes, MYT1L ⁶⁷ and ASH1L ⁸⁶ (Figure 5B). In the case of MYT1L, 5/38 of its target genes were nominally (p < 0.05) sex-DE in the vVMH, all male upregulated; similarly, all 14 genes in our ASD risk set that were also targets of MYT1L had a male-up direction of effect (regardless of DE significance). Moreover, MYT1L was itself male upregulated at an FDR of <0.1 in xVMH CRHR2 neurons (Figures 4B and 5D). These results provide further evidence that molecular sex differences in the VMH and ARC are highly relevant for NDDs and ASDs, suggesting a possible mechanism wherein sex-DE of a TF drives sex-DE of several additional disease-associated genes (Figure 5F).

Finally, given that the HYP plays a key role in energy metabolism, we also investigated gene sets relevant to BMI and obesity. GSEA did not reveal overrepresentation of BMI-associated genes in either vARC- or vVMH-enriched genes. However, GSEA using TF-target gene sets showed that both the male and female vARC, as well as the female vVMH, were enriched in target sets for STAT3, a TF necessary for leptin signal transduction. We also found female-specific enrichment of both the vARC and vVMH for targets of PPARA and PPARG, which play key roles in lipid and sterol metabolism (Figure 5B). Interestingly, both of these TFs appear to compete with estrogen receptor α for shared DNA-binding sites. ^87–89 In conclusion, these enrichments collectively illustrate that molecular sex differences in the tuberal HYP are organized by gene pathways and regulatory hubs critical to HYP functions, gonadal hormone signaling, and disease risk.

DISCUSSION

The tuberal HYP has been a focus of research investigating sex differences in neurobehavior and neuroendocrine processes, including an ample corpus of rodent research in these domains. However, the molecular anatomy of the human VMH/ARC remains relatively unexplored, especially in the context of sex differences. Here, we performed comprehensive spatial molecular profiling of two hypothalamic regions, the VMH and ARC, in a carefully selected set of healthy male and female control donors, enabling comparative transcriptomic analyses in the absence of critical brain or endocrinologic perturbations. Given that the ARC and VMH are small in size and lack clear anatomical boundaries, generating comprehensive molecular atlases for these regions in the human brain has been challenging to date. Using Visium, we were able to perform an in silico dissection of the VMH and ARC based on gene expression. While we did not observe discrete spatial clusters within the vARC or vVMH, we identified a preponderance of highly SVGs in the vARC, which overlapped with established marker genes of postmitotic human fetal ARC neurons. ⁵⁹ We hypothesized that higher-resolution assays, such as Xenium, would facilitate the identification of spatially restricted neuronal populations. Indeed, Xenium experiments recapitulated vARC SVG patterns and identified xARC clusters with distinct spatial localization.

Our spatial-molecular profiles highlight features of the HYP that are distinct from more commonly studied cortical regions, including unexpected LAMP5 expression in excitatory neurons of the VMH, complex/non-laminar cell-type organization in the ARC, and distinctive neuropeptide expression in these domains and their surroundings. We then achieved our primary goal of identifying signatures of sex-regulated gene expression in the VMH and ARC and localizing these to specific neuronal populations. Specifically, we highlight ESR1-expressing neurons in the ARC and CRHR2-expressing neurons in the VMH as hubs of molecular sex differences. Our approach provides a roadmap for using transcriptome-wide spatial assays to prioritize genes of interest for finer characterization at cellular resolution using imaging-based spatial transcriptomics. We highlight the melanocortin receptor MC4R to illustrate the utility of this prioritization approach. Whole-transcriptome Visium analysis at lower resolution detected nominally significant sex differences in MC4R expression. Given the relevance of melanocortins to appetitive/metabolic signaling in the rodent ventromedial HYP ⁵⁸ and subtle vVMH sex-DE, we included the gene in our follow-up Xenium assay. As with the vVMH, domain-level sex-DE of MC4R was not observed in the xVMH, while subsequent examination of individual cell types revealed significant sex-DE only in the xVMH CRHR2 neuron population. This demonstrates that trends observed in exploratory analyses of low spatial- or cell-type-resolution data can lead to molecular discoveries by guiding the design of focused, high-resolution follow-up assays.

Given abundant research suggesting sex-differentiated anatomy and molecular features of the VMH and ARC in model organisms, our primary objective was to examine the molecular impacts of sex on the human ARC and VMH. While a greater number of significant genes and larger FCs were observed in the ARC compared to the VMH, we identified biologically relevant sex-DE in both domains. These included genes with HYPrelevant functional roles and regulation: the vARC showed greater female expression of RAMP2 (logFC = − 1.49, FDR < 0.05; Table S7), which directly interacts with calcitonin-family receptors and glucagon receptors, ⁹⁰ consistent with the ARC’s role in metabolic regulation; the female vVMH and xVMH both showed greater expression of TAC1, a neuropeptide transmitter, and the male vVMH demonstrated elevated apelin receptor, APLNR (logFC = 1.22, p < 0.01). Interestingly, delivery of apelin into the rodent VMH increases lordosis (reproductive behavior) in females, ⁹¹ supporting a sex-differentiated role for this gene. In line with vVMH/vARC marker enrichment in retR-regulated genes, both the vARC and xARC demonstrated female upregulation of a retinol dehydrogenase, RDH10 (Figures 2E, 2F, and 3F), and male upregulation of the retR RXRG (Table S7). Unsurprisingly, sex-DEGs were also enriched for targets of gonadal hormone receptors. Consistent with rodent literature demonstrating the roles of androgens in sex-differential development of VMH populations ¹⁰ and in adult HYP sex-DE, ⁹² our vARC and vVMH sex-DE signatures were enriched for AR-regulated genes, including those induced downstream of protein kinase PKN1 (Data S5), providing a possible mechanism for the establishment and/or maintenance of sex-DE. Similarly, GSEAs implicated a role for progesterone in observed sex-DE, consistent with observations that progesterone modulates the morphology and electrophysiology of female mouse VMH neurons.⁹³ Interestingly, PGR expression itself was nominally greater in the male vVMH, while genes upregulated in the female vVMH were enriched for progesterone-responsive genes. This pattern could be due to sex differences in circulating progesterone itself, in the activity of PR co-TFs, or both.

At the domain level, transcriptome-wide sex-DE was well correlated between the vARC and vVMH, suggesting that sex similarly regulates expression in these two domains in humans. Monosynaptic circuits connecting the VMH-ARC have been traced in the rodent HYP ⁹⁴ ; if such circuits exist in humans, sex-DE concordance may be partially driven by the capture and subsequent quantification of mRNA from neurites of VMH neurons that project to the ARC. Consistent with Visium analysis, we identified robust sex-DE using Xenium at the levels of both domains (all cell types in a spatial area) and local neuron subtypes. Especially in the ARC, many sex-DEGs were shared across several xARC populations; for example, retinol dehydrogenase RDH10 was DE in 4/5 xARC neuron clusters. However, the major finding from our Xenium analyses is that sex-DE in the ARC and VMH appears to be driven predominantly by ESR1- and CRHR2-expressing neurons, respectively. Our results demonstrate that human sex differences in the tuberal HYP are specific not to its anatomic nuclei (e.g., the VMH or ARC) but to the particular cell types they contain—namely, xVMH CRHR2 and xARC ESR1 neurons.

We next explored relationships between HYP gene expression/sex-DE and genetic factors influencing BMI and neuropsychiatric disorders to ascertain possible roles for the VMH/ARC in neuropsychiatric disease risk and sex differences therein. Surprisingly, we did not find BMI-associated genes to be overrepresented in the VMH or ARC, despite the role of the HYP in appetite and metabolism. However, our method of gathering genes for this analysis yielded an improbable number of BMI-associated genes (over 8,000, i.e., a null expectation that >50% of query genes would be BMI associated; results using subsets of the ascertained genes for each phenotype are provided in Tables S11 and S7). Nonetheless, we did note sex-DE of the appetite-regulating and obesity-associated ⁵⁶ melanocortin receptor MC4R in the VMH and of CDH13—a receptor for the insulin-sensitizing hormone adiponectin—in the ARC. Interestingly, CDH13 also links to genetic risk for MDD (which occurs more often in females): CDH13 is repressed by a human-specific long non-coding RNA (lncRNA) FEDORA, itself shown to be sex-DE in the MDD cortex and to exert sex-differential behavioral effects when expressed in the mouse cortex. ^95–97 The most noteworthy finding, supported by multiple analyses, was the robust enrichment of male-upregulated genes of both the vVMH and vARC for ASD and NDD risk gene sets. Male-upregulated genes were overrepresented in targets of TFs encoded by genes mutated in syndromic ASDs/NDDs, including ZNF142, ⁸⁵ TCF20,⁹⁸ ASH1L, ⁸⁶ and MYT1L—a syndrome that includes obesity. ^67,99 Moreover, MYT1L was itself significantly male upregulated in TAC3-ESR1 KNDy neurons of the ARC and in CRHR2 neurons in the VMH. These findings highlight the VMH and ARC as candidate HYP regions susceptible (perhaps sex differentially) to common- and rare-variant-mediated neuropsychiatric disorder and NDD risk.

To support translational research efforts, we conducted cross-species analyses using available rodent datasets. ^12,50 Marker genes of the human vVMH and vARC strongly mapped to the corresponding domains in mice based on integrative studies with a previously published rodent dataset. Several marker genes, especially in the ARC, point to conserved physiologic roles implicit in the expression of well-characterized neuropeptides such as KISS1, GNRH, AGRP, and NPY. Likewise, domain-enriched genes were enriched in functional terms consistent with rodent studies of the VMH (e.g., regulation of blood glucose ^52,54) and ARC (e.g., primary cilia). Primary cilia are ubiquitous cell structures implicated in the role of the ARC in detecting nutritional status. Mice deficient for primary cilium assembly show obesity ²⁵; likewise, mutations disrupting primary cilium localization of MC4R, which is expressed in the ARC, are associated with obesity in humans.¹⁰⁰ These findings suggest that many broad-scope features of ARC/VMH biology are evolutionarily conserved between rodents and humans, with primary cilia in the ARC being a particularly strong example.

On the other hand, we observed unexpected marker, sex-DE, and GSEA findings relating to retinoid pathways in the human HYP. Despite generally shared molecular identities between the rodent and human ARC and VMH, retinoid pathway gene CYP26A1 was a unique marker to the human ARC. Additional convergent lines of evidence for retinoid activity in the human VMH and ARC included expression of both genes key to local synthesis of active retinoids (ALDH1A1 and RDH10) and enrichment among VMH- or ARC-specifically expressed genes for targets of retRs. Interestingly, however, the vARC did not display standout expression of CRABP1—a non-nuclear retinoid-binding protein capable of modulating signal transduction via kinase interactions¹⁰¹ —while its mouse ortholog was recently shown to be highly enriched in a POMC-and AGRP-negative, diet-responsive mouse ARC neuron population. ¹⁰² Similarly, divergence between mouse and rat in the regulation/expression of genes encoding retinoid synthesis enzymes and RetRs in the ARC has been observed. ¹⁰³ These findings collectively suggest that retinoid pathway gene expression and regulation are “evolutionarily plastic” in the tuberal HYP, in service of either species-specific (divergent) functions or a shared (conserved or convergent) function via distinct genes. Further molecular divergences with the rodent HYP were observed for sex hormone receptor expression and distribution. Notably, ESR1 expression in the human VMH and ARC diverged from previous findings in rodents. In mice, Esr1-expressing neurons in the VMH, especially in the ventrolateral subdivision, are a well-studied hub of rodent sex differences. ^1,3,8 By contrast, we observed uniformly sparse ESR1 expression in the human VMH (possibly due to anatomic considerations, discussed below). Notably, Esr1 expression patterns in the VMH vary even between mice and voles, suggesting that VMH Esr1 expression may track with species-specific social/parental behaviors¹⁰⁴ —which would explain our human VMH ESR1 pattern being distinct from mice and voles. ESR1 was, however, densely and highly expressed in a ventromedial pocket of the ARC. We validated both VMH and ARC components of this expression pattern using 3 complementary modalities: smFISH, Visium, and Xenium. ESR1-expressing cells in the human ARC included a subpopulation of TAC3+/KISS1+ neurons likely representing KNDy neurons previously described in the rodent ARC. ^5,105 KNDy neurons play a key role in hormone feedback and the LH/FSH cycle, ^9,28,29 so it is not surprising that the xARC population containing these neurons would show the most sex-DEGs. At the same time, it is difficult to infer what proportion of our ARC TAC3-ESR1 population comprised probable KNDy neurons, as we were not able to include KISS1 in our gene panel. While smFISH indicated it was likely a minority of ESR1 neurons, it is possible that gonadal hormones may have repressed key transcriptional markers of KNDy cells, resulting in lower KISS1 expression. However, this speculation is also based on mouse studies showing that KNDy marker genes (Kiss1 and Tac2) are repressed substantially by gonadal hormones, ^62,106,107 which we assume were circulating in our subjects.

While our findings provided clear evidence for a substructure within the human ARC, we could not detect spatial subdivisions in the human VMH. With Visium, VMH clusters were distinguished primarily by the extent of glial gene expression, and in Xenium, VMH neuronal clusters were intermingled. We also did not find clear expression of rodent VMH subpopulation marker genes, including those previously reported to be spatially localized, such as Esr1-expressing neurons in the posterior ventrolateral rodent VMH. ³ Supporting our findings, a single-nucleus RNA-seq (snRNA-seq) study of the developing human HYP through gestational week 20 (a stage at which nearly all neuronal populations have emerged) revealed only one postmitotic VMH cluster. ⁵⁹ However, certain subtle, but inconclusive, observations were consistent with descriptions of the rodent VMH as an excitatory neuronal core with a thin, cell-poor inhibitory shell. ^3,49 Transcriptomic data hinted at a shell when clustering Visium data at granular resolution (k = 31 clusters), yielding a thin, lateral vVMH domain in several samples (Figure S6, spots labeled “VMH31.2”). Compared to the other 4 vVMH clusters at k = 31, this cluster was enriched for PAQR6 (a membrane GPCR for progestins ¹⁰⁸), BCAS1, and CPLX1, with weaker but significant enrichments of GAD1 but not SLC17A6 (Table S4). Xenium analysis also identified a small, rare cell cluster (xVMH lateral border) that occupied all tissue areas except the interior of the VMH, only reaching the outermost margins of the VMH. However, the tissue-wide distribution of this cluster and the limited number of genes in Xenium make it difficult to assess whether this label was a catchall that merely included putative shell neurons. We also note that recent work in the human nucleus accumbens did not identify the clear core-shell organization described in rodents,¹⁰⁹ suggesting caution should be taken in assuming that regional core-shell organization is orthologous between rodent and human brain regions. There are several possible interpretations for the lack of a (readily evident) VMH substructure in our datasets. First, human VMH neuronal cell types may not appreciably exist in spatial niches, or such a phenomenon may only be discernible with very large, cell-resolution datasets and granular clustering. Second, human VMH cell types might be defined by different genes than rodent VMH cell types and for which spatial variability could not be detected within the human vVMH. Third, a cell-poor inhibitory shell would contribute only a small minority of cells to VMH domains at Visium resolution, such that excitatory marker gene expression still dominates the domain’s molecular profile. Finally, it is possible that molecularly defined VMH subdivisions exist in a specific anatomical plane across the anterior-posterior (AP) axis that was not captured in our study (discussed below).

In sum, these data represent a unique transcriptomic resource focused on the human tuberal HYP, an area important for endocrine, reproductive, and behavioral functions. Complementing recent human atlasing efforts in profiling a broader extent of the HYP, ^44,45 we provide a focused and spatially resolved single-cell atlas of the mediobasal HYP. We demonstrate that the molecular identities of the VMH and ARC are generally similar between mice and humans. Further, we show that autosomal gene expression differs across the sexes in both the VMH and ARC, often in a cell-type-specific manner. Finally, we show that sex differences in gene expression are associated with ASDs and NDDs, disorders with greater prevalence in males. Future studies should build upon this and other foundational molecular work in the human HYP by profiling its domains in-depth—that is, at depths similar to or greater than ours here—in order to power studies contrasting the effects of sex or physiology (BMI, reproductive status, etc.) on the HYP. To augment ongoing HYP research, we generated interactive web-based data browsers for the scientific community to explore these data at bit.ly/libdVisHYP (Visium) and bit.ly/libdXenHYP (Xenium).

Limitations of the study

We note several limitations in the present study. First, consistent with previous efforts to map the human HYP, ¹¹⁰ we only sampled one position along the AP axis per donor. We note that domain shape and size varied across donors, which we attribute to variability in cutting angle and potentially in AP sampling. We estimate that there is only a modest degree of AP sampling variation, as we screened tissue using smFISH for NR5A1, which, in adult rodents, is only expressed in anterior/middle VMH neurons, not ventrolateral sex hormone receptor-expressing neurons. ^3,111 Meanwhile, Esr1+ and Pgr+ neurons are most abundant in the posterior mouse VMHvl³ ; thus, sparse ESR1 expression in our VMH samples (Figure S27A) may be consistent with our data being biased toward the NR5A1+ anterior VMH. (Interestingly, PGR expression was diffuse and about 2-fold that of ESR1 in the vVMH [Figure S27A].) Finally, despite the known presence ^112,113 and important function in lactation ¹¹⁴ of dopamine neurons in the ARC, this population was not identified in our data due to insufficient dopaminergic marker expression. Furthermore, while heterogeneous gene expression within cell types along the AP axis has been observed in other areas of the rodent HYP, this phenomenon only involved a few genes. ¹¹⁵ Recent molecular atlasing of the human HYP identified several VMH clusters from snRNA-seq, all of which were represented in a single AP plane. ⁴⁵ This indicates that AP sampling variation should not affect our ability to detect (non-ESR1) VMH populations and is consistent with the identification of cells from all xVMH/xARC clusters in all Xenium samples. Focusing our Xenium sex-DE analyses on the cell clusters or “types” found across the entire sample set, we found sex differences far more widespread than could be supported by evidence of intra-regional AP expression gradients alone.

Second, there are methodological limitations to both spatial transcriptomic approaches that we employed. While Visium is transcriptome-wide, it is lower resolution and prohibits granular spatial-molecular organization of small tissue areas such as the VMH and ARC. We addressed this limitation by employing Xenium, which resolves single nuclei in space but at the expense of measuring only a fraction of the transcriptome. Our Xenium data cover only 366 genes, potentially preventing the identification of cell types, including tanycytes and cells of the VMH. In the former case, tanycyte labels were assigned to rare cells in the median eminence bordering the VMH/ARC, as would be expected, ¹¹⁶ but also to cells intermingled with connective and vascular clusters in many samples (Figure S31), illustrating the limitations of cell-resolution spatial transcriptomics as a method for cell-type discovery. Additionally, imaging-based methods such as Xenium are inherently limited in their ability to segment boundaries of complex multipolar cells with complex projections (like most neural cell types). To ensure that inherently imperfect cell segmentation was not the primary driver of our findings, we also performed marker and sex-DE analyses using only transcripts that were detected within a nucleus—that is, transcript sets that could be confidently attributed to individual cells (Tables S3 and S10).

Our findings regarding sex-DE, ESR1 expression, and features of putative ARC KNDy neurons could be potentially affected by menstrual stage and/or oral contraceptive use by female donors, though this information was neither systematically collected nor necessarily accurate (only available when reported by next of kin). For example, while we find MC4R to be male upregulated in the VMH of our cohort, it has been previously reported to be female upregulated in mice in an estrus-stage-dependent manner. ⁵⁷ Transient effects of estrogens/progestins are less likely to impact our findings on ESR1 distribution, as these patterns were consistent between sexes, and our smFISH experiment for ESR1 and KISS1 was performed on tissue from a male donor. Finally, given the key role of KNDy neurons in the female reproductive cycle, the molecular state of these cells presumably differs across the cycle. Extra caution is therefore needed in interpreting xARC ESR1 sex-DE findings, as these are especially likely to be driven by transient/stage-specific phenomena.

RESOURCE AVAILABILITY

Lead contact

Inquiries regarding the study or the resources provided should be directed to and will be fulfilled by the lead contact, Kristen R. Maynard (kristen.maynard@libd.org).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• FASTQ files from Visium (sequencing-based) spatial transcriptomics have been deposited at GEO as GEO: GSE280316 (see key resources table). Unfiltered transcript-level Xenium (imaging-based) spatial transcriptomic outputs, compatible with the 10× Xenium Explorer software, have been deposited at GEO as GEO: GSE280460 (see key resources table). Data can be interactively visualized and annotated in a web browser through our Samui ¹¹⁷ portals for the Visium data (with clustering available at four different resolutions from BayesSpace) at bit.ly/libdVisHYP (permanent URL: https://research.libd.org/spatial_HYP/) and for Xenium (with annotated cell clusters from Banksy, their higher-order cell groups, and domains) at bit.ly/libdXenHYP (permanent URL: https://research.libd.org/spatial_HYP/). Documentation for the data portals can be accessed through https://research.libd.org/spatial_HYP/. This paper analyzes existing, publicly available datasets: raw h5ad files with single-cell RNA-seq data from the adult mouse HYP¹² were downloaded from https://allen-brain-cell-atlas.s3-us-west-2.amazonaws.com/expression_matrices/WMB-10Xv3/20230630/WMB10Xv3-HPF-raw.h5ad and https://allen-brain-cell-atlas.s3-us-west-2.amazonaws.com/expression_matrices/WMB-10Xv2/20230630/WMB10Xv2-HY-raw.h5ad and then were filtered to cells that received cluster labels found in https://allen-brain-cell-atlas.s3-us-west-2.amazonaws.com/metadata/WMB-10X/20230630/views/cell_metadata_with_cluster_annotation.csv; TF-target gene sets (Transcription_Factor_PPIs, ChEA_2022, Rummagene_transcription_factors, TF_Perturbations_Followed_by_Expression, TF-LOF_Expression_from_GEO, and TRRUST_Transcription_Factors_2019) were obtained from the gene set libraries used by the Enrichr tool¹¹⁸ at https://maayanlab.cloud/Enrichr/#libraries on 12/06/2023. Our assertion for low expression of TF RNAs is based on an interactive plot from Lambert et al.¹¹⁹ illustrating the transcripts per million for each likely human TF in each tissue in data from the Human Tissue Atlas: http://humantfs.ccbr.utoronto.ca/.

• All original code has been deposited to GitHub (https://github.com/LieberInstitute/spatial_HYP), with contents at the time of submission archived through Zenodo (https://zenodo.org/records/14285059). Non-Bioconductor/CRAN R Packages: scCoco and the Allen Brain Atlas API interface it utilizes, cocoframer, are respectively available at https://github.com/lsteuernagel/scCoco and https://github.com/AllenInstitute/cocoframer.

• Supplemental data files; smFISH microscopy images; the Visium SpatialExperiment (including bundled H&E images) and Xenium SpatialFeatureExperiment objects as filtered, clustered, and analyzed; and the Xenium H&E images (not used in the study) are available through http://research.libd.org/globus/ via the Globus endpoint jhpce#HYP_suppdata. They are publicly available as of the date of publication.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Human postmortem brain hypothalamus tissue	Lieber Institute for Brain Development: See Table S1	N/A
Chemicals, peptides, and recombinant proteins
O.C.T Compound	Tissue-Tek Sakura	product code 4583
10% Neutral Buffered Formalin	Sigma-Aldrich	catalog no. HT501128
Opal Fluorophore 520	Akoya Biosciences	catalog no. FP1487001KT
Opal Fluorophore 570	Akoya Biosciences	catalog no. FP1488001KT
Opal Fluorophore 620	Akoya Biosciences	catalog no. FP1495001KT
Opal Fluorophore 690	Akoya Biosciences	catalog no. FP1497001KT
Fluromount-G	SouthernBiotech	catalog no. 0100-01
Critical commercial assays
Visium Spatial Gene Expression Slides	10X Genomics	Part Number 2000233
RNAscope Multiplex Fluorescent Kit, v2	Advanced Cell Diagnostics	catalog no. 323100
RNAscope 4-Plex Ancillary Kit	Advanced Cell Diagnostics	catalog no. 323120
RNAscope Probe- Hs-NR5A1-C1	Advanced Cell Diagnostics	catalog no. 553151
RNAscope Probe- Hs-OXT-C2	Advanced Cell Diagnostics	catalog no. 538811-C2
RNAscope Probe- Hs-NPY-C3	Advanced Cell Diagnostics	catalog no. 416671-C3
RNAscope Probe- Hs-MBP-C4	Advanced Cell Diagnostics	catalog no. 411051-C4
Xenium In Situ Gene Expression Slides & Reagent Kits	10X Genomics	Part Number 1000460
Xenium Human Brain Gene Expression Panel	10X Genomics	Part No.1000599
Xenium Custom Gene Expression Panel	10X Genomics	Part No. 1000561
RNAscope Probe- Hs-LAMP5-C2	Advanced Cell Diagnostics	catalog no. 487691-C2
RNAscope Probe- Hs-GAD1-01-C3	Advanced Cell Diagnostics	catalog no. 573061-C3
RNAscope Probe- Hs-GAD2-ver2-C3	Advanced Cell Diagnostics	catalog no. 415691-C3
RNAscope Probe- Hs-SLC17A6-C4	Advanced Cell Diagnostics	catalog no. 415671-C4
RNAscope Probe- Hs-KISS1-C1	Advanced Cell Diagnostics	catalog no. 507981
RNAscope Probe- Hs-ESR1-C2	Advanced Cell Diagnostics	catalog no. 310301-C2
RNAscope Probe- Hs-TAC3-C3	Advanced Cell Diagnostics	catalog no. 507301-C3
Software and algorithms
HALO	Indica Labs	https://indicalab.com/halo/
VistoSeg	Tippani et al. 120	https://doi.org/10.1017/S2633903X23000235
Loupe Browser	10X Genomics	https://www.10xgenomics.com/support/software/loupe-browser/downloads
Spaceranger v. 2.1.0	10X Genomics	https://www.10xgenomics.com/support/software/space-ranger/downloads/previous-versions
scCoco 0.0.0.900	R Package	Specific Github commit installed: lsteuernagel/scCoco@efffc50
cocoframer 0.1.1	R Package	Specific Github commit installed: AllenInstitute/cocoframer@1de30a8
R Programming Language 4.3–4.4 (see reproducibility info at the end of analysis scripts in this paper’s github/zenodo)	R Core Team	https://www.r-project.org
orthogene 1.10.1	R Package	https://www.bioconductor.org/packages/release/bioc/html/orthogene.html
BayesSpace 1.12.0	R Package	https://www.bioconductor.org/packages/release/bioc/html/BayesSpace.html
scran 1.30.2	R Package	https://www.bioconductor.org/packages/release/bioc/html/scran.html
scuttle 1.12.0	R Package	https://www.bioconductor.org/packages/release/bioc/html/scuttle.html
SpatialExperiment 1.12.0	R Package	https://www.bioconductor.org/packages/release/bioc/html/SpatialExperiment.html
data.table 1.15.4–1.17.4 (see reproducibility info at the end of analysis scripts in this paper’s github/zenodo)	R Package	https://github.com/Rdatatable/data.table
clusterProfiler 4.12.0	R Package	https://www.bioconductor.org/packages/release/bioc/html/clusterProfiler.html
nnSVG 1.8.0	R Package	https://www.bioconductor.org/packages/release/bioc/html/nnSVG.html
spatialLIBD 1.16.2	R Package	https://www.bioconductor.org/packages/release/bioc/html/spatialLIBD.html
escheR 1.2.0	R Package	https://www.bioconductor.org/packages/release/bioc/html/escheR.html
dreamlet 1.2.1	R Package	https://www.bioconductor.org/packages/release/bioc/html/dreamlet.html
limma 3.60.3	R Package	https://www.bioconductor.org/packages/release/bioc/html/limma.html
edgeR 4.2.0	R Package	https://www.bioconductor.org/packages/release/bioc/html/edgeR.html
fgsea 1.30.0	R Package	https://www.bioconductor.org/packages/release/bioc/html/fgsea.html
msigdbr 7.5.1	R Package	https://cran.r-project.org/web/packages/msigdbr/index.html
igraph 2.0.3	R Package	https://cran.r-project.org/web/packages/igraph/index.html
harmony 1.2.0	R Package	https://cran.r-project.org/web/packages/harmony/index.html
SpatialFeatureExperiment 1.6.1	R Package	https://www.bioconductor.org/packages/release/bioc/html/SpatialFeatureExperiment.html
Banksy 0.99.12	R Package	https://www.bioconductor.org/packages/release/bioc/html/Banksy.html
MoleculeExperiment 1.4.1	R Package	https://www.bioconductor.org/packages/release/bioc/html/MoleculeExperiment.html
sf 1.0-16	R Package	https://cran.r-project.org/web/packages/sf/index.html
Deposited data
Raw RNA-seq data (Visium)	This paper	GEO: GSE280316
Raw Xenium data	This paper	GEO: GSE280460
Supplemental Data files, smFISH microscopy images, the Visium SpatialExperiment (including bundled H&E images) and Xenium SpatialFeatureExperiment objects as filtered, clustered, and analyzed, and Xenium hematoxylin and eosin images (not used in the study)	This paper	http://research.libd.org/globus (endpoint jhpce#HYP_suppdata)
Interactive data browser (Visium)	This paper	https://bit.ly/libdVisHYP, research.libd.org/spatial_HYP, or https://samuibrowser.com/from?url=data.libd.org/samuibrowser/&s=V12D05-348_C1&s=V12D05-348_D1&s=V12D05-350_C1&s=V12D05-350_D1&s=V12D07-075_A1&s=V12D07-075_D1&s=V12Y31-080_A1&s=V13M13-362_A1&s=V13M13-362_D1&s=V13Y24-346_C1 (direct/permanent URL)
Interactive data browser (Xenium)	This paper	https://bit.ly/libdXenHYP, research.libd.org/spatial_HYP, or https://samuibrowser.com/from?url=data.libd.org/samuibrowser/&s=X36_5459A&s=X36_5459B&s=X36_8667C&s=X86_reg1&s=X86_reg2&s=X86_reg3&s=X97_reg1&s=X97_reg2&s=X97_reg3&s=X99_1225A&s=X99_1225B&s=X99_8741C&s=X99_8741D (direct/permanent URL)
Analysis and plotting code	This paper	https://github.com/LieberInstitute/spatial_HYP and, as frozen at time of final submission, Zenodo: https://zenodo.org/records/14285059
Other
Mouse hypothalamus single-cell RNAseq	https://doi.org/10.1038/s41586-023-06812-z	https://allen-brain-cell-atlas.s3-us-west-2.amazonaws.com/expression_matrices/WMB10Xv3/20230630/WMB-10Xv3-HPF-raw.h5ad, https://allen-brain-cell-atlas.s3-us-west-2.amazonaws.com/expression_matrices/WMB10Xv2/20230630/WMB-10Xv2-HY-raw.h5ad, and https://allen-brain-cell-atlas.s3-us-west-2.amazonaws.com/metadata/WMB-10X/20230630/views/cell_metadata_with_cluster_annotation.csv
TF-target libraries from Enrichr	https://doi.org/10.1093/nar/gkz446, https://doi.org/10.1093/nar/gkx1013, https://doi.org/10.1038/s42003-024-06177-7	https://maayanlab.cloud/Enrichr/#libraries (Transcription_Factor_PPIs, ChEA_2022, Rummagene_transcription_factors, TF_Perturbations_Followed_by_Expression, TF-LOF_Expression_from_GEO, TRRUST_Transcription_Factors_2019)
Interactive human tissue TF expression visualization	https://doi.org/10.1016/j.cell.2018.01.029	http://humantfs.ccbr.utoronto.ca/

STAR★METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Postmortem human brain tissue from 8 adult neurotypical donors of European ancestry was obtained at the time of autopsy following informed consent from legal next of kin, Office of the Chief Medical Examiner of the State of Maryland, under the Maryland Department of Health IRB protocol #12–24, from the Western Michigan University Homer Stryker MD School of Medicine, Department of Pathology under the WCG protocol #20111080, and from the NIMH Human Brain Collection Core, protocol #90-M-0142. Using a standardized strategy, all donors were subjected to clinical characterization and diagnosis. Macro- and microscopic neuropathological examinations were performed, and subjects with evidence of neuritic pathology or other neuropathological abnormalities were excluded. Details regarding tissue acquisition, processing, dissection, clinical characterization, diagnosis, neuropathological examination, RNA extraction and quality control (QC) measures have been previously published. ¹²¹ Demographics for the 8 neurotypical control donors included in the study are listed in Table S1. All subjects had BMIs between normal range and class I obesity (22.3–31.6; 7 of 8 donors non-obese with BMI <29) and no history of neuroendocrine disorders. Female subjects were premenopausal and not pregnant at time of death. Information on donor use of oral contraceptives was not systematically collected. Briefly, fresh frozen coronal brain slabs with clearly visible thalamus, hypothalamus, putamen, globus pallidus (internal and external segments), optic tract, anterior commissure, fornix, and posterior amygdala (MNI: − 7.04) were selected for VMH and ARC dissection. Using a hand-held dental drill, tissue blocks of approximately 10 × 20 mm were dissected, encompassing the hypothalamus, optic tract, anterior commissure and fornix. White matter (WM) tracts were used as landmarks to identify the approximate location of VMH and ARC for microdissection. Tissue blocks were stored in sealed cryogenic bags at − 80°C until cryosectioning.

METHOD DETAILS

Tissue processing and quality control

At the time of cryosectioning, tissue blocks were acclimated to the cryostat (Leica CM3050s) at − 14°C and mounted onto a chuck with OCT (TissueTek Sakura). Approximately 50 μm of tissue was trimmed from the block to achieve a flat surface, and several 10 μm sections were mounted onto pre-chilled glass slides for anatomical validation quality control (QC) experiments using single molecule fluorescent in situ hybridization (smFISH) with RNAScope technology (Advanced Cell Diagnostics [ACD]). RNAscope was performed as described below to verify the presence of neuroanatomical landmarks with probes specific for the VMH (Hs-NR5A1-C1, catalog no. 553151), paraventricular nucleus/supraoptic nucleus ([PVN/SON]; Hs-OXT-C2, catalog no. 538811), ARC (Hs-NPY-C3, catalog no. 416671), and WM tracts (Hs-MBP-C4, catalog no. 411051). Successive rounds of RNAscope were completed until the anatomical plane containing VMH and ARC was reached (regardless of the presence of PVN/SON). This location approximately corresponds to the area between the optic tract (as the ventro-lateral landmark) and the fornix (as the dorsal landmark). Blocks were then scored with a razor blade to isolate the VMH and ARC nuclei in ∼6.5 × 6.5 mm squares for the Visium assay. Adjacent 10 μm tissue sections were mounted onto pre-chilled Visium Spatial Gene Expression slides (part number 2000233, 10x Genomics) (Figure S3) and onto Superfrost plus glass slides, which were banked for additional RNAscope validation experiments. In some cases, tissue sections from the same block were collected onto multiple Visium arrays to ensure inclusion of target structures. In these cases, distance along the anterior-posterior axis between experiments did not exceed 100μm. Following Visium analysis, tissue blocks were remounted for cryosectioning, approximately 50 μm of tissue was trimmed off to achieve a flat surface, and several 10 μm sections were mounted onto pre-chilled glass slides and banked for validation studies. Blocks were then re-scored with a razor blade to isolate the VMH and ARC nuclei in ∼12 × 6.5 mm strips and mounted onto Xenium slides (Part Number 1000460, 10x Genomics). 3–4 tissue strips were fitted onto each Xenium array.

Visium data generation

Visium Spatial Gene Expression slides were processed according to the manufacturer’s protocol (protocol number CG000239, Rev G, 10X Genomics) as previously described. ^{122, 123} The tissue permeabilization step was performed for 18 min based on tissue optimization experiments in other brain regions. ^123,124 For each Visium slide, H&E staining was performed and the slide was coverslipped. Tissue sectioning and mounting to slide, sequencing library preparation, sequencing, H&E staining, and microscopy were performed across two sites (listed for each step per sample in Table S1). Briefly, microscopy images were acquired either at 40x on a Leica CS2 slide scanner equipped with a 20x/0.75NA objective and a 2x doubler, or on an Olympus VS200 slide scanner equipped with a 10x/0.4NA objective. Following removal of the coverslip, tissue was permeabilized, cDNA synthesized, and sequencing libraries generated. Libraries were loaded at 300 pM on a NovaSeq6000 System (Illumina) at the sites listed in Table S1. Sequencing was performed according to manufacturer’s instructions, targeting ∼60K reads per spot.

Xenium data generation

Xenium slides were fixed and permeabilized according to 10X Genomics’ Xenium In Situ for Fresh Frozen Tissues – Fixation & Permeabilization Protocol (CG000581, Rev C). Briefly, slides were warmed for 1 min prior to submersion in 3.7% formaldehyde for 30 min. Slides were then permeabilized with 1% sodium dodecyl sulfate solution (SDS) for 2 min and rinsed with PBS. This was followed by a 1 h incubation in 70% methanol on ice. Following PBS washes, slides were placed in their respective cassettes and incubated with 0.05% PBS-T. For probe hybridization, ligation, and amplification, the 10X Genomics protocol: Xenium In Situ Gene Expression was followed (CG000582, Rev E). The probe hybridization solution was then prepared. We used a combination of the off-the shelf “human brain,” probe panel designed by 10X Genomics (Xenium Human Brain Gene Expression Panel, Part No.1000599) and our custom probe panel (10X Genomics; Xenium Custom Gene Expression Panel, Part No. 1000561). Details of the custom probe panel design are described in section 3.8.1. Probe mix was applied to each slide, a cassette lid was applied to each slide cassette and the slides were incubated at 50°C overnight.

The next day, the slides were washed twice in 1X PBS-T followed by an incubated wash in post hybridization wash buffer (Xenium Post Hybridization Wash Buffer, Part No. 2000395, 10X Genomics). Three washes of 1X PBS-T were completed before addition of ligation mix to each slide.Slides were then incubated at 37°C for 2 h. After ligation, the slides were washed three times in 1X PBS-T, and then incubated in an amplification master mix at 30°C for 2 h. After amplification, slides were washed three times in TE buffer followed by three washes in 1X PBS and then incubated in diluted reducing reagent B for 10 min at room temperature. Slides were then washed in 70% ethanol followed by two washes of 100% ethanol. Autofluorescence solution was then added to each slide and incubated in the dark for 10 min at room temperature. Slides were then rinsed three times with 100% ethanol before drying on a 37°C preheated thermocycler for 5 min. Additional washes were performed before adding Xenium Nuclei Staining Buffer, which incubated for 1 min in the dark at room temperature. Slides were then washed three times in 1X PBS-T and stored at 4°C prior to being loaded on the Xenium analyzer at the Johns Hopkins Single Cell & Transcriptomics Core (SCTC). The solutions needed to run the Xenium analyzer were prepared according to the manufacturer’s instructions. Region selection was guided by our team with instrument operation performed by the SCTC Core.

Upon completion of the Xenium analyzer run, post-xenium H&E was completed following 10X Genomics Xenium In Situ Gene Expression - Post-Xenium Analyzer H&E Staining protocol (CG000613, Rev A) with the following modifications: reduced staining time in hematoxylin, from 20 min to 17 min, and reduced eosin staining time, from 5 min to 2 min. Briefly, slides were quenched in a 10mM sodium hydrosulfite solution for 10 min, slides were then washed four times in Milli-Q water before being stained with Mayer’s Hematoxylin for 17 min (Sodium hydrosulfite, Cat No. 157953–5G, Sigma Aldrich; Hematoxylin Solution, Mayer’s, Cat No. MHS16, Sigma Aldrich). After nuclear staining was completed, slides were rinsed three times in Milli-Q water before submerging into bluing buffer (Bluing Solution, Cat No. CS702, Dako). Slides were washed once in Milli-Q water, once in 70% ethanol, and once in 95% ethanol before being introduced to alcoholic eosin (Eosin Y Solution, Alcoholic, Cat No. 3801615, Leica). After cytoplasmic staining was completed, the slides were rinsed twice in 95% ethanol and twice in 100% ethanol before being cleared in two changes of xylene (Xylene, Reagent Grade, Cat. No. 214736, Millipore Sigma). Slides were then coverslipped using Vectamount mounting media and allowed to dry overnight before imaging (VectaMount Permanent Mounting Medium, Cat No. H-5000, Vector Laboratories Inc.). Imaging was completed at 20X and 40X using a Leica Aperio CS2 digital pathology slide scanner. While H&E images were not utilized in the current study, they have been made available with other supplemental data through a Globus endpoint (see resource availability).

Data was obtained from two slides each in two separate runs. The first two slides of Run 1 were acquired and post-processed by the Johns Hopkins SCTC using instrument firmware/xeniumranger version 1.5. The second two slides of Run 2 assayed at SCTC were acquired and post-processed with instrument firmware/xeniumranger 1.9. Due to changes in cell segmentation algorithms between these two Xenium versions, we uniformly re-postprocessed all data by using xenium-resegment –with-nuclei = TRUE in xeniumranger 1.7.

Single molecule fluorescence in situ hybridization (smFISH)

smFISH assays were performed with RNAscope technology utilizing the RNAscope Multiplex Fluorescent Reagent Kit v2 (catalog no. 323100; Advanced Cell Diagnostics [ACD]) in reference to the manufacturer’s protocol (document no. UM323100, rev B; ACD) with the following RNAscope probe sets: Anatomical quality control: Hs-NR5A1 (assigned Opal dye 570), Hs-OXT (assigned Opal dye 620), Hs-NPY (assigned Opal dye 690), Hs-MBP (assigned Opal dye 520) (catalog no. 553151, 538811-C2, 416671-C3, 411051-C4, respectively; ACD); Experiment 1: Hs-NR5A1 (assigned Opal dye 570), Hs-LAMP5 (assigned Opal dye 620), Hs-GAD1-01 or Hs-GAD2-ver2 (assigned Opal dye 690), Hs-SLC17A6 (assigned Opal dye 520) (catalog no. 553151, 487691-C2, 573061-C3 or 415691-C3, 415671-C4, respectively; ACD); Experiment 2: Hs-KISS1 (assigned Opal dye 520), Hs-ESR1 (assigned Opal dye 570), Hs-TAC3 (assigned Opal dye 690) (catalog no. 507981, 310301-C2, 507301-C3, respectively; ACD). Specifically, 10-μm tissue sections were fixed in 10% neutral buffered formalin for 30 min at room temperature (RT). After 2 PBS washes, the samples were sequentially dehydrated in 50%, 70%, 100%, and 100% ethanol solutions each for 5 min at RT, treated with hydrogen peroxide for 10 min at RT, washed 2 times again with PBS, and treated with protease IV for 30 min at RT. After another 2 PBS washes, sections were incubated in a solution of the four target probes for 2 h at 40°C, washed 2 times in wash buffer (catalog no. 310091; ACD), and stored in SSC for up to 48 h at 4°C. Afterward, the sections underwent amplification with AMP1 through AMP3 for 30, 30, and 15 min, respectively, at 40°C with 2 washes with wash buffer performed before each new reagent was added. In sequence, the sections underwent 4 cycles of 15-min incubation with HRP-Cx, 30-min incubation with a fluorescent Opal dye solution, and 15-min incubation with HRP blocker, all at 40° C and with 2 washes with wash buffer in between. Each cycle used a different fluorescent Opal dye (520, 570, 620, or 690; catalog no. FP1487001KT, FP1488001KT, FP1495001KT, and FP1497001KT, respectively; Akoya Biosciences) diluted 1:500 in TSA buffer. Finally, the tissue sections were treated with DAPI for 20 s at RT, decanted, and then coverslipped with Fluoromount G mounting medium (catalog no. 0100–01; Southern Biotechnology). Coverslipped slides were stored in the dark for at least 24 h before imaging. Slides were imaged as z-stacks (at least 10 per image) using a Nikon AXR confocal microscope system powered by the NIS-Elements imaging software with a Nikon APO lambda D 20x/0.80 objective and/or Nikon APO lambda S 40x/1.25 objective. For anatomical quality control, slides were imaged in a single plane with a Nikon Plan APO lambda D 2x/0.1 objective. All confocal images at the resolutions presented/analyzed are available with other supplemental data files through a Globus endpoint (see resource availability).

Visium H&E image segmentation and raw data processing

Multi-sample slide images acquired on the Leica CS2 were first processed using VistoSeg software, ¹²⁰ which divides the whole-slide Visium images into individual capture area images using the splitSlide function. The Olympus images were of single capture areas and did not require such processing. Images for each capture area were aligned in Loupe Browser (10x Genomics) with the gene expression spots captured. Slides were then processed using the Spaceranger v. 2.1.0 software from 10X Genomics, yielding feature counts for each spatial location (“spot”) per sample using the.json output from Loupe browser, the sample image, and FASTQ sequencing files.

QUANTIFICATION AND STATISTICAL ANALYSIS

Visium: General

All analyses were performed using R 4.4.1–4.4.2 unless otherwise noted. Analysis was performed within the SpatialExperiment R framework, ¹²⁵ coercing outside datasets in other formats to this framework where necessary using zellkonverter. ¹²⁶ Spatial expression and clustering plots were generated using SpatialLIBD ¹²⁷ or escheR. ¹²⁸ ggplot2 was used to generate heatmaps and scatter/volcano plots unless noted. Extended color palettes were generated using the package Polychrome. ¹²⁹

Visium: QC and normalization

Fraction of mitochondrial reads and number of unique genes identified were calculated for each spot. Plots of these values were manually examined for spatial patterning in each sample to ensure selection of filter thresholds that would not remove putative biological information (for example, WM tracts would be expected to have low gene diversity in a contiguous area, a matter of biological rather than technical variance). Using SpotSweeper, ¹³⁰ we first detected and removed high-end outliers relative to surrounding spots in terms of mitochondrial read percentage, low-end outliers for number of UMIs, or low-end outliers for number of unique genes. Subsequently, we filtered remaining spots for high mitochondrial percentage, low UMI count, and low unique gene count. We noted during exploratory analyses that the highest mitochondrial read fractions were largely in spots later identified as VMH—containing the only excitatory neurons of note in the samples—and spots along its lateral edge. We thus set a lenient threshold of 50% for mitochondrial read content in the final analysis. We plotted histograms of UMI and gene count distributions across all remaining spots and identified a clear drop-off in spot frequencies within the histograms at approximately ≤210 UMIs and ≤126 unique genes, suggesting the boundary between technical artifact and biologically low content. Indeed, such spots were almost exclusively localized to sample edges, and these thresholds were therefore used to remove low-content spots. For the retained spots, library sizes (number of reads) per spot were calculated, normalized, and log2 transformed ¹³¹ across all spots in all samples to get expression values per gene per spot.

Of 4,992 Visium spots possible in a sample, tissue occupied 4,621 ± 341 (mean ± standard deviation) spots. Spots with high mitochondrial read proportion (≥50%) were removed to preserve a potentially biological pattern of high mitochondrial counts adjacent to VMH (Figure S1C). Subsequent filtering for low unique molecular identifiers (UMIs) and/or gene diversity (≤210 UMIs, ≤126 unique genes) resulted in 4,508 ± 363 spots for downstream analysis. The filtered data included 6,059 ± 5057 UMIs/spot (mean ± SD) spanning 2,536 ± 1470 unique genes/spot, covering 45,074 spots and 30,361 unique genes altogether. Additional pre- and post-filtering metrics are included in Table S1.

Visium: Feature selection

We used the filtered data to identify highly-variable genes (HVGs, spatially-naive) and spatially variable genes (SVGs) using the packages scran¹³² and nnSVG, ⁴⁷ respectively. For HVGs, we used sample as the blocking factor and collected the top 10%ile and 20%ile of resultant genes using scran::getTopHVGs(…, prop = 0.1 or 0.2)¹³² (Table S2).

In the nnSVG approach, each sample is analyzed independently and genes are pre-filtered. Here, we filtered to genes with at least 3 counts in at least 0.5% of spots in a sample. We subsequently recalculated the logcounts for use within nnSVG to account for dropped genes. We extracted and concatenated the nominally significant SVGs from each sample (Table S2) and determined the experiment-wide SVGs by averaging their ranks across samples where the gene achieved nominal significance. We then filtered the union of samplewise SVGs to those that achieved nominal significance in three or more samples. These features were sorted in ascending order of mean rank (among samples where the feature achieved nominal significance for spatial variation) to collect the same number of features as contained in 10%ile and 20%ile HVG sets.

Visium: Selecting the number of clusters

In a preliminary subset of samples, we used each of the four feature sets described above for performing dimensionality reduction concurrent with batch correction using harmony with sample ID as the grouping variable. ⁴⁸ The default output of 50 reduced dimensions were appended to the SpatialExperiment object for downstream use separately for each feature set.

To determine an expected number of spatial domains, we turned to the Allen Brain Atlas for adult human ¹³³ and tabulated the number of unique anatomic areas denoted along the entire anterior-posterior span of the VMH, yielding 14 potential anatomically-defined domains. We assumed WM would comprise a single, homogeneous domain and added this to get a grand total of 15 potential anatomic areas: dorsal, central, and ventral VMH; ARC; periventricular nucleus, tuberal part; lateral hypothalamic area, tuberal part; posterior hypothalamic nucleus; dorsomedial nucleus; tuberomamillary nucleus; lateral tuberal nuclei; paraventricular nucleus; magnocellular nucleus of the lateral hypothalamic area; supraoptic nucleus; accessory secretory cells of lateral hypothalamus; and generic WM.

In this same preliminary subset of data, we also performed spatially-agnostic, weighted nearest-neighbors clustering using buildSNNgraph from scran. We utilized harmony batch-corrected principal components calculated on the 10%ile or 20%ile highly variable gene sets, and constructed two SNN graphs for each feature set, using k = 5 or k = 10 nearest neighbors, for a total of four initial neighbor graph spaces. From each of these four neighbor graphs, we divided the tree into 4, 5 … 47 clusters using igraph: cut_at and aggregated these assignments. We determined the optimal, spatially-agnostic clustering resolutions for maximal within-cluster concordance relative to between-cluster concordance using the fasthplus method. ^123,134 The values returned as seg_psi_est, seg2_psi1_est, and seg2_psi2_est represent the three estimated clustering values most optimal for a given graph. These 12 estimates included two values rounding to 9, three in the range 14.6–15.6, three ranging 17.0–19.8, and a maximum value of 30.8 leading us to select the recurrent minimum (9), the value consistent with anatomic expectations (15), the upper end of this range (20), and the highest value (31) as the number of clusters to generate in the final data.

We then utilized BayesSpace ¹³⁵ with parameters of k = 15 and 20,000 iterations for spatially-aware clustering of the harmony reduced dimensions for each of the four variable-feature sets. The resulting domain assignments from each feature set were comparable when plotted onto the spatial data and plots of the top principal component; however, UMAP plots labeled by domain assignment revealed that the majority of domains separated into multiple sectors of UMAP space, except when using the “10 %ile” nnSVG set. We therefore used this feature set and its derived spatial domains for all subsequent analysis.

Visium: Dimensionality reduction and clustering

For the analyses of the full dataset presented here, we only utilized an equivalently-derived “10 %ile” nnSVG feature set for dimensionality reduction by principal components analysis (PCA), UMAP, and sample-level batch correction using harmony (Data S1) (see Code Availability) with initialization of the lambda parameter—which alters the strength of correction—from the data itself, rather than requiring a user-defined initial value (an approach which we term “lambda null” or “lambda NA” throughout the methods and in certain Supplemental files). We used the harmony batch correction as input dimensions for clustering in BayesSpace at resolutions of k = 15, 20, and 31. All results presented reflect harmony BayesSpace clusters computed with k = 15 and 60,000 iterations. We produced additional clustering assignments with the nnSVG “10th %ile” feature set using reduced dimensions from MNN batch correction ¹³⁶ or harmony with its default lambda value of 1 as inputs to 60,000-iteration BayesSpace runs at k = 15, 20, and 31 (see Code Availability). Marker, sex-DE, and GSEA results are provided for clusters defined after using the development version of harmony (“lambda null”-enabled) and BayesSpace k = 15 (individual clusters), k = 15 (VMH/ARC clusters joined into single clusters), k = 20 and k = 31 (see resource availability and supplemental information). Since spots can contain multiple cells, we emphasize that we subsequently call these clusters “domains” to emphasize that they do not necessarily correspond to single cell types or subtypes as would be identified in single-cell or -nucleus approaches.

Visium: Marker gene identification

Marker genes were comprehensively determined using several complementary statistical approaches. We primarily utilized the registration_wrapper function in spatialLIBD ¹²⁷ to calculate t test statistics comparing log counts in each domain vs all others (this method being most analogous to the logFC comparisons assessed using Cohen’s d in scran). We performed select analyses using the scoreMarkers function from scran ¹³² to identify marker genes for select pairs of domains (e.g., to compare VMH subdomains) using multiple statistics: a Cohen’s d statistic for mean log fold-change (logFC) between compared groups; an AUC statistic, indexing the probability that a gene in a random spot drawn from domain/group 1 is greater than in a random spot drawn from domain/group 2; and the logFCs calculated when only considering spots from each domain/group where a given gene was detected. We focus on SpatialLIBD t test statistics in the main text for both global marker identification for comparisons between VMH or ARC subdomains. Putative VMH/ARC subdomains from the additional clustering approaches were identified based on marker genes and referred to throughout the supplement as (VMHorARC)(k).(digit), where each subdomain is assigned a sequential number arbitrarily (e.g., “VMH20.1” and “VMH20.2” are two subdomains of VMH identified from k = 20 BayesSpace analysis). To discern genes differentiating among VMH or ARC clusters from Visium, we additionally performed ‘one-vs-all-other’ analyses, wherein each cluster from a domain was separately compared to other clusters only from that domain by subsetting the data only to spots with VMH or ARC cluster labels before performing marker analysis using spatialLIBD as above. For k = 15 or 20, this amounted to a single pairwise comparison for each domain, as VMH and ARC each spanned two clusters at both of these clustering resolutions.

Visium: Spatially variable gene identification within VMH and ARC

To further explore whether we could identify spatial patterns of gene expression within the VMH and ARC specifically, we subset the data to those spots assigned a VMH or ARC domain as above and re-ran nnSVG on individual samples, only considering spots within ARC or within VMH. One sample was excluded from this analysis for VMH due to an impractically small number of VMH spots (sample v5459A_M, 330 spots) for SVG analysis. One sample was likewise excluded from ARC analysis for the same reason (sample v8667A_F, 207 spots). The threshold for gene inclusion for SVG analysis in a given sample was ≥1 count in at least 5% of spots within the domain being analyzed. The ranks of nominally significant genes per sample were tabulated and aggregated to their means across those samples where the gene achieved nominal significance, as was done for the feature selection steps described above. As the collective structure of VMH and ARC domains was similar at all clustering resolutions, we only performed these SVG analyses on the domains as defined by the k = 15 BayesSpace results.

Visium: Pseudobulk sex-differential expression analyses

We set out to test each Visium cluster/domain for sex differences in expression. To achieve this, spots with a given cluster/domain level were pseudobulk aggregated within-sample using dreamlet 1.3.1. We then subsetted to clusters found in at least 5 samples (at least 2 per sex). Pseudobulk counts were preliminarily normalized to total counts using TMM normalization. We then programmatically set a low-expression filter by detecting the smaller sex group (in terms of sample number) among the retained samples for a given cluster, detecting the second minimum in a histogram of gene-wise mean log counts. This value was used to perform low-expression filtering such that retained genes which adhered to an approximately normal distribution. TMM factors were then recalculated for the remaining genes, followed by DE analysis using voomLmFit, with donor as a blocking factor (which is internally passed to DuplicateCorrelation), adaptiveSpan set to true to identify the number genes to use for smoothing the mean-variance trend, sample.weights set to true to perform an analysis analogous to the voomLmFit predecessor voomWithQualityWeights, and fitting a model of ∼0+Sex to test for sex effects. This was run for all batch correction and clustering resolutions described above, with results available in the Supplement.

Visium: Gene set enrichment testing of markers and Sex-DEGs in GSEA-MSigDB and for transcription factor target genes

We performed comprehensive, threshold-free analyses of genes marking spatial domains to confirm expected functional roles and identify potentially unexpected ones. Using the t-statistics of all genes tested in marker analysis from spatialLIBD, ¹²⁷ we applied Gene Set Enrichment Analysis ^137,138 as implemented in the package fgsea ¹³⁹ to analyze marker enrichment in gene sets from MSigDB. ^137,140,141 We included “hallmark” sets, along with sets representing chromosomal regions; curated sets from publications, including pathway databases like KEGG; encompassing predicted microRNA (miRNA)/transcription factor (TF) targets; defined in ontologies; and reported human cell type markers. (These correspond to genesets “h.all”, “c1.all”, “c2.all”, “c3.all”, “c5.all”, and “c8.all”, respectively; all versioned “2022.1”. In the sole case of the k = 15 vARC.2 domain, sex-DE GSEA was run using the “2024.1” MSigDB gene libraries.) We also included the MSigDB “oncogenic signature”/”c6.all” dataset as a null comparator and to screen for transcriptomic changes suggestive of HPA tumors, which are relatively common and whose presence would run counter to our present goal of profiling healthy neurotypical hypothalamus.

We additionally aggregated several databases of tissue-nonspecific putative target genes of transcription factors (TFs) ascertained from model organism and human cell/tissue types. We collated TF-target sets from ChEA 2022, ¹⁴² TRRUST, ¹⁴³ those mined from publication supplements by Rummagene, ¹⁴⁴ and those curated in the Enrichr¹¹⁸ tool (Gene Expression Omnibus-mined results of TF perturbation and knockout experiments and TF protein-protein interactions). We subsequently filtered the results for each domain to those TFs expressed in that domain. Given that TFs are often lowly expressed at the mRNA level, ¹¹⁹ we used lenient filters to define TFs as expressed: for marker gene enrichment, we required the at least 1% of spots to have log counts >0 for the TF in ≥5 (majority) samples and a mean [% of sample’s domain spots with nonzero counts of the TF] >5%. For sex-DE enrichment testing, we considered a TF expressed if ≥ 3 (smallest group size, when considering limited ARC or VMH spots in a Visium sample from each sex—see Methods 7.6) samples had at least 0.25% of spots in a domain with log counts >0 and a dataset mean of at least 1% of spots in the cluster with log counts >0. Our rationale for this difference in parameters was that our marker-based statistics were for broad domains, such that regulatory signatures defining the domain should be relatively abundant, whereas for sex differences, the expectation is not necessarily that all (below-detectable resolution) subdomains or cell types/subtypes in the domain—ie., a much smaller number of spots than define the domain itself—drive the detected DE events.

Gene sets analyzed were limited to those containing 15–500 genes for MSigDb and TF-target analyses. We utilized signed, transcriptome-wide one-vs-all marker or male-vs-female DE t-statistics as our ranked list for the respective analyses. To retrieve enrichment statistics from every gene set for both male-upregulated and female-upregulated expression patterns, we explicitly ran analyses for the top (male-upregulated/positive t statistic) and bottom (female-upregulated/negative t statistic) of the rank lists, as fGSEA testing on both directions simultaneously will otherwise return data for the effect direction with greater enrichment, even if both directions are significant. Additional sex-DE GSEA analyses were performed using the absolute value of DE t-statistics to identify gene sets enriched for sex-differential expression as a general phenomenon. To ensure the ability to resolve extremely small p-values, we set the fgsea parameters eps to 0.0 to specify that p-values should be estimated for all enrichments, using 50,000 permutations for initial p-value estimations. To determine the most precise, independent enrichments for some of these analyses, we subsequently used the collapsePathways function in fgsea. For marker gene analysis, we only examined “positive” enrichment, i.e., enrichments corresponding to genes more specific to a given spatial domain relative to others. For sex-DEGs, we used the same parameters and gene sets as above, but examined both directions of enrichment for each set of spatial domain’s DEGs (as the test statistic’s sign indicates male or female upregulation).

Visium: GWAS-derived gene set enrichment testing and linkage disequilibrium score testing

To comprehensively examine whether neuropsychiatric disorder risk genes were enriched among markers or sex-DE genes, we collected genes reported in common-variant (genome-wide association study (GWAS)-based) and rare-variant studies across several modalities. GWAS-derived gene sets comprised those reported in GWAS studies, explicitly excluding genes assigned to loci on the sole basis of being the most proximal to a phenotype-associated variant, while retaining e.g., genes reported in the GWAS publications resulting from “post-GWAS” analyses such as MAGMA, quantitative trait locus (QTL) colocalization, etc. For schizophrenia and autism spectrum disorders, we also collected genes identified in, e.g., whole exome sequencing cohorts of patients and healthy sibling probands. We additionally collected transcriptome-wide association study (TWAS) analyses performed on GWAS summary statistics for the same traits (generally from the same GWAS studies); largely collected from webTWAS ^78–81 and TWASatlas, ⁸² as well as individual publications that performed TWAS. Finally, we gathered literature-mined, curated gene-disease pairs from DisGeNET. ⁸³ For major depressive disorder, these included several genes of historical interest that have been explicitly demonstrated not to play a role in heritable depression risk ¹⁴⁵ ; we thus removed the genes covered by ¹⁴⁵ from this set specifically. The sets from GWAS, TWAS, and DisGeNet and corresponding sources are provided in supplemental information. For the analyses presented, we analyzed enrichments solely in the union of genes collected for each diagnosis across these three source types. As some of these gene sets fell outside the above bounds of 15–500 genes, the largest gene set size was elevated to 8200 to exceed the largest disease-gene list (BMI, union of all gene-associated modalities, >8100 genes). These findings are depicted in Figure 5, with “*” indicating FDR<0.05 and “⋅” indicating uncorrected p < 0.05.

As a complementary approach, we applied sLDSC-SEG, ⁷⁷ which uses GWAS summary statistics and assigns each single nucleotide polymorphism (SNP) to a gene based on proximity in order to determine whether genomic intervals surrounding marker genes or sex-DEGs are enriched for heritability of disease traits. This innately imposes the assumption that a variant, if exerting an effect on a gene, acts on the nearest gene (whereas the genes linked to GWAS variants were determined by functional genomics-informed methods in our GSEA approach above). We collected marker genes (logFC>0 and FDR<0.05 in one-vs-all marker analysis) for each Visium cluster, including individual and combined vVMH/vARC clusters, as well as sex DE genes from these same Visium clusters including individual and combined (FDR<0.1). Genes on chromosomes X, Y, and MT were removed. The annotated hg19 start and end coordinates of each gene in each set were extended to 100kb up/downstream, and these sets of intervals were tested for enrichment in neuropsychiatric and physical traits/disorders. ^71,146–160 For each gene set, LDSC was run using the baseline LD model v2.2, which includes 97 annotations (e.g., GC content) to control for linkage between variants and broad genomic variables. Regression SNPs were those obtained from HapMap Project Phase 3, and reference SNPs those obtained from the European 1,000 Genomes cohort (both sets obtained from resources maintained as part of LDSC). The per-SNP heritability (h²_SNP) was z-normalized within gene set group (marker gene sets or sex-DEGs), with these z scores quantifying the contribution of a gene set to a GWAS trait/disease’s heritability after adjusting for the baseline model annotations. P-values were then derived from the z scores assuming a normal distribution, with multiple testing correction using the Benjamini-Hochberg FDR procedure, again separately for marker gene set tests and sex DE tests.

Visium: Spatial registration

We opted to map our Visium spatial domains to Allen Brain Atlas in situ hybridization (ISH) data from mouse ⁵⁰ using scCoco/coco-framer, as ISH signal within atlas regions should reflect mixed neuronal and glial content akin to that measured with Visium. We provided the top 150 markers per Visium spatial domain based on one-vs-all logFC, and set a minimum of 50 genes with adult mouse ISH data for a given domain’s comparison to proceed. scCoco then was used to score the expression of query markers within each Allen Atlas Common Coordinate Framework (CCF) area of the hypothalamus. Scoring proceeds for one query marker set at a time (i.e., 50–150 genes) by ranking the ISH expression of each query gene in 10 voxels from every CCF hypothalamic area, then converting the median rank of each marker’s expression to a value ranging 0–1 by taking 1 - (median rank/number of voxels queried). Finally, these converted values are averaged across all genes for each CCF region to get an enrichment score ranging from 0 (least overlap) to 1 (most overlap). Results using development harmony with BayesSpace at k = 15, 20, and 31 are provided in the Supplement.

We additionally utilized a recent mouse brain atlasing effort that included single-cell RNA-sequencing results from 45 sequencing libraries collected across over 25 donor mouse hypothalami from both sexes. ¹² The registration_wrapper command in spatialLIBD to identify marker genes of each cell “subclass” (a relatively coarse level of hierarchical clusters reported in the Yao data). We then tested the correlation of marker t statistics from our spatial domains (k = 15, 20, 31, and k = 15 with collapsed VMH and ARC domains) with each mouse subclass’ t-scores for the same genes. We tested correlations using the top 25, 50, 75, 100, 250, 500, 750, and 1000 spatial domain marker genes (ranked by t-statistic), as well as using all genes with one-to-one mouse orthologs. We additionally repeated the analyses with all mouse VMH subclasses re-labeled as “VMH” and all ARC subclasses relabeled “ARC” to test domain correlation to collective VMH/ARC mouse types. We limited these analyses to a total of 47 subclasses—those represented by at least 15 cells per sample in 22 or more of the samples. Pseudobulk expression matrices from these analyses were also created and utilized for comparison of retinoid gene expression described below.

Gene selection for xenium custom panel

Genes were selected for measurement on the Xenium platform in an initial subset of 7 Visium samples, all from donors in the final dataset. (One sample from the preliminary dataset was not included in the final data due to low read depth compared to all other samples; a second, adjacent tissue section sampled from the same donor was processed using Visium to collect higher-depth sequencing data and is included in the final dataset). This data subset was filtered using traditional QC approaches, removing spots with over 35% mitochondrial reads, under 600 UMIs, and/or under 450 unique genes detected. This data was integrated across samples with harmony, clustered using BayesSpace with k = 15, and analyzed for marker genes, VMH/ARC-local SVGs, and sex-DE, as described above, with the difference that dreamlet 1.0.0 was used for pseudobulk sex-DE analysis and a random effect of donor. Top marker genes of one or both VMH or ARC domains were visually inspected in sample-wise fashion for consistent patterns of specific or enhanced expression in that domain across samples. For SVGs, nnSVG results were aggregated across samples by taking the mean of ordinal gene ranks in samples nominally significant (p < 0.05) for spatial variability in at least 3 samples. The top 100 mean-ranked genes each for ARC and VMH were visualized in all samples to determine whether there was a consistent spatial pattern of expression within the ARC or VMH across donors, and visually robust genes were selected for inclusion. For sex-DE genes, the preliminary analysis with dreamlet 1.0.0 returned a much higher (compared to the final dataset) number of significant DEGs at FDR<0.05 due to algorithmic differences at that time, with approximately 150 in VMH and 400 in ARC. All genes meeting this significance threshold were plotted and visually examined in spotplots to determine whether the male/female pattern of expression difference was consistent. Further, visual inspection was used to rule out DE events that occurred due to genes predominantly detected at the edges and outside of VMH/ARC, which would indicate confounding due to differences in VMH/ARC boundaries relative to the neighboring domains primarily expressing the gene. Additional genes were selected on the basis of genes identified in model organism literature as marking VMH or ARC or their general relevance to sex differences in gene expression (e.g., androgen and estrogen receptors). We provided 10x Genomics with expression reference data for the panel design (balancing probing depth for each gene based on its expression) in the form of 1 male (v6197A_M) and 1 female (v6588B_F) Visium sample, with k = 15 BayesSpace cluster labels from our preliminary analysis. A table of the final panel of genes and the rationale(s) for inclusion in the study is provided in the Supplement.

Xenium: Data QC and preprocessing

Xenium data was initially assembled into an R SpatialFeatureExperiment (SFE) ¹⁶¹ object by importing xeniumranger outputs: cell boundaries, nucleus boundaries, and per-cell count data for transcript detections above quality score (QV) > 20. Cells and nuclei with segmentations containing holes as detected by the nngeo package were removed. Per-cell proportions of each type of negative control elements (negative control probes, negative control codewords, and unassigned/formerly BLANK codewords) were tabulated for each cell and added to the SFE. As the Xenium documentation does not address ‘Deprecated’ codewords or what they correspond to, this category of negative element was not considered further. Samples were then rotated or mirrored in space so that, for visualization, all tissue sections were oriented with the ARC in the bottom left-hand corner.

We then performed general quality control of the data as outlined in the documentation for the Voyager R package. Briefly, we first screened the data for nuclei that had either zero area or had an area ≥95% of their parenting cell’s segmentation (none were found). We then visualized the proportion of counts coming from each type of negative control element or of all elements considered jointly, confirming there were not marked spatial differences in negative element detection. We also examined the distribution of these proportions, finding that the vast majority of cells had 0 negative element reads and a small number of cells with ≥10% such reads. Following the Voyager approach, we used scuttle::detectOutliers to remove cells for which the proportion of reads contributed by any negative element (or their union) was >3 median absolute deviations above the sample’s median. Cells deemed outliers in this manner also did not show any clear spatial pattern, though they were more visually dense in more cell-dense areas. Finally, we removed nuclei and corresponding cells wherein the nuclear area was under 400 pixels, ² resulting in an initial dataset of 870,870 cells/nuclei.

Cell counts were normalized using both TMM normalization into log counts based on the number of total counts in a cell, and normalized without log transformation using scuttle::normalizeCounts(log = FALSE). PCA and UMAP dimensionality reduction was performed to visually inspect for batch effects of donor and Xenium run; remarkably, the non-log-normalized data (the data used for clustering with Banksy, as advised by its authors ¹⁶²) occupied a UMAP space shared across these potential batch covariates (Figure S19), and thus did not require batch integration for further analysis.

Xenium: Clustering with banksy

We subsequently set out to cluster the segmented Xenium cells using the Banksy package in R. ¹⁶² Per their documentation for multi-sample datasets, we placed an artificial buffer zone between the spatial coordinates of each tissue section, such that each section had unique x-coordinates and was “spaced” at least 1.5 tissue section widths apart from every other section. We then performed non-spatial clustering of the Xenium data with Banksy, using non-logarithmic “relative counts” normalization (scater::normalize-Counts with argument log = FALSE) as suggested in the Banksy documentation. We then performed Banksy dimensionality reduction of the relative counts with parameters kgeom = 6 (setting the physical 6 nearest cells to be defined as a cell’s neighborhood) and the azimuthal Gabor filter setting and lambda both set as 0 for spatially-unaware dimensionality reduction, followed by Leiden clustering with a resolution parameter of 2. We elected to forgo spatially-aware clustering in the Xenium data in order to examine whether genuinely distinct spatial clusters were present without biasing our analyses toward producing spatially-arranged clusters a priori.

Xenium: Cell type annotation and VMH/ARC domain smoothing

We subsequently removed 8 sample/donor-specific clusters (one of which was a VMH cluster) from the SFE object to perform pseudobulk one-vs-all marker detection for each of the 33 clusters found across the dataset using spatialLIBD as described above for Visium. Based on marker genes, we were able to assign each cluster a putative identity, including 3 main VMH clusters and 5 ARC clusters. One ARC cluster, marked by GAL expression, was present both in the ARC and along the entire medial longitudinal axis of the VMH; as this cluster was not restricted to the traditional atlas boundaries of ARC (ventral to VMH and extending only along the ventromedial portion of VMH), we did not include this cluster in the smoothing procedure to define the ARC domain. For smoothing of the VMH, we excluded a fourth, rare cluster (7.3k cells total) tracing the approximate edges of VMH. We note that this cluster is annotated as “VMH boundary” in recognition of the spatial pattern, but was not marked by any canonical VMH genes (SLC17A6, FEZF1, NR5A1).

In order to define the spatial boundaries of VMH and ARC, we then performed a two-step spatial k-nearest neighbors (kNN) approach. In brief, this procedure first quantified the proportion of each cell’s 10 (VMH) or 50 (ARC) spatially nearest neighbor cells (including those in sample-specific clusters) that were one of the three main VMH clusters or 4 main ARC clusters. A second round of kNN, obtaining the proportion of the 200 (VMH) or 500 (ARC) spatial neighbors labeled as VMH or ARC was performed, with intermediate VMH and ARC labels assigned based on thresholding the first kNN value for each region at 0.1, 0.2 … 1.0 (100 VMH/ARC combinations). We collected kNN values for the second step using each first-round threshold pair, and subsequently tested final definitions of ARC and VMH with thresholds of 0.1, 0.2 … 1.0 as applied to the second-round kNN smoothing (total of 10,000 combinations of VMH-ARC threshold pairs).

For each of these pairs, pseudobulk DE testing was done for marker genes between all cells passing the given final VMH or ARC threshold to quantify VMH/ARC marker gene enrichment in the respectively assigned domains. DE analysis was performed using registration_wrapper in spatialLIBD as we did for other gene marker analyses. Cells from any of the 33 retained clusters falling within the boundaries of the VMH or ARC domain were given that domains label as their ‘cluster’, while any cells from a VMH or ARC cluster but found outside the two domains were removed from analysis. One-vs-all analysis was then performed with spatialLIBD using these modified labels such that VMH and ARC domains were treated as single clusters alongside the unmodified (non-VMH, non-ARC neuron) clusters; enrichment statistics for the VMH and ARC domains were then retrieved. First-second round combination pairs with robust enrichment of VMH markers in VMH and of ARC markers in ARC were visualized to determine whether the parameters also resulted in contiguous and directly adjacent VMH and ARC domains. For smoothing parameter combinations where VMH and ARC boundaries were adjacent, we further performed pseudobulk DE analysis to determine whether expected VMH and ARC markers were enriched on their respective sides of the boundary zone. To achieve this, we used the sf package to define the convex and concave hulls of smoothed VMH/ARC domains, then extended these boundaries by 5% of the median distance between hull vertices using st_buffer in sf. We performed a second round of VMH/ARC pseudobulk DE testing only comparing intra-VMH and intra-ARC cells within the overlap region.

Based on the magnitudes of enrichment for known marker genes, domain continuity, and domain adjacency, we selected first-round KNN thresholds of 0.2 (VMH) and 0.1 (ARC) followed by second-round KNN thresholds of 0.2 (VMH) and 0.5 (ARC). Cells that passed both ARC and VMH thresholds in the first or second step were deconflicted by assigning the higher KNN score (e.g., if the KNN round 1 proportion of ARC neighbors was 0.25 and of VMH neighbors was 0.3, then the cell would be labeled VMH going into the second round of KNN analysis). The labels of VMH or ARC were then assigned to all cells, regardless of cluster, to demarcate the boundaries of the VMH and ARC. Refined labels were determined by performing a one-vs-others pseudobulk enrichment approach similar to that described for Visium VMH/ARC clusters in section 7.5: that is, we tested pseudobulk expression on data subsetted only to ARC clusters or only to VMH clusters. This again was performed using tools from spatialLIBD.

Domain labels were then cleaned by removing cells outside of the main domain area that received the label; in the case of one donor, Br8741, both Xenium samples were instead processed with a second round kNN threshold of 0.15 for VMH due to holes in the domain assignments at higher thresholds. This donor contained the donor-specific VMH cluster, which was excluded from the smoothing procedure, resulting in fewer remaining VMH neurons in the sample and consequently in lower KNN proportions for VMH neuron type-neighbors.

Xenium: Sex-differential expression analyses

We set out to test each cell cluster for sex differences in expression local to the VMH or ARC. To achieve this, we first extracted all cells within either the VMH domain or the ARC domain as defined above, then performed pseudobulk aggregation within samples for each cell cluster using dreamlet 1.3.1. We then subsetted to clusters for which there were at least 6 samples with at least 10,000 total transcript detections for that cluster within the domain of interest. Pseudobulk counts were preliminarily normalized to total counts using TMM normalization. We then programmatically set a low-expression filter by detecting the smaller sex group (in terms of sample number) among the retained samples for a given cluster in a given domain by detecting the second minimum in a histogram of gene-wise mean log counts per million to retain only those genes which adhered to an approximately normal distribution. Genes falling below this second minimum were discarded and TMM factors recalculated for the remaining genes. DE analysis was then performed with voomLmFit, using donor as a blocking factor (which is internally passed to DuplicateCorrelation), adaptiveSpan set to true to identify the number genes to use for smoothing the mean-variance trend, sample.weights set to true to perform an analysis analogous to the voomLmFit predecessor voomWithQualityWeights, and fitting a model of ∼0+Sex+Xenium Run to test for sex effects and control for potential variation between the two Xenium runs. The filtered pseudobulk object and full DE analysis output for each cluster within each domain were returned for visualization and downstream analyses. This same approach was also taken to conduct Visium-analogous, domain-level sex-DE analyses, treating cells with any of the 33 retained cluster labels and inside the VMH or ARC domain as “VMH” or “ARC”, respectively.

In order to validate our sex-DE findings within VMH and ARC domains, we performed two additional high-confidence analyses. In the first, we utilized the sf package in R to define the convex and concave hulls for VMH and ARC, respectively, and trimmed these boundaries back by 5% of their respective domain areas to obtain “high-confidence” domains. Cells falling within these interior domain boundaries were then analyzed as above for sex-DE. In a second analysis, we sought to rule out cell segmentation effects that could have resulted in transcripts being incorrectly attributed to a given cell type (e.g., astrocyte endfoot processes could potentially end up within neuronal segmentation boundaries). We therefore also performed sex-DE only considering transcripts overlapping nucleus segmentations by assigning a given nucleus the same cluster label as its parent cell segmentation. The nuclear count data was filtered to quality scores >20 and totaled per gene per nucleus using the MoleculeExperiment package in R, ¹⁶³ followed by the same pseudobulk per-cluster-per-domain procedure as described above.

smFISH puncta localization and quantification

For Experiment 2, HALO (Indica Labs) was used to segment and quantify fluorescent signals for each probe in single cells. Nikon.nd2 files were imported into HALO and viewed at the manufacturer’s recommended magnification (40x) for punctate probe signal quantification. The FISH-IF module was used to quantify RNA transcripts (copy counts) within each detected object (i.e., a nucleus with radially defined “membrane” at 3 μm to estimate a cell) in accordance with the manufacturer’s guidelines: HALO 3.3 FISH-IF Step-byStep guide (Indica labs, v2.1.4 July 2021). For each image, fluorescent detection parameters were optimized for DAPI and probe signal to ensure image faithfulness in accordance with the HALO 3.6 User Guide (Indica labs, February 2023). To ensure accurate counting of RNA transcript copies assigned to each cell for each of the 3–4 RNAScope probe channels, the median probe signal intensity value of positive cells per signal was input into HALO as the representative intensity value parameter for that channel’s copy intensity parameter consistent with previously published work. ¹⁶⁴

HALO.csv outputs were imported into R. As previously described, ¹⁶⁵ k-means clustering was performed on HALO variables Signal Area and Copy Count with a k = 3 to cluster “high,” “medium,” and “low/no” expressing cells) for each of the RNAScope probe channels. High and medium clusters were deemed positive. All cells were plotted as circles with x and y coordinates as determined by HALO, and colored by positivity for each channel. The ARC region of interest (ROI) was selected by drawing a polygon around the area expressing TAC3 and KISS1 for each donor. A data frame was then produced containing all information for each segmented cell in the ROI, including cell phenotypes (channel(s) meeting the positivity definition above). Phenotype occurrences were then tabulated to quantify cell phenotype counts and proportions. All HALO settings files are provided on GitHub.

ADDITIONAL RESOURCES

Interactive data browsers and their documentation can be found at: research.libd.org/spatial_HYP.

Supplementary Material

Supplemental information can be found online at https://doi.org/10.1016/j.celrep.2025.116877.

• Neuronal subpopulations of human ARC, but not VMH, appear to be spatially organized

• Regional sex-DE in ARC and VMH is respectively driven by ESR1+ and CRHR2+ neurons

• Autism spectrum disorder-associated gene expression ↑ in male (vs. female) VMH and ARC

• Retinoid genes and their regulatory targets are broadly expressed in human ARC/VMH