Transcriptional and functional divergence in lateral hypothalamic glutamate neurons that project to the lateral habenula and ventral tegmental area

Summary

The lateral hypothalamic area (LHA) regulates feeding and reward related behavior, but because of its molecular and anatomical heterogeneity, the functions of defined neuronal populations are largely unclear. Glutamatergic neurons within the LHA (LHA^Vglut2) negatively regulate feeding and appetitive behavior. However, this population comprises transcriptionally distinct and functionally diverse neurons that project to diverse brain regions, including the lateral habenula (LHb) and ventral tegmental area (VTA). To resolve the function of distinct LHA^Vglut2 populations, we systematically compared projections to the LHb and VTA using viral tracing, single-cell sequencing, electrophysiology, and in vivo calcium imaging. LHA^Vglut2 neurons projecting to the LHb or VTA are anatomically, transcriptionally, electrophysiologically, and functionally distinct. While both populations encode appetitive and aversive stimuli, LHb projecting neurons are especially sensitive to satiety state and feeding hormones. These data illuminate the functional heterogeneity of LHA^Vglut2 neurons, suggesting that reward and aversion are differentially processed in divergent efferent pathways.

Introduction

The lateral hypothalamic area (LHA) is a neuroanatomical hub that regulates diverse motivated behaviors, including feeding and reward (Abrahamsen et al., 1995; Berthoud and Münzberg, 2011; Rossi and Stuber, 2018; Stuber and Wise, 2016). LHA lesions abolish feeding, and electrical stimulation promotes food approach and consumption (Anand and Brobeck, 1951; Hoebel and Teitelbaum, 1962; Margules and Olds, 1962). However, recent studies have shown that the LHA comprises functionally, transcriptionally, and anatomically distinct neuronal populations. (González et al., 2016; Jennings et al., 2013, 2015; Lazaridis et al., 2019; Mickelsen et al., 2019; Nieh et al., 2015; Rossi et al., 2019). Because of this complexity, which cells constitute functionally distinct cell types and how they coordinate motivated behavior remain largely unknown.

Adding an additional layer of complexity to understanding LHA function, circulating signals, such as the anorexigenic adipose-derived hormone leptin and the orexigenic gut-derived hormone ghrelin, influence feeding behavior via both direct and indirect action on LHA cells. Leptin infusion into LHA reduces food intake and body weight (Leinninger et al., 2009), whereas ghrelin infusion increases food intake (Olszewski et al., 2003). Both leptin and ghrelin are thought to regulate behavior via influences on mesolimbic dopamine (Cone et al., 2014; Domingos et al., 2011), but the circuit mechanisms underlying this are still being investigated (Abizaid et al., 2006; Fulton et al., 2006; Leinninger et al., 2009; Liu et al., 2017) see (Zessen et al., 2012) for review. Currently, little is known about how feeding hormones influence activity dynamics of defined LHA neuronal populations in vivo.

One LHA neuronal population residing upstream of dopaminergic cells that has been shown to regulate feeding, reward, and aversion is glutamatergic neurons expressing the vesicular glutamate transporter type 2 (encoded by Slc17a6, hereafter referred to as Vglut2). LHA^Vglut2 activation suppresses feeding and is aversive (Jennings et al., 2013; Rossi et al., 2019). However, LHA^Vglut2 neurons show considerable heterogeneity in their in vivo response dynamics during rewarding and aversive events (Lazaridis et al., 2019; Rossi et al., 2019). LHA^Vglut2 neurons also project to diverse brain regions, including the lateral habenula (LHb; LHA^Vglut2→LHb) and ventral tegmental area (VTA; LHA^Vglut2→VTA), two brain regions known to regulate reward and aversion (Lammel et al., 2012; Trusel et al., 2019). Indeed, activation of both LHA^Vglut2→LHb and LHA^Vglut2→VTA projections promotes avoidance (de Jong et al., 2019; Nieh et al., 2016; Stamatakis et al., 2016). Despite this, little is known about in vivo activity dynamics of individual neurons projecting to the LHb or VTA. Even less is known about their anatomical, transcriptional, or electrophysiological properties.

Here, we systematically compared LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons. Using viral-mediated projection targeting, we found that LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons are distinguished by their anatomy, electrophysiological properties, and gene expression. In a series of in vivo two-photon calcium imaging experiments, we show that reward and aversion-associated stimuli (sucrose and quinine) are differentially encoded within these pathways. Finally, we demonstrate that satiety state and feeding hormones uniquely alter the sucrose reward encoding properties of LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons. Together, these results highlight the functional heterogeneity of LHA^Vglut2 neurons and show how reward and aversion are uniquely processed by divergent efferent pathways.

Results

LHA^Vglut2→LHb and LHA^Vglut2→VTA projections are anatomically distinct

To determine whether LHA^Vglut2 projections to the LHb and VTA are the same or unique cells, we used retrograde viral tracing. Cre-inducible, retrogradely trafficked viral vectors were microinjected into the LHb (retroAAV-DIO-eYFP) and VTA (retroAAV-DIO-mCherry) (Figure 1A–C and S1A–B). Fluorophore expression was examined within the LHA using fluorescence in situ hybridization. LHA^Vglut2→LHb and LHA^Vglut2→VTA projections showed unique expression patterns across the anterior-posterior LHA axis: LHA^Vglut2→LHb projections were more concentrated in the anterior LHA, whereas LHA^Vglut2→VTA projections were more concentrated in the posterior LHA (Figure 1D–F). LHA^Vglut2→VTA cells were found in greater numbers than LHA^Vglut2→LHb cells, but both were observed more frequently than double labeled cells (Figure 1G), indicating LHA^Vglut2→LHb and LHA^Vglut2→VTA projection pathways are largely anatomically distinct.

Figure 1: LHA^Vglut2→LHb and LHA^Vglut2→VTA projections are anatomically distinct.

A. Dual retrovirus strategy.

B-C. Examples of locally transduced cells at LHb and VTA injection sites. Blue, DAPI. Scale bars, 200 μm.

D-E. Confocal micrographs showing in situ hybridization of eYFP and mCherry expression in anterior (D) and posterior LHA (E). Approximate AP coordinates are shown. Scale bars: 200 μm top, 50 μm bottom. +, eYFP+; *, mCherry+; ^, eYFP+/mCherry+.

F. Relative density of LHA^Vglut2→LHb and LHA^Vglut2→VTA projections throughout the AP extent of the LHA (n=10 hemispheres; 5 mice; effect of Projection: F(2,18)=65.83, p=5.3e-9; no effect of AP location: F(6,54)=1.97, p=0.09; Interaction: F(12,108)=12.05, p=4.0e-15).

G. Proportion of cells classified as projecting to LHb, VTA, or both. CP, cerebral peduncle; EP, entopeduncular nucleus; MHb, medial habenula; SNC, substantia nigra pars compacta; SNR, pars reticulata; STN, subthalamic nucleus.

LHA^Vglut2→LHb and LHA^Vglut2→VTA projections are transcriptionally distinct

We next sought to test whether LHA^→LHb and LHA^→VTA projections exhibited unique transcriptional signatures by combining retrograde viral labeling and single-cell mRNA sequencing. Wild-type mice were injected with retroAAV-hSyn-eYFP in the LHb and retroAAV-CAG-tdTomato in the VTA to label projection neurons (Figure S1C). LHA tissue punches were collected, and single cells were processed via a Chromium v3 droplet based mRNA sequencing pipeline (Figure 2A; Hashikawa et al., 2020). Low quality cells and doublets were removed, and clustering was performed on 34,518 cells (Figure S1D–H). Of these, 17,378 were classified as neurons, which formed 32 subclusters (Figure 2B). Following quality control steps, we detected a median of 5,544 unique molecular identifiers (UMIs), 3,106 genes, 38,544 reads, and 5% mitochondrial reads per neuron. We detected eYFP (n=257 neurons) and tdTomato (n=1,595 neurons) transcripts in largely non-overlapping LHA neurons (Figures 2C and S1I). LHb projections were primarily glutamatergic, while VTA projections were more heterogeneous (Figure S1J). To determine if eYFP and tdTomato populations express unique marker genes, we performed logistic regression based differential gene expression analysis, revealing 106 differentially expressed genes (Figure 2D) and suggesting that LHA^→LHb and LHA^→VTA projections are transcriptionally distinguishable using multiple gene markers.

Figure 2: LHA^Vglut2→LHb and LHA^Vglut2→VTA projections are transcriptionally distinct.

A. Experimental pipeline. Fluorophore-encoding retrogradely trafficked AAVs were injected into LHb and VTA. LHA tissue punches were taken and single-cell RNA sequencing was performed.

B. UMAP showing 32 LHA neuronal subclusters (n=6).

C. Proportion of labeled cells expressing tdTomato (tdT), eYFP, or both.

D. DEG analysis comparing eYFP+ and tdTomato+ cells reveals 106 differentially expressed genes (orange and green; p<0.01, dashed line). Positive log fold change (logFC) indicates enrichment in eYFP+ cells. Highlighted genes were selected for in situ validation (See Figure 3).

E. eYFP+ and tdTomato+ cells are enriched in distinct neuronal subclusters.

F. Pax6 and Pdyn expression within neuronal subclusters.

G. Neuronal subclusters are distinguished by unique genes. Highlighted clusters were targeted for in situ validation.

Previous LHA sequencing results defined neuronal subclusters by combinatorial gene expression (Mickelsen et al., 2019; Rossi et al., 2019), and we extend these findings, demonstrating how LHA^→LHb and LHA^→VTA projection neurons segregate within defined subclusters. Clusters could first be classified based on expression of genes commonly used to target distinct LHA cell types, including Hcrt, Pmch, Slc32a1 (Vgat), and Vglut2, and secondary marker genes including neuropeptides, receptors, or transcription factors (Figure S1K–L). Clusters marked by Hcrt and Pmch also expressed low levels of Vglut2. eYFP and tdTomato labeled cells exhibited unique distributions within the neuronal subclusters (Figure 2E). Of the neuronal subclusters enriched with labeled cells, four also expressed high levels of Vglut2. They could be further differentiated by the presence of subcluster-specific gene markers, including Pax6 and Sostdc1 (LHA^Vglut2→LHb clusters); Hcrt, Pdyn and Pitx2 (LHA^Vglut2→VTA clusters); and Nptx2 (LHA^Vglut2→LHb and LHA^Vglut2→VTA clusters) (Figure 2F–G and S1K–L). The densest concentrations of labeled cells were found in distinct subclusters. Specifically, Pax6 robustly labeled Glut1, which was highly enriched with eYFP+ cells, whereas Pdyn labeled the glutamatergic hypocretin/orexin (Pdyn/Hcrt) cluster, which was enriched with tdTomato+ cells (Figure 2F). Nptx2 expressing cells were concentrated within both Glut1 and Pdyn/Hcrt clusters. The eYFP-rich Glut2 cluster was marked by Sostdc1, and the tdTomato-rich Glut14 cluster expressed Pitx2. Collectively, single-cell mRNA sequencing revealed that LHA^Vglut2→LHb and LHA^Vglut2→VTA projection populations express distinct molecular markers.

To cross-validate the observed transcriptional patterns, expression of genes that distinguish the two projections was assessed in situ with sequential single molecule hybridization chain reaction (HCR; von Buchholtz et al., 2021; Choi et al., 2018) for nine genes that characterized LHA^Vglut2→LHb and LHA^Vglut2→VTA enriched neuronal clusters (Vglut2, Vgat, Pax6, Pdyn, Nptx2, Sostdc1, Pitx2, eYFP, and tdTomato; Figure 3A–B). In each of three HCR rounds, mRNA from three genes were labeled and imaged. Between rounds, probes from the previous round were digested to permit new hybridization. Cells from all rounds were then aligned using brightfield images. mRNA puncta were quantified within single cells based on DAPI staining (HALO, Indica Labs). Each gene exhibited a unique spatial expression pattern within the LHA (Figures 3C–K and S2A–B). In agreement with previous results (Figure 1), eYFP and tdTomato expressing cells were spatially intermixed (Figure 3L). Proportions of virally labeled cells also expressing gene markers for LHA^Vglut2→LHb and LHA^Vglut2→VTA populations were similar between sequencing and HCR (Figure 3M–O). Consistently, the log fold-change between eYFP+ and tdTomato+ cells for the nine genes targeted for HCR was highly correlated across methodologies (Figure 3P). Based on the single-cell sequencing, we predicted that Nptx2 would label both projection populations with enrichment in Pdyn+ cells; however, HCR analysis indicates that this gene may be more highly expressed within the LHA^Vglut2→LHb neurons. We then asked whether the mRNA expression data could be used to identify eYFP+ and tdTomato+ cells using support vector machine (SVM) based classification. Both sequencing and HCR data performed equally well at classifying the two projections (Figures 3Q and S2C–D). These results confirm that LHA^Vglut2→LHb and LHA^Vglut2→VTA projections are transcriptionally distinct.

Figure 3: HCR validation of single-cell sequencing results.

A. Dual retrograde virus strategy used to profile LHA projection neurons.

B. Sequential HRC workflow. Three genes were targeted per round.

C-K. Example expression of the nine genes targeted. Rows correspond to HCR rounds. Scale bar, 200 μm. ^Λ eYFP+, < tdT+

L. Normalized spatial distribution of eYFP+ and tdTomato+ cells within LHA.

M-N. Proportion of eYFP+ and tdTomato+ cells expressing each of the 7 marker genes in HCR (M) and RNA sequencing (N) experiments.

O. Percent cells expressing each gene is highly correlated between sequencing and HCR experiments (r=0.81, p=0.00047).

P. LogFC values were highly correlated between RNA sequencing and HCR experiments (r=0.94, p=0.00015). Negative values indicate enrichment in tdTomato+ cells.

Q. HCR and RNA sequencing data are equally effective at classifying cells as eYFP+ or tdTomato+ (* Bonferroni adjusted p<0.05).

LHA^Vglut2→LHb and LHA^Vglut2→VTA projections are electrophysiologically distinct

To assess whether functional differences accompany the observed transcriptional differences, LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons were targeted for whole cell patch clamp recording (Figure 4A). Overall, LHA^Vglut2→LHb neurons were more active at rest and more excitable than VTA projectors (Figure 4B–J). LHA^Vglut2→LHb and LHA^Vglut2→VTA projections showed similar resting potentials (Figure 4C) and action potential amplitudes (Figure 4D). However, LHb projectors had lower membrane capacitance (Figure 4E), more hyperpolarized spike thresholds (Figure 4F), and smaller afterhyperpolarizations (Figure 4G), which likely underlie the increased spontaneous firing (Figure 4H), lower rheobase (Figure 4I), and increased spiking during positive current injections (Figure 4J). Consistent with the distinct electrophysiological signatures, these properties could be used to decode the projection target (Figure 4K).

Figure 4: LHA^Vglut2→LHb and LHA^Vglut2→VTA projections are electrophysiologically distinct.

A. Dual retrovirus strategy (left) and example cells (right).

B. Example whole cell recording of basal firing rate (left), current step injections (center; −20, 0, and +190 pA injections), and rheobase (right; 10 ms current injections, 10 pA steps).

C. Resting potential (46/47 cells; 11/10 mice; t(91)=0.01, p=0.99).

D. Action potential amplitude (45/46 cells; 11/10 mice; t(89)=0.01, p=0.99).

E. Membrane capacitance (46/47 cells; 11/10 mice; t(91)=2.69, p=0.008).

F. Action potential threshold (45/46 cells; 11/10 mice; t(89)=4.56, p=0.00002).

G. Afterhyperpolarization (AHP) amplitude (45/46 cells; 11/10 mice; t(89)=2.53, p=0.01).

H. Cell attached spontaneous firing rate (30/35 cells; 8/7 mice; t(63)=4.02, p=0.0002).

I. Rheobase (46/47 cells; 11/10 mice; t(91)=5.62, p=2e-7).

J. Current step injection (spikes per 800 ms sweep. 45/46 cells; 11/10 mice. Effect of Current: F(19,1691)=69.92, p<1e-6; Effect of Projection: F(1,89)=6.34, p=0.01; Interaction: F(19,1691)=3.77, p<1e-6). Mean±s.e.m.

K. Electrophysiological properties can be used to decode the projection target (p=0.002).

LHA^Vglut2→LHb and LHA^Vglut2→VTA projection neurons differentiate appetitive and aversive tastants in vivo

Previous studies have implicated both LHA^Vglut2→LHb and LHA^Vglut2→VTA projections in regulating reward and aversion (de Jong et al., 2019; Lazaridis et al., 2019; Nieh et al., 2016; Stamatakis et al., 2016); however, the underlying single neuron activity dynamics are largely uncharacterized. To address this, we combined projection- and cell-type-specific viral targeting with in vivo two-photon calcium imaging using the INTERSECT strategy to record activity dynamics from LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons during presentation of appetitive and aversive tastes (Fenno et al., 2014, 2020; Figure 5A and S3). Thirsty mice were randomly presented with sucrose and quinine solutions (75/25% of trials, respectively; Figure 5B). Because mice could not predict which solution would be delivered, they reliably consumed both tastants; however, licking ceased earlier on quinine trials compared with sucrose trials (Figure 5C). Many neurons showed excitatory responses that were timelocked to sucrose and quinine consumption (Figure 5D–H). Consistent with previous reports (de Jong et al., 2019; Lazaridis et al., 2019), neurons belonging to each projection pathway showed greater responses to quinine than to sucrose (Figure 5I). Interestingly, quinine potentiated the responses of LHA^Vglut2→VTA neurons significantly more than LHA^Vglut2→LHb neurons (Figure 5J–K), yielding greater ability to decode the delivered tastant from activity dynamics of individual LHA^Vglut2→VTA neurons compared with LHA^Vglut2→LHb neurons (Figure 5L). LHA^Vglut2→VTA neurons were also more responsive to water consumption than were LHA^Vglut2→LHb neurons (Figure S4). Together, these results suggest that LHA^Vglut2→LHb and LHA^Vglut2→VTA pathways respond to both positive and negative valence tastants, though responses are greatest to aversive stimuli.

Figure 5: LHA^Vglut2→LHb and LHA^Vglut2→VTA projections differentiate appetitive and aversive tastants.

A. Intersectional virus strategy.

B. Experimental design.

C. Average±s.e.m. lick rate for all trials (n=10 mice).

D. Example field of view for LHA^Vglut2→LHb (top) and LHA^Vglut2→VTA (bottom) recordings.

E. Calcium dynamics extracted from the cells in D.

F-G. Single LHA^Vglut2→LHb and LHA^Vglut2→VTA ROI activity dynamics for sucrose and quinine trials. Data are aligned to first lick after tastant delivery (dashed line). Right, mean responses. ROIs correspond to arrows in D. Heatmap scale bar: 2s.

H. Population activity dynamics for sucrose and quinine trials for LHA^Vglut2→LHb and LHA^Vglut2→VTA (n=5 mice/group) recordings. Mean±s.e.m. traces at bottom. Scale bar 2s.

I. AUC for LHA^Vglut2→LHb and LHA^Vglut2→VTA recordings for sucrose and quinine trials (effect of Tastant: F(1,268)=63.48, p=4.7e-14; no effect of Projection: F(1,268)=1.75, p=0.19; Interaction: F(1,268)=3.97, p=0.047). *, Tukey multiple comparisons test p<0.05. Mean+s.e.m.

J. ROI sucrose and quinine responses are correlated (LHb: Pearson r=0.86, p<1.0e-15; VTA: r=0.72, p<1.0e-15; *, F test to compare intercepts: F(1,267)=10.48, p=0.0014).

K. Percent ROIs in which quinine response magnitude is higher (↑), lower (↓), or unchanged (↔) relative to sucrose response (paired t-test, p<0.05). Populations compared with chi-squared test (X²=15.60, p=0.0004).

L. Activity of individual ROIs belonging to both projections can decode delivered tastant (Welch’s t test LHb model vs shuffle: t(275.4)=14.98, p<1.0e-15; VTA model vs shuffle: t(236.2)=20.84, p>1.0e-15; t test LHb model vs VtA model: t(268)=2.10, p=0.037).

Reward encoding in LHA^Vglut2→LHb and LHA^Vglut2→VTA projections depends on satiety state and is modified by feeding hormones

LHA neurons have long been known to alter food reward encoding according to satiety state (de Araujo et al., 2006; Burton et al., 1976; Rossi et al., 2019). To test whether sucrose reward encoding in LHA^Vglut2 projections is affected by satiety state, we recorded activity dynamics from LHA^Vglut2→LHb and LHA^Vglut2→VTA projections when mice were ad lib. fed and after a 24 hr fast (Figure 6). While fasting elevated total sucrose seeking (Figure S5A), consummatory lick patterning was similar between conditions (Figure S5B–C). Individual neurons from both LHA^Vglut2→LHb and LHA^Vglut2→VTA projection populations responded more vigorously to sucrose consumption when fasted compared to when fed (Figure 6A–D), but this difference could not be explained by differences in licking (Figure S5D–G). Notably, the projections exhibited reliable differences in their temporal dynamics; LHA^Vglut2→LHb responses tended to peak more quickly than LHA^Vglut2→VTA responses. Analysis of basal dynamics in vivo in the absence of sucrose consumption also supports the observed ex vivo electrophysiological differences in which LHA^Vglut2→LHb neurons are more active and excitable than LHA^Vglut2→VTA neurons (Figure S5H–K). Individual neuron responses in fasted and fed conditions were highly correlated in vivo (Figure 6E). In agreement with the increased response magnitudes after fasting, a greater proportion of both LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons were responsive when mice were fasted compared to when fed (Figure 6F), and activity dynamics of both populations could be used to decode the mouse’s satiety state (Figure 6G). Interestingly, when mice were fed, a greater proportion of LHA^Vglut2→LHb than LHA^Vglut2→VTA neurons responded to sucrose consumption but fasting eliminated this difference (Figure 6F). This is supported by the observation that fasting reduces dissimilarities between electrical properties of LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons ex vivo (Figure S6A–H). Decoding analysis revealed significantly greater ability to discriminate the projection target in fed mice relative to fasted mice using electrophysiological properties (Figure S6I). Synaptic input onto both LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons was uniformly affected by fasting (Figure S6J–S). Collectively, the results of fasting experiments suggest that satiety states robustly shape the activity of both LHA^Vglut2→LHb and LHA^Vglut2→VTA neuron populations, with fasting reducing the functional dissimilarity between pathways.

Figure 6: Satiety modifies reward encoding in LHA^Vglut2→LHb and LHA^Vglut2→VTA projections in vivo.

A. Field of view for LHA^Vglut2→LHb (top) and LHA^Vglut2→VTA (bottom) projections.

B. Single ROI examples (arrows in A) of activity dynamics during sucrose consumption when mice are fed and fasted. Data are aligned to first lick after sucrose delivery (dashed line). Bottom, mean response. Scale bars: 10% ΔF/F (bottom). Horizontal scale, 2s.

C. LHA^Vglut2→LHb and LHA^Vglut2→VTA population responses during sucrose consumption (n=5 mice/group). Bottom, mean±s.e.m. Vertical scale, 2% ΔF/F. Horizontal scale, 2s.

D. AUC is increased by fasting in both projection populations (effect of Fasting: F(1,578)=21.77, p=3.8e-6; no effect of Projection: F(1,578)=3.40, p=0.07; no Interaction: F(1,578)=0.15, p=0.70). *, Tukey multiple comparisons test p<0.05. Mean+s.e.m.

E. AUC values are correlated in fed and fasted ROIs (LHb: Pearson r=0.79, p<1.0e-15; VTA: r=0.75, p<1.0e-15; *, F test to compare slopes: F(1,287)=5.96, p=0.015).

F. Percent cells responding to sucrose consumption. Both projections showed more responsivity when fasted than when fed (LHb: X²=25.56, p=2.8e-6; VTA: X²=36.95, p=9.4e-9). When fed, LHb projectors are more responsive than VTA projectors to sucrose consumption (X²=12.58, p=1.9e-3), whereas projections show similar responsivity when fasted (X²=1.71, p=0.42).

G. Activity dynamics from single ROIs of both projections could be used to decode satiety state significantly better than shuffled data (LHb: Welch’s t(280.26)=7.13, p=8.5e-12; VTA: t(237.27)=6.18, p=2.7e-9).

Because satiety state alters encoding of sucrose rewards in LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons, we hypothesized that their activity dynamics might therefore be influenced by circulating feeding hormones, which are known to fluctuate based on energy state (Bagnasco et al., 2002; Jensen et al., 2013). Leptin and ghrelin exert opposing control over feeding via the LHA (Leinninger et al., 2009; Olszewski et al., 2003). mRNA for the leptin receptor, Lepr, and the ghrelin receptor, Ghsr, are found within many LHA cells, including a subset of LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons (Figure S7A–H). We tested whether these hormones influenced sucrose reward encoding by injecting leptin (1.5 mg/kg i.p. fasted), ghrelin (1.0 mg/kg i.p. fed), or vehicle control (saline fasted and fed) 30 min prior to imaging (Figure 7A–B). Consistent with their opposing effects on 24 hr food intake (Figure S7I–J), leptin increased the latency to retrieve sucrose rewards (Figure 7C), whereas ghrelin reduced it (Figure 7D) without disrupting lick patterning or reward consumption (Figure S7K–L) in both groups. Both LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons altered sucrose encoding after leptin, but in opposing directions. Leptin reduced response magnitude in LHA^Vglut2→LHb neurons and potentiated responses in LHA^Vglut2→VTA neurons (Figure 7E–H). In contrast, ghrelin potentiated LHA^Vglut2→LHb responses but had little effect on the responses of LHA^Vglut2→VTA neurons (Figure 7I–L). This pattern is also reflected in the proportion of neurons with significantly altered responses relative to vehicle control (Figure S7M). Thus, circulating feeding hormones tune reward encoding properties of molecularly defined LHA projection neurons in vivo, suggesting that LHA^Vglut2→LHb and, to a lesser extent, LHA^Vglut2→VTA neurons integrate peripheral signals to orchestrate feeding.

Figure 7: Feeding hormones differentially modify reward encoding in LHA^Vglut2→LHb and LHA^Vglut2→VTA projections.

A. Experimental design. Leptin was injected into fasted mice and ghrelin injected into fed mice.

B. Example fields of view.

C. Leptin increases the latency to lick following reward delivery for both groups (* effect of Leptin: F(1,390)=5.02, p=0.026; no effect of Projection: F(1,390)=0.15, p=0.70; no Interaction: F(1,390)=0.19, p=0.66).

D. Ghrelin reduces the latency to lick following reward delivery for both groups (* effect of Ghrelin: F(1,389)=6.39, p=0.012; no effect of Projection: F(1,389)=0.04, p=0.85; no Interaction: F(1,389)=0.04, p=0.85).

E. Population heatmap showing individual cell responses to vehicle and leptin injections.

F-G. Mean±s.e.m. response of LHA^Vglut2→LHb (F) and LHA^Vglut2→VTA (G) projections during sucrose consumption after leptin administration.

H. Leptin reduces evoked response magnitude in LHb projections and increases response magnitude in VTA projections (effect of Leptin: F(1,370)=10.50, p=0.0013; effect of Projection: F(1,370)=7.60, p=0.006; Interaction: F(1,370)=63.99, p=1.6e-14. * Sidak’s multiple comparisons test, p<0.001). V, vehicle; L, leptin.

I. Population heatmap showing individual cell responses to vehicle and ghrelin.

J-K. Mean±s.e.m. response of LHA^Vglut2→LHb (J) and LHA^Vglut2→VTA (K) projections during sucrose consumption after ghrelin administration.

L. Ghrelin reduces evoked response magnitude in LHb projections (no effect of Ghrelin: F(1,313)=1.56, p=0.21; no effect of Projection: F(1,313)=2.94, p=0.087; Interaction: F(1,313)=14.94, p=0.00013. * Sidak’s multiple comparisons test, p<0.001). V, vehicle; G, ghrelin. Mean+s.e.m.

Discussion

Here, we performed a systematic analysis of lateral hypothalamic glutamate neurons projecting to the LHb or VTA. We demonstrate that LHA^Vglut2→LHb and LHA^Vglut2→VTA projections are anatomically, transcriptionally, electrophysiologically, and functionally distinct from one another. LHA^Vglut2→LHb neurons are commonly found in anterior LHA, are more likely to express Pax6, and are more excitable than LHA^Vglut2→VTA neurons, which tend to reside in posterior LHA, more commonly express Pdyn/Hcrt, and are less excitable. Functionally, both LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons encode appetitive and aversive tastants but encode aversion with greater responses at the concentrations used in this study. Interestingly, VTA projectors show more pronounced quinine responses than LHb projectors. Both LHA^Vglut2→LHb and LHA^Vglut2→VTA projections show reliable excitatory responses to sucrose rewards, which are facilitated by fasting. However, LHA^Vglut2→LHb neurons are more responsive when mice are fed than are LHA^Vglut2→VTA neurons. Finally, feeding hormones differentially influence these populations such that leptin blunts LHA^Vglut2→LHb and potentiates LHA^Vglut2→VTA food reward encoding. In contrast, ghrelin potentiates LHA^Vglut2→LHb encoding but has little effect on LHA^Vglut2→VTA dynamics. Collectively, these results suggest functional divergence in which both LHA^Vglut2→LHb and LHA^Vglut2→VTA projections signal aversive events, but LHA^Vglut2→LHb neurons may play a greater role in guiding feeding.

Transcriptional characterization of LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons

To test whether LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons are molecularly distinct, we combined single-cell mRNA profiling with viral projection labeling. LHA^Vglut2→LHb and LHA^Vglut2→VTA projections tend to fall into transcriptionally distinct neuronal subclusters (Figure 2). Our results agree with previous LHA single-cell sequencing experiments that identified transcriptionally unique LHA subclusters without factoring in projection targets (Mickelsen et al., 2019; Rossi et al., 2019). These subclusters are classified based on the expression of Vgat or Vglut2, which are non-uniformly distributed within the LHA (Figure S2), as well as expression of subcluster specific markers (Wang et al., 2021). LHA^Vglut2→LHb neurons were concentrated in the cluster marked by Vglut2 and Pax6, whereas LHA^Vglut2→VTA neurons were concentrated in the cluster marked by Vglut2, Pdyn, and Hcrt. This pattern was confirmed in situ; however, we cannot conclusively say that the LHA^Vglut2→LHb and LHA^Vglut2→VTA projections further characterized here express Pax6 or Pdyn/Hcrt, respectively. Future studies are needed to test whether LHA^Pax6→LHb and LHA^{Pdyn/Hcrt→VTA} populations are functionally similar to LHA^Vglut2→LHb and LHA^Vglut2→VTA projections. Interestingly, sequencing identified more VTA-projecting than LHb-projecting cells. This is partly due to the relative heterogeneity of VTA projections that could be targeted with the nonspecific viral strategy. Numerous LHA cell types project to the VTA including those expressing GABA, glutamate, hypocretin, neurotensin, and melanin-concentrating hormone (Fadel and Deutch, 2002; Liu et al., 2017; Nieh et al., 2016; Woodworth et al., 2018). In contrast, LHA cells projecting to LHb are thought to be primarily glutamatergic (Lazaridis et al., 2019; Stamatakis et al., 2016). Indeed, our data agree with these observations (Figures 2E and S1J). The specific promoters used to target each pathway (i.e., hSyn or CAG) may also contribute to the increased proportion of VTA-projecting cells recovered (Jackson et al., 2016; Powell et al., 2015).

Notably, LHA^→LHb and LHA^→VTA projections were uniquely distributed within neuronal subclusters. Whereas LHA^→LHb neurons were relatively restricted within a few primarily glutamatergic subclusters, LHA^→VTA neurons were more broadly distributed across GABAergic and glutamatergic clusters (Figure 2E). The LHA^Vglut2 subcluster with the greatest concentration of VTA-projecting cells was marked by Pdyn and Hcrt, the gene encoding orexin/hypocretin. We used Pdyn for further validation because Hcrt was weakly expressed in many subclusters (Figure 2G); however, highly-expressing orexin neurons comprise a sizable subset of LHA^Vglut2→VTA neurons that largely overlaps with the LHA^Vglut2→VTA Pdyn+ population (Mickelsen et al., 2017). Thus, we consider LHA^→VTA neurons expressing Pdyn and Hcrt a single population. However, much research into this cell type has specifically focused on orexin neurons. LHA orexin neurons express Vglut2 (Rosin et al., 2003; also see Figure 2G) and project to the VTA (Fadel and Deutch, 2002; Iyer et al., 2018) where they directly and indirectly regulate dopamine neuron activity (Baimel et al., 2017; Borgland et al., 2008; Korotkova et al., 2003) to influence reward-related behavior and feeding (Harris and Aston-Jones, 2006; Harris et al., 2005; Nakamura et al., 2000; Teegala et al., 2020). Our results agree with known LHA^→VTA diversity (Nieh et al., 2015) as well as the anatomy and function of orexin neurons (de Lecea et al., 1998; Peyron et al., 1998), but future work must disentangle the contributions of LHA^{Pdyn/Hcrt→VTA} neurons from the wider LHA^Vglut2→VTA population.

Functional differences in reward and aversion encoding in LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons

Both LHA^Vglut2→LHb and LHA^Vglut2→VTA projections signal appetitive and aversive stimuli with predominately excitatory responses, though quinine elicits stronger responses than sucrose (Figure 5). Broadly, this aligns with single cell calcium imaging of LHA^Vglut2→LHb neurons (Lazaridis et al., 2019) and fiber photometry of LHA^Vglut2→VTA projections (de Jong et al., 2019) in which foot shocks elicit excitatory responses. We show for the first time that individual LHA^Vglut2 projection neurons encode both appetitive and aversive stimuli. LHA^Vglut2→VTA neurons showed comparatively greater responses to quinine than did LHA^Vglut2→LHb neurons, suggesting that VTA projections may be preferentially tuned for aversive stimuli. Interestingly, LHA^Vglut2→VTA neurons also showed greater responses than LHA^Vglut2→LHb neurons to water rewards when mice were thirsty (Figure S4), although activity of both populations was weaker for water than for sucrose or quinine. These results support the idea that VTA projectors are particularly responsive to highly salient stimuli, with the strongest responses being elicited by stimuli with negative valance. The broad tuning suggests that these neurons are not simply conveying reward or aversion signals but may represent higher level functions such as those associated with terminating ongoing actions, including feeding, consummatory behavior, and reward seeking.

In addition to responding to aversive events, both LHA^Vglut2→LHb and LHA^Vglut2→VTA projections were active during sucrose consumption, and response magnitudes increased when motivation to obtain food was elevated by fasting (Figure 6). In contrast to the observation that LHA^Vglut2→VTA neurons were more responsive to quinine, LHA^Vglut2→LHb neurons were more responsive than LHA^Vglut2→VTA projections to sucrose consumption when mice were fed. However, this difference was eliminated by fasting, suggesting that circuit level or systemic factors may differentially regulate LHA^Vglut2→LHb and LHA^Vglut2→VTA activity in response to satiety.

LHA^Vglut2→LHb and LHA^Vglut2→VTA projections exhibit pronounced electrophysiological differences, which support in vivo imaging results. LHA^Vglut2→LHb neurons were more excitable and more active at rest than LHA^Vglut2→VTA neurons (Figure 4). LHA^Vglut2→VTA neurons spontaneously fired at ~2 Hz (Figure 4H), which is consistent with recordings from identified orexin neurons (Li et al., 2002). LHA^Vglut2→LHb responses peaked sooner after sucrose consumption than did LHA^Vglut2→VTA responses (Figures 6 and 7). Differences in excitability may underlie these unique response profiles. The proportion of sucrose responsive neurons was greater in LHA^Vglut2→LHb than in LHA^Vglut2→VTA projection populations when mice were fed but fasting eliminated this difference (Figure 6F). This was mirrored ex vivo, as fasting reduced electrophysiological dissimilarities between the projection populations (Figure S6).

Hormonal modification of LHA^Vglut2→LHb and LHA^Vglut2→VTA reward encoding

Because LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons were more responsive to sucrose after a fast (Figure 6), we hypothesized that circulating hormones, which are opponently regulated by fasting (Bagnasco et al., 2002), may underlie these changes. Systemic administration of leptin and ghrelin uniquely affected sucrose encoding dynamics in LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons (Figure 7). Much evidence has demonstrated that leptin and ghrelin can influence behavior via the LHA and connected brain regions, yet the effects on neural activity in vivo are largely unknown. Lepr expressing neurons project to the VTA to regulate downstream activity (Liu et al., 2017) and food motivation (Leinninger et al., 2009; Schiffino et al., 2019). While LHA Lepr neurons are often considered exclusively GABAergic (Leinninger et al., 2009), previous reports have demonstrated that a subset of hypothalamic Lepr neurons express glutamatergic markers (Mickelsen et al., 2019; Romanov et al., 2017). Furthermore, leptin depolarizes LHA Lepr neurons and increases accumbens dopamine release in brain slices (Leinninger et al., 2009), which is consistent with our observation that leptin potentiated LHA^Vglut2→VTA sucrose dynamics in vivo. Similarly, intra-LHA ghrelin potentiates feeding (López-Ferreras et al., 2017; Olszewski et al., 2003) and striatal dopamine release (Cone et al., 2014), though it is unknown if LHb projections are involved. Whereas peripheral ghrelin tends to quickly increase food intake (Wang et al., 2014; Wren et al., 2001), leptin tends to reduce food intake more slowly (Barrachina et al., 1997). Despite this, both leptin and ghrelin can produce physiological changes within 30 min (Di Marzo et al., 2001; Wren et al., 2000). Consistently, we observed opposing effects on licking approximately 30 min after peripheral injection (Figure 7C,D), suggesting that head-fixed sucrose consumption may be more sensitive than free feeding assays.

Consistent with previous studies (Mickelsen et al., 2019; Romanov et al., 2017; Rossi et al., 2019), we detected only sparse Lepr and Ghsr expression with single-cell sequencing. We therefore used in situ hybridization to observe Ghsr and Lepr expression on projection neurons (Figure S7). Because LHA^Vglut2→LHb and LHA^Vglut2→VTA neurons express Lepr and Ghsr and because leptin and ghrelin act in LHA, it is possible that the observed changes in sucrose encoding are caused by direct action on the imaged neurons. However, the observed responses likely involve circuit level effects. Most LHA Lepr and Ghsr expression is on cells that were not identified LHb or VTA projections (Figure S7B,E). Furthermore, leptin and ghrelin can influence the electrical activity of other LHA cell types including orexin and melanin-concentrating hormone neurons (Liu et al., 2017; Yamanaka et al., 2003). Additionally, leptin and ghrelin act in both upstream (e.g., arcuate nucleus) and downstream (e.g., VTA) brain regions, so the net effects on neurocircuit dynamics are difficult to predict. Thus, we tested whether feeding hormones influence the net activity of molecularly and anatomically defined LHA projection neurons in vivo and demonstrate for the first time that leptin blunts LHA^Vglut2→LHb and facilitates LHA^Vglut2→VTA sucrose encoding. Conversely, ghrelin facilitates LHA^Vglut2→LHb dynamics while LHA^Vglut2→VTA responses are unaffected (Figure 7). Whether similar responses are observed in other LHA cell types remains to be tested.

Systemic leptin and ghrelin administration failed to fully recapitulate the sucrose encoding dynamics observed in fed and fasted states, respectively. Specifically, ghrelin administration in fed mice did not cause LHA^Vglut2→VTA neurons to respond to sucrose as if mice were fasted, and leptin given to fasted mice did not cause LHA^Vglut2→VTA neurons to respond as if fed. While circulating leptin and ghrelin levels vary with energy state, fasting causes widespread physiological changes extending beyond feeding hormones (Jensen et al., 2013). Thus, other factors that are not present in exogenous hormone manipulation may contribute to the effects of fasting on these pathways. Indeed, many other peripherally derived hunger and satiety signals, including insulin, glucagon, glucagon-like peptide-1, amylin, or cholecystokinin, may influence LHA^Vglut2→LHb and LHA^Vglut2→VTA activity. Future studies will be needed to test how they influence sucrose encoding.

Comparison of projection neurons to the LHA^Vglut2 population

Overall, responses observed in LHA^Vglut2→LHb and LHA^Vglut2→VTA projections are consistent with the known function and activity patterns of the wider LHA^Vglut2 population. Previous reports demonstrated that LHA^Vglut2 neurons encode positive and negative valence stimuli with excitatory responses (Lazaridis et al., 2019; Rossi et al., 2019). Both LHA^Vglut2→LHb and LHA^Vglut2→VTA projection populations, which comprise a subset of the LHA^Vglut2 population, primarily responded to appetitive and aversive stimuli with excitatory responses. Consistent with the proposed role of LHA^Vglut2 neuron activation in halting ongoing appetitive behavior, LHA^Vglut2→LHb and LHA^Vglut2→VTA projections showed elevated quinine responses, which were associated with truncated licking bouts. LHA^Vglut2→LHb and LHA^Vglut2→VTA projections also showed distinct temporal dynamics (Figures 6 and 7), which is in accordance with LHA^Vglut2 response heterogeneity (Lazaridis et al., 2019; Rossi et al., 2019). We speculate that the timing of responses within the LHA-LHb-VTA circuitry may be important for guiding reward related behavior.

In contrast to the general similarities between the LHA^Vglut2 population and the responses of LHb- and VTA-projecting subsets, fasting uniquely affects LHA^Vglut2→LHb and LHA^Vglut2→VTA projections. Both populations are more responsive when mice are fasted, and we previously showed that LHA^Vglut2 neurons are less responsive after feeding (Rossi et al. 2019). Importantly, some LHA^Vlglut2 neurons exhibit greater sucrose responses when fasted. We speculate that the small proportion of LHA^Vglut2 neurons projecting to LHb or VTA are among the cells that increased responding after fasting.

Collectively, LHA^Vglut2 neurons are thought to negatively regulate feeding because food intake is suppressed by bulk LHA^Vglut2 activation (Jennings et al., 2013; Rossi et al., 2019) and facilitated by inhibition (Jennings et al., 2013). Mice avoid LHA^Vglut2 stimulation and prefer LHA^Vglut2 inhibition (Jennings et al., 2013; Lazaridis et al., 2019; Rossi et al., 2019). While activation of either LHA^Vglut2→LHb (Stamatakis et al., 2016) or LHA^Vglut2→VTA (de Jong et al., 2019; Nieh et al., 2016) neurons is aversive, their relationship with feeding is less consistent. LHA^Vglut2→LHb manipulations influence food intake (Stamatakis et al., 2016), but LHA^Vglut2→VTA manipulations do not (Nieh et al., 2015, 2016). Consistently, LHA^Vglut2→LHb projections are probably more involved in guiding feeding because 1) they are sensitive to both leptin and ghrelin (i.e., leptin makes these neurons respond as if mice are sated, and ghrelin makes them respond as if mice are fasted), and 2) LHb projectors are more responsive to food rewards when fed. We therefore hypothesize that both LHA^Vglut2→LHb and LHA^Vglut2→VTA populations work in concert to orchestrate reward- and aversion-related behavior in general, whereas LHA^Vglut2→LHb neurons additionally regulate feeding.

Star Methods

Resource Availability

Lead Contact

Further information and requests for resources and reagents should be directed to Garret Stuber (gstuber@uw.edu).

Materials Availability

This study did not generate new unique reagents.

Data and Code Availability

The NCBI Gene Expression Omnibus accession number for the single-cell RNA sequencing data reported in this paper is GSE169176. All code used to analyze the data presented in this paper are available on the Stuber lab Github repository (www.github.com/stuberlab).

Experimental Model and Subject Details

Mice

Adult male and female mice were used for all experiments. Specific breakdown of age/sex are given for each experiment in the corresponding methods description. For single-cell sequencing and HCR experiments, C57BL/6J mice (Jackson Laboratory) were used. For all other experiments, Vglut2-ires-Cre (Jackson Laboratory strain #028863) (Vong et al., 2011) were used. Unless undergoing a fasting experiment, mice were maintained with ad lib. access to food and water. Mice were maintained on a 12 hr light-dark cycle. All experiments were conducted in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee and the University of Washington.

Method Details

Surgery

Mice were anaesthetized with isoflurane (maintained at <1.5%) and were placed in a stereotaxic frame (Kopf Instruments). Ophthalmic ointment was placed on the eyes, and lidocaine (1 mg/kg) was injected s.c. at the incision site. Following surgery, 5 mg/kg carpofen was administered i.p. All viruses were infused through stainless steel injectors at a rate of 100 nL/min. Injectors were left in place for at least 10 minutes following each infusion. The following coordinates were used (in mm relative to Bregma): −1.35 AP, 0.95 ML, −5.15 DV from brain surface for LHA; −3.08-3.10 AP, 0.55 ML, −4.25 DV from brain surface for VTA; and −1.5 AP, 1.2 ML, −3.15 from skull at 15 degrees from vertical for LHb. After surgery, mice were monitored and allowed to recover for at least two weeks prior to experimental manipulations.

Slice electrophysiology

Vglut2-Cre mice (n = 20, 10 male, 10 female, aged 3-4 months at time of surgery) were injected bilaterally with retroAAV-Ef1α-DIO-mCherry and retroAAV-Ef1α-DIO-eYFP in LHb and VTA (viruses counterbalanced between brain regions; 250-300 nL per injection). Approximately 6 weeks later mice were deeply anesthetized with pentobarbital and transcardially perfused with ice-cold sucrose cutting solution containing (in mM): 75 sucrose, 87 NaCl, 1.25 NaH₂P0₄, 7 MgCl₂, 0.5 CaCl₂, 25 NaHCO₃, 301-307 mOsm. Brains were rapidly removed, and coronal sections were taken through the LHA at 300 μm. Sections were incubated in aCSF at 32 °C containing (in mM): 126 NaCl, 2.5 KCl, 1.2 NaH₂P0₄, 1.2 MgCl₂, 2.4 CaCl₂, 26 NaHCO₃, 15 glucose, 303-308 mOsm. Fluorescence was visualized under 40x objective, and cells located in the LHA and expressing either mCherry or eYFP but not both were targeted for patching.

To assess excitability whole-cell current clamp recordings were performed by pulling borosilicate pipettes at 3.5-5.5 MΩ and backfilling with solution containing (in mM): 135potassium gluconate, 4KCl, 10 HEPES, 4ATP, 0.3GTP (pH 7.35, 285mOsm). Spontaneous synaptic input was assessed with whole-cell voltage clamp recordings with Cs-Methanesulfonate internal solution containing (in mM): 117 Cs-Methanesulfonate, 20 HEPES, 0.4 EGTA, 2.8 NaCl, 5 TEA, 5 ATP, and 0.5 GTP (pH 7.35, 285 mOsm). Recordings lasted 2-5 min at −70 mV and +10 mV holding potential for EPSCs and IPSCs, respectively. For fasting experiments, mice were deprived for 24 hr prior to sacrificing. All other methods were identical.

Single-Cell sequencing

C57BL/6J mice (n = 7, 4 male, 3 female, 6-8 weeks old at time of surgery) were injected bilaterally with retroAAV2-hSyn-eYFP in LHb and retroAAV2-CAG-tdTomato in VTA (300-400 μL). Approximately 6 weeks later, one mouse was used for histological analysis. The remaining six mice were deeply anesthetized with sodium pentobarbital and phenytoin sodium and transcardially perfused with 10 mL ice-cold NMDG-aCSF containing 96 mM NMDG, 2.5 mM KCl, 1.35 mM NaH2PO4, 30 mM NaHCO3, 20 mM HEPES, 25 mM glucose, 2 mM thiourea, 5 mM Na+ascorbate, 3 mM Na+pyruvate, 0.6 mM glutathione-ethyl-ester, 2 mM N-acetyl-cysteine, 0.5 mM CaCl2, 10 mM MgSO4; pH 7.35–7.40, 300-305 mOsm. Six 280 μm sections through the LHA were collected in NMDG-aCSF on a vibratome. Sections recovered in NMDG-aCSF containing 500 nM TTX, 10 μM APV, 10 μM DNQX (NMDG-aCSF-R) at room temperature for 40 minutes. After recovery, LHA tissue was isolated with tissue punches and incubated in NMDG-aCSF-R containing 1.0 mg/mL pronase for 50 minutes at room temperature. Following digestion, tissue from all animals was transferred to 1.0 mL NMDG-aCSF-R supplemented with 0.05% BSA (NMDG-aCSF-BSA), pooled, and mechanically dissociated with a patch pipet fire-polished to an internal diameter of 200-300 μm. The suspension was washed in 12 mL NMDG-aCSF-BSA, sedimented at 220 xg for 6 minutes at 18°C. Dead cells and cellular fragments were removed using a MACS^® Dead Cell Removal Kit (Miltenyi Biotec, Bergisch Gladbach, Germany) using 100 μL of dead cell removal beads and one MACs LS column following the standard manufacturer protocol. Following the final wash, cells were sedimented at 220 xg for 6 minutes at 4°C and resuspended in PBS containing 0.05% nuclease-free BSA for a final concentration of 700-800 cell/μL.

Single-cell RNA-sequencing was performed using Chromium Single Cell 3’ v3 chemistry (10X Genomics, Pleasanton, CA) following manufacturer recommendations unless otherwise noted. Approximately 16,000-18,000 pooled cells were loaded into an individual well of a Single Cell B Chip, with three wells used in total for three reactions. cDNA amplification and sample indexing were each performed using 11 cycles. Library size distribution and concentrations were determined with an Agilent 4200 Tapestation High Sensitivity D5000 ScreenTape and a Thermo Scientific Qubit dsDNA HS Assay, respectively. Each reaction library was sequenced on an individual lane of an Illumina HiSeq 4000 using 2x150 chemistry following the standard 10X Genomics v3 paired-end configuration by Genewiz, Inc. (South Plainfield, NJ). Between approximately 370-430 million reads were generated per lane with a mean quality score >32 across all lanes.

Demultiplexing was performed with 1 mismatch allowed using Illumina bcl2fastq v2.18.0.12. Initial read processing and generation of digital expression matrices was performed using 10X Genomics Cell Ranger v3.1.0 with 128 GB of memory on 24 cores. Sequences were aligned to a modified mm10 3.0.0 genome assembled with eYFP and tdTomato transgenic sequences. The reference sequences for eYFP and tdTomato were taken from Addgene plasmids #26972 and #59462, respectively and included the entire region from the start codon to the end of the 3’ UTR on the (+) strand (Fig S1G). Coverage was assessed with SAMtools v1.10 (Li et al., 2009).

HCR

C57BL/6J mice (n = 3, 1 male, 2 female, ~5 months old at time of surgery) were injected bilaterally with retroAAV2-hSyn-eYFP in LHb and retroAAV2-CAG-tdTomato in VTA (300-400 μL per injection). Six weeks later, mice were anesthetized and rapidly decapitated. Brains were quickly removed and fresh frozen at −80 °C. Sections through the AP extent of the LHA were cut at 20 μm, mounted on slides, and stored at −80 °C. 6-12 sections were quantified per mouse.

A total of nine genes were targeted in three sequential HCR rounds. The probe sets were as follows: Pax6, Sostdc1, eYFP (Round 1); tdTomato, Slc32a1, Slc17a6 (Round 2); and Nptx2, Pdyn, Pitx2 (Round 3). Sections were processed according to the sequential hybridization chain reaction (HCR) protocol (Molecular Instruments) with modifications noted below. Sections were fixed in 4% PFA at 4°C, dehydrated in serial ethanol washes, and treated with RNAscope Protease IV (ACDBio). Sections were then rinsed and hybridized overnight at 37°C with probes targeting the first gene set. Sections were then washed, and probes were amplified with fluorophores (Alexa 488, 546, and 647) overnight at room temperature. Next, sections were washed, and autofluorescence was quenched using a Vector TrueVIEW Autofluorescence Quenching kit (Vector labs cat# SP-8400). DAPI was applied in the first round only. Slides were coverslipped with VECTASHIELD Vibrance Antifade Mounting Medium and allowed to cure at RT for 2 hours before imaging. After the first two rounds of imaging, coverslips were removed, and sections were washed to remove mounting medium. The probes were then digested with DNase I (Sigma cat# 4716728001), and the next probe set was hybridized. Images were acquired using a Zeiss Axio Imager 2 microscope at 10x magnification.

RNAscope in situ hybridization

Vglut2-Cre mice (n = 5, 1 male, 4 female aged ~6wk at time of surgery) were injected bilaterally with retroAAV-Ef1α-DIO-mCherry in VTA and retroAAV-Ef1α-DIO-eYFP in LHb (400 nL per injection). Six weeks later, mice were anesthetized and rapidly decapitated. Brains were quickly removed and fresh frozen at −80 °C. Sections were cut at 20 μm, mounted on slides, and stored at −80 °C. Sections were fixed in 4% PFA at 4°C, dehydrated in ethanol, and processed with RNAscope (Advanced Cell Diagnostics). Sections were hybridized with the following probes: mCherry (cat. #431201), eYFP (cat. #312131), and either Ghsr (cat. #426141) or Lepr (cat. #402731). Following amplification, sections were counterstained with DAPI. Images were taken throughout the anterior-posterior extent of the LHA with a Zeiss Axio Imager 2 microscope at 10x magnification and fluorescence puncta were quantified via HALO software (Indica Labs). Cells were classified as mCherry- or eYFP-positive if they expressed five or more puncta. To assess overlap in Lepr, Ghsr, and Slc32a1 (Vgat), lateral hypothalamus of wild-type mice (n = 3, 2 male, 1 female, aged ~3 months) were processed as above, but sections were hybridized with Lepr, Ghsr, and Slc32a1 (cat #319191) probes.

In vivo 2p imaging

Vglut2-Cre mice (n = 10, 6 male, 4 female, aged 2-4 months at time of surgery) were injected unilaterally with 400 nL retroAAV-Ef1α-mCherry-IRES-FlpO (Addgene #55634) (Fenno et al., 2014) in either the LHb (n = 5) or VTA (n = 5). AAV8-Ef1α-CreOn/FlpOn-GCaMP6m (Addgene #137119) (Fenno et al., 2020) (400 nL) was then injected into the LHA. Next, a 0.6 mm diameter x 7.3 mm length GRIN lens (Inscopix) was slowly lowered into place above the LHA and affixed with glue. Finally, a head fixation ring was placed on the skull with the lens in the center, and everything was secured in place with dental cement. Mice recovered for 5 weeks before testing began.

Mice were acclimated to head fixation for two daily 15-min sessions prior to conducting the experiments. Generally, imaging sessions occurred in two phases beginning ~5 min after mice were headfixed: a 10-min baseline period in which basal calcium dynamics were monitored while mice remained at rest followed by a session in which tastants (~2 μL/trial) were delivered through a spout positioned directly in front of mouse’s mouth. Licks were recorded via a lickometer, and calcium dynamics were recorded for 25 s (125 frames) around the lick delivery. Rarely, mice failed to lick during this window. These trials were excluded from downstream analyses post hoc.

All experiments were conducted within 4 hr of the start of the dark cycle. For fasted experiments, mice were food restricted for 24 hr prior to testing; for fed experiments, mice had ad lib. access to food for at least 24 hr prior to testing. For sucrose/quinine and water tests, mice were water restricted for 24 hr prior to testing. For most sessions, 20 trials were conducted in which 10% sucrose solution was randomly delivered to each mouse without explicit cues. The exceptions to this were the sucrose/quinine session in which 60 trials were performed (15 trials 1.5 mM quinine, 45 trials 10% sucrose solution, random order) and the water session in which 20 trials of tap water were examined. For pharmacology experiments, mice were injected i.p. with 0.3 mL of either 1.5 mg/kg leptin (Sigma cat. #L3772) dissolved in sterile saline, 1.0 mg/kg ghrelin (Tocris cat. #1465) dissolved in sterile saline, or sterile saline alone (vehicle control) 30 min before imaging began each day (conducted in pseudorandom order with washout days after drug injections).

Imaging was conducted using an Olympus multiphoton microscope equipped with resonant scanners and GaAsP PMTs as described previously (Rossi et al., 2019). Calcium dynamics were imaged through the GRIN lens with a 20x air objective (Olympus, LCPLN20XIR, 0.45 NA, 8.3 mm WD). Two-photon excitation was accomplished using a Mai-Tai Deep See laser system (Spectra Physics) set to 920 nm wavelength. Data were collected at 5 Hz with Olympus FluoView software. Prior to each imaging session, FOVs were manually aligned with projections from the previous day.

Following the conclusion of experiments, mice were perfused with 4% PFA and the heads were post fixed for ~48 hr with lenses still in place before the brains were dissected. After post-fixation, brains were sliced coronally at 40 μm, mounted, and imaged with a Zeiss Axio Imager 2 microscope at 10x magnification to confirm lens placement.

Quantification and Statistical Analysis

Data analysis for slice electrophysiology

Data were processed using Clampfit v11.0.3.03 (Molecular Devices) and analyzed using GraphPad Prism v8.3.0. All tests were two-sided and corrected for multiple comparisons or unequal variance where appropriate. Spontaneous activity was assessed via cell-attached recordings and cells were considered spontaneously active if they fired <10 spikes/min. A greater proportion of LHA^Vglut2→LHb (30/30) than LHA^Vglut2→VTA (30/35) cells were spontaneously active (X²=4.64, p=0.03). To test whether electrical properties could be used to decode the projection target, support vector classification was performed using scikit-learn with GridSearch cross validation. Classifications were optimized using linear and rbf kernels with C parameter ranges 10⁻³, 10^−2, 10⁻¹, 10⁰, 10¹, 10²,10³ for both and gamma range 10⁻³, 10⁻², 10⁻¹, 10⁰, 10¹, 10²,10³ using the following neuron properties: whole cell spike rate, rheobase, spike threshold, spike duration, AHP duration, and total spikes elicited during current step injections.

Data analysis for single-cell mRNA sequencing and clustering

Clustering was performed using Seurat v3.1.2 (Stuart et al., 2019) and custom code in R v3.6.1 unless noted. Cells in which ≤ 1,000 unique genes and ≥ 15 percent mitochondrial reads were detected were removed from analysis. Doublets were removed using DoubletDetection v2.5.4 (Adam Gayoso and Jonathan Shor, 2020) on Python 3.7 run in R v3.6.1 via Reticulate v1.14. The first 50 dimensions were used in the doublet classifier with 100 iterations. A voter threshold of 0.5 and p-value threshold of 1⁻¹² were used in the classification. 34,518 cells were retained for the final analysis with a median of 2,250 unique genes and 4,783 unique transcripts, and a mean of 25,640 reads detected per cell.

Filtered data were transformed using a regularized negative binomial regression via SCTransform (Hafemeister and Satija, 2019; Stuart et al., 2019) with percent mitochondrial reads as a latent variable. Variable genes accounting for 20% of all genes in the dataset were used as input. Samples were integrated using canonical correlation analysis reduction (Butler et al., 2018; Stuart et al., 2019) with all genes used as anchor features. Pearson residuals from the regularized negative binomial regression and integration were calculated for all genes. Principal components were calculated on all genes, and the first 70 components were used as input for UMAP (McInnes et al., 2020) with 40 neighbors, a minimum distance of 0.50, and a spread of 0.40 as parameters. Clustering was performed using the Louvain algorithm with multilevel refinement (Rodriguez and Laio, 2014; Rotta and Noack, 2011) with a K parameter of 15 and a resolution of 0.025 (all cells) or 0.88 (neurons).

Feature discovery was performed on the Pearson residuals from the regularized negative binomial regression using a likelihood-ratio test for single-cell data as implemented in Seurat (Hafemeister and Satija, 2019; Macosko et al., 2015; McDavid et al., 2013). Canonical features were used for cluster cell-type assignment. Definitive cluster markers could not be obtained for two neuronal sub-clusters, and these clusters were removed from downstream analysis. Differential expression between eYFP+ and tdTomato+ cells was assessed using a logistic-regression based likelihood ratio test with Bonferroni corrected p values on log-transformed SCT counts for genes found in at least 5% of eYFP+ or tdTomato+ cells via Seurat.

Prepossessing, alignment, generation of digital expression matrices, SCTransform, sample integration, dimensionality reduction, clustering, and feature discovery were run on a Dell blade-based cluster running Linux RedHat Enterprise 7.7. All other steps were performed on an Apple MacBook Pro running macOS 10.15.3.

Data analysis for HCR

Following the final HCR round, images were registered based on the brightfield images using RNAscope HiPlex Image registration software (ACDBio). Cell segmentation (based on the DAPI image) and fluorescent puncta quantification were performed on the registered images using HALO software (Indica Labs). Data were then imported to a Python workflow in which the number of fluorescent puncta required to be classified as expressing a given gene was adjusted based on visual inspection of the expression patterns in tissue, distribution of puncta per cell, and background levels for all cells uniformly. Puncta filters were as follows: Pax6, 4; Sostdc1, 6; eYFP, 6; Vgat, 5; tdTomato, 4; Vglut2, 5; Nptx2, 10; Pdyn, 2; Pitx2, 6. To determine whether RNA expression data from HCR or single-cell sequencing experiments could be used to classify cells as eYFP+ or tdTomato+, SVM based classification was employed as described above. Classifications were optimized using linear and rbf kernels using filtered HCR count data or SCT adjusted sequencing count data corresponding to expression of Pax6, Sostdc1, Vgat, Vglut2, Nptx2, Pdyn, and Pitx2. To prevent the disparity between the number of eYFP+ and tdTomato+ cells from unevenly influencing the classification accuracy, the cell numbers were equalized by randomly selecting a subset from the larger, tdTomato+ population prior to each classification run. Training was performed on a random 50% subset of data for each of 1,000 runs. To assess how the number of genes influenced decoding accuracy, a random subset of 1-6 genes were removed from the training on each run.

Data analysis for in vivo imaging

Tastant trials were concatenated, and data were motion corrected using a planar hidden Markov model via SIMA (Kaifosh et al., 2014). ROIs were hand drawn and manually tracked across imaging sessions using standard deviation and mean projections. Neuropil contamination was removed using FISSA (Keemink et al., 2018), and signals were extracted for subsequent analysis using custom Python code.

Calcium events were identified and quantified as described previously (Rossi et al., 2019). For tastant trials, motion corrected and neuropil decontaminated fluorescence values were aligned to the first lick after tastant delivery and then normalized to the mean fluorescence value across the 3 s window immediately preceding the first consummatory lick (first lick after tastant delivery). Unless otherwise noted, statistical tests and classification were performed using a 6-s consummatory window beginning at the first lick after the tastant was delivered. To determine whether neurons were significantly responsive to tastant delivery, a paired t-test (p < 0.05) was conducted comparing the mean fluorescence in the 3-s window preceding the first lick after delivery and with the mean fluorescence in the 6-s window following the first lick after delivery for all trials. Because lick patterning after detection of sucrose and quinine was different, a shorter 3-s window was used to directly compare sucrose and quinine trials. Area under the curve (AUC) values were calculated for each neuron using the linear trapezoidal method within the consummatory window.

To assess whether calcium activity was correlated with lick rate, the normalized fluorescence values across the 6-s consummatory window on each trial were compared with the lick rate during the same period with a Pearson correlation. To test whether calcium dynamics of individual neurons could be used to decode the mouse’s satiety state or tastant delivered, support vector classification was performed using scikit-learn with GridSearch cross validation (Otis et al., 2017; Rossi et al., 2019). Classifications were optimized using linear and rbf kernels with C parameter ranges 10⁻³, 10^−2, 10⁻¹, 10⁰, 10¹, 10²,10³ for both and gamma range 10⁻³, 10⁻², 10⁻¹, 10⁰, 10¹, 10²,10³ using the normalized fluorescence values during the 6-s consummatory window.

Key Resources Table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Bacterial and Virus Strains
AAV8-Ef1α-CreOn/FlpOn-GCaMP6m	Fenno et al., 2020	Addgene Cat# 137119
retroAAV-Ef1α-mCherry-IRES-FlpO	Fenno et al., 2014; Addgene	Addgene Cat# 55634
retroAAV-Ef1α-DIO-mCherry	Addgene (packaged by UNC Vector Core)	Addgene Cat# 50462
retroAAV-Ef1α-DIO-eYFP	Addgene (packaged by UNC Vector Core)	Addgene Cat# 27056
retroAAV2-HSyn-eYFP	Deisseroth lab (packaged by UNC Vector Core)	UNC vector core (rgAAV-hSyn-eYFP)
retroAAV2-CAG-tdTomato	Addgene (packaged by UNC Vector Core)	Addgene Cat# 59462
Chemicals, Peptides, and Recombinant Proteins
RNAscope Protease IV	ACDBio	Cat# 322336
RNAscope DAPI	ACDBio	Cat# 320858
Vector TrueVIEW Autofluorescence Quenching kit	Vector Labs	Cat# SP-8400
DNase I	Sigma	Cat# 4716728001
Leptin	Sigma	Cat# L3772
Ghrelin	Tocris	Cat# 1465
Critical Commercial Assays
Sequential HCR in situ hybridization kit	Molecular Instruments	HCR v3.0 Bundle
RNAscope Multiplex Reagent Kit	ACDBio	Cat# 320850
Chromium Single Cell 3’ Library & Gel Bead Kit v3	10x Genomics	Cat# 1000092
Chromium i7 Multiplex Kit	10x Genomics	Cat# 120262
Chromium Single Cell B Chip	10x Genomics	Cat#1000154
Chromium Controller & Accessory Kit	10x Genomics	Cat# 120223
Dead Cell Removal Kit	Miltenyi Biotec	Cat# 130-090-101
Illumina HiSeq 4000	Illumina	Cat# N/A
Deposited Data
Raw sequencing data	This paper	GEO: GSE169176
Experimental Models: Organisms/Strains
Mouse: C57BL/6J	Jackson Laboratory	JAX: 000664; IMSR_JAX: 000664
Mouse: Vglut2-ires-Cre	Vong et al., 2011	JAX: 028863; IMSR_JAX: 028863
Oligonucleotides
RNAscope mCherry probe	ACDBio	Cat# 431201
RNAscope eYFP probe	ACDBio	Cat# 312131
RNAscope GHSR probe	ACDBio	Cat# 426141
RNAscope Lepr probe	ACDBio	Cat# 402731
RNAscope Slc32a1 probe	ACDBio	Cat# 319191
HCR Pax6 probe	Molecular Instruments	Accession# NM_001244198.2
HCR Sostdc1 probe	Molecular Instruments	Acession# NM_025312.3
HCR eYFP probe	Molecular Instruments	N/A
HCR tdTomato probe	Molecular Instruments	N/A
HCR Slc32a1 probe	Molecular Instruments	Acession# NM_009508
HCR Slc17a6 probe	Molecular Instruments	Acession# NM_001286502.1
HCR Nptx2 probe	Molecular Instruments	Acession# NM_016789.3
HCR Pdyn probe	Molecular Instruments	Acession# NM_011098.4
HCR Pitx2 probe	Molecular Instruments	Acession# NM_011098.4
Software and Algorithms
RNAscope HiPlex Image registration software	ACDBio	Cat# 300065
ImageJ (FIJI)	Imagej.nih.gov	Imagej.net ; RRID: SCR_002285
R	N/A	r-project.org ; RRID: SCR_001905
Seurat v3.1.2	Stuart et al., 2019	Satijalab.org ; RRID: SCR_016341
DoubletDetection v2.5.4	Gayoso et al., 2020	Github.com/JonathanShor/DoubletDetection; RRID: N/A
bcl2fastq v2.18.0.12	Illumina	Illumina.com; RRID: SCR_015058
Cell Ranger v3.1.0	10x Genomics	10xGenomics.com; RRID: SCR_017344
SAMtools v1.10	Li et al., 2009	Htslib.org; RRID: SCR_002105
pClamp v11.0.3	Molecular Devices	Moleculardevices.com ; RRID: SCR_011323
GraphPad Prism v9.0.0	GraphPad Software	Graphpad.com ; RRID: SCR_002798
Python v2.7	N/A	Python.org ; RRID: SCR_008394
SIMA v1.3.2	Kaifosh et al., 2013	Losonczylab.org
FISSA v0.7.2	Keemink et al., 2018	Github.com/Rochefort-lab/fissa
Scipy v1.2.1	SciPy.org	SciPy.org ; RRID: SCR_008058
Scikit-learn v0.20.3	Pedregosa et al., 2011	Scikit-learn.org; RRID: SCR_019053
HALO	Indica Labs	Indicalab.com ; RRID: SCR_018350