Hypothalamic control of interoceptive hunger

Summary

While energy balance is critical to survival, many factors influence food intake beyond caloric need or “hunger.” Despite this, some neurons that drive feeding in mice are routinely referred to as “hunger neurons,” whereas others are not. To understand how specific hypothalamic circuits control interoceptive hunger, we trained mice to discriminate fasted from sated periods. We then manipulated three hypothalamic neuronal populations with well-known effects on feeding while mice performed this task. While activation of ARC^AGRP neurons in sated mice caused mice to report being food-restricted, LH^VGAT neuron activation or LH^VGLUT2 neuron inhibition did not. In contrast, LH^VGAT neuron inhibition or LH^VGLUT2 neuron activation in fasted mice attenuated natural hunger, whereas ARC^AGRP neuron inhibition did not. Each neuronal population evoked distinct effects on food consumption and reward. After satiety- or sickness-induced devaluation, ARC^AGRP neurons drove calorie-specific feeding, while LH^VGAT neurons drove calorie-indiscriminate food intake. Our data support a role for ARC^AGRP neurons in homeostatic feeding and implicate them in driving a hunger-like internal state that directs behavior toward caloric food sources. Moreover, manipulations of LH circuits did not evoke hunger-like effects in sated mice, suggesting that they may govern feeding more related to reward, compulsion, or generalized consumption than to energy balance but also that these LH circuits can be powerful negative appetite modulators in fasted mice. This study highlights the complexity of hypothalamic feeding regulation and can be used as a framework to characterize how other neuronal circuits affect hunger and identify potential therapeutic targets for eating disorders.

eTOC

Using a behavioral paradigm for mice to discriminate feelings of “hunger” from satiety, Siemian et al. investigate the contributions of hypothalamic neuronal subtypes to interoceptive hunger. They show that arcuate nucleus AGRP neurons drive a hunger-like internal state, whereas GABAergic and glutamatergic lateral hypothalamic neurons do not.

Introduction

Regulating food intake is essential for survival. Food-oriented behavior in basic research is typically presented in terms of homeostatic feeding—food intake necessary to maintain normal body weight—or hedonic feeding—food intake motivated by gustatory sensation or reward^1,2. These distinctions are clearly applicable to humans, who eat to satisfy homeostatic energy demands but also for pleasure, such as ingesting desserts high in fat and sugar even after a large meal. Thus, understanding the neuronal circuits controlling homeostatic and hedonic feeding may help to address diseases associated with overeating and undereating such as obesity and anorexia nervosa, respectively.

Many neuronal circuits can simultaneously influence food intake and reward, suggesting extensive overlap between homeostatic and hedonic feeding systems¹. For example, activation of γ-aminobutyric acid (GABA)-expressing neurons in the lateral hypothalamus (LH; LH^VGAT neurons) evokes food intake and is rewarding³, whereas activation of LH glutamatergic (LH^VGLUT2) neurons decreases food intake and is aversive⁴. Further downstream, the activation of ventral tegmental area (VTA) GABAergic neurons inhibits reward consumption and drives aversion^5,6, while VTA dopaminergic neurons promote food seeking and reward⁷. Yet among these and other cell populations that modulate feeding, only arcuate hypothalamic agouti-related peptide-expressing (ARC^AGRP) neurons are routinely referred to as homeostatic feeding-related “hunger” neurons^8–16. This anthropomorphic terminology may have resulted from some interesting properties that ARC^AGRP neurons possess. For example, their activity increases during periods of caloric deficit and decreases in a calorie-dependent manner following the detection and ingestion of a food source^{10,14,17–19}. Moreover, activation of ARC^AGRP neurons recapitulates many of the behavioral and neuronal changes induced by food restriction^{9,11,15,20–25}. However, even the activity of these “homeostatic” feeding neurons has not been dissociated from overlapping reward processes. One study found that ARC^AGRP neuronal activation imparts a negative valence, since mice become averse to flavors and places repeatedly paired with it¹⁰. This finding predicted that the inhibition of ARC^AGRP neurons by food detection and consumption transmits a teaching signal via negative reinforcement (the removal of the negative stimulus). However, the inability of food-restricted mice to acquire simple operant tasks to cease ARC^AGRP neuronal stimulation does not support negative reinforcement as the primary driver of their effects on food intake^8,10. Alternatively, another study suggested that ARC^AGRP neurons drive feeding through a long-lasting positive valence signal, as mice develop a preference for flavors and foods accessible following pre-stimulation of ARC^AGRP neurons (and not concurrent stimulation)⁸. Moreover, mice self-administered short bursts of optogenetic activation of ARC^AGRP neurons when food was also accessible⁸. ARC^AGRP neuronal activation was also recently shown to potentiate a dopamine response to food, further depicting the overlap between circuits for homeostasis and reward²⁶. Thus, while these studies together may form a unified model to account for the contribution of ARC^AGRP neuronal activity to food-oriented behavior, their findings do more to characterize how ARC^AGRP neurons influence reward rather than hunger.

“Hunger” is a subjective experience, and in rodents is an inferred motivational state generally defined according to the operation that produces it—food deprivation—or the effect it produces—food intake²⁷. While humans can easily answer whether they are “hungry” or “sated,” the same does not apply to rodents. Mice and rats eat for reasons other than food deprivation^28–30 and feeding does not always follow food deprivation^31–35. Therefore, measures of food intake do not distinguish hunger from the multiple factors that affect feeding and, by extension, ARC^AGRP neuronal activity may not truly encode for hunger. Furthermore, as described nearly 50 years ago, hypothalamic stimulation-induced behavior tends to be unusually rigid²⁷. For instance, rats do not switch to consume a different food when an LH stimulation-paired food is removed³⁶, and LH^VGAT neuron activation in mice can evoke consummatory behavior even when food is not available³. Similar findings have been reported for ARC^AGRP neuron stimulation, which drives stereotypic behaviors beyond feeding in the absence of food such as locomotion, digging, grooming, and marble burying³⁷. Therefore, to distinguish homeostatic from hedonic feeding-related circuits, and whether these circuits mediate hunger-like states, calls for investigations of behavior extending beyond simple food intake have been raised¹.

Drug discrimination experiments have been used for decades to understand the interoceptive effects of pharmacological compounds and to gain a more detailed understanding of their actions that extend beyond their simple reinforcing effects^38,39. For example, both cocaine and heroin are readily self-administered^40–43, but they do not produce a common discriminative stimulus^44,45. That is, each drug triggers a reinforcing, yet distinct, interoceptive experience. A similar scenario may exist among neuronal populations that drive feeding, where some evoke a sensation of hunger and others do not. However, only a small number of studies have used discrimination techniques to examine differences in the interoceptive features produced by food satiety or food restriction^46–48. In these studies, some pharmacological manipulations that alter food intake in rats also change the interoceptive cue associated with food restriction or satiety, whereas others do not, suggesting divergent feeding-associated drives^49–52. However, such methods have not been used to understand the subjective effects associated with cell type- or circuit-specific manipulations that evoke or suppress feeding. Here we apply an operant “hunger discrimination” paradigm in combination with optogenetics and chemogenetics to determine the roles of ARC^AGRP, LH^VGAT, and LH^VGLUT2 neurons in regulating interoceptive hunger cues in mice. In addition, we directly compare the feeding- and reward-associated effects of these neuronal populations and assess whether they drive calorie-specific or indiscriminate feeding.

Results

Hypothalamic control of hunger- and satiety-associated interoceptive cues

We first determined whether mice could discriminate periods of food restriction (22-h food deprivation, ‘fasted’) from periods of food satiety (1-h food deprivation, ‘sated’). On fasted training days, food was removed from the home cage 22 h before the operant session. Responses on one designated active lever (e.g., left) were reinforced with sucrose pellets (fixed ratio 15, 5 pellets maximum), whereas responses on the opposite lever reset the response requirement on the active lever. On sated training days, food was removed from the home cage 1 h before the operant session, and the lever contingencies were reversed. Responses on the opposite lever (e.g., right) were then reinforced with sucrose pellets, whereas responses on the other lever reset the requirement on the now-active lever (Figure 1A). Sated and fasted sessions were administered according to a generally double alternating schedule (i.e., sated, sated, fasted, fasted)⁵³. Within the first 50 sessions, average responding of all mice approached the 80% threshold for feeding status-appropriate responding (Figure 1B). Individual mice qualified for testing by performing greater than 80% appropriate responding in 5 consecutive or 6 out of 7 consecutive sessions^47,49–57 (Figure 1C). The mean number of sessions until meeting these qualification criteria was 60 (median: 50, range: 17 – 182, Figure 1D). Mice were then given two free-choice test sessions, one in the sated condition and one in the fasted condition, during which sucrose pellets could be earned by responses at either lever to verify stimulus control (Figures 1A and E; top panel) and engagement in the task during both sessions (Figure 1E; bottom panel). Following these tests, mice were randomly assigned within each genotype to receive Cre recombinase-dependent experimental (ChR2-YFP + hM4D-mCherry) or control (YFP + mCherry) viral cocktail injections in the ARC of Agrp^Cre mice^58,59 or in the LH of Vgat^Cre and Vglut2^Cre mice⁶⁰ with optical fibers implanted above the ARC or LH, respectively (Figures 2A–C). Using this previously validated dual-virus approach⁶¹, we next examined the effects of activating and inhibiting these specific hypothalamic populations on interoceptive hunger.

Figure 1. Discriminative control by feeding condition in mice.

(A) Schematic of the operant hunger discrimination paradigm. During training sessions, responses on one designated active lever depending on feeding condition were reinforced (e.g., 22-h food-restricted, left lever active; 1-h food-restricted, right lever active) with sucrose pellets on an FR15 schedule, whereas responses on the opposite lever reset active lever requirements. During testing sessions, responses on both levers were reinforced. All sessions lasted for 20 min or until five pellets were earned.

(B) Group-averaged data for all mice (n = 35) on the first 25 sessions of each trial type. Condition-appropriate responding increased as training progressed.

(C) Mice qualified for training by performing >80% responses on the condition-appropriate lever (shown as dotted lines) for five consecutive sessions or six out of seven consecutive sessions. This graph shows training data from a representative mouse during training sessions 40 through 56. On sessions 51 through 56, the mouse exceeded the 80% threshold each session and qualified for testing.

(D) Acquisition curve depicting the number of sessions required for each mouse to reach training criteria. (n = 35 mice)

(E) Performance of mice during free-choice test sessions (n = 12 Agrp^Cre mice, 12 Vgat^Cre mice, and 11 Vglut2^Cre mice). Upper panels, condition-appropriate lever responding; lower panels, response rate.

Figure 2. Hypothalamic control of hunger- and satiety-associated interoceptive cues.

(A) Surgery schematic for experimental and control viral cocktail injections into the ARC of Agrp^Cre mice or the LH of Vgat^Cre and Vglut2^Cre mice.

(B) Representative images showing optical fiber placement in the ARC of Agrp^Cre mice or in the LH of Vgat^Cre and Vglut2^Cre mice. Scale bar = 500 μm.

(C) 20× confocal images demonstrating colocalization of ChR2-YFP (green) and hM4D-mCherry (red) in the ARC of Agrp^Cre mice and in the LH of Vgat^Cre and Vglut2^Cre mice. Scale bar = 50 μm.

(D) ARC^AGRP activation in sated mice triggers hunger. Upper panel, one-way repeated-measures ANOVA revealed a significant effect of photostimulation in ARC^AGRP:ChR2 mice (n = 6; F(1.62, 8.11) = 14.12, p = 0.003; Dunnett’s post-test: no stim vs. 10 Hz,*p = 0.022; no stim vs. 20 Hz, ***p = 0.001) on fasted-associated lever responding. No significant effects were observed in YFP controls (n = 6; paired t test: t(5) = 0.53, p = 0.62). Lower panel, no effects of photostimulation on response rate were observed in ARC^AGRP:ChR2 (F(1.98, 9.92) = 0.29, p = 0.75) or ARC^AGRP:YFP mice (t(5) = 0.37, p = 0.73).

(E) ARC^AGRP inhibition in fasted mice does not decrease hunger. Upper panel, two-way mixed-model ANOVA found no significant effects of ARC^AGRP inhibition on fasted-associated lever responding (n = 6 mice per group; group × treatment interaction: F(1, 10) = 1.81, p = 0.21) or response rate (lower panel, F(1, 10) = 0.70, p = 0.42).

(F) LH^VGAT activation in sated mice does not evoke hunger. Upper panel, no significant effects of photostimulation in LH^VGAT:ChR2 mice (n = 6; F(1.71, 8.56) = 4.32, p = 0.055) or LH^VGAT:YFP mice (n = 6; t(5) = 1.61, p = 0.17) were observed. Lower panel, no effects of photostimulation on response rate were observed in LH^VGAT:ChR2 (F(2.39, 11.95) = 2.59, p = 0.11) or LH^VGAT:YFP mice (t(5) = 0.82, p = 0.45).

(G) LH^VGAT inhibition in fasted mice decreases hunger. Upper panel, two-way mixed-model ANOVA found a significant effect of LH^VGAT inhibition on fasted-associated lever responding (n = 6 mice per group; group × treatment interaction: F(1, 10) = 10.47, p = 0.0089; Bonferroni’s post-test **p = 0.0013). Lower panel, no effects of inhibition were observed on response rate (F(1, 10) = 0.18, p = 0.68).

(H) LH^VGLUT2 activation in fasted mice decreases hunger. Upper panel, one-way repeated-measures ANOVA revealed a significant effect of photostimulation in LH^VGLUT2:ChR2 mice (n = 5; F(2.08, 8.31) = 7.06, p = 0.016; Dunnett’s post-test: no stim vs. 2 Hz,*p = 0.042; no stim vs. 5 Hz,*p = 0.016; no stim vs. 10 Hz,*p = 0.016; no stim vs. 20 Hz, *p = 0.015) on fasted-associated lever responding. No significant effects were observed in YFP controls (n = 6; paired t test: t(5) = 0.58, p = 0.59). Lower panel, photostimulation significantly decreased response rate in LH^VGLUT2:ChR2 (F(2.07, 8.28) = 10.60, p = 0.005; Dunnett’s post-test: no stim vs. 20 Hz, *p = 0.013) but not LH^VGLUT2:YFP mice (t(5) = 0.92, p = 0.40).

(I) LH^VGLUT2 inhibition in sated mice does not induce hunger. Upper panel, two-way mixed-model ANOVA found no significant effects of LH^VGLUT2 inhibition on fasted-associated lever responding (n = 5 LH^VGLUT2:ChR2 and n = 6 LH^VGLUT2:YFP; group × treatment interaction: F(1, 9) = 0.0035, p = 0.95) or response rate (lower panel, F(1, 9) = 1.20, p = 0.30).

See also Figure S1.

Following surgical recovery and brief retraining, mice were exposed to free-choice test sessions while we manipulated neuronal activity with optogenetics or chemogenetics to assess the potential bidirectional effects of these three specific hypothalamic cell types on the perceived feeding status. Photostimulation of sated ARC^AGRP:ChR2/hM4D mice significantly increased fasted-associated lever responding in a frequency-dependent manner (Figure 2D; top panel). At the highest frequency tested (i.e., 20 Hz), and the one most commonly used in ARC^AGRP studies^8,10,15,23, mice exceeded the 80% fasted-appropriate responding threshold, indicating the induction of an interoceptive discriminative stimulus similar to the 22-h fasted condition. However, chemogenetic inhibition of ARC^AGRP neurons in fasted ARC^AGRP:ChR2/hM4D mice via 1 h CNO pretreatment (1 mg/kg, i.p.) did not significantly decrease fasted-associated lever responding (Figure 2E; top panel). In a separate test, 4 h CNO pretreatment also did not significantly decrease fasted-associated lever responding (83.11 ± 7.78%). Neither ARC^AGRP neuronal activation nor inhibition affected the response rate (Figures 2D–E; bottom panels), suggesting that the changes in lever responding during manipulations of ARC^AGRP neurons were not triggered by changes in motor coordination or motivation to obtain the sucrose pellets. Furthermore, no changes in lever choice or response rate were observed during photostimulation or clozapine N-oxide (CNO) injection in ARC^AGRP:YFP/mCherry control mice (Figures 2D–E; top and bottom panels), demonstrating no off-target effects of these manipulations^62–64. Thus, while ARC^AGRP neurons appear sufficient to evoke a hunger-like interoceptive state in sated mice, they are not necessary for maintaining this state in fasted mice.

Contrasting results were observed during manipulations of lateral hypothalamic neurons. Photostimulation of sated LH^VGAT:ChR2/hM4D mice did not significantly increase fasted-associated lever responding, and these mice did not approach 80% fasted-lever responding at any of the photostimulation frequencies tested (Figure 2F; top panel). However, chemogenetic inhibition of LH^VGAT neurons in fasted LH^VGAT:ChR2/hM4D mice did significantly decrease fasted-associated lever responding (Figure 2G; top panel). Neither manipulation affected the response rate in these sessions, and no changes in lever responding or response rate in LH^VGAT:YFP/mCherry control mice were observed (Figures 2F–G; top and bottom panels). These results suggest that LH^VGAT neurons are not sufficient to evoke hunger in sated mice but are necessary for maintaining a hunger-like interoceptive state in fasted mice.

We last examined the effects of LH^VGLUT2 neuronal manipulations in this task. Since LH^VGLUT2 neuron activation has previously been shown to decrease food intake^4,9, we photostimulated fasted LH^VGLUT2:ChR2/hM4D mice during testing. Strikingly, we observed that all of the photostimulation frequencies tested significantly decreased fasted-associated lever responding (Figure 2H; top panel). However, higher frequencies also decreased the response rate (Figure 2H; bottom panel), suggesting effects on motivation or motor coordination, or the induction of other aversive-like effects⁴. However, no effects of chemogenetic inhibition of these neurons in sated LH^VGLUT2:ChR2/hM4D mice were observed (Figure 2I; top panel), and neither manipulation affected the behavior of LH^VGLUT2:YFP/mCherry control mice (Figures 2H–I; top and bottom panels). Thus, LH^VGLUT2 neurons appear to be potent appetite suppressors, but at higher stimulation frequencies, they also seem to disrupt normal operant behavior. Together, these results indicate that the activity of LH^VGLUT2 neurons is sufficient but not necessary for satiety-like interoceptive states.

To further characterize this operant model in mice, we also tested several pharmacological compounds previously shown to alter food intake in rodents^58,65–70 and/or cause weight loss in humans^71,72. Similar to the findings of previous studies in rats^49–52, some of these compounds (the orexigenic hormone ghrelin and the anorexigenics liraglutide or combination of naltrexone/bupropion) significantly changed feeding status-appropriate responding, whereas others (the anorexigenics rimonabant, lorcaserin, and or phentermine) did not (Figure S1). Together, these results show that out of many neuronal or pharmacological manipulations previously shown to alter food intake, only some appear to affect interoceptive hunger. Thus, simple measurements of food intake are likely an unreliable indicator of interoceptive hunger.

Hypothalamic control of food intake and reward-related behavior

We next directly compared the effects of these three hypothalamic populations on feeding and reward-related behaviors. In a context distinct from the previous operant experiments, we first measured the free-access intake of food pellets that have identical composition to standard chow but novel shape and texture (Figure 3A). The novelty of the test chamber and food, combined with testing during the light cycle, was by design to make baseline food intake very low. We found that ARC^AGRP neuronal activation in sated mice triggered significant intake of the novel food pellets in the novel context (Figure 3B), and that this effect became more robust during a second photostimulation test (Figure 3C) following three days of habituation to the food pellets and context without photostimulation. Chemogenetic inhibition of ARC^AGRP neurons in food-restricted mice decreased intake of home cage standard chow by 4 h post-food access (Figure 3D), indicating that inhibition of these neurons may cause an effect more similar to early satiety than a general suppression of feeding. CNO did not affect feeding in control mice (Figure 3D), suggesting that no off-target effects occurred⁶³. In the real-time place preference assay, activation of ARC^AGRP neurons did not evoke significant rewarding or aversive effects (Figures 3E–F and S2A). Thus, ARC^AGRP neuronal activity bidirectionally modulates food intake but does not appear to affect reward-related behaviors in the absence of food.

Figure 3. Hypothalamic control of food intake and reward-related behavior.

(A) Flowchart for testing Agrp^Cre and Vgat^Cre mice in assays of food intake and reward. Blue arrow segments indicate tests with optogenetic activation of neurons and red arrow segments indicate tests with chemogenetic inhibition of neurons.

(B) ARC^AGRP activation in sated mice triggered consumption of novel food pellets in a novel context (two-way mixed-model ANOVA group × epoch interaction: F(2, 20) = 8.32, p = 0.0023; Bonferroni’s post-test: **p = 0.001) as well as (C) food pellet consumption following habituation (F(2, 20) = 42.81, p < 0.0001; Bonferroni’s post-test: ****p < 0.0001). n = 6 mice per group.

(D) ARC^AGRP inhibition in food-restricted mice decreased standard chow intake by 4 h post-food access (three-way mixed-model ANOVA revealed a significant main effect of time (F(2,18) = 193.9, p < 0.0001), with significant interactions of time × group (F(2, 18) = 5.07, p = 0.18) and treatment × group (F(1, 9) = 8.73, p = 0.16). Tukey’s post-test revealed between-group differences in food intake during CNO treatment at 4 h post-food access, *p = 0.015). n = 6 ARC^AGRP:hM4D and n = 5 ARC^AGRP:YFP.

(E) ARC^AGRP activation did not evoke reward- or aversive-like effects in the real-time place preference test (n = 6 mice per group; two-way mixed-model ANOVA group × chamber interaction: F(2, 20) = 2.54, p = 0.10).

(F) Group average location RTPP heat maps for ARC^AGRP:ChR2 (n = 6) and ARC^AGRP:YFP (n = 6) mice. Scale is percentage of time spent in location.

(G) LH^VGAT activation in sated mice did not evoke food pellet intake under novel food and context conditions (F(2, 20) = 0.85, p = 0.44) and (H) only triggered intake post-habituation (F(2, 20) = 9.74, p = 0.001; Bonferroni’s post-test: ***p = 0.0001). n = 6 mice per group.

(I) LH^VGAT inhibition in food-restricted mice decreased standard chow intake throughout the 4 h feeding test (three-way mixed-model ANOVA revealed significant main effects of time (F(2,20) = 391.5, p < 0.0001), treatment (F(1, 10) = 64.00, p < 0.0001), and group (F(1, 10) = 15.59, p = 0.0027), with significant interactions of time × treatment (F(2, 20) = 19.01, p < 0.0001) and treatment × group (F(1, 10) = 45.02, p < 0.0001). Tukey’s post-test revealed between-group differences in chow intake during CNO treatment at 1-, 2-, and 4-h post-food access (****p < 0.0001). n = 6 mice per group.

(J) LH^VGAT activation triggered significant reward-like effects in the real-time place preference test (n = 6 mice per group; two-way mixed-model ANOVA group × chamber interaction: F(2, 20) = 45.13, p < 0.0001). Bonferroni’s post-test revealed between-group differences in time spent in the stimulation-paired and -unpaired chambers (****p < 0.0001) but no differences in the time spent in the hallway compartment (p = 0.25).

(K) Group average location RTPP heat maps for LH^VGAT:ChR2 (n = 6) and LH^VGAT:YFP (n = 6) mice. Scale is percentage of time spent in location.

(L) Flowchart for testing Vglut2^Cre mice. Blue arrow segments indicate tests with optogenetic activation of neurons and red arrow segments indicate tests with chemogenetic inhibition of neurons.

(M) LH^VGLUT2 activation in food-restricted mice significantly decreased standard chow intake (F(2, 16) = 5.88, p = 0.012; Bonferroni’s post-test: **p = 0.0035). n = 5 LH^VGLUT2:ChR2 and n = 6 LH^VGLUT2:YFP.

(N) LH^VGLUT2 inhibition in sated mice did not change standard chow intake over 4 h (three-way mixed model ANOVA only revealed a significant main effect of time: F(2, 16) = 160.5, p < 0.0001) but (O) did increase intake of a novel non-caloric gel (F(1, 8) = 5.98, p = 0.04; Bonferroni’s post-test: **p = 0.0025). n = 5 LH^VGLUT2:ChR2 and n = 6 LH^VGLUT2:YFP.

(P) LH^VGLUT2 activation triggered significant aversive-like effects in the real-time place preference test (n = 5 LH^VGLUT2:ChR2 and n = 6 LH^VGLUT2:YFP; two-way mixed-model ANOVA group × chamber interaction: F(2, 16) = 21.93, p < 0.0001). Bonferroni’s post-test revealed between-group differences in time spent in the stimulation-paired and -unpaired chambers (****p < 0.0001) but no differences in the time spent in the hallway compartment (p = 0.86).

(Q) Group average location RTPP heat maps for LH^VGLUT2:ChR2 (n = 5) and LH^VGLUT2:YFP (n = 6) mice. Scale is percentage of time spent in location.

See also Figures S2–S4.

By contrast, LH^VGAT neuronal activation did not evoke consumption of the novel food pellets (Figure 3G), only triggering intake following habituation (Figure 3H). Moreover, inhibition of LH^VGAT neurons in food-restricted mice decreased home cage standard chow intake across the entire testing period (Figure 3I), indicating a more generalized suppressive effect on feeding as compared to ARC^AGRP neuronal inhibition. Furthermore, photostimulation of LH^VGAT neurons also triggered significant rewarding effects in the real-time place preference assay (Figures 3J–K and S2B), consistent with previous studies showing the robust effects on reward-like behavior mediated by these neurons³.

Activation of LH^VGLUT2 neurons in food-restricted mice significantly decreased standard chow intake (Figures 3L–M) in agreement with previous findings⁴. Chemogenetic inhibition of LH^VGLUT2 neurons in sated mice did not alter standard chow intake in sated or food-restricted mice (Figures 3N and S3A) nor did it affect the intake of several other caloric and palatable food sources tested (Figures S3B–D), in contrast to previous studies that reported slight yet significant increases in food intake during LH^VGLUT2 inhibition^4,9. We speculated that LH^VGLUT2 neuronal inhibition may drive the consumption of food that mice would normally not eat in large amounts. With access to a palatable non-caloric gel, LH^VGLUT2 inhibition increased gel intake, whereas control mice ate very little (Figure 3O), perhaps reflecting the negative feedback normally occurring during ingestion of such a calorie-free food source. Moreover, we observed that photostimulation of LH^VGLUT2 neurons triggered significant real-time place avoidance (Figures 3P–Q and S2C), consistent with previous reports⁴.

To characterize any potential overt behavioral changes, locomotor impairments, or alterations in anxiety-like behavior triggered by the functional inhibition of these hypothalamic populations, we tested sated mice in the open field paradigm. Chemogenetic inhibition of ARC^AGRP neurons increased time spent in the center of the apparatus but did not affect total distance or maximum speed (Figures S4A–D), indicating anxiolytic-like effects consistent with previous findings²². LH^VGAT neuronal inhibition decreased both total distance and maximum speed (Figures S4E–H), whereas inhibition of LH^VGLUT2 neurons increased total distance (Figures S4I–L). These results support the opposing behavioral roles of these LH populations^3,4 and suggest that locomotor activity, in addition to food intake and reward, is under opposing regulation by LH^VGAT and LH^VGLUT2 circuits.

Hypothalamic feeding circuits are distinguished by calorie-specific versus indiscriminate food intake

Since activation of both ARC^AGRP and LH^VGAT neurons triggers food intake^3,8,10,15,25, and we have demonstrated here that only ARC^AGRP activation evokes a hunger-like interoceptive cue, we last compared the effects of activating these two specific hypothalamic cell types on caloric and non-caloric food intake under states of satiety- or sickness-induced devaluation. We predicted that the hunger-like, homeostatic drive evoked by ARC^AGRP activation would direct behavior specifically toward caloric food under conditions of devaluation, whereas the motivational drive evoked by LH^VGAT neuronal activation would drive indiscriminate food intake regardless of caloric content. It was previously shown that ARC^AGRP neuronal responses are rapidly trained by the caloric content of gel¹⁸, thus we used two gels of different flavor and caloric content (one sweetened with sucrose and one sweetened with sucralose). Mice should rapidly learn and prefer the calorie-containing gel but also find the artificially-sweetened non-caloric gel palatable⁷³, as opposed to non-palatable control targets such as plastic or cellulose pellets which are rarely consumed^4,74. For satiety-induced devaluation (Figure 4A), mice had simultaneous access to both caloric (strawberry flavor) and non-caloric (orange flavor) gel for 30-min test sessions. While ad libitum-fed mice normally displayed preference for the caloric gel over the non-caloric gel (Figures 4B–C, ‘Not prefed’), 1-h pre-exposure to the caloric gel significantly decreased caloric gel intake during the test session but did not increase non-caloric gel intake (Figures 4B–C, ‘CG prefed’). Strikingly, ARC^AGRP neuronal activation following 1-h caloric gel pre-exposure selectively restored caloric gel intake during the test session (Figures 4B, ‘CG prefed + stim’), whereas activation of LH^VGAT neurons triggered increases in both caloric and non-caloric gel intake (Figure 4C, ‘CG prefed + stim’). Thus, ARC^AGRP neuronal activity appears to guide specific intake of caloric food even under conditions of sensory and caloric satiety, whereas LH^VGAT activity drives indiscriminate food intake unrelated to caloric content.

Figure 4. Hypothalamic circuits for feeding are distinguished by calorie-specific versus indiscriminate food intake.

(A) Schematic representation of satiety-induced devaluation. In test 1 (Not prefed), mice were housed with ad libitum chow and had 30-min simultaneous access to the caloric and non-caloric gels. In test 2 (CG prefed), mice had 1 h of access to caloric gel in the home cage prior to the 30-min choice session. Test 3 (CG prefed + stim) was the same as test 2 except that photostimulation was delivered during the 30-min choice session. n = 6 mice per group for all tests.

(B) ARC^AGRP activation triggered calorie-specific gel intake following satiety-induced devaluation in the two-gel choice assay. A three-way mixed-model ANOVA revealed a significant test × calorie × group interaction, F(2, 20) = 7.92, p = 0.0029) and that among the two-factor interactions, the test × group interaction (p = 0.0004) accounted for the most variation. Thus, follow-up two-way (test × group) mixed-model ANOVAs within each gel revealed a significant test × group interaction for caloric gel (F(2, 20) = 13.47, p = 0.0002) but not non-caloric gel (F(2, 20) = 1.13, p = 0.34). Bonferroni’s post-tests revealed that all mice displayed a significant decrease in caloric gel intake following 1 h caloric gel pre-exposure (CG prefed) as compared to normal ad libitum fed conditions (Not prefed, p = 0.0001) and that photostimulation increased caloric gel intake following devaluation in ARC^AGRP:ChR2 as compared to ARC^AGRP:YFP control mice (CG prefed + stim, ****p < 0.0001).

(C) LH^VGAT activation triggered indiscriminate gel intake following satiety-induced devaluation in the two-gel choice assay. A three-way mixed-model ANOVA revealed no significant test × calorie × group interaction (p = 0.63) but that among the two-factor interactions, the test × group interaction (p = 0.004) accounted for the most variation. Thus, follow-up two-way (test × group) mixed-model ANOVAs within each gel revealed significant test × group interactions for both caloric gel (F(2, 20) = 4.95, p = 0.018) and non-caloric gel (F(2, 20) = 6.39, p = 0.0072). Bonferroni’s post-tests revealed that all mice displayed a significant decrease in caloric (p = 0.0005) but not non-caloric (p > 0.99) gel intake following 1 h caloric gel pre-exposure, but that photostimulation increased both caloric (***p = 0.0008) and non-caloric (**p = 0.0019) gel intake following devaluation in LH^VGAT:ChR2 as compared to LH^VGAT:YFP control mice.

(D) Schematic representation of sickness-induced devaluation. Mice were given access to caloric gel for 1 h followed by i.p. injection with LiCl. The following day, mice were exposed to the caloric gel during optogenetic stimulation and caloric gel consumption was measured. Three days later, LiCl-induced devaluation was repeated with a non-caloric gel. n = 6 mice per group for all tests.

(E) ARC^AGRP activation triggered calorie-specific gel intake following LiCl-induced devaluation in single-gel test sessions (F(1, 10) = 7.47, p = 0.021; Bonferroni’s post-test **p = 0.0014).

(F) LH^VGAT activation evoked indiscriminate gel intake following LiCl-induced devaluation in single-gel test sessions (caloric gel: F(1, 10) = 11.78, p = 0.0064; Bonferroni’s post-test: ***p = 0.0005).

(G) ARC^AGRP activation did not affect non-caloric gel intake (F(1, 10) = 2.18, p = 0.17).

(H) LH^VGAT activation evoked indiscriminate non-caloric gel intake following LiCl-induced devaluation in single-gel test sessions: F(1, 10) = 9.13, p = 0.013; Bonferroni’s post-test: ***p = 0.0007).

To further elucidate the effects of manipulating these neuronal populations on devalued food intake, we used sickness-induced devaluation with lithium chloride (LiCl; Figure 4D). When administered after animals consume a novel food, LiCl induces gastrointestinal malaise and decreases subsequent intake of that food by learned negative associations (i.e., conditioned taste aversion)^75–77, as opposed to the active physiological satiety cues in the previous test (Figures 4A–C). For this, mice were exposed to a novel caloric gel for 1 h (peach flavor) and then immediately given an intraperitoneal (i.p.) injection of LiCl (6 mEq/kg) to induce gastrointestinal malaise and taste aversion. The following day, mice were again exposed to the caloric gel during photostimulation, and activation of either ARC^AGRP or LH^VGAT neurons significantly restored intake of the caloric gel relative to their respective control mice (Figures 4E and G). This procedure was repeated 3 days later using a novel non-caloric gel (cherry flavor). While no changes from control mice were observed during ARC^AGRP neuronal activation, photostimulation of LH^VGAT neurons significantly increased non-caloric gel intake following LiCl devaluation (Figures 4F and H). Together, these results demonstrate that both ARC^AGRP and LH^VGAT neurons can drive food intake despite conditions of devaluation. Our findings support a role for ARC^AGRP neurons in driving calorie-specific food intake. Here, we demonstrate for the first time that ARC^AGRP neurons encode a discriminable hunger-like interoceptive state to mediate homeostatic feeding. In contrast, the calorie-independent food intake observed during activation of LH^VGAT neurons suggests that these neurons likely mediate generalized consummatory or compulsive-like feeding.

Discussion

Understanding whether behavior reflects the internal state of an animal is a fundamental challenge of neuroscience. We, as humans, realize that many behaviors rely on internal feelings, such as reflexive withdrawal from pain or eating when hungry. However, these behaviors are not always exclusive to the preceding internal state. One may withdraw a hand due to surprise as opposed to pain or eat despite a lack of hunger. Then again, the lack of a behavioral response does not necessarily indicate the absence of a particular internal state. For instance, a boxer may not eat despite extreme hunger when trying to make a weight class or persist in a match despite being injured and in physical pain. As cutting-edge techniques continue to identify specific circuits involved in behavioral control, understanding the effects of these manipulations on the internal sensations of the subject will be required to draw accurate conclusions regarding behavior.

To this end, we assessed three hypothalamic neuronal populations with well-characterized effects on food intake for their ability to modulate interoceptive hunger- and satiety-like states. Between ARC^AGRP and LH^VGAT neurons that trigger feeding when activated, only ARC^AGRP neurons evoked significant increases in hunger-like responding. This finding, coupled with ARC^AGRP neurons driving calorie-specific food intake, strongly supports a role for ARC^AGRP neurons in mediating homeostatic feeding and hunger. Importantly, the lack of aversive effects of ARC^AGRP neuronal activation during the RTPP test in our study are not at odds with previous reports, which only observed aversive effects after several pairing sessions¹⁰. In contrast, activation of LH^VGAT neurons increased intake of familiar but not novel food, induced rewarding effects, and triggered calorie-indiscriminate feeding, suggesting these neurons are more involved in generalized consummatory behavior rather than homeostatic feeding. Notably, LH^VGAT neuronal activation did not interfere with the responding rate in this task, suggesting that the lack of hunger-like responding did not result from a generalized disruption of operant performance. Interestingly however, LH^VGAT but not ARC^AGRP inhibition decreased hunger-like responding, which accords with the more robust decrease in food intake by LH^VGAT inhibition. This decrease was also importantly not due to a disruption of operant behavior, as the lever pressing rate was unaffected by LH^VGAT inhibition. The inability of ARC^AGRP inhibition to suppress physiological hunger in the operant paradigm is not altogether surprising, as our study and those of others show that ARC^AGRP neurons are generally less amenable to inhibition than activation via optogenetic or chemogenetic methods^8,22,25. The explanation for this phenomenon is currently unclear. Since the activity of only 800 ARC^AGRP neurons is sufficient to induce feeding¹⁵, it is possible that our chemogenetic viral strategy did not inhibit enough ARC^AGRP neurons to attenuate feeding. However, this is not likely due to the co-administration of AAVs, which has been demonstrated in many studies to remain a viable viral strategy without between-virus interference^78–80. Instead, the gradual suppression of feeding observed during ARC^AGRP neuronal inhibition in the food intake experiment functionally validates the strategy indicating that there was sufficient chemogenetic inhibition of these neurons and accords with previous reports that the suppression of feeding by ARC^AGRP neuronal inhibition is related to premature cessation of feeding rather than a generalized decrease in feeding^9,25,81–83. Of note, our chemogenetic inhibition of ARC^AGRP neurons in this study caused the exact same pattern of suppressed feeding as previously observed²⁵. As such, current evidence suggests that the more subtle behavioral effects of ARC^AGRP inhibition are not related to technical issues but are simply a byproduct of the primarily unidirectional appetitive effects of ARC^AGRP circuitry on food intake.

To tie these results to the prevailing mechanistic understanding of ARC^AGRP neuronal activity, previous studies suggest that during physiological hunger, ARC^AGRP neurons are highly active and release neurotransmitters such as GABA, AGRP, and neuropeptide Y (NPY) to downstream brain regions such as the paraventricular hypothalamus (PVH)⁸⁴. Once this happens, food seeking and intake occur even in the absence of concurrent ARC^AGRP neuronal activity^8,84, suggesting that these effects are long-lasting and dependent on brain regions and neuronal circuits outside the ARC. When a mouse detects and/or consumes a food source, ARC^AGRP neuronal activity drops proportionately to the perceived and actual energy content of the food, such that more caloric foods induce larger and longer-lasting inhibition of AGRP neuronal activity^16–18,85. This activity then gradually recovers prior to triggering a subsequent feeding bout, and this cycle continues until enough calories have been consumed to meet current homeostatic need. In this light, chemogenetic inhibition of ARC^AGRP neurons is likely either (a) not powerful enough to suppress ongoing ARC^AGRP activity caused by 22-h food restriction or (b) incapable of reversing the effects triggered in downstream brain regions like the PVH to reduce initial feeding bouts. Instead, ARC^AGRP inhibition seems capable of potentiating the natural decrease in neuronal activity following food intake and/or attenuating the recovery of neuronal activity to decrease feeding in later bouts. In other words, ARC^AGRP inhibition does not affect the rate of feeding early in a meal but may cause an individual to “feel full” earlier within a meal or longer after a meal, akin to bariatric surgery in humans, effectively limiting overall food consumption within and/or across meals. In our study, since mice were food-restricted for 22 h prior to the “fasted” tests in the hunger discrimination task, ARC^AGRP neuron inhibition did not alter natural hunger cues. However, it is likely that ARC^AGRP inhibition would decrease the amount of food required for a mouse to report feeling sated in this task. Interestingly, the more robust effects of LH^VGAT inhibition on feeding and hunger-like responding in food restricted mice support previous predictions that both homeostatic and hedonic circuits are active during all feeding conditions, albeit likely to different degrees depending on the food source and physiological conditions¹. In this context, during acute manipulations, circuits in the LH may be necessary but not sufficient for hunger, whereas circuits in the ARC may be sufficient but not necessary for hunger.

One point warranting further discussion is the effect of hypothalamic activation on calorie preference during the gel intake tests in the current study. Previous work showed that taste-blind mice can develop preference for caloric sweeteners but not for non-caloric sweeteners, and that sucrose intake uniquely induced dopamine release in the ventral striatum, suggesting that calories can directly influence reward circuitry independent of taste⁸⁶. Since ARC^AGRP activation does not generally evoke dopamine release but instead potentiates dopamine responses selectively to food, this likely explains why mice maintained specific preference for the caloric gel. In contrast, the generalized increase on dopamine release by LH^VGAT activation⁸⁷ likely explains why mice did not discriminate between gels, as dopamine signaling was artificially increased enough to drive intake of the non-caloric gels and lower caloric discrimination. Relatedly, learning a behavioral task is also heavily dopamine-dependent, and the effects of ARC^AGRP activity on learning have been a subject of intense interest in previous studies^8,10. However, these studies mainly examined the role of ARC^AGRP activity on learning the caloric value of a food rather than learning the consequences of an action. Since we carefully controlled the amount of training in each physiological condition in the hunger discrimination task, we do not think that hypothalamic activation in the current study influenced the expression of a differentially learned behavior to obscure test results. Instead, behavior must have been guided by the elicited interoceptive information from the specific neuronal type being manipulated.

Our findings also further elucidate a role for LH^VGLUT2 neurons as a powerful brake on food intake. Known to trigger aversion and decreases in feeding⁴, LH^VGLUT2 neurons were more recently shown to encode satiety state and undergo modulation following chronic intake of a high fat diet⁸⁸. Together, these pieces support a role for LH^VGLUT2 neuronal activity in the promotion of satiety cues. However, higher activation frequencies of LH^VGLUT2 neurons caused a decrease in the ability of the mice to perform the hunger discrimination task, likely due to aversive-like effects. Although LH^VGLUT2 activation at low frequencies may also induce aversive effects⁴ that secondarily suppress hunger, our results at least reveal overlap between motivational and homeostatic mechanisms. Though we did not find evidence that acute chemogenetic LH^VGLUT2 inhibition increased intake of caloric food sources, contrary to the small effects previously reported⁹, imaging and ablation studies suggest that changes in LH^VGLUT2 neuron activity generally precede behavioral consequences^88,89, and thus more chronic inhibition of these neurons may be required to observe changes in food intake. Moreover, LH^VGLUT2 neuron inhibition did increase non-caloric gel intake. Since control mice consumed very little of the gel as it was novel and non-nutritive, the increase during LH^VGLUT2 inhibition supports a role for these neurons in normally inhibiting inappropriate food intake (e.g., if a food is devoid of nutrients). In concept, this role could be readily translated to pathological conditions like overeating when normal satiety feedback signals are insufficient to cease feeding beyond appropriate amounts⁸⁸. Nevertheless, the changes in open field behavior by LH^VGLUT2 neuron inhibition, which were opposite to those during LH^VGAT inhibition, also corroborate the acute functional inhibition of these neurons.

Future work will be needed to further unravel interactions between ARC^AGRP neurons and circuits within the LH. It is currently known that ARC^AGRP neurons directly innervate both LH^VGAT and LH^VGLUT2 neurons⁹, but how the activity of these circuits might be differentially affected remains less clear. Brain slice electrophysiological recordings using ChR2-assisted circuit mapping (CRACM) have only been performed between ARC^AGRP and LH orexin-expressing neurons, which were shown to be inhibitory synapses⁹⁰. LH orexin neurons are mainly glutamatergic, but some may be GABAergic⁹¹. Moreover, activation of ARC^AGRP neurons was shown to inhibit and induce long-term depression in postsynaptic neurons of several other projection regions⁹²; while the LH was not studied, it is likely that the same phenomenon occurs in both LH^VGAT and LH^VGLUT2 neurons. Deciphering the significance of this on feeding drives will likely prove an interesting line of future research.

Relatedly, the higher order brain regions responsible for the conscious perception of interoceptive hunger driven by these hypothalamic neuronal populations remain largely unstudied. Although ARC^AGRP neurons only send direct axonal projections to subcortical brain regions⁹³, chemogenetic activation of ARC^AGRP neurons evoked increased activity in several higher brain areas, including the hippocampus and multiple cortical regions⁹⁴. Yet, the only detailed characterization of ARC^AGRP activity on individual cortical neurons has been performed in the insular cortex, a critical site for interoception^95,96. ARC^AGRP activation in sated mice caused insular cortical neurons to respond to visual cues similarly to physiological food restriction^20,21. The pathway from ARC^AGRP neurons to the insular cortex was trisynaptic, passing through the paraventricular nucleus of the hypothalamus and the basolateral amygdala, indicating that several layers and types of information are likely integrated in this system. Characterizing how ARC^AGRP neurons and LH neuronal subpopulations affect cortical processing as it pertains to interoception will likely be a fruitful line of future research. Furthermore, conducting similar “hunger discrimination” paradigms using alternative reinforcement strategies such as shock avoidance, which removes the potential motivational confound of food in the current and previous studies, will be an important component to reveal the effects of these and other neuronal populations on interoceptive drives.

In summary, this is the first study to use an operant discrimination task to determine how specific neuronal circuits modulate interoceptive hunger and satiety states. While we confirm previous theories that ARC^AGRP neurons evoke hunger and homeostatic feeding, we also find complex roles for feeding circuits in the LH that can influence interoceptive hunger cues in food-restricted mice. Therefore, our study provides a versatile framework for future circuit-specific investigations of the interoceptive sensations of hunger or other complex motivational states.

STAR methods

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by Yeka Aponte (yeka.aponte@nih.gov).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• All data reported in this paper will be shared by the lead contact upon request.

• This paper does not report original code.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Experimental model and subject details

Animals

All experimental protocols were conducted in accordance with U.S. National Institutes of Health Guidelines for the Care and Use of Laboratory Animals and with the approval of the National Institute on Drug Abuse Animal Care and Use Committee. Heterozygous Agrp^Cre (Agrp^tm1(cre)Lowl; C57BL/6J background; Strain 12899, The Jackson Laboratory, ME, USA), Vgat^Cre (Slc32a1^tm2(cre)Lowl; C57BL/6J background; Strain 28862, The Jackson Laboratory), and Vglut2^Cre (Slc17a6^tm2(cre)Lowl; C57BL/6J background; Strain 28863, The Jackson Laboratory) male and female mice were used in this study. Mice were maintained at the National Institute on Drug Abuse Intramural Research Program animal facility under standard housing conditions. From the onset of operant training, mice were individually housed under a 12-hour light-dark cycle at 20 – 24 °C and 40 – 60% humidity with free access to water and food (“chow”; PicoLab Rodent Diet 20, 5053 tablet, LabDiet/Land O’Lakes Inc., MO, USA) unless otherwise stated.

Hunger discrimination training procedure

Daily discrimination sessions were conducted in six standard two-lever operant conditioning chambers (Coulbourn Instruments, LLC, PA, USA) housed inside sound-attenuating, ventilated cubicles. 20-mg sucrose pellets (PicoLab Rodent Test Diet sucrose rewards 5TUT) reinforced lever pressing and were delivered by a pellet dispenser into an extended pellet delivery trough (Coulbourn) located between the two retractable response levers. A house light located in the back panel of the operant chamber was illuminated during active response periods during experimental sessions. Experimental contingencies and data collection were executed via Graphic State v4 software (Coulbourn).

To begin operant training, mice were food-restricted to approximately 90% of their free-feeding bodyweight (fed 2.5 to 3.0 g chow per day) and trained to lever press for sucrose pellets. Mice were weighed prior to sessions performed five to seven days per week (one session daily). Initially, a single lever press was reinforced with a 20-mg sucrose pellet, and response requirements were gradually increased according to individual performance until 15 lever presses (fixed ratio 15; FR15) were required to produce sucrose pellets^49–51. When responding reliably occurred at both levers, mice were given free access to food for 3 days before discrimination training began. Mice were then trained to discriminate between periods of fasting and satiety. Under ‘fasted’ conditions, food was removed 22 h before the training session. Mice were placed into the operant chambers for a 10-min habituation period after which the house light was illuminated, both levers were extended into the chamber, and 15 lever presses at one designated active lever (e.g., left) were reinforced with a 20-mg sucrose pellet under the FR15 reinforcement schedule. Incorrect (e.g., right) lever presses reset the response requirement on the active lever⁵³ but had no programmed consequence otherwise. The session continued until five reinforcers were earned or 20 min elapsed. Under ‘sated’ conditions, food was removed 1 h before the training session and contingencies were reversed. The previously ‘incorrect’ lever (e.g., right) was now reinforced with sucrose pellets under the FR15 schedule, and the responses at the other lever (e.g., left) now reset the response requirement at the ‘sated’ lever. Conditions were administered according to a roughly double-alternating schedule (i.e., fasted, fasted, sated, sated) according to individual performance. Discrimination training continued for each mouse until they emitted at least 80% condition-appropriate responding over the entire session for either five consecutive daily sessions or six out of seven consecutive sessions.

We used the incorrect lever to reset the correct lever responses to discourage ‘trial-and-error’ strategies. Our pilot experiments followed previously used rat protocols that punished 15 incorrect lever responses with an 8-s timeout period and never reset the responses at the correct lever⁵⁰. However, we observed that mice adopted one of two strategies under these training conditions: (1) responding at one lever until sucrose or the timeout consequence was administered, then adjusting responding accordingly (stay or switch), or (2) alternating back and forth between levers, performing 2 to 3 responses at each lever for the entire session. Both strategies relied on external information, as opposed to the interoceptive cues of food deprivation or satiety and did not lead to consistent condition-appropriate responding. The adjustment of not reinforcing or punishing the inactive lever but requiring continuous bouts of responding at the active lever at a high FR schedule (FR15) required the mice to use interoceptive information to commit to one lever or the other and demonstrate reliable stimulus control. Insufficient stimulus control (i.e., performing too many incorrect responses) thus became very clear.

Two generalization tests (one ‘fasted’ and one ‘sated’) were performed following the qualification criteria described above to ensure stimulus control prior to surgery. Mice were placed in the operant chamber for a 10-min habituation period. During the active period, the house light was illuminated but responses at either lever were reinforced with sucrose pellets under the FR15 schedule (each lever had an independent FR15 schedule). Generalization tests lasted until mice earned 5 reinforcers or 20 min elapsed, whichever occurred first. Appropriate discriminative performance for at least two training days (one ‘fasted’ and one ‘sated’) was required between generalization tests. After these initial two generalization tests, stereotaxic surgery was performed.

Stereotaxic viral injection

Mice were anesthetized with isoflurane and placed onto a stereotaxic apparatus (David Kopf Instruments, CA, USA). After exposing the skull by a minor incision, small holes (< 1 mm diameter) were drilled bilaterally for virus injection. For all experiments, 50 nl of an adeno-associated virus cocktail was injected bilaterally (rate: 30 nl/min) into the ARC (bregma: −1.70 mm; midline: ±0.25 mm; dorsal skull surface: −5.85, −5.80 and −5.75 mm; Agrp^Cre mice) or LH (bregma: −1.23 mm; midline: ±1.00 mm; dorsal skull surface: −5.15 mm; Vgat^Cre and Vglut2^Cre mice) by a pulled glass pipette (20 – 30 μm tip diameter) with a micromanipulator (Narishige International USA Inc., NY, USA) controlling the injection speed.

One of two viral cocktails was injected: (1) Experimental: rAAV2/9-EF1α-double floxed-hChR2(H134R)-EYFP-WPRE-HGHpA (Addgene viral prep 20298-AAV9) + rAAV2/9-hSyn-DIO-hM4D(Gi)-mCherry (Addgene viral prep 44362-AAV9), or (2) Control: rAAV2/9-EF1α-DIO-EYFP (Addgene viral prep 27056-AAV9) + rAAV2/9-hSyn-DIO-mCherry (Addgene viral prep 50459-AAV9). All viruses were injected at a titer of 5.0 × 10¹² GC/ml.

For optogenetic targeting of the ARC, optical fibers were implanted unilaterally above the ARC (bregma: −1.70 mm; midline: +0.25 mm; dorsal skull surface: −5.60 mm), and for optogenetic targeting of the LH, optical fibers were implanted bilaterally above the LH (bregma: −1.23 mm; midline: ±1.10 mm, 5° angle; dorsal skull surface: −4.80 mm). Fiber implants were affixed to the skull with cyanoacrylate adhesive and C&B Metabond Quick Adhesive Cement System (Parkell, Inc., NY, USA). Subsequently, mice were given one week for post-surgical recovery before training resumed. Initially, n = 36 mice were trained, and surgeries were performed on all mice. However, one Vglut2^Cre mouse injected with the experimental virus cocktail died during post-operative recovery. This mouse was removed from all Figure 1 analyses. A Vglut2^Cre mouse injected with the control virus cocktail died following testing on the hunger discrimination paradigm but before the free-access feeding and RTPP studies. Optogenetic and chemogenetic testing did not occur until four weeks post-surgery to allow sufficient viral transduction time.

Optical manipulations

Optical fiber implants were coupled to patch cords which were connected to lasers (Doric Lenses Inc., Quebec, Canada) via rotary joints mounted over behavioral testing areas. Laser output was controlled by Doric Neuroscience Studio software (v5.1). For photostimulation experiments, 450-nm laser diodes were used to deliver 5-ms pulses of 10- to 15-mW light. For generalization tests, free-access feeding tests, and gel feeding tests, light pulses were delivered for 1 s (2, 5, 10, or 20 Hz, as indicated) followed by a 3-s break¹⁵, with the sequence repeating for the duration of the test. For real-time place preference experiments, 20-Hz photostimulation was delivered constantly while the mouse occupied the photostimulation-paired area. These photostimulation protocols are commonly used and do not induce significant off-target effects such as heat in brain tissue⁶⁴.

Drugs

Rimonabant (Cat. No. 9000484), bupropion (Cat. No. 10488), lorcaserin (Cat. No. 15521), liraglutide (Cat. No. 24727), and phentermine (Cat. No. 14207) were purchased from Cayman Chemical (MI, USA); naltrexone hydrochloride (Cat. No. N3136) was purchased from Millipore Sigma (MO, USA), and ghrelin (Cat. No. AS24159) was purchased from Anaspec (CA, USA). Clozapine N-oxide (CNO; Cat. No. 4936) was purchased from Tocris Bioscience (Bristol, UK), and lithium chloride (LiCl; Cat. No. L0600) was purchased from Teknova, Inc. (CA, USA). Rimonabant was dissolved in 8% Tween-80 in 0.9% saline and administered subcutaneously (s.c.). Naltrexone and bupropion were combined in a mixture of 1 part per weight naltrexone and 10 parts per weight bupropion, dissolved in saline and administered i.p. lorcaserin, liraglutide, phentermine, ghrelin, and CNO were dissolved in saline and administered i.p., except liraglutide, which was administered s.c. LiCl was diluted in sterile water and administered i.p. All drugs were prepared freshly the day of use and administered in a volume of 10 ml/kg, except LiCl which was administered in a volume of 30 ml/kg.

Generalization tests

Generalization tests were performed to compare the discriminative stimulus effects of ARC^AGRP, LH^VGAT, or LH^VGLUT2 neuronal activation or inhibition to natural states of food deprivation or satiety. Following stereotaxic surgery and recovery, mice continued training until condition-appropriate (>80% correct lever presses) discriminative performance for at least two training days (one ‘fasted’ and one ‘sated’) was met before the first generalization test, and this requirement for appropriate responding across two training sessions separated all generalization tests. For ‘sated’ condition generalization tests, ARC^AGRP and LH^VGAT experimental mice were tested with the following optogenetic photostimulation protocols: no stim, 2 Hz, 5 Hz, 10 Hz, and 20 Hz; control mice were only tested under no stim and 20 Hz photostimulation. Due to the delay in ARC^AGRP photostimulation effects on feeding, ARC^AGRP photostimulation began 10 min before the active responding period, whereas LH^VGAT photostimulation began concurrently with the start of the active responding period. LH^VGLUT2 mice received ‘sated’ generalization tests with inhibitory chemogenetic manipulation, and thus were given 1-hr pretreatments of saline (i.p.) or 1 mg/kg clozapine N-oxide (CNO, i.p.). For ‘fasted’ generalization tests, ARC^AGRP and LH^VGAT mice were tested with inhibitory chemogenetic manipulations (saline or 1 mg/kg CNO, i.p., 1-hr pretreatment). LH^VGLUT2 experimental mice were tested with optogenetic photostimulation protocols (no stim, 2 Hz, 5 Hz, 10 Hz, and 20 Hz), which began concurrently with the start of the active period; LH^VGLUT2 control mice were tested under no stim and 20 Hz photostimulation. All tests were arranged in pseudorandom order.

Following optogenetic and chemogenetic testing in YFP/mCherry control mice, mice were shuffled into three new groups for pharmacological testing to avoid effects of prior testing or genotype. To facilitate data collection, some well-performing mice were used to test more than two compounds and replaced slower, poorly performing mice for these tests. All mice were first given saline treatment prior to control ‘fasted’ and ‘sated’ tests to re-verify discriminative control. Six compounds were tested in total, and each group of mice was tested with two compounds. The first group received rimonabant (1 – 10 mg/kg, s.c.) and a mixture of 1-part naltrexone to 10-parts bupropion (0.3 mg/kg naltrexone:3 mg/kg bupropion – 3 mg/kg naltrexone:30 mg/kg bupropion, i.p.) under ‘fasted’ conditions. The second group received ghrelin (0.3 – 1 mg/kg, i.p.) under ‘sated’ condition and lorcaserin (1 – 10 mg/kg, i.p.) under ‘fasted’ condition. The third group received liraglutide (0.03 – 0.3 mg/kg, s.c.) and phentermine (3 – 10 mg/kg, i.p.) under ‘fasted’ conditions. All compounds were administered 1 h prior to testing, and test orders were arranged pseudorandomly.

Following generalization testing, all mice were housed for 1 week with ad libitum access to standard chow in home cages before further behavioral testing.

Free-access feeding with optogenetics

Mice were placed into standard rat housing cages that were empty except for one plastic weigh boat secured to the floor that contained 20-mg grain food pellets (“food,” Cat. No. 1815928-372; LabDiet) of identical composition to the standard chow (PicoLab Rodent Diet 20) but with different shape and texture. Tests were 90 min in duration and pellet consumption was assessed at the end of each of 3 consecutive 30-min epochs: pre-photostimulation, photostimulation, and post-photostimulation. Ad libitum fed ARC^AGRP and LH^VGAT experimental and control mice received two of these tests. The first test was performed during the first exposure of the mice to this testing apparatus and food pellets (‘novel context’). Then, mice were habituated to the apparatus and food pellets in 1-h sessions across three days before the second test (‘habituated’).

LH^VGLUT2 mice were food-restricted and habituated to the context and food source until food intake reliably occurred within 30 min. Then, the food-restricted LH^VGLUT2 experimental and control mice were given the 90-min optogenetic test described above.

Free-access feeding with chemogenetics

Mice were food-restricted for 3 days, with 2-h/day access to standard chow presented on the floor of the home cage, which elicited consistent, large amounts of food intake during tests. For tests, mice were weighed and injected with saline or CNO (1 mg/kg, i.p.) 1 h prior to food delivery. Approximately 5 g of food (1 to 2 pieces of standard chow) was then presented on the floor, and the pellets were collected and weighed at 1, 2, and 4 h post-access. Water was available throughout the test. One ARC^AGRP:YFP/mCherry mouse was excluded from this test due to malocclusion.

LH^VGLUT2 experimental and control mice were also tested under ad libitum fed conditions as described above. Additional palatable food tests included lard (100% calories from fat; Armour, Conagra Brands, IL, USA), sugar cubes (100% calories from sugar; Domino Sugar, MD, USA), and peanut butter chips (~50% calories from fat, ~40% calories from carbohydrates, ~10% calories from protein; Reese’s, The Hershey Company, PA, USA); these tests were separated by at least three days and 1-h CNO pretreatment was always given. Non-caloric gel (orange flavor; Snack Pack, Conagra Brands) tests lasted 1 h, and mice were tested twice on consecutive days, once with CNO pretreatment and once with saline pretreatment conditions. Mice were not habituated to any of these foods prior to the test.

Real-time place preference

Real-time place preference tests were performed in apparatuses composed of two identical rectangular chambers connected by a small hallway. A thin layer of fresh rodent bedding covered the entire floor of the apparatus. Ad libitum fed mice were connected to patch cords, placed in the hallway section of the apparatus, and then, the test began immediately. ANY-maze video tracking software v5 (Stoelting Co., IL, USA) tracked the location of the mouse throughout the 20-min test and triggered 20-Hz photostimulation when the mouse entered one of the chambers, which was kept constant for all mice. Photostimulation ended when the mouse left the chamber. ANY-maze software was used to analyze the time spent in each chamber and average speed of the mice in each chamber.

Gel devaluation tests

Prior to testing, ARC^AGRP and LH^VGAT experimental and control mice were habituated to non-caloric lemon gel (Jello, Kraft Heinz Company, IL, USA) in home cages for 3 d to reduce neophobia to gel foods. Tests were performed in empty rat cages and gels were presented in plastic weigh boats secured to the floor.

For satiety-induced devaluation, mice received 30-min ‘free choice’ test sessions over three consecutive days with simultaneous access to 2 – 3 g each of caloric (strawberry flavor; Snack Pack brand) and non-caloric (orange flavor; Snack Pack brand) gels. In the first test, ad libitum fed mice were placed in the chamber, and gel intake was determined at the end of 30 min. In the second test, mice were pre-exposed to the caloric gel in the home cage for 1 h prior to the test. In the third test, mice were pre-exposed to the caloric gel in the home cage for 1 h prior to the test and received 20-Hz photostimulation (1-s on, 3-s off) throughout the test.

For sickness-induced devaluation, mice received two 1-h ‘forced choice’ tests over consecutive days with access to 3 g of either caloric (peach flavor; Jello brand) or non-caloric (cherry flavor; Jello brand) gel. In the first session, mice were placed in the chamber for 1 h with access to caloric gel, then immediately injected with 6.0 mEq/kg LiCl (0.2 M (8.48 mg/ml) LiCl injected at 30 ml/kg, i.p.) to induce gastrointestinal malaise and returned to home cages^77,97. Gel intake was determined. In the second session, mice were returned to the test chamber with access to caloric gel and received 20-Hz photostimulation (1-s on, 3-s off) for 1 h; gel intake was determined at the end of the session. Mice were given three days of recovery and then this two-day procedure was repeated for the non-caloric gel.

Open field test

Open field tests were conducted in 30 × 30-cm clear acrylic arenas with a thin layer of bedding on the chamber floor. Mice naive to the chambers were pretreated with 1 mg/kg CNO (i.p.) and 60 min later were gently placed inside the chambers. Total locomotion, time spent in the center area, and maximum speed over 30 min were measured with ANY-maze video tracking system v5 (Stoelting).

Histology

Mice were deeply anesthetized with isoflurane and transcardially perfused with 1x phosphate buffered saline (PBS) followed by 4% paraformaldehyde (PFA) in 1x PBS. Whole brains were removed and post-fixed in 4% PFA until further processing. Samples were cryoprotected in 30% sucrose in 1x PBS, frozen on dry ice, and mounted in Tissue-Tek O.C.T. Compound (Sakura Finetek USA, Inc., CA, USA). Coronal brain sections (50-μm thick) were collected in 1x PBS using a Leica Biosystems CM3050 S cryostat (Wetzlar, Germany). Samples containing ChR2/hM4D were stained with anti-GFP and anti-DsRed for visualization. Briefly, free-floating sections were blocked for 2 h in 1x PBS with 0.03% Triton X-100 and 3% normal goat serum (block solution). Sections were then incubated in a cocktail of primary antibodies in block solution overnight at 4°C (1:1000 each; chicken anti-GFP, Cat. No. GFP-1020, Aves Labs, OR, USA; rabbit anti-DsRed, Cat. No. 632496, Clontech/Takara Bio USA, CA, USA). Sections were washed for 6 × 5 min in 1x PBS before incubating with secondary antibodies in block solution for 2 h (1:500 each; Alexa Fluor 488 goat anti-chicken, Cat. No. A11039; Alexa Fluor 647 goat anti-rabbit, Cat. No. A21245; Invitrogen/Thermo Fisher Scientific, CA, USA). Subsequently, sections were counterstained with DAPI in 1x PBS (1:5000; 4’,6-Diamidino-2-Phenylindole, Dilactate; Cat. No. D3571, Invitrogen/Thermo Fisher Scientific) and washed in 1x PBS. Sections were mounted with Fluoromount-G aqueous mounting medium (Electron Microscopy Sciences, PA, USA) onto Superfrost Plus glass slides (VWR International, PA, USA). Images were taken with an AxioZoom.V16 fluorescence microscope and LSM 700 laser scanning confocal microscope using Zen 2012 software (Carl Zeiss Microscopy LLC, NY, USA).

Quantification and statistical analyses

Graphs and statistics were generated with Prism 8 software (GraphPad Software, CA, USA). All data are plotted as mean ± s.e.m, and significant effects were noted if p < 0.05. Data were analyzed with unpaired Student’s t-tests, or one-, two-, or three-way repeated-measures or mixed-model ANOVAs as indicated in the figure legends. Geisser-Greenhouse correction was applied to operant responding data. Dunnett’s, Bonferroni’s, or Tukey’s post-tests were performed following significant ANOVA values to determine pairwise differences between conditions. Sample sizes were chosen based on similar prior experiments that yielded significant results with similar sizes^3,4,49,51.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
chicken polyclonal anti-GFP	Aves Labs	Cat#GFP-1020; RRID:AB_10000240
rabbit polyclonal anti-DsRed	Takara Bio, Inc.	Cat#632496; RRID:AB_10013483
goat polyclonal anti-chicken Alexa Fluor 488	Thermo Fisher Scientific	Cat#A11039; RRID:AB_2534096
goat polyclonal anti-rabbit Alexa Fluor 647	Thermo Fisher Scientific	Cat#A21245; RRID:AB_2535813
Bacterial and virus strains
rAAV2/9-EF1α-double floxed-hChR2(H134R)-EYFP-WPRE-HGHpA	Karl Deisseroth, Stanford University	Addgene 20298-AAV9; RRID:Addgene_20298
rAAV2/9-hSyn-DIO-hM4D(Gi)-mCherry	25	Addgene 44362-AAV9; RRID:Addgene_44362
rAAV2/9-EF1α-DIO-EYFP	Karl Deisseroth, Stanford University	Addgene 27056-AAV9; RRID:Addgene_27056
rAAV2/9-hSyn-DIO-mCherry	Bryan Roth, University of North Carolina	Addgene 50459-AAV9; RRID:Addgene_50459
Chemicals, peptides, and recombinant proteins
rimonabant	Cayman Chemical	Cat#9000484
bupropion	Cayman Chemical	Cat#10488
lorcaserin	Cayman Chemical	Cat#15521
liraglutide	Cayman Chemical	Cat#24727
phentermine	Cayman Chemical	Cat#14207
naltrexone hydrochloride	Millipore Sigma	Cat#N3136
ghrelin	Anaspec	Cat#AS24159
clozapine N-oxide (CNO)	Tocris Bioscience	Cat#4936
lithium chloride (LiCl)	Teknova, Inc.	Cat#L0600
Experimental models: Organisms/strains
Mouse: Agrptm1(cre)Lowl(AgrpCre)	The Jackson Laboratory	RRID:IMSR_JAX:012899
Mouse: Slc32a1tm2(cre)Lowl (VgatCre)	The Jackson Laboratory	RRID:IMSR_JAX:028862
Mouse: Slc17a6tm2(cre)Lowl (Vglut2Cre)	The Jackson Laboratory	RRID:IMSR_JAX:028863
Software and algorithms
GraphicState v4	Coulbourn Instruments	Cat#GS4.0; https://www.coulbourn.com/category_s/363.htm
Neuroscience Studio v5.1	Doric Lenses, Inc.	RRID:SCR_018569; https://neuro.doriclenses.com/products/doric-neuroscience-studio
ANY-maze behavioral tracking software v5	Stoelting Co.	RRID:SCR_014289; https://www.anymaze.co.uk/index.htm
Prism 8	GraphPad	RRID:SCR_002798; https://www.graphpad.com/
Other
PicoLab Rodent Diet 20 (5053) – 2.5 g Tablet	LabDiet	Cat#1815928
PicoLab Rodent Diet 20 (5053) Pellet – 20 mg Pellet	LabDiet	Cat#1815928–372
Sucrose Reward Pellet 5TUT – 20 mg Pellet	LabDiet	Cat#1811555
Armour brand lard	Conagra Brands	N/A
sugar cubes	Domino Sugar	N/A
Reese’s brand peanut butter chips	The Hershey Company	N/A
Snack Pack brand sugar-free orange flavored gelatin	Conagra Brands	N/A
Snack Pack brand strawberry flavored gelatin	Conagra Brands	N/A
Jello brand sugar-free lemon flavored gelatin	Kraft Heinz Company	N/A
Jello brand peach flavored gelatin	Kraft Heinz Company	N/A
Jello brand sugar-free cherry flavored gelatin	Kraft Heinz Company	N/A

Highlights.

• Mice can be trained to report periods of fasting from satiety

• In sated mice, ARC^AGRP neuronal activation evoked fasted-associated responding

• In fasted mice, LH^VGAT inhibition or LH^VGLUT2 activation drove satiety-like effects

• Only ARC^AGRP neuronal activation drove calorie-specific feeding