Main

We hypothesized that the rapid regulation of AGRP neurons by food cues is likely to arise from an afferent neural pathway that is distinct from those that cause sustained inhibition. One such afferent to AGRP neurons is the GABAergic ventral dorsomedial hypothalamic (vDMH) projection that expresses leptin receptor (LEPR) and prodynorphin (PDYN)12. These neurons (hereafter referred to as DMHLEPR neurons) are activated by the sensory detection of food and their activation is proportional to caloric value and palatability12. To precisely determine the relationship between food cue detection and changes in neuronal activity, we trained mice to associate an arbitrary cue (light) with access to food. Food-restricted mice were trained on a two-alternative forced-choice task (2AFC) in which the illumination of a light source was randomly presented at one of two active ports, indicating food availability in the illuminated port14. With training, mice learned to poke the correct port associated with this initially arbitrary visual cue (light) to receive the food reward (Ensure) (Fig. 1a). Using this paradigm in combination with fibre photometry, we found that in trained mice, DMHLEPR neurons are robustly activated and their activation follows visual cue presentation (green line) as opposed to nose poking for ‘ensure’ (black line) (Fig. 1b, Extended Data Fig. 1a). As anticipated, this neural response to the otherwise arbitrary light cue was absent in naive mice—thus it developed over time as the mouse learned to associate the light cue with food availability (Fig. 1c, d, Extended Data Fig. 1a, b).

Fig. 1: Regulation of DMHLEPR neurons by learned food cues.
figure 1

a, Illustration of the 2AFC behavioural task. Randomized cue lights are associated with the appearance of food (beige). b, Representative heat map of a well-trained mouse using DMHLEPR neuron fibre photometry to observe cue responses during a single behavioural session. Single trials are sorted by response time. The green line indicates light onset (time = 0), the black line indicates time when the mouse poked correctly. c, Example traces in a mouse when completely naive to the 2AFC task (top) and after it has become proficient at the task (bottom). d, Mean peak amplitude of DMHLEPR neuron fibre photometry responses in naive and trained mice in the 2AFC task. Peak amplitude was taken within a 2-s window following cue-light presentation. n = 4 mice; unpaired two-tailed t-test, *P = 0.0241. Data represent mean ± s.e.m.

Source data

DMHLEPR neurons are clearly rapidly activated by the sensory detection of food cues12 (Fig. 1b, Extended Data Fig. 1a). However, it is unknown whether DMHLEPR neurons are also capable of responding to nutrients in the gastrointestinal tract, in a similar way to AGRP neurons9,10. To address this, we infused both caloric (Ensure and glucose) and non-caloric (saline and saccharin) substances directly into the stomach. DMHLEPR neurons were activated by infusions of caloric substances, but not by non-caloric substances (Extended Data Fig. 1c–g). However, compared with sensory detection of food, the DMHLEPR neuron response to gut infusions was slower and smaller in magnitude (approximately 5 min and 5% ΔFn) compared to the pellet drop response (about 1 s, about 20% ΔFn) (Extended Data Fig. 1d–f, h). Of note, this regulation appears to not be consequential for satiation, as inhibition of DMHLEPR neurons did not increase the amount of food eaten during a meal (Extended Data Fig. 1i, Supplementary Information).

As DMHLEPR neurons are not necessary for regulating processes that occur over longer timescales—that is, satiation (Extended Data Fig. 1i) or satiety12—we next investigated the basis for and function of the rapid regulation of DMHLEPR neurons by food cues. To find the monosynaptic afferents to DMHLEPR neurons, we used EnvA pseudotyped, G-deleted rabies tracing15,16 and channelrhodopsin (ChR2)-assisted circuit mapping. Notably, whereas many sites provide GABAergic input, only one site provides glutamatergic input—the lateral hypothalamus (LH) (Extended Data Fig. 2, Supplementary Information). As neurons in the LH expressing vesicular glutamate transporter 2 (VGLUT2) (LHVGLUT2 neurons) are excitatory and would activate DMHLEPR neurons, and as the general LHVGLUT2 population has been shown to be involved in sensory cue detection17, we characterized the LHVGLUT2 → DMHLEPR circuit in greater depth.

The LH is a well-known but poorly understood central regulator of motivational drive, reward learning, and feeding behaviour18,19. Previous studies have demonstrated that LHVGLUT2 neurons, en masse, potently suppress food intake and are aversive20,21,22,23. However, as LHVGLUT2 neurons are extremely diverse, both genetically and in their projection targets24,25, we sought to test the effects of activating the DMH-projecting subset of LHVGLUT2 neurons on feeding and valence (that is, positive–rewarding versus negative–aversive). As these LHVGLUT2 neurons activate DMHLEPR neurons, which in turn inhibit AGRP neurons, we hypothesized that their activation would produce effects consistent with AGRP neuron inhibition—namely, suppress food intake and be appetitive during times of caloric deficiency (when AGRP neurons are active). To test this, we bilaterally injected adeno-associated virus (AAV) expressing Cre-dependent-ChR2 into the LH of Vglut2-IRES-Cre mice and placed a single optic fibre within the midline above the DMH to selectively stimulate LHVGLUT2 → DMH terminals (Fig. 2a). We then measured food intake while optically stimulating LHVGLUT2 → DMH terminals, beginning at the onset of the dark cycle (when AGRP neuron activity is high and mice are inclined to eat). Consistent with our proposed circuit, we found that LHVGLUT2 → DMH terminal stimulation significantly decreased food intake (Fig. 2b). To evaluate the valence of the LHVGLUT2 → DMH circuit, we used a real-time place-preference assay (RTPP) in which mice were optically stimulated in one side of a behavioural arena and could roam between the stimulation and non-stimulation side. When ad libitum-fed mice were placed in the RTPP arena at the onset of the light cycle (that is, when AGRP activity is low8), the mice were agnostic to LHVGLUT2 → DMH stimulation (Fig. 2c, Extended Data Fig. 3a). We then repeated the RTPP assay at the onset of the dark cycle (that is, when AGRP activity is high8). Under these conditions, ad libitum-fed mice showed a robust preference (about 80%) for LHVglut2 → DMH stimulation, which is expected to inhibit AGRP neuron activity (Fig. 2d, Extended Data Fig. 3a). Finally, to drive AGRP activity to a maximal level and to control for the time of day, we food-restricted the mice overnight and repeated the RTPP assay at the onset of the light cycle. Remarkably, the mice displayed a robust preference (approximately 80%) for LHVGLUT2 → DMH stimulation (Fig. 2e, Extended Data Fig. 3a). Notably, total locomotor activity was not significantly different between X-coloured fluorescent protein (XFP)- and ChR2-expressing groups (Extended Data Fig. 3b–d). Consistent with these LHVGLUT2 → DMH neurons not promoting aversion, we found that, unlike other LHVGLUT2 neurons, they do not send collaterals to sites known to promote aversive behaviours20,21,22,23 (Extended Data Fig. 3e–g). In total, these findings suggest that the DMH-projecting LHVGLUT2 neurons are functionally and anatomically distinct from previously studied LHVGLUT2 neurons20,21,22,23,26. Furthermore, as AGRP neurons promote hunger and aversion1,2,3,4,6,27,28, the observed effects (decreased hunger and being appetitive) are consistent with these LHVGLUT2 neurons being upstream of the DMHLEPR (GABAergic) → AGRP circuit.

Fig. 2: Functional characterization of a LHVGLUT2 → DMHLEPR neuronal circuit.
figure 2

a, Experimental design schematic and representative image. AAV-DIO-ChR2 was bilaterally injected into LHVGLUT2 neurons and a single optic fibre was placed in the midline above the DMH. Scale bar, 500 μm. n = 8 mice. b, Optogenetic stimulation of LHVGLUT2 → DMH terminals decreases nighttime food intake in ad libitum-fed mice. n = 8 mice; repeated-measures two-way ANOVA; Sidak’s multiple comparison test; **P = 0.0016, ***P = 0.0002. ce, Real-time place preference: optogenetic stimulation (stim.) of LHVGLUT2 → DMH terminals has no effect when mice are calorically replete (P = 0.6531) (c) and is appetitive when mice are calorically deficient at night (P = 0.008) (d) and during fasting (P = 0.0129) (e). n = 7 (ChR2) and 7 (XFP). Two-tailed unpaired t-test, *P ≤ 0.05, **P ≤ 0.01; NS, not significant. f, Experimental design (top) and representative image (bottom). AAV-DIO-GCaMP6s was injected into LHVGLUT2 neurons. Optic fibre was placed over the DMH to perform fibre photometry in LHVGLUT2 terminals. Coloured arrows indicate light wavelengths delivered and collected. Scale bar, 500 μm. n = 9 (mice). gj, LHVGLUT2 → vDMH axons are rapidly activated upon the sensory detection of food (g) and are scalable by caloric value (h, i) and palatability (j). Heat map (g) represents the trial-by-trial response in a representative mouse from 15 mg pellet presentation (time = 0). Line in hj represents mean ± s.e.m. k, Mean peak response within the first 5 s following pellet presentation. One-way ANOVA; Friedman test, ***P = 0.0007. n = 9 mice. Data represent mean ± s.e.m.

Source data

The LH is known to participate in many aspects of feeding behaviour18,19,24 and LHVGLUT2 → nucleus accumbens (NAc) neurons have been found to be necessary for signalling cue–outcome associations17. With this in mind, we tested whether the DMH-projecting LHVGLUT2 neurons were responsive to food pellet drops. We used axon-fibre photometry to record LHVGLUT2 → DMH axonal Ca2+ activity in freely moving mice (Fig. 2f) while presenting food pellets of increasing caloric value and palatability7,12 (Fig. 2g–k). LHVGLUT2→DMH axons were rapidly activated upon the presentation of a food item, were scalable to caloric content, and the responses were transient (returning to baseline within 10–20 s) (Fig. 2h–k, Extended Data Fig. 3i, j). In addition, we found that the response was mostly specific for food in that a much smaller and much more delayed response was seen when a non-edible object was presented (Extended Data Fig. 3h, i) or when water was presented to water-deprived mice (Extended Data Fig. 3k–m). In sum, our findings suggest that LHVGLUT2 → DMH neurons are activated by sensory cues preceding food ingestion, and that they probably provide this information to DMHLEPR neurons.

If the LH does indeed provide food cue information to the DMHLEPR → AGRP pathway, then selectively inhibiting LHVGLUT2 → DMH afferents should attenuate the food cue response in DMHLEPR neurons. We used two distinct methods to determine whether LH → DMH afferents were necessary for the cue-evoked response in DMHLEPR neurons. First, we used a retrograde AAV-FlpO virus to selectively express FlpO in LH neurons that project to the DMH. To record Ca2+ activity while simultaneously inhibiting LH → DMH afferents, we injected Flp-dependent AAV-hM4Di into the LH and Cre-dependent AAV-GCaMP6s into the DMH of Lepr-IRES-Cre mice (Fig. 3a). Flp-dependent AAV-hM4Di efficiently hyperpolarized LH neurons (Extended Data Fig. 4a, b). Notably, inhibiting LH → DMH afferents attenuated the food cue-evoked response in DMHLEPR neurons by about 70% (Fig. 3b, c). Clozapine-N-oxide (CNO) injections in GFP-expressing control mice had no effect (Extended Data Fig. 4c, d). To specifically investigate the role of LHVGLUT2 neurons in mediating DMHLEPR neuron food cue responses, we used Vglut2-IRES2-FlpO::Lepr-IRES-Cre mice to express Flp-dependent AAV-hM4Di in LHVGLUT2 neurons and Cre-dependent AAV-GCaMP6s in DMHLEPR neurons (Fig. 3d). Selectively inhibiting LHVGLUT2 neurons also attenuated the rapid DMHLEPR cue-evoked response by approximately 68% (Fig. 3e, f). Collectively, these studies show that LHVGLUT2 neurons drive food cue-evoked responses in DMHLEPR neurons.

Fig. 3: LH afferents to the DMH are necessary for rapid food cue-evoked responses in DMHLEPR and AGRP neurons.
figure 3

a, Top, retrograde AAV-FlpO and AAV-DIO-GCaMP6s were injected into the DMH of Lepr-IRES-Cre mice and AAV-fDIO-hM4Di was injected into the LH to inhibit LH → DMH neurons. Bottom, sample images of hM4Di expression (left) and GCaMP6s expression and fibre placement (right). Scale bar, 200 μm. n = 4 mice. b, Example traces of vehicle and CNO recording session within the same animal. Vehicle and CNO recording sessions were split so that treatment comparisons were done within the same day. Vehicle 1.1 represents the first vehicle recording session and vehicle 1.2 represents the second recording session. Vehicle 2.1 represents vehicle recording session on the second day and CNO 2.2 represents the second recording session with CNO injection. The black vertical line represents cue presentation; the purple box represents the analysed 2-s period; the blue horizontal line represents peak of vehicle control from the first recording session. c, LH → DMH inhibition attenuates the cue response in DMHLEPR neurons. Normalized responses are calculated as ΔF/F0 for the second recording session (R2) divided by ΔF/F0 for the first recording session (R1). n = 4 mice. Two-tailed t-test, **P = 0.0035. d, Top, AAV-fDIO-hM4Di was injected into the LH and AAV-DIO-GCaMP6s was injected into the DMH of Vglut2-IRES2-FlpO::Lepr-IRES-Cre mice. Bottom, sample images of AAV-fDIO-hM4Di expression (left) and AAV-DIO-GCaMP6s expression and fibre placement (right). Scale bar, 500 μM. n = 4 mice. e, Example traces from one mouse for experiment described in d (labelled as in b). f, LHVGLUT2 inhibition attenuates the cue response in DMHLEPR neurons. n = 4 mice; two-tailed t-test, *P = 0.0009. g, Top, AAV-fDIO-hM4Di was injected into the LH and AAV-DIO-GCaMP6s was injected into the ARC of Vglut2-IRES2-FlpO::Agrp-IRES-Cre mice. Bottom, sample images of expression and fibre placement: AAV-fDIO-hM4Di (right) and AAV-DIO-GCaMP6s (left). Scale bar, 500 μm. n = 4 mice. h, Example traces in one mouse for vehicle and CNO (labelled as in b). i, LHVGLUT2 inhibition decreases the cue response in AGRP neurons. n = 4 mice; two-tailed t-test, *P = 0.027. Data represent mean ± s.e.m.

Source data

To determine whether LHVGLUT2 neurons do indeed cause food cue activation of AGRP neurons, we used Vglut2-IRES2-FlpO::AGRP-IRES-Cre mice to restrict expression of Flp-dependent AAV-hM4Di in LHVGLUT2 neurons and Cre-dependent AAV-GCaMP6s in AGRP neurons (Fig. 3g). Selective inhibition of LHVGLUT2 neurons robustly attenuated the rapid, cue-evoked response in AGRP neurons (approximately 76% decrease in response magnitude) but did not affect sustained AGRP inhibition during food consumption of large food pellets (Fig. 3h, i, Extended Data Fig. 4f). Thus, LHVGLUT2 neurons have an essential role in food cue-evoked inhibition of AGRP neurons but appear to not be involved in gut nutrient-mediated regulation of AGRP neurons.

We next explored the function of this food cue-mediated regulation of the LHVGLUT2 → DMHLEPR → AGRP circuit. In mice that learned the task, CNO–hM4Di inhibition of LH → DMH neurons had no effect on task performance (Extended Data Fig. 4e, g–j). This lack of effect may be owing to habit-like performance in such highly trained mice29. To determine whether this food cue-regulated circuit has a role in the learning of this task, we expressed Cre-dependent AAV-tetanus toxin (TeNT) in DMHLEPR neurons to eliminate evoked synaptic release within the circuit (Fig. 4a). We then trained naive mice on the 2AFC task (Fig. 1a) until they mastered the task (or a maximum of 21 training days). We calculated the number of correct responses in addition to the errors made, including misses (failing to poke within the response window) and false alarms (poking in the incorrect port). We also assessed the response time (how long after cue presentation the poke was performed). TeNT-mediated DMHLEPR silencing caused a significant delay in task acquisition (Fig. 4b, c). In addition, the TeNT-mediated DMHLEPR-silenced mice took many more days of training to reduce their misses, false alarms and response times (Extended Data Fig. 5a–c). The higher number of misses and false alarms and, in particular, the longer response times suggest that in DMHLEPR-silenced mice, the cue light is less effective in motivating food-seeking behaviour. Consistent with the view that the DMHLEPR → AGRP neuron circuit is not involved in either satiety12 or satiation (Extended Data Fig. 1i), body weight (Extended Data Fig. 5d) and post-fast refeeding (Extended Data Fig. 5e) were not affected in TeNT-mediated DMHLEPR-silenced mice.

Fig. 4: Afferent modulation of AGRP neurons is required for learning a cue-initiated food acquisition task.
figure 4

a, Top, AAV-DIO-TeNT injection into DMHLEPR neurons. Bottom, example images of TeNT expression. Scale bar, 500 μm. n = 5 (mice). b, DMHLEPR neuron silencing attenuates correct responses. n = 5 (GFP) and 5 (TeNT) mice. Red line is the line of best fit (***P < 0.0001, plateau = 97.49 (GFP), 132.4 (TeNT); tau = 6.928 (GFP), 20.80 (TeNT)). Two-way repeated measures ANOVA; main effect of days, ****P < 0.0001; main effect of group, ****P < 0.0001. c, Silencing DMHLEPR neurons increases time to reach learning criterion (>80% correct across three consecutive days); 5 out of 5 GFP and 4 out of 5 TeNT mice reached the learning criterion. n = 5 (GFP) and 5 (TeNT) mice. Two-tailed, unpaired t-test; *P = 0.0238. d, Top, AAV-DIO-TeNT injection into DMHLEPR neurons. Bottom, example images of TeNT expression. Scale bar, 500 μm. n = 6 mice. e, Silencing DMHLEPR neurons does not affect behavioural performance in learning to obtain water. n = 4 (GFP) and 6 (TeNT) mice. Red line is the line of best fit (plateau = 129.3, tau = 11.53). Two-way repeated measures ANOVA; main effect of days, ****P < 0.0001; main effect of group, P = 0.7203. f, Silencing DMHLEPR neurons does not affect learning (>80% correct across three consecutive days) when water-deprived mice are trained to receive water rewards. n = 4 (GFP) and 6 (TeNT) mice. Two-tailed, unpaired t-test; P = 0.4375. Data represent mean ± s.e.m.

Source data

Although AGRP neuron-regulating DMHLEPR neurons project only to the arcuate nucleus12 (ARC), the DMH contains other LEPR-expressing neurons that project elsewhere12. Thus, the impairment in learning caused by TeNT silencing of DMHLEPR neurons could be owing to the silencing of DMHLEPR neurons that project to other parts of the brain (Extended Data Fig. 5f). To address this, we used Pdyn-IRES-Cre mice to express TeNT selectively in DMH neurons expressing PDYN (DMHPDYN neurons). It was previously established that PDYN is expressed by a major subset of DMHLEPR neurons and that DMHPDYN neurons send long-range projections exclusively to the ARC, where they inhibit AGRP neurons12 (Extended Data Fig. 6). These mice were then trained on the 2AFC task. Of interest, silencing DMHPDYN neurons modestly increased body weight in TeNT-expressing mice (Extended Data Fig. 5l) but did not affect post-fast refeeding (Extended Data Fig. 5m). As body weight was not increased when DMHLEPR neurons were silenced (Extended Data Fig. 5d), this body weight effect may be owing to the silencing of LEPR-negative DMHPDYN neurons. Of note, silencing DMHPDYN neurons, similar to silencing DMHLEPR neurons, significantly altered task performance in that it delayed task acquisition, increased mistakes made and increased response time (Extended Data Fig. 5g–k). Given that DMHPDYN neurons project selectively to AGRP neurons, these findings, along with those from the DMHLEPR-silencing study (Fig. 4a–c, Extended Data Fig. 5a–c), strongly support the view that the DMH → AGRP neuronal circuit has a key role in the learning of a sensory cue–food acquisition task.

To specifically test the role of the LH → DMH segment of the circuit in learning the 2AFC task, we selectively silenced DMH-projecting LH neurons by bilaterally injecting a retrograde AAV-Cre virus into the DMH and a Cre-dependent TeNT AAV into the LH (Extended Data Fig. 5n). Indeed, silencing these neurons significantly impaired the rate of task acquisition and significantly increased misses and response times (Extended Data Fig. 5o–s). Of note, TeNT-mediated silencing of LH → DMH neurons did not affect body weight or post-fast refeeding (Extended Data Fig. 5t, u), suggesting that this LH → DMH projection is not necessary for satiety or satiation as seen when silencing DMHLEPR neurons (Extended Data Fig. 5d, e). Collectively, these data (Fig. 4a–c, Extended Data Fig. 5a–u) show that the LHVGLUT2 → DMHGABAergic → AGRP neuron circuit has an important role in promoting the mastery of this caloric-deficiency-driven, sensory cue-initiated food-acquisition task.

To assess the possibility that TeNT expression in DMHLEPR neurons somehow interferes non-specifically with deprivation state-motivated learning of an operant task, we generated a cohort of TeNT-expressing water-deprived mice and determined their ability to learn the identical task with water, instead of food, as the reward (Fig. 4d). We hypothesized that the principles of learning in these two tasks would be identical, with the sole exception that water cue regulation of aversive thirst neurons6,30,31,32, instead of food cue regulation of hunger neurons, would promote learning. As the DMH → AGRP circuit is involved in food cue regulation of hunger neurons, and not in water cue regulation of thirst neurons, this experiment controls for nonspecific effects. As shown in Fig. 4e, f and Extended Data Fig. 5v–x, TeNT expression had no effect on learning in the water-oriented version of the task. Thus, the DMHLEPR → AGRP neuronal circuit is specific for learning food cue-initiated food-acquisition tasks, which is congruous with the very specific role of AGRP neurons in regulating food intake but not water intake.

Discussion

These findings lead us to propose the following model. Caloric deficiency activates AGRP neurons (reviewed in ref. 11) and this causes the aversive feeling of hunger6. Environmental cues instructive for food acquisition engage the LHVGLUT2 → DMHLEPR → AGRP neuron circuit and this transiently reduces AGRP neuron activity. As AGRP neurons have been proposed to transmit a negative-valence teaching signal6, these ‘appetitive’ falls in aversive AGRP neuron activity, over time, increase the incentive salience13 of food cues, thereby facilitating the learning of food acquisition tasks. If it holds, this model implies the existence of a very important general neurobiological mechanism for how homeostatic deficiency states promote the learning of tasks directed at acquiring the cognate goals (such as caloric deficiency promoting food acquisition tasks or dehydration promoting water acquisition tasks)—by providing the ‘substrate’—the deficiency-activated aversive drive neurons6,30,31,32—on which the inhibitory sensory cue-regulated afferent neurons can operate.

Methods

No statistical methods were used to predetermine sample size. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment.

Experimental subjects

Vglut2-IRES-Cre (JAX 016963)33, Vglut2-IRES2-FlpO (JAX 030212) (unpublished, donating investigator H. Zeng (Allen Institute for Brain Science)), AGRP-IRES-Cre (JAX 012899)34, Pdyn-GFP35 and wild-type mice (JAX101045) were obtained through Jackson Laboratories or in house. Lepr-IRES-Cre36 mice were maintained as previously described12. All mice were maintained on a mixed genetic background unless otherwise noted. Pdyn-IRES-Cre27 mice were maintained as previously described and kept on a congenic C57 BL/6J background. The National Institute of Health and Beth Israel Deaconess Medical Center Institutional Animal Care and Use Committee approved all animal care and experimental procedures. Mice were housed at 22–24 °C, 20–30% humidity with a 12:12 light:dark cycle with standard mouse chow (Teklad F6 Rodent Diet 8664) and water was provided ad libitum, unless otherwise stated. All diets were provided as pellets. For all behavioural studies, we used male mice between 8 and 20 weeks of age. For electrophysiogical recordings, we used male mice between 8 and 12 weeks of age.

Brain tissue preparation

Mice were terminally anaesthetized with chloral hydrate (Sigma-Aldrich C8383) and transcardially perfused with phosphate-buffered saline (PBS) followed by 10% neutral buffered formalin (Fisher Scientific SF100). Brains were extracted then cryoprotected in 20% sucrose. All brains were sectioned coronally on a freezing sliding microtome (Leica Biosystems) at 40 μm and collected in four equal series.

Immunohistochemistry

Tissue sections were washed with 0.1 M phosphate-buffered saline (pH 7.4) then blocked in 5% normal donkey serum/0.2% Triton X-100 in PBS for 1 h at room temperature. Sections were then incubated overnight at room temperature in blocking solution containing: rat anti-mCherry (1:3,000, Invitrogen M11217) and chicken anti-GFP (1:1,000, Invitrogen A10262). Secondary detection was performed with Alexa Fluor 488 or 594 conjugated donkey anti-chicken or donkey anti-rat (1:1,000, Invitrogen) for 1 h at room temperature. After secondary incubation, sections were washed and mounted onto gelatin-coated slides and fluorescent images were obtained with an Olympus VS120 slide scanner microscope. Our inferred Bregma coordinates on all histological images were adopted from a stereotaxic atlas37.

Stereotaxic surgeries and viral injections

For viral injections, six-to-eight week old male mice were anaesthetized with a ketamine (100 mg kg−1) and xylazine (10 mg kg−1) cocktail diluted in 0.9% saline and placed into a stereotaxic apparatus (Kopf model 940). Subcutaneous injection of sustained-release meloxicam (4 mg kg−1) was provided as postoperative care. A pulled glass micropipette (20–40 μm diameter tip) was used for stereotaxic injections of AAV. For electrophysiological experiments, bilateral injections (25 nl) of purified AAV (6.24 × 1012 viral genomes ml−1) were injected into the NAc (from bregma: +1.3 AP, ± 0.5 ML, −4.25 DV), BNST (from bregma: +0.14 AP, ± 0.75ML, −4.9 DV), MPO (from bregma: +0.4 AP, ± 0.25 ML, −4.75 DV), LH (from bregma: −1.3 AP, ± 0.9 ML, −5.3 DV), VTA (from bregma: −3.15 AP, ± 0.6 ML, −4.74 DV), IPN (from bregma: −3.4 AP, ± 0.0 ML, −4.8 DV), and PAG (from bregma: −4.16 AP, ± 0.3 ML, −2.5 DV). For optogenetic experiments, bilateral injections (15 nl) of AAV9-EF1α-DIO-ChR2(H134R)-eYFP purchased from the University of Pennsylvania School of Medicine Vector Core (donating investigator, K. Deisseroth; AV-9-20298P; Addgene: 20298; 2.7 × 1013 viral genomes ml−1) were injected into the LH (coordinates as above). For in vivo fibre photometry experiments, AAV1-hSyn-DIO-GCaMP6s (University of Pennsylvania Vector core; Addgene 100845-AAV1; 1.6 × 1013 viral genomes ml−1) was injected unilaterally into either the LH (25 nl, coordinates as above), DMH (25 nl, from bregma: −1.80 AP, ± 0.3 ML, −5.2 DV), or ARC (150 nl, from bregma: −1.45 AP, ± 0.25 ML, −5.85 DV). For retrograde chemogenetic silencing studies, AAV6-CAG-FlpO (Boston Children’s Hospital Viral Vector Core, modified from Addgene 67829; 2.88 × 1014 viral genomes ml−1) was bilaterally injected into the DMH (15 nl, coordinates as above) and AAV8-nEF-fDIO-hM4Di-mCherry (Boston Children’s Hospital Viral Vector Core, modified from Addgene 44362 and 55644; 9.98 × 1013 viral genomes ml−1) was bilaterally injected into the LH (25 nl, coordinates as above). For Flp-dependent chemogenetic studies in combination with fibre photometry, AAV8-nEF-fDIO-hM4Di-mCherry was bilaterally injected into the LH (25 nl, coordinates as above) or, the DMH (25 nl, coordinates as above). For tetanus toxin mediated silencing studies, AAVDJ-CMV-DIO-eGFP-2A-TeNT (Stanford; GVVC-AAV-71; 3.6 × 1012 viral genomes ml−1) was injected bilaterally into the DMH (15 nl, coordinates as above). Finally, for projection specific tetanus toxin mediated silencing studies, AAVDJ-CMV-DIO-eGFP-2A-TeNT was injected bilaterally into the LH (20 nl, coordinates as above) and rAAV2-hSyn-Cre (Addgene 105553; 1.2 × 1013 viral genomes ml−1) was injected bilaterally into the DMH (15 nl, coordinates as above). Mice were allowed to recover for a minimum of three weeks before the initiation of any experiments. Following each experimental procedure, accuracy of AAV injections were confirmed via post hoc histological analysis of mCherry, YFP, or GFP fluorescent protein reporters, viral expression of each individual surgery is catalogued in detail within Extended Data Figs. 711. All subjects determined to be surgical ‘misses’ were those with absent or low reporter expression and were removed from the experimental dataset. In addition, mice were excluded from the data set if reporter expression was primarily outside of the area of interest. Anatomical boundaries were drawn by using the DAPI signal so that we could clearly discern landmarks within a histological section. On the basis of the landmarks present, we then inferred the A/P coordinate in the atlas37 and traced the outline of different nuclei and superimposed the outline on our histological images Extended Data Fig. 12.

Optic fibre implantation

Optic fibre implantations were performed during the same surgery as viral injections (above). For optogenetic photostimulation of LH → DMH terminals, ceramic ferrule (Precision Fibre Products) optical fibres (200 μm diameter core, 0.39 NA, multimode; Thorlabs) were implanted within the midline over the DMH (from bregma: −1.8 AP, ± 0 ML, −4.7 DV). For LHVGLUT2 → DMH axon fibre photometry, a metal ferrule (Precision Fibre Products) optic fibre (400 μm diameter core, 0.5 NA, multimode; Thorlabs) was implanted unilaterally over the DMH (from bregma: −1.8 AP, ± 0.3 ML, −5.1 DV). For DMHLEPR and ARCAGRP cell body fibre photometry, a stainless steel ferrule optic fibre (same as above) was implanted unilaterally over the DMH (coordinates as above) or the ARC (from bregma: −1.45 AP, ± 0.25 ML, −5.8 DV). Fibres were fixed to the skull using dental acrylic and mice were allowed to recover for three weeks before the start of acclimation to behavioural testing.

Monosynaptic rabies mapping

Lepr-IRES-Cre mice were unilaterally injected with a 1:1 mixture of AAV8-EF1α-DIO-TVA-mCherry (University of North Carolina Vector Core, donating investigator N. Uchida) and AAV8-CAG-FLEX-Rabies G (Stanford; GVVC-AAV-59) into the DMH (15 nl, coordinates as above) (Extended Data Fig. 2a). Mice recovered for three weeks after TVA and RG transduction to ensure adequate levels of TVA and RG viral expression. Following recovery, mice underwent a second surgery in which mice were injected with SADΔG-EGFP (EnvA) rabies (Salk Gene Transfer Targeting and Therapeutics Core) into the DMH (15 nl, coordinates as above). Mice recovered for 7 days to allow for the retrograde transport of rabies virus and EGFP expression before perfusion and histological processing. Sites of afferent input to DMHLEPR neurons were assessed by the presence of EnvA-eGFP positive neurons.

Rabies collateral mapping

Vglut2-IRES-Cre mice were unilaterally injected with AAV8-EF1α-DIO-TVA-mCherry into the LH (15 nl) and allowed to recover for three weeks (Extended Data Fig. 3e–g). Then, SADΔG-EGFP (EnvA) rabies was unilaterally injected into the DMH (15 nl). Mice were allowed to recover for seven days to allow for the retrograde transport of rabies virus and EGFP transgene expression before perfusion and histological processing. Comprehensive examination of SADΔG-EGFP (EnvA) axonal and retrograde transduction was assayed using immunohistochemistry followed by imaging the entire brain for the presence of EGFP expression.

Electrophysiology

Mice were deeply anaesthetized, and intracardially perfused with ice-cold dissection buffer (in mM: 2.5 KCl, 1.25 NaH2PO4, 20 HEPES, 10 MgSO4•7H2O, 0.5 CaCl2•2H2O, 92 choline chloride, 25 glucose, 2 thiourea, 5 sodium ascorbate, 3 sodium pyruvate, and 20 NaHCO3) bubbled with 95% O2, 5% CO2. Brains were then rapidly removed and immersed in ice-cold dissection buffer. DMH sections were dissected and 300 μm thick coronal slices were prepared using a vibrating microtome (Campden 7000smz 2). Slices recovered for 10 min in a 35 °C submersion chamber filled with oxygenated dissection buffer. Slices were then transferred to a secondary 35 °C submersion chamber filled with oxygenated artificial cerebrospinal fluid (ACSF; in mM: 125 NaCl, 2.5 KCl, 25 NaHCO3, 2 CaCl2, 1 MgCl2, 1.25 NaH2PO4, 25 glucose) and allowed to recover for an additional 15 min. Slices were then kept at room temperature in oxygenated ACSF for ≥ 30 min until use (Extended Data Fig. 2b–m).

Channelrhodopsin-2 assisted circuit mapping

To isolate optically evoked inhibitory post-synaptic currents (oIPSCs) and excitatory post-synaptic currents (oEPSCs), slices were placed in a submersion chamber and perfused at 4 ml min−1 with oxygenated ACSF. Cells were visualized with a Scientifica SliceScope Pro 1000 microscope equipped with infrared differential interference contrast optics. DMHPDYN GABAergic neurons were identified by GFP fluorescence ventral to the DMC (dorsomedial hypothalamic nucleus compacta). Open-tip resistances for patch pipettes were between 2–4 MΩ and were backfilled with a Cs-based internal solution containing (in mM): 135 CsMeSO3, 10 HEPES, 1 EGTA, 3.3 QX-314 (Cl salt), 4 Mg2-ATP, 0.3 Na2-GTP, and 8 Na2-phosphocreatine with pH adjusted to 7.3 with CsOH and osmolarity adjusted to about 295 mOsM by the addition of sucrose. oEPSCs were isolated with membrane potential clamped at Vh = −70 mV and oIPSCs were isolated with membrane potential clamped at Vh = 0 mV. Bath solutions for pharmacological isolation of excitatory or inhibitory currents in whole-cell voltage clamp recordings contained SR95531 (10 μM, gabazine), kynurenic acid (1 mM), tetrodotoxin (TTX, 1 μM), and 4-Aminopyridine (4-AP, 500 μM). To photostimulate ChR2-positive fibres, an LED light source was used (470 nM, Cool LED pE-100). The blue light was focused onto the back aperture of the microscope objective (40×) producing wide-field exposure around the recorded cell of 10–15 mW per mm2 as measured using an optical power meter (PM100D, Thorlabs). A programmable pulse stimulator, Master-8 (A.M.P.I.) and pClamp 10.2 and 10.6 software (Molecular Devices, Axon Instruments) controlled the photostimulation output. The oIPSC/oEPSC detection protocol consisted of one blue-light pulse (5 ms pulse length) at 30 s intervals for at least 6 consecutive sweeps. Changes in series and input resistance were monitored throughout the experiment by giving a test pulse every 30 s and measuring the amplitude of the capacitive current. Cells were discarded if series resistance rose above 25 MΩ (Extended Data Fig. 2b–m). All electrophysiology data was analysed using Clampfit 10.2 and 10.6.

Optogenetic behavioural experiments

In vivo photostimulation of LHVGLUT2→DMH terminals was conducted by firmly attaching a fibre optic cable (1.25 m long, 200 μm core diameter, 0.63 NA; Doric Lenses) with ceramic split sleeves (Precision Fibre Products) (Fig. 2, Extended Data Fig. 3a–d). Mice were acclimated by connecting them to a ‘dummy’ fibre optic cable three days before the initiation of the experiment. Mice were stimulated with blue light (465 nM LED; Plexon) at 20 Hz, 5 ms pulses for 1 s with a 3 s recovery period (LED off) during stimulation trains to avoid ChR2 desensitization, neuronal transmitter depletion, and tissue heating. Light pulse trains were programmed using a waveform generator (National Instruments) that provided TTL input to the blue light LED. The light power exiting the fibre optic cable measured by an optical power meter (Thorlabs) was 7–8 mW in all experiments. After completion of photostimulation experiments, mice were perfused for assessment of surgical accuracy of both ChR2-expression and optic fibre tip location via histological analysis as described in ‘Stereotaxic surgeries and viral injections’.

Food intake studies

To test the sufficiency of LHVGLUT2→DMH neurons for satiety, mice were tested under conditions of physiological hunger at the onset of the dark cycle (Fig. 2b). For dark cycle feeding, mice with ad libitum access to food were photostimulated for 5 min before the onset of the dark cycle (a time when mice often eat) and photostimulation continued throughout the duration of the study (three hours). For post-fast refeeding assays (Extended Data Figs. 1, 5), mice were food restricted for 24 h then, were given ad libitum access to food. Food was then weighed each hour to determine the amount consumed during the experimental manipulation.

Real-time place preference assays

Mice were placed in a custom-made behavioural arena (transparent acrylic, 50 × 50 × 25 cm) for 20 min (Fig. 2, Extended Data Fig. 3a–d). One counterbalanced side of the arena was designated as the photostimulation side. The mouse was placed in the stimulation side at the onset of the experiment and each time the mouse crossed to the non-stimulation side of the arena, the photostimulation immediately stopped until the animal crossed back into the stimulation side. Behavioural data was recorded with Ethovision software (Noldus Information Technologies). To test photostimulation preference during different hunger states, ad libitum-fed mice were placed within the arena immediately before the onset of the dark cycle. Following a one-week rest period, ad libitum fed mice were then placed in the same arena and tested for photostimulation preference at the onset of the light cycle. Following a one-week rest period, mice were then fasted overnight then, placed in the arena and tested for photostimulation preference at the onset of the light cycle. Each RTPP assay was counterbalanced within animal and within day.

In vivo fibre photometry

Fibre photometry was performed on a rig constructed as follows: A 465-nm LED (PlexBright LED, Plexon) was used as the excitation source which was passed through a fluorescence mini cube (excitation: 460–490 nm, detection: 500–550 nm; Doric Lenses), and transmitted onto the sample via a fibre optic cable (1 m long, 400 μm diameter, 0.48 NA; Doric Lenses). The optic fibre was coupled to the implanted optic fibre with a ceramic mating sleeve (Precision Fibre Products). Light intensity was measured as 100–200 μW at the end of the patch cord and was kept constant across sessions for each mouse. Emitted light was collected by a photodetector (2151; Newport). The signal was digitized at 1 kHz with a data acquisition card (National Instruments) and collected with a custom MATLAB (MATLAB2016a; MathWorks) script (Figs. 13, Extended Data Figs. 1, 3, 4).

For LHVGLUT2→DMH axon fibre photometry recordings (Fig. 2, Extended Data Fig. 3h–j), mice underwent a 25-min recording session within their home cage that consisted of: eight ‘small’ chow (15 mg; Bio-Serv) trials, two ‘large’ chow (500 mg; Bio-Serv) trials, and one peanut butter (Reese’s peanut butter chips; Hershey) trial which were dropped into a Pyrex Petri dish. Each mouse underwent only one session per day and was food restricted (85% of ad libitum body weight) for at least one week before beginning the experiment in which it was habituated to the pellet drops within the Petri dish. All trials were pooled to calculate mean peak response (0–5 s following food presentation) to each food presentation. For object drop experiments, mice underwent a 12-min recording session within their home cage that consisted of 10 total trials of non-edible object drops. Objects consisted of uniform, white plastic marbles (BC Percision, Hungry Hungry Hippos Marbles). After an object drop session, a 500 mg chow pellet was dropped as a positive control. Each mouse underwent a single session per day. All trials were pooled to calculate mean z-score response to each object drop.

For water presentation experiments (Extended Data Fig. 3k–m), mice were water restricted (85% of original body weight) for at least one week before beginning the experiment. Mice were then habituated to receiving water in a ceramic dish within their home cage and were given free access for 5 min. All sessions were pooled to calculate the mean peak response (0–30 s) and the time to peak response.

Data were analysed using a custom MATLAB (MATLAB2016b; MathWorks) script (Figs. 13, Extended Data Figs. 1, 3, 4). Fluorescence traces were down-sampled from 1 kHz to 100 Hz and smoothed using a 1-s running average. The fractional change in fluorescence was calculated as ΔF/F = (F − F0)/F0, in which F0 was the mean of all data points from the baseline before each trial. In home cage pellet drop or water presentation experiments, F0 was the average for 5 s before the food drop/water presentation. In operant chamber silencing experiments, F0 was the average for 1 s before cue presentation. All trials in a single session were averaged then the mean peak amplitude was taken for quantification.

Two-alternative forced-choice task

To determine if LHVGLUT2→DMH neurons and DMHLEPR neurons are necessary for food cue responses (Figs. 13, Extended Data Figs. 1, 4), mice underwent training in a three-nosepoke operant chamber (Bpod r2; Sanworks) controlled by a custom MATLAB (MATLAB2016a; Mathworks) script. In brief, food restricted mice (85% of ad libitum body weight) were trained to associate a light presentation with Ensure (10 μl; Ensure PLUS, vanilla) and were required to nose poke and hold their snouts within the port for 200 ms before Ensure was delivered. Light delivery was randomized between the left and right nose poke and, mice had a 10 s response window with a 10 s inter-trial interval. The required learning criterion was a success rate of ≥ 80% across three consecutive days. After mice learned the task, mice were then attached to a fibre optic cable as described above. To determine the necessity of LH afferents on food-cue responses (Fig. 3, Extended Data Fig. 4), mice underwent two, 11-min recording sessions on the same day separated by a 10-min ‘break’ period. The first session was always a saline run and mice were injected with saline 10 min before the onset of the recording session. The second session was a clozapine-n-oxide (CNO; 1 mg kg-1; 0.5% body weight volume) or saline injection 10 min before the onset of each recording session. Comparisons between vehicle and CNO recordings were made within day; therefore, a mouse received two saline injections or, one saline and one CNO injection in a single recording session. A TTL pulse triggered at the onset of each trial determined cue onset. All trials were pooled to calculate the mean peak response (0–2 s following cue presentation) and were normalized to the first recording session within the same day.

Intragastric catheter surgery

Mice with DMHLEPR photometry neural activity signal larger than 10% ∆F/F to chow during fast re-feed were implanted with intragastric catheters38 (Extended Data Fig. 1). During surgery, mice were anaesthetized with isoflurane (1.5–3%) and treated post-operation with buprenorphine (1 mg kg−1 subcutaneously) analgesia. A midline incision on the abdomen was made through skin and muscle layers. Micro-Renathane catheter tubing 6–7cm in length (Braintree Scientific, MRE-033, 0.033 × 0.014 in) was anchored with epoxy spheres on each end (Devcon Clear Epoxy Adhesive, 92926, Lowes). The catheter was inserted into the fundus of the stomach through a puncture hole and secured with surgical mesh (5-mm diameter piece, Bard, 0112660). The other end of the catheter was directed out of an intrascapular incision. A metal cap made out of 27G blunt needle was placed in the exposed end for seal. The gastric catheter was flushed with sterile water immediately, and daily, after surgery to prevent blockage. Mice were fed with gel chow diet and given at least 1–1.5 week for recovery before experimentation. Daily body weight was monitored until stable pre-surgical weight was regained.

Gastric infusion

Upon recovery, gastric infusions of liquid substances listed below were performed in a counterbalanced experimental design, under both overnight fasted and sated conditions (Extended Data Fig. 1). The intrascapularly exposed end of the gastric catheters were connected to tubing and a syringe driven by an infusion pump (Harvard Apparatus, 70-3007). At a rate of 0.1 ml min−1, 1-ml infusions were performed38 over the course of 10 min. Fibre photometry recordings were collected via a lock-in amplifier (TDT) and the software Synapse (TDT). Each trial was approximately 27 min (>7 min baseline recording, followed by a 10 min gastric infusion and a 10 min chow refeed). Each mouse underwent infusions of the following infusates: 0.9% isotonic saline, Ensure Original Nutrition Shake (vanilla), 1% saccharine, 25% d-glucose (equal caloric content as Ensure).

Each photometry recording data point was normalized against the average of the last 5 min of baseline period to produce the normalized traces. For normalized and delta (ΔFn/Fn, %) quantified comparisons, we use 1-min averages at the end of baseline period (t = −1 to 0 min), 5 min into infusion (t = 4 to 5 min), and 10 min into infusion (t = 9 to 10 min). Owing to the transient nature of neural response to food, comparison of delta (ΔFn/Fn, %) maximum magnitude between ensure infusion and food presentation were made using 30 s averages at the end of ensure infusion (t = 9.5–10 min) and at the beginning of food presentation after saline infusion (t = 10–10.5 min).

TeNT-mediated silencing

AAV-DIO-TeNT was injected into the DMH of Lepr-IRES-Cre mice, Pdyn-IRES-Cre mice, or wild-type mice (Fig. 4, Extended Data Fig. 5). Littermate controls were used for AAV-DIO-TeNT and AAV-DIO-GFP behavioural groups. Following three weeks, mice were then placed within the Bpod arena and trained everyday as described above. Mice were either food or water restricted and maintained at ≥85% ad libitum body weight. Food and water restriction was performed in separate cohorts of mice. For food-learning assays, mice were trained for a total of 21 days. For water-learning assays, mice were trained for a total of 14 days. Mice were excluded if they were non-learners meaning that they did not increase their performance rate for five consecutive days (one GFP-expressing mouse (in water-deprived group) and one TeNT-expressing mouse (in food-deprived group) were removed based on this criterion). Mice were placed in the Bpod for a total of 20 min and allowed to perform as many trials as possible with an intertrial interval of 5 s and a response window of 10 s. Mice were given either 10 μl of Ensure or 5 μl of water for food and water-learning assays, respectively. Each day, performance in the task was quantified with a custom MATLAB script (Fig. 4, Extended Data Fig. 5). It should be noted that Pdyn-IRES-Cre and wild-type behavioural cohorts (Extended Data Fig. 5–u) were different strains than our Lepr-IRES-Cre cohort; this was owing to the limited availability of mice during the COVID-19 pandemic. As such, they performed slightly differently on the 2AFC task. Therefore, their learning criterion was lowered to >70% correct responses across three consecutive days.

Quantification and statistical analysis

Statistical analyses were performed using Prism 5 and Prism 8 (GraphPad) software and are described in the figure legends in all cases. No statistical method was used to predetermine sample size, nor were randomization and blinding methods used. Statistical significance was defined as P < 0.05. All data presented met the assumptions of the statistical test employed. As mentioned in sections above, experimental mice were excluded if histological validation revealed poor or absent reporter expression or poor fibre optic placement in the region of interest. These criteria were established before data collection. n values reflect the final number of validated mice per group.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this paper.