Hunger states switch a flip-flop memory circuit via a synaptic AMPK-dependent positive feedback loop

SUMMARY

Synaptic plasticity in response to changes in physiologic state is coordinated by hormonal signals across multiple neuronal cell types. Here, we combine cell type-specific electrophysiological, pharmacological, and optogenetic techniques to dissect neural circuits and molecular pathways controlling synaptic plasticity onto AGRP neurons, a population that regulates feeding. We find that food deprivation elevates excitatory synaptic input, which is mediated by a presynaptic positive feedback loop involving AMP-activated protein kinase. Potentiation of glutamate release was triggered by the orexigenic hormone ghrelin and exhibited hysteresis, persisting for hours after ghrelin removal. Persistent activity was reversed by the anorexigenic hormone leptin, and optogenetic photostimulation demonstrated involvement of opioid release from POMC neurons. Based on these experiments, we propose a memory storage device for physiological state constructed from bistable synapses that are flipped between two sustained activity states by transient exposure to hormones signaling energy levels.

INTRODUCTION

Neurons that express Agouti-related protein (Agrp) are a molecularly-defined population localized in the hypothalamic arcuate nucleus, and increasing their electrical activity is sufficient to rapidly induce voracious feeding behavior (Aponte et al., 2011). They are intermingled with a separate and functionally opposed population that is delineated by expression of Pro-opiomelanocortin (Pomc), which inhibits feeding (Aponte et al., 2011). These neurons form a core circuit regulating food intake and energy expenditure in which AGRP neurons inhibit POMC neurons (Cowley et al., 2001). Both populations behave as interoceptive sensory neurons, modulating their electrical activity in response to hormonal signals of metabolic state. In addition, AGRP and POMC neurons receive synaptic inputs that are also sensitive to the adipocyte-derived anorexigenic hormone leptin (Pinto et al., 2004). Thus, there is a close interaction between synaptic and hormonal regulation of these neurons.

Nevertheless, the influence of synaptic input on the activity of AGRP neurons is not known. In light of the capacity of AGRP neurons to activate voracious feeding (Aponte et al., 2011), the relationship between excitatory synaptic input and neuronal activity is expected to be behaviorally important. Synaptic regulation of neuron activity is also of general interest in physiological circuits because synapses have a potential memory capacity through synaptic plasticity (Gordon and Bains, 2006). Here, we investigate these issues by first mapping a signaling pathway for synaptic plasticity at AGRP neurons and then selectively manipulating this pathway to test its influence on AGRP neuron electrical activity. The properties of this pathway endow this circuit with a memory mechanism.

We find that food deprivation results in persistent upregulation of excitatory synaptic input to AGRP neurons, which is also sufficient to increase their firing rate. We further examined this persistent synaptic activity and identified presynaptic signaling pathways and neural circuits that are responsible for its induction, maintenance, and offset. Plasticity was induced by the gut-derived orexigenic hormone ghrelin, which triggered a positive feedback loop dependent on AMP-activated kinase that maintained elevated synaptic activity even after ghrelin was removed. Conversely, persistent activity was switched off by leptin, acting through a previously undescribed opioid receptor-dependent pathway, which optogenetic experiments showed could be mediated by opioid release from POMC neurons. Based on these experiments, we propose a memory device for physiological state constructed from bistable synapses that are flipped between sustained high and low activity states by transient exposure to hormones associated with energy deficit and surfeit, respectively.

RESULTS

Deprivation-induced synaptic plasticity

To investigate a potential role for synaptic plasticity in hunger, synaptic inputs onto AGRP neurons were compared between food-deprived and ad libitum fed mice during the early light period, when mice consume little. AGRP neurons were identified for whole cell patch clamp recordings in brain slices as the subpopulation of neuropeptide Y (NPY)-expressing neurons (Hahn et al., 1998) located in the arcuate nucleus of Npy-sapphireFP transgenic mice (Pinto et al., 2004). Because two measures of post-synaptic plasticity, AMPA-R/NMDA-R synaptic current ratio (Figure S1A) and the rectification of glutamatergic synaptic currents (Figure S1B) did not show significant differences between fed and food-deprived mice, we focused on presynaptic plasticity.

We examined presynaptic function by measuring the frequency of miniature excitatory post synaptic currents (f_mEPSC). mEPSCs reflect spontaneous neurotransmitter release from the presynaptic terminal (Fatt and Katz, 1952) under pharmacologic conditions that block action potentials and, by extension, network activity (tetrodotoxin, TTX, 1 μM). For AGRP neurons, we found that f_mEPSC was approximately doubled after deprivation or when measured at the beginning of the dark period (DP) in fed mice (fed: 1.4 ± 0.1 s⁻¹, n = 62; deprived: 3.0 ± 0.2 s⁻¹, n = 62; fed(DP): 3.1 ± 0.3 s⁻¹, n = 26; F_2,147 = 30.5, P < 0.001; Figure 1A,B), while mEPSC amplitudes were not significantly different (F_2,145 = 2.2, P = 0.11; Supplementary Fig. 1C). Thus, in AGRP neurons, f_mEPSC increased in association with the animal’s tendency to consume food, either after food deprivation or at the start of the dark period. Furthermore, deprivation-induced synaptic upregulation was cell type-selective as POMC neurons, which are intermingled with AGRP neurons (Figure 1C) but are functionally opposed, experienced marked reduction in f_mEPSC with deprivation (fed: 3.3 ± 0.3 s⁻¹, n = 30; deprived: 1.8 ± 0.2 s⁻¹, n = 22; unpaired t-test P < 0.001; Figure 1D), consistent with previous experiments measuring synaptic inputs to these neurons (Sternson et al., 2005).

Figure 1. Deprivation-induced synaptic plasticity in AGRP and POMC neurons.

(A) mEPSCs from fed and food-deprived (dep) mice.

(B) f_mEPSC in AGRP neurons from deprived mice in the light period or from fed mice at transition to the dark period (DP) were both significantly increased over f_mEPSC from fed mice in the light period.

(C) Fluorescence micrograph of arcuate nucleus from pomc-topazFP, npy-sapphireFP double transgenic mouse. POMC neurons (green) and AGRP neurons (blue) are intermingled.

(D) In POMC neurons, f_mEPSC is decreased by food-deprivation.

(E) Elevated spontaneous firing rate in AGRP neurons from deprived mice (n = 11) is reduced to the level of fed mice (n = 12) by CNQX.

***P < 0.001. Data are represented as mean ± s.e.m. See also Figure S1.

Activation of AGRP neurons increases feeding behavior (Aponte et al., 2011), so we considered the impact of elevated spontaneous synaptic input on AGRP neuron firing rate. Because neuron firing is measured in the absence of TTX, we first recorded the frequency of spontaneous excitatory synaptic currents under this condition, which showed a significant deprivation-induced increase (fed: 2.4 ± 0.3 s⁻¹, n = 9; deprived: 5.4 ± 0.7 s⁻¹, n = 10; unpaired t-test P = 0.009). In line with elevated excitatory synaptic input, AGRP neurons from deprived mice also have a higher spontaneous firing rate (f_AP) than neurons from fed mice. This difference was eliminated by blocking excitatory synaptic activity with the glutamate receptor antagonist CNQX (2 μM) (+/−CNQX: F_1,21 = 15.8, P < 0.001; fed/dep: F_1,21 = 3.2, P = 0.09; interaction: F_1,21 = 6.2, P = 0.02; Figure 1E). A previous report showed that AGRP neuron firing rate was elevated after food-deprivation in the absence of all synaptic input (Takahashi and Cone, 2005). For pharmacological conditions similar to those used in this prior study, we find comparable f_AP in slices from deprived mice (f_AP, CNQX/picrotoxin: 1.2 ± 0.4 s⁻¹ , n = 10), however our results indicate that in the presence of synaptic inhibition, a more physiologically relevant condition, excitatory synaptic upregulation is necessary for increased electrical activity. Together with AGRP neuron photostimulation-evoked feeding (Aponte et al., 2011), these experiments suggest a relationship between synaptic activity, AGRP neuron firing, and feeding behavior. Because of the potential behavioral importance of this synaptic control point, we further investigated synaptic regulation of these neurons.

Synaptic upregulation requires ryanodine-sensitive calcium stores

To investigate the mechanism of synaptic upregulation at AGRP neurons, we first manipulated intracellular Ca²⁺ buffering in brain slices with the membrane permeable Ca²⁺ buffer, BAPTA-AM (25 μM). This treatment reduced f_mEPSC in slices from deprived mice to the level observed in BAPTA-AM treated slices from fed mice (fed: 1.1 ± 0.1 s⁻¹, n = 24; deprived: 1.4 ± 0.1 s⁻¹, n = 27; Figure 2A, left). These data implicate Ca²⁺-dependent processes in synaptic plasticity observed after food-deprivation.

Figure 2. Role of calcium and ghrelin in synaptic plasticity at AGRP neurons.

(A) Comparison of f_mEPSC in AGRP neurons from fed and food-deprived (dep) mice with no treatment (nt, data from Figure 1B) or treated with BAPTA-AM, ryanodine, or ghrelin (left panel). The right panel shows treatments in the presence of CdCl₂, which blocks VGCCs. All pairwise interactions were tested and P values were corrected with Holm’s method. Significant differences are denoted by any interaction across the red dashed line; interactions on the same side of the line are not significant (P > 0.05). Left and right panels were analyzed separately.

(B) Caffeine increases f_mEPSC (n = 8), which is blocked by ryanodine pretreatment (red, n = 3).

(C) Time course of ghrelin-mediated f_mEPSC increase in an AGRP neuron.

(D) Ghrelin increases f_mEPSC in AGRP neurons (n = 6), which is blocked by d-Lys³-GHRP6 (blue, n = 5) or ryanodine (red, n = 7).

(E) For fed mice, ghrelin injection increased f_mEPSC relative to saline treatment.

(F) I.c.v. d-Lys³-GHRP6 (blue) during deprivation blocked the f_mEPSC increase observed with i.c.v. saline (black).

***P < 0.001. Data are represented as mean ± s.e.m.

One potential pathway for Ca²⁺ is through voltage gated Ca²⁺ channels (VGCCs). Blockade of VGCCs with CdCl₂ (200 μM) reduced f_mEPSC in both fed and deprived mice (fed: 0.7 ± 0.1 s⁻¹, n = 26; deprived: 1.8 ± 0.3 s⁻¹, n = 21; Figure 2A, right); however, f_mEPSC in AGRP neurons from deprived mice was still significantly greater than that observed in fed mice. Thus, VGCCs contribute to f_mEPSC in AGRP neurons, but they are not required for deprivation-induced synaptic upregulation. Moreover, in the presence of CdCl₂, f_mEPSC in AGRP neurons from deprived mice was still reduced by BAPTA-AM to the level observed in slices from fed mice (0.6 ± 0.2 s⁻¹, n = 11; Figure 2A, right), which indicated that a different source of Ca²⁺ was responsible for elevated f_mEPSC in deprived mice.

We next perturbed Ca²⁺ release from internal stores by blocking ryanodine receptors (RyR) with ryanodine (10 μM). Under these conditions, f_mEPSC in AGRP neurons from deprived mice was reduced to the level observed for fed mice (fed: 1.5 ± 0.1 s⁻¹, n = 26; deprived: 1.6 ± 0.1 s⁻¹, n = 21; Figure 2A, left), and this was also the case in the presence of CdCl₂ (deprived: 0.8 ± 0.1 s⁻¹, n = 10; Figure 2A, right). This indicates that internal stores serve as a source of Ca²⁺ required for elevated f_mEPSC in deprived mice. In addition, caffeine (10 mM), which activates RyR-mediated Ca²⁺ release, increased f_mEPSC in AGRP neurons from fed mice to the level observed after deprivation, and this was blocked by ryanodine pretreatment (Figure 2B). It has been shown previously that increased spontaneous neurotransmitter release is associated with Ca²⁺ release from internal stores in the presynaptic terminal (Emptage et al., 2001). Our experiments are consistent with a role for Ca²⁺ release from internal stores in deprivation-induced plasticity.

Ghrelin mediates deprivation-induced synaptic plasticity

Synaptic plasticity was associated with conditions that increase food intake such as deprivation or dark-period onset. Therefore, we tested the role of the gut-derived orexigenic hormone ghrelin, which is elevated in circulation after food deprivation (Tschop et al., 2000). We found that ghrelin treatment (30 nM) of brain slices from fed mice (n = 23) increased f_mEPSC to levels similar to those observed in food-deprived mice (n = 17, Figure 2A). This synaptic upregulation was rapid (Figure 2C,D) and was blocked by pretreatment with a ghrelin receptor (Ghsr1) antagonist, d-Lys³-GHRP6 (100 μM) (Howard et al., 1996) (Figure 2D). Moreover, pretreatment with ryanodine also blocked the ghrelin-mediated increase in f_mEPSC (Figure 2D), consistent with our proposal that RyR-mediated Ca²⁺ release is needed for this pathway.

Next, we tested the necessity and sufficiency of ghrelin for deprivation-induced synaptic plasticity. We found that the upregulation of f_mEPSC could also be induced by intraperitoneal administration of ghrelin (1 μg/g) to ad libitum fed mice in the early light period, only 30 min before preparing brain slices (fed/saline: 1.4 ± 0.1 s⁻¹, n = 20; fed/ghrelin: 2.8 ± 0.3 s⁻¹, n = 24; unpaired t-test P < 0.001; Figure 2E). In addition, the deprivation-induced increase of f_mEPSC onto AGRP neurons was blocked by intracerebroventricular (i.c.v.) administration of d-Lys³-GHRP6 (0.35 nmol, 3 times during 24 h deprivation period; saline: 2.5 ± 0.5 s⁻¹, n = 8; d-Lys³-GHRP6: 1.2 ± 0.1 s⁻¹, n = 14; unpaired t-test P = 0.003; Figure 2F). Together, these results are consistent with a key role for ghrelin and Ghsr1 signaling for upregulation of synaptic activity in AGRP neurons after food deprivation.

Presynaptic AMPK signaling mediates synaptic upregulation and AGRP neuron activation

Ghrelin-evoked feeding requires both AMP-activated protein kinase (AMPK) signaling (Lopez et al., 2008) and AGRP neurons (Luquet et al., 2007). Furthermore, we showed that, in brain slices, excitatory synaptic input is necessary for elevated firing in AGRP neurons after food-deprivation. Thus, the influence of ghrelin on presynaptic activity suggested the possibilities that 1) AMPK regulates presynaptic activity and 2) presynaptic upregulation of glutamate release activates AGRP neuron firing. To test this, we examined the role of AMPK for ghrelin- and deprivation-induced synaptic activation and its impact on electrical activity in AGRP neurons using pharmacological activation and inhibition of AMPK signaling.

A small molecule activator of AMPK, AICAR, is transported into cells and phosphorylated to generate ZMP (Corton et al., 1995), a mimetic of adenosine monophosphate, which allosterically activates AMPK. AICAR (500 μM) increased f_mEPSC in AGRP neurons in brain slices from fed but not food-deprived mice (+/− AICAR: F_1,14 = 8.5, P = 0.011; fed/dep: F_1,14 = 1.7, P = 0.21; interaction: F_1,14 = 14.8, P = 0.002; Figure 3A, dose response in Supplementary Fig. 2A). Conversely, the AMPK antagonist, Compound C (Cpd C, 150 μM), reduced f_mEPSC in brain slices from deprived but not fed mice and also blocked subsequent ghrelin-mediated activation of f_mEPSC (treatment: F_2,12 = 5.4, P = 0.012; fed/dep: F_1,12 = 3.5, P = 0.09; interaction: F_2,12 = 5.2, P = 0.013; Figure 3B, dose response in Supplementary Fig. 2B). Pharmacological control experiments testing alternative modes of action for AICAR and Cpd C on f_mEPSC did not reveal off-target effects (Supplementary Figs. 2C-H). These results show that AMPK is a necessary and sufficient signaling component mediating synaptic upregulation after food deprivation or application of ghrelin in brain slices.

Figure 3. AMPK signaling mediates deprivation- and ghrelin-induced synaptic activity.

(A) AICAR increases f_mEPSC in AGRP neurons from fed (n = 8) but not food deprived (n = 8) mice.

(B) Cpd C reduces f_mEPSC in AGRP neurons from deprived (n = 8) but not fed (n = 6) mice and blocks the ghrelin-mediated increase in f_mEPSC.

(C) AGRP neuron dialysis with the AMPK activator ZMP (3 mM) in the patch pipette internal solution (int.) does not increase f_mEPSC relative to AGRP neurons recorded with standard internal solution (nt, data from Fig. 1B).

(D) Targeted AMPK inhibition with Cpd C (int.) in AGRP neurons does not significantly (P > 0.05) change f_mEPSC after neuron dialysis (15 min).

(E) AGRP neuron firing rate is increased by bath-applied AICAR after intracellular blockade of AMPK with Cpd C (int.).

(F) AGRP neuron firing is decreased by bath-applied Cpd C after dialysis with Cpd C (int.).

(G) Targeted AMPK inhibition with Cpd C (int.) does not block ghrelin-mediated f_mEPSC increase.

(H) Ghrelin activates firing in AGRP neurons from fed (n = 8) but not deprived (n = 7) mice with postsynaptic AMPK blockade by Cpd C (int.).

(I) Glutamate receptor blockade prevents ghrelin activation of AGRP neuron firing.

(J) Paired pulse ratio (PPR) in AGRP neurons from fed and food deprived mice in 2 mM Ca²⁺.

(K) PPR in AGRP neurons from fed mice (0.5 mM external Ca²⁺) is reduced by ghrelin and this is reversed by Cpd C. Average EPSC responses from two cells are also shown (inset).

(L,M) PPR in AGRP neurons from deprived mice (0.5 mM external Ca²⁺) is unaffected by ghrelin and increased by Cpd C. (M, inset) Average EPSC response from one cell is shown.

(N) Inhibition of CAMKK with STO-609 in AGRP neurons from fed mice blocks ghrelin-mediated but not AICAR-mediated increase of f_mEPSC.

(O) AICAR does not increase f_mEPSC in the presence of ryanodine.

(P) 8-Br-cADP ribose blocks the ghrelin-mediated increase of f_mEPSC.

(Q) Diagram of the signaling pathway supported by these experiments. Pointed and “T” arrows represent activation and inhibition, respectively.

n.s. P > 0.05, *P < 0.05, **P < 0.01, ***P < 0.001. Data are represented as mean ± s.e.m. See also Figure S2.

Modulation of f_mEPSC indicated that AMPK may act presynaptically. Because presynaptic AMPK signaling has not been reported previously, we performed additional experiments to rule out a postsynaptic mechanism by manipulating AMPK in the postsynaptic patch clamped AGRP neuron. Dialysis of AGRP neurons with the membrane impermeant AMPK activator, ZMP (3 mM) (Corton et al., 1995) in the patch pipette did not increase f_mEPSC in brain slices from fed mice (1.2 ± 0.1 s⁻¹, n = 11; Figure 3C and Supplementary Fig. 2I). In brain slices from deprived mice, dialysis with a high concentration of Cpd C (500 μM) did not decrease f_mEPSC (break-in: 3.1 ± 0.4 s⁻¹, n = 7; 15 min cell dialysis: 3.0 ± 0.4 s⁻¹, n = 7; paired t-test P = 0.25; Figure 3D). These results indicate that AMPK signaling in the postsynaptic AGRP neuron does not influence f_mEPSC, which is consistent with a presynaptic AMPK signaling pathway.

Because deprivation-induced upregulation of excitatory synaptic input increases the action potential firing rate (f_AP) of AGRP neurons (Figure 1E), we next determined the contribution of AMPK-mediated presynaptic activation to electrical activity in AGRP neurons. For this, we blocked AMPK signaling in the postsynaptic AGRP neuron with dialysis of Cpd C from the patch pipette while perturbing the presynaptic pathway with subsequent bath-application of AMPK activators and inhibitors. In brain slices from fed mice, AICAR increased f_AP in AGRP neurons under conditions of postsynaptic AMPK inhibition (control: 0.11 ± 0.05 s⁻¹; AICAR: 0.9 ± 0.3 s⁻¹, n = 5; paired t-test P = 0.033; Figure 3E). Conversely, AMPK blockade by extracellular but not intracellular Cpd C reduced f_AP in slices from deprived mice (control: 1.3 ± 0.4 s⁻¹; Cpd C: 0.2 ± 0.1 s⁻¹, n = 5; paired t-test P = 0.037; Figure 3F). Together, these experiments show that AMPK signaling is sufficient to increase AGRP neuron firing rate to levels observed in deprived mice and that this is not due to AMPK in the postsynaptic AGRP neuron.

Ghrelin activation of synaptic activity, which is sensitive to Cpd C-mediated AMPK inhibition (Figure 3B), was also unaffected by Cpd C targeted to the postsynaptic AGRP neuron (control: 1.4 ± 0.1 s⁻¹, ghrelin: 2.5 ± 0.2 s⁻¹, n = 6; paired t-test P < 0.001; Figure 3G), consistent with a presynaptic effect. Further supporting this, mEPSC amplitude was not significantly changed by ghrelin treatment (control: 17.8 ± 0.9 pA; ghrelin: 19.2 ± 1.6 pA, n = 6; paired t-test, P = 0.34). In addition, ghrelin elevated f_mEPSC in AGRP neurons dialyzed with high BAPTA (10 mM, control: 1.7 ± 0.3 s⁻¹, ghrelin: 2.9 ± 0.1 s⁻¹, n = 4; paired t-test P = 0.007, Supplementary Fig. 2J). These experiments indicate that neither postsynaptic AMPK nor postsynaptic Ca²⁺ signaling was necessary for upregulation of f_mEPSC by ghrelin, which are in agreement with a presynaptic mechanism.

Consistent with elevated excitatory synaptic input frequency, AGRP neuron firing was increased by ghrelin in slices from fed but not food-deprived mice with AMPK signaling blocked in the postsynaptic AGRP neuron (+/−ghrelin: F_1,13 = 5.1, P = 0.041; fed/dep: F_1,13 = 8.1, P = 0.014; interaction: F_1,13 =6.9, P = 0.014; Figure 3H). Also, ghrelin-mediated activation of f_AP in AGRP neurons was prevented by excitatory synaptic blockade with CNQX (control/CNQX: 0.1 ± 0.03 s⁻¹; ghrelin/CNQX: 0.1 ± 0.05 s⁻¹; n = 9; paired t-test P = 0.57; Figure 3I). Both of these results indicate that ghrelin acts directly on presynaptic terminals, which is consistent with a report showing ghrelin binding sites associated with axonal boutons in the arcuate nucleus (Cowley et al., 2003). Our data also indicate a presynaptic site of action for ghrelin activation of AGRP neuron firing rate and not a direct effect on AGRP neurons as has been previously suggested (Andrews et al., 2008).

To further examine the presynaptic mechanism described here, we also considered the influence of this signaling pathway on activity-evoked synaptic release. The ratio of synaptic responses (EPSC₂/EPSC₁, paired-pulse ratio, PPR) from a pair of electrical stimuli (50 ms interval) to afferent axons is a property associated with presynaptic function, where lower PPR values (synaptic depression) are indicative of higher synaptic release probability (Dobrunz and Stevens, 1997). A positive relationship might be expected between release probability and elevated f_mEPSC. However, despite divergent f_mEPSC levels, we found that PPR in AGRP neurons was not significantly different in brain slices from fed and food-deprived mice (fed: 0.67 ± 0.09, n = 15; dep: 0.59 ± 0.08, n = 12; unpaired t-test P = 0.53; Figure 3J). In both conditions, low PPR was not due to glutamate receptor desensitization, as similar results were obtained under partial receptor block by γ-D-glutamylglycine (2 mM), a competitive glutamate receptor antagonist (Supplementary Fig. 2K). Next, we considered the possibility that because synapses onto AGRP neurons from fed mice show high release probability, they may have limited capacity for further reduction of PPR after food deprivation or treatment with ghrelin. To test this, we shifted synaptic responses from depressing to facilitating by reducing external [Ca²⁺] from our standard conditions (2 mM) to 0.5 mM, which resulted in PPR > 1 in AGRP neurons from fed mice (Figure 3K). We found that treatment of brain slices with ghrelin under these conditions reduced PPR and switched synaptic responses from facilitating to depressing and that this was reversed by treatment with Cpd C (control: 1.3 ± 0.05, ghrelin: 0.88 ± 0.03, ghrelin+Cpd C: 1.4 ± 0.07; n = 7; Figure 3K). Conversely, even in 0.5 mM Ca²⁺, synaptic depression was observed in AGRP neurons from deprived mice, which was not significantly altered by ghrelin (control: 0.80 ± 0.09; ghrelin: 0.80 ± 0.06; n = 3; paired t-test P = 0.98; Figure 3L). Notably, AMPK inhibition with Cpd C significantly increased PPR which produced facilitating synaptic responses in AGRP neurons from food deprived mice (control: 0.85 ± 0.05; Cpd C: 1.3 ± 0.1, n = 4; paired t-test P = 0.007; Figure 3M). These experiments indicate that a ghrelin- and AMPK-sensitive signaling pathway in presynaptic terminals influences both spontaneous release and evoked release, however the observable effect on evoked release might be limited at physiologic [Ca²⁺] due to high baseline release probability.

A signaling pathway for AMPK-mediated synaptic activity

Elevated AMPK activity is typically associated with energy deficit and a high cellular AMP/ATP ratio. In our experiments, brain slices were maintained in an energy rich, high glucose solution for several hours. Thus, it is unlikely that cellular depletion of ATP after food-deprivation is the mechanism for AMPK activation, and we considered other pathways. Ca²⁺/calmodulin-dependent protein kinase kinase (CAMKK) is a direct activator of AMPK (Hawley et al., 2005; Hurley et al., 2005; Woods et al., 2005) and has been shown to be important for ghrelin-mediated feeding (Anderson et al., 2008). Inhibition of CAMKK with the inhibitor STO-609 (3-15 μM, dose response in Supplementary Fig. 2L) blocked the ghrelin-mediated increase in f_mEPSC onto AGRP neurons from fed mice, however AICAR was still capable of increasing f_mEPSC after treatment with STO-609, confirming that CAMKK signaling is upstream of AMPK (control: 1.7 ± 0.2 s⁻¹, STO-609: 1.8 ± 0.2 s⁻¹; ghrelin: 1.8 ± 0.1 s⁻¹, AICAR: 2.6 ± 0.2 s⁻¹, n = 10; Figure 3N). We note that higher concentrations of STO-609 blocked AICAR-mediated synaptic upregulation (Supplementary Fig. 2L), in line with known “off-target” AMPK inhibition by high STO-609 doses (Hawley et al., 2005). However, dose response curves for f_mEPSC and Western blots of brain slice lysates show a “dosage window” in brain slice preparations, consistent with STO-609 inhibition of CAMKK without AMPK blockade (Supplementary Figs. 2L-N).

We also investigated the downstream signaling pathway by which AMPK increases f_mEPSC. Ryanodine blockade of AICAR upregulation of f_mEPSC showed that RyRs are downstream of AMPK (Figure 3O). Furthermore, AMPK has been shown previously to elicit Ca²⁺ release in smooth muscle tissue via the endogenous RyR potentiator, cADP ribose (Evans et al., 2005). The cADP ribose antagonist, 8-Br-cADP ribose (100 μM), blocked the ghrelin-mediated increase in f_mEPSC (Figure 3P), indicating that cADP ribose production is likely the link from AMPK to RyR activation.

Collectively, these experiments are consistent with a deprivation-induced, ghrelin-sensitive pathway where signaling through Ghsr1, a Gα_q/11-coupled pathway that mobilizes Ca²⁺ from IP₃-sensitive internal stores (Smith, 2005), leads to activation of CAMKK, which signals through AMPK to release Ca²⁺ from ryanodine-sensitive internal stores, thereby increasing spontaneous excitatory synaptic activity and release probability (Figure 3Q). Elevated synaptic glutamate release results in increased AGRP neuron firing, a response that has been shown to elicit elevated food intake in mice (Aponte et al., 2011)

Persistently elevated synaptic activity results from an AMPK-dependent positive feedback loop

In light of this ghrelin-dependent pathway, it was puzzling that elevated synaptic activity in brain slices from food-deprived mice persisted for hours after preparation and incubation of brain slices in ghrelin-free artificial cerebrospinal fluid (aCSF). We found that persistent activity could be sustained even after Ghsr1 blockade by d-Lys³-GHRP6 or by the Ghsr1 inverse agonist [d-Arg¹,d-Phe⁵,d-Trp^7,9,Leu¹¹]-Substance P (SP*, 100 μM) (Holst et al., 2003) (Supplementary Fig. 3), indicating that ghrelin-mediated or ghrelin-independent constitutive Ghsr1 signaling were not responsible for the sustained activity.

To explore this further, we examined the potential for ghrelin to elicit persistent activity in brain slices from fed mice by transient exposure to ghrelin (5 min) followed by washout and treatment with SP* to block subsequent Ghsr1 signaling. This sequence led to elevation of f_mEPSC and demonstrated marked hysteresis, remaining high after ghrelin washout and treatment with SP* (n = 6; Figure 4A).

Figure 4. Synaptic hysteresis resulting from an AMPK-dependent positive feedback loop.

(A) Ghrelin upregulation of f_mEPSC shows hysteresis; f_mEPSC is sustained after ghrelin washout and Ghsr1 blockade with SP*. Duration of each transition is in parentheses (minutes).

(B) Caffeine increases f_mEPSC which remains elevated after washout. Synaptic activity returned to baseline after treatment with Cpd C, consistent with a positive feedback loop.

(C-F) Procedure for testing duration of persistent activity by transient exposure of brain slices to ghrelin or caffeine (5 min), transfer to a wash solution (10 min), transfer again to a solution containing either aCSF alone or with d-Lys³-GHRP6 (3-5 h).

(D,E) After transient ghrelin exposure and prolonged d-Lys³-GHRP6 incubation, f_mEPSC remained elevated but was rapidly (10 min) reduced by (D) Cpd C or (E) STO-609. Control brain slices that were not treated (nt) with ghrelin but otherwise incubated as above (D) had low f_mEPSC.

(F) After transient exposure to caffeine and prolonged incubation in caffeine-free aCSF (3-5 h), f_mEPSC was still elevated but was rapidly (10 min) reduced by Cpd C, consistent with the operation of a positive feedback loop.

n.s. P > 0.05, *P < 0.05, **P < 0.01, ***P < 0.001. Data are represented as mean ± s.e.m. See also Figure S3.

Hysteresis, a memory capacity observed in diverse biological processes, frequently involves positive feedback (Ferrell, 2002). We considered the possibility that, in this system, hysteresis resulted from positive feedback in the AMPK-dependent signaling pathway (Figure 3Q). This ghrelin-responsive pathway results in Ca²⁺ release from ryanodine-sensitive internal stores, but also involves the upstream Ca²⁺-sensitive kinase, CAMKK, raising the possibility for positive feedback to sustain the pathway in the absence of ghrelin. Consistent with the high cooperativity associated with positive feedback (Ferrell, 2002), dose response curves for AICAR, Cpd C, and STO-609, showed switch-like ultrasensitivity (Koshland et al., 1982) (Supplementary Figs. 2A,B,L).

To further explore the possibility of positive feedback, we reasoned that persistent synaptic activity should also be sensitive to activation and blockade at any node in the loop. To test this, we released Ca²⁺ from ryanodine-sensitive internal stores using a brief exposure (5 min) to caffeine, which we have shown elevates f_mEPSC (Figure 2B). Under these conditions, synaptic activity also exhibited hysteresis; after washout of caffeine, f_mEPSC was still elevated (Figure 4B). Subsequent blockade of the “upstream” AMPK node of the pathway with Cpd C reversed f_mEPSC to pre-caffeine levels (Figure 4B; compare with Supplementary Fig. 2F), as expected from a positive feedback loop. Toggling from the high to the low activity state in the absence of ghrelin shows that these synapses are bistable.

Duration of the persistently activated state

To measure the duration of the persistently activated state, brain slices from fed mice were treated briefly with ghrelin (30 nM, 5 min), washed by transfer to a ghrelin-free bathing solution, and then transferred again for incubation with the Ghsr1 antagonist, d-Lys³-GHRP6 (100 μM), to ensure the absence of ongoing ghrelin signaling for 3-5 h after which they were used for electrophysiology in d-Lys³-GHRP6 containing solution (Figure 4C). Remarkably, AGRP neurons from these slices still showed elevated f_mEPSC levels that were not significantly different from those observed in ghrelin treatment experiments which involved continuous exposure to the hormone (Figures 2D and 3G). Furthermore, f_mEPSC was still rapidly reversed by AMPK blockade with Cpd C (ghrelin/wash/d-Lys³-GHRP6: 2.7 ± 0.2 s⁻¹, +Cpd C: 1.1 ± 0.1 s⁻¹, n = 3; paired t-test P = 0.023; Figure 4D) or CAMKK inhibition by STO-609 (3 or 15 μM; Figure 4E). As a control, we incubated slices with d-Lys³-GHRP6 without pre-exposure to ghrelin; these showed low f_mEPSC and no effect of Cpd C (d-Lys³-GHRP6, 3-5 h: 1.3 ± 0.02 s⁻¹, +Cpd C: 1.2 ± 0.1 s⁻¹, n = 3; paired t-test P = 0.51; Figure 4D).

In addition, we tested brief caffeine exposure (10 mM, 5 min) in brain slices from fed mice followed by incubation in caffeine-free aCSF for 3-5 hours (Figure 4C). This treatment also elevated f_mEPSC, which was rapidly reversed by Cpd C (caffeine/wash: 2.9 ± 0.3 s⁻¹, +Cpd C: 1.9 ± 0.1 s⁻¹, n = 6; paired t-test P = 0.026; Figure 4F). Thus, synaptic activation persists for at least several hours after induction, is AMPK-sensitive throughout this time frame, and can be induced upstream and downstream of AMPK signaling.

Counter-regulation of persistent synaptic activity

Next, we investigated the pathway by which persistent synaptic upregulation is switched off. We observed that f_mEPSC from mice that were food-deprived and then re-fed was still elevated 24 h after re-feeding but was reduced to baseline by 48 h (Figure 5A); notably, mEPSC amplitude did not change significantly throughout this time course (data not shown, F_4,187 = 1.22, P = 0.31). Consistent with the elevated excitatory synaptic input, the action potential firing rate of AGRP neurons was also elevated 24 h after re-feeding, which required glutamatergic synaptic input (24 h re-fed: 1.5 ± 0.2 s⁻¹; 24 h re-fed/CNQX: 0.2 ± 0.1 s⁻¹, n = 6; paired t-test P = 0.002; Figure 5B).

Figure 5. Persistent synaptic upregulation is reversed by leptin-mediated opioid release.

(A) Synaptic activity before, during, and after food-deprivation and re-feeding (fed and deprived data from Figure 1C, 1h: n = 21, 24h: n = 27, 48h: n = 23). P values for multiple pair-wise comparisons were adjusted with Holm’s correction. The significant differences are denoted by any interaction across the red dashed line.

(B) AGRP neuron firing rate 24 h after re-feeding was still elevated, which was dependent on glutamatergic synaptic input.

(C) Injection of leptin in deprived mice reduced f_mEPSC in AGRP neurons relative to saline injection.

(D) DAMGO reduced f_mEPSC in AGRP neurons under VGCC block (CdCl₂). The f_mEPSC remained at this level during a 15 min wash, but was increased by treatment with AICAR.

(E) f_mEPSC in AGRP neurons from deprived mice treated with DAMGO are insensitive to Cpd C.

(F) NTX pretreatment of deprived mice blocks leptin-mediated reduction in f_mEPSC observed with saline pretreatment.

(G) Co-injection of ghrelin and NTX, but not ghrelin alone, leads to elevated f_mEPSC in brain slices prepared after 3 h.

*P < 0.05, **P < 0.01, ***P < 0.001. Data are represented as mean ± s.e.m.

This timescale led us to examine the role of the hormone leptin, which is associated with long-term regulation of energy homeostasis. Leptin reduces ghrelin-induced feeding (Nakazato et al., 2001) and deprivation-induced AMPK phosphorylation in the hypothalamus (Minokoshi et al., 2004). Leptin administration (1 μg/g) to deprived mice three hours before brain slice preparation decreased f_mEPSC to levels found in fed mice (i.p. saline: 2.8 ± 0.4 s⁻¹, n = 9; i.p. leptin: 1.7 ± 0.2 s⁻¹, n = 10; unpaired t-test P = 0.024; Figure 5C). Leptin (100 nM) treatment of brain slices from deprived mice did not affect f_mEPSC onto AGRP neurons (2.7 ± 0.3 s⁻¹, n = 11), so we considered an indirect interaction.

One possibility is that leptin activation of POMC neurons in the arcuate nucleus (Cowley et al., 2001) could release a neuropeptide modulator of AMPK activity. Because α-MSH is a POMC neuron-derived neuropeptide that reduces feeding through melanocortin receptor signaling, we tested the influence of MTII (500 nM), a synthetic agonist of melanocortin receptors, on f_mEPSC. However, this treatment did not decrease f_mEPSC onto AGRP neurons in brain slices from deprived mice (control: 3.0 ± 0.7 s⁻¹, MTII: 3.1 ± 0.6 s⁻¹, n = 6; paired t-test P = 0.93).

We next considered the possibility that an opioid such as β-endorphin could reverse persistent synaptic upregulation in AGRP neurons. β-endorphin is also produced in POMC neurons and is a μ-opioid receptor (MOR) agonist (Silva et al., 2001). Brain slices from food-deprived mice were treated with a synthetic selective MOR agonist, DAMGO (1 μM). This was done in the presence of CdCl₂ because MORs can also influence synaptic release by inhibition of VGCCs (Endo and Yawo, 2000), which we have shown to be unrelated to deprivation-induced plasticity (Figure 2A). Under these conditions, we found that, after treatment and washout of DAMGO (see Methods), f_mEPSC in AGRP neurons was similar to levels from fed mice. Furthermore, these synapses were now sensitive to AICAR, which increased f_mEPSC (control: 2.2 ± 0.3 s⁻¹; DAMGO: 0.9 ± 0.2 s⁻¹; wash: 1.0 ± 0.2 s⁻¹; AICAR: 1.8 ± 0.3 s⁻¹, n = 9; Figure 5D). Conversely, treatment with Cpd C after washout of DAMGO did not change f_mEPSC, indicating that AMPK had been fully inactivated by MOR signaling (control: 2.3 ± 0.2 s⁻¹; DAMGO: 1.1 ± 0.2 s⁻¹; wash: 1.1 ± 0.3 s⁻¹; Cpd C: 1.1 ± 0.3 s⁻¹, n = 5; Figure 5E).

The negative regulation of AMPK by opioid receptor signaling raised the question of whether tonic opioid activity was required to prevent ghrelin-independent activation of the AMPK-mediated positive feedback loop. Incubation of brain slices from ad libitum fed mice with the opioid receptor antagonist naltrexone (NTX, 500 nM) for 3-5 h did not increase f_mEPSC or induce sensitivity to Cpd C (NTX: 1.3 ± 0.2 s⁻¹; NTX/Cpd C: 1.4 ± 0.1 s⁻¹, n = 3; paired t-test P = 0.84), indicating that AMPK is not disinhibited by NTX blockade of opioid receptors. This suggests additional control pathways to restrain this presynaptic positive feedback loop in the absence of ghrelin.

We also tested the leptin-opioid relationship in vivo by treating food-deprived mice with NTX (10 μg/g) followed by leptin. This combination blocked leptin reduction of f_mEPSC in brain slices from deprived mice (vehicle/leptin: 1.6 ± 0.1 s⁻¹, n = 26; NTX/leptin: 3.0 ± 0.2 s⁻¹, n = 34; unpaired t-test P < 0.001; Figure 5F), which confirms that leptin downregulates AMPK-mediated synaptic activity through an opioid receptor-dependent mechanism.

Next, we examined the persistence of the ghrelin-induced upregulation of f_mEPSC after ghrelin injection to ad libitum fed mice. Because injected ghrelin levels are reported to decline to baseline over 1-2 h (Tschop et al., 2000), we looked at the effect of ghrelin treatment on f_mEPSC 3 h post-injection, during which time mice were not exposed to food. Based on our analysis of persistent activity at these synapses, we expected f_mEPSC to be elevated as it is 30 min after ghrelin administration (Figure 2E). Instead, we found that f_mEPSC was similar to uninjected fed mice (ghrelin, 3h: 1.5 ± 0.2 s⁻¹, n = 22). We tested the possibility that the opioid tone in fed mice was sufficient to switch off elevated synaptic activity once injected ghrelin was cleared. Consistent with this, co-treatment of ad libitum fed mice with ghrelin and NTX, led to f_mEPSC that was still elevated three hours after ghrelin treatment (ghrelin/NTX 3 h: 2.7 ± 0.4 s⁻¹, n = 10; unpaired t-test P = 0.009; Figure 5G). Fed mice treated solely with NTX did not increase f_mEPSC after 3 h (1.5 ± 0.3 s⁻¹, n = 11). These experiments are consistent with the persistence of ghrelin-induced synaptic upregulation and indicate that the baseline opioid tone in fed mice can reverse this after ghrelin levels fall.

POMC neurons release an opioid that resets persistent synaptic activity

While opioid receptors are involved in the leptin-mediated inactivation of presynaptic AMPK, it was unclear if POMC neurons could serve as the source of these opioids. We tested this using an optogenetic approach. We expressed the light-activated cation channel, channelrhodopsin-2 (ChR2) (Boyden et al., 2005) in POMC neurons, rendering them selectively photoexcitable. POMC neurons were targeted using the Cre recombinase (Cre)-dependent viral vector rAAV-FLEX-rev-ChR2-tdtomato, which we have described previously for POMC neuron photostimulation (Aponte et al., 2011; Atasoy et al., 2008).

In brain slices from food-deprived Pomc-Cre;Npy-sapphireFP double transgenic mice expressing ChR2, we photostimulated the area containing the arcuate nucleus with focused laser light in a 4×10 spatial pattern of stimulation sites (Figure 6A). Our expectation was that a neuromodulator would be released which could potentially downregulate AMPK signaling. Next, the slice was treated with TTX to record mEPSCs and CdCl₂ in order to eliminate any contribution of VGCCs. POMC neuron photostimulation led to a significant reduction in f_mEPSC in AGRP neurons, and this effect was blocked if photostimulation was performed in the presence of NTX (500 nM) (control: 1.3 ± 0.2 s⁻¹, n = 15; NTX: 2.2 ± 0.2 s⁻¹, n = 14; unpaired t-test P = 0.003; Figure 6B). When NTX was present during photostimulation, AGRP neurons from deprived mice maintained insensitivity to AICAR (Figure 6C). However, after photostimulation in the absence of NTX, f_mEPSC could be increased in response to subsequent AICAR administration (+/− AICAR: F_1,8 = 43.3, P < 0.001; +/−NTX: F_1,8 = 3.6, P = 0.096; interaction: F_1,8 = 36.8, P < 0.001; Figure 6C). These results indicate that a POMC neuron-derived opioid, likely β-endorphin, reverses AMPK-mediated persistent synaptic upregulation.

Figure 6. POMC neurons release an opioid that resets persistent synaptic activity.

(A) Epifluorescence micrograph of brain slice with POMC neurons expressing ChR2-tdtomato. Blue circles: photostimulation sites; ARC: arcuate nucleus; 3V: third ventricle; D: dorsal; V: ventral.

(B) ChR2-mediated photostimulation of POMC neurons in brain slices from deprived mice in the presence or absence of NTX. Photostimulation of POMC neurons reduces f_mEPSC unless performed in the presence of NTX.

(C) Subset of neurons in (B) subjected to AICAR after photostimulation. POMC neuron photostimulation in the absence of NTX (n = 5) rendered f_mEPSC in AGRP neurons sensitive to AICAR, while those photostimulated in the presence of NTX (n = 5) were insensitive to AICAR, indicating that a POMC neuron-derived opioid, inactivates AMPK.

**P < 0.01. Data are represented as mean ± s.e.m.

DISCUSSION

Homeostasis requires increased food intake as energy stores are depleted. In this study, we provide evidence for a dynamic neural circuit (Figure 7A) with a reversible memory storage capacity that regulates feeding behavior in response to energy deficit. Previous work showed a core circuit in which AGRP neurons inhibit POMC neurons and electrical activity is regulated by hormones that signal physiological state (Cowley et al., 2001; Cowley et al., 2003). We report here that a key control point for this circuit is excitatory synapses onto AGRP neurons. These synapses endow this circuit with unexpected memory properties based on a presynaptic AMPK-dependent positive feedback loop and regulation by a previously undescribed POMC neuron-derived opioid pathway.

Figure 7. SR Flip-flop model of a neural circuit with synaptic memory of physiological state.

(A) A core circuit in which AGRP neurons synaptically inhibit POMC neurons and are regulated by circulating hormones is controlled by ghrelin-responsive excitatory synapses. These synapses give this circuit a memory property based on an AMPK-dependent positive feedback loop (inset), which can be reversed by POMC neuron output, likely β-endorphin.

(B) A heuristic for the logic of this circuit is the SR flip-flop memory storage circuit. In the analogy with the neural circuit reported here, the set signal is ghrelin, which activates the green NOR gate, representing the conglomeration of AGRP neurons and their ghrelin-sensitive excitatory presynaptic terminals. The reset signal is leptin which interacts with POMC neurons represented as the blue NOR gate. Notably, when R and S are both high, the circuit does not support memory, and this condition is consistent with the case of ghrelin treatment of fed mice where opioid signaling is sufficiently high to prevent persistent synaptic upregulation (Figure 5G).

The investigation of synaptic regulation of AGRP neurons provided insight into the mechanism of ghrelin-evoked feeding. AGRP neurons are required for ghrelin’s orexigenic properties (Luquet et al., 2007). Previous reports concluded that ghrelin activates AGRP neurons directly (Andrews et al., 2008; Cowley et al., 2003). A model of ghrelin-induced feeding by Andrews, et al. (2008) suggests that Ghsr1 signaling in AGRP neurons elicits firing by AMPK-mediated alteration of fatty acid oxidation and regulation of mitochondrial uncoupling proteins. However, it is unclear how these metabolic pathways influence electrical activity, except perhaps by increasing available intracellular energy stores.

Our findings support an alternative model in which ghrelin acts at presynaptic receptors to increase glutamate release and activate AGRP neurons through ionotropic glutamate receptors. This is unlikely related to increased excitatory synapse number based on previous ultrastructural quantification following ghrelin treatment (Pinto et al., 2004). We suggest that metabolic pathways for ghrelin action in AGRP neurons (Andrews et al., 2008) are likely responsible for altering neuropeptide gene expression or addressing metabolic requirements to maintain AGRP neuron firing but are not directly involved in the increase of neuron firing following ghrelin administration.

We also find that AMPK participates in a positive feedback loop, resulting in synaptic hysteresis. Hysteresis for AMPK activation was investigated previously in Xenopus oocytes, but, while non-reversing activation was observed with transient antimycin treatment, this was attributed to irreversible metabolic block, not bistability, and other activators of AMPK did not show hysteresis (Martianez et al., 2009). Thus, hysteresis for AMPK activity is not elicited in all cells.

Hysteresis is a hallmark of bistability (Bhalla and Iyengar, 1999), which confers memory capacity in biological systems, often in conjunction with positive feedback loops (Ferrell, 2002). Other kinase-dependent positive feedback loops have been considered previously for long term information storage in the central nervous system (Bhalla and Iyengar, 1999; Tanaka and Augustine, 2008). The AMPK-mediated feedback described here persistently upregulates activity in presynaptic terminals for at least 5 h, which could serve as a memory of the hormone ghrelin in the form of sustained synaptic activity onto AGRP neurons.

Synapses with positive feedback also require a separate inhibitory signal to reverse synaptic upregulation, for example once energy balance is restored. Leptin is well suited for this role as it signals energy reserves by falling rapidly during food deprivation and rises with re-feeding (Boelen et al., 2006). We find that leptin is sufficient to counter-regulate elevated presynaptic activity onto AGRP neurons through an opioid receptor-dependent pathway. Leptin activates POMC neurons (Cowley et al., 2001), and it is likely that β-endorphin is the opioid released from POMC neurons to downregulate elevated synaptic activity. Interestingly, rapid synaptic plasticity at AGRP neurons was previously described in leptin-deficient ob/ob mice after leptin administration (Pinto et al., 2004); future work could investigate if this involves the opioid-dependent mechanism described here.

Previous studies showed selective ablation of only the β-endorphin-encoding portion of the pomc gene yields mice that are hyperphagic and overweight (Appleyard et al., 2003). This was unexpected in the context of β-endorphin intracranial administration, which activates multiple opioid pathways and leads to overeating (Grandison and Guidotti, 1977), but it is consistent with the anorexigenic function of POMC neurons. In light of the experiments reported here, β-endorphin ablation appears to disable a key regulatory system that may include ghrelin-mediated synaptic activity onto AGRP neurons, consequently potentiating ghrelin’s feeding stimulatory function.

The circuit described here could be further elaborated by considering the origin of the ghrelin-sensitive glutamatergic inputs. Currently, these neurons are not known, but they should originate from Ghsr1-expressing populations, which are clustered in several brain regions (Zigman et al., 2006). Retrograde tracing studies (Li et al., 1999) (DeFalco et al., 2001) indicate that a subset of these brain regions project to the arcuate nucleus and to AGRP neurons. Future work could examine the identity of these presynaptic neurons, which would provide an entry point for genetically-encoded tools to perturb presynaptic pathways with kinase mutants or optogenetic tools. In addition, upstream populations projecting to AGRP neurons might also be expected to influence feeding behavior, possibly extending control of AGRP neuron electrical activity beyond the contexts of hormonal and homeostatic regulation discussed in this study.

Why might neural circuits responding to hormonal cues require a memory capacity?

These experiments provide evidence of a circuit which has the capacity to store a memory of specific hormonal states (Figure 7A). Heuristically, its properties are analogous to the set/reset (SR) flip-flop memory storage circuit from digital electronics which consists of two interconnected NOR logic gates (Horowitz and Hill, 1989) (Figure 7B). Such a memory mechanism is consistent with the logic for hormonal control of synaptic upregulation and counter-regulation described in this study.

SR flip-flop memory has useful features for physiologic regulation. In this model, two hormones (“set”: ghrelin and “reset”: leptin) separately signal deficit or surfeit. The memory properties of the SR flip-flop act as an “until” loop, maintaining activity in response to S (even after S is turned off) until R is true. By using two discrete signals that reflect different physiological states, the circuit detects a state change. This model does not have a set point, but instead, it has a set range which is defined by the respective thresholds for activation of S and R. The operation of this circuit differs from homeostatic set point control theory in which a physiological parameter is continuously compared to a reference value and results in commensurate adjustment of physiological or behavioral output. This two component circuit model is, however, evocative of Walter Cannon’s original formulation of the homeostasis concept which referred to the balance of opposing factors (Cannon, 1929). In this circuit, though, we add the provision that these signals operate as discrete Boolean logical operations (see Figure 7B) due to an AMPK-dependent positive feedback loop. Importantly, this circuit displays memory properties only under sufficient deprivation to significantly reduce R (leptin levels or opioid tone), which is what we observe experimentally.

A potential advantage of a physiological set range for energy balance is that it avoids constant readjustment of energy stores as required for defending a set point, which may facilitate pursuit of other behaviors when energy repletion is not critical. Moreover, modulation of the response threshold for ghrelin, leptin, or opioids could shift this range and, potentially, food intake and body weight. Another implication of this state sensing property is that it could, in principle, simplify requirements for the regulatory control of ghrelin and leptin release from peripheral endocrine cells such that they need not reflect the integrated energy balance of the organism but instead to simply report a state of deficit and surfeit, respectively. For example, high ghrelin levels resulting from food-deprivation have been shown to fall within 1-2 h of re-feeding, due to negative regulation by nutrients such as carbohydrates (Tschop et al., 2000), even though energy homeostasis may not be achieved in this short time. However, the ghrelin-initiated synaptic upregulation and the corresponding firing rate elevation in AGRP neurons were maintained after 24 hours of re-feeding (Figure 5A,B), during which energy stores are restored. The operation of this memory system extends the responsiveness of the circuit beyond the lifetime of the triggering hormone until a signal of energy surfeit resets the circuit, in this case by opioid receptor activation. Such a mechanism may also be important in neural circuits governing behavioral responses to other physiological conditions. Furthermore, the relative ubiquity of these signaling pathways may extend the findings of positive feedback and set/reset bistability to other neuromodulators in the brain.

EXPERIMENTAL PROCEDURES

Experimental protocols were conducted according to U.S. National Institutes of Health guidelines for animal research and were approved by the Institutional Animal Care and Use Committee at Janelia Farm Research Campus. Food deprivation was for 24 h.

For several experiments, the experimenter was blinded to the identity of the pharmacological reagent or to the deprivation state of the animal [Figure 1B,D,E; Figure S1; Figure 2A (except ghrelin treatment); Figure 3A,B; Figure 5F; Figure 6].

Electrophysiology, photostimulation, methods are provided in Supplemental Experimental Procedures.

In figures, bars represent separate groups of cells, and symbols connected by lines represent manipulations applied across a group of cells. Data are represented as mean ± s.e.m. P values for pairwise comparisons were calculated by two-tailed Student’s t-test. P values for comparisons across more than two groups were adjusted with Holm’s correction (Holm, 1979). Tests involving 1-way ANOVA or 2-way ANOVA with one factor repeated measures were calculated with SigmaPlot (Systat). n.s. P > 0.05, *P < 0.05, **P < 0.01, ***P < 0.001.

Highlights.

• -Ghrelin and food deprivation induce synaptic plasticity at AGRP neurons
• -An AMPK-dependent positive feedback loop results in synapse bistability
• -Leptin and POMC opioids switch off AMPK, positive feedback, and persistent activity
• -Bistable synapses can store a reversible memory of hormonal state in hunger circuits