To examine neural circuits for feeding initiation, we identified neurons directly postsynaptic to gustatory sensory axons in the central brain. We utilized the full adult fly brain (FAFB) electron microscopy (EM) volume (Zheng et al., 2018) to manually reconstruct neurons postsynaptic to 17 labellar GRN axons in the right hemisphere that likely correspond to sugar-sensing GRNs (Engert et al., 2021). Fifteen second-order taste neurons and their synapses were fully reconstructed (Figure 1A–B and Figure 1—figure supplement 1) in the CATMAID platform (Saalfeld et al., 2009). The second-order neurons do not receive inputs from all candidate sugar GRNs (Figure 1—figure supplement 1B); however, as sugar GRNs make extensive axonal-axonal connections (Engert et al., 2021), the impact of these differences is unclear.

To assess the completeness of our second-order collection, we compared these 15 second-order neurons with the recently released Flywire dataset, a dense, machine learning based reconstruction of FAFB neurons (Dorkenwald et al., 2022; Eckstein et al., 2020). This comparison revealed that we identified 14 of the 15 neurons with the most synapses from sugar GRNs. These second-order neurons represent 12 unique cell types (Figure 1—figure supplement 1F). The 15 second-order neurons we manually reconstructed receive 21% of sugar GRN synaptic outputs. We note that the distribution of second-order neurons has a very long tail, likely due in part to small neural fragments that are challenging to reconstruct. These second-order neurons have not previously been characterized, except for G2N-1, which was identified as a candidate second-order gustatory neuron based on anatomical proximity to sugar-sensing GRNs (Miyazaki et al., 2015). Each of the second-order neurons is a local SEZ interneuron with arbors that overlap extensively with sugar GRN termini.

To test whether second-order gustatory neurons participate in feeding behaviors, we identified split-Gal4 lines that provide specific genetic access to individual second-order cell types, using NBLAST comparisons (Costa et al., 2016) to a library of SEZ split-Gal4 lines (Sterne et al., 2021). This provided split-Gal4 matches for seven second-order neurons (Figure 1C and Figure 1—figure supplement 2A). Additionally, we used intersectional approaches to gain genetic access to two additional second-order neurons, Cleaver (Figure 1—figure supplement 2A) and Zorro (Figure 1C). These genetic reagents are exquisitely specific for each of the nine second-order gustatory neurons, providing the opportunity to evaluate their function.

As activation of sugar-sensing GRNs on the proboscis labellum causes the fly to extend its proboscis to initiate feeding (Dethier, 1976), we tested whether activation or inhibition of second-order taste neurons influences this behavior. We expressed the red-shifted channelrhodopsin CsChrimson (Klapoetke et al., 2014) selectively in each second-order taste neuron, activated each with 635 nm light, and examined the proboscis extension response (PER). Remarkably, optogenetic activation of seven of the nine second-order taste neurons elicited proboscis extension (Figure 1D and Figure 1—figure supplement 2B). Moreover, inhibiting the activity of each second-order neuron individually, by optogenetic activation of the anion channelrhodopsin GtACR1 (Mohammad et al., 2017), reduced proboscis extension to 50 mM sucrose in food-deprived flies, for six of the seven second-order neurons that elicited PER upon activation (Figure 1E). At a higher sucrose concentration (100 mM), neural inhibition of only two of the second-order neuron classes decreased proboscis extension (Figure 1—figure supplement 2C). These studies argue that multiple second-order neurons contribute to normal feeding initiation behavior and suggest that the partial redundancy of these second-order neurons ensures robust feeding.

How does activation of a diverse set of second-order neurons drive proboscis extension? Proboscis motor neurons 4, 6, 7, and 9 are involved in extending different segments of the proboscis for feeding initiation (McKellar et al., 2020). We focused on the well-studied motor neuron 9 (MN9), which is necessary and sufficient for extension of the rostrum, the largest portion of the proboscis (Gordon and Scott, 2009; McKellar et al., 2020). We located MN9 in the FAFB EM volume by examining large SEZ neurons that lack synaptic output. To identify a pathway from taste detection to proboscis extension, we reconstructed presynaptic partners of MN9 and postsynaptic partners of second-order taste neurons.

This strategy identified a minimal pathway from taste detection to proboscis extension, composed of interconnected second-order neurons and third-order neurons each receiving inputs from a subset of second-order neurons, and feedforward premotor neurons (Figure 2A and Figure 2—figure supplement 1A). The third-order neurons represent a small subset based on comparisons to Flywire automated reconstructions (Figure 2—figure supplement 1B). They include one previously characterized neuron, the putative feeding command neuron, Fdg (Flood et al., 2013; Figure 3B), and a set of descending neurons, Bract, that project to the ventral nerve cord (Sterne et al., 2021). The premotor neurons are strongly connected to MN9, representing approximately 13% of the synaptic input onto MN9 (Figure 2—figure supplement 1C). There are direct connections between three second-order neurons (G2N-1, Zorro, and FMIn) and premotor neurons, and additional paths via third-order neurons to premotor neurons. The neurons in the feeding initiation circuit are predicted to be cholinergic, except for Usnea and Phantom which are predicted to be GABAergic (Figure 2—figure supplement 1D), based on machine learning classifications (Eckstein et al., 2020).

To investigate the function of deeper layers of this circuit, we identified split-Gal4 lines that selectively label two third-order neurons and one premotor neuron (Figure 2C) using NBLAST comparisons with SEZ split-Gal4 lines (Sterne et al., 2021). Optogenetic activation of third-order or premotor neurons with CsChrimson revealed that each cell type elicits robust proboscis extension (Figure 2D). However, acute inhibition of the third-order or premotor neurons with GtACR1 did not influence PER to 50 mM sucrose (Figure 2E), consistent with multiple pathways to proboscis motor neurons. Thus, by combining EM tracing studies with precise neural manipulations afforded by split-Gal4 lines, we have elucidated a neural circuit that promotes feeding initiation upon sweet taste detection.

To examine how taste information is processed by the feeding initiation circuit to guide feeding decisions, we monitored taste-induced activity of each neuron in the circuit. The proboscis was stimulated with water, sugar, or bitter taste solutions, while monitoring GCaMP6s calcium activity (Chen et al., 2013) in live flies (Harris et al., 2015). Eight of the ten neural classes responded to sugar taste presentation in food-deprived animals, and not to water or bitter solutions (Figure 3). Two second-order cell types responded to water taste detection: Usnea responded specifically to water and Phantom responded equally to water and sugar detection. Usnea and Phantom are reciprocally connected with GRNs (Figure 1—figure supplement 1D), suggesting that these second-order cell types may tune GRN responses in the presence of water. One third-order neuron, Fdg, did not respond to proboscis taste stimulation, but did respond to optogenetic activation of sugar-sensing GRNs (Figure 3—figure supplement 1A), suggesting that Fdg may respond to pharyngeal or leg taste detection. Together, these studies reveal that sugar taste is processed by a multilayered neural circuit to initiate feeding in food-deprived flies.

To test whether responses in the proboscis extension circuit are altered based on specific nutrient needs, we examined taste responses in flies that were thirsty rather than hungry. High hemolymph osmolality is a key signal of thirst that acts on central neurons to promote water consumption (Jourjine et al., 2016). We mimicked a thirsty-like state by increasing hemolymph osmolality, which enhanced water responses in water-sensing GRNs (Figure 3—figure supplement 1B). In four of the five central neurons tested, response profiles were similar in food-deprived and thirsty-like flies (Figure 3 and Figure 3—figure supplement 1C). However, one second-order neuron, Clavicle, responded to water and to sugar taste detection in a thirsty-like state but only to sugar in a hungry state (Figure 3 and Figure 3—figure supplement 1C). These results suggest that state-dependent responses to water at a single node (Clavicle) may tune the responsiveness of the pathway to bias acceptance of more dilute sugar solutions. However, the majority of neurons uncovered here responded selectively to sugar taste detection regardless of whether the animal was hungry or thirsty.

Given that two second-order neurons, Usnea and Phantom, responded to water taste stimulation despite synaptic connectivity to candidate sugar GRNs, we examined the connections of the second-order feeding initiation neurons from all GRNs of the right hemisphere (Engert et al., 2021). These GRNs have previously been clustered into candidate taste categories (sugar, water, bitter, high salt, and low salt) based on their morphology and GRN-GRN connectivity. Remarkably, we found that the second-order neurons that receive sugar GRN inputs also receive inputs from candidate water and high-salt (ppk23-positive, Glut-positive) GRNs but do not receive inputs from candidate bitter or low salt (Ir94e) GRNs (Figure 3—figure supplement 1D). The connectivity is consistent with our calcium imaging studies showing responses to sugar taste detection but not bitter taste. However, responses to sugar and water were not consistent with the predicted connectivity for each neuron, suggesting the possibilities of state-dependence, network interactions, and/or errors in GRN modality categorization. As GRN category assignments in the EM dataset were based on anatomy and connectivity alone, some GRNs may be misclassified, leading to errors in assessing sensory inputs (Engert et al., 2021). While the connectivity suggests the exciting possibility that taste integration may occur at second-order neurons, further functional studies will be necessary to illuminate the taste categories that activate or inhibit individual second-order neurons under different nutritive states. Nevertheless, our studies demonstrate that most second-order neurons respond to sugar taste stimulation but not to water or bitter tastes.

How is sugar taste information integrated with hunger state to promote feeding initiation in food-deprived flies? Hunger modulates sugar GRN activity (Inagaki et al., 2012); however, whether sensory gating is the only mechanism for hunger regulation or whether modulation of central neurons contributes to an altered network state in hungry animals has not been examined. To comprehensively investigate how taste detection is integrated with hunger state to initiate feeding, we optogenetically activated each neuron in the PER circuit in either fed or food-deprived flies and examined behavior. Optogenetic activation has the advantage of bypassing changes in sugar sensory detection that propagate through the circuit, enabling the evaluation of central circuit changes.

We reasoned that activating neurons upstream of or at the node(s) where hunger modulation occurs would cause differences in proboscis extension rates between hungry and fed flies, whereas activating neurons beyond the site where hunger impinges would not. Indeed, CsChrimson-mediated activation of sugar GRNs caused higher proboscis extension rates in food-deprived flies than in fed flies, whereas activation of MN9 elicited the same proboscis extension rate in food-deprived and fed flies (Figure 4A–C). Moreover, activation of two second-order neurons, G2N-1, and Clavicle, increased proboscis extension in food-deprived flies, whereas activation of all other neural classes did not (Figure 4A and D). Thus, hunger signals act on sensory neurons to increase detection sensitivity and on a specific set of second-order interneurons to amplify sugar pathway activation and promote feeding.

Animals evaluate both internal nutritive state and food quality to decide whether to initiate feeding. To investigate how food quality alters feeding initiation, we examined how the detection of bitter compounds is integrated with sugar taste information in the feeding initiation circuit. Previous studies have demonstrated that bitter compounds inhibit sugar-sensing gustatory neurons to prevent feeding (Chu et al., 2014; French et al., 2015; Jeong et al., 2013; Meunier et al., 2003), but have not addressed how downstream neural circuitry modulates appetitive feeding behaviors in response to bitter taste detection. To investigate central mechanisms of bitter modulation, we examined whether pathways from bitter GRNs intersect with the feeding initiation pathway.

As bitter GRNs do not directly synapse with neurons in the feeding initiation circuit, we asked whether second-order bitter neurons synapse onto the feeding initiation pathway. We reconstructed neurons downstream of bitter GRNs (Engert et al., 2021) in the EM volume and identified a second-order bitter neuron, Scapula, which receives over 150 synapses from bitter GRNs and is the second-most strongly connected cell type with bitter GRNs (Figure 5A and Figure 5—figure supplement 1). Scapula synapses directly onto two feeding initiation premotor neurons, Roundup and Rounddown, but not onto second- or third-order appetitive taste neurons.

Because bitter taste detection inhibits proboscis extension, we hypothesized that Roundup and Rounddown would be inhibited by Scapula to prevent proboscis extension to sugar in the presence of bitter compounds. Consistent with this hypothesis, Scapula is predicted to release glutamate (Figure 2—figure supplement 1D), which is often an inhibitory neurotransmitter in Drosophila (Liu and Wilson, 2013). To test this, we monitored activity in Roundup by in vivo calcium imaging while activating sugar GRNs, bitter GRNs, or both, using optogenetics to bypass sensory modulation. Roundup responded to optogenetic activation of sugar GRNs but not bitter GRNs (Figure 5B), as expected based on its response to taste compounds (Figure 3B). Upon co-activation of sugar and bitter GRNs, the Roundup response was dramatically decreased compared to the response to sugar GRN activation alone, arguing that bitter signals suppress the feeding initiation pathway. To test whether this bitter suppression reflects a central mechanism acting at Roundup, we monitored activity in a second-order neuron, G2N-1, directly upstream of Roundup, and found that its response upon co-activation of sugar and bitter GRNs was indistinguishable from its response to sugar GRN activation alone (Figure 5C). Together, the EM and imaging studies demonstrate that sugar and bitter tastes are integrated at feeding initiation premotor neurons, providing a central mechanism to reject sweet foods laced with bitter compounds.

Previous studies in Drosophila have identified gustatory neurons, motor neurons, and three candidate interneurons that influence feeding initiation (Flood et al., 2013; Gordon and Scott, 2009; Kain and Dahanukar, 2015; McKellar et al., 2020; Miyazaki et al., 2015; Talay et al., 2017). Here, by elucidating a complete sensorimotor circuit with synaptic resolution, we provide a comprehensive view of the neural pathway that elicits proboscis extension, the first step in feeding. The feeding initiation pathway is a local circuit, with three- and four-synaptic relays to motor output. Each neuron elicits proboscis extension upon optogenetic activation, demonstrating that each neuron participates in a pathway for the behavior. Inhibiting activity of single second-order neurons reduced the behavioral response, whereas inhibiting activity of third-order or premotor neurons did not. As inhibiting activity of single neurons did not abolish proboscis extension, this demonstrates that additional paths not requiring the individual neural cell type contribute to this innate behavior. These results are consistent with the circuit connectivity, which reveals that there are multiple routes between sugar GRNs and MN9 for proboscis extension. The multiple paths from second-order neurons to premotor neurons may enable proboscis extension to be recruited in different contexts to ensure robust feeding. More generally, multiple circuit paths may enhance behavioral flexibility by facilitating sensory tuning and multisensory integration (Miroschnikow et al., 2018; Ohyama et al., 2015).

Consistent with our finding that there are multiple pathways for proboscis extension, EM circuit tracing of sensorimotor paths for feeding in Drosophila larvae revealed both direct GRN-motor neuron synapses as well as indirect pathways (Miroschnikow et al., 2018). The functional significance of this circuit organization has not been explored. Although we did not identify direct GRN-motor neuron synaptic connections for proboscis extension in adult Drosophila, other feeding subprograms such as pharyngeal pumping may utilize direct sensory to motor connections and remain to be discovered.

Most neurons in the proboscis extension pathway respond to sugar taste detection, but not to water or bitter tastes, in food-deprived flies, demonstrating a direct line from sweet taste detection to the motor output for feeding. How water taste modulates proboscis extension in thirsty flies will require further study. Importantly, we identified and characterized three second-order neurons that respond to water taste detection: one is selective for water taste, another responds to both water and sugar tastes, and the third shows state-dependent sugar and water taste responses. Further study of these second-order neurons and their connectivity will be critical to evaluate the degree of separation or convergence of water and sugar pathways for feeding initiation. Moreover, we find that second-order neurons that respond to sugar receive inputs from sugar GRNs as well as candidate water and high-salt GRNs. This hints at the exciting possibility that second-order neurons integrate multiple taste modalities and may show responses that vary based on nutritive state. Furthermore, functional studies will be essential to assess taste integration by second-order neurons.

Of the interneurons identified here, only G2N-1 and Fdg have previously been implicated in feeding. G2N-1 was identified as a candidate sugar-sensing second-order neuron based on its anatomical proximity to gustatory axons alone (Miyazaki et al., 2015); here, we elucidated its functional role in taste detection and feeding initiation. Fdg was isolated as a ‘feeding command neuron,’ able to elicit multiple steps in feeding, including proboscis extension (Flood et al., 2013). In our calcium imaging studies, Fdg did not respond to proboscis taste stimulation but did respond to optogenetic activation of sugar GRNs. This suggests that Fdg may receive gustatory signals from GRNs on the pharynx or legs. Our studies demonstrate that Fdg is a third-order neuron in the feeding initiation pathway, with synaptic connections to Bract descending neurons. A description of the reconstruction of all FlyWire DNs is in preparation (K. Eichler, M. Costa, G. Card, G. Jefferis, personal communication). As Bract synapses with proboscis premotor neurons and ventral cord circuits, Fdg and Bract are well-poised to coordinate proboscis extension with other steps in feeding.

The architecture of the circuit provides a platform to investigate how taste signals are transformed in the brain to drive behavior. In this study, we focused on MN9, the rostrum protractor motor neuron that elicits proboscis extension, as a key readout of proboscis extension behavior. However, proboscis extension involves not only rostrum protraction but also extension of the haustellum and opening of the labellum, controlled by additional motor neurons (McKellar et al., 2020). We hypothesize that the connectivity among second- and third-order neurons may coordinate the precise temporal activation of different muscle groups for coordinated extension. Moreover, proboscis extension is followed by ingestion and then meal termination (Dethier, 1976; Pool and Scott, 2014). Continued expansion and exploration of this pathway will provide the opportunity to examine how different feeding subprograms are timed and coordinated to elicit feeding in natural environments.

Studies in C. elegans, Drosophila, and mammals have demonstrated that a key site of hunger regulation is at the peripheral chemosensory neurons, altering sensitivity of detection (Chalasani et al., 2010; Kawai et al., 2000; Root et al., 2011; Savigner et al., 2009; Sengupta, 2013). For example, dopamine enhances the sensitivity of Drosophila sugar-sensing gustatory neurons to promote proboscis extension at lower sucrose concentrations in hungry animals (Inagaki et al., 2012; Marella et al., 2012). Hunger modulation of taste processing beyond sensory neurons has been more challenging to evaluate, both because of lack of knowledge of central networks and because changes at the sensory level propagate through the network.

To isolate the role of central brain neurons in hunger modulation, we used the precise genetic access available in Drosophila to activate each node of the feeding initiation pathway and examined the behavioral response elicited in fed and food-deprived flies. These studies pinpoint the site of hunger modulation to sensory neurons and two second-order neurons. Although caveats of artificial stimulation exist, the consistent changes seen across different light intensities for neural manipulation early in the pathway, but not downstream, argue that these results are robust. These studies demonstrate that hunger acts at a few critical nodes to modulate feeding initiation: sensory neurons increase detection sensitivity and second-order neurons amplify pathway activation. It will be interesting to examine whether hunger modulation of sensory and second-order neurons occurs independently or over different time scales to adjust behavioral responses as starvation increases. In addition, the specific hunger signals that act on central neurons and their mechanism of modulation may now be explored.

While previous studies have demonstrated interactions between sweet and bitter taste modalities at the level of sensory neurons (Chu et al., 2014; French et al., 2015; Jeong et al., 2013; Meunier et al., 2003 LeDue et al., 2016) and through feedback from the mammalian gustatory cortex (Jin et al., 2021), this study reveals a third circuit strategy for weighing sweet and bitter tastes: a local inhibitory network. Inhibitory interactions between bitter and sugar pathways at the level of premotor neurons provide an elegant strategy to weigh incoming sugar and bitter taste information and adjust behavioral probability. In addition, by blocking activity at specific muscles, bitter detection may specifically change behavior to direct the proboscis away from a hazardous food source. The existence of a local inhibitory circuit for bitter-sweet integration has been recently postulated based on studies of mammalian taste circuitry (Jin et al., 2021) and may be a shared strategy across species. These multiple circuit mechanisms for suppression of sweet attraction by bitter signals may reflect the evolutionary importance of robust bitter taste avoidance.

By examining a complete sensorimotor pathway, we elucidate how a complex nervous system orchestrates the decision to initiate feeding and illuminate central modules that integrate taste detection with internal state. These central controls afford independent amplification and suppression of feeding and stand apart from sensory modulation as mechanisms that dynamically tune behavior. As sensory modulation may suffer from finite amplification and incomplete suppression, central modulation provides a strategy to bypass those limits, allowing a broader range and different temporal dynamics of modulation.

Statistical tests for behavioral assays were performed in Prism. For analysis of Proboscis Extension Response assays, Fisher’s Exact Test was used in comparing the fraction of PER responses in experimental versus control flies. Statistical analysis of calcium imaging was carried out in R and Python. For imaging experiments carried out in a block design with three treatments, Quade tests were carried out in R using the PMCMRplus package (Pohlert, 2021). Quade test was chosen because it is more powerful than Friedman for a block-design experiment with three treatments (Conover, 1999). Other statistical analyses of calcium imaging were carried out in Python using the SciPy (Virtanen et al., 2020) and scikit-posthocs packages (Terpilowski, 2018).