To systematically characterize gustatory inputs and outputs, we traced gustatory axons in the FAFB volume (Zheng et al., 2018). Tracing was performed manually using the annotation platform CATMAID (Saalfeld et al., 2009). The GRNs from the proboscis labellum send axons through the labial nerve to the SEZ (Figure 1A). The labial nerve is a compound nerve, carrying sensory axons from the labellum, maxillary palp, and eye, as well as motor axons innervating proboscis musculature (Hampel et al., 2017; Hartenstein et al., 2018; Miyazaki and Ito, 2010; Nayak and Singh, 1983; Rajashekhar and Singh, 1994). Different sensory afferents occupy different domains in the SEZ, with labellar gustatory axons terminating in the ACSC (Hartenstein et al., 2018; Miyazaki and Ito, 2010; Thorne et al., 2004; Wang et al., 2004; Figure 1A). Therefore, to trace gustatory axons, we began by tracing neurites in the right labial nerve, readily identifiable in the EM dataset (Figure 1B and C), and selected fibers that terminated in the anterior central SEZ to trace synaptic completion (Zheng et al., 2018).

In tracing axons, we found that neurites with small- to medium-sized diameters in the dorsomedial labial nerve (Figure 1C) projected along a single neural tract (Figure 1D) to the anterior central region of the SEZ. This neural tract served as an additional site to select arbors for reconstruction. Individual fibers followed along the same tract and showed variation in terminal branching (Figure 1E). In total, we identified 87 axonal projections in the right hemisphere. Tracing from the left labial nerve and neural tract in the left hemisphere, we identified 57 additional projections. Misalignments in the EM volume precluded identification of additional GRNs in the left hemisphere. Because there are 90–104 GRNs per labellum (Jaeger et al., 2018; Stocker, 1994), we estimate that we have identified 83–96% of the GRN fibers from the right labellum and 54–63% from the left. The projections from the left and right labial nerves are symmetric and converge in a dense web in the anterior central SEZ (Figure 1F). This arborization pattern recapitulates the labellar sensory projections of the ACSC (Hartenstein et al., 2018). We confirmed that the reconstructed neurites overlap with the known projection pattern of sugar and bitter GRNs in the registered fly brain template (Figure 1—figure supplement 1; Bogovic et al., 2020), demonstrating that we have identified and traced GRNs.

In addition to the skeleton reconstructions, we manually annotated pre- and postsynaptic sites. The presence of T-shaped structures characteristic of presynaptic release sites (‘T bars’), synaptic vesicles, and a synaptic cleft was used to identify a synapse, consistent with previous studies (Zheng et al., 2018). Synapses are sparse along the main neuronal tract and abundant at the terminal arborizations (Figure 1E). Each GRN has a large number of pre- and postsynaptic sites intermixed along the arbors (Figure 1E and G–I), characteristic of fly neurites (Bates et al., 2020a; Meinertzhagen, 2018; Olsen and Wilson, 2008; Takemura et al., 2017). On average, a GRN contains 175 (±6 SE) presynaptic sites and 168 (±6 SE) postsynaptic sites, with individual GRNs showing wide variation in pre- and postsynapse number (Figure 1—figure supplement 1B). GRNs are both pre- and postsynaptic to other GRNs, with each GRN receiving between 2% and 66% (average = 39%) of its synaptic input from other GRNs (Figure 1—figure supplement 1C). The large number of synapses between GRNs suggests that communication between sensory neurons may directly regulate sensory output.

Drosophila GRNs comprise genetically defined, discrete populations that are specialized for the detection of specific taste modalities (Wang et al., 2004; Cameron et al., 2010; Jaeger et al., 2018). As the EM dataset does not contain molecular markers to distinguish between GRNs recognizing different taste modalities, we set out to identify subpopulations of reconstructed GRNs based on their anatomy and connectivity.

We performed hierarchical clustering of GRN axons to define different subpopulations based on their morphology and synaptic connectivity. GRNs of the right hemisphere were used in this analysis as the dataset is more complete. Each traced skeleton was registered to a standard template brain (Bogovic et al., 2020), and morphological similarity was compared pairwise using NBLAST in an all-by-all matrix (Costa et al., 2016). Then, GRN-GRN connectivity was added for each GRN skeleton and the resulting merged matrix was min/max scaled. We then used Ward’s method to hierarchically cluster GRNs into groups (Ward 1963). We chose six groups as the number that minimizes within-cluster variance (Figure 2—figure supplement 1A; Braun et al., 2010). Each group is composed of 7–23 GRNs that occupy discrete zones in the SEZ and share anatomically similar terminal branches (Figure 2).

To evaluate whether the different groups represent GRNs detecting different taste modalities, we compared the anatomy of each group in the right hemisphere with that of known GRN classes using mean NBLAST scores. We registered EM reconstructed GRN projections and GRN projections from immunostained brains to the same standard brain template for direct comparisons (Bogovic et al., 2020). For each group, we performed pairwise NBLAST comparisons with bitter (Gr66a; Wang et al., 2004; Thorne et al., 2004), sugar (Gr64f; Dahanukar et al., 2007), water (Ppk28; Cameron et al., 2010; Chen et al., 2010), and candidate low-salt (Ir94e; Croset et al., 2016; Jaeger et al., 2018) GRN projections. There is not a specific genetic marker for candidate high-salt projections as Ppk23 labels both bitter and high-salt GRNs (Jaeger et al., 2018). These comparisons (see section ‘NBLAST analysis for taste modality assignment’) yielded a GRN category best match for each group, illustrated by overlays in the three-dimensional standard fly brain template (Figure 3). Groups 1 and 2 best match bitter projections, forming a characteristic medial ringed web (Figure 3A and B). Group 3 projections show greatest similarity to low-salt GRNs, with distinctive dorsolateral branches (Figure 3C). Groups 4–6 are anatomically very similar, and identity assignments are tentative. Groups 4 and 5 best match sugar GRNs (Figure 3D and E). Because group 4 contains a dorsolateral branch seen in Gr64f projections and not seen in group 5 projections, we hypothesize that group 4 is composed of sugar GRNs and that the remaining group 5 is composed of high-salt GRNs. Group 6 best matches water GRNs (Figure 3F). Thus, morphological and connectivity clustering suggests molecular and functional identities of different GRNs.

An identical clustering analysis of GRNs from the left hemisphere yielded seven groups of 4–15 neurons (Figure 3—figure supplements 1–2). Groups 1 and 2 best match bitter projections and group 6 best matches low-salt projections (see section ‘NBLAST analysis for taste modality assignment’), with anatomy consistent with known projection patterns. Other groups are not well-resolved (see section ‘NBLAST analysis for taste modality assignment’), arguing that a more complete dataset is necessary to resolve GRN categories in the left hemisphere.

As GRNs have a large number of synaptic connections with other GRNs (Figure 1—figure supplement 1C), we examined whether synapses exist exclusively between neurons of the same group, likely representing the same taste modality, or between multiple groups. The all-by-all connectivity matrix illustrated blocks of connectivity within groups, with fewer connections between groups (Figure 4A). To quantify this, we summed all GRN-GRN connections within and between groups. This analysis revealed that most synapses are between neurons of the same group (79%), while only 21% of the synapses are between GRNs of different groups (Figure 4B). For example, group 4 neurons receive 1468 synapses from other group 4 neurons and 38 from group 3, 156 from group 5, and 130 from group 6 neurons. Focusing on connections of five or more synapses between GRN pairs, representing high-confidence connections (Buhmann et al., 2021; Li et al., 2020; Takemura et al., 2013; Takemura et al., 2015), resulted in the elimination of some but not all between-group connections (Figure 4C), with between-group connections representing only 10% of all GRN connections.

The large numbers of chemical synapses between GRNs within a group may provide a mechanism to amplify signals of the same taste modality. In contrast, weak connectivity between GRNs of different groups may serve to integrate taste information from different modalities before transmission to downstream circuitry. We note that misclassification of individual GRNs in the clustering analysis may result in over- or underestimates of GRN connectivity within and between groups.

Neurotransmitter predictions of GRNs, in general, do not predict a clear majority neurotransmitter (Figure 4—figure supplements 1 and 2; Eckstein et al., 2020). This suggests that GRNs may release multiple neurotransmitters or that neurotransmitter predictions should be considered uncertain until further testing (Eckstein et al., 2020).

To examine whether the small number of connections between GRNs of different taste modalities results in cross-activation of GRNs detecting different primary tastant classes, we tested if activation of one GRN class results in propagation of activity to other GRN classes in vivo. To test for interactions between GRNs of different taste modalities, we undertook calcium and voltage imaging studies in which we monitored the response of a GRN class upon activation of other GRN classes.

We expressed the calcium indicator GCaMP6s in genetically defined sugar-, water-, or bitter-sensitive GRNs to monitor excitatory responses upon artificial activation of different GRN classes. To ensure robust and specific activation of GRNs, we expressed the mammalian ATP receptor P2X2 in sugar, water, or bitter GRNs, and activated the GRNs with an ATP solution presented to the fly proboscis while imaging gustatory projections in the brain (Yao et al., 2012; Harris et al., 2015). Expressing both P2X2 and GCaMP6s in sugar, water, or bitter GRNs elicited strong excitation upon ATP presentation (Figure 5A–B and G–H, Figure 5—figure supplement 1C and D, Figure 5—figure supplement 2C and D, Figure 5—figure supplement 3C and D), demonstrating the effectiveness of this method. Activation of sugar or water GRNs did not activate bitter cells, nor did bitter cell activation elicit responses in sugar or water axons (Figure 5—figure supplement 1E–H, Figure 5—figure supplement 2E and F, Figure 5—figure supplement 3G and H). Similarly, we did not observe responses in sugar GRNs upon water GRN activation (Figure 5C and D, Figure 5—figure supplement 2I and J) or responses in water GRNs upon sugar GRN activation (Figure 5I and J, Figure 5—figure supplement 3E and F). To examine whether interactions between modalities are modulated by the feeding state of the fly, we performed the activation and imaging experiments in both fed and starved flies (Figure 5—figure supplements 1–6). These experiments did not reveal feeding state-dependent interactions between GRN populations. To examine whether inhibitory interactions might exist between two GRN classes, we expressed the voltage indicator ArcLight (Cao et al., 2013), which reliably reports hyperpolarization, in sugar GRNs while activating water GRNs via P2X2 and vice versa. These experiments revealed no change in voltage in one appetitive gustatory class upon activation of the other (Figure 5E–F and K–L, Figure 5—figure supplement 7). Overall, despite the potential for crosstalk between different modalities revealed by EM, we observed no communication between appetitive GRNs by calcium or voltage imaging of gustatory axons.

Experimental animals were maintained on standard agar/molasses/cornmeal medium at 25°C. For imaging experiments requiring food-deprived animals, flies were placed in vials containing wet kimwipes for 23–26 hr prior to the experiment. For behavioral experiments, flies were placed on food supplemented with 400 μM trans-retinal for 24 hr prior to the experiment.

Neuron skeletons were reconstructed in a serial sectioned transmission EM dataset of the whole fly brain (Zheng et al., 2018) using the annotation software CATMAID (Saalfeld et al., 2009). GRN projections were identified based on their extension into the labial nerve and localization to characteristic neural tracts in the SEZ. Skeletons were traced to completion either entirely manually or using a combination of an automated segmentation (Li et al., 2019) and manual tracing as previously described (Hampel et al., 2020). Chemical synapses were annotated manually and neurons were traced to synaptic completion using criteria previously described (Zheng et al., 2018). Skeletons were reviewed by a second specialist, so that the final reconstruction presents the consensus assessment of at least two specialists. Skeletons were exported from CATMAID as swc files for further analysis, and images of skeletons were exported directly from CATMAID. FAFB neuronal reconstructions will be available from Virtual Fly Brain (https://fafb.catmaid.virtualflybrain.org/).

GRNs were hierarchically clustered based on morphology and connectivity using NBLAST and synapse counts. First, GRN skeletons traced in FAFB were registered to the JRC2018U template (Bogovic et al., 2020) and compared in an all-by-all fashion with NBLAST (Costa et al., 2016). NBLAST analysis was carried out with the natverse toolkit in R (Bates et al., 2020b; R Development Core Team, https://www.r-project.org/). The resulting matrix of ‘normalized’ NBLAST scores was merged with a second matrix containing all-by-all synaptic connectivity counts for the same GRNs. The resulting merged matrix was min–max normalized such that all values fall within the range of 0 and 1. The merged, normalized matrix was hierarchically clustered using Ward’s method (Ward 1963) in Python (Python Software Foundation, https://www.python.org/) with SciPy (Virtanen et al., 2020). The number of groups was chosen based on analysis of Ward’s joining cost and the differential of Ward’s joining cost.

Connectivity data of GRNs was exported from CATMAID for further analysis, and connectivity diagrams were generated using Cytoscape (Shannon et al., 2003).

GRN skeletons traced in FAFB were registered to the JRC2018U template and summed in Fiji to create a composite stack of the combined morphologies of all individual GRNs in a given group (as assigned by morphology and connectivity clustering). The morphology of the composite stack for each group was compared to an image library of GRN projection patterns using NBLAST (Costa et al., 2016). The image library contained projection patterns of Gr66a-Gal4, Gr64f-Gal4, Ir94e-Gal4, and Gr64f-Gal4 brains, three per genotype, registered to the JRC2018U template, prepared as described (see section ‘Immunohistochemistry’). Group identity was assigned based on the top hit from the image library. Following NBLAST analysis, the anatomy of each group was compared to the projection pattern of its top hit using VVDViewer.

NBLAST of groups in the right hemisphere against known GRN categories yielded the following top GRN matches (mean NBLAST score): group 1, Gr66a-Gal4 #1 (0.77986); group 2, Gr66a-Gal4 #1 (0.83017); group 3, Ir94e-GAL4 #2 (0.73743); group 4, Gr64f-Gal4 #2 (0.80821); group 5, Gr64f-Gal4 #2 (0.81091); and group 6, Ppk28-Gal4 #1 (0.80059). NBLAST of groups in the left hemisphere against known GRN categories yielded the following top GRN matches (NBLAST score): group 1, Gr66a-Gal4 #3 (0.86974); group 2, Gr66a-Gal4 #3 (0.88230); group 3, Gr64f-Gal4 #2 (0.85942); group 4, Gr64f-Gal4 #2 (0.84788); group 5, Gr64f-Gal4 #2 (0.87164); group 6, Ir94e-Gal4 #2 (0.79400); and group 7, Gr64f-Gal4 #2 (0.78896).

For imaging studies of GRNs, mated females, 10–21 days post eclosion, were dissected as previously described (Harris et al., 2015), so that the brain was submerged in artificial hemolymph (AHL) (Wang et al., 2003) while the proboscis was kept dry and accessible for taste stimulation. To avoid occlusion of taste projections in the SEZ, the esophagus was cut. The front legs were removed for tastant delivery to the proboscis. AHL osmolality was assessed as previously described (Jourjine et al., 2016) and adjusted according to the feeding status of the animal. In fed flies, AHL of ~250 mOsm was used (Wang et al., 2003). The AHL used for starved flies was diluted until the osmolality was ~180 mOsm, consistent with measurements of the hemolymph osmolality in food-deprived flies (Jourjine et al., 2016).

Calcium transients reported by GCaMP6s and GCaMP7s were imaged on a 3i spinning disk confocal microscope with a piezo drive and a ×20 water immersion objective (NA = 1). For our studies of GRNs, stacks of 14 z-sections, spaced 1.5 µm apart, were captured with a 488 nm laser for 45 consecutive time points with an imaging speed of ~0.3 Hz and an optical zoom of 2.0. For better signal detection, signals were binned 8 × 8, except for Gr64f projections, which underwent 4 × 4 binning.

Voltage responses reported by ArcLight were imaged similarly to the calcium imaging studies. To increase imaging speed, the number of z planes was reduced to 10, and the exposure time was decreased from 100 to 75 ms, resulting in an imaging speed of ~0.7 Hz. To maintain a time course comparable to that of the calcium imaging experiments of GRNs, the number of time points was increased to 90. Signals were binned 8 × 8 in each experiment.

Taste stimuli were delivered to the proboscis via a glass capillary as previously described (Harris et al., 2015). For GRN studies, each fly was subjected to three consecutive imaging sessions, each consisting of a taste stimulation at time point 15, 25, and 35 (corresponding to 30, 50.5, 71.5 s). During the first imaging session, the fly was presented with a tasteless 20% polyethylene glycol (PEG, average molecular weight 3350 g/mol) solution, acting as a negative control. PEG was used in all solutions except water solutions as this PEG concentration inhibits activation of water GRNs (Cameron et al., 2010). This was followed in the second session with stimulations with 100 mM ATP in 20%PEG. In the last imaging session, each fly was presented with a tastant acting as a positive control in 20% PEG (Gr64f: 1 M sucrose; Gr66a: 100 mM caffeine, 10 mM denatonium benzoate; ppk28: H₂O; ppk23: 1 M KCl in 20% PEG).

Image analysis was performed in Fiji (Schindelin et al., 2012). Z stacks for each time point were converted into maximum z-projections for further analysis. After combining these images into an image stack, they were aligned using the StackReg plugin in Fiji to correct for movement in the xy plane (Thévenaz et al., 1998).

For our exploration of interactions between GRN subtypes, one region of interest (ROI) was selected encompassing the central arborization of the taste projection in the left or right hemisphere of the SEZ in each fly. Whether the projection in the left or right hemisphere was chosen depended on the strength of their visually gauged response to the positive control. The exception was Gr66a projections, in which the entire central projection served as ROI. If projections did not respond strongly to at least two of the three presentations of the positive control, the fly was excluded from further analysis. If projections responded to two or more presentations of the negative control, the fly was excluded from further analysis. A large ROI containing no GCaMP signal was chosen in the lateral SEZ to determine background fluorescence.

In calcium imaging experiments, the first five time points of each imaging session were discarded, leaving 40 time points for analysis with taste stimulations at time points 10, 20, and 30. The average fluorescence intensity of the background ROI was subtracted at each time point from that of the taste projection ROI. F0 was then defined as the average fluorescence intensity of the taste projection ROI post background subtraction of the first five time points. ΔF/F (%) was calculated as 100% * (F(t) - F0)/F0. Voltage imaging experiments were analyzed similarly, with 10 initial time points discarded for a total of 80 time points in the analysis and tastant presentations at time points 20, 40, and 60.

Graphs were generated in GraphPad Prism. To calculate the max ΔF/F (%) of GCaMP responses, the ΔF/F (%) of the three time points centered on the peak ΔF/F (%) after the first stimulus response were averaged. The average ΔF/F (%) of the three time points immediately preceding the stimulus onset were then subtracted to account for changing baselines during imaging. ArcLight data was similarly analyzed, except that five time points centered on the peak ΔF/F (%) and five time points prior to stimulus onset were considered.

Statistical analysis was performed using GraphPad Prism (GraphPad Software, La Jolla, CA). All reported values are mean ± SEM. Data were analyzed using ANOVA followed by Tukey’s multiple comparisons for multiple comparisons of parametric data.

To visualize GRN projections with light microscopy, males of Gr64f-Gal4, Gr66a-Gal4, Ir94e-Gal4, or Ppk28-Gal4 were crossed to virgins of UAS-Syt-HA, 20XUAS-CsChrimson-mVenus (attP18). Dissection and staining were carried out by FlyLight (Gr64f-Gal4 and Gr66a-Gal4) or in house (Ir94e-Gal4 and Ppk28-Gal4) according to the FlyLight ‘IHC-Polarity Sequential Case 5’ protocol (https://www.janelia.org/project-team/flylight/protocols). Samples were imaged on an LSM710 confocal microscope (Zeiss) with a Plan-Apochromat ×20/0.8 M27 objective. Images were then registered to the 2018U template using CMTK (https://www.nitrc.org/projects/cmtk) and manually segmented with VVDViewer (https://github.com/takashi310/VVD_Viewer; Otsuna et al., 2018; Kawase and Rokicki, 2022) in order to remove any nonspecific background.