ABSTRACT
The insect central complex and vertebrate basal ganglia are forebrain centres involved in selection and maintenance of behavioural actions. However, little is known about the formation of the underlying circuits, or how they integrate sensory information for motor actions. Here, we show that paired embryonic neuroblasts generate central complex ring neurons that mediate sensory-motor transformation and action selection in Drosophila. Lineage analysis resolves four ring neuron subtypes, R1-R4, that form GABAergic inhibition circuitry among inhibitory sister cells. Genetic manipulations, together with functional imaging, demonstrate subtype-specific R neurons mediate the selection and maintenance of behavioural activity. A computational model substantiates genetic and behavioural observations suggesting that R neuron circuitry functions as salience detector using competitive inhibition to amplify, maintain or switch between activity states. The resultant gating mechanism translates facilitation, inhibition and disinhibition of behavioural activity as R neuron functions into selection of motor actions and their organisation into action sequences.
INTRODUCTION
The selection and maintenance of a behavioural action is the outcome of specific neural interactions mediated by circuits that have evolved to coordinate adaptive behaviours. Essential to such circuits is their ability to organize sequences of actions by facilitating appropriate motor programmes while inhibiting competing ones. The underlying action selection mechanisms involve defined centres in the forebrain of arthropods and vertebrates including, respectively, the central complex (CX) and the basal ganglia (BG)1
In vertebrates, the BG mediate action selection by two main pathways that project from medium spiny GABAergic interneurons (MSN) of the striatum to the outer and inner shell of the Globus Pallidus. Striatal MSN that express dopamine D1 receptors form the direct pathway, whereas MSNs expressing dopamine D2 receptors constitute the indirect pathway. Although initially thought to act antagonistically, functional imaging studies reveal simultaneous activity of both direct and indirect pathways during behavioural sequences2–4, suggesting that both are involved, with different roles, in selecting and enhancing motor programs while inhibiting competing ones5–10. However, despite these advances, little is known about the underlying neural mechanisms, or how they are formed.
We recently identified extensive correspondences in neural organization and function of the BG with the CX of mandibulate arthropods1,10. In Drosophila, the CX comprises five centres: the protocerebral bridge (PB), fan-shaped body (FB), ellipsoid body (EB), noduli (NO) and lateral accessory lobes (LAL) (Figure 1). Similar to the BG, the CX, in particular its EB, plays an essential role in a multitude of behavioural manifestations, including goal-directed locomotion, spatial orientation, learning and memory, as well as a role in sleep, attention, arousal, and decision-making11–27. All of these behaviours require the selection and maintenance of motor actions, or their active suppression. Thus, within the context of the proposed homology of CX and BG circuitries, the many tools available for targeted cell and circuit-specific manipulations of the EB in Drosophila28, makes the invertebrate system ideal for investigating neural mechanisms and computations underlying action selection.
Here we use genetic tracing from the embryo into the adult to show that all four subtypes of EB ring neurons (henceforth called R neurons, R1-4) in the adult Drosophila CX derive from embryonic neuroblast ppd5. Anatomical and genetic analyses reveal that R neurons establish a connectivity network among clonally related sister cells. Calcium-imaging and electrophysiological recordings demonstrate that R2-4 neurons are inhibitory GABAergic and are inhibited by GABA-A receptor signalling. We show that targeted manipulations of subtype-specific R neurons reveal that EB microcircuitry is involved in sensory-motor transformation required for the selection and maintenance of motor actions and their organisation into action sequences. A neural model of the EB R neuron circuit based on these structural and functional data suggests computational mechanisms mediating sensory-motor transformation and motor action selection in Drosophila.
RESULTS
R neurons derive from embryonic neuroblast ppd5
In Drosophila, the EB is generated by R neurons. Their layer-specific projections form a toroidal neuropil located dorso-anteriorly straddling the midline of the protocerebrum29–31. Based on their layer-specific projections, R neurons are classified into four subtypes (R1-4) each resolved by subtype-specific Gal4 lines30,31 (Fig. 1 and Supplementary Fig. 1). In order to reconstruct the origin, formation and circuit assembly of the EB, we first examined pox neuro (poxn), which is expressed in adult R neurons14 as well as in embryonic basal forebrain lineages that co-express Engrailed32 (en) (Supplementary Fig. 2).
Protocerebral poxn+ lineages derive from embryonic stem cells, neuroblasts ppd5 and ppd8, that are distinguished by en and Dachshund (Dac) expression33, with Dac restricted to ppd8 (Fig. 2a-f). To determine whether adult EB poxn+ cells share their origin with embryonic lineages, we used genetic tracing34 to identify lineage relationships from the time of embryogenesis up to the formation of the adult brain employing a heritable membrane-tethered fluorescent marker (Fig. 2g).
en-Gal4 tracing labelled R neurons that are immunopositive for anti-Poxn (Fig. 2h). Similarly, poxn-Gal4 tracing identified poxn>tub>mCD8::GFP labelled R neurons and their neuropil, all of which are immunopositive for anti-Poxn (Fig. 2i). In contrast, Dac>tub>mCD8::GFP did not label R neurons and showed no overlap between protocerebral poxn+ cells and membrane-tethered GFP (Fig. 2j). Tracing with en-Gal4, poxn-Gal4 and Dac-Gal4 recapitulated endogenous en, poxn and Dac expression. We never observed de novo expression coinciding with poxn protocerebral cells, nor did we detect any evidence for cell migration into this basal forebrain region. Additionally, GFP labelling of subtype-specific Gal4 lines together with genetic tracing and anti-Poxn immunolabeling identified poxn expression in GFP-labelled R neuron subtypes (Fig. 2k-m and Supplementary Fig. 3a-d). In contrast, we did not detect anti-Dac immunoreactivity in poxn+ cells (Supplementary Fig. 3e-n). Because Dac expression distinguishes between neuroblasts ppd5 and ppd8, our data identify R1-4 neurons as lineage-related sister cells derived from a pair of bilateral symmetric stem cells, embryonic neuroblast ppd5.
R neurons form a connectivity network among sister cells
EB cells multiply during development and, similar to observations on grasshopper CX development35, establish an orthogonal axon scaffold that is later reshaped by morphogenesis into the EB’s characteristically toroidal architecture29–31. To determine R neuron connectivity, we utilized the protocerebral poxn enhancer element14 and generated poxn-Gal4 strains that either target subsets or the majority of poxn+ R neurons (Supplementary Fig. 4a,b). This allowed us to determine their connectivity pattern. We combined poxn-Gal4 (Fig. 3), or R1-4-specific Gal4 lines c105, c232 and EB1(Fig. 4a-d) with the presynaptic marker synaptotagmin::GFP and the postsynaptic marker DenMark16. Co-labelling revealed apposed Syt::GFP and DenMark punctae across layers and modules (Fig. 3b-d and Fig. 4a-d; Supplementary Movie 1). Super-resolution microscopy with structured illumination confirmed Syt::GFP and DenMark labelled apposing punctae in EB R neuropil (Supplementary Fig. 4c and Supplementary Movie 2). Clonal analysis identified ramified terminal synapses densely distributed along R neuron subtype- and layer-specific arborisations (Fig. 4e). These data indicate that Syt::GFP, DenMark labelled apposing punctae visualize synaptic connections of lineage-related R neurons (Fig. 4f).
To gain further insights into EB connectivity we next performed electron microscopy analyses of the EB neuropil. At low-resolution (Supplementary Fig. 4d) dense neuronal profiles are resolved in any segment of the EB, each encompassing an outer and two inner domains, which define the concentric arrangement of modules (also called ‘wedges’26). Supplementary Figure 4e demonstrates the location of synapses (right panel, in purple), which across sections allows an estimate of synaptic densities. Our analysis revealed that domains of the inner EB ring have smaller profiles with several presynaptic sites, many of which appose large profiles, also possessing presynaptic sites (Supplementary Fig. 4e,f). In addition, we also detected pairs of synapses opposing larger boutons, thus revealing synaptic convergences (Supplementary Fig. 4g). More rarely, tetrad synapses terminating onto four postsynaptic profiles contain dense core vesicles, indicating that neuromodulatory elements likely participate in EB functionality (Supplementary Fig. 4h).
To further substantiate connectivity among R neurons, we used GFP reconstitution across synaptic partners (GRASP). GRASP utilizes the expression of two split-GFP halves that only fluoresce when reconstituted37. To visualise potential connections between R neurons, we used split-GFP fragments UAS-CD4::spGFP1-10 and LexAop-CD4::spGFP11 (ref. 38), combined with R2/4-specific EB1-Gal4 (Fig. 4g and Supplementary Fig. 5a) and NV45-LexA::VP16 specific to R3 neurons (Fig. 4h and Supplementary Fig. 5b). When EB1-Gal4 and NV45-LexA were combined for GRASP, reconstituted GFP fluorescence was detectable around the EB’s circumference and intersecting modules, with most intense GFP signal observed at the intersection of R3 and R2/4 layers (Fig. 4n). However reconstituted GFP fluorescence was not detected in respective controls (Supplementary Fig. 5d). In contrast, super-resolution microscopy of GRASP signals revealed punctate (Fig. 4i-k) and elongated, grid-type GFP expression extending across R2-R4 layers and modules (Fig. 4l,m and Supplementary Movies 3 and 4). Of note, high-resolution confocal microscopy of control EB1-Gal4>UAS-mCD8::GFP revealed not only intense labelling of R2/4 layers but also spoke-like projections towards the inner ring (Fig. 4g inset and Fig. 4n), while NV45-LexA>AOP-mCD8::GFP controls revealed intense labelling of R3 layers but also projections within the outer ring (Fig. 4h inset and Fig. 4n). When compared to control patterns (Fig. 4g,h), the detected GRASP signal thus also suggests inter-layer connectivity between EB1 and NV45 (Fig. 4l-n). Importantly, the suggested connections were observed with confocal and super-resolution microscopy not only for specific layers, as visualised for example with c105>Syt::GFP, DenMark (Fig. 3a,e) and EB1>Syt::GFP, DenMark (Supplementary Fig. 4c), but also for modules as seen with poxn>Syt::GFP, DenMark (Fig. 3a-d and Fig. 4o). Taken together, these data suggest connectivity among R neurons that extends across layers and modules of the EB neuropil.
GABAergic inhibition of R neurons
We next investigated physiological properties of the R neuron circuitry. Previous studies indicate a majority of R neurons are GABAergic29,39. We therefore tested whether poxn+ R neurons are GABAergic. GABA expression was identified in 76% of c105 R1 neurons, 70% of c507 and 93% of c232 R3/4d neurons, 73% of c819 and 94% of EB1 R2/4m neurons (Supplementary Table 1). These data suggest that c232 and EB1 R neurons are inhibitory GABAergic interneurons. To test this hypothesis, we carried out functional imaging using the bioluminescent photoprotein GFP-aequorin (GA)40. We expressed GA in c232 neurons and measured photon emission related to calcium-induced bioluminescence activity in open-brain preparations (Fig. 5a,b). Direct bath application of 250□3n picrotoxin, a noncompetitive blocker of GABAA receptor chloride channels, resulted in photon emission (up to 300 photons/s) that persisted for several minutes (Fig. 5c). These data suggest that picrotoxin caused a widespread disinhibition of GABAergic c232 neurons.
To corroborate these observations, we carried out EB circuit- and layer-specific knock down of the Drosophila GABAA receptor, Receptor of dieldrin (Rdl) and combined it with expression of UAS-GA (c232>GA, Rdl-IR). To simulate input activity of varying strength, we implanted electrodes into open-brain preparations (Fig. 5d, inset) and injected current of 20V in three stimulations, followed by another three 40V stimulations, each of which lasted 5ms separated by 2min rest (Fig. 5d). When compared to c232>GA controls, 40V stimuli but not 20V stimuli resulted in increased photon emission in c232>GA, Rdl-IR (Fig. 5d; p<0.01 two-way ANOVA, n=14). These data suggest that the circuit-specific Rdl knockdown increased the electric stimulus induced-response of c232 neurons when compared to c232>GA controls. Moreover, consistent with a disinhibitory effect of c232>Rdl-IR, the observed GA activity recordings suggest c232 circuit-specific inhibition among c232 neurons.
To gain further insight into physiological properties of GABAergic R neurons, we performed whole-cell patch recordings from GFP-labelled EB1 R neurons. These revealed resting membrane potentials of −52.9 ± 4.0mV (mean ± standard deviation, SD) and input resistances of 1185 ± 407MΩ (n=20). Following bath-application of 250μM picrotoxin, all of the six recorded neurons exhibited individual spikes followed by bursting activity that lasted between 10-40s (Fig. 5e). Bursts (0.5-10Hz frequency) comprised up to 60 individual spikes, with an instantaneous spike frequency ranging from 30-100Hz, followed by strong depolarisation which suppressed further activity. To test their stimulus-salience response, we injected increasing amounts of depolarising current (up to 40pA) into EB1 R neurons, which evoked repetitive firing with successively more spikes triggered by higher currents, up to a maximum frequency of 80Hz (Fig. 5f), suggesting that spike frequency increased proportional to stimulus salience/strength. Together, these data suggest that c232 and EB1 R neurons are inhibitory GABAergic interneurons that are inhibited by GABAA receptor signalling.
Sensory-motor transformation by inhibitory R neurons
Ellipsoid body R neurons have been shown to play a role in goal-directed locomotion, spatial orientation learning and memory, as well as in sleep, attention, arousal, and decision-making13,14,16–18,20–26. All of these observations suggest that these neurons are involved in the assessment of sensory information leading to the determination of a behavioural action. To further address this, we made use of temperature-related behaviour which in Drosophila is known to elicit goal-directed locomotion41 (Fig. 6).
Temperature sensation in Drosophila is mediated by thermal receptors, with sensors for innocuous warmth located in the brain42. Four warmth-activated thermosensory neurons called Anterior Cells (AC) express Transient Receptor Potential A1 (TRPA1) essential for temperature integration and the selection of preferred temperatures; these neurons send connections to the subesophageal ganglion, to two antennal lobe glomeruli and to the superior lateral protocerebrum (SLP)43,44. The SLP is innervated by Fm-neurons of the FB45 which itself is connected with the EB (Fig. 1b). Since none of the known thermal receptors are expressed in the EB, we asked whether integrated temperature information is received and processed by R neurons.
To resolve this question, we expressed GCaMP6f (ref. 46) in EB1 R neurons. GCaMP-fluorescence of preparations exposed to temperatures ramps ranging from 18°C-35°C showed a response in 7 out of 9 brains (Fig. 5g). GCaMP-fluorescence detection revealed a saccade of small peaks starting around 23°C, indicative of activation of individual neurons, which by >30°C focused into fluorescence spiking for both EB1>GCaMP6 expressing neurons and their ring neuropil (Fig. 5g). Spiking occurred instantaneously and across the whole population, with GCaMP-fluorescence intensity 4-7-fold above baseline for ring neuropil and cell bodies, respectively, and rapidly declined over 2 minutes, followed by inactivity even at noxious 35°C (Fig. 5g). The pattern and time course of this calcium increase resembles the steadily increasing activity seen in electrophysiological recordings in response to application of picrotoxin (Fig. 5e) or to increasing amounts of depolarizing current (Fig. 5f). These results suggest that EB1 R neurons assess stimulus input of sensory information, such as temperature changes, by threshold-dependent neuronal activity.
GABAergic inhibition of R neurons mediates motor actions
The selection and gating of behavioural actions comprise their initiation, maintenance and termination as well as their assembly into activity sequences, which together generate adaptive behaviour9. The most basic selection is between action and inaction. The failure of such selectivity is illustrated by pathologies of human motor control and motivational disorders, such as the indecisiveness of motor actions seen in Parkinsonism, and the diminution of voluntary motor actions seen in abulia47. These behavioural symptoms exemplify disease-specific alterations in the initiation, maintenance or termination of actions and their goal-directed assembly into action sequences. Analogous pathologies occur in Drosophila. These are characterized by anatomical alterations or defective neuromodulation of CX circuits11,18. Given its proposed role in action selection, EB dysfunctions are likewise expected to affect, for example, the frequency in which flies initiate and maintain behavioural actions, such as bouts of walking.
We measured such behavioural alterations using open-field assays48, employing video-assisted motion tracking to record freely moving Drosophila (Fig. 6a). Flies were kept in 35mm diameter arenas and after 1hr of adaptation were recorded for 60 minutes, with a short pulse of mechanical stimulation applied after 30 minutes (see supplementary methods). Based on these recordings, we generated raster plots and trajectories for each individual fly and extracted parameters detailing their behaviour including action initiation and maintenance and their response to sensory stimulation (Fig. 6b-d), each of which are integral to the action selection process49. To examine R neuron functions, we utilized c105, c232 and EB1-Gal4 drivers to express specific UAS-transgenes targeted to R1, R3 and R2/4 neurons, respectively (Supplementary Fig. 1). We first expressed UAS-Rdl-IR to knock down the Drosophila GABAA receptor, Rdl, which identified alterations for R neuron subtypes (Fig. 7 and Supplementary Table 2). Rdl-IR expression resulted in increased activity and walking distance, which for R1 and R3 was accompanied with less frequent initiation of actions but their increased bout length. EB1>Rdl-IR flies were also faster and paused less. Mechanical stimulations did not reveal impaired behavioural responses but an increase in the amplitude was observed for R1 and R3 that could be attributed to decreased pre-stimuli speed (Supplementary Fig. 6). These data suggest that GABAergic inhibition of R neurons coordinates voluntary motor actions and their assembly into action sequences.
R neurons mediate the selection and maintenance of behavioural actions
A Action selection circuitry is readily engaged by external stimuli that drive an animal’s behaviour49. Previous studies identified a key role of the insect EB in a variety of behavioural activities relating to sensory stimuli16–18,20–22,26,27, suggesting that stimulus-induced action selection requires EB function. We hypothesized that modifying EB R neuron circuitry should result in an acute impact on behaviours that relate to salient external stimuli. To test this we focused on temperature-related behaviour in the knowledge that Drosophila actively avoid temperatures above 29°C, an aversive temperature that readily elicits goal-directed locomotion41 (Fig. 6b and Supplementary Fig. 7a). To utilise this activity in our analysis of the role of R neurons, we expressed either Tetanus-Toxin-Light-Chain (TNT), encoding an inhibitor of synaptic transmission50, or NaChBac encoding a bacterial depolarisation-activated sodium channel51 to increase neuronal excitability (Supplementary Fig. 8). In order to rule out any potential developmental effects, we also utilized thermogenetic activation using the cation channel TRPA1 (ref. 43) to activate neurons or the dynamin mutant ShibireTS to block vesicle endocytosis52.
We first determined thermal preference behaviour for c105>TNT, c232>TNT and EB1>TNT flies, which revealed peaks of around 25°C (Supplementary Fig. 7a) as preferred temperature41. In addition, we ruled out that R neurons regulate fixed action patterns such as startle-induced negative geotaxis (SING) which triggers an innate escape response53, or that R neuron>TNT flies were impaired in leg coordination, or subject to motor abnormalities, or to muscle fatigue. Except for a single meta-thoracic neuron cluster in c105, none of the lines show Gal4 expression in muscles or motor neurons (Supplementary Fig. 1), and SING analysis revealed no significant differences (Supplementary Fig. 7b), which would be the case if motor neurons or leg coordination were affected53,54.
Next we analysed open-field behaviour of c105, c232 or EB1-Gal4 flies misexpressing UAS-TNT or UAS-NaChBac. Recordings at 31°C revealed behavioural alterations of c105>TNT flies that initiated more, albeit slower, actions of shorter duration (bout length) which were intermitted by longer pauses (Interbout Interval, IBI length); these alterations, however, were unrelated to their arousal state which remained unchanged (Fig. 8 and Supplementary Fig. 9). In contrast to R1-related behavioural alterations, we did not observe any significant changes relating to R2-R4 neurons. This R1 subtype-specific modulation of behaviour was largely recapitulated by misexpression of ShibireTS (Supplementary Fig. 10). No significant alterations in the behaviour of >NaChBac expressing flies were observed at 31°C (Supplementary Fig. 11). In contrast, TRPA1-mediated acute activation of EB-R neurons at 31°C identified behavioural alterations almost exclusively restricted to R3 neurons (Fig. 9). Thus, c232>TRPA1 flies were characterized by significantly decreased basal activity that contrasted with more actions initiated which were not maintained very long and intermitted by extended pauses; these behavioural alterations were unrelated to changes in the arousal state of c232>TRPA1 flies, which remained unaltered (Supplementary Fig. 6). Together these data establish specific roles of R neuron activity in the selection and maintenance, and thus gating of behavioural actions.
A computational model of EB action selection circuitry
The sum of our data supports the proposition that R neurons form GABAergic inhibition circuitry among inhibitory sister cells that underlies the selection and maintenance of behavioural actions. To further explore the computational features of such circuitry within the CX, we developed a neural model relying on leaky integrator units to simulate the mean field activity of neural populations55 (see supplementary methods). We constrained the model architecture to replicate known connectivity of the CX, focussing on afferent and efferent EB projections (Fig. 10a). The EB receives and integrates heterogeneous sensory information10. For instance, stimulus-induced visual activity in the PB’s and the FB’s columns convey information about a visual feature’s locality in relation to the head and compound eyes22,26. In four simulations used as case studies (Fig. 10), noisy information about sensory cues is converted into the activation of PB modules. The features characterizing these stimuli, such as signal intensity and localization in space, are then propagated in parallel to EB modules, activating them dependent on the strength of the signal, as has been shown for visual cues26, and weighed by the relative connections. Our present study suggests GABAergic R3 neuron projections are connected to GABAergic R2/4 neurons. Therefore we simulated three EB layers (R1, R2/4 and R3), each divided into 8 modules per hemisphere (as for PB, FB and LAL)45,56–58 that are characterized by undifferentiated lateral inhibitions across layers and within each layer (Fig. 10a).
We tested a range of parameters to explore the robustness of emerging functions (Supplementary Fig. 12). In particular, we focused on whether the known neurophysiology of the EB is consistent with the proposed functional role of a sensory-driven selector. We opted for fixed lateral inhibition between R neurons, while exploring the consequences of several different configurations of excitatory inputs from the PB and FB to the EB, and of inhibitory outputs from EB to LAL (Supplementary Fig. 12). Our simulations show that parallel input connectivity, coupled with internal lateral inhibitions are sufficient to realize sensory integration, to detect the highest saliency, to cancel competing signals and noise, and to adapt to changes in the environment such as novel sensory input (Fig. 10b-e and Supplementary Fig. 12). The first two case studies show dynamics of several noisy competing stimuli encoded in the activation of PB modules (Fig. 10b). These stimuli are propagated to EB modules where lateral inhibitions mediate competition among all active modules. The resulting computation among R neurons mediates salience detection by winner-takes-all functionality and the selection of only one active module, cancelling both noise and competing signals59 and yet allowing adaptation to changes in the topography and configuration of new salient sensory cues (Fig. 10c and Supplementary Fig. 12). Our model also predicts that during persistent stimulus input, such as elevated temperature, the EB circuitry enables continued salience detection and the competitive selection of one active module (Fig. 10d). At the end of a winner-takes-all competition, and depending on the synaptic strength of R neuron connections, our simulations show that lateral inhibition can maintain a selected activity even after input changes (Fig. 10e).
DISCUSSION
Our results identify a pair of bilateral symmetric, engrailed-expressing embryonic neuroblasts, ppd5, that generate all four subtypes of ellipsoid body R neurons (R1-R4) in the central complex of the adult brain in Drosophila. Lineage-related R neurons form GABAergic microcircuitry that when dysfunctional impair the selection and maintenance of behavioural actions. The present results identify the organization of EB microcircuitry as a salience detector, using competitive inhibition to amplify different activity levels between modules thereby implementing a winner-takes-all mechanism for selecting action output.
A key feature of action selection is the evaluation and transformation of sensory representations into motor representations49. Consistent with previous findings16–18,20–22,26, our results demonstrate that R neurons assess sensory information and mediate the coordination of voluntary motor actions and their assembly into action sequences. Our computational model accounts for the observed results of genetic manipulations and the ensuing behavioural observations, suggesting the R neuron circuitry functions in sensory-motor action selection. The predictions of our model are supported by functional imaging studies showing that specific EB module activity correlates with the azimuth of a salient visual cue26. The observed activity “bump” was largely confined to one module but its position within the EB ring did not match a retinotopic position of the visual cue. Rather it represented the fly’s orientation relative to the visual landmark, with module-restricted activity switching along the EB toroid according to changes in the fly’s angular orientation aligned to the sensory cue26.
Layers and modules in the ellipsoid body have been implicated in a variety of behavioural manifestations that identify key roles of the EB in response not only to visual cues for spatial orientation15,19,22,24,26 and memory formation16,17,21, but also in response to olfactory cues for ethanol-induced locomotion20 and tactile cues for arousal18 and turning behaviour23,25 as well as in relation to sleep drive27. Our findings further suggest that R neurons mediate weighing and selection of sensory representations independent of their modality. In support of this, the present functional imaging and electrophysiological data identify threshold-dependent stimulus responses of R3 and R2/4 neurons, the spiking initiation of which was dependent on stimulus salience, and whose spiking frequency was proportional to stimulus strength (Fig. 5). Moreover, the model simulations predict that dependent on synaptic strength, inhibitory R neuron activity can maintain a selected activity even after inputs change. In accordance with this, an EB module-specific activity “bump” has been shown to persist as after a salient visual cue disappeared26. Thus, the selection and maintenance of an active module can be regarded as convergence of an internal representation of a behaviourally relevant sensory cue and the agent’s reference to it, such as angular orientation26 and movement intent25,60. The availability of an EB module-specific activity for a prolonged time can guide successive behaviours and in this sense exhibits essential features of short-term memory61, which in the case of visual short-term memory has been shown to require R neuron activity16.
The R neuron circuitry is embedded in a connectivity network with the PB, FB and the LAL45,56–58, the latter a premotor command centre projecting to descending neurons that innervate central pattern generators executing motor actions62. As a node of convergence, situated several synapses downstream of sensory neurons and several synapses upstream of motor ganglia, the position and connectivity of the EB-LAL interface implicates it in the translation of sensory representations into motor representations. Experiments in locusts63 and silk moths64 identified that LAL modules preserve information about values and saliency associated with selected sensory representations. Our simulations confirm these findings and predict that the EB exerts selection on premotor commands in the LAL by module-specific columnar projections10,64 that retain stimulus values by a one-to-one correspondence. As a consequence, inhibitory projections from the EB towards the LAL can reverberate the selection process and perform a surround gating of the LAL. As a result, only one motor command is selected at a time while the release of multiple motor actions are avoided at the same time (Fig. 10c-e and Supplementary Fig. 12).
In line with these predictions, our results show that artificial manipulations of R neurons are themselves sufficient to initiate, increase or abrogate motor actions and their organization into action sequences (Figs 7–9, Supplementary Figs 6 and 9-11). Disinhibited behavioural activity was observed with Rdl-IR expression for each of the EB’s concentric layers, which was inversely correlated with action initiation as flies less frequently initiated behavioural activity. A similar inverse correlation was also observed with specific overexpression of TNT, Shibire and TRPA1 in R neurons. Although all EB layers respond to Rdl-IR, the resulting behavioural manifestations reveal subtype-specific differences that are also seen for TNT, Shibire and TRPA1. TRPA1-mediated behavioural changes suggest a specific role of R3 and R2/4 neurons in the initiation and termination of motor action, while c105-specific TNT and Shibire expression caused activity changes indicative of an essential role of R1 neurons as the EB output10.
A comparable organization into partially segregated functional units also characterizes inhibitory GABAergic striato-pallidal projections that constitute the direct and indirect pathways of the vertebrate BG10. In mice, optogenetic manipulations have shown that direct pathway stimulation facilitates motor actions, whereas indirect pathway stimulation decreases motor activity2, both of which are cooperatively active during voluntary movement3,4. Furthermore, functional imaging of direct and indirect pathways in mice identified their differential yet concomitant activities underlie the parsing and concatenation of action sequences65–67. These observations lend support to the prevailing model of BG-mediated action selection; namely, by disinhibition of a selected motor program and inhibition of competing actions5,7–10,68.
In Drosophila, our genetic manipulations and neural computations identify facilitation, inhibition and disinhibition of motor activity as differential R neuron functions underlying the initiation, maintenance and termination of actions and their organisation into action sequences. These findings suggest that R neuron circuitry exhibits homologies to direct and indirect pathway activities including their disease-related dysfunction1,10. Focal lesions and dysfunction of direct and indirect pathways are associated with movement disorders, such as Parkinson’s and Huntington’s disease and dystonia5,47, as well as neuropsychiatric disorders including impaired memory formation, attention deficits, affective disorders and sleep disturbances6–8,69. Comparable behavioural deficits are seen in flies with altered R neuron function16–18,20,26, suggesting homologous pathological manifestations. It stands to reason that evolutionarily conserved mechanisms underlie action selection in health and disease10, including the coordination of higher motor control and motivation.
METHODS SUMMARY
For extended methods see Supplementary Information.
Drosophila stocks and genetics
Flies were maintained on standard cornmeal medium at 25°C in a 12hr:12hr light/dark cycle. For lineage tracing, w1118 (control) and Gal4 strains were crossed to UAS-mCD8::GFP, tub-FRT-CD2-FRT-Gal4, UAS-FLP/CyO GMR Dfd YFP. For clonal analysis, lines y, w, FRT19A; +; + and w, FRT19A, tubP-Gal80, hs-FLP; UAS-nucLacZ, UAS-mCD8::GFP / CyO; tubP-Gal4 / TM6B Hu were used. For GRASP analysis, R2/4-specific EB1-Gal4 and R3-specific NV45-lexA::VP16 driver were combined with UAS-CD4::spGFP1-10 and LexAop-CD4::spGFP11, For functional imaging, UAS-GFP-aequorin or 20XUAS-IVS-GCaMP6 were combined with R3/4 c232-Gal4 or R2/4 EB1-Gal4. For behavioural experiments, UAS-Rdl-IR, UAS-TNT, UAS-NaChBac, UAS-ShibireTS, or UAS-TRPA1 were crossed to R neuron-specific Gal4 lines, and controls generated by crossing driver and responder lines to w1118. Further details are provided in the extended methods section of Supplementary Information.
Super-resolution and electron microscopy
Structured Illumination Microscope and Multi-photon microscope (Nikon) were used for super resolution microscopy; 3D reconstruction and video assembly was carried out using Nikon software. For electron microscopy, 80 nm sections were cut on a Leica Ultracut S Ultramicrotome and visualized in a Phillips CM12 transmission electron microscope.
Functional imaging and electrophysiology
Bioluminescence imaging was performed in open brain preparations of tethered flies. GCaMP6 imaging and electrophysiology were carried out on isolated brains maintained in saline solution. Whole cell current clamp recordings were performed using glass electrodes with 8-12 MΩ resistance filled with intracellular solution and an Axon MultiClamp 700B amplifier, digitised with an Axon DigiData 1440A (sampling rate: 20kHz; filter: Bessel 10kHz) and recorded using pClamp 10 (Molecular Devices, CA, USA).
Behavioural analysis
Single, 3-5 day old flies were placed in an open-field arena, and their behaviour was recorded with a Logitech c920 camera at 10 frames per second. 36 flies, 12 of each genotype, were recorded simultaneously for 1 hour. After 30 minutes, mechanical stimulation was delivered through shaft-less motors (Precision Microdrives) controlled by custom-made DART software70. These behavioural experiments were carried out in a temperature-controlled chamber at either 25°C or 31°C. Thermal preference behaviour was carried out using a linear temperature gradient from 18°C to 31°C.
Neural network modelling
Ellipsoid body R neuron modelling was performed in MATLAB using custom scripts. Network dynamics were captured with a firing rate model based on a leaky integrator defined by a continuous-time differential equation.