Presynaptic inhibition of dopamine neurons controls optimistic bias

Regulation of reward signaling in the brain is critical for appropriate judgement of the environment and self. In Drosophila, the protocerebral anterior medial (PAM) cluster dopamine neurons mediate reward signals. Here, we show that localized inhibitory input to the presynaptic terminals of the PAM neurons titrates olfactory reward memory and controls memory specificity. The inhibitory regulation was mediated by metabotropic gamma-aminobutyric acid (GABA) receptors clustered in presynaptic microdomain of the PAM boutons. Cell type-specific silencing the GABA receptors enhanced memory by augmenting internal reward signals. Strikingly, the disruption of GABA signaling reduced memory specificity to the rewarded odor by changing local odor representations in the presynaptic terminals of the PAM neurons. The inhibitory microcircuit of the dopamine neurons is thus crucial for both reward values and memory specificity. Maladaptive presynaptic regulation causes optimistic cognitive bias.


Introduction
Regulation of reward signaling in the brain is critical for maximizing positive outcomes and for avoiding futile costs of the behaviors at the same time. Across animal phyla, dopamine neurons are primarily involved in reward processing (Brembs et al., 2002;Tobler et al., 2005;Liu et al., 2012;Ichinose et al., 2017). In the fruit fly Drosophila melanogaster, a subset of dopamine neurons in the protocerebral anterior medial (PAM) cluster mediates the reinforcement property of sugar reward (Burke et al., 2012;Liu et al., 2012). In olfactory learning, dopamine input to the mushroom body (MB) causes changes in preference of a simultaneously presented odor by modulating the output of odor-representing MB intrinsic neurons, Kenyon cells (KCs) (Séjourné et al., 2011;Boto et al., 2014;Cohn et al., 2015;Owald et al., 2015;Louis et al., 2018;Hige et al., 2015;Bilz et al., 2020). Such associative presentations of odor and electric shocks were reported to change the activity of MB-projecting dopamine neurons (Riemensperger et al., 2005). Recent studies (Hattori et al., 2017;Cervantes-Sandoval et al., 2017;Takemura et al., 2017) suggest that axon terminals of the dopamine neurons locally integrate olfactory inputs to function as multiple independent units, though such subcellular reward processing has yet to be examined.

Results and discussion
To understand neuronal mechanisms for the regulation of reward processing, we here focused on gamma-aminobutyric acid (GABA) signaling in the PAM neurons. Six GABA receptor genes are identified in the fly genome. We silenced the expression of each receptor gene in the PAM cluster neurons by targeting transgenic RNAi (Ni et al., 2011) and tested their appetitive olfactory memory ( Figure 1A). We found increased memory performance by downregulation of a metabotropic GABA receptor, GABA-B-R3 ( Figure 1A and   We next examined endogenous GABA-B-R3 expression in the adult brain using the intronic CRISPR-Mediated Integration Cassette (CRIMIC) insertion of T2A-GAL4 (Lee et al., 2018). The T2A self-cleaving peptide between the target protein and GAL4 allows bi-cistronic translation by a ribosome skipping mechanism (Diao and White, 2012). GABA-B-R3 was expressed broadly in the brain, including the majority of the PAM cluster neurons ( Figure 1C-D), whereas the expression was weak in KCs ( Figure 1D and Figure 1-figure supplement 1C). There was no notable morphological alteration in the brain of knock-down flies (data not shown). Consistently, adult stage-specific GABA-B-R3 silencing in the PAM neurons using Tub-GAL80 ts (McGuire et al., 2003) similarly enhanced appetitive memory performance ( Figure 1B). Without transgene induction, their appetitive memory was indistinguishable from the controls ( Figure 1B).
Increased learning speed and/or performance plateau may underlie the enhanced appetitive memory in the GABA-B-R3 knock-down flies. We attempted to distinguish these possibilities by characterizing their memory acquisition ( Figure 1E and Figure 1-figure supplement 1D). The performance of R58E02-GAL4/UAS-GABA-B-R3-RNAi flies reached a significantly higher asymptote than control genotypes without changing the acquisition speed ( Figure 1E and Figure 1-figure supplement 1E-F). In a learning theory, the magnitude of reinforcement is the determinant for the plateau of the acquisition curve (Rescorla, 1972), suggesting that sugar reward was perceived more strongly with enhanced dopaminergic activity in the GABA-B-R3 knock-down flies. Live calcium imaging at terminal branches of the reward-related PAM neurons (i.e., PAM-g5 and -a1) revealed the augmented sugar responses upon downregulating GABA-B-R3 ( Figure 1F-H). We thus conclude that GABA-B-R3 signaling is required for negative regulation of the sugar reward.
We visualized the localization of GABA-B-R3 proteins using a GFP-tagged reporter (Sarov et al., 2016). GABA-B-R3 proteins were heavily localized to the presynaptic terminals of the PAM neurons ( . We thus hypothesized that presynaptic inhibition of dopamine neurons within the MB controls the gain of reward signals. A single pair of the GABAergic anterior paired lateral (APL) neurons was reported to massively innervate the entire MB and to be involved in olfactory learning (Liu and Davis, 2009). Differential labeling of the PAM and APL neurons revealed that their ramifications abut on each other ( Figure 2C-D). Consistently, we found enhanced reward memory in knock-down flies for glutamic acid decarboxylase 1 (Gad1) and vesicular GABA transporter (VGAT) in the APL neuron ( Figure 2E). This result not only underscores the importance of GABA metabolism in the APL neurons, but suggests the role of the inhibitory microcircuit in the MB for the gain control of the reward value. We therefore examined the local inhibition hypothesis by comparing sugar responses in the dendrites and presynaptic terminals of the PAM neurons ( Figure 2F and Figure 2-figure supplement 1B-C). The enhanced calcium activity upon GABA-B-R3 knock-down was much more pronounced in the presynaptic terminals ( Figure 2G-H). Therefore, GABAergic signals from the APL neurons negatively control the reward gain at the output site of the PAM neurons through GABA-B-R3 signaling in the MB.
To quantify the local activity regulation in the PAM terminals, we measured calcium influx at active zones using the ratiometric calcium sensor Brp::GCaMP6s::mCherry (Kiragasi et al., 2017). This sensor is composed of GCaMP fused to calcium insensitive mCherry and targeted to active zones using the short fragment of Brp, enabling the measurement of local calcium influx at active zones (Kiragasi et al., 2017). Immunolabelling confirmed the localization of the sensor       . Furthermore, we found that Brp:: GCaMP6s::mCherry signals had temporal fluctuations, which was amplified by silencing GABA-B-R3 ( Figure 3B and Figure 3C). GABA-B-R3 in the PAM terminals may thus stabilize the basal presynaptic activity. This suggests that GABA inhibition contributes to the robustness of activity against local perturbations.
System robustness is often related to refined regulation of activity patterns by which information is efficiently coded (Hesse and Gross, 2014). As active-zone calcium in the PAM terminals is likely to reflect local input in the MB (see Figure 2G and H), we hypothesized that GABA-B-R3 controls the spatial representation of odor information in the PAM terminals. In control flies, odor presentations barely changed the distribution of active-zone calcium with a marginal increase of the overall signal intensity ( Figure  . These stimulus-specific presynaptic responses are likely to reflect differential input sites of odor and sugar signals, that is, pre-and post-synaptic sites of the PAM neurons, respectively. Strikingly, GABA-B-R3 knock-down flies responded to an olfactory stimulation much more strongly ( Figure 3D-E), as if in response to a global sugar stimulation (Figure 3-figure supplement 1D). These results suggest that the local control of PAM presynaptic activity by GABA-B-R3 regulates the odor responses, possibly by refining its spatial representation.
To further characterize how GABA-B-R3 regulates the structure of presynaptic activity in PAM terminals, we examined the patterns of local peaks of Brp::GCaMP6s::mCherry signals by introducing a spatial measure. We found that the average peak area (i.e., cluster size) became larger upon knocking down GABA-B-R3 both before and during the odor stimulation ( Figure 3F-G). Note that we did not observe a clear correlation between the cluster size and intensity of pixels in the cluster (data not shown), suggesting that the cluster size measure provides information orthogonal to the intensity value of calcium signals. We also quantified the size of high-intensity pixels using another measure, which we call peak size (see Materials and methods; Figure 3-figure supplement 1F). We further found that intense Brp::GCaMP6s::mCherry signals are more spatially clustered in the knock- individual. The abscissa and ordinate represent the mean and SD of the signal distribution, respectively. The clear inverse correlation between the mean and the variance in wild-type terminals (r = À0.82, p<10 À6 ) may represent individually defined unique set points of activity levels. This structured individual difference is disrupted by GABA-B-R3 RNAi (r = 0.53, p=0.03). The start and end of an arrow represent the probability distribution of activezone calcium before and during the odor stimulation, respectively. Scale bar, 0.5 in Log 10 (GCaMP6s/mCherry). (F) Local peaks in the PAM terminals are color-coded for their cluster sizes (right). Schematic example of the algorithm finding the cluster structure applied to a one-dimensional system (Left). Spatial distribution of calcium intensity (Left upper). Finding steepest paths from each pixel to local peaks by computing the gradient (i.e., the difference of the intensity values between neighboring pixels)(Left middle). Note that in two-dimensional systems we analyzed, each pixel had four neighboring pixels. Clustering based on the paths to local peaks (Left lower). Pixels having the same destination (i.e., local peak) are clustered together. (G) The average area per peak is significantly larger in the GABA-B-R3 knock-down flies during baseline activity (*p<0.05, Mann-Whitney test, N = 22, 17) and odor stimulation (*p<0.05, t-test, N = 15, 11). The online version of this article includes the following source data and figure supplement(s) for figure 3: Source data 1. Temporal variance of spatially averaged intensities in control and GABA-B-R3 knock-down flies. Source data 2. Probability distribution of active-zone calcium intensity before and during odor stimulation. Source data 3. Average cluster size in control and GABA-B-R3 knock-down flies.   down PAM terminals, further corroborating the fine spatial regulation of calcium signals by GABA-B-R3. Since KC and PAM terminals form mutual synapses in the MB lobe (Takemura et al., 2017;Cervantes-Sandoval et al., 2017), the overall disinhibition in the knock-down terminals may impair the selective delivery of dopaminergic reward signals to odor-activated KCs.
To test if presynaptic GABA-B-R3 signaling controls odor representations, we examined the specificity of memory using odor generalization. Conditioned odor approach of wild-type flies drops by increasing the blend ratio of a contaminant odor to the trained odor Chen et al., 2017). Strikingly, the downregulation of GABA-B-R3 in the PAM neurons resulted in a broader generalization profile, indicating that reward memory in the knock-down flies is less specific to the learned odor ( Figure 4A-B). Moreover, we found similarly broadened generalization profiles by downregulating GAD1 expression in the APL neurons ( Figure 4C-D). We altogether conclude that local GABAergic inhibition of the PAM neurons regulates the intensity and specificity of reward memory ( Figure 4E).
Our results indicate that presynaptic modulation of the PAM neurons is a critical component for determining the magnitude of dopaminergic reward signals. Notably, abolition of the local GABAergic input to the PAM terminals not only enhanced the internal reward intensity but compromised memory specificity (Figures 1 and 4). These behavioral alterations can be explained by a dual physiological role of GABA-B-R3, that is, the gain control and the spatial segmentation of dopaminergic reward signals in the PAM terminals ( Figure 4E). As the behavioral traits caused by the downregulation of GABA-B-R3 are characteristic in optimism (Carver et al., 2010;Solvi et al., 2016), presynaptic control of reward signals may underlie such a cognitive bias. It would be fruitful to examine if a similar subcellular modulation of punishment-mediating neurons conversely leads to the pessimistic bias (Sharot et al., 2009;Bateson et al., 2011;Sharot et al., 2012;Kregiel et al., 2016;Solvi et al., 2016;Zidar et al., 2018).

Behavioral assays
The conditioning and testing protocols were as described previously Yamagata et al., 2016). Briefly, for a normal sugar learning experiment ( Figures 1A-B and 2E and Figure 1-figure supplement 1A-B), a group of approximately 50 flies in a training tube alternately received octan-3-ol (3OCT; Merck) and 4-methylcyclohexanol (4MCH; Sigma-Aldrich) for 1 min in a constant air stream with or without dried 2 M sucrose paper. For varied training duration protocol ( Figure 1E and Figure 1-figure supplement 1D), flies received two odors and dried sugar alternately for defined duration (10-120 s) with an interval of 1 min between two odors. For odor generalization protocol ( Figure 4A-D), flies were trained with an odor, that is, they alternately received 4MCH and paraffin oil (Sigma-Aldrich), or 3OCT and paraffin oil, for 1 min in a constant air stream with or without dried 2 M sucrose paper. Then the conditioned response of the trained flies was measured. For the normal protocol ( Figures 1A-B, E and 2E, Figure 1-figure supplement 1A-B, D), flies were given a choice between CS+ and CS-for 2 min in a T maze. For generalization protocol ( Figure 4A-D), flies were given a choice between a 'trained' odor with a respective mixture ratio of a contaminant odor (2-methylcyclohexanol [2MCH]; Sigma-Aldrich or 1-octen-3-ol [1OCT]; Sigma-Aldrich) and the solvent for 2 min in a T maze. All odors were diluted to 10% in the paraffin oil and placed in a cup with a diameter of 3 mm (OCT) or 5 mm (MCH). The memories were tested immediately after training unless otherwise stated. A learning index was then calculated by taking the mean preference of the two reciprocally trained groups. A half of the trained groups received reinforcement together with the first presented odor, and the other half with the second odor to cancel the effect of the order of reinforcement.

Brain dissection, immunohistochemistry, and sample mounting
Dissection of fly brains was performed as previously described (Kondo et al., 2020) with minor modifications. Brains of female ( Figures 1C-D

Fly preparation and in vivo calcium imaging
Flies were treated as described in Hiroi et al., 2013;Shiozaki and Kazama, 2017 with some modifications. The fly was briefly (<1 min) anesthetized on ice and placed in a custom-made holding device on a Peltier plate (CP-085, Scinics) held at 4˚C. The head capsule was fixed to the dish by UV curing optical adhesives (NOA68, Thorlabs). The proboscis was glued onto the capsule to eliminate brain movement. Forelegs interfering sugar feeding during recordings were removed. A small window on the top of the head capsule was opened using sharp forceps in Drosophila saline (103 mM  A laser scanning confocal microscope (A1R, Nikon) equipped with a 30Â/1.1 water immersion objective (Apo LWD 25Â, Nikon) and a Piezo nanopositioner (Nano-F450, MCL Inc) combined with a Nano-Driveone controller (MCL Inc) was used for live imaging. GCaMP6s and mCD8::RFP or mCherry were sequentially excited at 488 and 561 nm, respectively. The emission light was collected onto GaAsP detectors using dichroic mirrors and emission filters (BP500-550 and BP570-620). Transverse sections of the MB lobes and the superior medial protocerebrum were scanned at a resolution of 0.5 mm/pixel (512 Â 128 pixels) at 333 ms/frame ( Figure 1F-H 1) averages using the resonant scanning mode. The pinhole was set to 2.5 AU (561 nm). For 3D imaging, two z sections (ca. 100 mm interval) were scanned using a piezoelectric motor. To record calcium responses to sugar and odors, images were acquired for 20 or 30 s and saved for later image processing. For sugar stimulation, a droplet of 500 mM sucrose deposited on a tip of Microloader pipette tip (Eppendorf) was presented to the proboscis for 3 s using a micromanipulator (UN-3C, Narishige). Sugar stimulation to flies was monitored with a USB camera (Grass-hopper3, FLIR) mounted with a zoom lens (MACRO ZOOM 0.3ÂÀ1 Â 1:4.5, Computar) and captured by FlyCapture2 (FLIR). Odor stimulation has been made manually using a 50 ml syringe containing a piece of filter paper (1 Â 2 cm 2 ) soaked with pure or 10 times diluted 4MCH and 3OCT. For each stimulation,~15 ml odor contained air was delivered to a fly in 3 s through a 4 mm silicon tube placed ca. 10 mm away from the fly head.

Data analyses
All the acquired images were first processed with Fiji. An object in each recording was stabilized by TurboReg plugin (Thévenaz et al., 1998) using mCD8::RFP or mCherry signal. ROIs to be involved in later calculations were defined by mCD8::RFP signal in the left or right hemisphere. In Figures 1F-H and 2F-H and Figure 2-figure supplement 1C, GCaMP6s signal was used as a fluorescent F value. The DF/F 0 was calculated as: where F t and F 0 denote fluorescent values at time frame t and baseline (i.e.,~7 frames before stimulation), respectively. To highlight MB compartments that responded to stimulations ( Figure 1F), a time series projection of the DF/F 0 during stimulation (for 3 s) was thresholded and superimposed on a projection image of mCD8::RFP signal at respective frames. After XY registration, the Brp::GCaMP6s signal was divided by mCherry (GCaMP6s/mCherry) to normalize the calcium signal by Bruchpilot abundance. An ROI for the a1 compartment of the MB was defined by mCherry signal. Pixels devoid of an mCherry fluorescence value were censored. The image stacks were then imported to Matlab (MathWorks) and log-transformed.
To evaluate the spatial pattern of the calcium intensity in the PAM-a1 terminals, we computed the size of the area of each peak ( Figure 3F) in the following manner: (i) For each pixel in an image, the pixel with the largest intensity value among the four neighboring pixels was identified. If the given pixel had a larger intensity value than those of the neighboring pixels, we recorded the pixel as a local peak. We ignored pixels that had the background intensity value. This procedure yielded the steepest path to a local peak from each pixel. (ii) We clustered the pixels based on the local peak connected by the paths. Note that this simple algorithm is a discrete version of the gradient ascent algorithm and used for clustering with height information elsewhere (Ezaki et al., 2017). (iii) We counted the number of pixels belonging to the same cluster and computed the average cluster size for each fly based on nine frames before stimulation (or three frames during the stimulation). Because the average cluster size may be affected by the size and shape of the PAM-a1 terminals in the recorded images, we normalized it by a null model. The average cluster size for the null model was obtained as follows. (iv) For each image, we shuffled the intensity values across the pixels. (v) We computed the average cluster size for the shuffled image by performing (i)-(iii). (vi) We repeated (iv) and (v) for 1000 runs. (vii) Finally, we calculated the average cluster size for the null model by averaging the results over the 1000 runs and nine frames before stimulation (or three frames during the stimulation). This value was used for normalization in Figure 3F.
In addition to the cluster size analysis above, we used another measure (i.e., peak size) to quantify the spatial structure of the calcium intensity. The peak size was computed for each image as follows. First, we collected the pixels that had an intensity value larger than 95 percentile of the entire pixels in each image. Then, we identified the clusters of these selected pixels by checking if multiple pixels (peaks) were adjacent to each other at the top, bottom, left, or right. Finally, we counted the number of pixels in each cluster and computed the average. Similarly to the cluster size measure above, we normalized this value by using a null model. We computed the peak size for 1000 randomized images that were obtained by shuffling the intensity values in the pixels in the original image. The results were averaged over the 1000 null data, which we used for the normalization.

Statistics
Statistics were performed by Eclipse (Eclipse foundation) and Prism5 (Graphpad). For the data points that did not violate the assumption of normality and homogeneity of variance (D'Agostino and Brown-Forsythe test), parametric statistics were applied. The data points that were significantly different from the normal distribution were analyzed with nonparametric statistics. The significance level of statistical tests was set to 0.05. For details, see Supplementary file 1.
To estimate the acquisition curve dynamics ( Figure 1E), hyperbola curve fitting was applied: where A and B are constants, t is the training duration, and LI is learning index. 'A' denotes the theoretical maximum value of LI (i.e., plateau) and 'B' the training duration required to reach the half of the maximum (i.e., acquisition speed). To test the statistical significance of observed differences in A and B (DA obs. and DB obs. ) between genotypes, we performed permutation tests (Knijnenburg et al., 2009); we randomly shuffled the experimental dataset by reassigning the group labels and fitted a hyperbola function to the data to calculate the differences in A and B between groups (DA perm. and DB perm. ). The procedure was repeated over 2000 runs to generate the null distributions of DA perm. and DB perm. for testing the statistical significance of DA obs. and DB obs . . Transparent reporting form