Neural Correlates of Odor Learning in the Presynaptic Microglomerular Circuitry in the Honeybee Mushroom Body Calyx

Abstract The mushroom body (MB) in insects is known as a major center for associative learning and memory, although exact locations for the correlating memory traces remain to be elucidated. Here, we asked whether presynaptic boutons of olfactory projection neurons (PNs) in the main input site of the MB undergo neuronal plasticity during classical odor-reward conditioning and correlate with the conditioned behavior. We simultaneously measured Ca2+ responses in the boutons and conditioned behavioral responses to learned odors in honeybees. We found that the absolute amount of the neural change for the rewarded but not for the unrewarded odor was correlated with the behavioral learning rate across individuals. The temporal profile of the induced changes matched with odor response dynamics of the MB-associated inhibitory neurons, suggestive of activity modulation of boutons by this neural class. We hypothesize the circuit-specific neural plasticity relates to the learned value of the stimulus and underlies the conditioned behavior of the bees.


Introduction
Mechanistic understanding of the neural processes of learning and memory requires the identification of the cellular compositions where learning-related plasticity takes place. In searching for such "memory traces", the classical olfactory conditioning whereby an animal learns the association between an odor and a reward or punishment has been used widely (Wilson and Stevenson, 2003;Davis, 2004). Thanks to its robust behavioral performance and relatively simple brain circuitry, the honeybee has been studied intensively as a model organism for this type of olfactory learning (Menzel, 2001;Giurfa and Sandoz, 2012). To date, a rich body of evidence supports the interpretation that the sequential and parallel activity of the two olfactory centers, the antennal lobe (AL) and the mushroom body (MB), play a key role in the formation and recall of some forms of olfactory memory in the brain of bees (Hammer and Menzel, 1998;Menzel, 2012) and flies (Heisenberg, 2003;Yu et al., 2004;Scheunemann et al., 2012;Scholz-Kornehl and Schwärzel, 2016). The exact localization of learning-related plasticity in that circuitry, however, remains elusive.
In this study, we investigated the network activity of the terminal boutons of the olfactory projection neurons (PNs) in the MB calyx on learning. The MB calyx is a main input site and is formed by the dendritic arbors of the intrinsic Kenyon cells (KCs; Mobbs, 1982). Parallel axonal fibers of these KCs form lobe structures, along which each cell forms numbers of en passant output synapses. The calyx can be anatomically separated into three subdomains. One such domain is the lip region, the main olfactory input site, containing numerous synaptic inputs from the PN boutons to the dendritic arbors of the KCs (Ganeshina and Menzel, 2001). The lip is also innervated by the octopaminergic neuron that conveys the reinforcing properties of the rewarding unconditioned sugar stimulus during associative learning (Hammer, 1993). These circuit conditions make the lip region a unique site of odor and sugar reinforcement convergence. An additional information flow into the lip occurs through the GABAergic inhibitory feedback neurons in the protocerebrum-calyx tract (PCT) that originate in the MB lobes and project back into the calyx (Grünewald, 1999a). These neurons alongside with other MB output neurons (MBONs) exhibit learningdependent plasticity (Mauelshagen, 1993;Grünewald, 1999b;Okada et al., 2007;Strube-Bloss et al., 2011;Filla and Menzel, 2015). In Drosophila, it is believed that memory information is mainly stored in the output synapses between the KCs and MBONs (Heisenberg, 2003), and different MBON channels undergo learning-dependent plasticity by their own rules (Aso et al., 2014;Hige et al., 2015). In honeybees, however, both the anatomic convergence of the CS (odor) and the US (sugar reward) path-ways and the learning-specific plasticity of the KC spines (Szyszka et al., 2008) indicate associative olfactory plasticity also in the input site of the MB. Within the lip, these neural compositions locally form numbers of synaptic hubs, called microglomeruli (Ganeshina and Menzel, 2001). The PN boutons form the core of these microstructures by integrating the inhibitory input from recurrent PCT neurons (Grünewald, 1999a;Filla and Menzel, 2015), local excitatory feedback from the KCs (Ganeshina and Menzel, 2001), and the octopaminergic reward input (Hammer, 1993;Sinakevitch et al., 2013). This specific circuit configuration of the presynaptic PN boutons prompted us to think that they may play a key role in olfactory learning and are subject to experiencedependent plasticity.
Here, we studied neural plasticity of the PN boutons in the MB calyx with an emphasis on its relation to the learning rate on the level of the individual animal. The learning performance was monitored by the conditioned proboscis extension response (PER) while odor responses in individual PN boutons were measured by Ca 2ϩ imaging. During learning, individual boutons showed an increase or decrease of their Ca 2ϩ responses to both the rewarded (CSϩ) and the unrewarded stimulus (CS-). No significant differences between CSϩ and CS-odors were observed in the way the responses were up-or downregulated. However, the response pattern similarity decreased after learning. Moreover, the amount of neural plasticity in absolute terms (i.e., unsigned increase and decrease) induced for the CSϩ is strongly and positively correlated with the learning rate across individual animals. The temporal profile of the induced changes matched the Ca 2ϩ response dynamics of the inhibitory GABAergic feedback neurons, suggestive of associative activity modulation of boutons by these inhibitory neurons. We hypothesize that the observed plasticity in olfactory presynaptic terminals relates to the learned value of the stimulus, signifies short-term memory, and underlies the conditioned response (CR) behavior of the bees.

Preparation and dye loading
Bees were prepared as described previously (Yamagata et al., 2009). In short, foraging female worker bees were collected, chilled and fixed in recording chambers with wax. The head capsule was opened and a mixture of the solid Ca 2ϩ -sensitive dye fura-dextran (10,000 MW, Invitrogen) and the lysine fixable dye tetramethylrhodamine-dextran (10,000 MW, Invitrogen) was injected into the brain aiming for the soma cluster of the PNs of the lateral AL tract (l-ALT). Then the head capsule was closed, and the bees were fed until satiation and kept at 17-20°C for 8 -24 h. Before measurements, the legs and wings were cut, and the abdomen, thorax, and mandibles were immobilized with wax. The antennae were fixed with n-eicosane and the calyces of the MB were exposed for measurements. Kwik-Sil adhesive (World Precision Instruments, Inc.) was poured into the head capsule to completely stabilize the brain. After sealing the gaps with Vaseline, the recording chambers were filled with Ringer solution (130 mM NaCl, 7 mM CaCl 2 , 6 mM KCl, 2 mM MgCl 2 , 160 mM sucrose, 25 mM glucose, and 10 mM HEPES, pH 6.7, 500 mosmol).

Odor stimulation and conditioning
In most bees, 2-octanone and octanal (Sigma) diluted to 10 Ϫ2 with paraffin were used. 1-Hexanol and 2-octanol (Sigma) were used either with or without dilution (10 Ϫ2 and 5 ϫ 10 Ϫ2 , respectively). Whole during experiments, the bees were exposed to a constant air stream. Injection of the odorant (40 l soaked with 1 ϫ 2 cm filter paper) into the constant air stream was switched on and off by a computer-controlled solenoid valve (Galizia et al., 1997;Peele et al., 2006;Haehnel andMenzel, 2010, 2012). Odors were presented for 3 s at an interstimulus interval and an intertrial interval of 90 s.
Before starting conditioning, the proboscis extension response (PER) to the sugar stimulus of each bee was checked and only responding animals were moved to the recording site and given some minutes to rest. The protocol for the classical conditioning experiment (Bitterman et al., 1983;Matsumoto et al., 2012) followed the design given in Figure 1A. In the pre-training phase (PRE), each bee was exposed to two different odors for five to eight times in a pseudo-random order. After an interval of 3 min, bees were conditioned to one of the odors (CSϩ) by forward pairing the odor with an unconditioned reward stimulus (US) consisting of a drop of 30% sucrose. Conditioning always started with the rewarded odor (CSϩ) and the unrewarded control odor (CS-) was alternately presented (5-10 trials). After a 15-min retention, bees were again exposed to both odors for at least five trials (five to eight trials) in a pseudo-random order (POST). In all three phases, we monitored the animals' conditioned response (CR) as expressed in the PER by visual inspection. Only a complete extension of the proboscis was regarded as the CR. At the end of each experiment, the sugar response of each bee was tested again and only responding animals were included in the analyses.

Ca 2؉ imaging
Ca 2ϩ measurements were performed at room temperature with a sampling rate of 5 Hz, using a TILL-Photonics imaging setup mounted on a fluorescence microscope (Zeiss Axioskop). Fura was alternately excited at 340 and 380 nm. Exposure times were 15 and 60 ms, respectively. Each measurement started 3 s before stimulus onset and lasted for 10 s. Images were acquired through a 60ϫ/0.9 NA water dipping objective (Olympus), a 410-nm dichroic mirror and a 440-nm-long pass filter with an Imago CCD camera (640 ϫ 480 pixels, 4ϫ binned on chip to 160 ϫ 120). Pixel size was 1.47 ϫ 1.47 m, which allowed a resolution of single boutons of PNs. Approximately 1/40th of the calyx ring neuropil was imaged focusing on the frontal margin of the calyx. We performed imaging during the PRE and the POST phase while Ca 2ϩ responses were not recorded during the training phase due to movement artifacts of the brain by sugar stimulation. For imaging during the POST phase, animals were selected according to their behavioral performance during training. Consequently, only a small fraction of non-learners was imaged.

Confocal microscopy
After Ca 2ϩ measurements, the brain was dissected and fixed in 4% formaldehyde in Millonig's buffer overnight at 4°C. The brain was then rinsed in saline, dehydrated in ethanol, cleared in methyl salicylate, set into a chamber filled with methyl salicylate and observed with a confocal microscope (Leica TCS SP2; Leica). The excitation wavelength was 543 nm using a green HeNe laser. The entire brain was scanned with a 10ϫ/0.4 NA air objective (Olympus). Where necessary, the AL was scanned with a 20ϫ/ 0.70 NA air objective (Olympus) and the MB calyx was scanned with a 63ϫ/1.32-0.6 NA oil objective (Olympus). Morphologic images were acquired as an averaged raw fluorescence image of 380 nm during the measurements and was later unsharp mask-filtered in Photoshop (Adobe).

Identification and pre-processing of bouton activity
Recorded videos of Ca 2ϩ responses were preprocessed in IDL (RSI) using custom scripts as described in Yamagata et al. (2009). A mean of 15 frames during a single odor stimulation was calculated and displayed as a false-color image (Fig. 2). A spatial low-pass filter (5 ϫ 5 pixels) was applied to these images for better visualization. Individual boutons were determined as isolated activity spots (21 pixels) in the false-color images. For each bouton a response trace was calculated by averaging the signal of an activity patch without any filtering and correction.
Data were exported to Python or MATLAB (MathWorks Inc). Subsequent steps in the analyses were performed either using standard libraries for numerical/scientific programming in Python (NumPy, SciPy) of MATLAB. Overall, the observed Ca 2ϩ -dependent odor responses were dominated and overrepresented by excitatory signals. While an excitatory stimulation can lead to an arbitrary large increase in the Ca 2ϩ signal, the inhibitory effect on the Ca 2ϩ response is bounded, i.e., it can only suppress the relatively small amount of spontaneous Ca 2ϩ activity. Therefore, before analyzing odor response profiles in more details, we transformed all data using the hyperbolic tangent function. The function provides a useful transformation to reduce the effect of outlying values of a variable while compressing large values (Extended Data Fig. 2-1;Godfrey, 2009). In our data, this transformation enhances small changes around zero in the Ca 2ϩ signal and specifically improves the resolution of small inhibitory responses. This step facilitated the application of symmetric criteria for the classification of odor response profiles and their changes.

Signal quality
To analyze pronounced odor-dependent variations in response strength and signal quality, the signal quality of each animal as its average signal-to-noise ratio (SNR) was calculated according to Equation 1, with a b (t) representing the time-varying activity trace of bouton b for a total of N b boutons per animal. () and 2 () indicate the mean and the variance, respectively: SNRs for each animal and stimulus type were calculated as an average of responses during both TRAIN and POST phases. Based on the bimodal distribution of these values we defined an empirical threshold to distinguish between weak (low SNR) and strong (high SNR) signals.

ON and OFF responses
Excitatory and inhibitory ON and OFF responses were checked during respective response intervals, i.e., 1 s after odor on-and offset (five frames), for each bouton based on the trial-averaged Ca 2ϩ activities (Table 1). During a baseline period of 2.6 s before odor onset, we computed the mean a 0 and SD a0 across time. Excitatory responses were detected if the Ca 2ϩ activity exceeded a 0 Ϯ 2.5 a0 in at least three of the five frames during the response interval. To capture the fewer, weaker and mostly delayed inhibitory ON responses, we extended the considered frames, requiring subthreshold values in at least three of the 15 frames whole during the 3-s odor presentation. A threshold of 2.5 a0 proved to be suitable for separating the noisy signals that appeared during spontaneous activity from those values that are likely related to odor stimulation.

Odor response plasticity
To categorize response plasticity into four classes (up, down, both, none; Table 1, right), positive and negative thresholds were set as c 0 Ϯ 2.5 c0 , where c 0 and c0 define the average and standard deviation of baseline activity of each bouton with respect to odor type (CSϩ, CS-). Bouton responses were then regarded as changed if the activity during and 1s after odor stimulation exceeded the threshold at least once during the time window. This threshold provided a suitable compromise for filtering most of the changes that appeared during spontaneous activity while keeping those that might be related to odor stimulation.
A single measure of neural response plasticity (⌬NR) was also assigned to each animal and stimulus type (⌬NR ϩ for CSϩ and ⌬NRfor CS-). For this, we computed the sum of the absolute change across all frames that exceeded the aforementioned threshold on a per bouton basis. We then averaged the absolute change across boutons for each individual animal and stimulus type.

Behavioral plasticity
The learning effect was quantified by calculating the change in the behavioral performances during the training phase (⌬CRTRAIN) and the POST training phase (⌬CRPOST) for each individual animal by computing the difference between its PER-activity in all but the first CSϩ trials CR t CSϩ and in the same number of CS-trials CR t CSϪ divided by the number of trials N t . Thus, ⌬CR is the difference between the empirical probabilities of a CR p(CR) in CSϩ and CS-trials.
⌬CR ϭ 1 For the training phase (TRAIN), we excluded the first CSϩ trial from this calculation because only the subsequent trials of the training can be considered as measures in both acquisition and test trials. We defined bees as non-learners if they did not show a CR in any of the CSϩ trials during both phases TRAIN and POST following the arguments presented by (Pamir et al., 2011).

Correlation of behavioral and neural plasticity
We calculated Spearman's rank order correlation and the corresponding p value as a test for positive linear correlations between measures of behavioral performances (⌬CRTRAIN, ⌬CRPOST) and neural plasticity (⌬NR ϩ , ⌬NR -). We performed correlation analyses separately for the full data set consisting of all 18 animals and for the subset of animals with recordings that showed high SNR according to the definition given in the previous paragraph (eight animals in the case of CSϩ and 10 animals for CS-). Because of the low number of samples, we applied jackknife resampling as an estimate for the robustness and dependence on outliers of the correlation results.

Bouton pattern correlation
For each bee, we computed the spatial bouton response pattern by averaging the Ca 2ϩ activity within 1s after odor onset. The similarity of the response pattern between the CSϩ and CS-odors before and after training was quantified by the linear correlation coefficient (R). We computed Fisher's z-transform of the R for the statistical comparison, and back-transformed to obtain the average linear correlation coefficient. We used a one-sided Wilcoxon signed-rank test under the null-hypothesis that the pattern correlation does not reduce by learning. For control, we compared CSϩ/CS-odor response pattern correlation during the baseline period of 2.6 s before odor onset.

Behavioral performance of learning
We trained honeybees under the microscope in a differential conditioning paradigm as depicted in Figure 1A. One of two odors (CSϩ) was forward paired with the sucrose reward (US) while another (CS-) was presented without sucrose. We mainly used 2-octanone and octanal for odors (Fig. 1B). The performance of CR was evaluated by the probability of the proboscis extension response (PER) of the bee population to CSϩ [p(CR); Fig. 1C]. Since the CS preceded the US by 2 s, memory acquisition during the training phase was also examined (TRAIN). None of the bees showed a PER in any of the odor stimulation trials before training (PRE). Bees that did not show a behavioral response to any of the CSϩ trials in the training phase or the post training phase (POST) were classified as non-learners (five bees; see Materials and Methods). Including these non-learners, an asymptotic level of p(CR) Ϸ 0.7 was reached after five rewarded trials. The response level decreased gradually during the 10 unrewarded extinction trials (five trials each for CSϩ and CS-) in the POST phase (Fig. 1C, POST). We also observed PER to the CS-odor during training but not test, suggestive of a behavioral indication for CS-effect. A high behavioral performance during training phase implied a high performance after training phase (Pamir et al., 2014) with a significant correlation ( ϭ 0.61, p Ͻ 0.007). We observed similar learning performances by different odors, indicative of rather good behavioral performances by bees even in this condition.

Odor response profiles of PN boutons
To measure the odor-evoked Ca 2ϩ responses from terminal boutons of a defined class of PNs in the MB calyx, the l-ALT PNs were dye-loaded with the Ca 2ϩsensitive dye fura-dextran via their somata ( Fig. 2A). The dye was mixed with the lysine fixable dye tetramethylrhodamine-dextran that enabled to identify recorded neurons subsequently by a confocal microscopy ( Fig.  2B,C). The activity patches were defined for the localizations of boutons in Ca 2ϩ imaging data ( Fig. 2D; see Materials and Methods;Yamagata et al., 2009).
Considering the noisy nature of the small activities of boutons, multiple trials of odor-evoked Ca 2ϩ responses were examined before and after olfactory learning (Fig.  1B). Irrespective of the absolute signal strengths, nearly all of the recorded animals showed odor-and phasespecific spatial response patterns across boutons (Fig.  2E), a finding expected from the combinatorial odor processing and associative plasticity in the AL and the MB (Joerges et al., 1997;Faber et al., 1999;Faber and Menzel, 2001;Szyszka et al., 2005;Krofczik et al., 2008;Fernandez et al., 2009;Yamagata et al., 2009;Galizia and Rössler, 2010;Rath et al., 2011;Arenas et al., 2012). The trial variances of these responses were reasonably small such that the Euclidean distances of Ca 2ϩ responses between given two trials within the same phase (i.e., PRE or POST) were significantly smaller than that between phases [i.e., PRE vs POST; Extended Data Fig. 2-2A; Kolmogorov-Smirnov (KS) test, p Ͻ 0.0001]. Accordingly, we conducted further analyses based on the trialaveraged bouton activities to improve the S/N.
To acquire an overview of the odor response profiles of boutons in the untrained condition, we categorized their responses (ON, OFF, ON and OFF, and none) by stimulus Types of excitatory odor responses of untrained animals: all 1652 boutons were categorized by their response profiles of (1) only ON responses, (2) only OFF responses, (3) both ON and OFF responses, or (4) no detected responses to CSϩ and/or CS-odors. Types of excitatory bouton response changes by learning: all 1652 boutons were categorized by their response profiles of (1) only up-regulated, (2) only down-regulated, (3) both up-and down-regulated, or (4) no detected change to CSϩ and/or CS-odors.
type (CSϩ, CS-, and both; Table 1; see Materials and Methods). Although boutons were visually identified by the responses (Fig. 2D), ϳ17% of the activities appeared too sparse or too noisy to satisfy our rather conservative criteria (none and both group). The fraction of ON, OFF, and ON and OFF responses were roughly equally distributed in both odors to be CSϩ and CS-, suggestive of their similarity in odor responses before learning. 1-Octanal elicited generally stronger odor responses than 2-octanone, which is represented as the bimodal distribution of signal qualities (Extended Data Fig. 2-2C). Temporal properties of odor responses were analyzed by calculating time-resolved activities for each bouton. Individual bouton response traces of a representative bee are exemplified in Figure 2F. A measure for neural plasticity (⌬NR) was calculated by taking the differences between PRE and POST phases. Figures 3, 4 show all recordings of 1652 boutons from 18 bees in response to CSϩ (Fig. 3) and CS- (Fig. 4) odors. The boutons typically exhibited phasic-tonic activities that may or may not last longer than an odor presentation. Bouton response patterns were rather homogeneous within an animal but diverse across animals (Figs. 3, 4). Accounting for this, the Euclidean distances of individual bouton activities within an animal were significantly smaller than those between animals (KS test, p Ͻ 10 Ϫ4 ; Extended Data Fig. 2-2B). Within the first second after odor onset, PN bouton activities were clearly dominated by excitatory responses (Figs. 3, 4). In contrast, inhibitory responses of boutons arose with rather slower kinetics. We compared their latencies, i.e., the first time point of the Ca 2ϩ traces that exceeded the defined thresholds (see Materials and Methods). Most excitatory responses started within the first two frames after odor onset with an average of 0.22 s (Figs. 5A,C) while that of inhibitory responses distributed more widely and were much slower, with an average of Ϸ 1 s (Figs. 5B,C). These distinction suggest separated mechanisms for excitatory and inhibitory inputs onto the PN boutons, presumably through feed-forward excitatory inputs mainly from PN axons and feed-back inhibitory inputs mainly from GABAergic PCT neurons (Grünewald, 1999a;Filla and Menzel, 2015).

Plasticity of bouton activities and their correlation with behavior
We sought for neural correlates of learning behavior in the PN boutons. To grasp a coarse overview of the neural plasticity that takes place in the boutons, we first categorized boutons by their response profiles during learning (Table 1). A majority (ϳ50%) of them changed the responses only to one direction, i.e., either increased or decreased their responses, while ϳ13% of boutons showed changes in both directions. Response reduction appeared more often (Ͼ30%) than response increase, suggesting a dominant learning effect through some inhibitory mechanisms. The probability of each type of change was roughly equal between CSϩ and CS-odors, suggesting no apparent contrasts between rewarded and unrewarded odors in this regard. We therefore tested whether the overlap between the spatial bouton response patterns to the CSϩ and CS-odors is reduced as a result of learning (see Materials and Methods) and found that, indeed, there is a significant reduction (p Ͻ 0.05, Wilcoxon signed rank test) of the response pattern correlations with an average reduction from R ϭ 0.30 to R ϭ 0.22 (Extended Data Fig. 2-3).
We hypothesized that the neural plasticity in the PN boutons correlates with the behavioral plasticity. We assigned the unsigned neural plasticity ⌬NR, which only considers absolute changes of bouton activities (see Materials and Methods). Likewise, we quantified the changes in behavior (⌬CR) of each individual bee during and after training (see Materials and Methods). We found a clear and highly significant positive correlation between learning induced neural plasticity of CSϩ odor and behavioral performance ( Fig. 6A; ϭ 0.76, p Ͻ 0.0002). This correlation was even more pronounced for a subset of animals that showed Ca 2ϩ responses with a high SNR (see Materials and Methods; ϭ 0.84, p Ͻ 0.004). In contrast, the CS-plasticity did not show a significant correlation with the behavior (Fig. 6B), neither when considering all animals ( ϭ 0.29, p Ͼ 0.12) nor for a subgroup of high SNR bees ( ϭ 0.09, p Ͼ 0.39).
The robustness of the correlation was confirmed by a jackknife approach, where neural and behavioral data from all but one animal was iteratively resampled (Fig.  6C,D). Irrespective of signal SNR, the correlation of CSϩ odor was narrowly distributed around the above stated values. The corresponding p values suggested statistical significance in all cases. In contrast, CS-odor exhibited much lower correlations with high p values. Taken together, we concluded that the observed plasticity in the PN bouton correlates with the learned value of a stimulus to alter the behavior of the bees.
Finally, we also assessed the plasticity in temporal dynamic of odor responses by calculating the difference of excitatory odor response between the PRE and POST phases (Fig. 5B). Interestingly, the time course resembled temporal kinetics of the odor response of the MB extrinsic feedback neurons previously reported (Haehnel and Menzel, 2010). This finding suggests that the observed plasticity in the PN boutons can be explained by the altered input strength from the inhibitory feedback neurons onto the boutons.

Discussion
Here, we studied neural plasticity of the PN boutons in the MB calyx in relation to the learning rate at the level of the individual bee. To this end, we combined Ca 2ϩ imaging and associative olfactory learning. A majority of bees readily performed learning behavior under our experimental conditions (Fig. 1). Both at PRE and POST phases, odor-unique and repetitive spatiotemporal population responses of PN boutons were observed (Figs. 2-4) as shown previously (Szyszka et al., 2005;Yamagata et al., 2009). We found upregulation and downregulation of bouton activities to CSϩ and CS-odors, which occurred equally frequently by odors (Table 1). The absolute amount of neural plasticity was strongly and positively correlated with the behavioral learning success across  individuals for the CSϩ but not the CS-odor (Fig. 6). The absence of correlated neural plasticity in CS-may be relevant to (1) the neural plasticity linked to the behavioral CS-effect (Fig. 1C); (2) the hidden neural adaptations that are not directly expressed in behavioral change, e.g., retardation of reversal learning or decreased generaliza-tion in the course of differential conditioning (Extended Data Fig. 2-3; Menzel et al., 2007); and (3) neural changes that are not accompanied with behavioral changes, which is exemplified and well documented with neural activity in PCT neurons in the MB during PER conditioning to visual cues or visual context (Filla and . Moreover,  the temporal profile of the induced changes by odors matched that of the inhibitory GABAergic PCT neurons (Haehnel and Menzel, 2010), suggestive of a functional linkage between PN boutons and PCT neurons for learning. We thus hypothesize that the observed plasticity in olfactory presynaptic terminals of the PN boutons relates to the learned value of the stimulus, signifies short-term memory and underlies the CR behavior.
Each single bouton of PN receives olfactory input from the AL and forms the core of a microglomerular structure by contacting with several postsynaptic KCs, inhibitory presynaptic feedback PCT neurons and modulatory VUM reward neurons (Hammer, 1993;Ganeshina and Menzel, 2001;Schröter and Menzel, 2003;Sinakevitch et al., 2013). These neuroanatomical properties suggest a separate regulation of excitatory and inhibitory inputs onto the PN boutons. Indeed, the majority of the excitatory responses match the phasic-tonic profiles and low response latencies of PNs, known from previous Ca 2ϩ imaging data recorded in PNs of the AL (Peele et al., 2006;Fig. 5A) and the typical intracellularly and extracellularly recorded spike responses in PNs (Krofczik et al., 2008;Strube-Bloss et al., 2011;Brill et al., 2013). The average inhibitory response, on the other hand, matches previously recorded data from inhibitory PCT neurons (Haehnel and Menzel, 2010;Fig. 5B). Interestingly, the difference between the average excitatory profiles (PRE-POST) is closely related to the temporal structure of the average inhibitory response and also resembles those observed in the PCT neurons (Haehnel and Menzel, 2010;Fig. 5B). Thus, decreased responses in the PN boutons are likely to be the consequence of activity modulation by the GABAergic PCT neurons. Likewise, a part of activity increase in boutons may also be explained by the inhibitory modulation of odor responses of PCT neurons on learning (Grünewald, 1999b). In bees, GABAergic signaling in the MB calyces are reported to be important for the discrimination of similar odors (Devaud et al., 2015). It is thus conceivable that our observed plasticity in the PN boutons contributes to enhance odor discriminability (Extended Data Fig. 2-3), thereby underpinning an odor-specific CR.
The clear correlation between the absolute amount of plasticity with the behavioral change (Fig. 6) highlights the importance of both upregulation and downregulation of bouton activities in learning. Seen in the context of the olfactory pathway in insects, olfactory memory is distributed across various sites: from the pre-and postsynaptic sites in the glomeruli of the AL to the pre-and postsynaptic sites in the MB input (the lip of the calyx), and the MB output lobes (Heisenberg, 2003;Menzel, 2012). In those neural circuits in bees, the increase and decrease of neural changes for the learned stimulus is a wide spread property of neurons. For example, spike activity recorded extracellularly from PNs increase or decrease their responses to the learned odor with similar probability (Denker et al., 2010), which potentially obscures associative changes as recorded with Ca 2ϩ imaging at the postsynaptic sites in AL glomeruli (Peele et al., 2006). Fernandez et al. (2009) andRath et al. (2011) also found that the total response strength of PNs measured as Ca 2ϩ signals does not change after conditioning although the response patterns changed. Although MB KCs predominantly increase their responses for the CSϩ and decrease it for the CS-, also respective opponent changes were found (Szyszka et al., 2008). MBONs are well known for their opponent changes in the course of learning for both an olfactory cue and a visual context (Haehnel and Menzel, 2010;Strube-Bloss et al., 2011Filla and Menzel, 2015). Similar opponent changes were also found in MBONs in Drosophila (Cohn et al., 2015;Owald et al., 2015). Neural plasticity without altering overall excitation may therefore be a general mechanism seen in the insect brain, presumably to achieve effective memory encoding in the limited coding space with less energy consumption.
Although we could not image the boutons during the acquisition phase due to technical limitations, we assume that synaptic plasticity develops already during acquisition. This idea is supported by the finding of neural changes already 15 min after the last training trial. Hammer and Menzel (1998) also found that US substitution by octopamine injection into the MB calyx paired with an odor leads to the acquisition of appetitive memory. The synaptic plasticity observed in PN boutons is likely to be a component of short-term memory (Menzel, 1999(Menzel, , 2012. In this respect, the olfactory input to the MB shows characteristics of some particular MB extrinsic neurons, e.g., the PE1 neuron (Mauelshagen, 1993;Okada et al., 2007) and other A1/A2 output neurons (Menzel and Manz, 2005) but not of all MB extrinsic neurons that may develop their learning-dependent plasticity only 3 h after the last learning trial (Strube-Bloss et al., 2011) or even later (Haehnel and Menzel, 2012). Both short-term memory and consolidation to long-term memory may thus be the characteristics of the MB, although their neuroanatomical substrates may not overlap completely. In flies, a segregation of local circuits within the MB is well established for short-term and long-term memory storage and encoding (Trannoy et al., 2011;Yamagata et al., 2015Yamagata et al., , 2016. Within each of the animals, the temporal structures of individual bouton responses were often closely related to each other. Due to technical limitations, we observed a specific, locally restricted patch of the MB calyx of each animal, comprising an estimated 8% of the total area. We may thus assume that this small fraction of calyx consists of boutons from a few and possibly similarly tuned PNs. Recent evidence from the fruit fly suggests that PNs with similar tuning properties have a tendency to converge onto the same KCs (Gruntman and Turner, 2013). PNs also form multiple boutons along their axons in the calyces (Zwaka et al., 2016), which exhibit correlated activity to odors (Yamagata et al., 2009).
While we observed animals with a low neural plasticity score ⌬NR ϩ that also showed a stable behavioral performance ⌬CR, we did not observe animals with a high neural plasticity score accompanied by a weak behavioral performance (Fig. 6A). A likely explanation is that we captured only a fraction of the neural changes that appeared in a specific animal due to the undersampling of a small fraction of the calyx. In addition, the staining procedure may have resulted in a varying amount of back-filled PNs. Presumably through these and other related issues, the quality of odor-evoked signals differed across preparations and for different odors, and physiologic changes may have been concealed. These reasons can lead to a reduced but not to an overestimated ⌬NR ϩ , which could explain the observation in Figure 6A.
In the MB calyx, associative plasticity is conveyed by multiple network properties, i.e., modulation of the presynaptic sites in the AL glomeruli of PNs (Faber et al., 1999;Fernandez et al., 2009;Rath et al., 2011), the postsynaptic sites in the spines of the MB intrinsic KCs (Szyszka et al., 2008), and the postsynaptic sites of the inhibitory feedback neurons in the PCT (Grünewald, 1999b;Haehnel et al., 2012;Filla and Menzel, 2015). The octopamine immunoreactive reward pathway of the VUM neuron with presynaptic terminals to both the PN boutons and the KC spines (Ganeshina and Menzel, 2001;Sinakevitch et al., 2013) also changes its response properties to the forward paired odor (Hammer, 1993). We emphasize that the observed plasticity in the PN boutons is not a simple copy of these neural correlates of olfactory memory by the following reasons. (1) Local synaptic profiles surrounding the PN boutons receive octopaminergic modulatory inputs from the VUM neurons (Ganeshina and Menzel, 2001;Sinakevitch et al., 2013), indicative of their capability as a site of inducing associative plasticity. Indeed, Hammer and Menzel (1998) demonstrated that a focal microinjection of octopamine into the MB calyx paired with an odor leads to the formation of appetitive memory. In Drosophila, microglomeruli in the MB calyx constructed by PN boutons, KC spines and GABAergic interneurons (Yasuyama et al., 2002;Butcher et al., 2012) also undergo activity-dependent physiologic plasticity (Pech et al., 2015).
(2) Boutons of the PNs are the convergence site of two forms of learning related plasticity (i.e., from the AL via PNs and from the MB Lobes via PCTs), certainly storing different contents from the memory in these brain regions. Thus, the PN boutons are expected to act as a hub of multiple memory information and sensory inputs. It is, therefore, reasonable to see associative plasticity in the PN boutons in both increasing and decreasing manner, presumably depending on which of its inputs (excitatory from the AL, inhibitory from recurrent neurons) are more effective. Although a clear correlation between conditioned behavior and neural plasticity (Fig. 6) suggests a specific role of PN boutons for mediating learned value of odors, it may not be the sole function of the microglomeruli. In bees, experience and age-dependent structural plasticity has been reported (Groh et al., 2012;Muenz et al., 2015), indicating a convergence of state information and a resultant circuit rewiring. Future studies employing a state-of-the-art genetic technique that makes it possible to target specific synapses (Hayashi-Takagi et al., 2015) would provide valuable insights into the functional significance of this unique microcircuitry.