Calcium in Kenyon Cell Somata as a Substrate for an Olfactory Sensory Memory in Drosophila

Animals can form associations between temporally separated stimuli. To do so, the nervous system has to retain a neural representation of the first stimulus until the second stimulus appears. The neural substrate of such sensory stimulus memories is unknown. Here, we search for a sensory odor memory in the insect olfactory system and characterize odorant-evoked Ca2+ activity at three consecutive layers of the olfactory system in Drosophila: in olfactory receptor neurons (ORNs) and projection neurons (PNs) in the antennal lobe, and in Kenyon cells (KCs) in the mushroom body. We show that the post-stimulus responses in ORN axons, PN dendrites, PN somata, and KC dendrites are odor-specific, but they are not predictive of the chemical identity of past olfactory stimuli. However, the post-stimulus responses in KC somata carry information about the identity of previous olfactory stimuli. These findings show that the Ca2+ dynamics in KC somata could encode a sensory memory of odorant identity and thus might serve as a basis for associations between temporally separated stimuli.


INTRODUCTION
Odorants evoke odorant-specific spiking patterns across olfactory receptor neurons (ORNs), which drive odorant-specific neural activity patterns in different brain areas. In mammals, odorants evoke activity first across ORNs in the olfactory epithelium and then in glomeruli in the olfactory bulb, followed by responses in the olfactory cortex, the amygdala, and other brain areas (Uchida et al., 2014). Similarly, in insects, activity across ORNs first drives responses in olfactory glomeruli in the antennal lobe, and later in higher brain regions such as the mushroom bodies (Galizia, 2014). In neurons, activity (e.g., membrane potential or changes in cytosolic Ca 2+ concentration) in the dendrite is often different from activity in axon terminals. Therefore, in order to understand information processing along neural signaling chains, it is necessary to record neural activity in subcellular compartments. Ca 2+ imaging is a suitable technique: Ca 2+ enters the cytosol through ligand-gated and voltage-dependent Ca 2+ channels (as a function of local membrane potential) and/or through release from intracellular stores (as a function of second messenger cascades, including Ca 2+ itself), and its concentration is restored through buffers, membrane pumps, and carriers (Augustine et al., 2003). Ca 2+ has multiple functions: in presynaptic terminals Ca 2+ triggers vesicle release (Stanley, 1997); at post-synaptic sites Ca 2+ is involved in long-term synaptic plasticity (Khodakhah and Armstrong, 1997); in somata, Ca 2+ influences gene transcription (Lyons and West, 2011). Therefore, recording the dynamics of cytosolic Ca 2+ across neuronal compartments sheds light on different cellular processes.
The olfactory system of flies consists of several layers of neurons ( Figure 1A). ORNs that express the same receptor type coalesce in the same glomerulus (Vosshall et al., 2000;Couto et al., 2005). Within a glomerulus, all ORNs converge onto uniglomerular projection neurons (PNs) (Kazama and Wilson, 2009). Excitatory and inhibitory local neurons interconnect glomeruli and mediate gain control Silbering and Galizia, 2007;Root et al., 2008) and decorrelate odorantevoked activity patterns (Wilson et al., 2004). Lateral inhibition via local neurons and short-term synaptic depression at the ORN-PN synapse shorten PN's odorant-evoked responses, increase their sensitivity for concentration changes (Wilson et al., 2004;Bhandawat et al., 2007), and increase their odorant-specificity (Olsen and Wilson, 2008;Silbering et al., 2008). PNs connect the antennal lobe with the mushroom body calyx ( Figure 1A). The mushroom body-intrinsic Kenyon cells (KCs) respond to fewer odorants and generate fewer spikes than individual PNs (Perez-Orive, 2002;Szyszka et al., 2005;Turner et al., 2008). This sparsening of KC responses is caused by the divergent connectivity between PNs and KCs (Jortner et al., 2007;Caron et al., 2013;Eichler et al., 2017) and by inhibitory feedback (Demmer and Kloppenburg, 2009;Lei et al., 2013;Lin et al., 2014). KCs send their axons to the mushroom body lobes, where they receive input from dopaminergic neurons (Aso et al., 2010;Burke et al., 2012). In standard classical conditioning, the simultaneous presentation of an odorant and a reward or punishment (reinforcer, encoded by dopaminergic neurons) induces an associative odor memory in flies (Tully and Quinn, 1985), which is encoded as a change in synaptic strength between KC axon terminals and mushroom body output neurons. The formation of this memory depends on cytosolic Ca 2+ signaling and requires coincidence detection of an elevated Ca 2+ concentration in KC axon terminals (for the odorant) and activated dopamine receptors (for the reinforcer) by an adenylyl cyclase (McGuire, 2001;Schwaerzel et al., 2002;Tomchik and Davis, 2009;Gervasi et al., 2010;Séjourné et al., 2011;Aso et al., 2014;Cohn et al., 2015;Hige et al., 2015;Owald et al., 2015). However, this model of associative odor learning cannot explain associative odor learning during trace conditioning. In olfactory trace conditioning, the reinforcer arrives after the odor stimulus has already terminated, and there is no overlap between the two stimuli. In this situation, the reinforcer arrives when the cytosolic Ca 2+ in KC axon terminals is already back to baseline (Galili et al., 2011;Shuai et al., 2011;Szyszka et al., 2011;Dylla et al., 2013Dylla et al., , 2017. Where, then, is coincidence detection possible for associative trace conditioning? Trace conditioning requires an as-yet elusive sensory memory of the odorant. Such sensory odor memory is a form of short-term, non-associative memory, which may also be used in other contexts: honey bees use a sensory odor memory to solve delayed-matching-to-sample tasks (Giurfa et al., 2001), and flies and moths use a sensory odor memory to continue their odor source search after losing an attractive odor plume (Vickers, 2005;van Breugel and Dickinson, 2014;Saxena et al., 2017). Behavioral data show that these sensory odor memories can last for several seconds. Previous studies have proposed that such sensory odor memories could be encoded by prolonged changes in Ca 2+ concentrations (Tomchik and Davis, 2009;Galili et al., 2011;Szyszka et al., 2011;Yarali et al., 2012) or synaptic modification (Drew and Abbott, 2006;Izhikevich, 2007;Cassenaer and Laurent, 2012), but the cellular localization of sensory odor memories is still unknown.
Because in Drosophila, associative odor learning requires coincidence detection of an elevated cytosolic Ca 2+ concentration (representing the odorant) and activated dopamine receptors (representing the reinforcer) (Tomchik and Davis, 2009;Gervasi et al., 2010), we here characterized cytosolic Ca 2+ signals across neurons and neuronal compartments in the olfactory system of Drosophila. We used the genetically encoded Ca 2+ reporter GCaMP to record odorant-evoked cytosolic Ca 2+ signals in ORN axon terminals (in antennal lobe glomeruli), two compartments of PNs (dendrites and somata in the antennal lobes), and two compartments of KCs (dendrites and somata in the mushroom bodies). We focused on how odorant-evoked Ca 2+ signals develop during the presence of odorants and after their offset in experimentally naïve, unconditioned flies. In particular, we tested whether stimulusoutlasting Ca 2+ signals could encode sensory odor memories. We found that all neuronal compartments show odorant-specific Ca 2+ activity patterns during the stimulus time. Furthermore, all compartments showed prolonged odorant-evoked Ca 2+ concentration changes after stimulus offset. However, only the somata of KCs showed Ca 2+ signals with a pattern that could encode odorant identity for several seconds after stimulus offset. Therefore, prolonged Ca 2+ signals in KC somata encode a sensory odor memory which could be used to associate an odorant with a delayed reinforcer during trace conditioning.

Nomenclature
To avoid terminological confusion, we define the terms used in this paper here. Classical conditioning: pairing a conditioned stimulus with a reinforcer. Standard conditioning (synonym: delay conditioning): the most frequent form of classical conditioning with overlapping conditioned stimulus and the reinforcer. Trace conditioning: classical conditioning with a temporal gap between the offset of the conditioned stimulus and the onset of the reinforcer. Trace conditioning requires a sensory odor memory (synonym: trace). A sensory odor memory differs from a working memory, as a sensory odor memory does not require attention.

In Vivo Calcium Imaging
Female flies (aged 2-14 days after eclosion) were prepared either for antennal lobe recordings as described previously (Silbering and Galizia, 2007;Galili et al., 2011), or for accessing the layer of KC somata from the posterior side of the fly's head, similar to the methods described elsewhere (Murthy and Turner, 2010;Honegger et al., 2011;Campbell et al., 2013). In both preparation methods, the fly head was fixed with a composition of low melting wax (hard wax, soft-sticky wax, and myristic acid, composition ratio 2:1:2, melted and mixed), and the antennae were shielded from the Ringers' solution with a plastic coverslip which had a window to gain access to the head capsule. Gaps between the cover slip window and the head were sealed with silicone elastomer (Kwik Sil, World Precision Instruments, Sarasota, FL, United States). Through the window the cuticle was cut, opened, and removed, so that the brain was visible.
Calcium imaging was done with experimentally naïve, unconditioned flies. Calcium imaging of ORNs and PNs in the antennal lobe was performed as described previously (Galili et al., 2011) with a fluorescence microscope (either BX-50 WI, Olympus, Tokyo, Japan, or AXIO Examiner.D1, Zeiss, Jena, Germany), equipped with a 40× water-dip objective (NA 1.0; Zeiss, Jena, Germany). The excitation wavelength (475-nm, Monochromator Polychrome II or Polychrome V, Till Photonics, Gräfelfing, Germany) was filtered with a 500-nm short-pass filter and reflected onto the sample by a 495-nm dichroic mirror. The emission light was filtered by a 505-nm long-pass filter, before being captured by the CCD-Camera (either Imago QE at the Olympus Microscope or PCO. Imaging SensiCam at the Zeiss Microscope, both: Till Photonics, Gräfelfing, Germany). On-chip binning of pixels (4 × 4 in the Imago QE and 8 × 8 in the SensiCam) resulted in a resolution of 160 × 120 pixels, corresponding to 145 × 109 µm on the preparation (Olympus Microscope) or 172 × 130 pixels, corresponding to 242 × 183 µm on the preparation (Zeiss Microscope). The recording rate was 5 Hz for 35 s (175 frames). The exposure time was adjusted between 120 and 180 ms and the basal fluorescence was adjusted either by adding gray filters into the excitation light beam (5, 10, and 32% transmission, at the Polychrome II, Olympus Microscope), or by tuning the intensity and bandwidth of the Polychrome V monochromator (at the Zeiss Microscope). During the recording, odorant stimulation was controlled by the acquisition software of the imaging system (Till Vision, Till Photonics). Odorants were applied for 10 s in a pseudo-randomized order. The interval between recordings was 2 min (interstimulus interval, ISI, 2 min).
The KC and PN somata recordings were performed with the Zeiss LSM 510 Confocal Microscope to reduce scattered light. Excitation light was 488-nm (Argon laser), the objective 40× water-dip (NA 1.0; Zeiss, Jena, Germany). Odorant stimulation was controlled by a stimulus control device (cRIO-9074 combined with IO module NI-9403, National Instruments) with custom-written software (by Stefanie Neupert), which was synchronized by the image acquisition software. Odorants were applied for 10 s in a pseudorandomized order. The interval between recordings (ISI) was 2 min.

Odorants
Odorants (Sigma-Aldrich, Deisenhofen, Germany) were diluted in 10 mL of mineral oil (MO, Sigma-Aldrich) in 100 mL rolledflange glass bottles (Fisher Scientific GmbH, Schwerte, Germany), which were sealed with silicon-Teflon septa (Schmidlin Labor Service GmbH, Schwäbisch Gmünd, Germany). The bottles were connected to a custom-built, computer-controlled olfactory stimulator  via syringes (1.2 mm external diameter) through the septum. Fresh odorant solutions were prepared every 1-4 weeks. A constant airstream (3 L min −1 ) was applied to the fly's antennae through a glass tube (inner diameter 6.2 mm), which was located approximately 8 mm away from the fly. This constant airstream was the sum of a carrier airstream (1.2 L min −1 ) and six channels (each 0.3 L

Data Analysis
The imaging data were analyzed with custom routines written in Python 2.7 1 , R 2 , and IDL (RSI, Boulder, CO, United States).

Data Preprocessing
First, the image sequences obtained by Ca 2+ imaging were movement corrected (anatomical landmark-based for ORN and PN recordings in the antennal lobe; affine and nonlinear registration for PN and KC somata and KC dendrites) within and between measurements. The relative fluorescent change F/F was calculated for each time point i of the recording as F/F = (F i -F B )/F B , where F i is the absolute fluorescence of the i th frame and F B is the background fluorescence, which was calculated as the average fluorescence of 15 frames before odorant stimulation (frames 10-25). Response traces were corrected for bleaching, by fitting and subsequently subtracting an exponential decay function F(t) = ae bt +c to the average light intensity change  in each region of interest (ROI) over time (Galizia and Vetter, 2004).

Region of Interest Selection in Imaging Data
The ROIs were determined manually in each fly using the interactive calcium signal data analysis suite ILTIS 3 or IDL routines. Glomerular responses were calculated by averaging the fluorescent light intensity of 7 × 7 pixels (corresponding to 6.3 × 6.3 µm) in the center of a glomerulus. Glomeruli were identified based on anatomical cues and on their response profiles (Silbering and Galizia, 2007;Silbering et al., 2008;Galili et al., 2011). For single soma resolution imaging, somata responding to at least one of the five odorants during the whole recording were selected. The choice of whether a soma was active or not was based on visual observation, but conservative in the 3 https://github.com/grg2rsr/ILTIS sense that even very weakly responding cells were selected. This means that somata that did not respond to any of the odorants could not be selected, resulting in an overestimate of responding cells as compared to when all KCs are considered. ROI size was 8-10 pixel diameter and placed centrally on the soma. KC dendrites in the calyx did not show any clear pattern of single cell resolution. Hence, we spatially downsampled the recordings and treated each 4 × 4 pixel bin as one ROI.
Only ROIs with a signal larger than 0.75 F/F to any of the 5 odorants were included in later analyses. The following number of flies and glomeruli/somata/ROIs were used in the analyses:

Figures 2, 3, 7:
ORN axons: N = 9 flies, n = 85 glomeruli (glomeruli per fly: 11, 11, 10, 5, 10, 10, 7, 10, 11). PN dendrites: N = 10 flies, n = 88 glomeruli (glomeruli per fly: 9,5,8,9,11,10,10,12,7,7). show that odor and post-odor response patterns are reproducible and stable in PN somata, KC dendrites, and KC somata (see high correlation values during the odor and after odor offset). In PN somata, the odor response pattern changes into a distinct post-odor response at odor offset (see the gray areas right and below the correlated odor pattern rectangles). In KC dendrites, and even more in KC somata, the pattern change is smooth and the post-odor response pattern retains a similarity to the odor response pattern. Correlation values were defined as being significant when p < 0.005 and are color coded (color scale bar, right). Non-significant values are shown in gray. Purple and blue frames mark the time windows used to calculate the time-resolved correlation, shown in (B).
(B) Time-resolved correlation between the odor response pattern (marked by the purple square above the graph) and the post-odor response pattern (marked by the blue square) of an odor response with all time points of another response to the same odorant (purple and blue traces, respectively). The odor response pattern breaks down at odor offset in PN somata, and yields to a distinct post-odor response pattern which evolves at odor offset. Conversely, KC somata show a smooth (Continued)  9,5,8,9,11,10,10,12,7,7).

Categorization of Response Dynamics
In order to categorize calcium responses according to their response dynamics, we formed three categories into which a response could fall: on (only responding during the stimulus), off (only responding after the offset of the stimulus), or prolonged (sustained responses starting with the stimulus but outlasting it) ( Figure 2C). For our categorization, we defined thresholds that allowed us to unambiguously classify each recorded response into one of the three categories: We defined on or prolonged responses those which crossed a threshold of 2.57 × SD of the pre-stimulus signal (corresponding to a significance level of p < 0.005). If they declined below 37% (corresponding to 1/e) of the maximum response value within 5 s after odorant offset, they were classified as on, otherwise as prolonged. We defined off responses those which crossed the response threshold 5 s after odorant offset but not during the stimulus.

Correlation Matrices and Traces
Time-resolved correlation matrices were constructed to compare the change of patterns and their similarity during the recording. This was done by first combining data from all animals for either a given odorant (as in Figure 3 for 1-butanol and Supplementary Figure S3) or all odorants (as in Figure 6

Classification Analysis
Classification analysis was performed by training a support vector machine (SVM) with a 2-s averaged response pattern sliding across the total time of the recording, and testing all other time points afterwards. To assess overall classification performance, the average fraction of correct classifications during the 10 s stimulus window was calculated. All SVM classifications were performed with a ν-SVM (python module sklearn.svm.NuSVC with a ν-parameter of 0.9). Chance classification levels were obtained by shuffling all labels and repeating the classification 250 times, and the upper 95% confidence interval is taken as a threshold for significant classification (Combrisson and Jerbi, 2015).

RESULTS
We studied the odor-specificity of odorant-evoked activity during and after olfactory stimulation in different neuronal populations and compartments in the fly brain. We measured odorant-evoked Ca 2+ concentration changes in ORN axons, PN dendrites, and PN somata (in the antennal lobes) and in KC dendrites and KC somata (in the mushroom bodies) (Figure 1A). Stimulus onset was defined as t = 0 s. Responses during stimulation were called "odor response, " activity after stimulation "post-odor response" (Figure 1B). We defined "odor response pattern" as the averaged response of 5 frames from t = 1-2 s, and "post-odor response pattern" as the 5 frames from t = 15-16 s (Figure 1B, bottom). We chose this time point as post-odor response pattern because trace conditioning experiments have shown that flies still have a sensory memory of the odorant at this time point (Galili et al., 2011). We categorized Ca 2+ responses according to their response dynamics, forming three threshold-based categories: on (only responding during the stimulus), off (only responding after the offset of the stimulus), or prolonged (sustained responses starting with the stimulus but outlasting it) ( Figure 2C, see methods for details on the construction of thresholds used for categorization).

Glomerular Post-odor Response Patterns in the Antennal Lobe Are Dissimilar to Odor Response Patterns
The Ca 2+ responses in ORN axons and PN dendrites (Figure 2A) in the antennal lobe were odor-and glomerulus-specific ( Figure 2B), with a tendency of PN responses to be more phasic than ORN responses and with more pronounced off-responses. Across glomeruli, axonal ORN activity patterns and dendritic PN response patterns generally involved the same glomeruli, in line with previous publications (Wang et al., 2003;Silbering and Galizia, 2007). In most glomeruli (three example glomeruli are shown in Figure 2B), positive responses decreased rapidly after odorant offset and reached baseline within our recording time of 20 s for both ORNs and PNs. Acetic acid (AceA) and propanoic acid (ProA) induced negative responses in some glomeruli during the stimulus, which were sometimes followed by positive off responses. The diversity of response dynamics across odorants was large in ORNs, with all three categories (on, off , and prolonged) taking up a large share (medians of 25, 44, and 27% for on, off , and prolonged, respectively, Figures 2C,D). Conversely, on responses dominated in PNs (75%, Figure 2D).
As the spatial activity pattern across glomeruli contains information about odorant identity (Guerrieri et al., 2005), a stable activity pattern over an extended period of time after stimulus offset would be indicative of a possible sensory memory. Thus, we performed a spatiotemporal analysis by calculating the correlation between stimulus evoked activity patterns along the entire recording ( Figure 3A). The resulting time-resolved correlation matrices (Figure 3A) show the correlation between the response patterns at each single time point of each recording, where each pixel displays the Pearson's correlation coefficient r of the two response vectors at these particular time points.
For ORNs, the odorant-evoked glomerular patterns during the 10-s-long stimuli were correlated (as shown by the compact dark red "field" in the plot of Figure 3A), showing that the odorant-evoked patterns were largely invariant during the entire stimulation period and reproducible in different trials. However, the activity patterns after stimulus offset were not correlated to the odor responses anymore (Figure 3A, gray areas to the right and below the odor response "field").
Although post-odor response patterns in ORNs were not correlated to the odor response patterns, they were fairly stable within themselves ( Figure 3A). However, they were neither odorant-specific nor reproducible (no increased correlation between the post-odor response patterns of two 1-butanol trials in Figure 3D).
Next, we analyzed how the spatiotemporal odor response pattern (t = 1-2 s) and the post-odor response pattern (t = 15-16 s, i.e. 5 s after odorant offset) developed over time.
To this end, we calculated the correlation to these time windows across different trials (Figure 3B, purple and blue frames in Figure 3A mark the correlation values plotted in 3B). We found that the ORN odor response pattern (purple trace) was stable during the odor stimulus, but immediately collapsed at odorant offset. On the other hand, the post-odor response pattern (blue trace) gradually evolved into a distinct and dissimilar pattern.
Similarly, PN dendrite response patterns were also correlated between two 1-butanol stimulations, but when compared to ORNs the correlation values were lower and decreased during the 10-s stimulation (Figures 3A,B). This shows that the odor response pattern of PN dendrites was both less stable during odorant stimulation and less reproducible from trial to trial. After odorant offset, the post-odor response pattern was not correlated to the odor response pattern (Figures 3A,B).
Across different odorants, we found that similarity was higher in ORNs compared to PNs (Figures 3C-E). This confirms previous observations of higher odorant-specificity in PNs (Olsen and Wilson, 2008;Silbering et al., 2008) as well as higher response variability in PNs (Jeanne and Wilson, 2015). However, for all tested odors and both ORNs and PNs, the correlations between the odor and post-odor response patterns were low (diagonal entries, Figure 3E). This indicates that no information about odor identity was maintained in the glomerular response pattern after odorant offset, and ORNs and PNs do not contain any sensory memory in their Ca 2+ response activity.
Do post-odor responses of ORN axons and PN dendrites in the antennal lobe depend on the stimulus duration? We compared post-odor responses after a 10-s-long stimulus with post-odor responses after shorter stimuli (0.2, 0.4, 1, 3, and 6 s) (Figure 4 and Supplementary Figure S2). In both ORN axons and PN dendrites, stimuli of different length activated the same glomeruli (Figure 4A), and accordingly odor response patterns were correlated (Figure 4B and Supplementary Figure  S2). After odorant offset, Ca 2+ responses of both ORNs and PNs rapidly changed for all stimulus durations. Additionally, post-odor response patterns for different stimulus durations were more similar to each other with increasing stimulus durations ( Figure 4B and Supplementary Figure S2).

Post-odor Response Patterns in KCs Are Similar to Odor Response Patterns
The Ca 2+ responses in PN dendrites consist of both pre-and post-synaptic activity (Ca 2+ influx through cholinergic receptors and Ca 2+ influx through voltage-gated Ca 2+ channels). In the cell bodies (somata), however, Ca 2+ influx is driven by cell depolarization only, and thus reflects a further processing step with respect to PN dendrites. However, PN somata cannot be attributed to identified glomeruli, and therefore, we analyzed patterns across non-identified "anonymous" units. Similarly, we analyzed "anonymous" responses in the mushroom bodies, recording KC dendrites and KC somata (Figures 5A,B). As before, we categorized the responding cells/ROIs again into on, off , and prolonged responding units (Figures 5B,C). In all three cell/compartment types, most units showed on responses. Off responses were rare in PN and KC somata, and almost nonexistent in the KC dendritic region. KC somata had a significantly larger proportion of prolonged responses than KC dendrites. Many of the observed prolonged responses, particularly in PN and KC somata, lasted even longer than the 20 s recording time after odorant offset ( Figure 5B).
To investigate whether the prolonged responses in PN somata, KC dendrites and KC somata maintained the odor-specific pattern after odorant offset, we performed the same correlation analyses as done for glomerular signals and quantified the timeresolved pattern similarity between two stimulations (trial 1 and trial 2) of the same odorant ( Figure 6A and Supplementary Figure S3).
Similar to the 1-butanol responses in PN dendrites (Figure 3A), odor response patterns and post-odor response patterns in their somata (PN somata) were reproducible and odor-specific ( Figure 6A and Supplementary Figure S3). However, odor response patterns were uncorrelated to post-odor response patterns (Figure 6A and Supplementary Figure S3). Odor response patterns (purple trace, Figure 6B) broke down within 3-4 s after odorant offset, while the post-odor response pattern (blue trace) developed after the end of the odorant stimulation and lasted for several seconds. This suggests that it is rather unlikely that Ca 2+ activity in PN somata harbors a sensory memory for specific odors.
In KC dendrites, the odor response patterns and post-odor response patterns were reproducible over stimulus repetitions ( Figure 6A). Unlike all upstream compartments, the post-odor response patterns were correlated to the odor response pattern ( Figure 6A). The correlation trace (Figure 6B, purple trace) showed a decline of pattern similarity within 3 s after odorant offset (correlation values from 0.5 to < 0.2), but did not reach baseline. This indicated that Ca 2+ activity in KC dendrites maintained a pattern correlation to the odor response pattern after odorant offset, which could contribute to a sensory memory.
Similar to the other analyzed cellular compartments, KC somata also showed reproducible odor and post-odor response patterns ( Figure 6A and Supplementary Figure S3). Here, we found an elevated similarity between post-odor and odor response patterns: After odorant offset, the similarity to the odor response patterns changed only gradually, maintaining a correlation to the odor response pattern (Figures 6A,B, purple trace). Furthermore, the post-odor response pattern already developed during the odorant stimulation ( Figure 6B, blue trace). This pattern remained correlated to the odor response pattern for at least 20 s after odorant offset (until the end of the recording time), i.e., for a behaviorally relevant time scale (Galili et al., 2011). Together, these results indicated that Ca 2+ activity in KC dendrites and even more in KC somata had an elevated similarity between odor and post-odor response pattern. This Ca 2+ activity could represent the neural substrate for a sensory memory in trace conditioning.

Post-odor Response Patterns Are More Reproducible and Odorant-Specific in KC Somata
How specific were the Ca 2+ responses for each odor? We analyzed the correlation matrices for all odor pairs for the odor response patterns across trials (t = 1-2 s, Figure 6C) and the post-odor response patterns across trials (t = 15-16 s, Figure 6D) (for the corresponding correlation matrices over time of the single odor repetitions see Supplementary Figure S3). The odor response patterns were reproducible across trials for all tested odorants ( Figure 6C and Supplementary Figure S3), as shown by the high values for repeated odorant stimulations in all three compartments. Compared to KC somata, the correlation between odor response patterns for repeated stimulus presentations was lower in PN somata (except 4-methylcyclohexanol, MCH). Postodor response patterns increased in their reproducibility from PN somata, over KC dendrites to KC somata (Figure 6D), indicating that the neural networks in the brain modify response patterns towards more stereotypic, odor-specific activity patterns.
If post-odor responses encode an odorant-specific sensory memory, then those post-odor response patterns should be similar to the immediate odor-responses. We therefore quantified the pattern similarity between odor responses and post-odor responses ( Figure 6E). There was no such correlation for PN somata. For KC dendrites, correlations were elevated for acetic acid (AceA) and propanoic acid (ProA) (both within and between odorants and stimulus repetitions). In KC somata, however, the correlations were odorant-specific and reproducible for most odorants ( Figure 6E). Thus, KC somata post-odor response patterns could encode an odorant-specific sensory memory. Their post-odor responses were most reproducible, most odorant-specific and most similar to the corresponding odor response patterns as compared to all other compartments studied here.

Prolonged Response Patterns in KC Somata Can Identify the Past Odors' Identity
In olfactory trace conditioning, insects learn the association between the odorant and the reinforcer during the postodor response (because this is when the reinforcer is given), but during test they respond to the actual odorant stimulus (Galili et al., 2011;Szyszka et al., 2011). Therefore, when the brain forms an associative memory between an already terminated odorant stimulus and a current punishment or reward, the post-odor activity pattern must allow the recognition of future presentations of the same odorant. Therefore, we asked whether the activity patterns in the measured cell compartments contained sufficient information to identify the odor response , and the average correct classification during the stimulus was calculated (Y -axis). Shaded bars denote chance classification rates obtained by performing the analysis 250 times on label permuted datasets. The upper bound of the shaded area denotes the 95% confidence interval of such chance classification. Only KC somata were above this chance classification rate for an extended period of time. We overlaid the behaviorally observed learning performance (Data from Galili et al., 2011, with 0 learning scores coinciding to the chance level of classification, right ordinate axis). See Materials and Methods for number of flies and glomeruli/somata/ROIs. pattern itself. We trained a SVM with the response pattern during odor presentation, 5 s after odor offset, or 15 s after odor offset, and then tested for correct stimulus identification over time ( Figure 7A). When trained with the odor response patterns, classification was strong for all compartments during the stimulus and decayed shortly after ( Figure 7A, left panel). When trained with a post-odor response pattern, most neuronal compartments' patterns were not suitable to recognize the odor response pattern (Figure 7A, middle panel). However, the SVM trained on the KC somata post-odor responses resulted in high classification also during the odor responses ( Figure 7A, middle and right panel).
Training the SVM with a sliding window and testing with the odor response pattern confirmed these results: after stimulus offset, odorant classification by ORNs, PNs, and also KC dendrites quickly dropped to or below chance level. However, odorant classification by KC somata remained high for at least 20 s (dark green line in Figure 7B). We compared the resulting timecourse of the odorant classification success to behavioral learning performance during trace conditioning, where the reinforcer (punishment) was given at different time points after odorant onset (X-axis in Figure 7B). Learning was quantified as the choice between the persistent odorant stimulus and a solvent after conditioning (right Y-axis in Figure 7B). We found that the SVM odorant classification performance closely reproduced the behavioral odorant recognition performance (Galili et al., 2011).

DISCUSSION
We measured odorant-evoked cytosolic Ca 2+ signals in multiple successive layers of the olfactory system in Drosophila: in ORN axons, PN dendrites, PN somata in the antennal lobes, and KC dendrites and KC somata in the mushroom body calyx. In all neurons and neuronal compartments, we recorded the Ca 2+ responses to a common set of odorants. Our analysis focuses on the evolution of the spatiotemporal response patterns and their odorant specificities during and after the odorant stimulus.

Transformation of Immediate Odorant Responses Along the Olfactory Pathway
We found that the stimulus-specificity and separability of the immediate odorant-evoked response patterns across neurons successively increased with each processing layer (Figures 3C,  6C). This is in accordance with previous reports that found lateral inhibition via GABAergic interneurons increased odorant-specificity of PNs (Olsen and Wilson, 2008;Silbering et al., 2008). The combination of a divergent projection of PNs to KCs and feedback inhibition of KCs leads to the formation of sparser and more odorant-specific KC ensemble responses, as compared to PN ensemble responses (Szyszka et al., 2005;Honegger et al., 2011;Lin et al., 2014). These findings are also in line with theoretical predictions for pattern based feature extraction systems, where such a divergence-convergence motif leads to increased pattern separability (Luo et al., 2010;Stevens, 2015). Note that the Ca 2+ signals in ORN axons and PN dendrites in the antennal lobe and in KC dendrites in the mushroom body originate from multiple, overlapping neurons. In the antennal lobe, all ORNs or PNs that innervate the same glomerulus have highly correlated response profiles (Kazama and Wilson, 2009). Therefore, the glomerular Ca 2+ signals in ORNs and PNs reflect the response of individual ORNs and PNs. However, the dendritic Ca 2+ signals in KCs reflect the summed activity of different KCs with different response properties. Indeed, the dendritic KC responses were less odorant-specific than the response of individual KC somata ( Figure 6C).
Post-odor responses also occurred in the mushroom body in the dendrites and somata of KCs (Figure 6). Compared to PN somata and KC dendrites, KC somata had the largest fraction of prolonged responses (Figures 5B,C). In contrast to PN somata, the KC somata prolonged responses were stimulusspecific ( Figure 6D).
To analyze whether these post-stimulus responses are informative about the odorant identity, we used a classifier-based decoding analysis to predict the identity of past odorant stimuli in a time resolved manner. With this approach, we found that only the Ca 2+ signals of the KC somata allowed predicting the previously presented odorant's identity on a behaviorally relevant time scale (Figure 7B). Thus, cytosolic Ca 2+ responses in KC somata contain all the information necessary to encode odorant identity after the stimulus offset. This shows that cytosolic Ca 2+ forms a potential substrate for short-term sensory odor memory.
Animals need sensory odor memories in many situations. Such a memory is required for odor plume-tracking when insects continue their odor source search after losing an attractive odor plume (Vickers, 2005;van Breugel and Dickinson, 2014;Saxena et al., 2017), or for the olfactory working memory in honey bees when they solve delayed-matching-to-sample tasks (Giurfa et al., 2001), or to learn associations between temporally separated stimuli, as in trace conditioning (Galili et al., 2011;Szyszka et al., 2011). Our data show that Ca 2+ levels in KC somata are a suitable substrate for such sensory odor memories. Our data do not exclude that Ca 2+ signals in other neurons, not measured here, or signals other than Ca 2+ may encode sensory odor memories (Galán et al., 2006;Dylla et al., 2013;Betkiewicz et al., 2017).

Mechanisms of Prolonged Ca 2+ Levels in Kenyon Cell Somata
Which cellular and molecular mechanisms could cause the prolonged Ca 2+ elevation? A current analysis of cockroach KCs by Demmer and Kloppenburg (2009) found Ca 2+ currents and Ca 2+ -dependent currents that are unusual for insect neurons: the same neurons had both low voltage-activated inward Ca 2+ currents and high voltage-activated Ca 2+ -dependent repolarizing currents, both of large amplitude. These low voltageactivated, inward Ca 2+ currents could lead to a nonlinear Ca 2+ increase in response to odorant-induced post-synaptic potentials, and the duration of the post-odor Ca 2+ response might reflect the time required to restore baseline Ca 2+ concentrations by Ca 2+ buffers and pumps. The high voltage-activated Ca 2+ -dependent repolarizing currents could mediate the typical fast adaptation of KC spike responses to odorants Turner et al., 2008;Demmer and Kloppenburg, 2009;Palmer et al., 2013;Saha et al., 2013;Gupta and Stopfer, 2014;Kropf and Rössler, 2018). Indeed, in a modeling study, Betkiewicz et al. (2017) predicted that odorant-induced prolonged Ca 2+ responses exist in KCs, and they suggested that the prolonged Ca 2+ -dependent repolarizing currents mediate adaptation and encode a sensory short-term memory for odorants in KCs.

The Role of Ca 2+ in Associative Learning During Classical Conditioning
Classical conditioning requires the detection of a predictive temporal relationship between a stimulus and a reinforcer (punishment or reward). In the standard Drosophila olfactory conditioning paradigm, the molecular component detecting the immediate odorant-reinforcer coincidence is thought to be the adenylyl cyclase rutabaga in KCs (Dudai et al., 1983;Levin et al., 1992): here, the odor information is encoded by the cytosolic Ca 2+ in KCs, and the reinforcer is encoded by the activation of dopamine receptors in the KC axons (Gervasi et al., 2010). The coincident increase in Ca 2+ and dopamine receptor activation induces a change in synaptic strength between KCs and mushroom body output neurons (McGuire, 2001;Schwaerzel et al., 2002;Tomchik and Davis, 2009;Gervasi et al., 2010;Séjourné et al., 2011;Aso et al., 2014;Cohn et al., 2015;Hige et al., 2015;Owald et al., 2015). However, this mechanism cannot explain trace conditioning.
Trace conditioning differs from the standard classical conditioning paradigm in that the odorant and reinforcer are separated by a temporal gap (Galili et al., 2011;Szyszka et al., 2011). Thus, a sensory odor memory is required to bridge the gap between odorant stimulus and reinforcer. The phasic nature of odorant-evoked KC spiking and KC Ca 2+ activity in the lobes is unsuitable to encode a sensory memory Turner et al., 2008;Saha et al., 2013;Gupta and Stopfer, 2014;Dylla et al., 2017). Standard conditioning and trace conditioning also differ in the underlying neuronal mechanisms: in vertebrates, different brain regions are involved in standard and trace conditioning (Solomon et al., 1986;Woodruff-Pak and Disterhoft, 2008), and different molecular constituents have been identified in Drosophila, where rutabaga is not necessary for trace conditioning (Shuai et al., 2011).
Could the prolonged and odorant-specific Ca 2+ signals in the KC somata serve as a sensory odor memory in trace conditioning? Our data demonstrate that Ca 2+ activity patterns across KC somata encode the necessary odorant information, and therefore could serve as a substrate for sensory odor memory during trace conditioning. Our experiments did not address the molecular nature of a possible coincidence detector. One possibility could be that a protein kinase C (Choi et al., 1991) or a non-rutabaga adenylyl cyclase (Adams et al., 2000) in the KC somata would act as coincidence detector for odorant-induced Ca 2+ signaling and reinforcer-induced dopamine signaling. Dopaminergic neurons (PPL2ab neurons) innervate the mushroom body calyx (Mao and Davis, 2009), and application of dopamine increases cAMP in KCs in the calyx (Tomchik and Davis, 2009). Paired application of acetylcholine (simulating the odor stimulus) and dopamine synergistically increases cAMP in KCs in the lobes, but not in the calyx (Tomchik and Davis, 2009). However, it has not been tested whether the pairing of a prolonged Ca 2+ signal with a delayed dopamine input (as is the case in trace conditioning) would induce a synergistic cAMP response. It is therefore conceivable that a non-rutabaga adenylyl cyclase (Adams et al., 2000) in the calyx might serve as a coincidence detector for the odorantinduced prolonged Ca 2+ signal and for the reinforcer-induced dopamine receptor activation in trace conditioning.

CONCLUSION
Our data suggest that, in addition to the molecular separation (Shuai et al., 2011), trace and standard conditioning in Drosophila might be spatially separated, reminiscent of the situation in vertebrates (Solomon et al., 1986;Woodruff-Pak and Disterhoft, 2008). This adds an interesting thought to how memories are organized in brains: seemingly equivalent memories (here, the association of an odorant with a reinforcer) might be localized in parallel and different compartments, reflecting the mode with which they were learned (trace or standard conditioning).
The molecular and spatial separation of trace and standard conditioning opens the possibility to further dissect the physiological and molecular processes underlying trace conditioning, including specific phenotypes such as Rac (Shuai et al., 2011), using pharmacological alterations during in vivo imaging, or optogenetical shunting of KC activity at different time points during a trace-conditioning paradigm. Such studies, readily performed in Drosophila, will lead to a more mechanistic understanding of a brain's capability to bridge temporal gaps between stimuli and form associations across them.

AUTHOR CONTRIBUTIONS
AL performed the experiments. AL, PS, and CG designed the experiments. GR and AL analyzed the data. JN and AH advised about the data analysis. AL, GR, PS, and CG wrote the paper. All the authors edited the paper.

FUNDING
This work was supported by the Bundesministerium für Bildung und Forschung (BMBF, Grant No. 01GQ0931 to AL, JN, CG, AH, and PS) and by the German Science Foundation (SPP 1392 to GR and CG).

ACKNOWLEDGMENTS
We thank Fabian Hersperger for his experimental support and Stefanie Neupert for programming the odorant stimulation controller.

SUPPLEMENTARY MATERIAL
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fncel. 2018.00128/full#supplementary-material FIGURE S1 | The olfactory stimulator produced odorant pulses, with steep odorant on-and off-sets. Traces show photoionization detector signals (PID, Model 200a, Aurora Scientific Inc.) during stimulation with the tracer substance ethyl acetate (undiluted, ionization potential 10.01 eV) for different pulse lengths (0.2, 0.4, 1, 3, 6 s in channel 1, and 10 s in channel 2; indicated by the bar below each graph, single measurements). For PID traces to other odorants applied with a similar olfactory stimulator, see . Figure 4: Post-odor response patterns of ORN axons and PN dendrites differ from odor response patterns. Corresponding graphs as in Figures 3A,B, for cross correlations between stimuli of increasing length (abscissa, from top to bottom), with a stimulus of 10 s length (ordinate, in each graph). Figure 6: Odor and post-odor responses evolve differently in different brain areas. Corresponding graphs as in Figure 6 for cross correlations between responses to repeated presentation of the same odorant, in PN somata, KC dendrites and KC somata [columns (A-C), respectively]. See methods for odorant abbreviations.