Smell and Aftersmell: Fast Calcium Imaging Dynamics of Honey Bee 1 Olfactory Coding 2

10 Odour processing exhibits multiple parallels between vertebrate and invertebrate 11 olfactory systems. Insects, in particular, have emerged as relevant models for olfactory studies 12 because of the tractability of their olfactory circuits. Here, we used fast calcium imaging to track 13 the activity of projection neurons in the honey bee antennal lobe (AL) during olfactory 14 stimulation at high temporal resolution. We observed a heterogeneity of response profiles and 15 an abundance of inhibitory activities, resulting in various response latencies and stimulus-16 specific post-odour neural signatures. Recorded calcium signals were fed to a mushroom body 17 (MB) model constructed implementing the fundamental features of connectivity between 18 olfactory projection neurons, Kenyon cells (KC), and MB output neurons (MBON). The model 19 accounts for the increase of odorant discrimination in the MB compared to the AL and reveals 20 the recruitment of two distinct KC populations that represent odorants and their aftersmell as 21 two separate but temporally coherent neural objects. Finally, we showed that the learning-22 induced modulation of KC-to-MBON synapses can explain both the variations in associative 23 learning scores across different conditioning protocols used in bees and the bees' response 24 latency. Thus, it provides a simple explanation of how the time contingency between the 25 stimulus and the reward can be encoded without the need for time tracking. This study 26 broadens our understanding of olfactory coding and learning in honey bees. It demonstrates 27 that a model based on simple MB connectivity rules and fed with real physiological data can 28 explain fundamental aspects of odour processing and associative learning. 29


Introduction
The logic of olfactory coding and learning has been extensively studied at the behavioural and neural levels in various insect models (Adam et al., 2022;Galizia, 2014;Giurfa, 2015;Jefferis et al., 2007).Among them, the honey bee Apis mellifera has played a pivotal role in our understanding of these processes due to its behavioural accessibility and tractability of its nervous system.In bees, olfactory perception and learning have been typically investigated using the proboscis extension reflex (PER) (Giurfa, 2007;Menzel, 1999) Our understanding of the neural processing subtending olfactory coding and learning relies on decades of neuroanatomical and neurophysiological studies performed in honey bees (Paoli and Galizia, 2021).Olfactory processing starts at the peripheral level, when volatile chemicals interact with the olfactory receptors expressed on the dendritic membrane of the olfactory sensory neurons (OSNs).The biochemical nature of odorant-receptor interactions allows for a certain molecule to bind to multiple receptors with different affinities, resulting in the activation of an OSNs sub-population with odorant-specific response intensities and latencies (Münch and Galizia, 2016).Olfactory sensory neurons innervate the first olfactory processing centre, the antennal lobe (AL), where the neurons expressing the same olfactory receptor converge onto one of ~160 glomeruli, the anatomical and functional units of the AL (Flanagan and Mercer, 1989).Thus, odorant detection results in the stimulus-specific activation of a subset of glomeruli, creating a stereotypical map of odour-induced glomerular responses (Galizia et al., 1999;Sachse et al., 1999).The signal processed in the AL is forwarded by ~800 output neurons -the projection neurons (PNs) -to higher-order brain centres: the mushroom bodies (MBs), central, paired structures dedicated to multisensory integration, memory storage and retrieval (Heisenberg, 2003;Stopfer, 2014), and the lateral horn of the protocerebrum, a diffuse bilateral structure involved in valence coding of odorants (Jeanne et al., 2018;Roussel et al., 2014;Strutz et al., 2014).The MB architecture is defined by the layout of its ~185 000 intrinsic neurons, the Kenyon cells (KCs).Each KC extends its dendritic arborisation within the MB input regions, termed the calyces, where it receives input from multiple PNs, and projects its axon to the MB output region, termed the pedunculus, where it bifurcates into the vertical (α and γ) and the medial (β) lobes (Mobbs, 1982;Strausfeld, 2002).Kenyon cells integrate, among others, the excitatory input of olfactory PNs and the inhibitory input of recurrent GABAergic feedback neurons (Ganeshina and Menzel, 2001;Grünewald, 1999;Rybak and Menzel, 1993;Zwaka et al., 2018).The latter plays a critical role in shaping KCs olfactory responsiveness and maintaining a sparse output over a wide range of odorants and concentrations (Papadopoulou et al., 2011;Stopfer, 2014).
Antennal lobe, mushroom body and lateral horn are innervated by the VUMmx1 neuron (Hammer, 1993), an unpaired octopaminergic neuron conveying appetitive reward-related information to the olfactory circuit and mediating reward-based olfactory memory formation.
The activity in the Kenyon cells is integrated by MB output neurons (MBONs).These neurons exhibit which show learning-related plasticity and provide valence-loaded information to pre- In this study, we significantly improved the temporal resolution of calcium imaging recordings of olfactory neurons in the bee brain by means of a resonant scanning multiphoton microscope.This allowed us to record calcium activity at a <10-ms resolution, preserving temporal information of the olfactory code and yielding a more realistic representation of odour trajectories in the AL.First, we observed that odour representation changes dynamically during and after an olfactory stimulation, resulting in specific odour and after-odour images.Then, we showed that most glomerular response profiles presented an inhibitory component, suggesting that glomerular activity is strongly shaped by local inhibition, with only a minor influence -if any -of local excitatory interneurons.
Calcium imaging analysis was combined with a modelling approach to investigate how Here, we show that an MB neural network model based solely on three simple connectivity rules abstracted from insect studies -sparse connectivity, feedback inhibition, and learning-induced synaptic modulation -is sufficient to explain fundamental MB processing features such as improved olfactory discrimination and associative learning.The model also predicts that the presence of post-stimulus inhibitory and excitatory activity leads to the recruitment of a second pool of KC, resulting in a distinct after-odour representation in the MB.
Additionally, the model's performance in appetitive learning and across-stimuli generalisation was coherent with empirical measurements obtained from behavioural protocols.Finally, using real physiological datasets as input, we showed that such a model is robust to noise and biological variability across multiple stimulus repetitions.
First, we assessed the stability of odorant representations across time and trials given the duration of each imaging session (30 min) and the multiple stimulus repetitions.For each individual, we calculated the Pearson's correlation coefficients across pairs of trials of the glomerular response vectors before, during and after olfactory stimulation.Figure 1E shows that across-trial correlation is elevated during odour arrival and still ~0.5 after stimulus termination.This shows that, in our experimental conditions, odour coding was stable throughout the entire imaging period.
All odorant response traces were pooled together (Fig. 1F) and clustered according to their direction with respect to the pre-stimulus baseline (i.e., excitatory or inhibitory) and duration (Fig. 2, see Materials and Methods), allowing for a statistical description of glomerular responses across individuals and odorants, and for the identification of classes of recurrent profiles (Fig. 2A,B).The most common profiles comprised inhibitory (group 1) and excitatory (group 4) responses lasting the entire olfactory stimulations and terminating after odour offset.In some cases, they could result in prolonged inhibition (group 2) or excitation (group 6) lasting after stimulus offset or could be followed by a post-stimulus excitatory (group 3) or inhibitory activity (group 5).Approximately 15% of responses consisted of short phasic excitation followed by prompt signal termination (group 7) or even by an inhibitory response (group 8).About 15% of all recorded glomerular traces showed no detectable odorant-induced activity (group 9).Of all traces, 48% were excitatory, 37% inhibitory, and 15% unresponsive (Fig. 2C,D).Moreover, 251 glomerular responses out of 546 (46%) displayed an inhibitory component either during or after olfactory stimulation.
The heterogeneity of PNs response profiles could be due to the diversified temporal pattern of the ONSs input (Kim et al., 2023) and to second-order processing mediated by lateral connections within the AL (Girardin et al., 2013;Krofczik et al., 2009).We assessed the contribution of these two components by measuring the latency of excitatory and inhibitory glomerular responses (i.e. the timepoint t at which a response profile exceeds the threshold of one standard deviation from the mean pre-stimulus activity).Figure 2E shows that excitatory responses (groups 4 to 8) all share a similar onset (313 ± 22 ms, n = 263), which is shorter than that of inhibitory profiles (groups 1 to 3;

Model input and output: Simulating olfactory coding in the mushroom body
For each recorded individual AL (n = 8), an MB simulation was generated and fed with the time series of the recorded glomerular activity obtained for that individual.Thus, for each odorant and individual, we simulated the time course of the firing of a virtual KC ensemble (Fig. 4A).The analysis of KC population turnover (Fig. 4B) showed that, before odorant onset, the population varies randomly between adjacent time points while it stabilises during olfactory stimulation, with a turnover rate of ~10%.The dynamics of KC recruitment across bees and odorants (Fig. 4C) indicates that two odorant-specific sets of cells are recruited during an olfactory stimulation: the first one at stimulus onset, the second one at odour termination (see Both populations are rather stable, with a turnover rate below the pre-stimulus baseline for up to 15 s after stimulus offset (turnover ratepre-stim = 0.61 ± 0.14; turnover rateON_0-5s = 0.17 ± 0.05; turnover rateOFF_5-10s = 0.25 ± 0.04; turnover rateOFF_10-15s = 0.33 ± 0.08).
Correlation analysis of stimulus-induced responses (Fig. 5D,E) showed a high correlation of neural activity across time points during olfactory stimulation and during the post-odour window both in the recorded PN population and in the simulated KCs.This indicates that the neural representation of an odorant is sufficiently stable to provide a neural image, which is coherent with itself during a 5-s stimulation window and during a ~10-s post-stimulation window.
Furthermore, the across-odorants correlation is rather high in the PN space (Pearson's correlation coefficient, r = 0.4), but proximal to zero in the KC space, supporting the notion that the MB network increases stimulus discriminability, not only during but also after odour offset.
To further corroborate this idea, we compared the temporal relative trajectories obtained for the three odorants in the measured PN space and in the modelled KC space (Fig. 4F).In both cases, response trajectories diverged at stimulus onset and returned to the centre of space after stimulus termination.However, the principal component analysis showed that odorants are better separated from each other and that they remained separated for a longer time in the KC space compared to the PN space.This indicates improved and prolonged stimulus discriminability in the MB with respect to the AL (Fig. S1).
We next investigated the discriminatory power of the system across a larger set of odour response profiles.For this, we pooled together all response profiles acquired during the calcium imaging analysis (Fig. 1F) and we artificially combined them to simulate 100 odorant response vectors, each comprising 30 different glomerular profiles (see Methods).Using these artificially combined odorant response profiles, we simulated the relative KC response time series and measured the correlation across all response vectors in the PNs (experimental data) and KCs spaces (modelled data) (Fig. 4G).Notably, while response vectors calculated from PN data were highly correlated (r = 0.71 ± 0.08), such a correlation significantly decreased in the modelled KC space (r = 0.52 ± 0.10; Student's t-test between r distributions of PN and The calcium signal dynamics and modelling analysis presented here (Fig. 4A-F) suggest that such an ISI-dependent effect is the consequence of the strong temporal dynamics of the odorant neural representation.As the neural representation of the CS changes in time, a reward system recruited at different time points will interact with a different population of CSrecruited KCs.Because learning results from the coincidental activation of CS and US elements (Pavlov, 1927), an individual should respond to the different time points of an olfactory stimulation with different strengths, depending on the US arrival time experienced during conditioning.To test this hypothesis, learning-related MBON plasticity was modelled for three different CS/US temporal contingencies reflecting three conditioning paradigms, wellstudied in honey bee (Giurfa and Sandoz, 2012) (Fig. 5A).Having a fixed CS stimulation interval from t = 0 to 5 s, and a learning window (US) of 3 s, we simulated MBON learning with three different ISIs.Backward pairing occurred when the US initiated at t = -2 s (i.e. 2 s before CS onset), early pairing when it started at t = 1 s (1 s after CS onset), and delay pairing when it started at t = 4 s (4 s after CS onset).First, we trained the model with the experimental PN response profile recorded for either 1-hexanol or peppermint oil (see Fig. 1).One odorant was used as CS and the other as a novel odorant to test the specificity of the learning obtained under these experimental conditions.After learning, we modelled the MBON action potential (AP) probability upon the presentation of the CS (e.g.1-hexanol) and the novel odorant (e.g.peppermint) and repeated this operation for the response profiles of all individuals.To account for the biological variability in the neural activities elicited by repetitions of the same stimulus, learning was assessed against the KC response profiles resulting from five CS and five novel odorant presentations, which were not included in the training dataset.As expected, we found that a trained MBON showed a larger decrease in firing probability when presented with the CS than with the unfamiliar stimulus, demonstrating learning (Fig. 5B and C).Interestingly, the model predicts weak learning in the case of backward pairing -that is when US and CS are partially overlapping, but the US begins before the CS -whereas both early and delay conditioning produced a CS-specific down-regulation of MBON firing probability.The model also predicts that high synaptic plasticity sensitivity (low spt values) facilitates non-specific learning, resulting in a strong generalisation effect.Conversely, MBON plasticity under high spt values requires a robust and stable KC activation, a situation that is not compatible with a learning interval that is only partially overlapping with the stimulation window (note that measurements are provided at 20 Hz, and in this context, a spt = 40 requires a synapse to be activated during two whole seconds -although not continuously -to be switched off) (Fig. S2).A simulation of the action potential probability of an MBON after olfactory conditioning revealed that differentiation between the CS and a novel odorant was observable whenever the CS anticipated the US (ISIs of +1s and +4s) but less successful in the backward protocol (IS of -2s).Generalisation errors were also observed in a similar way as in behavioural experiments.

Stimulus representation is dynamic (see
Notably, in this study, we adopted a backward contingency with a 1-second overlap between the US and the CS (Fig. 5A).This differs from the more traditional configuration, where the CS is delivered after US termination, that is, without any temporal overlap.In this case, the model would predict no learning at all, as observed in honey bee behavioural experiments Felsenberg et al., 2014;Hellstern et al., 1998).Hence, we tested the model's prediction in the form of backward conditioning where US and CS are partially overlapping.In this case, the model predicts a weak appetitive association (see Fig. 5B,C,E).This result is coherent with a previous behavioural report, where bees subject to a similar protocol showed a weak response to the conditioned odorant (Felsenberg et al., 2014).In the early and delay configurations, a CS-specific associative memory was formed, both in the model (Fig. 5C) and in the behavioural tests (Fig. 5E).However, the model predicts a difference in response latency and latency variability depending on relative CS/US arrival (ISI of +1 or +4 s): a learning window placed closer to stimulus onset (early protocol) resulted in early and temporally precise MBON responses, whereas a delayed learning window produced a similar learning score, but with higher response latencies and temporal uncertainty Experimental results based on the same odorants that were fed to the model confirmed the model's predictions.They showed that the two forward CS/US contingencies (i.e.early and delay protocols, with the CS anticipating the US) yielded comparable results in terms of learning score and CS-specificity while differing in terms of CS latency and precision (Fig. 5C-F).These observations support the idea that associative learning occurs upon coincident detection of the CS and US neural elements.In fact, because the CS neural representation changes dynamically during the olfactory stimulation, a positive association with the initial odour signature will trigger an earlier and temporally precise conditioned response.
Conversely, the association with the later part of the olfactory signature will elicit a delayed response with greater temporal variability.

Encoding time without encoding time
It has been reported that mammals can encode the CS/US interval time, allowing them to modulate the latency of a conditioned response (Kirkpatrick and Balsam, 2016).However, it is still unclear whether also smaller, less complex insect brains are capable of time tracking.is learnt at the time of the US presentation is the KC configuration of the CS at this particular time.In other words, the MBON learns which subpopulation of KC is active during reward arrival, rather than the expected reward time.Interestingly, because the KC representation of the CS does not evolve linearly with time (Fig. 4A-C), the CS/US time contingency is not encoded perfectly.Remarkably, the model captures the experimental measurements of the proboscis response latency in bees, which indeed tend to respond earlier than the expected reward (Fig. 5D,F).
As often in insect research, heuristics can provide simple solutions to apparently complex cognitive problems.In this case, a mechanism based on the temporal dynamics of the conditioned stimulus, explains the bees' systematic errors at encoding proper CS/US contingency.Whether, in addition to this process, other mechanisms enable bees to track time contingencies remains to be seen.local interneurons observed in the AL of bees, which seems to differ from that occurring in flies, thus highlighting the fact that convergence, but also differences, can characterise the functional architectures of olfactory systems in different insect species.

Methods Experimental Model
Experiments were performed on honey bees Apis mellifera reared in outdoor hives at the experimental apiary of the Research Centre on Animal Cognition (CNRS, Toulouse, France) situated in the campus of the University Paul Sabatier.In all cases, honey bee foragers (>3-week old) were used.No institutional permission is required for experimental research on honeybees

Projection neurons labelling
Honey bee foragers were collected the day before the experiment at an artificial feeder, to which they were previously trained, and projection neurons were labelled for calcium imaging analysis as previously described (Paoli et al., 2017;Sachse and Galizia, 2002).Shortly, bees were fixed in a 3D-printed holder with soft dental wax, antennae were blocked frontally with a drop of eicosane (Sigma Aldrich, CAS: 112-95-8).A small window was opened in the head cuticle, and glands and tracheas were displaced to expose the injection site.The tip of a borosilicate glass needle coated with Fura-2-dextran (Thermo Fisher Scientific Inc.) was inserted between the medial and lateral mushroom body calyces, where medial (m-ACTs) and lateral antenno-cerebral tracts (l-ACTs) cross.After dye injection, the head capsule was closed to prevent brain desiccation, and bees were fed ad libitum with a 50% sucrose/water solution.On the following day, antennal lobes were exposed to allow optical access, and the brain was covered in transparent two-component silicon (Kwik-Sil, WPI).Although the injection procedure may result in variable levels of PN labelling, the reproducibility of response amplitudes and the lack of labelling bias for specific glomeruli indicate that the loading procedure is reproducible.

Calcium imaging analysis and signal processing
Undiluted solutions of 1-hexanol, 1-heptanol and peppermint oil (Sigma-Aldrich) were delivered to the bees using an Arduino Uno ® -controlled automated olfactometer (Bestea et al., 2022; Raiser et al., 2017).Odorants were alternated and presented 20 times on a 5/25 second ON/OFF configuration.Calcium imaging recordings were conducted with a straight Leica SP8 scanning microscope (Leica Microsystems, Germany) equipped with a SpectraPhysics InSight X3 multiphoton laser tuned at 780 nm for Fura-2 excitation.All images were acquired with a water immersion 16x objective (Leica HC FLUOTAR 16x/0.6IMM CORR, Leica Microsystems, Germany), at 64x64 pixel resolution and ~127 Hz.
Calcium imaging data were analysed with custom-made MATLAB (MathWorks Inc.) scripts.The baseline signal, calculated as the mean fluorescence during the one second before stimulus onset, was used to calculate baseline-subtracted and normalised stimulusinduced glomerular activity (∆F/F).The normalised activity was multiplied by -1 to display excitatory/inhibitory responses as positive/negative relative changes (-∆F/F).Time series of glomerular activity were averaged across stimulus repetitions for response profile clustering analysis.Glomeruli were hand-selected based on a morphological and functional map.
Glomerular response profiles were calculated as the mean intensity of an area of 5-by-5 pixels around the centre of the glomerulus across time.Stimulus representation stability (Fig. 1E) was assessed with a Person's correlation analysis between vectors of the mean glomerular response of different trials.Glomerular response vectors' stability was calculated before, during and after olfactory stimulation.

Mushroom Body Modelling
We used Matlab R2018a to investigate how the recorded PN activity would affect KC responses of a MB network model.

Projection neurons to Kenyon cells connectivity
Real projection neurons' activity acquired via calcium imaging analysis provided the input to a simple model of the mushroom body neural circuits.For each experimentally measured honey bee, approx.26 projection neurons (this number depends on the number of detected glomeruli in the relative calcium imaging experiment) converge onto 1000 modelled KCs.
Connections between PNs and KCs were modelled as pseudo-random, with each KC receiving connections from approx.8 PNs (30% of the input PN population), as observed in bees For every time step t, the integrated input to each KCs (KCepsp) was calculated as the matrix product of the input neuron activity vector (PNactivity) and the logical connectivity matrix.This integrated input of each KC is analogous to its excitatory post-synaptic potential (EPSP).

KCepsp(t) = PNactivity(t) × W_PN_to_KC.
Winner-takes-it-all inhibitory feedback element Whether a KC will fire an action potential depends not only on the sum of its input signals but also on the inhibitory activity of the MB feedback neurons (Rybak and Menzel, 1993;Zwaka et al., 2018).These neurons act on the whole KC population, providing that the KCs receiving the highest summed input will produce an action potential (AP).To account for such effect, and based on known MB neurophysiology (Honegger et al., 2011;Peng and Chittka, 2017;Turner et al., 2008), we imposed that only the 10% most active KCepsp (i.e.receiving the strongest summed input) will fire an action potential.
If KC(i)epsp(t) > 10 th percentil of KCepsp(t), then KC(i)activity(t) = 1; else KC(i)activity(t) = 0 This outputs a binary vector of firing pattern across the whole KC population (KCactivity), where ones correspond to action potential and zeroes to silent KC.This operation is repeated for all epochs to obtain the time course of the activity of the KC population before, during and after odour arrival.To ensure the robustness of the model and account for the variability of the initial parameters, the MB simulation was run 10 times for each experimental AL input dataset.
The results were then averaged over the 10 MB network simulations.

KC-to-MBON model and learning-induced plasticity
Evidence in Drosophila and in the honey bee suggests that the default state of an appetitive MBON -i.e. an MBON departing from an MB lobe innervated by a dopaminergic neuron conveying reward information -is to be broadly responsive to neutral stimuli (Okada et al., 2007;Owald et al., 2015).Under the assumption that similar principles are highly conserved among insects, we modelled an MBON receiving connections from all KCs and with all connections having an initial synaptic weight = 1 to reflect a generalised odour tuning.This provides a broadly-tuned MBON that responds equally to any pattern of KC activity.Upon learning, MBON synapses that are repeatedly recruited are switched off (= 0).As a consequence, a trained MBON retains a high AP probability in the presence of an unfamiliar stimulus, but a lower one in response to the conditioned odorant (or to an odorant that is represented by a similar KC population).Two parameters guide learning: a synaptic plasticity threshold (spt) and a learning window.The variable spt defines the number of firing events needed for a given KC to induce its output synaptic depression (i.e.KC(i)_to_MBON synaptic weight switch from 1 to 0).For the learning tests, a spt value of 15 was adopted.This value falls in the middle of the range tested in our simulations (from 1 to 40).It is stringent enough to hinder plastic modulation of rarely recruited KCs, while still requiring repetitive activation (i.e. a synapse must be active at least 750 ms during a 3-s learning window) before allowing learning-induced plasticity.
The learning window provides a temporal restriction of such a learning rule, analogous to the reward/US delivery window in an appetitive conditioning protocol.Different CS/US contingencies were tested.Inter-stimulus intervals (ISI) of -2, 1 and 4 were used, where such values indicate the onset of a 3-s US window with respect to the onset of a 5-s CS window.
Throughout the manuscript, we referred to the three learning protocols as backward (ISI = -2), early (ISI = 1), and delay (ISI = 4).The training was modelled using the mean glomerular responses to the first 5 stimulations of 1-hexanol or peppermint oil; memory retention was tested against the next 5 stimulations of both odorants (one acting as the conditioned stimulus, the other one as the novel odorant).

Odorant response profiles simulation
The response profile database (Fig. 1F) was repeatedly subsampled to generate 100 odorant response maps, each comprising 30 glomerular responses.The proportion of response types (inhibitory, excitatory, non-responsive, etc.) was maintained constant and equal to the mean proportions across the entire population to generate a batch of similar glomerular response maps (representing perceptually similar odorants (Guerrieri et al., 2005)) (Fig. 4G).Projection neurons (or KC) response vectors used for the Pearson correlation analysis are constituted by the average response for each PN (or KC) during olfactory stimulation (t = 1 to 4 s).

Proboscis extension response latency
Bees were collected in the afternoon at the institute's outdoor beehives, kept in customprinted plastic cages in groups of 16 individuals, and provided with 240 microliters of 50% sugar/water solution (an average of 15 l for each individual).On the next morning, they were harnessed in plastic tubes, blocked in place with tape, and fed 3 l of sugar solution.Three hours later, they were exposed to an absolute olfactory conditioning protocol (Villar et al., 2020).During conditioning, bees were placed in front of the odour delivery device (the same one used for the calcium imaging analysis) and exposed to clean air for 15 s (familiarisation phase), to the odorant for 5 s, and to clean air again for another 20 s.Three different intervals were selected for delivering the sugar reward: either from 2 s before odour onset to 1 s after (backward conditioning), from 1 to 4 s from odour onset (early conditioning) or from 4 to 7 s from odour onset (delay conditioning).Bees were exposed to four rewarded trials, with a 10minute inter-trial-interval. Memory was tested 1 hour after the last conditioning trial.During the memory test, videos were acquired with a commercial video camera at 30 fps.A LED connected to the odour delivery device and in synch with the olfactory stimulation was used as a marker for the starting timepoint for the measurement of proboscis extension latency.
Peppermint oil and 1-hexanol were used as olfactory stimuli.They were used non-diluted and their role as conditioned stimulus (CS) or novel odorant (NOd) was balanced between bees.
During the test, half of the bees were exposed first to the CS and then to the NOd, while the remaining half was tested in the opposite order.A 50% sugar/water solution (w/w) was used as unconditioned stimulus (US).

Statistical Analysis
For MBON learning simulations, a total of 16 MB networks were generated and trained with the PN response profiles to 1-hexanol (n = 8) or peppermint oil (n = 8).After training, each MBON was tested against the response profiles of the CS and of the novel odorant to assess AP firing probability to the learned and the unfamiliar stimulus.For each simulated MBON, 5 memory tests against the 5 repetitions of the CS and of the NOd were performed.To quantify stimulus response to the conditioned/novel stimulus, the mean firing probability during the time-window of stimulus arrival was calculated.The difference between the responses to the CS and to the NOd was tested with a Kruskal-Wallis statistical test (Fig. 5C).The latency of MBON response to the CS was calculated as the latency to reach the 90% of minimal firing probability upon stimulation.The difference in latency between protocols was tested with a Kruskal-Wallis test; the difference in latency variance was assessed with a Barlett's test (Fig.

5D).
Proboscis extension response was used to assess memory retention 1 h after the last absolute conditioning trial.Conditioned stimulus specificity was tested with a McNemar test.
The latency of CS-specific PER was measured as the first frame after stimulus onset, where the proboscis trespasses the imaginary line between the open mandibles (Fig. 4E).
Differences between response latencies after early and delay conditioning were assessed with a Kruskal-Wallis test as well as with Barlett's test for difference in variance among datasets (Fig. 4F).(B) For each timepoint t of a calcium imaging recording, the measured activity level for each glomerulus (represented by a PN in the model) is projected onto all its synaptic connections with the KC population.Each KC, i.e. each column in the scheme, integrates all excitatory postsynaptic potentials.Finally, recurrent inhibitory feedback is simulated by imposing that only the 10% most active KCs will generate an action potential (AP) (red cells), while all others remain silent (grey cells).(C) A single appetitive MBON receiving input from all KCs was modelled.Based on a synaptic plasticity threshold parameter, all synapses from KCs to MBON that are activated with a

Figures
, a protocol that relies on pairing a neutral olfactory stimulus (the conditioned stimulus or CS) with a positive reinforcement of sugar solution (the unconditioned stimulus or US) (Bitterman et al., 1983; Giurfa and Sandoz, 2012; Takeda, 1961).In the classical version of PER conditioning, the CS precedes and partially overlaps US presentation.This results in high levels of specific memory for the conditioned odorant.Conversely, when the reward precedes the stimulus (backward conditioning), no positive association can be established (Felsenberg et al., 2014; Hellstern et al., 1998).It also enables the study of more sophisticated cognitive processes such as trace learning (Paoli et al., 2023a; Szyszka et al., 2011) and patterning discrimination (Deisig et al., 2001; Devaud et al., 2015).In the former, CS and US are not overlapping but separated by a stimulus-free temporal gap, whereas in the latter, bees are trained to respond in opposite ways to a two-odorant mixture compared to its individual components (Deisig et al., 2001; Devaud et al., 2015).Overall, the olfactory conditioning of PER provides a robust read-out for investigating the dynamics of olfactory memory formation (Giurfa and Sandoz, 2012; Villar et al., 2020) and olfactory perception (Guerrieri et al., 2005).
motor areas (Aso et al., 2014a; Okada et al., 2007; Schmalz et al., 2022; Strube-Bloss and Rössler, 2018).The neural representation of odorants in the AL has been extensively described by means of functional calcium imaging (Paoli and Haase, 2018).In vivo imaging allowed observing that each stimulus is represented with a specific pattern of excitatory and inhibitory responses across the glomeruli of the AL, (Galizia et al., 1999; Sachse and Galizia, 2002; Sachse et al., 1999), and that perceptually similar odorants elicit similar glomerular response patterns (Guerrieri et al., 2005).Calcium imaging of AL activity has been typically conducted at low temporal resolutions (~100-200 ms) (Locatelli et al., 2016; Mertes et al., 2021; Nouvian et al., 2018; Sachse and Galizia, 2003).This is partially justified by the slow dynamics of fluorescent calcium sensors, which resulted -in general -in the compression of olfactory representation into a spatial vector of glomerular response intensity, with a concomitant loss of information on neural response dynamics.While calcium signal decay is relatively slow (>100 ms), its onset is fast (<10 ms) (Helassa et al., 2015; Moreaux and Laurent, 2007), and provides the possibility for investigating parameters such as glomerular response latency (Junek et al., 2010; Paoli et al., 2018), signal frequency components (Paoli et al., 2016) and interglomerular information transfer (Chen et al., 2023; Paoli et al., 2023b).
odorant representation evolves from the AL to the MB.Electrophysiology experiments have provided insight on the phasic and sparse activity of Kenyon cells as well as into its oscillatory nature (Laurent and Davidowitz, 1994; Perez-Orive et al., 2002; Stopfer, 2014).However, the unavailability of an imaging method allowing the visualisation of olfactory coding in a KC ensemble prevented us from further understanding how odorants are represented within the MB, e.g.how the neural representation of an odorant is transformed from the AL to the MB or to what extent the similarity among odorant representation is maintained in the KC space.One way to address these questions is by using neural network models constructed via the abstraction of common features of MBs across insect species (e.g.fruit fly, locust, honey bee) to reproduce cognitive tasks such as stimulus discrimination or learning.Models can be built by simulating multiple neurons interacting with each other according to physiological rules such as the dynamics of action potentials or Hebbian synaptic plasticity (Eschbach et al., 2020; Finelli et al., 2008; Gkanias et al., 2022; Huerta et al., 2004; Smith et al., 2008).Other models simulate the MB neural network by considering the statistics of connectivity within the neuropil, e.g. the ratio of PNs to KCs, the average number of synaptic connections, the neuronal firing rate (Ardin et al., 2016; Buehlmann et al., 2020; Le Moël et al., 2019; Peng and Chittka, 2017; Springer and Nawrot, 2021; Wystrach, 2023).Here, we followed the second approach and built a simplified but realistic neural network model of the MB based on the neuroanatomical and functional properties of the insect's olfactory circuit.The model comprises three layers of neurons: (1) a MB input layer to provide the model experimentally acquired time-series of PN activity; (2) a MB intrinsic layer, where the input signal is distributed to a population of modelled KCs based on neuroanatomical and physiological data; (3) a MB output layer, where one appetitive MB output neuron (MBON) receives input from the KC layer and can be subject to learning-induced plasticity.Generally, the input for this type of model is simulated based on the spatial and temporal statistics of the odour-induced glomerular activity (Eschbach et al., 2020; Finelli et al., 2008; Gkanias et al., 2022; Huerta et al., 2004; Le Moël and Wystrach, 2020; Peng and Chittka, 2017; Smith et al., 2008; Springer and Nawrot, 2021).In this case, we fed the model with the time series of the PN responses recorded via in vivo calcium imaging analysis from multiple individuals exposed to three different odorants at the MB working frequency of 20 Hz (Cassenaer and Laurent, 2007; Laurent and Naraghi, 1994).
351 ± 47 ms, n = 201) (Kruskal-Wallis test, p<0.05,Tukey-Kramer multiple comparison correction).Moreover, the latency of short excitatory responses' termination (groups 7 and 8; 346 ± 47 ms, n = 60) is coherent with the onset of inhibitory profiles.These findings indicate that odour representation in the AL is shaped initially by the excitatory input delivered by the OSNs and reshaped -approx.40 ms later -by local inhibition.Such an olfactory tuning is different from what is proposed in Drosophila, where both excitatory and inhibitory local neurons contribute to moulding the neural correlate of an olfactory input (Chou et al., 2010; Olsen et al., 2007).A mushroom body neural network model: Key principles We constructed a simple but realistic MB neural network model based on the known connectivity of this structure in the insect brain.The model architecture relies on three main principles (Fig. 3): First, in each brain hemisphere, ~800 AL projection neurons (PNs) diverge onto ~185,000 KCs (Strausfeld, 2002).Neuroanatomical studies in bees (Szyszka et al., 2005) and flies (Ashok Litwin-Kumar et al., 2017; Caron et al., 2013) suggest that each KC is randomly innervated by approx.7-10 PNs.Thus, we generated a MB network where ~25 PNs (i.e., the average number of glomeruli imaged during a calcium imaging experiment) diverge onto 1000 KCs, with each KC being innervated by ~8 PNs.Second, recurrent inhibitory neurons such as the A3 feedback neurons in the bee (Mobbs, 1982; Rybak and Menzel, 1993; Szyszka et al., 2005; Zwaka et al., 2018), the APL neuron in fly (Lin et al., 2014), and the giant GABAergic neuron in the locust (Papadopoulou et al., 2011), modulate MB KCs firing rate so that less than 20% of KCs are active upon olfactory stimulation, and only ~5% are stimulus-specific (Honegger et al., 2011; Turner et al., 2008).Hence, we enriched the model with a winner-takes-it-all feedback inhibition mechanism by forcing that only 10% of KCs receiving the largest summed input generate an action potential.Third, as in bees and locusts, MB have a 20-Hz oscillatory cycle (Cassenaer and Laurent, 2007; Laurent and Naraghi, 1994; Popov and Szyszka, 2020), we fed the model with experimentally recorded calcium signals resampled at a 20 Hz-sampling frequency.This approach allowed testing if a simple neural architecture -so far challenged with simulated datasets (Ardin et al., 2016; Buehlmann et al., 2020; Peng and Chittka, 2017) -could be used for understanding how real odour activity recorded at the level of PNs is transformed in the MB, and if the rules governing the model were sufficient to support appetitive olfactory conditioning.
KC data: p ≈ 0, n = 4774), indicating that the signal transformation operated by the proposed MB neural architecture induces a strong decorrelation among odour signatures.Overall, the proposed MB network can process experimentally acquired input data and produce physiologically plausible KC response patterns coherent with in vivo KC activity measurements (Lüdke et al., 2018).Moreover, it accounts for the expansion of the coding space from the AL to the MB (Lin et al., 2014; Papadopoulou et al., 2011; Stopfer, 2014), enhancing inter-stimuli decorrelation and confirming previous theoretical works based on artificial datasets (Olshausen and Field, 2004; Peng and Chittka, 2017).Importantly, these emergent properties of the model architecture are based on the known neuroanatomy and physiology of the insect brain and have been obtained without any optimisation aiming at reproducing specific features of olfactory coding.The neural network model predicts appetitive behaviour Appetitive classical conditioning relies on the coincidental activation of the neural elements representing the conditioned stimulus (CS) and the unconditioned stimulus (US).The latter is mediated by various neuromodulators, such as specific dopaminergic, as shown in the fly (Burke et al., 2012; Liu et al., 2012), or octopaminergic neurons, as demonstrated in the honey bee (Hammer, 1993).The coincidental activation of CS and US neural elements induces plastic modulations in the MB output neurons (MBONs) (Hige et al., 2015; Okada et al., 2007; Owald et al., 2015).Here, we assessed if the proposed MB model could reproduce empirical measurements of appetitive learning by introducing a time window, during which the weights of the recruited KC-to-MBON synapses could be downregulated from 1 to 0 (Ardin et al., 2016).This rule reflects MBON learning-induced plasticity, according to which they display a broadly tuned response, which is reduced in presence of a learned stimulus (Amin and Lin, 2019; Aso et al., 2014a; Aso et al., 2014b; Cognigni et al., 2018; Cohn et al., 2015; Hige et al., 2015; Lyutova et al., 2019; Okada et al., 2007; Owald et al., 2015).To implement such a rule, we introduce a synaptic plasticity threshold (spt) parameter, which defines the number of firing events of a given KC upon which its synaptic output weight is reduced from 1 to 0. In other words, it determines how active a synapse should be during the learning window before being switched off.It is well known that the inter-stimulus interval (ISI) -i.e. the time elapsed from CS to US onsets -can influence the efficiency of Pavlovian learning (Domjan, 2015; Holland, 1980).

Fig. 1
Fig. 2A, groups 1-3,5,7,8), underlining the dominant role of inhibitory AL interneurons in Previous studies have shown that the duration of the inter-stimulus interval experienced during learning can influence the conditioned response's latency in bumble bees (Boisvert and Sherry, 2006) and honey bees (Szyszka et al., 2011).Both studies indicate that bees trained with a longer ISI exhibited a larger latency in their response to the conditioned stimulus.This provided clear evidence that CS-related information can modulate response timing.However, the appetitive response largely anticipated the expected reward time, and a clear interpretation of time tracking in insect brains is still missing Our model shows that this time contingency is naturally encoded, simply because the olfactory representation in the KCs evolves through time in a deterministic fashion.Thus, what

(
Szyszka et al., 2005) and flies (Ashok Litwin-Kumar et al., 2017; Caron et al., 2013).PNto-KC connections were assumed to be binary: either 0 (not connected), or 1 (connected), and the connectivity scheme is stable in a given MB model and not subject to plastic changes.This was implemented as a logic connectivity matrix (W_PN_to_KC) of size number_of_PNs × number_of_KCs.Experimental PNs activity (PNactivity) was resampled at 20 Hz for modelling analysis (50 ms temporal resolution) to match the 20-Hz oscillatory cycle detected in the MB (Cassenaer and Laurent, 2007; Laurent and Naraghi, 1994; Popov and Szyszka, 2020).

Figure 1 .
Figure 1.Projection Neurons Calcium Imaging Analysis (A) For each AL, the response maps to all three odorants were overlaid to highlight glomerular structures.Glomeruli were hand-labelled for time-series extraction.V, ventral; L, lateral.(B) Exemplary AL odorant responses during olfactory stimulation with three odorants.Colorbar indicates the relative change of activity during olfactory stimulation with respect to the pre-stimulus baseline.Circular areas indicate identified glomeruli according to (A) (C,D) Temporal profiles of the glomerular regions identified in (A) for the three odorants.Each line in (C), labelled from 1 to 28, refers to the glomerular ID in (A).Temporal profiles represent the average activity of 20 stimulations.Dashed lines in (C,D) limit the begin and the end of the olfactory stimulation.(E) Matrix of Pearson's correlations between pairs of glomerular response vectors across all repetitions.The upper right part of the matrix shows correlation scores among glomerular activity during olfactory stimulation (t = 1 to 4 s after odorant onset) across trials; the bottom left part shows correlation scores among glomerular activity before stimulation (t = -1 to 0 s) across trials.On the right, the mean (± s.e.m.) across-trial correlation before, during and after (t =6 to 9s) stimulation is shown.(F) All glomerular responses from 8 ALs to 3 odorants were pooled together to provide an overview of the complexity of response profiles (n = 546).

Figure 2 .
Figure 2. Clustering of Projection Neurons' Response Profiles (A) All glomerular responses were clustered with supervision according to their activity (excitatory, inhibitory, non-responsive) during and after stimulus arrival.(B) Mean ± s.e.m. of all curves of the relative groups in (A).The green patch indicates the stimulus delivery interval.(C) Average curves of all response groups from (B) are superimposed.The grey patch indicates the stimulus delivery interval.(D) Relative amount of all response categories.(E) Latency of glomerular responses for excitatory and inhibitory profiles (groups 1 to 8).For groups 7 and 8, the latency of short excitatory response's termination was calculated.Letters refer to significative groups after Kruskal-Wallis statistical test and Tukey-Kramer correction.On the left side, exemplary traces of phasic and tonic excitatory responses and an inhibitory response are shown.The orange line indicates the latency of response onset or termination.
Figure 3.A Simple Neural Network Model for Olfactory Coding and Learning.(A)The synaptic connectivity between AL PNs and KCs is built as a pseudorandom logic matrix where each column represents KC dendritic arborisation and each row PN axon terminals.Each KC receives synaptic contacthere represented as a white circlefrom 30% of the PN population.(B) For each timepoint t of a calcium imaging recording, the measured activity level for each glomerulus (represented by a PN in the model) is projected onto all its synaptic connections with the KC population.Each KC, i.e. each column in the scheme, integrates all excitatory postsynaptic potentials.Finally, recurrent inhibitory feedback is simulated by imposing that only the 10% most active KCs will generate an action potential (AP) (red cells), while all others remain silent (grey cells).(C) A single appetitive MBON receiving input from all KCs was modelled.Based on a synaptic plasticity threshold parameter, all synapses from KCs to MBON that are activated with a

Figure 4 .
Figure 4. Olfactory Representation in Modelled Kenyon Cells (A) The first row displays the recorded PN responses of one honey bee to three odorants.The second row shows the transformation of the PN activity operated by the model to simulate the activity of 1000 KCs.In each plot, KCs are ordered for response strength during the onset and offset to better visualise activity clusters.(B) Turnover rate of recruited KCs across adjacent time points.Light blue traces refer to individual bees (mean of 10 MB simulations).Thick, dark blue traces indicate average curves across bees.(C) Cumulative recruitment of KC shows two main recruitment events at stimulus onset and offset.To observe odorant-related KC recruitment dynamics, the cumulative sum was initiated at stimulus onset.Light blue traces refer to individual bees (mean of 10 MB simulations).Thick, dark blue traces indicate average curves across bees.(D,E) Correlation matrix among time points of measured PN activity (C) and simulated KC activity (E).Mean responses to the three odorants were concatenated to allow observing within and across odorant correlations.(F) PCA of odorants' trajectories in the measured PN space (top) and in the modelled KC space (bottom) for four exemplary bees.Trajectories comprise a 10-s interval, ranging from odour onset (t = 0 s, light) to 5 s after offset (t = 10 s, dark).(G) The raster plot shows a set of 100 artificially combined glomerular responses (top), each simulating the neural representation of 100 similar odorants in the PN space.A second raster plot (middle) shows the representation of the same odorants in the modelled KC space.Note that