Inhibitory muscarinic acetylcholine receptors enhance aversive olfactory learning in adult Drosophila

Olfactory associative learning in Drosophila is mediated by synaptic plasticity between the Kenyon cells of the mushroom body and their output neurons. Both Kenyon cells and their inputs from projection neurons are cholinergic, yet little is known about the physiological function of muscarinic acetylcholine receptors in learning in adult flies. Here, we show that aversive olfactory learning in adult flies requires type A muscarinic acetylcholine receptors (mAChR-A), particularly in the gamma subtype of Kenyon cells. mAChR-A inhibits odor responses and is localized in Kenyon cell dendrites. Moreover, mAChR-A knockdown impairs the learning-associated depression of odor responses in a mushroom body output neuron. Our results suggest that mAChR-A function in Kenyon cell dendrites is required for synaptic plasticity between Kenyon cells and their output neurons.


Introduction
Animals learn to modify their behavior based on past experience by changing connection strengths between neurons, and this synaptic plasticity is often regulated by metabotropic receptors. In particular, neurons commonly express both ionotropic and metabotropic receptors for the same neurotransmitter, where the two may mediate different functions (e.g. direct excitation/inhibition vs. synaptic plasticity). In mammals, where glutamate is the principal excitatory neurotransmitter, metabotropic glutamate receptors (mGluRs) have been widely implicated in synaptic plasticity and memory (Jö rntell and Hansel, 2006;Lüscher and Huber, 2010). Given the complexity of linking behavior to artificially induced plasticity in brain slices (Schonewille et al., 2011;Yamaguchi et al., 2016), it would be useful to study the role of metabotropic receptors in learning in a simpler genetic model system with a clearer behavioral readout of synaptic plasticity. One such system is Drosophila, where powerful genetic tools and well-defined anatomy have yielded a detailed understanding of the circuit and molecular mechanisms underlying associative memory (Busto et al., 2010;Cognigni et al., 2018;Hige, 2018). The principal excitatory neurotransmitter in Drosophila is acetylcholine, but, surprisingly, little is known about the function of metabotropic acetylcholine signaling in synaptic plasticity or neuromodulation in Drosophila. Here, we address this question using olfactory associative memory.
Flies can learn to associate an odor (conditioned stimulus, CS) with a positive (sugar) or a negative (electric shock) unconditioned stimulus (US), so that they later approach 'rewarded' odors and avoid 'punished' odors. This association is thought to be formed in the presynaptic terminals of the~2000 Recent work indicates that mAChR-A is required for aversive olfactory learning in Drosophila larvae, as knocking down mAChR-A expression in KCs impairs learning (Silva et al., 2015). However, it is unclear whether mAChR-A is involved in olfactory learning in adult Drosophila, given that mAChR-A is thought to signal through G q , and in adult flies G q signaling downstream of the dopamine receptor Damb promotes forgetting, not learning (Berry et al., 2012;Himmelreich et al., 2017). Moreover, it is unknown how mAChR-A affects the activity or physiology of KCs, where it acts (at KC axons or dendrites or both), and how these effects contribute to olfactory learning.
Here, we show that mAChR-A is required in KCs for aversive olfactory learning in adult Drosophila. Surprisingly, genetic and pharmacological manipulations of mAChR-A suggest that mAChR-A is inhibitory and acts on KC dendrites. Moreover, mAChR-A knockdown impairs the learning-associated depression of odor responses in an MB output neuron, MB-MVP2, that is required for aversive memory retrieval. We suggest that dendritically acting mAChR-A is required for synaptic depression between KCs and their outputs.

mAChR-A expression in KCs is required for aversive olfactory learning in adult flies
Drosophila larvae with reduced mAChR-A expression in KCs show impaired aversive olfactory learning (Silva et al., 2015), but it remains unknown whether mAChR-A in KCs also functions in learning in adult flies. We addressed this question by knocking down mAChR-A expression in KCs using two UAS-RNAi lines, 'RNAi 1' and 'RNAi 2' (see Materials and methods). Only RNAi 2 requires co-expression of Dicer-2 (Dcr-2) for optimal knockdown. To test the efficiency of these RNAi constructs, we expressed them pan-neuronally using elav-GAL4 and measured their effects on mAChR-A expression levels using quantitative real-time polymerase chain reaction (qRT-PCR). Both RNAi lines strongly reduce mAChR-A levels (RNAi 1: 39 ± 8% of elav-GAL4 control, or 61 ± 8% below normal; RNAi 2: 43 ± 10% of normal; mean ± s.e.m.; see Figure 1A). We then examined whether knocking down mAChR-A in KCs using the pan-KC driver OK107-GAL4 affects short-term aversive learning in adult flies. We used the standard odors used in the field (i.e. 3-octanol, OCT, and 4-methylcyclohexanol, MCH; see Materials and methods). Under these conditions, both UAS-RNAi transgenes significantly reduced aversive learning, whether training against MCH or OCT ( Figure 1B,C and Figure 1-figure supplement 1). Interestingly, knocking down mAChR-A did not affect learning when we trained flies with a more intense shock (90 V instead of 50 V, Figure 1-figure supplement 1), suggesting that mAChR-A may only be required for learning with moderate intensity reinforcement, not severe reinforcement. Consistent with this, knocking down mAChR-A had no effect on naïve avoidance of MCH and OCT ( Figure 1D; see Materials and methods) or flies' reaction to electric shock (Figure 1-figure supplement 1), showing that the defect was specific to learning, rather than reflecting a failure to detect odors or shock.
Given that mAChR-A is expressed in the larval MB and indeed contributes to aversive learning in larvae, it is possible that developmental effects underlie the reduced learning observed in mAChR-A knockdown flies. To test this, we used tub-GAL80 ts to suppress RNAi 1 expression during development. Flies were grown at 23˚C until 3 days after eclosion and were then transferred to 31˚C for 7 days. Adult-only knockdown of mAChR-A in KCs reduced learning ( Figure 1E), just as constitutive knockdown did, indicating that mAChR-A plays a physiological, not purely developmental, role in aversive learning. To further verify that GAL80 ts efficiently blocks RNAi expression (i.e. that GAL80 ts is not leaky), flies were grown at 23˚C without transferring them to 31˚C, thus blocking RNAi expression also in adults. When tested for learning at 10 days old, these flies showed normal learning ( Figure 1E). mAChR-A is required for olfactory learning in g KCs, not ab or a 0 b 0 KCs Kenyon cells are subdivided into three main classes according to their innervation of the horizontal and vertical lobes of the MB: g neurons send axons only to the g lobe of the horizontal lobes, while the axons of ab and a 0 b 0 neurons bifurcate and go to both the vertical and horizontal lobes (ab axons make up the a lobe of the vertical lobe and b lobe of the horizontal lobe, while a 0 b 0 axons make up the a 0 lobe of the vertical lobe and b 0 portion of the horizontal lobe). These different classes  The following source data and figure supplements are available for figure 1: Source data 1. Source data for Figure 1A.  play different roles in olfactory learning (Guven-Ozkan and Davis, 2014;Krashes et al., 2007). To unravel in which class(es) mAChR-A functions, we used a Minos-mediated integration cassette (MiMIC) line to investigate where mAChR-A is expressed (Venken et al., 2011). The MiMIC insertion in mAChR-A lies in the first 5' non-coding intron, creating a gene trap where GFP in the MiMIC cassette should be expressed in whichever cells endogenously express mAChR-A. Because the GFP in the original mAChR-A MiMIC cassette produced very little fluorescent signal (data not shown), we used recombinase-mediated cassette exchange (RMCE) to replace the original MiMIC cassette with a MiMIC cassette containing GAL4 (Venken et al., 2011). These new mAChR-A-MiMIC-GAL4 flies should express GAL4 wherever mAChR-A is endogenously expressed. To reveal the expression pattern of mAChR-A, we crossed mAChR-A-MiMIC-GAL4 and 20xUAS-6xeGFP flies. mAChR-A-MiMIC-GAL4 drove GFP expression throughout the brain, consistent with previous reports (Blake et al., 1993;Croset et al., 2018;Davie et al., 2018;Hannan and Hall, 1996) and with the fact that the Drosophila brain is mostly cholinergic. In the mushroom bodies, GFP was expressed in the ab and g lobes, but not the a 0 b 0 lobes ( Figure 2A). No GFP signal was observed with an inverted insertion where GAL4 is inserted in the MiMIC locus in the wrong orientation (data not shown). Consistent with these MiMIC results, two recently reported databases of single-cell transcriptomic analysis of the Drosophila brain (Croset et al., 2018;Davie et al., 2018) confirm that mAChR-A is more highly expressed in ab and g KCs than in a 0 b 0 KCs ( Figure 2-figure supplement 1). However, mAChR-A is still clearly present in a 0 b 0 KCs' transcriptomes, suggesting that mAChR-A-MiMIC-GAL4 may not reveal all neurons that express mAChR-A.
The higher expression of mAChR-A in ab and g KCs compared to a 0 b 0 KCs suggests that learning would be impaired by mAChR-A knockdown in ab or g, but not a 0 b 0 , KCs. To test this, we expressed mAChR-A RNAi in different KC classes. As expected, aversive olfactory learning was reduced by  knocking down mAChR-A in ab and g KCs together using MB247-GAL4, but not by knockdown in a 0 b 0 KCs using c305a-GAL4. To examine if ab and g KCs both participate in the reduced learning observed in mAChR-A knockdown flies, we sought to limit mAChR-A RNAi expression to either ab or g neurons. While strong driver lines exist for ab neurons, the g GAL4 drivers we tested were fairly weak (H24-GAL4, MB131B, R45H04-GAL4, data not shown), perhaps too weak to drive mAChR-A-RNAi enough to knock down mAChR-A efficiently. Therefore, we used MB247-GAL4, which was strong enough to affect behavior, and blocked GAL4 activity in either ab or g KCs by expressing the GAL80 repressor under the control of R44E04-LexA (ab KCs) or R45H04-LexA (g KCs) (Bräcker et al., 2013). These combinations drove strong, specific expression in ab or g KCs (Figure 2-figure supplement 2). Learning was reduced by mAChR-A RNAi expression in g, but not ab, KCs ( Figure 2B). These results suggest that mAChR-A is specifically required in g KCs for aversive olfactory learning and short-term memory.

mAChR-A suppresses odor responses in g KCs
We next asked what effect mAChR-A knockdown has on the physiology of KCs, by expressing GCaMP6f and mAChR-A RNAi 2 together in KCs using OK107-GAL4 (this driver and RNAi combination was also used for behavior in Figure 1C). Knocking down mAChR-A in KCs increased odorevoked Ca 2+ influx in the mushroom body calyx, where KC dendrites reside ( Figure 3). This result is somewhat surprising because mAChR-A is a G q -coupled receptor whose activation leads to Ca 2+ release from internal stores (Ren et al., 2015), which predicts that mAChR-A knockdown should decrease, not increase, odor-evoked Ca 2+ influx in KCs. However, some examples have been reported of inhibitory signaling through G q by M 1 -type mAChRs (see Discussion), and Drosophila mAChR-A may join these as another example of an inhibitory mAChR signaling through G q .
Because mAChR-A is required for aversive learning in g KCs, not ab or a 0 b 0 KCs (Figure 2), we next asked how odor responses in ab, a 0 b 0 and g KCs are affected by mAChR-A knockdown. ab, a 0 b 0 and g KC dendrites are not clearly segregated in the calyx, so we examined odor responses in the axonal lobes. Indeed, although odor responses in all lobes were increased by mAChR-A knockdown, only in the g lobe was the effect statistically significant for both MCH and OCT ( Figure 3). This result is consistent with the behavioral requirement for mAChR-A only in g KCs. However, we do not rule out the possibility that mAChR-A knockdown also affects ab and a 0 b 0 odor responses in a way that does not affect short-term memory, especially as ab and a 0 b 0 odor responses were somewhat, although not consistently significantly, increased. Although the DF/F traces from the g lobe had higher signal-to-noise ratio (SNR) than some other lobes (Figure 3-figure supplement 1) due to its larger size (averaging over more pixels) or shallower z-depth (less light scattering), a power analysis revealed that all lobes had SNRs high enough to detect an effect as large as that observed in the g lobe (Figure 3-figure supplement 1). However, note that we do not exclude the possibility that ab-or a 0 b 0 -specific (as opposed to pan-KC) knockdown of mAChR-A might significantly increase ab or a 0 b 0 KC odor responses. Do increased odor responses in g KCs prevent learning by increasing the overlap between the g KC population representations of the two odors used in our task (Lin et al., 2014)? When GCaMP6f and mAChR-A-RNAi 2 were expressed in all KCs, mAChR-A knockdown did not affect the sparseness or inter-odor correlation of KC population odor responses ( Figure 4A-C) even though it increased overall calyx responses. To focus specifically on g KCs, we expressed GCaMP6f and mAChR-A-RNAi 1 only in g KCs, using mb247-Gal4, R44E04-LexA and lexAop-GAL80, the same driver and RNAi combination used in the behavioral experiments in Figure 2B. GCaMP6f was visible mainly in the g lobe ( Figure 4D). g-only expression of mAChR-A-RNAi 1 increased odor responses in the calyx (here, dendrites of g KCs only) and, in the case of OCT, in the g lobe ( Figure 4E,F). Note that g KC odor responses were increased by both RNAi 1 ( Figure 3A,B) and RNAi 2 ( Figure 4E,F). As with pan-KC expression, g-only expression of mAChR-A-RNAi 1 did not affect the sparseness or inter-odor correlation of g KCs ( Figure 4G-I). Thus, mAChR-A knockdown does not impair learning through increased overlap in KC population odor representations.

KC odor responses are decreased by an mAChR agonist
RNAi-based knockdown of mAChR-A might induce homeostatic compensation that obscures or even reverses the primary effect of reduced mAChR-A expression. To test the acute role of mAChR-A in regulating KC activity, we took the complementary approach of pharmacologically activating mAChR-A. Initially, we bath-applied 10 mM muscarine, an mAChR-A agonist (Drosophila mAChR-B is 1000-fold less sensitive to muscarine than mAChR-A is [Collin et al., 2013], and mAChR-C is not expressed in the brain [Davie et al., 2018]). Muscarine strongly decreased odor responses in all subtypes of KCs ( Figure 5A,B, Figure 5-figure supplement 1). However, muscarine did not significantly affect the amplitude of odor responses in PN axons in the calyx ( Figure 5C), suggesting that the effect of muscarine on KCs arose in KCs, not earlier in the olfactory pathway. KCs can be silenced by an inhibitory GABAergic neuron called the anterior paired lateral (APL) neuron (Lin et al., 2014;Masuda-Nakagawa et al., 2014;Papadopoulou et al., 2011), so we asked whether muscarine reduces KC odor responses indirectly by activating APL, rather than directly inhibiting KCs. We applied muscarine to flies with APL-specific expression of tetanus toxin (TNT), which blocks inhibition from APL and thereby greatly increases KC odor responses. In these flies, APL is labeled stochastically, so hemispheres where APL was unlabeled served as controls (Lin et al., 2014) (see Materials and methods). Muscarine decreased KC odor responses both in  (6) for control, 12 (6) for knockdown. *p<0.05, **p<0.01, 2-way repeated-measures ANOVA with Holm-Sidak multiple comparisons test. (G) Example activity maps (single optical sections from a z-stack) of g KC odor responses to MCH and OCT in control (mb247-GAL4 > GCaMP6f, R44E04-LexA > GAL80) and knockdown (mb247-GAL4 > GCaMP6f, mAChR-A-RNAi 1, R44E04-LexA > GAL80) flies. Note the gaps in baseline GCaMP6f signal due to lack of ab and a 0 b 0 KCs labeled. Scale bar, 10 mm (H) Sparseness of g KC population responses is not affected by mAChR-A knockdown (p=0.76, two-way repeated-measures ANOVA). (I) Correlation between g KC population responses to MCH and OCT is not affected by mAChR-A knockdown (p=0.32, t-test). DOI: https://doi.org/10.7554/eLife.48264.015 The following source data is available for figure 4: Source data 1. Source data for control hemispheres and hemispheres where APL synaptic output was blocked by tetanus toxin (Figure 5D), and the effect of muscarine was not significantly different between the two cases ( Figure 5E). This result indicates that muscarine does not act solely by activating APL or by enhancing inhibition on KCs (e.g. increasing membrane localization of GABA A receptors).
To test mAChR-A function even more acutely, we locally applied muscarine to the MB calyx by pressure ejection (Figure 6, Figure 6-figure supplement 1). Red dye included in the ejected solution confirmed that the muscarine remained in the calyx for several seconds but did not spread to Peak DF/F during the odor pulse before and after muscarine in control hemispheres where APL was unlabeled (left, n = 6 hemispheres from 6 flies) and hemispheres where APL expressed tetanus toxin (TNT) (right, n = 6 hemispheres from 5 flies). *p<0.05, **p<0.01, ***p<0.001 by two-way repeated measures ANOVA with Holm-Sidak multiple comparisons test. (E) (Response (peak DF/F during the odor pulse) after muscarine) / (response before muscarine), using data from (D). No significant differences were observed (p>0.05, two-way repeated measures ANOVA with Holm-Sidak multiple comparisons test). DOI: https://doi.org/10.7554/eLife.48264.017 The following source data and figure supplement are available for figure 5: Source data 1. Source data for Figure 5. the MB lobes ( Figure 6B). Surprisingly, applying muscarine to the calyx in the absence of odor stimuli increased GCaMP6f signal in the calyx and a lobe, with small increases in the b and g lobe that were not statistically significant ( Figure 6A,C). It also decreased GCaMP6f signal in the a 0 and b 0 lobes around 1-2 s after application ( Figure 6A), although this effect was also not statistically significant. The increased Ca 2+ in the calyx most likely did not reflect increased excitability, as applying muscarine to the calyx did not increase the calyx odor response ( Figure 6D,E). If anything, it likely decreased the calyx odor response, because the Ca 2+ increase induced by muscarine alone (no odor) lasted~6-7 s and thus would have continued into the odor pulse in the muscarine +odor condition. If the odor response was unaffected by muscarine, the muscarine-evoked and odor-evoked increases in GCaMP6f signal should have summed. Instead, the peak calyx DF/F during the odor pulse was the same before and after locally applying muscarine, suggesting that the specifically odor-evoked increase in GCaMP6f was decreased by muscarine. Indeed, applying muscarine to the calyx suppressed odor responses in KC axons ( Figure 6D,E). Although muscarine did not significantly affect peak DF/F during the odor in the a lobe, muscarine most likely did decrease a lobe odor responses, by the same logic as for calyx odor responses (see above). Given that calyx muscarine suppresses a 0 b 0 axonal odor responses, the decrease in a 0 b 0 KC GCaMP6f signal in the absence of odor likely reflects suppression of spontaneous action potentials ( Figure 6A,C), as a 0 b 0 KCs have the highest spontaneous spike rate out of the three subtypes (Groschner et al., 2018;Turner et al., 2008). The effect of muscarine on a 0 b 0 KCs is consistent with single-cell transcriptome analyses showing that a 0 b 0 KCs express mAChR-A, albeit at a lower level than ab or g KCs (Figure 2-figure supplement 1) (Croset et al., 2018;Davie et al., 2018). The increase in calyx Ca 2+ induced by muscarine alone (without odor) might reflect Ca 2+ release from internal stores triggered by G q signaling, which then inhibits KC excitability (thus smaller odor responses). Note that muscarine on the calyx is unlikely to reduce KC odor responses via presynaptic inhibition of PNs, because bath muscarine does not affect odor-evoked Ca 2+ influx in PNs in the calyx ( Figure 5C), although we cannot rule out Ca 2+ -independent inhibition.

mAChR-A localized to the MB calyx can rescue learning in a mAChR-A hypomorphic mutant
We next asked where mAChR-A exerts its effect. To visualize the localization of mAChR-A, we created a new construct with mAChR-A tagged with FLAG on the C-terminus under UAS control. When we overexpressed FLAG-tagged mAChR-A in KCs using OK107-GAL4, we only observed anti-FLAG staining in the calyx ( Figure 7A), suggesting that mAChR-A is localized to the calyx. To test whether the FLAG tag or overexpression might cause the mAChR-A to be mis-localized, we tested whether mb247-GAL4 > mAChR A-FLAG overexpression could rescue learning in a mAChR-A mutant background. The original MiMIC allele with a GFP insertion in the 5' UTR intron of mAChR-A contains a stop cassette and polyadenylation signal, and indeed, it is a strongly hypomorphic allele: qPCR shows almost total lack of mAChR-A mRNA in the 'MiMIC-stop' allele ( Figure 7B). Flies homozygous for the 'MiMIC-stop' allele are viable but show impaired learning, while learning is significantly improved by using mb247-GAL4 to overexpress mAChR-A-FLAG in ab and g KCs ( Figure 7C), indicating that overexpressed mAChR-A-FLAG can support learning. These flies ('MiMIC-stop', mb247 >mAChR A-FLAG) also show anti-FLAG staining only in the calyx (Figure 7-figure supplement 1). These results suggest that mAChR-A exerts its effect on learning in KC dendrites, consistent with the effect of locally applying muscarine to KC dendrites.

mAChR-A knockdown prevents training-induced depression of MBON odor responses
The finding that mAChR-A functions in KC dendrites raises the question of how mAChR-A can affect learning. While learning-associated plasticity in KC dendrites has been observed in honeybees, In Drosophila, olfactory associative memories are stored by weakening the synapses between KCs and output neurons that lead to the 'wrong' behavior. For example, aversive memory requires an output neuron downstream of g KCs, called MBON-g1pedc>a/b or MB-MVP2. MB-MVP2 leads to approach behavior (Aso et al., 2014b), and aversive conditioning reduces MB-MVP2's responses to the aversively-trained odor (Hige et al., 2015;Perisse et al., 2016). We tested whether knocking down mAChR-A would prevent this depression. We knocked down mAChR-A in KCs using OK107-GAL4 and UAS-mAChR-A-RNAi 1, and expressed GCaMP6f in MB-MVP2 using R12G04-LexA and lexAop-GCaMP6f ( Figure 8A). We trained flies in the behavior apparatus and then imaged MB-MVP2 odor responses (3 hr after training to avoid cold-shock-sensitive memory). Because overall response amplitudes were variable across flies, for each fly we measured the ratio of the response to MCH (the trained odor) over the response to OCT (the untrained odor). Consistent with previous published results (Hige et al., 2015;Perisse et al., 2016), in control flies not expressing mAChR-A RNAi, the MCH/OCT ratio was substantially reduced in trained flies relative to mock-trained flies ( Figure 8B). This was not because the OCT response increased, because there was no difference between trained and mock-trained flies in the ratio of the response to OCT over the response to isoamyl acetate, a 'reference' odor that was absent in the training protocol. This was also not because of any general decrease in odor responses, as shown by analyzing absolute response amplitudes to MCH, OCT and isoamyl acetate (Figure 8-figure supplement 1). In contrast, in flies expressing mAChR-A RNAi in KCs, the MCH/OCT ratio was the same between trained and mock-trained flies ( Figure 8B), indicating that the mAChR-A knockdown impaired the learning-related depression of the KC to MB-MVP2 synapse. This result suggests that mAChR-A function in KC dendrites is necessary for learningrelated synaptic plasticity in KC axons.

Discussion
Here, we show that mAChR-A is required in g KCs for aversive olfactory learning and short-term memory in adult Drosophila. Knocking down mAChR-A increases KC odor responses, while the mAChR-A agonist muscarine suppresses KC activity. Knocking down mAChR-A prevents aversive learning from reducing responses of the MB output neuron MB-MVP2 to the conditioned odor, suggesting that mAChR-A is required for the learning-related depression of KC->MBON synapses.
Why is mAChR-A only required for aversive learning in g KCs, not ab or a 0 b 0 KCs? Although our mAChR-A MiMIC gene trap agrees with single-cell transcriptome analysis that a 0 b 0 KCs express less mAChR-A than do g and ab KCs (Croset et al., 2018;Davie et al., 2018), transcriptome analysis indicates that a 0 b 0 KCs do express some mAChR-A (Figure 2-figure supplement 1). Moreover, g and ab KCs express similar levels of mAChR-A (Crocker et al., 2016). It may be that the RNAi knockdown is less efficient at affecting the physiology of ab and a 0 b 0 KCs than g KCs, whether because the knockdown is less efficient at reducing protein levels, or because ab and different intrinsic properties or a different function of mAChR-A such that 40% of normal mAChR-A levels is sufficient in ab and a 0 b 0 KCs but not g KCs. This interpretation is supported by our finding that mAChR-A RNAi knockdown significantly increases odor responses only in the g lobe, not the ab or a 0 b 0 lobes. Alternatively, g, ab and a 0 b 0 KCs are thought to be important mainly for short-term memory, long-term memory, and memory consolidation, respectively (Guven-Ozkan and Davis, 2014; Krashes et al., 2007); as we only tested short-term memory, mAChR-A may carry out the same function in all KCs, but only its role in g KCs is required for short-term (as opposed to long-  term) memory. Indeed, the key plasticity gene DopR1 is required in g, not ab or a 0 b 0 KCs, for shortterm memory (Qin et al., 2012). It may be that mAChR-A is required in non-g KC types for other forms of memory besides short-term aversive memory, such as appetitive conditioning or other phases of memory like long-term memory. Our finding that mAChR-A is required in g KCs for aversive short-term memory is consistent with our finding that mAChR-A knockdown in KCs disrupts training-induced depression of odor responses in MB-MVP2, an MBON postsynaptic to g KCs required for aversive short-term memory (Perisse et al., 2016). However, the latter finding does not rule out the possibility that other MBONs postsynaptic to non-g KCs may also be affected by mAChR-A knockdown in KCs.
mAChR-A seems to inhibit KC odor responses, because knocking down mAChR-A increases odor responses in the calyx and g lobe, while activating mAChR-A with bath or local application of muscarine decreases KC odor responses. Some details differ between the genetic and pharmacological results. In particular, while mAChR-A knockdown mainly affects g KCs, with other subtypes inconsistently affected, muscarine reduces responses in all KC subtypes. What explains these differences? mAChR-A might be weakly activated in physiological conditions, in which case gain of function would cause a stronger effect than loss of function. Similarly, pharmacological activation of mAChR-A is likely a more drastic manipulation than a 60% reduction of mAChR-A mRNA levels. Although we cannot entirely rule out network effects from muscarine application, the effect of muscarine does not stem from PNs or APL ( Figure 5C,D) and locally applied muscarine would have little effect on neurons outside the mushroom body.
How does mAChR-A inhibit odor-evoked Ca 2+ influx in KCs? Given that mAChR-A signals through G q when expressed in CHO cells (Ren et al., 2015), that muscarinic G q signaling normally increases excitability in mammals (Caulfield and Birdsall, 1998), and that pan-neuronal artificial activation of G q signaling in Drosophila larvae increases overall excitability (Becnel et al., 2013), it may be surprising that mAChR-A inhibits KCs. However, G q signaling may exert different effects on different neurons in the fly brain, and some examples exist of inhibitory G q signaling by mammalian mAChRs. M 1 / M 3 /M 5 receptors acting via G q can inhibit voltage-dependent Ca 2+ channels (Gamper et al., 2004;Kammermeier et al., 2000;Keum et al., 2014;Suh et al., 2010), reduce voltage-gated Na +currents (Cantrell et al., 1996), or trigger surface transport of KCNQ channels (Jiang et al., 2015), thus increasing inhibitory K + currents. Drosophila mAChR-A may inhibit KCs through similar mechanisms.
What is the source of ACh which activates mAChR-A and modulates odor responses? In the calyx, cholinergic PNs are certainly a major source of ACh. However, KCs themselves are cholinergic  and release neurotransmitter in both the calyx and lobes (Christiansen et al., 2011). KCs form synapses on each other in the calyx (Zheng et al., 2018), possibly allowing mAChR-A to mediate lateral inhibition, in conjunction with the lateral inhibition provided by the GABAergic APL neuron (Lin et al., 2014).
What function does mAChR-A serve in learning and memory? Our results indicate that mAChR-A knockdown prevents the learning-associated weakening of KC-MBON synapses, in particular for MBON-g1pedc>a/b, aka MB-MVP2 (Figure 7). One potential explanation is that the increased odorevoked Ca 2+ influx observed in knockdown flies increases synaptic release, which overrides the learning-associated synaptic depression. However, increased odor-evoked Ca 2+ influx per se is unlikely on its own to straightforwardly explain a learning defect, because other genetic manipulations that increase odor-evoked Ca 2+ influx in KCs either have no effect on, or even improve, olfactory learning. For example, knocking down GABA synthesis in the inhibitory APL neuron increases odor-evoked Ca 2+ influx in KCs (Lei et al., 2013;Lin et al., 2014) and improves olfactory learning (Liu and Davis, 2009).
The most intuitive explanation would be that mAChR-A acts at KC synaptic terminals in KC axons to help depress KC-MBON synapses. Yet overexpressed mAChR-A localizes to KC dendrites, not axons, and functionally rescues mAChR-A hypomorphic mutants, showing that dendritic mAChR-A suffices for its function in learning and memory. Does this show that mAChR-A has no role in KC axons? Our inability to detect GFP expressed from the mAChR-A MiMIC gene trap suggests that normally there may only be a small amount of mAChR-A in KCs. It may be that with mAChR-A-FLAG overexpression, the correct (undetectable) amount of mAChR-A is trafficked to and functions in axons, but due to a bottleneck in axonal transport, the excess tagged mAChR-A is trapped in KC dendrites. While our results do not rule out this possibility, a general bottleneck in axonal transport seems unlikely as many overexpressed proteins are localized to KC axons (Trunova et al., 2011). We feel it is more parsimonious to take the dendritic localization of mAChR-A-FLAG at face value and infer that mAChR-A functions in KC dendrites.
How can mAChR-A in KC dendrites affect synaptic plasticity in KC axons? mAChR-A signaling might change the shape or duration of KC action potentials (Allen and Burnstock, 1990;Ghamari-Langroudi and Bourque, 2004), an effect that could potentially propagate to KC axon terminals (Juusola et al., 2007;Shu et al., 2006). Such changes in the action potential waveform may not be detected by calcium imaging, but could potentially affect a 'coincidence detector' in KC axons that detects when odor (i.e. KC activity) coincides with reward/punishment (i.e. dopamine). This coincidence detector is generally believed to be the Ca 2+ -dependent adenylyl cyclase rutabaga (Levin et al., 1992). Changing the waveform of KC action potentials could potentially affect local dynamics of Ca 2+ influx near rutabaga molecules. In addition, rutabaga mutations do not abolish learning (mutants have~40-50% of normal learning scores) (Yildizoglu et al., 2015), so there may be additional coincidence detection mechanisms affected by action potential waveforms. Testing this idea would require a better understanding of biochemical events underlying learning at KC synaptic terminals.
Alternatively, mAChR-A's effects on synaptic plasticity may not occur acutely. Although we ruled out purely developmental effects of mAChR-A through adult-only RNAi expression ( Figure 1E), knocking out mAChR-A for several days in adulthood might still affect KC physiology in a notentirely-acute way. For example, as with other G-protein-coupled receptors (Wang and Zhuo, 2012), muscarinic receptors can affect gene expression (von der Kammer et al., 1998), which could have wide-ranging effects on KC physiology, for example action potential waveform, expression of key genes required for synaptic plasticity, etc. Another intriguing possibility is suggested by an apparent paradox: both mAChR-A and the dopamine receptor Damb signal through G q (Himmelreich et al., 2017), but mAChR-A promotes learning while Damb promotes forgetting (Berry et al., 2012). How can G q mediate apparently opposite effects? Perhaps G q signaling aids both learning and forgetting by generally rendering synapses more labile. Indeed, although damb mutants retain memories for longer than wildtype, their initial learning is slightly impaired (Berry et al., 2012); damb mutant larvae are also impaired in aversive olfactory learning (Selcho et al., 2009). Although one study reports that knocking down G q in KCs did not impair initial memory (Himmelreich et al., 2017), the G q knockdown may not have been strong enough; also, that study shocked flies with 90 V shocks, which also gives normal learning in mAChR-A knockdown flies (Figure 1-figure supplement 1).
Such hypotheses posit that mAChR-A regulates synaptic plasticity 'competence' rather than participating directly in the plasticity mechanism itself. Why should synaptic plasticity competence be controlled by an activity-dependent mechanism? It is tempting to speculate that mAChR-A may allow a kind of metaplasticity (Abraham, 2008) in which exposure to odors (hence activation of mAChR-A in KCs) makes flies' learning mechanisms more sensitive. Indeed, mAChR-A is required for learning with moderate (50 V) shocks, not severe (90 V) shocks. Future studies may further clarify how muscarinic signaling contributes to olfactory learning.

Fly strains
Fly strains (see below) were raised on cornmeal agar under a 12 hr light/12 hr dark cycle and studied 1-10 days post-eclosion. Strains were cultivated at 25˚C unless they expressed temperature-sensitive gene products (GAL80 ts ); in these cases, the experimental animals and all relevant controls were grown at 23˚C. To de-repress the expression of RNAi with GAL80 ts , experimental and control animals were incubated at 31˚C for 7 days. Subsequent behavioral experiments were performed at 25˚C. Experimental animals carried transgenes over Canton-S chromosomes where possible to minimize genetic differences between strains. Details of fly strains are given in the Key Resources Table. UAS-mAChR-A-FLAG plasmid was generated by Gibson assembly of fragments using the NEBuilder HiFi Master Mix (NEB). Fragments were created by PCR using Phusion High-Fidelity DNA Polymerase (NEB). The full-length mAChR-A cDNA was purchased from GenScript (clone ID OFa11160). The vector was pTWF-attB, a gift from Prof. Oren Schuldiner (Yaniv et al., 2012). This vector consists of a FLAG tag in the C-terminal of the inserted gene and an attB site for site-specific integration of the transgene. PCR and Gibson assembly were carried out following the manufacturer's recommendations with the following primers: For mAChR-A: tgggaattatcgacaagtttgtacaaaaaagcaggctATGGAGCCGGTCATGAGTC and cactttgtacaagaaagctgggtaATTGTAGACGCCGCGTAC For pTWF-AttB: aaagctgggtaCTTGTACAAAGTGGTGAGCTCC and agcctgcttttttgtacAAACTTG TCGATAATTCCC Transgenes were injected into the attP2 landing site using jC31 integration (by BestGene).

Behavioral analysis
Behavioral experiments were performed in a custom-built, fully automated apparatus (Claridge-Chang et al., 2009;Lin et al., 2014;Parnas et al., 2013). Single flies were housed in clear polycarbonate chambers (length 50 mm, width 5 mm, height 1.3 mm) with printed circuit boards (PCBs) at both floors and ceilings. Solid-state relays (Panasonic AQV253) connected the PCBs to a 50 V source. Air flow was controlled with mass flow controllers (CMOSens PerformanceLine, Sensirion). A carrier flow (2.7 l/min) was combined with an odor stream (0.3 l/min) obtained by circulating the air flow through vials filled with a liquid odorant. Odors were prepared at 10 fold dilution in mineral oil. Therefore, liquid dilution and mixing carrier and odor stimulus stream resulted in a final 100 fold dilution of odors. Fresh odors were prepared daily.
The 3 l/min total flow (carrier and odor stimulus) was split between 20 chambers resulting in a flow rate of 0.15 l/min per half chamber. Two identical odor delivery systems delivered odors independently to each half of the chamber. Air or odor streams from the two halves of the chamber converged at a central choice zone. The 20 chambers were stacked in two columns each containing 10 chambers and were backlit by 940 nm LEDs (Vishay TSAL6400). Images were obtained by a MAKO CMOS camera (Allied Vision Technologies) equipped with a Computar M0814-MP2 lens. The apparatus was operated in a temperature-controlled incubator (Panasonic MIR-154) maintained at 25˚C.
A virtual instrument written in LabVIEW 7.1 (National Instruments) extracted fly position data from video images and controlled the delivery of odors and electric shocks. Data were analyzed in MATLAB 2015b (The MathWorks) and Prism 6 (GraphPad).
A fly's preference was calculated as the percentage of time that it spent on one side of the chamber. Training and odor avoidance protocols were as depicted in Figure 1. The naïve avoidance index was calculated as (preference for left side when it contains air) -(preference for left side when it contains odor). During training, MCH was paired with 12 equally spaced 1.25 s electric shocks at 50 V (Tully and Quinn, 1985). The learning index was calculated as (preference for MCH before training) -(preference for MCH after training). Flies were excluded from analysis if they entered the choice zone fewer than four times during odor presentation.
Movies were motion-corrected in X-Y using the moco ImageJ plugin (Dubbs et al., 2016), with pre-processing to collapse volume movies in Z and to smooth the image with a Gaussian filter (standard deviation = 4 pixels; the displacements generated from the smoothed movie were then applied to the original, unsmoothed movie), and motion-corrected in Z by maximizing the pixel-by-pixel correlation between each volume and the average volume across time points. DF/F, activity maps, sparseness and inter-odor correlation were calculated as in Lin et al. (2014). Briefly, movies were smoothed with a 5-pixel-square Gaussian filter (standard deviation 2). Baseline fluorescence was taken as the average fluorescence during the pre-stimulus period. Frames with sudden, large axial movements were discarded by correlating each frame to the baseline image and discarding it if the correlation fell below a threshold value, which was manually selected for each brain by noting the constant high correlation value when the brain was stationary and sudden drops in correlation when the brain moved. DF/F was calculated for each pixel as the difference between mean fluorescence during the stimulus period vs. the baseline fluorescence (DF), divided by the baseline fluorescence. For pixels where DF did not exceed two times the standard deviation over time of that pixel's intensity during the pre-stimulus period, the pixel was considered non-responsive. We excluded nonresponsive flies and flies whose motion could not be corrected.
Inter-odor correlations were calculated by first aligning the activity maps of each odor response by maximizing the inter-odor correlations of baseline fluorescence, and then converting image matrices of the activity maps of each odor response into linear vectors and calculating the Pearson correlation coefficients between each 'odor vector'. A threshold for baseline fluorescence was applied as a mask to the activity map to exclude pixels with no baseline GCaMP6f signal. Population sparseness was calculated for activity maps using the following equation (Vinje and Gallant, 2000;Willmore and Tolhurst, 2001):

Statistics
Statistical analyses were carried out in GraphPad Prism as described in figure legends and Supplementary file 1. In general, no statistical methods were used to predetermine sample sizes, but where conclusions were drawn from the absence of a statistically significant difference, a power analysis was carried out in G*Power to confirm that the sample size provided sufficient power to detect an effect of the expected size. The experimenter was blind to which hemispheres had APL neurons expressing tetanus toxin before post-experiment dissection ( Figure 5) but not otherwise. The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Data availability All data generated or analysed during this study are included in the manuscript and supporting files. Source data files have been provided for Figures 1-8