Pairing-Dependent Plasticity in a Dissected Fly Brain Is Input-Specific and Requires Synaptic CaMKII Enrichment and Nighttime Sleep

In Drosophila, in vivo functional imaging studies revealed that associative memory formation is coupled to a cascade of neural plasticity events in distinct compartments of the mushroom body (MB). In-depth investigation of the circuit dynamics, however, will require an ex vivo model that faithfully mirrors these events to allow direct manipulations of circuit elements that are inaccessible in the intact fly. The current ex vivo models have been able to reproduce the fundamental plasticity of aversive short-term memory, a potentiation of the MB intrinsic neuron (Kenyon cells [KCs]) responses after artificial learning ex vivo. However, this potentiation showed different localization and encoding properties from those reported in vivo and failed to generate the previously reported suppression plasticity in the MB output neurons (MBONs). Here, we develop an ex vivo model using the female Drosophila brain that recapitulates behaviorally evoked plasticity in the KCs and MBONs. We demonstrate that this plasticity accurately localizes to the MB α′3 compartment and is encoded by a coincidence between KC activation and dopaminergic input. The formed plasticity is input-specific, requiring pairing of the conditioned stimulus and unconditioned stimulus pathways; hence, we name it pairing-dependent plasticity. Pairing-dependent plasticity formation requires an intact CaMKII gene and is blocked by previous-night sleep deprivation but is rescued by rebound sleep. In conclusion, we show that our ex vivo preparation recapitulates behavioral and imaging results from intact animals and can provide new insights into mechanisms of memory formation at the level of molecules, circuits, and brain state. SIGNIFICANCE STATEMENT The mammalian ex vivo LTP model enabled in-depth investigation of the hippocampal memory circuit. We develop a parallel model to study the Drosophila mushroom body (MB) memory circuit. Pairing activation of the conditioned stimulus and unconditioned stimulus pathways in dissected brains induces a potentiation pairing-dependent plasticity (PDP) in the axons of α′β′ Kenyon cells and a suppression PDP in the dendrites of their postsynaptic MB output neurons, localized in the MB α′3 compartment. This PDP is input-specific and requires the 3′ untranslated region of CaMKII. Interestingly, ex vivo PDP carries information about the animal's experience before dissection; brains from sleep-deprived animals fail to form PDP, whereas those from animals who recovered 2 h of their lost sleep form PDP.


Introduction
A neutral experience (conditioned stimulus [CS]) can be remembered as positive or negative if closely followed by rewarding or punishing reinforcement (unconditioned stimulus [US]). The ability to form this type of "associative" memory is phylogenetically conserved; Drosophila form robust associative memories (Tully and Quinn, 1985), most of which are encoded and stored in the mushroom body (MB) (de Belle and Heisenberg, 1994). The MB is a higher brain structure made of 15 distinct compartments. Each compartment is built on a scaffold of axons of one of the three main types of Kenyon cells (KCs; ab , a9b 9, and g ). The KCs connect to MB output neurons (MBONs), which project out of the MB to bias behavior (Aso et al., 2014a,b;. The KC!MBON synapses are modulated by dopaminergic neurons.
During aversive olfactory associative learning, an odor (the CS) activates a sparse group of KCs, such that this odor identity is represented across all MB compartments (Turner et al., 2008;Lin et al., 2014). Simultaneously, dopaminergic neurons from the protocerebral posterior lateral (PPL1) cluster are activated by the US, encoding negative prediction errors in MB compartments (Riemensperger et al., 2005;Schroll et al., 2006;Claridge-Chang et al., 2009;Mao and Davis, 2009;Aso et al., 2012). When KC activation and the dopaminergic signal coincide within a compartment, the KC!MBON synapses in that compartment are depressed, biasing the circuit output to aversion (Sejourne et al., 2011;Cohn et al., 2015;Hige et al., 2015;.
Many studies have investigated the properties of this circuitry using in vivo calcium imaging in intact animals (for review, see Adel and Griffith, 2021). In contrast, explanted brains have been used mostly for establishing connectivity between neurons or interrogating a specific biochemical pathway; only a few studies have attempted to understand memory circuit logic ex vivo (Wang et al., 2008;Ueno et al., 2013Ueno et al., , 2017Suzuki-Sawano et al., 2017). In the best-developed paradigm, Ueno et al. (2013) observed a potentiation of KC responses in the tips of the MB vertical lobes which they termed "long-term enhancement" (LTE). This laid the groundwork for developing ex vivo models of this circuit, but there were major differences between LTE and associative memory observed in intact animals. The most significant were that the plasticity was not specific to the a9b 9 KCs and that dopamine release by the US was not observed; it was only seen after CS1US coincidence (Ueno et al., 2013. In this study, we establish an ex vivo paradigm that resolves these discrepancies and exhibits the cardinal features of associative learning. We show that pairing odor and punishment pathway activation in dissected brains results in a localized potentiation of the a9b 9 KCs and suppression of their postsynaptic MBONs in the a93 compartment. Because both KC potentiation and MBON suppression are strictly dependent on temporal coincidence of the CS and US, we term this paradigm "pairing-dependent plasticity" (PDP). We show that, like the CS specificity of associative memories, PDP is specific to the subset of odor-representing projection neurons activated during the artificial training. We also provide evidence that dopamine is released by activation of the US pathway and does not require CS1US coincidence.
This ex vivo paradigm can be used for obtaining new mechanistic insight into memory formation at the molecular and circuit levels. We present data indicating that the 39UTR of the CaMKII gene is critical for short-term memory (STM) formation and that the primacy of a9 compartment plasticity in learning is because of differences in input/response relationships between a and a9. Finally, we demonstrate that the ability of the ex vivo brain to be plastic can be influenced by prior in vivo experience, as we report that brains of sleep-deprived flies fail to form PDP, but as little as 2 h of recovery sleep rescues this learning impairment.
Ex vivo imaging and electrical stimulations. Brains from 4-to 8-dold female flies were dissected in ice-cold HL3 medium (70 mM NaCl, 115 mM sucrose, 5 mM KCl, 20 mM MgCl 2 , 1.8 mM CaCl 2 , 10 mM NaHCO 3 , 5 mM trehalose, 5 mM HEPES, osmolarity: 395.4 mOsm; pH 7.3) (Stewart et al., 1994). Brains were then transferred to an imaging chamber containing fresh HL3 saline and immobilized using tungsten pins over the optic lobes. In the case of paired stimulations of the antennal lobe (AL) and the ascending fibers of the ventral nerve cord (AFV), the brain was dissected with the ventral nerve cord attached. The ventral nerve cord was then cut using sharp scissors near its base, leaving one end of the AFV free. Glass suction microelectrodes were used to apply the electrical stimulation to either the AL or the AFV. Because odor information is randomly encoded in the MB (Bhandawat et al., 2007;Turner et al., 2008;Honegger et al., 2011;Lei et al., 2013), we did not target specific AL glomeruli across different animals. However, the properties of the stimulation and the electrode were kept the same across all experiments unless noted otherwise. Based on the size of the AL electrode tip and the distribution of AL calcium responses to AL stimulation (see Fig. 3B-D), we estimate that 20%-25% of the ipsilateral AL projection neurons are activated with our AL stimulation protocol. Brains were always perfused with fresh saline throughout the experiment with a flow rate of ;2 drops per second.
In pre-induction and post-induction AL stimulations, 5 trains of 20 pulses at 100 Hz were applied. Pulse width is 1 ms and interpulse interval is 9 ms. Intertrain interval is 15 s. Stimulation strength is 100 (low input stimulation) or 200 (high input stimulation) mAmps. During induction, 12 AL stimulation trains were applied with 5 s intertrain interval. Regarding the AL electrode size, we noticed a relationship between the diameter of the AL electrode tip and the input current such that applying a 100 mAmps input stimulation via a large electrode tip diameter has a similar effect on baseline KC calcium responses and on the resultant plasticity as applying a 200 mAmps via a small electrode tip diameter. For example, with a 100 mAmps input stimulation, PDP can be formed in the MB a3 compartment if the diameter of the AL electrode is significantly enlarged. Therefore, to minimize variability within the same dataset, all electrodes used in any experiment in this study were made at the same time using a p-97 micropipette puller (Sutter Instruments) before data collection. To minimize variability across the different experiments, we kept the AL electrode size approximately one-fourth of the AL size.
AFV stimulation was similar to AL stimulation, but the diameter of the AFV stimulation electrode was large enough to suck in the free end of the AFV. In GRAB DA2m experiments, AFV stimulation strength was adjusted to be above the threshold of KC GRAB DA2m responses; it varied between 500 and 1000 mAmps. We noted that AFV stimulation under these conditions did not produce a GRAB DA2m response in horizontal lobes, suggesting that, while PPL1 neurons were stimulated, PAM neurons were not. Whether this reflects differences in circuitry or differences in relative excitability (we are using HL3) is unknown. In AL pairing experiments, we noted in tests of the AFV electrode that there were MB GCaMP responses in some animals, but not all (2 of 8 had no response). In no case, however, did AFV stimulation alone cause PDP, suggesting that AFV-stimulated MB calcium increases are not able to act as a CS. Plasticity has been observed to be dependent on calcium entry pathway previously in mammalian brain (Deisseroth et al., 1998).
Images were captured using an ORCA Flash4.0 V3 sCMOS camera at 200 frames per second (except in the MBON imaging experiment using MB027B split-GAL4 . 20Â-UAS-GCaMP6f, frame rate was reduced to 2 frames per second) with a 40Â water immersion lens on an Olympus upright microscope BX50W1. Images were collected as 512 Â 512 resolution and a binning factor of 2 Â 2. Imaging was done using the HCImage Live software. Excitation of the used florescent sensors was done using the CoolLED pE-4000 LED source. For GCaMP and GRAB DA2m , the 470 nm LED channel was used with an excitation filter Chroma 450/50 and emission filter FF01-525/45-25. jRCaMP excitation was done using the 550 nm LED channel and the excitation filter FF01-530/43-25 and an emission filter FF01-607/36-25. Calcium traces at every frame were calculated as DF/F 0 = (F -F 0 )/F 0 , where F is the florescence value at a given frame and F 0 is the florescence value at baseline. Peak response to a stimulation was calculated by subtracting the average DF/F 0 during the last second before a stimulation from the peak DF/F 0 during stimulation. The responses to AL stimulations before and 15 min after induction were averaged to calculate the preand post-responses, respectively. PDP values or DF/F relative change was calculated as (postpre)/pre.
The isolated pulse stimulator model 2100 and the perfusion system ValveLink 8.2 were triggered using a custom program written and controlled by the pClamp 11 software and the Digidata1550A digitizer.
Behavioral experiments. Aversive learning experiments were performed in an environmental room in red light at 25°C with 65% humidity. Flies (mixed males and females) were between 4-and 14-d-old. Flies were given at least 10 min acclimation period in room before training or testing. Data for each experiment were pooled from at least three independent experimental days. The learning assays were performed as described by the Quinn laboratory (Tully and Quinn, 1985). The US was provided as 12 1 s 90 or 60 V shocks during the 1 min CS-US pairing; 10% 4-methylcyclohexanol and 3-octanol OCT were used as the CS odors. Flies were then given a 2 min rest. Testing involved 2 min of simultaneous exposure to CS odors, after which flies choosing either odor were counted. A performance index was calculated for each trial as (number of flies choosing the conditioned odor) -(number of flies choosing the not-conditioned odor)/(total number of flies). This performance index was averaged between reciprocal trials where one of the odors was conditioned in one trial and the other odor was conditioned in the other to calculate the Learning Index. To confirm the CaMKII UDel flies' sensitivity to electric shock, 2 min preference tests were performed during which flies chose between a stimulus vial (24 spaced 1 s 90 V shocks) and a neutral vial.
Sleep assay. Sleep deprivation was done in 25°C incubators on a 12 h light, 12 h dark cycles. Mated female flies were individually loaded into glass sleep tubes containing a food mixture of 5% sucrose and 2% agar. Drosophila Activity monitors system (TriKinetics) was used to measure sleep. Sleep was defined as inactivity bouts of 5 or more minutes (Hendricks et al., 2000;Shaw et al., 2000). Flies were sleep-deprived by turning on the shaker between ZT12 and ZT24. Drosophila Activity monitor data were analyzed using a custom MATLAB program (Donelson et al., 2012).
Immunostaining. Adult fly brains were dissected in cold Schneider's Insect Medium (Sigma, S0146), and then fixed in 4% PFA solution for 30 min at room temperature. Fixed brains were washed 3 Â 30 min in 0.5% Triton-PBS (PBST) solution, blocked in 10% normal goat serum solution for 1 h, and incubated in mouse anti-CaMKII antibody (1:10 000, Cosmo) for 3 d. CaMKII antibody solutions were removed, and samples washed in PBST solution for 3 Â 30 min. Samples were then incubated in AlexaFluor-633 anti-mouse antibody (Invitrogen) overnight, then washed 3 Â 30 min in PBST solution and mounted in the Vectashield mounting medium. Images were taken under a 20Â objective lens with the same settings using a Leica SP5 confocal microscope. The images were analyzed by ImageJ software. For the intensity of MB regions, the middle slices were selected as the representative pictures, and mean intensity of all MB lobes was quantified.
Immunoblotting. A total of 100 5-d-old Canton-S WT or CaMKII UDel flies (mixed males and females) were frozen on dry ice and vortexed to remove heads. Fly heads were separated from the fly bodies using a sieve. The heads were then homogenized in loading buffer (4Â Bolt LDS, Invitrogen, Novex with 5% b -mercaptoethanol added) and heated for 10 min. Proteins were separated by SDS-PAGE (Bolt, Bis-Tris Protein Gels, Invitrogen) and transferred to a nitrocellulose membrane (GE Healthcare). Membrane was blocked (Blocking Buffer for Fluorescent Western Blotting, Rockland Immunochemicals) and then incubated with Anti-dCaMKII Clone 18 (1:1000, CosmoBio) and antiactin mAb C4 (1:1000, Millipore). The secondary antibody used was DyLight 680 mouse. Membrane was imaged using ChemiDoc system from Bio-Rad. Intensity of bands was calculated using Adj. Volume (Int) within the ImageLab 6.0 software. Intensity of the CaMKII band was normalized to that of actin in the same lane.
Experimental design and statistical analysis. Female flies were used in all imaging experiments because of their larger size. Both males and females aged 4-8 d were used in other experiments unless noted otherwise. All statistical analyses were performed in Prism 9 software. All tests were two-tailed, and confidence levels were set at a = 0.05. Normality of statistical data was determined via the Shapiro-Wilk test (a = 0.05). Parametric tests were used for all experiments, except the immunoblotting experiment where nonparametric tests were used. The statistical tests, p values, sample sizes, and other statistical information for each experiment are listed in the figure legends. Post hoc analysis information for Figures 2G and 5F is listed in Tables 1 and 2, respectively.

Artificial aversive training induces a potentiation PDP in KCs
Previous studies had demonstrated that enhanced calcium responses similar to those reported in the MB in vivo after aversive training can be achieved in dissected brains (Wang et al., 2008;Tomchik and Davis, 2009;Ueno et al., 2013). To investigate whether this plasticity recapitulates the formation of aversive memory, and to optimize our protocol, we dissected the CNS (brain with attached ventral nerve cord) of 4-to 8-d-old mated female flies expressing the Ca 21 indicator GCaMP6f in the MBs using the KC driver (VT030559-GAL4). Dissected brains were pinned in an imaging bath chamber, and the cervical connective toward the base of the ventral nerve cord was cut to allow electrical stimulation of the AFV connecting the ventral nerve cord to the brain (Fig. 1A). We used glass suction electrodes to stimulate the AL projection neurons and the AFV to activate the odor pathway and the electric shock pathway, respectively (Fig. 1B). Given that in vivo aversive STM formation is correlated with enhanced calcium responses in the a93 MB compartment but not in the a3 compartment (Krashes et al., 2007;Wang et al., 2008;Cervantes-Sandoval et al., 2013;Zhang et al., 2019), we decided to focus our calcium imaging on these two regions (Fig. 1C). We found that both a93 and a3 compartments respond only to ipsilateral AL stimulation, but AFV stimulation could generate a calcium response in both compartments of both MBs. Therefore, we imaged the MB ipsilateral to the stimulated AL.
To induce the formation of an aversive STM trace, we paired 12 trains of stimulation of both the AL electrode and the AFV electrode, thus activating the CS and the US pathways at the same time (paired induction). As controls, we repeated the same induction paradigm but separated the AL stimulation from the AFV stimulation by 30 s (unpaired induction) or omitted the AFV stimulation (AL alone induction) or omitted the AL stimulation (AFV alone induction) or simply allowed the brain to rest for 15 min (no induction). We compared the change in calcium response to AL stimulation before induction and 15 min after induction (Fig. 1D). We predicted that, if this ex vivo model truly recapitulated short-term aversive training, only the paired induction should produce an enhancement of the calcium responses in the a93 and none of the stimuli should potentiate a3 (Krashes et al., 2007;Wang et al., 2008;Davis, 2011;Cervantes-Sandoval et al., 2013). Indeed, this was the result we obtained (Fig. 1E,F). Interestingly, we noticed that the repetitive unpaired activation of KCs results in suppression of their calcium responses (Fig. 1F). This is reminiscent of desensitization or habituation; a nonassociative plasticity/ memory described as a decrement in neural responses to uninteresting, frequently encountered stimuli whose predictive value is negligible (Wilson and Linster, 2008;Wilson, 2009). These nonassociative processes have been documented in insects (Cho et al., 2004;Das et al., 2011;Semelidou et al., 2018), rodents (Wilson, 1998;Best and Wilson, 2004), and humans (Ferdenzi et al., 2014;Pellegrino et al., 2017).
Dopamine replaces the artificial US stimulus but does not replace the CS Multiple behavioral and in vivo imaging studies have shown that dopaminergic neurons encode the US valence information in the MB, with PPL1s providing aversive reinforcement and PAMs providing appetitive reinforcement (Schwaerzel et al., 2003;Riemensperger et al., 2005;Schroll et al., 2006;Kim et al., 2007;Claridge-Chang et al., 2009;Mao and Davis, 2009;Aso et al., 2010Aso et al., , 2012Aso et al., , 2014bPech et al., 2013;Yamagata et al., 2016;Handler et al., 2019). These studies suggest that the US alone is sufficient to evoke dopamine release. However, mechanistic studies using the LTE paradigm reported no dopamine release after US pathway stimulation alone. Strong dopamine release was only seen after coincident activation of both the CS and the US pathways, and it was concluded that dopamine release is downstream of the CS1US coincidence and does not encode the primary US information (Ueno et al., , 2020. To directly address this discrepancy, we expressed a G-protein-coupled receptor activation-based dopamine sensor (GRAB DA ) (Sun et al., 2020) in the KCs. We then dissected the fly's CNS and activated the US pathway by stimulating the AFV. We also expressed the calcium indicator jRCaMP1a in the KCs to use as a reference for the strength of activation ( Fig. 2A,B). We observed a robust dopamine release in the a93 and a much weaker release onto the a3 compartment in response to the same AFV stimulation used in our induction experiments (Fig. 2C,D). This result shows that dopamine release occurs in response to the US stimulus alone and does not require CS1US coincidence.  Figure 5F. N1, Sample size of first group; N2, sample size of second group; t, value of the t distribution. We next asked whether dopamine perfusion can replace electrical stimulation of the AFV. We paired electrical stimulation of the AL with perfusion of different concentrations (1, 5, and 10 mM) of dopamine and analyzed the KC response to AL stimulation before and after the pairing (Fig. 2E,F). Similar to the plasticity seen after AL1AFV induction, we found that only the paired AL1 DA (5 or 10 mM) induction induced an enhancement PDP in the a93; pairing with 1 mM DA was not sufficient to induce PDP ( Fig. 2G; for detailed statistical data for the post hoc analysis, see Table 1). Also, in agreement with our observation that dopamine release in a3 was very weak compared with the a93 compartment on AFV stimulation, our protocol did not induce ex vivo PDP in the a3 compartment (Fig. 2G). We also found that PDP (induced by AL 1 10 mM DA) lasts for at least 1 h after induction (Fig. 2H).
Although these results supported a PDP model with direct dopamine modulation of KCs, we were concerned with the lack of synaptic specificity in the AL1DA induction, as dopamine was perfused onto the whole preparation. To address this concern, we repeated the experiment but replaced dopamine perfusion with activation of the dopaminergic neurons in the PPL1 cluster, which carry the negative valence US information to the MB (Schroll et al., 2006;Claridge-Chang et al., 2009;Mao and Davis, 2009;Aso et al., 2010Aso et al., , 2012Burke et al., 2012;Aso and Rubin, 2016). We expressed the ATP-gated channel P2X 2 (Yao et al., 2012) in the PPL1 dopaminergic cluster using the TH-LexA driver. We paired AL stimulation with application of 2.5 mM ATP to activate PPL1s (Fig. 2I,J). Again, we observed an enhancement PDP in the a93 but not in the a3 compartment (Fig. 2K).
Together, our findings agree with previous behavioral studies and in vivo imaging studies and show that dopamine is released in response to the US alone and carries US valence information to the MB to allow associative learning. Potentiation PDP induced in our ex vivo paradigm also localizes to a9 cells only, congruent with the memory traces reported in vivo with aversive STM.
PDP ex vivo plasticity is input-specific Associative memory is CS-specific. A fly that is aversively trained against an odor shows aversive behavior to that odor and closely similar odors only but does not generalize the aversion to all odors (Barth et al., 2014). Olfactory information is sparsely transmitted to KC dendrites, with the identity of the odor being determined by population coding (Marin et al., 2002;Perez-Orive et al., 2002;Wong et al., 2002;Murthy et al., 2008;Turner et al., 2008;Honegger et al., 2011). To mimic the application of two different odors to the fly brain, we placed two suction glass electrodes onto the same AL and spaced them as far as possible from each other to maximize the probability that each electrode stimulates a distinct subset of projection neurons. Theoretically, a certain degree of overlap between the two subsets could be allowed as long as the two activated populations are sufficiently distinct. To test whether the two electrodes activated distinct populations of projection neurons, we expressed the calcium indicator (GCaMP6f) in the projection neurons and divided the AL into multiple ROIs, asking whether the pattern of responses was equivalent for the two electrodes (Fig. 3A,B). We found that spacing resulted in activation of distinct populations of projection neurons, allowing us to provide two hypothetical odors to the MB (Fig. 3C,D).
We then asked whether the observed ex vivo PDP is specific to the olfactory input activated during induction. To challenge the olfactory-input specificity of our preparation even more, we used the less specific induction method, AL1DA, in which dopamine is impartially perfused to the whole brain to eliminate any specificity on the US pathway side (Fig. 3E). We recorded KC calcium responses to both electrode stimulations before induction (pre). Then, we delivered a train of stimulations through electrode 1, followed by a train of stimulations through electrode 2. Dopamine perfusion started 5 s before the stimulation train for only one of the two electrodes (paired) and stopped at the last pulse of the train, while normal HL3 saline was perfused during the activation of the other electrode (unpaired). We then allowed the brain to rest for 15 min and tested the KC response to electrode 2 first, then to electrode 1 to eliminate any bias because of the order of stimulation (Fig. 3F). An increased Top, In the paired induction, the ipsilateral AL is activated by 5 stimulation trains with 15 s intertrain interval followed by resting for 1 min (pre-induction). Twelve trains of stimulations are then applied via the AL electrode, coincident with 60 s of dopamine perfusion. The brain is then rested for 15 min before being tested by 5 trains of AL stimulations like those applied during the pre-induction (post-induction). Bottom, The unpaired induction: same as paired induction, except that the AL stimulation and dopamine perfusion are separated by 30 s during the induction stage. The stimulation train is 20 pulses at 100 Hz; each pulse is 1 ms with 9 ms interpulse interval. AL stimulation strength: 100 mAmps. G, Mean relative change of the calcium responses in the a93 (left) and the a3 (right) compartments after DA perfusion alone (gray), unpaired AL1DA induction (green), and paired AL1DA induction (blue). Three different DA concentrations were used: 1 mM (left), 5 mM (middle), and 10 mM (right). Data are mean 6 SEM. Two-way ANOVA: lobe effects, F (1,68) = 2.724, p = 0.1034; induction effects, F (8,68) = 9.584, p , 0.0001. Tukey's post hoc results are listed in Table 1 min}, respectively. I, Schematic of the AL1PPL1 induction protocol, showing electrode placement above the AL, and perfusion of 2.5 mM ATP to activate the P2X 2 channels expressed in the PPL1 dopaminergic neurons (driven by TH-LexA). J, Paired and unpaired induction protocols. Top, In the paired induction, the ipsilateral AL is activated by 5 stimulation trains with 15 s intertrain interval followed by resting for 1 min (pre-induction). We then apply 12 trains of stimulations to the AL electrode, coincident with 60 s of 2.5 mM ATP perfusion. The brain is then rested for 15 min before being tested by 5 trains of AL stimulations like those applied during the pre-induction (post-induction). Bottom, The unpaired induction: same as paired induction, except that the AL stimulation and ATP perfusion are separated by 30 s during the induction stage. The stimulation train is 20 pulses at 100 Hz; each pulse is 1 ms with 9 ms interpulse interval. AL stimulation strength: 100 mAmps. K, Mean relative change of the calcium responses in the a93 (left) and the a3 (right) compartments after unpaired AL1ATP induction (green), and paired AL1ATP induction (blue). Data are mean 6 SEM. Two-way ANOVA (a = 0.05; n = 8 in each condition): lobe effects, F (1,28) = 9.053, p = 0.0055; induction effects, F (1,28) = 8.284, p = 0.0076. Sidak post hoc tests: in the a93 compartment, **t (28) = 3.446, p , 0.0036; In the a3 compartment, ns t (28) = 0.624, p = 0.7862. calcium response was observed in a93 after activation of the paired subset compared with the unpaired one (Fig. 3G,H). This result demonstrates that even with the broader bath-dopamine induction paradigm, the plasticity achieved in our paradigm is specific to the CS input activated during ex vivo training.

Artificial aversive training induces a suppression PDP in a93
MBONs While the a9 branches of the KCs are potentiated in response to an aversively trained odor in intact animals, responses in their postsynaptic MBONs are suppressed (Sejourne et al., 2011;Zhang et al., 2019). To eliminate the possibility that our paradigm potentiates memory-relevant KCs via a memory-irrelevant epiphenomenon, we needed to demonstrate that our ex vivo training could produce a suppression PDP in MBONs. Therefore, we repeated the same AL1DA induction described previously (Fig. 2F), but this time we expressed 20xGCaMP6f with the MB027b split-GAL4 driver to examine the responses in the a93 MBONs (MBON-a93ap and MBON-a93m) (Tanaka et al., 2008;Sejourne et al., 2011;Aso et al., 2014a) (Fig. 4A). We found that the a93 MBON response to AL stimulation is suppressed after our artificial learning paradigm (Fig. 4B,C), agreeing with in vivo results. These findings show that both the enhancement and suppression PDPs achieved in our paradigm are learning-specific and are encoded as enhancement in the presynaptic KCs and a suppression in the postsynaptic MBONs.
The 39UTR mRNA of CaMKII is important for ex vivo PDP We next turned our attention to exploring the utility of this paradigm for understanding memory formation, choosing problems at several different levels of analysis: molecular components of the memory machinery, organizational principles of the circuitry, and the role of brain state in gating plasticity. At the molecular level, previous studies have demonstrated that CaMKII is important for synaptic plasticity and memory formation in many species (Kelleher et al., 2004;Giese and Mizuno, 2013), including Drosophila melanogaster (Griffith et al., 1993;Koh et al., 1999;Ashraf et al., 2006;Malik et al., 2013;Mitchell et al., 2021). Recently, we found that the long 39UTR region of CaMKII mRNA is responsible for the activity-dependent synthesis of CaMKII in presynaptic terminals at the larval neuromuscular junction (Kuklin et al., 2017) and for the basal accumulation of axonal CaMKII protein in MB (Chen et al., 2022). We hypothesized that the loss of the 39UTR would impair associative plasticity by decreasing synaptic CaMKII protein levels and disrupting the signaling machinery that is triggered by the CS1US coincidence. We tested this idea and used our ex vivo PDP preparation to gain insight into whether this effect is upstream or downstream of the CS1US coincidence.
First, we asked whether loss of the 39UTR of CaMKII mRNA affects CaMKII levels. We used animals in which the CaMKII gene was engineered using CRISPR/Cas9 to lack the 39UTR (CaMKII Udel , Fig. 5A) (Chen et al., 2022). Immunostaining to quantify CaMKII protein levels in the MB neuropil in both WT and CaMKII UDel showed that CaMKII levels are specifically decreased in synaptic regions (Fig. 5B). We also used Western blotting to quantify total CaMKII levels and found a substantial decrease, normalized to actin, in CaMKII UDel flies compared with WT (Fig. 5C). This is in agreement with a previous study which found that mice expressing a mutant form of CaMKII lacking the 39UTR show decreased levels of CaMKII in the dendritic, but not the somatic, region of hippocampal neurons (Miller et al., 2002).
We then tested the impact of the 39UTR deletion on aversive STM. We found that CaMKII UDel flies showed a significant impairment in immediate memory performance compared with WT flies (Fig. 5D). This impairment was not because of sensorimotor dysfunctionalities as both WT and CaMKII UDel flies respond to electric shock (Fig. 5E). We then tested the effect of the CaMKII mRNA 39UTR deletion on potentiation PDP in KCs. We used the previously described AL1DA induction protocol and examined KC responses to AL stimulation both before and 15 min after either paired or unpaired induction. In the a93 compartment, although both WT and CaMKII UDel genotypes showed statistically significant PDP relative to the unpaired induction, the PDP in the WT flies was .5 times stronger than in the CaMKII UDel flies ( Fig. 5F; for detailed statistical data for the post hoc analysis, see Table 2). We also noticed that the change in signal in the a3 compartment in WT flies after paired induction was statistically different from the unpaired induction. However, this does not translate into formation of PDP in that compartment as this change was not different from zero (one- Figure 3. PDP is CS-pathway input-specific. A, Schematic of the dissected adult brain showing the placement of two electrodes (E1 and E2) onto the surface of the same AL. B, Basal calcium signals in the AL projection neurons (genotype: w-;GH146-GAL4/1;UAS-GCaMP6f/1) showing two ROIs highlighted in blue (ROI near E1) and red (ROI near E2). C, Calcium responses in each ROI to stimulation from E1 (blue trace) or E2 (red trace). D, Quantification of C. Dashed line is plotted at DF/F = 0.05. E, Same schematic as in A but with dopamine perfusion. F, Induction protocol. The ipsilateral AL is activated by 3 stimulation trains with 15 s intertrain interval followed by resting for 1 min (pre-induction). Twelve trains of stimulations to electrode 1 (E1) are then delivered, followed by a 30 s rest, then 12 trains of stimulations to electrode 2 (E2). Dopamine [10 mM] is perfused in the chamber for 60 s paired with either E1 or E2 stimulation. The brain is then rested for 15 min before being tested by 3 trains of E2 stimulation followed by 3 trains of E1 stimulation (post-induction). G, Representative trace of a prep in which dopamine perfusion is paired with E1 stimulations. H, Mean relative change of the calcium responses in the a93 compartment. Data are mean 6 SEM. Unpaired t test, two-tailed; n = 8: ***t (14) = 4.505, p = 0.0005. sample t test, p = 0.947). This is likely because of a technical difference as the AL input stimulation in this experiment was stronger (200 mAmps vs 100 mAmp in previous experiments), which led to a stronger suppression after the unpaired AL stimulation in the unpaired induction (see also below). These results show that the 39UTR mRNA of CaMKII is important for memory and its loss impairs learning-induced plasticity. It also indicates that the effect on the learning circuit is downstream of the CS1US coincidence and does not block the CS1US coincidence detection machinery itself.
Stronger inputs to the MB can recruit the a3 compartment into the STM circuit At the circuit level, it has been hypothesized that memory in flies, like humans (McClelland et al., 1995;Dudai, 2012), undergoes systems consolidation: initial potentiation of the a9b 9 lobes with a time-dependent transfer of potentiation to ab . The MB a93 and a3 compartments are adjacent to each other. Both respond to odor (Turner et al., 2008) and AL stimulation (Fig. 6B), and both use the same coincidence detector, rutabaga (Livingstone et al., 1984;Levin et al., 1992;Mao et al., 2004;Gervasi et al., 2010). This raises the question of how the two compartments are able to play different roles in the memory circuit and why only the a9b 9 cells show immediate PDP both ex vivo and in vivo. To determine whether this might be because of differences in the intrinsic properties of the two classes of cells, we examined the basal responses of both compartments to a ramp of AL stimulations without any pairings (Fig. 6A). We found that a93 axons are recruited first, with very weak AL stimulation, while a3 only starts responding at higher stimulation strengths (Fig. 6B). There was also a distinct difference in the maximum response levels: a93 compartment responses plateau at a much lower level than a3 (Fig. 6B), likely explained by the fact that the a3 compartment receives more axons than a93 (F. Li et al., 2020). Previous studies reported that odors activate a small percentage of KCs (;5%-12%), and that the odor responses in a9b 9 cells are stronger than those of ab cells (Turner et al., 2008;Inada et al., 2017). This suggests that actual odor encoding is similar to the weak AL stimulation used in our experiments (100 mAmp), recruiting only a93 but insufficient to recruit a3. We hypothesized that an increase in the AL stimulation intensity during ex vivo induction might be sufficient to recruit a3 and allow PDP in that compartment as well. We repeated both AL1AFV and AL1DA inductions, but this time with a 200 mAmp AL stimulation intensity. Indeed, under these conditions, we found that PDP occurred in both the a93 and a3 compartments (Fig. 6C,D). This result agrees with predictions from our previously published theoretical model of the associative learning circuit in which the CS1US coincidence triggers a recurrent loop between KCs and dopaminergic neurons, which increases dopamine release only onto cells receiving both inputs together (Adel and Griffith, 2021). The weak response of the a3 compartment to weak AL stimulation is subthreshold for the MB coincidence detector, and insufficient for forming PDP in a3. Stronger AL stimulation puts a3 over threshold and triggers the coincidence detector in a axons, thus increasing dopamine gain and forming PDP/memory.
Our model further predicts that there should be a certain dopamine concentration above which dopamine alone is sufficient to potentiate both a93 and a3 compartments without any need for pairing, as this hypothetical high dopamine concentration bypasses the need for CS1US coincidence detector activation of the gain control machinery that increases local dopamine release. To test this prediction, we applied either 10 mM dopamine (used in our previous experiments) or a high dopamine concentration (200 mM) to the dissected brains without pairing with AL stimulation. Indeed, 200 mM dopamine alone was sufficient to potentiate both a93 and a3 responses to AL stimulation (100 mAmps) 15 min after dopamine application (Fig. 6E). A similar effect of high dopamine was reported by Ueno et al. (2013) but interpreted differently. We show that pairing-independent plasticity resulting from very high dopamine, while possibly implemented by the same machinery responsible for associative learning, does not truly represent a memory trace as it lacks specificity to CSactivated synapses, likely representing a generalized potentiated state of the MB. In contrast, the PDP seen with lower dopamine concentrations relies on an interplay between the CS and US signals to gate dopamine gain only at the synapses activated by the CS during training.

PDP is blocked by sleep deprivation and rescued by rebound sleep
Many of the most interesting questions about memory formation are ones that involve the interaction of plasticity with the animal's internal state. To test the ability of this preparation to retain traces of previous experience and allow interrogation of memory mechanisms at this level, we looked at the effect of sleep on subsequent ex vivo plasticity. Sleep is linked to memory across phyla. In Drosophila, perturbations of sleep have been shown to impair both memory formation (Bushey et al., 2007;Seugnet et al., 2008Seugnet et al., , 2009X. Li et al., 2009;Donlea et al., 2014;Seidner et al., 2015) and   Figure 2F. C, Mean relative change of the calcium responses in the MBONs' dendrites in the a93 compartment. Data are mean 6 SEM. Unpaired t test, twotailed; n = 6: **t (10) = 3.495, p = 0.0058. Dissel et al., 2015;Donlea, 2019). To determine whether the plasticity we observe in the MB is sleep-dependent, we mechanically sleep-deprived entrained (12:12 light:dark) flies for 12 h during the ZT12-ZT24 night period before using the AL1DA paradigm to induce the artificial memory at ZT0. Control flies were housed on a different shaker in the same incubator, but the shaker was turned off. A third group of flies received shaking for only the last 2 h of the night (ZT22-ZT24) to control for the acute physical stress that may result from the shaker. Figure 7A shows a schematic of our experimental design, and Figure 7B shows the minutes of sleep for each group during the relevant time windows. We found that sleep-deprived flies did not exhibit KC enhancement PDP, while both the sleep control group and the stress control group had normal PDP (Fig. 7C).
It is important to note that the stress control we used in this study is meant to control for acute stress, not the chronic stress that flies might experience when put on a shaker for 12 h. In many studies, the control for nighttime sleep deprivation via prolonged shaking is to sleep-deprive flies during the 12 h light period. We chose not to do this for several reasons. First, it would necessitate testing the flies for PDP/memory formation at ZT12 rather than ZT0, and STM has been shown to be influenced by the clock (Lyons and Roman, 2009;Flyer-Adams et al., 2020). Second, a recent study found that sleep deprivation during the light period causes significant sleep deprivation that is discharged as rebound sleep the next day (Wiggin et al., 2020); this would invalidate using daytime shaking as a control since it produces significant sleep debt. While shaking for the last 2 h of the night is not as stressful as 12 h of shaking, it is a better control for sleep deprivation since flies start waking up naturally in that time window, which means that there is very little sleep loss (as shown in Fig. 7B).
We then asked whether allowing sleep-deprived flies a period of recovery sleep could restore their ability to learn. As little as 2 h of rebound sleep was sufficient to restore the ability to induce the same level of a93 PDP observed in the control groups (Fig.  7C). Interestingly, the 2 h minimum of rebound sleep that we found in our experiment to be required for rescuing the ex vivo plasticity is similar to the period reported in vivo to be required for sleep-deprived flies to restore the ability to form STM (X. Li et al., 2009). These findings demonstrate that our ex vivo paradigm not only forms the plasticity that is correlated with STM, but also accurately recapitulates the dynamics of the learning memory circuit and its interplay with the sleep circuit. These data also suggest that the effect of sleep deprivation on the memory circuit is downstream to sensory processing of both the CS and the US information and CS1US coincidence. This gradual recovery of the ability to form memory (Fig. 7D) may support previous indications that sleep deprivation likely impairs memory formation by downregulating dopamine receptors or other downstream molecules in the dopamine signaling pathway (Seugnet et al., 2008).

Discussion
Drosophila neural circuits are traditionally studied by relating in vivo genetic and chemical manipulations with their consequent behavioral outcomes, from which circuit information can then be inferred (Olsen and Wilson, 2008;Simpson, 2009). More  Table 2. Striped bars represent the CaMKII Udel genotype.
recently, the advent of in vivo calcium imaging allowed for tracing neural activity in actively behaving flies. Over more than a decade of such in vivo studies, the general circuit mechanisms of associative memory have been discovered, but there are limitations imposed by imaging the brain of an active intact fly (for review, see Adel and Griffith, 2021). These include the relatively low signal-tonoise ratio, the inaccessibility of multiple brain regions because of restrictions on imaging angles, the difficulty of doing acute pharmacological studies, and the possible confounds of studying the brain of a movement-restricted fly experiencing ongoing stress. Taking inspiration from the way the LTP hippocampal slice model revolutionized our understanding of mammalian memory (Bliss and Lomo, 1973), we provide here an ex vivo model of Drosophila memory which can overcome these limitations and offer a powerful preparation for studying Drosophila memory circuits. Importantly, this model provides a framework for investigating the dynamics of neural circuits in the fly brain.
Most of the previous studies investigating the associative learning circuit ex vivo have focused on mapping connectivity (Cohn et al., 2015;Barnstedt et al., 2016;Felsenberg et al., 2017Felsenberg et al., , 2018Zhao et al., 2018) or characterizing a specific biochemical pathway (Tomchik and Davis, 2009;Handler et al., 2019;Ueno et al., 2020). Only a few ex vivo studies (Wang et al., 2008;Ueno et al., 2013Ueno et al., , 2017Suzuki-Sawano et al., 2017) have focused on understanding MB circuit logic. In the LTE model developed by Ueno et al. (2013), pairing a stimulation of the CS and US pathways induced a potentiation of KC responses in the tips of the MB vertical lobes, but LTE did not fully recapitulate other characteristics of associative memory observed in intact flies. Here we develop a modified ex vivo model that resolves these discrepancies, showing that the paired activation of odor and punishment pathways induces appropriate plasticity at multiple nodes in the circuit: potentiation of KCs and suppression of MBONs. Several mechanisms for encoding those opposite forms of the plasticity have been proposed, including spike timing-dependent plasticity and activation of distinct dopaminergic receptors (for review, see Adel and Griffith, 2021). Spike timingdependent plasticity mechanisms appear less likely as MBON suppression was shown to not require MBON spiking (Hige et al., 2015). Perhaps the strongest model so far comes from Handler et al. (2019) who showed that differences in the order of KCs activation and dopaminergic input activate distinct dopaminergic receptors, DopR1or DopR2, which encode MBON suppression or potentiation, respectively. It is important to note that both Hige et al. (2015) and Handler et al. (2019) studied the plasticity in MB medial lobes, while our study focused on MB vertical lobes, so our paradigm may be useful in gaining greater mechanistic insight into this sign transformation in the vertical lobes.
We show that PDP is localized to the MB a93 compartment and not in a3, in alignment with most imaging studies in intact flies. Importantly, in this ex vivo preparation, punishment information is relayed to the MB through dopaminergic release from the PPL1 subset. Bath application of dopamine in our preparation does not interfere with the specificity of associative learning since PDP is exclusively formed in the cells that were active during the dopamine application. These data settle several inconsistencies between previous ex vivo studies (Ueno et al., 2013 and the majority of in vivo reports (Schwaerzel et al., 2003;Riemensperger et al., 2005;Schroll et al., 2006;Kim et al., 2007 genesis of the discrepancies was not because of any inherent difference between intact and ex vivo brains but was rather a consequence of technical considerations, including stimulation strength, dopamine concentration, and the sensor tools used (for a more complete discussion, see Adel and Griffith, 2021).
An ex vivo preparation that recapitulates the cardinal features of the circuits underlying associative memory formation should be useful for mechanistic studies at the molecular, cellular, and systems levels. We used our model to ask a new question about the innerworkings of the circuit at each of these levels. At the molecular level, we demonstrated the importance of normal levels of CaMKII by manipulating the 39UTR of CaMKII mRNA. Deletion of this region of the CaMKII gene drastically reduces the amount of CaMKII protein in synaptic regions and blunts the ability to form STM and to generate a potentiation PDP in KC axons. Our data argue that the role of this molecule is downstream of the CS1US coincidence detector, as we observe a much weaker PDP in CaMKII Udel flies. Whether the behavioral defect is due solely to the KC PDP defect is not completely clear since CaMKII likely has active roles at other circuit nodes (Mitchell et al., 2021).
At the cellular level, we asked why STM and PDP form in the a93 but not the nearby a3 compartment when both compartments respond to odors (Turner et al., 2008) and AL stimulation, and both receive dopaminergic input from the same PPL1 cluster . Previous work found that real odors cause activity in only 5%-12% of KCs and elicit a much higher spike rate in the a9b 9 KCs than in the ab KCs. We found that low-intensity AL stimulation (100 mAmps) elicits a stronger response in the a93 than in the a3 compartment, while high-intensity ALstimulation (200 mAmps) causes strong responses in the a3 compartment and recruits it to the learning circuit. Coupling this with our observation of lower dopamine release in a3 suggests a model in which odor presentation during associative learning causes subthreshold responses in ab cells such that the CS1US coincidence detector is not triggered, while the stronger responses in the a9b 9 cells bypass this threshold, allowing plasticity in the a9b 9 cells only. This notion is in alignment with the previous finding that a9b 9 cells have a lower firing threshold than ab cells (Inada et al., 2017). Further, It is possible that long-term memory and the enhancement memory trace in the ab KCs after repetitive space training (Yu et al.,  and ZT24 and then tested between ZT0 and ZT2 the next day; Sleep-deprived 1 Recovery, same as Sleep-deprived, but flies were allowed to recover sleep between ZT0 and ZT2 and then tested between ZT2 and ZT4; Stress Control, flies are undisturbed during the night between ZT12 and ZT22 and then the shaker was turned on between ZT22 and ZT24 before testing between ZT0 and ZT2 the next day. Testing was done by the AL1DA induction described in Figure 2F. 2006) require a gradual potentiation of the ab KC responses with every training session such that the responses bypass the coincidence detection threshold after several training sessions. Whether repetition of AL1DA pairings recruits PDP in the a3 compartment remains unclear. It is also yet to be determined whether shortcutting the circuit and recruiting ab cells in the first training session reduces the need for multiple spaced training sessions in long-term memory formation.
In conclusion, we looked at the ability of the effects of prior experience, or brain state, on the memory circuit to be retained in the ex vivo preparation. Excitingly, we found that sleepdeprived flies could not form PDP, but that as little as 2 h of rest before dissection allowed the brain to recover PDP formation. The complete abolition of PDP in sleep-deprived flies at first and the gradual recovery in plasticity afterward (Fig. 7D) suggest that sleep converges on the memory circuit upstream of the CS1US coincidence detector. Whether this involves regulation of dopamine receptors in the MB during sleep remains to be determined. The ability to retain in some functional way the internal state of the brain will allow this preparation to be used to understand how memory formation is altered by global system alterations.