key: cord-0831589-1rwjqdcj
authors: Otto, Nils; Pleijzier, Markus W.; Morgan, Isabel C.; Edmondson-Stait, Amelia J.; Heinz, Konrad J.; Stark, Ildiko; Dempsey, Georgia; Ito, Masayoshi; Kapoor, Ishaan; Hsu, Joseph; Schlegel, Philipp M.; Bates, Alexander S.; Feng, Li; Costa, Marta; Ito, Kei; Bock, Davi D.; Rubin, Gerald M.; Jefferis, Gregory S.X.E.; Waddell, Scott
title: Input Connectivity Reveals Additional Heterogeneity of Dopaminergic Reinforcement in Drosophila
date: 2020-08-17
journal: Curr Biol
DOI: 10.1016/j.cub.2020.05.077
sha: eff9df37b7bf9b42bfc4b6c896886b8be805f7ca
doc_id: 831589
cord_uid: 1rwjqdcj

Different types of Drosophila dopaminergic neurons (DANs) reinforce memories of unique valence and provide state-dependent motivational control [1]. Prior studies suggest that the compartment architecture of the mushroom body (MB) is the relevant resolution for distinct DAN functions [2, 3]. Here we used a recent electron microscope volume of the fly brain [4] to reconstruct the fine anatomy of individual DANs within three MB compartments. We find the 20 DANs of the γ5 compartment, at least some of which provide reward teaching signals, can be clustered into 5 anatomical subtypes that innervate different regions within γ5. Reconstructing 821 upstream neurons reveals input selectivity, supporting the functional relevance of DAN sub-classification. Only one PAM-γ5 DAN subtype γ5(fb) receives direct recurrent feedback from γ5β′2a mushroom body output neurons (MBONs) and behavioral experiments distinguish a role for these DANs in memory revaluation from those reinforcing sugar memory. Other DAN subtypes receive major, and potentially reinforcing, inputs from putative gustatory interneurons or lateral horn neurons, which can also relay indirect feedback from MBONs. We similarly reconstructed the single aversively reinforcing PPL1-γ1pedc DAN. The γ1pedc DAN inputs mostly differ from those of γ5 DANs and they cluster onto distinct dendritic branches, presumably separating its established roles in aversive reinforcement and appetitive motivation [5, 6]. Tracing also identified neurons that provide broad input to γ5, β′2a, and γ1pedc DANs, suggesting that distributed DAN populations can be coordinately regulated. These connectomic and behavioral analyses therefore reveal further complexity of dopaminergic reinforcement circuits between and within MB compartments.

Correspondence scott.waddell@cncb.ox.ac.uk

Otto et al. use electron microscope resolution connectomics to describe the structure and organization of neurons providing synaptic input to functionally discrete subtypes of dopaminergic neurons. The nanoscale anatomy reveals further anatomical and functional specialization of dopaminergic neurons, which is confirmed with behavioral experiments.

In adult Drosophila, anatomically discrete dopaminergic neurons (DANs) innervate adjacent compartments of the mushroom body (MB) [2] . In some cases, different combinations of DANs serve discrete roles. However, there are instances where multiple functions have been assigned to DANs that innervate the same compartment. For example, DANs innervating the g5 compartment reinforce short-term courtship memories and appetitive memories with sugar, and they also signal the absence of expected shock, to extinguish aversive memory [7] [8] [9] [10] [11] . Similarly, DANs innervating the b 0 2a compartment have roles such as controlling thirst state-dependent water seeking and water memory expression, sugar reinforcement, and hunger-dependent modulation of carbon dioxide avoidance [9, 10, [12] [13] [14] . Moreover, individual PPL1-g1pedc DANs, which innervate the g1 compartment in both hemispheres, are required to reinforce aversive memories with electric shock, high heat, and bitter taste, and also provide hunger state-dependent motivational control of sugar memory expression [5, 6, 15, 16] . For an individual DAN to multi-task it must function in different modes. However, where a compartment is innervated by multiple DANs, different neurons Report in the population could perform discrete functions, and/or the group might function together in different modes. Here we used connectomics to investigate the organization of neurons providing input to DANs innervating the g5, g1, and b 0 2a compartments to better understand how valence-specific reinforcement is generated.

Determining the Nanoscale Structure of Reinforcing Dopaminergic Neurons We used a recent EM dataset of a full adult fly brain (FAFB) [4] to identify, manually trace, and reconstruct the nanoscale anatomy of 20 DANs in the protocerebral anterior medial (PAM) cluster whose presynaptic arbors innervate the g5 compartment (PAM-g5 DANs [2] ) in the fly's right brain hemisphere (we also reconstructed 9 PAM-g5 DANs in the left MB). We identified 8 right hemisphere PAM-b 0 2a DANs and reconstructed 6. We also reconstructed the 2 protocerebral posterior lateral (PPL)1-g1pedc DANs that innervate the g1 compartments of each MB (Figures 1A and S1A-S1D; Video S1). STAR Methods and the revision status table in Methods S1 detail metrics of quality control.

We noticed when reconstructing PAM-g5 DANs that their dendrites occupied different areas of the superior medial protocerebrum (SMP) [17] , that their somata were connected via 2 neurite tracts, and that each g5 DAN had a contralateral projection crossing the midline of the brain in an upper, middle, or lower commissure ( Figures 1A, 1C -1G, S1A, S1B, and S1E-S1H). We therefore used unbiased anatomical clustering to explore suborganization of PAM-g5 DANs ( Figure 1B ). This grouped PAM-g5 DANs into 5 discrete clusters of 1-7 neurons (Figures 1C-1G) . Importantly, a different clustering criterion produced an identical result ( Figure S1I ). Although we did not trace the finest axonal branches of PAM-g5 DANs, it was evident that their major presynaptic arbors occupy different areas of the g5 compartment. We therefore named the PAM-g5 subtypes according to their defining morphological feature ( Figures 1C-1G ; Video S1).

The four reviewed PAM-b 0 2a DANs could also be clustered into three groups with commissure crossed, overall morphology, and region of compartment innervation again being the distinguishing features ( Figures 1B, 1H-1J , S1C, and S1E-S1H; Video S1). Tracing also identified a ''non-canonical'' PAM-b 0 2a DAN, which mostly innervates b 0 2a but also extends axons into g5 (Figure 1J ). The PAM-b 0 2a DAN dendrites were largely intermingled with those of PAM-g5 DANs (Figures 1C-1J ; Video S1), consistent with their roles in reinforcing appetitive memories.

Reconstructing the individual right hand PPL1-g1pedc DAN revealed that its dendrites occupy locations distinct from those of PAM-g5 and PAM-b 0 2a DANs ( Figure 1K ). This suggests that PPL1-g1pedc receives mostly different input, consistent with it signaling aversive rather than appetitive valence. The PPL1-g1pedc DAN dendrite has four major arbors that extend into orthogonal locations in the brain. Postsynapses also clustered in each of these locations ( Figures 1L, 1M , and S1H; Video S2).

Receiving branch-specific information may represent a solution for how a single PPL1-g1pedc DAN isolates and prioritizes its discrete roles in reinforcement and state-dependent control [5, 6] .

We next traced 821 neurons providing input to postsynapses identified in the dendrites of the PAM-g5, PAM-b 0 2a, and PPL1-g1pedc DANs (Figure 2A ). Since connectivity is dense and manual tracing is labor-intensive, we traced between 45% and 97% of the inputs to the postsynapses annotated on all the reconstructed DANs. We prioritized upstream tracing to retrieve a comparable coverage of inputs to each of the DAN subclasses. Sampling criteria and metrics of quality control are detailed (Methods S1).

Reassuringly, input selectivity of PAM-g5, PAM-b 0 2a, and PPL1-g1pedc DANs largely reflected the relative overlap of their dendritic fields. All the DANs receive unique inputs (Figures 2B, 2C, and S2A-S2D). However, a greater share of the 553 identified inputs to PAM-g5 DANs (181) also provided inputs to PAM-b 0 2a DANs than to the PPL1-g1pedc DAN (84). Likewise, more of the 281 PAM-b 0 2a DANs inputs connect to PAM-g5 DANs (181) than to the PPL1-g1pedc DAN (67). In contrast more of the 275 traced PPL1-g1pedc DAN inputs were unique (168) than also contacted PAM-g5 DANs (84) or PAM-b 0 2a DANs (67). Lastly, 5% (44) of the traced input neurons synapsed onto all three classes of traced DANs. These common input neurons suggest that activity in the valence-specific arms of the DAN Figure S1 and Videos S1, S2, and S4. Report system may be coordinated. It is, however, also possible that the different DANs respond in unique ways to the same input neurotransmitters.

Our sampling suggests that despite there being 20 PAM-g5 DANs and one PPL1-g1ped DAN, there are only approximately twice as many inputs to all PAM-g5 DANs compared to the PPL1-g1pedc DAN ( Figure 2B) . However, the different DAN types have markedly different weighting of inputs, assuming that synapse number correlates with input strength (Figures 2D and  S2E ). Whereas the PPL1-g1pedc DAN receives weakly connected inputs, each input to the PAM-g5 DANs or PAM-b 0 2a DANs is more strongly connected and contributes a larger proportion of the individual neuron's postsynaptic budget. Plotting a connectivity matrix for the most completely traced DANs reveals that certain groups of inputs preferentially synapse onto different PAM-g5 and PAM-b 0 2a DANs, demonstrating that all individual PAM-g5 and PAM-b 0 2a DANs have an element of input specificity ( Figures 2E  and S2F ). Nevertheless, a matrix comparing input structure between DANs reveals significant similarities in input between particular groups of PAM-g5 and PAM-b 0 2a DANs, which is more organized than random connectivity ( Figure 2F ; Methods S1). DAN clustering based on dendritic connectivity was identical using two different methods ( Figure S2G ). Moreover, clustering based on input connectivity correlated well with the prior clustering using full neuron morphology ( Figure 2H ).

The observed connectivity could be confounded by the incompleteness of reconstruction. However, plotting the synapses identified following standard and extensive review suggests that each iteration of review adds synapses that are evenly distributed across a DAN's dendritic arbor ( Figure S2H ). Nevertheless, we tested whether connectivity clustering resulted from unintentional bias in input neuron tracing by repeated clustering following random downsampling of input connectivity from 5% to 50%. Cluster content remained largely robust in these analyses with DANs clustering within the same groups across 10,000 simulations where synapses were removed (Figures S2I-S2M ). The stability of DAN clustering based on input structure and the similarity of connectivity and morphology clustering suggest that information conveyed by selective input is likely to be maintained in the activity of different DAN subtypes.

We next clustered the input neurons based on their morphology using a three-step approach (Methods S1). Using soma location and primary neurite layout revealed 20 coarse clusters ( Figures 2I and S2N ; Video S3). These could be further decomposed into 285 fine clusters distinguished by the anatomy of smaller neurites. Follow-up analyses were directed toward four classes of input neurons, for which we could postulate a functional role: MBONs; lateral horn (LH)-associated neurons that include lateral horn output neurons (LHONs); subesophageal output neurons (SEZONs), potential gustatory projection neurons that ascend from the SEZ; and OTHERS, a variety of neurons conveying information from other brain areas ( Figures  2I and 2J ).

Annotating the DAN clustering with input neuron identity revealed that the g5b 0 2a MBONs specifically provide feedback (fb) input from the MB to our previously defined PAM-g5(dd) subtype ( Figure 2K ). We therefore renamed PAM-g5(dd) neurons as PAM-g5(fb). Other MBONs also provide selective input to different DANs. In contrast, as a group the SEZONs and LHONs provide input to the PPL1-g1pedc DAN and all PAM-g5 and PAM-b 0 2a DANs, although the relative proportions vary considerably.

Toward assigning functional relevance to the identified input pathways, we used NBLAST [18] to screen traced neuronal skeletons against a collection of confocal volumes of GAL4 and split-GAL4 lines for those that potentially drive expression Percentage of shared inputs across all three groups is 16%, 15%, and 8% for PPL1-g1pedc, PAM-b 0 2a, and PAM-g5 DANs, respectively. (D) Bar chart showing DANs have many inputs with very low edge weight and each representing a small fraction of their overall postsynaptic budget. Inputs contributing more of the postsynaptic budget (to the right of the graph) are more abundant for PAM-g5 and PAM-b 0 2a DANs; PPL1-g1pedc distribution is strongly left shifted (bars show mean ± SD). (E) DANs can be clustered by input connectivity (rows correspond to F). Heatmap shows every DAN has a group of unique input neurons represented by unique blocks in each row. Clustering of DANs mostly depends on lesser number of shared inputs compressed to the left edge of the heatmap. (F) A matrix where DANs are grouped by the similarity of their input connectivity has clear structure, i.e., significantly more organized than random connectivity (comparison to null model, p < 0.0001; see Methods S1). (G) Representation of traced input neurons labeled using the unique and common input anatomy determined in (B) (see also Figures S2A-S2D) . (H) Tanglegram comparing DAN clustering by morphology (from Figure 1B) and clustering by input connectivity (left of E). Connectivity and morphology are not significantly independent of each other (Pearson's correlation between the corresponding distance matrices, r = 0.604; Mantel test, p < 10 À7 ; p w < 10 À7 within only g5 or b 0 2a group). Figure 2L ). We identified >50 driver lines with putative expression in the SEZONs, LHONs, MBONs, and OTHERS groups of DAN inputs.

Pairing odor exposure with optogenetic activation of PPL1-g1pedc or PAM-g5, PAM-b 0 2a DANs can produce either aversive or appetitive odor memories, respectively [22] . We therefore assumed that neurons providing significant input to reinforcing neurons should generate similar phenotypes if artificially engaged, instead of the relevant DANs. We combined the GAL4 driver lines with the red-light activated UAS-CsChrimson [23] optogenetic trigger and screened them for their potential to reinforce olfactory memories ( Figures 3A and S3A ). Whereas activation of some GAL4 lines produced appetitive odor memories, others produced aversive memories, and some had no consequence. We next correlated the identity of neurons labeled in each GAL4 line with their implanted memory performance and their respective DAN connectivity ( Figure 3B ; Video S3; Data S1). Figure 3A ), whereas the OTHERS15 line formed aversive memory and preferentially synapsed with PPL1-g1pedc ( Figures 3A and 3B ).

The g5b 0 2a and g4g5 MBONs both have a dendrite in the g5 compartment [3, 11] . In the stimulus replacement screen, these MBONs appeared to convey opposite valence. Whereas MBON-g5b 0 2a activation formed appetitive memory, MBON-g4g5 reinforced aversive memory ( Figure 3A ; confirmatory 30 min memory experiment in Figure S3B ). Connectivity supported these behavioral results. MBON-g4g5 is the strongest MBON input to PPL1-g1pedc, but does not connect to PAM- and PAM-g5(uc). With few exceptions, the relatively sparse connectivity of MBONs to our traced DANs was largely maintained when other traced neurons were included as potential interneurons between them ( Figure 3D ). Most notably, the apparent bias of connection of MBON-g4g5 to PPL1-g1pedc and selectivity of MBON-g3b 0 1 remained and new selective clusters to PAM-g5(fb), PAM-g5(uc), and PAM-b 0 2a groups became apparent. No indirect connections were detected between MBON-g4g5 and PPL1-g1pedc in our dataset. Moreover, only one additional b 0 2a DAN emerged downstream of MBON-g4g5 when putative indirect connectivity was considered. In contrast, although MBON-g5b 0 2a was directly connected to only the PAM-g5(fb) DANs, it was indirectly connected to most of the traced DANs, and all of the inputs to the unique PAM-g5(da) neuron come from neurons downstream of MBON-g5b 0 2 ( Figure 1G ). Axo-axonic synapses are frequent (>11,000) in the DAN input network. For example, MBONs frequently make reciprocal synapses on the axons of SE-ZONs and LHONs and the different classes of DAN input neurons are also highly interconnected within cluster (317).

Synthetic activation of the SEZON lines also produced different learning phenotypes. Activating SEZON01 neurons formed aversive memory and these connect to PPL1-g1pedc and some PAM-g5, but not to g5 DANs reinforcing sugar memory (see below). Stimulating SEZON03 neurons formed appetitive memory and these SEZONs synapse onto PAM-g5, but not PPL1-g1pedc. SEZON02 neuron activation did not implant significant memory of either valence and appears to connect weakly to all three classes of traced DANs ( Figures 3A, 3B , and S3K).

Despite their specificity, we expect some of our identified GAL4 drivers to express in our traced neurons of interest, and additional similar neurons in a fascicle. For example, a SEZON line could label a collection of ascending neurons representing both tasteful and distasteful gustatory stimuli [26] . Labeling such a mixed population with contradictory value could explain the inability of SEZON2 to reinforce a memory with clear valence. We therefore used the dominant temperature-sensitive UAS-Shibire ts1 transgene [27] to block neurotransmission from SE-ZONs during training with sugar or bitter taste reinforcement ( Figures 3E, 3F , and S3E-S3I). We included R58E02-GAL4 as a positive control for sugar memory, which expresses in the majority of PAM DANs [28], and MB320C-GAL4 for bitter learning because it labels PPL1-g1pedc [29] . Blocking R58E02 or (B) Connectivity matrix between DANs ordered according to morphological cluster identity and neurons labeled in 10 GAL4 lines, corresponding to 11 input clusters (MBON-g5b 0 2a and MBON-g4g5, SEZON01-03, LHON01-02, LHON-AD1b2, and OTHERS15-16). Valence of memory formed is reflected by input connectivity. (C) Direct MBON-DAN connectivity matrix. We identified several MBONs to provide input to specific DANs (note: we traced all inputs to 7 PAM-g5, and 2 b 0 2a DANs with extensive review; Figure S2 ; Methods S1). Numbers indicate total synapse counts between MBONs and DANs. (D) Adding other traced DAN input neurons creates potential for indirect connectivity between some MBONs and specific subsets of DANs. Indirect connectivity matrix showing the number of DAN input neurons that are downstream of MBONs with at least 3 synapses between each. Columns are normalized by their sum. (E) Olfactory learning with sucrose reinforcement. Schematic: experimental timeline and temperature shift protocol. Blocking neuron output during training abolished 30 min appetitive memory specifically in SEZON03-GAL4; UAS-Shi ts1 and R58E02-GAL4; UAS-Shi ts1 flies (mean ± SEM, p < 0.0241 and 0.0089, respectively; one-way ANOVA, with Dunnett's post hoc test, n = 10). (F) Olfactory learning with bitter (DEET) reinforcement. Schematic: experimental timeline and temperature shift protocol. Blocking neuron output during training impaired immediate aversive memory in MB320C-, SEZON01-, and SEZON02-GAL4; UAS-Shi ts1 , but not in SEZON03; UAS-Shi ts1 flies (mean ± SEM, p < 0.0009, 0.0344, and 0.0170, respectively; one-way ANOVA with Dunnett's post hoc test, n = 12). (G) Connectivity matrix to specific branches of the PPL1-g1pedc dendrite reveals classes of input neurons have branch specificity. See also Figure S3 , Data S1, and Video S3. (C) 0804-GAL4 labels 5 PAM-g5 DANs per hemisphere, previously ''g5 narrow,'' that occupy the lower commissure. Scale bar, 20 mm. (D) 0104-GAL4; UAS-GFP labels PAM-g5 DANs, previously named ''g5 broad,'' that cross the midline in the upper and middle commissures. 0104 also labels some other PAM DANs [9] . (E) Table summarizing DAN expression in GAL4 lines used for behavior, modified from [9] . R48B04GAL80 refines the 0104-GAL4 expression [12] , shown in the Data S1. (F) Aversive olfactory memory extinction. Schematic: experimental timeline and temperature shift protocol. Blocking neuron output during odor re-exposure impaired memory extinction in R58E02-, VT006202-, and 0804-GAL4; UAS-Shi ts1 , but not in MB315C-or 0104-GAL4 ± GAL80; UAS-Shi ts1 flies. Bars show mean ± SEM. Asterisks denote p < 0.035 (wild-type) and p < 0.0176 (0104); one-way ANOVA with Tukey's post hoc test, n = 10-12. Report SEZON03 neurons during training abolished 30 min memory reinforced with sugar, but blocking SEZON01 and SEZON02 neurons had no impact. In contrast, blocking MB320C, SEZON01, or SEZON02, but not SEZON03, neurons impaired 30 min memory after bitter learning. These data support a role for SEZONs in relaying positive and negative gustatory valence to DAN subtypes.

Analyzing the location of identified inputs to PPL1-g1pedc confirmed that it receives branch-specific information (Figures 3G, S3J, and S3K; Video S4). Whereas all aversively reinforcing input from SEZONs connects to the SMP arbor, OTHERS15 neurons, which can also produce aversive learning, connect to the PPL1-g1pedc arbor in the ventral and dorsal crepine (CRE) [17] . MBONs in general and the strong MBON-g4g5 input also mostly connect to the PPL1-g1pedc CRE v and CRE d branches.

Since the CRE branches are closest to the primary axon, input from other MB compartments may be particularly salient to PPL1-g1pedc ( Figure S3J ). Interestingly, the strongest input from MBON-g5b 0 2a to the PAM-g5(fb) DANs is similarly placed on the DAN dendrite ( Figure S4A ).

Based on prior findings, we hypothesized that PAM-g5(fb) DANs, which receive recurrent feedback from MBON-g5b 0 2a, would be required for memory revaluation [11] and other PAM-g5 DANs receiving input from SEZONs labeled by SEZON03 would be required to reinforce sugar memory [9] . Testing this model required locating GAL4 drivers that label g5 DAN subsets that at least partially correspond to functionally relevant subtypes. We therefore used commissure crossing to select GAL4 drivers expressing in subsets of PAM-g5 DANs [9] . We reasoned that drivers labeling the lower commissure might express in g5(fb)-DANs ( Figure 1E ) while others labeling the upper commissure could include g5-DANs connected to sugar-selective SEZONs. We identified VT006202-GAL4 ( Figure 4A ), which expresses in g5-DANs in all commissures; MB315C-( Figure 4B ) and 0804-GAL4s ( Figure 4C ), which express in 8 and 3-5 g5 DANs, respectively, in the lower commissure; and 0104-GAL4 ( Figure 4D ), which only expresses in upper commissure g5 DANs [2, 9, 21, 30] . We also used genetic intersection with GAL80 to restrict 0104-GAL4 expression to g5 broad and b 0 2 m ( Figure 4E ; [9, 12] ).

We next tested whether blocking the neurons labeled in these GAL4s with UAS-Shi ts1 disrupted aversive memory extinction and/or sugar learning ( Figures 4F and 4G ). The PAM DAN expressing R58E02-GAL4 served as control. To assay extinction of aversive memory ( Figures 4F and S4C ), flies were differentially conditioned by pairing one of two odors with shock [31]. Then 30 min after training they received two un-reinforced exposures of the previously shock paired odor (CS+) or the other odor (CSÀ) at 15 min interval [11] . They were then tested 15 min later for olfactory memory. As previously established, only CS+ re-exposure diminished memory performance, demonstrating odorspecific memory extinction. We blocked subsets of DANs specifically during odor re-exposure by training flies at permissive 25 C, transferring them to restrictive 32 C immediately after training, then returning them to 25 C after the second odor reexposure. Blocking R58E02, VT006202, and 0804 neurons abolished memory extinction, whereas memory was still extinguished in flies with blocked MB315C or 0104 neurons (+ and À GAL80). No extinction was observed in any line when flies were re-exposed to the CSÀ odor after training. These data support a role for the 0804-GAL4 group of lower commissure PAM-g5(fb) DANs in memory extinction ( Figure 4F ).

In contrast, when these neurons were selectively blocked during sugar conditioning [32] ( Figures 4G, S4D , and S4E), 30 min memory was impaired in R58E02, VT00602, and 0104 (+ and -GAL80) flies expressing UAS-Shi ts1 , but was unaffected in UAS-Shi ts1 -expressing MB315C and 0804 flies ( Figure 4G ). These data demonstrate that memory extinction and sucrose reinforcement are dissociable in the g5 DANs and support the selective role for PAM-g5(fb) DANs in memory extinction and SE-ZON-connected PAM-g5 DANs in sucrose reinforcement (Figure 4H) . We therefore propose that other PAM-g5 DAN subtypes may serve different reward-related functions.

The morphologically distinct g5 DAN subtypes innervate different regions of the g5 compartment where they could depress or potentiate different parts of the KC-MBON network (Figures 1C-1G and S4B) [29, 33-35]. We do not currently understand the full relevance of the sub-compartment architecture. However, since the g-lobe dorsal (gd) KCs carry visual information and the main g KCs are olfactory [36], connections of these two streams of KCs to g5 MBONs could be independently modified by PAM-g5(da) ( Figure 1G ) and PAM-g5(fb) DANs ( Figure 1E ), whose processes are confined to the respective subregions of the g5 compartment ( Figure S4F ). DAN stratification may therefore maintain modality specificity of olfactory memory revaluation [11] . It is interesting to note that larvae only have one DAN per MB compartment [37] and that multiple DANs per compartment are an adultspecific specialization [2] . We expect this expansion reflects the additional behavioral demands of the adult fly [13, 38] , and our work here suggests the larger number of DANs in each compartment provides additional functional capacity to the compact anatomy of the adult MB. We propose that the elaboration and specialization of g5 DANs may permit the adult fly to individually represent the values of a broad range of rewarding events.

Detailed methods are provided in the online version of this paper and include the following: Figure 1 . Open circles represent 2 g5 DANs that were identified but not further analyzed. See also Figure S4 , Data S1, and Video S4. 

We are grateful to Yaling Huang and Suewei Lin for confocal imaging during the COVID-19 shutdown in the UK. We also thank Marion Sillies, Tom Clandinin, Wes Grueber, and Wolf Huetteroth for access to confocal stacks. Members of the Waddell and Jefferis groups contributed to discussion throughout this project. This work was largely funded by a Wellcome Collaborative Award 

The new split-GAL4 Drosophila lines described in this study were produced by Masayoshi Ito. They are available on request from the Lead Contact and will be sent from the Waddell lab or from the Janelia Research Campus, via K. Ito and G. M. Rubin.

The datasets and code used for analyses in R and Python are mostly available through public repositories as indicated in this Methods section of the manuscript. Any other code is available on request and without restriction. Neuronal morphologies and connectivity data will be publicly available through VFB: https://v2.virtualflybrain.org/org.geppetto.frontend/geppetto and NeuroMorpho.org: http://neuromorpho.org/ and neuronal skeletons can be requested from the authors in .swc format.

Fly strains All Drosophila strains were raised at 25 C on standard cornmeal-agar food at 50%-60% relative humidity in [3, 37, 49] , tracing followed the centerline of a neuron's profiles through the dataset to reconstruct neurite morphology and annotate synaptic sites. We used an established and tested iterative approach [42] where initial reconstruction is followed by a systematic proofreading by at least two expert reviewers (> 500 h of tracing experience). We also took advantage of recent automatic segmentation efforts of the FAFB dataset [50], where flood-filling algorithms create volumetric segmentations of the EM data. These segmentations are then skeletonised to produce neuron fragments that can be joined together to expedite reconstruction. Human proof-reading is still required to remove incorrect merges of skeletons. In this study auto-segmentation was only used to aid the tracing of DAN input neurons to identification (see below). Synaptic sites were identified based on three, previously described criteria [51] and reviewed as above: an active zone with (1) Tbar(s) and (2) Report has been estimated that the tracing approach employed finds 99.8% of presynapse and 91.7% of postsynapses [42] . The probability of identifying false-positive postsynapses is 2.2% and negligible for presynapses.

DAN identifying, tracing and quality control DANs were first identified by selecting potential profiles in the midline commissure between left and right hemisphere MBs. These profiles were traced until axonal branches could be identified in the MB compartment of interest. We exhausted all possible profiles between the two MB compartments and in doing identified, both PPL1-g1pedc DANs, and right-hand side (RHS) b 0 2a DANs, and g5 DANs. Although the number of g5 DANs has been estimated to be between 8 and 21 [2] we nevertheless considered the possible existence of unilateral g5 DANs, which would not have a process extending across the midline. To do this we also sampled neural profiles in the descending tract where the processes of g5 DANs enter the MB lobe. However, we did not identify additional g5 DANs. Following identification DANs were traced and reviewed (as described above). Full details of tracing and review are provided in the Revision Status Table, Methods S1. In brief, the RHS PPL1-g1pedc DAN was completed and subjected to standard expert review. The g1pedc dendritic field was further extensively reviewed. 20 PAM-g5 DANs on the RHS and 9 on the left were reconstructed and the RHS neurons received standard expert review strategy as described above. 7 PAM-g5 DANs received further extensive reviewed of their dendritic field. Of the 4 PAM-b 0 2a DANs, 2 underwent both standard and additional extensive review, one only received standard review and a fourth was only partially reviewed. Any neurons that were not reviewed to this standard were excluded from the analyses. We note that it was more challenging to reconstruct DANs than many other neurons in the Drosophila brain. DAN dendrites are very thin and have a dark/granular texture, which increases the likelihood of missing branches and synapses. We therefore scrutinized completion and postsynapse annotation for 7 PAM-g5 DANs (representing all morphological clusters), 2 PAM-b 0 2a DANs and the PPL1-g1pedc DAN. Following this extended reconstruction and revision effort, we are confident that we have annotated all identifiable postsynapses on these selected DANs. Comparing data obtained from the regular review protocol to that from our extended review effort showed that regular review captured $30% of the postsynapses on more than 60% of all cable. We also analyzed the placement of old (regular review) and new (added following extensive review) synapses, by measuring their geodesic (along-the-arbor) distance to the dendritic root ( Figure S2H ). This analysis showed that each round of additional review adds new synapses that are distributed along the arbor. Lastly, we assessed whether uneven tracing of input connectivity altered the clustering of DANs by randomly downsampling (see below) the 9 extensively reviewed neurons to a level of inputs traced for the other regularly reviewed neurons. DANs could be similarly clustered following the downsampling, demonstrating that our DAN clustering results are unlikely to vary greatly with additional tracing of more input neurons.

When a postsynapse was annotated on a DAN, a single-(seminal) node profile was placed in the center of the presynaptic cell, unless a neuron or fragment was already present. To reconstruct upstream neurons from these seminal nodes we randomized the sampling order from each postsynapse within the total population on a neuron-by-neuron basis. For the reviewed PAM DANs (18 g5 and 4 b 0 2a) we typically traced over 85% of the input neurons to identification from annotated postsynapses on a DAN arbor (see Revision Status Table, Methods S1). For the PPL1-g1pedc DAN, we traced from 50% of annotated DAN postsynapses to identify the input neurons. Tracing inputs to this collection of g5, b 0 2a and g1pedc DAN postsynapses recovered 821 upstream neurons, some of which connect to multiple DANs in the traced groups. The tracing of the upstream neurons also varies in level of completeness but all neurons were traced to identify their microtubule containing backbone and were followed to a soma to retrieve their gross morphology. 

To analyze and draw conclusions from differences and similarities in large amounts of connectivity or morphology data, the information is represented in the form of distance matrices between each data point in space. Euclidean distance is the direct (bee-line) distance between two points in a Cartesian coordinate system. Manhattan distance between two data points in a Cartesian system is the sum of distances between the coordinates. Ward's clustering criterion Ward's method was used for agglomerative hierarchical clustering (part of R base package). Each datapoint starts in its own cluster and pairs of clusters are merged, moving up the hierarchy. At each step the pair of clusters with minimum within-cluster variance are merged. Connectivity data, as well as morphology data was clustered using Ward's criterion to compare to clusters formed using average linkage criterion.

Average linkage (also known as unweighted pair group method with arithmetic mean, UPGMA) is another criterion for agglomerative hierarchical clustering. With average linkage clustering pairwise dissimilarities between each element in cluster 1 and 2 are computed and the average of these dissimilarities are considered as the distance between the two clusters. Clusters separated by the smallest distance are merged during clustering, moving up the hierarchy. Both morphology and connectivity data were clustered using average linkage to compare to data clustered with Ward's criterion.

To compensate for different levels of completeness of tracing, DANs were simplified to their longest tree with 200 branch points (the minimum number of branch points throughout the PAM DAN population). Morphological similarity matrices were calculated using NBLAST [18] . Hierarchical clustering was primarily performed using base R functions, taking Euclidean distance matrices of similarity scoring, with average linkage clustering criterion. Morphology clustering was performed with Ward's and average linkage criteria for comparison.

Clustering DANs by input connectivity Connectivity information was retrieved from CATMAID after synapse annotation and upstream tracing of input neurons. Only neurons upstream of the dendritic region of DANs with > 50 sampled profiles were included in the analyses. Before clustering the number of synapses annotated on each DAN was normalized to reduce bias in clustering that could arise from the varying levels of tracing completeness and/or natural differences in the number of inputs to the different DANs. Hierarchical clustering was primarily performed using the Manhattan distance between upstream connectivity profiles of DANs with Ward's clustering criterion. Connectivity data was also clustered using the average linkage criterion for comparison. Report method measures how similar observations are to their own cluster and how dissimilar to other clusters. The average silhouette width ranges from 0 and 1, with 1 indicating observations are well clustered. To validate DAN clustering the average silhouette width was calculated using the nbclust R package [47, 53] . Input neuron morphology clustering Morphology clustering of upstream neurons was performed using hierarchical clustering with average linkage criterion. This involved a multi-step approach to account for varying levels of tracing and for the morphological diversity of 821 neurons (Methods S1).

Coarse clustering was performed taking the soma tract as the primary feature of neuron identity. Subsequently the larger primary clusters were subclustered by splitting neurons into the primary neurite and its complement/remainder. Similarity matrices were calculated using NBLAST and an element-wise mean (80:20) was used for clustering. For fine clusters, weighting methods were selected iteratively depending on overall sub-cluster morphology.

Tanglegrams to compare clustering of 2 feature spaces Tanglegrams were generated to visually compare clustering dendrograms produced by different criteria (e.g., average linkage versus Ward's) or clustering based on morphology versus those produced using connectivity. Dendrogram layouts were determined to minimize edge crossing (i.e., minimize Manhattan distance between corresponding DANs) using dendextend [40] .

Mantel test to determine dependence of 2 feature spaces The Mantel test was used to compare 2 sample spaces -here neuron morphology distance matrices obtained from all-by-all NBLAST and distances based on connectivity were used. To create distance matrices for connectivity, connectivity matrices were normalized by the postsynaptic budget of DANs. The implementation of the Mantel test was based on [54]. Pearson's correlation between the two observed datasets was calculated, then one of the matrices was shuffled 10 7 times and each event tested for correlation with the observed data. The number of events where the correlation is higher than between the two original datasets was divided by the amount of comparisons (10 7 ) to create a p value. When p values were lower than the significance level, it was concluded that the null model of independence between the two feature spaces could be rejected (see Methods S1).

To verify that clustering into the observed DAN groups does not result from bias in the relative completeness of tracing of the input network to each DAN, down-sampled the datasets so that all DANs were randomly stripped of 5%-50% of their input connections. Clustering resulting from 10,000 repetitions of this down-sampling were then compared to the clustering obtained from the full dataset. In addition, to exclude that more exhaustive reviewing of a few exemplary DANs might skew the clustering, we created a dataset of 10,000 repetitions where the connectivity of only the exhaustively reviewed DANs was reduced to the average of all the remaining DANs. Clustering obtained with this normalized dataset was also compared to that retrieved using the full dataset.

Cluster similarity analyses In 10,000 iterations of resampling a dataset with reduced connectivity, each neuron has a different likelihood to cluster with the same original group that it did in the 100% connectivity dataset or, with any of the other original groups. We therefore also calculated the average likelihood (over the 10,000 trials) that a downsampled DAN clustered with the same group that it clustered with in the full 100% connectivity dataset. These values were then plotted as a stacked bar plot.

The Fowlkes-Mallows Index measures the similarity of the content between two different clusterings. The performance of the first clustering is compared to that of the second clustering (which is assumed to be perfect). Exact matches/good performance result in an FMI = 1 [55].

To identify SEZONs and LH-associated neurons, the cable length within the respective neuropil mesh (3D bounding box) was calculated. A cut off of > 60 nodes within the defined neuropil region was required for classification.

Dendrogram representations of neurons were created as in [11] . Dendrograms are 2D representations of 3D neuronal reconstructions which preserve the topology of neuron and visualize specific synapses on specific branches. The neato layout (Graphviz, https:// graphviz.gitlab.io/ [43]) attempts to minimize a global energy function, equivalent to statistical multi-dimensional scaling to represent the neuron morphology as a graph. Code available (https://github.com/markuspleijzier/AdultEM/tree/master/Dendrogram_code) using the Graphviz library with Python bindings provided by NetworkX, (https://networkx.github.io/ [44]).

Euclidean distances between MBON-g5b 0 2a postsynapses and the closest DAN presynapse were measured and marked with the identity of the morphological DAN clusters (Figure 1 ). The Euclidean distances were then thresholded to within 2 mm and the synapses ll OPEN ACCESS Current Biology 30, 3200-3211.e1-e8, August 17, 2020 e6 Report identified to be under that threshold were plotted on a neato dendrogram. The plot in Figure S4F therefore shows all postsynapses within a 2 mm diffusion distance from a dopaminergic presynapse.

Edge weight distributions describe how many upstream neurons contribute a given number of presynapses to a connection with a postsynaptic neuron (frequency versus number of synapses). Normalizing by the total number of postsynapses details the percentage of the total postsynaptic budget a given number of synapses represents. For example, if a neuron makes 10 presynapses onto a postsynaptic neuron, which has a total of 100 postsynapses, then that upstream neuron contributes 10% of the postsynaptic budget.

Identified MBONs were collapsed by type. The number of synapses between DANs and MBONs was normalized by the number of all DAN-MBON connections of the given DAN. Connectivity matrices can be calculated for single branches of a neuron after defining the relevant branchpoints in CATMAID. For the PPL1-g1pedc DAN, we manually split the dendrite into 4 postsynaptic clusters, as defined from cluster analyses, and recorded the specific connectivity to each of these clusters/branches.

A connectivity similarity score between 2 DANs was defined as one minus half of the Manhattan distance between their normalized connectivity patterns (normalized connectivity patterns of DANs shown in Figure 2E ).

Statistical analysis of DAN connectivity -comparison to a null model of random connectivity A DAN input connectivity matrix was first randomized 10 4 times, respecting both DAN postsynaptic budget and input neuron presynaptic budget, so that after randomization each row sum and each column sum remained the same as in the observed data (i.e., each DAN gets the same number of inputs and each input neuron has the same number of outputs). Then the Manhattan distance between upstream connectivity profiles of DANs in the observed data and those in simulated random matrices, both normalized by DAN postsynaptic budget were calculated and means of these distances were compared to obtain a p value describing the similarity of these means. A p value lower than the significance level concluded that the null model of randomized connectivity could be rejected (see Methods S1).

The odors used for US substitution, sucrose learning and DEET learning experiments were 10 À3 dilutions of 3-octanol (OCT) and 4methylcyclohexanol (MCH) in mineral oil. For extinction experiments odor concentrations of 10 À6 were used to avoid pre-exposure effects [11] . Experiments were performed at 23 C and 55%-65% relative humidity, except for electric shock learning which occurred at 70% relative humidity.

In both the behavioral screen and follow-up experiments, neurons were artificially activated to substitute for an unconditioned stimulus in the training chamber of a T-maze. Prior to the experiments, 80-120 1-5 day old mixed sex flies were housed on standard food supplemented with 1% all-trans-Retinal for 3 days before a 20 -28 h starvation period in vials containing 2 mL 1% agar as a water source and a 2x4 cm strip of Whatman filter paper. During training, groups of flies were exposed to the CS-for 2 min followed by 30 s rest with fresh air, then 2 min of CS+ odor with optogenetic activation of the genetically encoded Channel Rhodopsin with red light exposure. Three red (620-630nm) LEDs (Multicomp, p/n OSW-4338) with 3 W maximum power were mounted on the training arm of a T-maze and 1ms pulses were driven at 1.2V with a stimulation frequency of 500Hz, which is flicker free red-light that flies cannot see. For screening, immediate memory testing followed. Flies were transferred back into their starvation vials after training before testing 30 min memory.

Appetitive olfactory learning with sucrose reward Prior to the experiments, 80-120 3-8 day old mixed sex flies were starved for 20-28 h in vials containing 2 mL 1% agar as a water source and a 2x4 cm strip of Whatman filter. Flies were transferred to 32 C 30 min before training. During training, groups of flies were exposed to the CS-odor with dry paper for 2 min followed by 30 s of fresh air, then 2 min of CS+ odor exposure with dry sugar paper. Flies were either tested immediately after training or were transferred back into 25 C starvation vials after training prior to testing 30 min memory.

Aversive olfactory learning with bitter reinforcement Flies were aversively trained with DEET as previously described [16] . In brief, prior 80-120 3-7 day old mixed sex flies were starved for 20-24 h in vials containing 2 mL 1% agar and a 2x4 cm strip of Whatman filter paper. Training and immediate testing were performed at 32 C. During training groups of flies were exposed to the CS-odor with 1% agar on filter paper for 2 min followed by 30 s fresh air, then 2 min of CS+ odor with 0.4% DEET, 3 M xylose and 100 mM sucrose in 1% agar on filter paper. Flies were tested for their odor preference immediately after training.

e7 Current Biology 30, 3200-3211.e1-e8, August 17, 2020 Report Aversive memory extinction Extinction memory was tested as described [11] . In brief, mixed sex groups of 80-120 flies were transferred into vials with 2 mL cornmeal medium and a 2x4 cm strip of Whatman paper for 18-26 h before training. Aversive olfactory conditioning in the T-maze was conducted as previously described [31] . Flies were exposed to the CS+ odor for 1 min paired with twelve 90 V electric shocks at 5 s intervals. Following 45 s of clean air, the CS-odor was presented for 1 min without shock. Immediately after training flies were transferred to 32 C. 30 min later flies were re-exposed twice to either the CS-or CS+ odor with a 15 min interval. Flies were then returned to permissive 23 C and tested 15 min later for memory performance.

To test memory performance flies were loaded into the T-maze and transported to the choice point where they were given two min to choose between the CS+ and CS-odors in the dark. A Performance Index was calculated as the number of flies in the CS+ arm minus the number in the CS-arm, divided by the total number of flies [31] . MCH and OCT, were alternately used as CS+ or CS-and a single sample, or n, represents the average performance score from two reciprocally trained groups. Statistical analysis was carried out with GraphPad, Prism (v8.1). All experiments were analyzed with a one-way ANOVA. Silhouette Method to determine k in Figure 1L 1a  5a  4a  3a  2a   1a  5a  4a  3a  6a  1b  5b  4b  3b  2b 1b 5b 4b 3b 6b

Training 

Do the right thing: neural network mechanisms of memory formation, expression and update in Drosophila

The neuronal architecture of the mushroom body provides a logic for associative learning

Mushroom body output neurons encode valence and guide memory-based action selection in Drosophila

A complete electron microscopy volume of the brain of adult Drosophila melanogaster

A neural circuit mechanism integrating motivational state with memory expression in Drosophila

Specific dopaminergic neurons for the formation of labile aversive memory

Dopamine neurons modulate pheromone responses in Drosophila courtship learning

Persistent activity in a recurrent circuit underlies courtship memory in Drosophila

Sweet taste and nutrient value subdivide rewarding dopaminergic neurons in Drosophila

Distinct dopamine neurons mediate reward signals for short-and long-term memories

Integration of parallel opposing memories underlies memory extinction

Neural correlates of water reward in thirsty Drosophila

A neural mechanism for deprivation state-specific expression of relevant memories in Drosophila

3209 Report integration of conflicting sensory information in Drosophila

Converging circuits mediate temperature and shock aversive olfactory conditioning in Drosophila