We identified the neurons comprising the Drosophila mushroom body (MB), an associative center in invertebrate brains, and provide a comprehensive map describing their potential connections. Each of the 21 MB output neuron (MBON) types elaborates segregated dendritic arbors along the parallel axons of ∼2000 Kenyon cells, forming 15 compartments that collectively tile the MB lobes. MBON axons project to five discrete neuropils outside of the MB and three MBON types form a feedforward network in the lobes. Each of the 20 dopaminergic neuron (DAN) types projects axons to one, or at most two, of the MBON compartments. Convergence of DAN axons on compartmentalized Kenyon cell–MBON synapses creates a highly ordered unit that can support learning to impose valence on sensory representations. The elucidation of the complement of neurons of the MB provides a comprehensive anatomical substrate from which one can infer a functional logic of associative olfactory learning and memory.https://doi.org/10.7554/eLife.04577.001
One of the key goals of neuroscience is to understand how specific circuits of brain cells enable animals to respond optimally to the constantly changing world around them. Such processes are more easily studied in simpler brains, and the fruit fly—with its small size, short life cycle, and well-developed genetic toolkit—is widely used to study the genes and circuits that underlie learning and behavior.
Fruit flies can learn to approach odors that have previously been paired with food, and also to avoid any odors that have been paired with an electric shock, and a part of the brain called the mushroom body has a central role in this process. When odorant molecules bind to receptors on the fly's antennae, they activate neurons in the antennal lobe of the brain, which in turn activate cells called Kenyon cells within the mushroom body. The Kenyon cells then activate output neurons that convey signals to other parts of the brain.
It is known that relatively few Kenyon cells are activated by any given odor. Moreover, it seems that a given odor activates different sets of Kenyon cells in different flies. Because the association between an odor and the Kenyon cells it activates is unique to each fly, each fly needs to learn through its own experiences what a particular pattern of Kenyon cell activation means.
Aso et al. have now applied sophisticated molecular genetic and anatomical techniques to thousands of different transgenic flies to identify the neurons of the mushroom body. The resulting map reveals that the mushroom body contains roughly 2200 neurons, including seven types of Kenyon cells and 21 types of output cells, as well as 20 types of neurons that use the neurotransmitter dopamine. Moreover, this map provides insights into the circuits that support odor-based learning. It reveals, for example, that the mushroom body can be divided into 15 anatomical compartments that are each defined by the presence of a specific set of output and dopaminergic neuron cell types. Since the dopaminergic neurons help to shape a fly's response to odors on the basis of previous experience, this organization suggests that these compartments may be semi-autonomous information processing units.
In contrast to the rest of the insect brain, the mushroom body has a flexible organization that is similar to that of the mammalian brain. Elucidating the circuits that support associative learning in fruit flies should therefore make it easier to identify the equivalent mechanisms in vertebrate animals.https://doi.org/10.7554/eLife.04577.002
Neural representations of the sensory world give rise to appropriate innate or learned behavioral responses. Innate behaviors are observed in naïve animals without prior learning or experience, suggesting that they are mediated by genetically determined neural circuits. Responses to most sensory stimuli, however, are not innate but experience-dependent, allowing an organism to respond appropriately in a variable and uncertain world. Thus, most sensory cues acquire behavioral relevance through learning. In Drosophila melanogaster, a number of different forms of learning have been observed in response to sensory stimuli (Siegel and Hall, 1979; Liu et al., 1999, 2006; Masek and Scott, 2010; Schnaitmann et al., 2010; Ofstad et al., 2011; Vogt et al., 2014). In associative olfactory learning, exposure to an odor (conditioned stimulus, CS) in association with an unconditioned stimulus (US) results in appetitive or aversive memory (Quinn et al., 1974; Tempel et al., 1983; Tully and Quinn, 1985). Olfactory memory formation and retrieval in insects require the mushroom body (MB) (Heisenberg et al., 1985; de Belle and Heisenberg, 1994, Dubnau et al., 2001; McGuire et al., 2001), an associative center in the protocerebrum (Figure 1 and Video 1).
Olfactory perception in the fly is initiated by the binding of an odorant to an ensemble of olfactory sensory neurons in the antennae, resulting in the activation of a distinct and topographically fixed combination of glomeruli in the antennal lobe (Figure 1A,B; reviewed in Vosshall and Stocker (2007); Masse et al. (2009)). Most antennal lobe projection neurons (PNs) extend dendrites to a single glomerulus and project axons that bifurcate to innervate two brain regions, the lateral horn and the MB (Stocker et al., 1990; Wong et al., 2002; Jefferis et al., 2007). The invariant circuitry of the lateral horn is thought to mediate innate behaviors, whereas the MB translates olfactory sensory information into learned behavioral responses (Heisenberg et al., 1985). The PN axons synapse onto the dendrites of the Kenyon cells (KCs) in the MB calyx; the parallel axons of the KCs form the MB lobes. Odors activate sparse subpopulations of KCs distributed across the MB without spatial preference (Turner et al., 2008; Honegger et al., 2011; Campbell et al., 2013). Anatomical and physiological studies reveal that each KC receives on average 6.4 inputs from a random combination of glomeruli; that is, knowledge of a single input to a KC provides no information about the identity of the additional inputs, and connections differ in different flies (Murthy et al., 2008; Caron et al., 2013; Gruntman and Turner, 2013). Thus, the calyx of the MB discards the highly ordered structure of the antennal lobe. A restoration of order must therefore be imposed downstream to link the KC representation to an appropriate behavioral output.
Three classes of KCs extend parallel fibers that form the γ, α′/β′, and α/β lobes of the MB, where they form synapses with a relatively small number of MB output neurons (MBONs; Figure 1C) (Crittenden et al., 1998; Ito et al., 1998; Strausfeld et al., 2003; Lin et al., 2007; Tanaka et al., 2008; Busch et al., 2009). The MBONs have dendrites in the MB lobes and project axons to neuropils outside of the MB. Modulatory input neurons, including dopaminergic neurons (DANs) and octopaminergic neurons (Nassel and Elekes, 1992; Tanaka et al., 2008; Busch et al., 2009; Mao and Davis, 2009), also innervate the MB lobes. The MBONs and DANs send their processes to stereotyped locations, defining spatially restricted ‘subdomains’ in each lobe (Ito et al., 1998; Tanaka et al., 2008; Mao and Davis, 2009; Pech et al., 2013). However, these studies did not establish the precise anatomical relationships between the subdomains; knowledge of these relationships will be required to understand the structure and logic of MB circuits.
The DANs are the most prevalent modulatory neurons in the MB and dopamine is thought to act locally to modify KC–MBON synapses (Aso et al., 2010; Waddell, 2013). In accord with this model, DAN activity is required during learning (Schwaerzel et al., 2003; Aso et al., 2010, 2012; Burke et al., 2012; Liu et al., 2012) and exogenous activation of DAN subpopulations can serve as an US in associative learning paradigms (Schroll et al., 2006; Claridge-Chang et al., 2009; Aso et al., 2010, 2012; Burke et al., 2012; Liu et al., 2012). In addition, D1-like dopamine receptors in the KCs are necessary to form olfactory memories (Kim et al., 2007).
Different populations of DANs are activated by USs of different valence; see Figure 1A of the accompanying paper (Aso et al., 2014) for summary (Riemensperger et al., 2005; Mao and Davis, 2009; Liu et al., 2012; Das et al., 2014). Genetic manipulation has also implicated specific subsets of MBONs in the mediation of learned appetitive and aversive behaviors (Sejourne et al., 2011; Pai et al., 2013; Placais et al., 2013; Aso et al., 2014). These experiments implicate the DANs as the source of the learning cue and the MBONs as the mediators of behavioral output. The elucidation of the connections between KCs, DANs, and MBONs should provide insight into a problem shared by invertebrate and vertebrate nervous systems: how is meaning imposed on an unstructured ensemble of neurons and how is imposed valence translated into an appropriate behavioral response?
In this study, we developed new genetic reagents and used them to identify the cell types and projections of the neurons comprising the MB lobes. These data provide insight into the potential connections in the MB and suggest how the MB may mediate learned behaviors. We found that the MB lobes are composed of ∼2200 neurons that include 7 KC, 21 MBON, and 20 DAN cell types. The MBONs of a given type exhibit spatially stereotyped dendritic arbors in the MB lobes that form 15 compartments that collectively tile the lobes. Each DAN cell type projects axons to one or at most two of the compartments defined by the MBONs. The alignment of DAN axons with compartmentalized KC–MBON synapses creates an isolated unit for learning that can transform the disordered KC representation into ordered MBON output. The MBON axons project to five discrete neuropils outside of the MB, providing loci for the convergence of all the information necessary for learned associative responses. The elucidation of the full complement of MB neurons and the details of their projections provide an anatomical substrate from which we can infer a functional logic of olfactory learning and memory.
We designed a genetic approach to examine the architecture of the MB circuit and identified most, if not all, of the neurons innervating the MB lobes. We screened adult brains from 7000 GAL4 lines driven by known enhancers (Pfeiffer et al., 2008; Jenett et al., 2012) to identify lines containing neurons that innervate the MB lobes. These GAL4 drivers typically label many other neurons, making it difficult to disambiguate the projection patterns of the labeled MB neurons. We therefore identified lines with overlapping expression patterns for the MB neurons and used the split-GAL4 strategy (Luan et al., 2006; Pfeiffer et al., 2010) to identify lines with more restricted expression in the MB (Figure 2). After screening 2500 such intersections, we obtained more than 400 split-GAL4 combinations that had strong expression in the MB neurons. Most split-GAL4 lines that drive expression in MBONs or DANs contain a small number of neurons that share virtually identical morphologies and exhibit bilateral symmetry, and this profile is maintained across different individuals (Figure 2—figure supplements 3–6 and data not shown). The split-GAL4 lines that label the KC contain far greater cell numbers (75–600) (Figure 2—figure supplement 2). These split-GAL4 lines allowed us to classify the MB neurons into cell types (Figure 3 and Videos 2–4). We operationally define a cell type as a single neuron (per hemisphere) or a group of neurons that is not further subdivided in any of the 7000 GAL4 lines. Moreover, neurons within most cell types exhibit indistinguishable morphology. Importantly, we identified on average 12 GAL4 drivers that label the same cell type within the set of 7000 lines, indicating that the screen was near saturation in our large GAL4 collection.
We employed an independent approach, photoactivatable GFP (PA-GFP) tracing (Patterson and Lippincott-Schwartz, 2002; Datta et al., 2008; Ruta et al., 2010), to verify the results of the split-GAL4 experiments and determine whether the neuron types identified in our screen represent the full complement of MBONs and DANs. Photoactivation of the MB labels all neurons that express PA-GFP and project to the MB (Figure 4A; for limitations, see ‘Materials and methods’). Photoactivation of the MB lobes in flies expressing PA-GFP pan-neuronally (except in the KCs) resulted in labeling of eight individual neurons and five clusters of neuronal cell bodies (Figure 4, see ‘Materials and methods’). The number and position of these PA-GFP labeled neurons matched well with the cells identified in the split-GAL4 lines.
We performed a more refined analysis by photoactivating individual subdomains of the MB lobes (Figure 5). By labeling processes of specific MBONs or DANs, we could decorate individual subdomains of the MB lobes (Figure 1C and see below) (Tanaka et al., 2008), allowing focal photoactivation and subsequent identification of the full complements of neurons innervating each lobe subdomain. Photoactivation of individual subdomains confirmed the results obtained from the genetic approach (Figure 5), but the photoactivation experiments also revealed two MBONs not identified in the split-GAL4 lines. These MBONs were subsequently identified in the VT-GAL4 collection, allowing us to characterize their projections (Figure 5—figure supplement 1).
The split-GAL4 approach identified 20 DAN types of the PPL1 and PAM clusters that innervate the MB. We identified about 30% less DANs in the PAM cluster in our collection of split-GAL4 lines compared to the number estimated by PA-GFP and anti-dopamine immunoreactivity (Figure 5) (Liu et al., 2012). However, these additional DANs exhibit innervation patterns similar to those of the split-GAL4 lines (see below), and therefore we assume that they either represent closely related cell types or that some of our split-GAL4 drivers are stochastic in their expression and fail to label all members of a cell type. Taken together, these data indicate that each cell type defined by our criteria likely represents an irreducible group of equivalent cells, and that the split-GAL4 screen and PA-GFP tracing identified perhaps all of the neurons in the MB lobes.
These complementary analyses allowed us to make a comprehensive list of cell types comprising the MB lobes (Table 1). We selected 92 split-GAL4 lines representing the best examples for single cell types as well as combinations of related cell types; these split-GAL4 lines will facilitate further anatomical and functional characterization of the MB cell types (see Figure 2—figure supplements 1–6, Supplementary file 1, www.janelia.org/split-gal4, and ‘Materials and methods’). In this study, we focus on the three major classes of neurons that provide the input and output of the MB lobes: 7 types of KCs, 21 types of MBONs, and 20 types of DANs (Figure 3 and Videos 2–4).
Each MB contains ∼2000 KCs that are sequentially generated from four neuroblasts (Ito et al., 1997; Lee et al., 1999; Zhu et al., 2003; Lin et al., 2007). The dendrites of the KCs form the MB calyx and their parallel axons form the three MB lobes (Figure 1) (Crittenden et al., 1998). The main calyx primarily receives olfactory input from the antennal lobe, whereas the smaller ventral and dorsal accessory calyces are thought to receive non-olfactory input (see Figure 1C) (Tanaka et al., 2008; Butcher et al., 2012). The KCs have been divided into three classes, γ, α′/β′, and α/β, with each class projecting axons to the eponymous lobe (Crittenden et al., 1998; Lee et al., 1999). The split-GAL4 screen and the analysis of the axonal projection patterns of single cells revealed that these three classes of KCs divide into seven cell types (Figure 3A and Video 2). Each of the four neuroblasts contributes to each of the seven cell types and the dendrites of the KCs generated from the different neuroblasts remain segregated in the main calyx (Lin et al., 2007). The parallel axon fibers of each of the seven types of KCs occupy specific layers within the γ, α′/β′, and α/β lobes (Figures 6 and 7). Two KC types divide the γ lobe into the main and dorsal (d) layers, two types divide the α′/β′ lobe into the middle (m) and anterior–posterior (ap) layers, and three KC types divide the α/β lobe into the posterior (p), core (c), and surface (s) layers (Figures 6 and 7, also see Figure 1C). Examination of single cell morphologies suggests that each KC may form en passant synapses with target MBONs along the length of its axon, providing each MBON with access to a large number of KC inputs. Five of the seven types of KCs elaborate their dendrites in the main calyx, whereas two types of KCs (γd and α/βp) have dendrites exclusively in the ventral and dorsal accessory calyces, respectively (Figures 6A and 7A) (Lin et al., 2007; Tanaka et al., 2008; Butcher et al., 2012).
The five KC types (γmain, α′/β′ap, α′/β′m, α/βc, and α/βs) that receive olfactory information are each represented by hundreds of neurons per hemisphere and have their dendrites in the main calyx. Each KC cell type sends axonal projections to a spatially segregated layer in the lobes. The dendritic arbors of each KC type also tend to be found in the same regions of the calyx (Lin et al., 2007; Leiss et al., 2009), but those dendritic zones are largely overlapping and individual KCs within a given cell type exhibit variable dendritic projection patterns (Figures 6 and 7). Moreover, the KCs receive input from an apparently random collection of glomeruli (Murthy et al., 2008; Caron et al., 2013; Gruntman and Turner, 2013). These features are in sharp contrast to most neuronal cell types in the olfactory pathway of the fly that are thought to consist of one to ten neurons that exhibit stereotyped projections (Yu et al., 2010), suggesting that their input and output connections are genetically predetermined. These observations suggest a unique function of the KCs in the processing of olfactory information (see ‘Discussion’).
The MBONs extend dendrites that overlap with the KC axons in the MB lobes and project axons outside the MB. By determining the polarity of each cell type using high-resolution confocal imaging along with an analysis of the expression of the presynaptic reporter synaptotagmin-smGFP-HA (Syt::smGFP-HA; Figure 8), we identified 34 MBONs that comprise 21 different cell types (Table 1, Figure 3B, Video 3). We employed immunostaining to identify MBON types as either cholinergic, glutamatergic, or GABAergic (Figure 9, Table 1). MBONs that use the same neurotransmitter extend dendrites to adjacent regions of the lobes; cholinergic MBONs in the vertical (α and α′) lobes, glutamatergic MBONs in the medial (β, β′, and γ) lobes, and GABAergic MBONs in an area of the lobes at the intersection between these two regions (Figure 9 and Video 5).
Fourteen MBON cell types consist of only one cell per hemisphere, six types contain two cells, and one type eight cells per hemisphere. In split-GAL4 lines with expression in more than one neuron, single-cell resolution was achieved by using the multicolor flp-out strategy (MCFO; Nern et al., in preparation, Figure 8). Single cell analysis revealed that each member of an MBON type exhibits indistinguishable morphology as assessed by light microscopy, and these stereotyped projection patterns are invariant across flies (see below for all cell types).
The 21 MBON types elaborate dendritic arbors in insular, segregated domains of the lobes that we call compartments. MBON dendritic arbors within each compartment exhibit little, if any, overlap with arbors in neighboring compartments (Figure 10). Computational alignment of the dendritic arbors of each of the MBON types within a single reference brain revealed that these compartments collectively tile the MB lobes with minimal overlap (Figure 10G,I,K). The alignment reveals gaps between arbors at four compartment borders; staining of the MB lobes for the presynaptic marker Bruchpilot (Figure 10—figure supplement 1) suggests that these gaps represent areas of reduced synaptic density. Two-color labeling experiments confirmed that the dendritic arbors of different MBONs are segregated in spatially stereotyped compartments (Figure 11A–C). We observed ensheathing glia at the borders between the MB lobes but not between the MBON compartments in each lobe (Figure 11J–L).
The MB lobes are divided into 15 distinct compartments containing the segregated dendritic arbors of one or a small number of MBONs (Figure 1C, Figure 12 and Figure 13). These compartments tile the MB lobes, revealing a general organizational principle of the MB output. This organization is in accord with an earlier proposal by Tanaka et al. (2008) that each of the γ, α′/β′, and α/β lobes is divided into five domains. 13 of the 21 MBON types extend dendrites to a single compartment, and 8 MBON types project to two compartments (Figure 12A–C). Most of the MBON types innervate KCs from each of the layers within a compartment, but eight types restrict their dendritic arbors to specific layers (Figure 11D, 12A,C,D).
The identification of the full complement of 21 MBON types highlights the extensive convergence of 2000 KCs onto just 34 MBONs, a number even smaller than the number of glomeruli in the AL. Thus, the high-dimensional KC representation of odor identity is transformed into a low-dimensional MB output. This suggests that the MBONs do not represent odor identity but instead provide a representation that may bias behavioral responses (see ‘Discussion’).
Two clusters of dopaminergic neurons (PPL1 and PAM) have previously been shown to project axon terminals to specific regions within the MB lobes and transmit information about reward and punishment to the MB to guide learning (Schwaerzel et al., 2003; Claridge-Chang et al., 2009; Mao and Davis, 2009; Aso et al., 2010, 2012; Burke et al., 2012; Liu et al., 2012). Our split-GAL4 screen identified over 100 DANs of 20 types (Figure 3C, Table 1 and Video 4). Each DAN type contains a small number of neurons: DAN types from the PPL1 cluster contain one or two cells per hemisphere and DAN types from the PAM cluster contain up to ∼20 cells per hemisphere (Table 1, see Figures 14–16 for each cell type, see ‘Materials and methods’ for classification).
The axon terminals of the DANs project to specific compartments and, similar to the dendrites of MBONs, tile the entire MB lobes (Figure 10H,J,L, 14–16 and Video 4). 17 of the 20 DAN types project to only a single compartment (Figure 12). Two-color labeling and multicolor flp-out of DANs innervating neighboring compartments show clear segregation of their axon termini (Figure 11E,F and data not shown). Two-color labeling experiments revealed overlap of the axon termini of DAN types and the dendritic arbors of cognate MBON types that innervate the same compartment (Figure 11G–I). Computational alignment to a single reference brain extends these observations to all DAN types and further demonstrates that the DAN axon termini tile the MB lobes (Figure 10H,J,L). Thus, different MBON types have access to largely equivalent input from the KCs but are modulated by different DANs (Figures 12 and 13). In classical learning paradigms, different DAN types respond to different unconditioned stimuli (US) (Riemensperger et al., 2005; Mao and Davis, 2009; Burke et al., 2012; Liu et al., 2012). The compartmental organization we observe suggests that the DANs may convey information about the US to specific MBON types. Dopamine release in specific compartments may modify local KC–MBON synapses to bias behavioral output.
We identified three MBONs and one DAN type that appear to interconnect different compartments of the MB lobes (Figure 12A, Figure 14—figure supplement 1E, and Figure 17A–D). Each of these three MBONs projects to the compartments of other MBONs but not back to the compartment occupied by its own dendrites. Thus information flows in one direction, and these MBON connections could create a multi-layered feedforward network (Figure 17J, see ‘Discussion’). 12 of the MBON types receive input from the KCs but not from other MBONs and therefore read out KC activity as a single output layer. Most of the MBONs having dendrites in the α′/β′ and γ lobes provide such a single-layer readout. In contrast, the glutamatergic MBON whose dendrites arborize in the γ4 compartment projects axons into the γ1 and γ2 compartments as well as to neuropils outside the MB (Figure 17A,G–I). Thus, the γ1 and γ2 MBONs have access to direct KC input as well as to the input provided by the γ4 MBON. The outputs of these two γ lobe compartments thus reflect a two-layer feedforward network. The readout of the α/β lobe is even more elaborate because the γ1 MBON projects to all of the compartments of the α/β lobe (Figure 17C,E,F), and the β1 MBON projects to the α1, α2, and α3 compartments (Figure 17B). The α1, α2, and α3 MBONs thus represent a four-layer feedforward network with access to the KC activity of the α/β and γ lobes, both directly and indirectly through MBON intermediaries (Figure 17J).
The DAN cell type that projects axons to the γ4 compartment (Figure 17D) provides another form of communication between compartments. This cell type extends dendrites to the γ1 and γ2 compartments, where they appear to receive direct inputs from the axonal termini of MBON-γ4>γ1γ2 (Figure 17G–I), creating a recurrent loop involving modulatory dopamine input.
The MBON axons that innervate other compartments within the MB lobes have access to dopamine inputs and thus can potentially be modified by learning. Adaptive multi-layer or ‘deep’ feedforward networks are known to be capable of more complex readout functions than single-layer readouts (Bishop, 2006). Thus, the four-layer α lobe readout system may support more sophisticated neuronal computations than the one- and two-layer γ or α′/β′ lobe systems.
The 34 MBONs represent the sole outputs of the MB lobes. Computational alignment of single-cell images allowed us to localize the axon termini of the MBONs outside the MB. The axon terminals of the MBONs (Figure 18) converge onto five neuropils: the crepine (CRE; a region surrounding the horizontal/medial lobes), the superior medial protocerebrum (SMP), the superior intermediate protocerebrum (SIP), the superior lateral protocerebrum (SLP), and the lateral horn (LH) (Figure 18E). Most MBON types project axons to several of these neuropils (Figure 18—figure supplement 1). Within a neuropil, the axon terminals of each MBON type exhibit distinct and confined projection patterns, suggesting that different MBON types may synapse onto different neurons (Figure 19, also see Figures 14–16 for individual MBON projection patterns). However, we also observed significant overlap between the axon terminals of different MBONs, suggesting that in certain cases MBONs may converge onto the same post-synaptic target neuron (Figure 20A,D,E). These results, obtained by computational alignment, were confirmed for a subset of MBONs by two-color labeling experiments (Figure 20J–L; Video 6).
Computational alignment of single cell images of each DAN type identified the dendritic arbors of the DANs outside the MB (Figure 18D). Interestingly, 90% of the dendritic arbors of the DANs reside in four of the neuropils targeted by the MBONs: CRE, SMP, SIP, and SLP (Figures 18E and 19). Because DANs are activated by unconditioned stimuli (US), these neuropils are likely targets of the US input (see ‘Discussion’). We observed cases where the dendrites of different DAN types in a single neuropil overlap (Figure 20B,F,G), suggesting that they may share upstream input. Furthermore, we observed overlap between MBON terminals and DAN dendrites within each neuropil, implying that some MBONs may synapse on specific DANs (Figure 20C,H,I). In some cases, the axon terminals of an MBON overlap with the dendrites of the DAN type innervating the same compartment (Figure 20H). This may provide feedback that regulates dopamine release and hence learning. In other cases, axon termini of MBONs from one compartment overlap with the dendrites of DAN types projecting to different compartments (Figure 20I). This provides a pathway through which the activity of one MBON could modulate the synapses between KCs and MBONs of other compartments.
The MBONs are thought to elicit learned behavioral responses, and it is therefore of interest that four MBON types project to the lateral horn (Figure 18—figure supplement 1), a brain region responsible for innate odor responses (Heisenberg et al., 1985). These cholinergic MBONs may exploit LH neurons capable of eliciting innate behaviors to generate learned olfactory responses (Sejourne et al., 2011). This connection may also serve to modulate innate behavioral responses as a consequence of learning. Most of the MBON outputs, however, do not target the LH but converge onto the neuropils surrounding the MB (Figure 18). These convergent loci also receive projections from the antennal lobe and the lateral horn and are likely to be major sites of integration and processing of information within the fly brain (Aso et al., 2014).
We have identified the neurons that comprise the MB and characterized their projection patterns. These data provide a comprehensive map describing the potential connections of neurons in the MB lobes. The MB lobes are innervated by approximately 2000 KCs, 34 MBONs of 21 cell types, and about 130 modulatory DANs of 20 cell types that fall into two families, PAM and PPL1. These data, extending the level of detail and completeness of previous studies of MB anatomy (Tanaka et al., 2008; Mao and Davis, 2009), describe the neuronal architecture of the MB and provide insight into the logic of MB function.
Most of the cell types in the olfactory circuit—the MBONs and the DANs as well as the olfactory sensory neurons, the PNs, and the LH neurons (Vosshall and Stocker, 2007; Masse et al., 2009)—comprise small groups containing from one to 20 neurons (Tanaka et al., 2004; Tanaka et al., 2008, 2012; Yu et al., 2010). Furthermore, neurons within these groups share highly stereotyped projection patterns that superimpose across different individuals (this work) (Marin et al., 2002; Datta et al., 2008). This suggests that the input and output connections of these neurons are genetically determined (Ruta et al., 2010; Kohl et al., 2013; Fisek and Wilson, 2014), although these may be subject to plasticity. In contrast, the KCs of the MB comprise 2000 neurons that define only seven distinct cell types, receive unstructured, rather than stereotyped, inputs (Murthy et al., 2008; Caron et al., 2013; Gruntman and Turner, 2013), and connect to multiple MBONs through plastic synapses (Cassenaer and Laurent, 2007, 2012). These distinctive features of the KCs afford the MB the ability to contextualize novel sensory experiences, consistent with its role in mediating learned olfactory associations and behavior.
Neural representations of odor exist at multiple stations of the olfactory circuit. In the representation of odors by the antennal lobe projection neurons, every odor can be thought of as a point in an approximately 50 dimensional space, each dimension corresponding to a particular glomerulus. The KC representation of odor identity has a dimension at least an order of magnitude larger, not only because there are many more KCs than antennal lobe glomeruli but also because the KCs mix projection neuron inputs nonlinearly (Gruntman and Turner, 2013). The dimension of the KC representation not only allows for a far greater capacity to respond appropriately to a large number of odors, it can also enhance performance in more complex decision-related tasks. Interestingly, the lack of structure in the input to the KCs, the small number of inputs, and the sparseness of their activity all appear to be tuned to maximize this dimension (Perez-Orive et al., 2002; Huerta et al., 2004; Caron et al., 2013; Gruntman and Turner, 2013) (Ann Kennedy, Columbia University Thesis). The fly has therefore evolved an olfactory circuit that exhibits structural and functional features predicted to optimize its ability to contextualize and respond appropriately to a rich array of olfactory experiences.
The convergence of a large number of KCs onto a small number of MBONs indicates that the dimension of the MBON representation is significantly smaller than that of the KCs. The dimension of the MBON can be no greater than their number (34) and is likely to be considerably smaller because there are only 21 different MBON cell types. Thus, rather than providing a general representation of odor identity, the activity of the individual MBONs is more likely to encode a set of ‘state variables’ that collectively bias behavioral responses to sensory stimuli. This bias is likely to reflect the combined effects of external experiences and the internal state of the fly (see accompanying paper) (Aso et al., 2014, Krashes et al., 2009; Bracker et al., 2013). Consistent with this view, the odor responses of individual MBONs differ between flies and these differences appear to depend on plasticity (Hige et al., unpublished).
It is tempting to associate individual MBONs with specific behaviors, but the ultimate bias in behavioral response may be represented across the full set of MBONs as a population code. Thus, the high-dimensional representation of odor identity in the KCs may be transformed into a low-dimensional representation that dictates behavioral bias. Just as an individual odor is represented by an ensemble of KCs, a given behavioral bias is likely to be represented by an ensemble of MBONs. In accord with this view, activation or inactivation of combinations of MBON cell types results in more robust effects on behavior than those observed with individual MBONs (see the accompanying paper) (Aso et al., 2014).
We have shown that the MB lobes are divided into compartments that receive input from the KCs and specific DANs and transmit information to a small number of MBONs. Each compartment therefore receives specific dopaminergic input capable of modifying the synapses between the KCs and specific MBONs. These compartments may reflect basic computational units of the MB. The specificity of the dopamine input and its ability to direct learning may therefore transform the unstructured KC representation of odor to an ordered MBON representation encoding behavioral bias. Specific subpopulations of DANs may react to different features in the external and internal world (Mao and Davis, 2009; Liu et al., 2012; Tomchik, 2013; Das et al., 2014). DAN activity may modify the activity of MBONs through learning to provide a representation that is predictive of the implications of an olfactory stimulus. Learning, in this view (Sutton and Barto, 1998), is a transition from a reactive DAN representation to a predictive MBON representation.
The dendritic and axonal projection patterns of the MBONs and DANs suggest feedforward and feedback circuit motifs in the MB lobes. The interactions among the different MBON types may form a multi-layer feedforward readout network (Figure 17). Processing through such a network significantly expands the computational capacity of a readout system, which can be valuable for more complex learning strategies. Consider a learning scenario in which a fly associates an odor with a strongly aversive US. A conditioned aversive response to the initial odor could occur through plastic changes affecting an ensemble of output neurons, and a low response threshold could allow these responses to generalize to related odors. Upon further experience, some odors within this category might be identified as ‘safe’ or even appetitive. Ethologically important exceptions could be learned if one of the neurons that interconnect compartments (MBON-γ4>γ1γ2, MBON-γ1pedc>α/β, and MBON-β1>α) became responsive to a ‘safe’ odor and inhibited the original trained ensemble of aversive MBONs. Thus, the layered MBON network provides an efficient mechanism for modifying and updating previous learning.
Interestingly, the dendrites of the DANs overlap with MBON axons in four of the five MBON projection zones (Figures 19 and 20). Thus, MBONs may modify the activity of the DANs that modulate their own activity and plasticity, resulting in a recurrent loop. This could provide positive or negative feedback to a specific compartment. Positive feedback might enhance learning to particularly salient stimuli, whereas negative feedback might suppress dopamine release once the correct response has been learned. These recurrent connections may also allow output from one compartment to modulate learning in other compartments, as some MBONs appear to target DANs that innervate non-cognate compartments.
The axonal terminals of the MBONs are largely confined to five discrete neuropils in the brain, the SIP, SMP, CRE, SLP, and the LH (Figures 18 and 19). The CRE, SMP, SIP, and SLP may be sites of convergence for all of the signals relevant for classical conditioning. These convergence zones are major sites of arborization for the dendrites of the DANs and axons of the MBONs. Since DANs are activated in response to an aversive or appetitive US, these neuropils are also likely to receive input from neurons transmitting information about the nature of the US. A US, by definition, elicits an innate behavioral response consistent with its valence, suggesting that the outputs from these neuropils may convey motor commands. Interestingly, dendrites of neurons projecting to the fan-shaped body in the central complex, a brain region coordinating motor actions (Strauss, 2002), arborize in these neuropils (Hanesch et al., 1989; Young and Armstrong, 2010). We therefore postulate an evolutionary primitive circuit in the MBON convergence zones, in which US inputs activate motor command neurons to elicit innate responses. Evolution may have built upon this simple reflex circuit, incorporating pathways from the MB that generate learned behaviors in response to a CS (Figure 21).
The vertebrate brain consists of interconnected structures comprised of large collections of equivalent neurons whose number often increases with evolutionary complexity. This is in sharp contrast to most brain structures in invertebrates that consist of small number of neurons with stereotyped projections that suggest determined connections. The MB represents an exception. The MB lobes are formed by the axons of a large number of equivalent neurons, the KCs, and as with vertebrate cortical neurons, the number of KCs increases in species with more complex behaviors (Strausfeld, 2012). Moreover, the inputs to the KCs from olfactory projection neurons are not determined or stereotyped but appear random. Thus, the MB diverges from the highly ordered neural architecture typical of the invertebrate nervous system. The MB may therefore represent an evolutionary primitive brain structure homologous in form and function to structures in the vertebrate brain (Schurmann, 1974; Laurent, 2002; Tomer et al., 2010; Farris, 2011). The elucidation of the inputs and outputs of the MB may now permit an understanding of how learning links an abstract representation to a specific behavior, and this may provide insight into higher associative functions in both invertebrate and vertebrate brains.
Split-GAL4 and LexA transgenes used enhancers, selected based on GAL4-line expression patterns (Jenett et al., 2012), and were constructed as previously described (Pfeiffer et al., 2010). VT999036 was from the Vienna Tiles collection and a gift of Barry Dickson. Promoter regions corresponding to the following GAL4 constructs were amplified by PCR: TH-GAL4 (Friggi-Grelin et al., 2003), Ddc-GAL4 (Li et al., 2000), HL9-GAL4 (Claridge-Chang et al., 2009), and Tdc2-GAL4 (Cole et al., 2005). All fragments were amplified from genomic DNA except for the upstream region of HL9, which was amplified from the HL9-GAL4 plasmid in order to conserve the mutated exon B start site. 5′-XbaI and 3′-FseI sites were added to the fragments upstream of GAL4. Downstream fragments were amplified with added 5′-SpeI (TH) or NheI (Tdc2, Ddc and HL9) sites and 3′-NotI sites. These fragments were then cloned into the corresponding sites on pBPp65ADZpUw and pBPZpGAL4DBDUw vectors (Pfeiffer et al., 2010) that had been modified to add new restriction sites as follows: the p65ADZp and ZpGAL4DBD segments of these vectors were amplified with the addition of a 5′-XbaI and a 3′-AvrII site and then cloned into pBDP (Pfeiffer et al., 2008) at 5′-EcoRI and 3′-NotI. Downstream fragments were cloned into the modified AD and DBD vectors at AvrII NotI (TH) or NheI NotI (Tdc2, Ddc and HL9) sites. To generate Trh-p65ADZp and Trh-ZpGAL4DBD, the Trh promoter region was amplified from genomic DNA using primers SM(1A) and BI(1S) (Alekseyenko et al., 2010) and cloned into pBPp65ADZpw and pBPZpGAL4DBDw using the Gateway system as previously described (Pfeiffer et al., 2008, 2010).
pJFRC225-5xUAS-IVS-myr::smGFP-FLAG in VK00005 and pJFRC200-10xUAS-IVS-myr::smGFP-HA in attP18 are described by Viswanathan et al. (unpublished); smGFP is a non-fluorescent, mutated GFP fused with multiple copies of an epitope tag (either HA, V5, or FLAG) for immunolabeling with various fluorescent dyes. pJFRC51-3xUAS-Syt::smGFP-HA in su(Hw)attP1 and pJFRC216-13xLexAop-myr::smGFP-V5 in su(Hw)attP8 were generated by standard methods using vectors described in Pfeiffer et al. (2010). Using a Syt::smGFP-HA construct with only three copies of the UAS sequence was necessary to decrease expression to a level that generated a >5-fold enrichment of signal in presynaptic boutons relative to other cellular subdomains. UAS-nuclearLacZ (UAS-nlsLacZ) for cell counting was previously described (Baker et al., 1996).
Multicolor flp-out (MCFO) is a stochastic method that labels individual cells in different colors using a set of three UAS-STOP-epitope constructs that each expresses a different epitope when the STOP cassette is removed. The STOP cassettes in these constructs are each flanked by FRT sites that are removed in a stochastic way by limited expression of flp recombinase (Struhl and Basler, 1993). Reagents for MCFO are described in Nern et al., in preparation.
NSyb-QF was generated by PCR amplifying the 1.9 kb EcoRI fragment of NSyb from pGWB-NSyb (gift of Julie Simpson), adding the Drosophila Synthetic Core Promoter (DSCP) (Pfeiffer et al., 2008) by overlap-extension PCR and cloning the resulting fragment between the MluI and EcoRI sites of pQUAST (Potter et al., 2010) replacing the QUAS and hsp70-promoter sequences. A QF coding sequence (QFrco; gift of Christopher Potter) was then cloned into this vector using EcoRI and BamHI. Transgenic flies were obtained by standard P-element mediated transgenesis (Genetic Services, Inc.). A total of four independent transformants were identified and in all cases most neurons were positive for QF as assessed by multiple lines bearing QUAS-PA-GFP; we used the line with the highest expression levels. A small group of neurons, including a subset of neurons within the PPL1 cluster, however, did not produce detectable QF activity in any of the NSyb-QF inserts (data not shown). To make MB247-QS, a 247 bp sequence of Mef2 encoding the MB247 enhancer was PCR amplified from genomic DNA, the DSCP sequence added by overlap-extension PCR, and the MB247–DSCP fragment cloned between the MluI and EcoRI sites of pQUAST. The coding sequence of QS was then added using EcoRI and NotI. Five independent transformants were recovered, and all exhibited suppression of QF activity in α/β and γ KCs as assessed with NSyb-QFrco. To generate QUAS-C3PA-GFP and QUAS-SPA-GFP, the C3PA-GFP and SPA-GFP coding sequences, flanked with EcoRI-CAAC (a Drosophila Kozak sequence) at the 5′-end and NotI at the 3′-end, were cloned into pQUAST (Potter et al., 2010). To generate 10xUAS-C3PA-GFP and 10xUAS-SPA-GFP, the C3PA-GFP and SPA-GFP coding sequences were blunt-end cloned into pJFRC-MUH (Pfeiffer et al., 2008) that had been digested with NotI and XhoI. Transgenic flies were obtained by phiC31-integrase mediated transgenesis (Genetic Services, Inc.) with insertion in attP40, attP2, VK00027, VK00005 for C3PA-GFP and in attP40 for SPA-GFP.
To identify enhancer fragments that drive expression in the MB cell types, we screened a database of the adult brain expression patterns of 7000 GAL4 driver lines (Pfeiffer et al., 2008; Jenett et al., 2012). We then generated approximately 400 transgenic lines that express either the transcription activation domain (p65ADZp) or the DNA binding domain (ZpGAL4DBD) of GAL4 under the control of one of the selected enhancers using the vectors described in Pfeiffer et al., (2010). We also generated p65ADZp and ZpGAL4DBD lines using the control regions from the genes encoding the enzymes for synthesizing monoamine neurotransmitters: tyrosine hydroxylase, dopamine decarboxylase, tryptophan hydroxylase, and tyrosine decarboxylase. To assay the expression pattern produced by an intersection of two enhancers, we visualized GAL4 activity in the progeny of a cross between a line expressing p65ADZp under one enhancer and a line expressing ZpGAL4DBD under the other enhancer.
We screened the expression patterns observed in female brains of more than 2500 different p65ADZp-ZpGAL4DBD combinations, each chosen based on our anatomical analyses of the original GAL4 lines as likely sharing expression in a particular cell type. For screening expression patterns generated by p65ADZp and ZpGAL4DBD combinations, we crossed males carrying pJFRC200-10XUAS-IVS-myr::smGFP-HA in attP18; the ZpGAL4DBD transgene in attP2 with virgin females carrying the p65ADZp transgene in either su(Hw)attP8, attP40, or VK00027 and examined expression in 3- to 10-day old female progeny. For screening, we performed immunohistochemistry as described below in Terasaki 60-well microtiter plates (Thermo Scientific, Waltham, MA) containing 8 µl of solution. To obtain polarity and higher resolution information on selected lines, split-GAL4 lines were crossed to pJFRC51-3xUAS-Syt::smGFP-HA in su(Hw)attP1; pJFRC225-5xUAS-IVS-myr::smGFP-FLAG in VK00005 and four females brains plus two ventral nerve cords (VNCs) were dissected per line and immunolabeled in 2.0 ml tubes as described below. We further characterized lines we intended to use in behavioral experiments. First, we compared the intensity and specificity of lines that drive expression in the same cell types by immunostaining in parallel, mounting on the same glass slide, and imaging with identical confocal settings. Second, we screened for expression in non-neuronal tissues by crossing lines with pJFRC2-10xUAS-mCD8::GFP in VK00005 and examining the bodies of F1 progeny by stereo fluorescence microscopy; we observed fluorescence in non-neuronal tissues in 17 out of the 176 split-GAL4 lines (see Figure 2—figure supplement 7 for examples).
Figure 2—figure supplements 2–6 and Supplementary file 1 document the lines used in this and the accompanying paper (Aso et al., 2014). These include KCs (Figure 2—figure supplement 2), DANs (Figure 2—figure supplements 3 and 4), and MBONs (Figure 2—figure supplement 5). We also made split-GAL4 lines for a variety of other modulatory input cell types that are putatively serotonergic, GABAergic, octopaminergic, and peptidergic (Figure 2—figure supplement 6), but we did not characterize these lines further in this study as their projections are not limited to localized areas of the MB lobes. Confocal image stacks documenting the expression patterns of the 92 selected lines in adult female brains and VNCs are available online (http://www.janelia.org/split-gal4).
Because the observed expression pattern depends to some extent on the reporter used, and in particular on its site of genomic insertion, we assayed the expression of the 92 selected lines (Supplementary file 1) independently with reporter constructs inserted at each of the two chromosomal sites (attP18 and VK00005) that we intended to employ in future behavioral experiments. Consistent expression patterns were observed when using different UAS reporters; but for some lines, the intensity of expression observed in both the targeted MB neurons and off target cells varied significantly with the reporter used. pJFRC2-10xUAS-IVS-mCD8::GFP and pJFRC225-5xUAS-IVS-myr::smGFP-FLAG in VK00005 expressed more strongly in the targeted MB neurons than pJFRC200-10XUAS-IVS-myr::smGFP-HA in attP18, but the two reporters (pJFRC2 and pJFRC225) in VK00005 occasionally visualized off-targeted cells that were not visible with pJFRC200 in attP18.
The small number of MB extrinsic neurons expressing split-GAL4 in most lines, generally 1 to 14 cells per brain hemisphere, permitted the use of simple visual inspection to judge the completeness and reproducibility of the expression pattern. However, for lines expressing in subsets of the KCs, often several hundred cells, we also employed computer-assisted cell counting. KC nuclei were visualized with UAS-nlsLacZ and their membranes with pJFRC225-5xUAS-IVS-myr::smGFP-FLAG. The cell body cluster of the KCs was imaged at high-resolution (0.1 µm × 0.1 µm × 0.1 µm voxels). To improve separation of neighboring nuclei, myr::smGFP-FLAG signals were subtracted from the nlsLacZ channel. Nuclei were segmented using the 3D component analyzer plugin of the Fluorender (Voxelpress; http://www.voxelpress.com/); first, we counted nuclei-sized (mean ± 2SD) nlsLacZ-stained objects at low threshold, deleted the detected volumes, and then counted the remaining nuclei at higher threshold. The counting results were visualized by randomly assigning colors and numbers to each counted object. This procedure gave results consistent with previous manual counting of c739 and c305a drivers (Aso et al., 2009). The signal intensity of nlsLacZ varied between nuclei in the same sample; we counted all detectable objects.
Brains of 3- to 7-day old females were dissected in saline (108 mM NaCl, 5 mM KCl, 5 mM HEPES, 5 mM trehalose, 10 mM sucrose, 4 mM NaHCO3, 1 mM NaH2PO4, 2 mM CaCl2, 1 mM MgCl2, pH-7.3) and mounted dorsal-anterior up in a Sylgard-coated Petri dish. Photoactivation was performed using a two-photon laser scanning microscope (Ultima, Prairie Technologies, Middleton, WI) with an ultrafast Ti:S laser (Chameleon Vision, Coherent, Santa Clara, CA) modulated by Pockels Cells (Conoptics, Danbury, CT). A water immersion objective (60X/1.0 NA, Olympus, Japan) was used for both visualization and photoactivation. Emitted photons were detected by a GaAsP detector (Hamamatsu Photonics, Japan) for green fluorescence and a PMT for red fluorescence. The laser was tuned to 925 nm for visualizing the samples (with an intensity of ∼0.7–2 mW as measured after the objective) and 710 nm for photoactivation. The Pockels Cells bias voltage was adjusted to obtain maximum signal-to-noise ratio when tuned at 710 nm, and photoactivation laser intensity at 710 nm was adjusted to be between 2 and 3 mW as measured after the objective. The photoactivation scan was performed with a pixel size of 0.24–0.39 μm and pixel dwell time of 2 μs. Each pixel was scanned four times successively using the frame-averaging function of the microscope software (PrairieView, Prairie Technologies); this was repeated with each repetition separated by 10–30 s. A target volume, such as the MB lobes (Figure 4) or a single MB lobe compartment (Figure 5), was divided into 8–15 z-slices with 2–5 μm steps, depending on the size of the target as well as the orientation of the sample. The number of repetitions of the photoactivation scan depended on the expression levels of PA-GFP as well as the depth of the photoactivation target. For photoactivation of the MB lobes, the photoactivation scan was repeated 30–60 times. For photoactivation of a single compartment, the photoactivation scan was repeated 90–120 times. All photoactivated samples were prepared for confocal microscopy as follows: fixed for 45 min at RT using 2% PFA/PBL (2% paraformaldehyde in 75 mM lysine, 37 mM sodium phosphate buffer, pH 7.4), washed multiple times in PBS containing 0.3% Triton X-100 (PBST), blocked with 10% normal goat serum in PBST, incubated in primary antibodies overnight at 4°C, washed multiple times in PBST, and incubated in secondary antibodies overnight at 4°C (or for more than 3 hr at RT) before final washes with PBST. The samples were mounted using either VECTASHIELD (Vector Labs, Burlingame, CA) or SlowFade Gold (Life Technologies, Grand Island, NY) and confocal imaging was performed using an LSM510 with a Plan-Neofluar 40X/1.3 objective (Zeiss, Germany). The following primary and secondary antibodies were used: rabbit-anti-DsRed (1:1000, Clontech), nc82 (1:10, Developmental Studies Hybridoma Bank), mouse anti-tyrosine hydroxylase (1:100, EMD Millipore, Germany), rat anti-Dopa decarboxylase (1:400, a gift from Jay Hirsch), rabbit anti-TexasRed (1:500, Invitrogen, Life Technologies), Alexa Fluor 568 goat anti-rabbit (1:200, Life Technologies), Alexa Fluor 633 goat anti-rat (1:200, Life Technologies), and Alexa Fluor 633 goat anti-mouse (1:200, Life Technologies).
We used flies expressing PA-GFP pan-neuronally with the exception of the KCs, generated with a Synaptobrevin-GAL4 driver (NSyb-GAL4 2-1; gift of Julie Simpson) in combination with MB247-GAL80, to identify the MB extrinsic neurons by photoactivation of the MB lobes (Figure 4). Generation of the photoactivation mask for the MB lobes was guided by a red fluorescent protein expressed in KCs (MB247-DsRed, a gift of Andre Fiala). In initial experiments, a volume of ∼800 μm × ∼600 μm × ∼220 μm covering most of the central brain was imaged at a pixel size of 0.39 μm × 0.39 μm, with z-step size of 3 μm before and after photoactivation of the MB lobes in a hemisphere. This identified five clusters of MB extrinsic neuron cell bodies, each of which reproducibly contained more than two cells (n = 3 brains, data not shown, see Figure 4). Neurons found reproducibly, but not within these clusters, include MB-APL, MB-DPM, two MBONs located by the contralateral spur (MBON-γ4>γ1γ2 and MBON-β1>α; MV2), and two MBONs located in the contralateral anterior lateral region (MBON-γ3 and MBON-γ3β′1). We examined the number of neurons in each of the five clusters except the anterior PAM cluster (see below) by comparing higher resolution images of areas containing each cluster taken before and after photoactivation. Two or three clusters were examined per sample, and a volume of ∼200 μm × ∼200 μm × ∼100 μm was imaged at 0.39 μm × 0.39 μm × 1 μm resolution for each cluster. Images taken before and after photoactivation were aligned using a subpixel registration algorithm (Guizar-Sicairos et al., 2008) and correlation between registered images. Slight shifts in sample orientation precluded the use of a simple image calculation to identify photoactivated neurons. Instead, the photoactivated cell bodies were visually identified from aligned images assisted by a custom MATLAB interface. This analysis identified two neurons with their cell bodies located in regions that do not contain any of the neurons identified in the split-GAL4 lines. One was located ventral to the calyx (MBON-γ4γ5) and the other was located ventral lateral to the antennal lobe (MBON-γ1γ2) (see below, and Figure 5G and Figure 5—figure supplement 1). They represent the only two cell types identified by PA-GFP tracing that were not found in the split-GAL4 lines.
It is important to note that we were not able to reproducibly visualize neurons with elaborate arbors over multiple neuropils and only diffuse and sparse processes in the MB lobes, such as octopaminergic neurons and SIFamide peptidergic neurons (data not shown). This indicates a limitation for the PA-GFP tracing experiments, at least using the parameters described above, in detecting neurons with a large cytoplasmic volume by just photoactivating PA-GFP molecules within a small fraction of the total volume.
We performed photoactivation of individual compartments to more specifically visualize the MB extrinsic neurons innervating each compartment (Figure 5). PA-GFP was expressed pan-neuronally using the Q-system (Potter et al., 2010) except in the α/β and γ KCs (NSyb-QF, MB247-QS, QUAS-PA-GFP) and each compartment was demarcated by a red fluorescent protein (myr::tdTomato) expressed using a split-GAL4 driver. The photoactivation mask was generated using this red marker, and in some cases where the split-GAL4 driver labels two compartments, this was further restricted using basal fluorescence of PA-GFP in the α′/β′ KCs (for example, to demarcate the β2 compartment using MBON-β2β′2a, Figure 5A). Upon photoactivation, the brains were fixed and immunostained for tdTomato and tyrosine hydroxylase or dopa-decarboxylase as described above. The confocal images were processed by a custom MATLAB code to identify photoactivated cells. For each image, the fluorescent intensity of the photoactivated GFP signals was measured by drawing a region-of-interest around tdTomato positive neurons (i.e., neurons whose processes had been used as target for the photoactivation), and the green channel of the image was thresholded by a pixel intensity value representing mean minus 2× standard deviations of all the pixels within the regions-of-interest. A cell was counted as photoactivated if most of the pixels within the cell, as determined visually, were over this threshold value. Finally, we determined whether these cells are dopaminergic by examining co-localization with tyrosine hydroxylase or dopa-decarboxylase. Photoactivation of each of the 15 compartments identified all the MBONs included in the split-GAL4 lines (Figure 5, and data not shown). We observed the same number of PA-GFP positive cell bodies for each MBON cell type as were labeled by the corresponding MBON split-GAL4 line (e.g., Figure 5A,B) and therefore these experiments confirmed that the split-GAL4 lines label all neurons within each MBON type. For the MBONs with dendrites in α′1 and α′3 compartments and cell bodies near KC cell bodies, it was not possible to assess whether there are more neurons of these types, because the cell bodies of photoactivated α′/β′ KCs were not distinguishable from those of the MBONs.
Photoactivation of the MB lobes resulted in labeling of over 100 cells in the anterior medial cluster including the PAM-DANs (cluster 5 in Figure 4) and it was not feasible to accurately count them by the methods described above. We therefore employed two independent approaches to examine the MB neurons in this cluster. We first performed photoactivation of the MB lobes with flies carrying R58E02-GAL80 (Liu et al., 2012) transgene that suppresses NSyb-GAL4-mediated PA-GFP expression in all PAM-DANs. We observed a dramatic reduction in the number of photoactivated cells in the anterior medial cluster ipsilateral to the photoactivated lobes as compared to flies without R58E02-GAL80 (data not shown). This suggests that many of the MB extrinsic neurons in the anterior medial cluster are PAM-DANs. Moreover, we observed no photoactivated cells in the contralateral anterior medial cluster when using R58E02-GAL80, indicating that the MB extrinsic neurons in this region are all PAM-DANs. We then characterized the MB extrinsic neurons in the anterior medial cluster by photoactivation of individual compartments as described above. We identified all the MBONs in this cluster and confirmed their numbers. We observed that there are more photoactivated GFP positive neurons in this cluster than those labeled by the split-GAL4 lines expressed in PAM-DANs (see for example Figure 5C,D). These are likely dopaminergic as confirmed by the immunostaining (Figure 5E and data not shown).
It is important to note that the identification of neurons using PA-GFP tracing critically depends on the signal-to-noise ratio of photoactivated fluorescence relative to basal fluorescence of PA-GFP molecules. Neurons may not be detected if they express insufficient levels of PA-GFP, as observed in some cases using NSyb-enhancer, or if the photoactivated volume contains only a small fraction of a neuron's processes. For example, we observed little expression of NSyb-QF in a subset of DANs in the PPL1 cluster as well as in PAM-α1 DANs (data not shown). Thus, we could not assess whether the split-GAL4 lines label all the PPL1-DANs and PAM-α1 DANs.
We used dye-filling to visualize one of the MB extrinsic neurons identified by PA-GFP that was not included in the split-GAL4 collection (Figure 5G). Flies were generated in which PA-GFP was expressed pan-neuronally except in the KCs (genotype in Figure 5G legend). No red fluorescent protein was expressed. Photoactivation was targeted to the medial lobes using the absence of basal PA-GFP fluorescence in the KCs to demarcate the MB lobes. Upon photoactivation, the brain was treated with collagenase (2 mg/ml, Sigma) for 30 s, the glial sheath removed using fine forceps, and the brain was mounted again in Sylgard-coated Petri dish in saline. A fire-polished, pulled glass pipette (0.5 mm ID, 1.0 mm OD, Sutter) was backfilled with a TexasRed dye (lysine-fixable 3000 MW, Invitrogen) dissolved in saline. A two-photon microscope was used to guide the pipette to the photoactivated cell body, and the dye was injected into the cell body by iontophoresis using over one hundred 3 ms pulses of 20 Vdc applied every 0.5 s. The dye was allowed to diffuse for an additional 10 min and then the brain was fixed and immunostained with anti-TexasRed antibody and nc82. Four brains were examined and in all cases we identified one neuron with the same morphology.
Dissection and immunohistochemistry of fly brains were done as previously described with minor modifications (Jenett et al., 2012). Brains and VNCs of 3- to 10-day old female flies were dissected in Schneider's insect medium and fixed in 2% paraformaldehyde in Schneider's medium for 55 min at room temperature (RT). After washing in PBT (0.5% Triton X-100 in PBS), tissues were blocked in 5% normal goat serum (or normal donkey serum, depending on the secondary antibody) for 90 min. Subsequently, tissues were incubated in primary antibodies diluted in 5% serum in PBT for 2–4 days on a nutator at 4°C, washed three times in PBT for 30 min or longer, then incubated in secondary antibodies diluted in 5% serum in PBT for 2–4 days on a Nutator at 4°C. Tissues were washed thoroughly in PBT four times for 30 min or longer and mounted on glass slides for imaging (see below for the mounting protocol).
The following antibodies were used: rabbit anti-GFP (1:1000; Invitrogen; A11122), mouse anti-nc82 (1:33.3; Developmental Studies Hybridoma Bank, Univ. Iowa) (Hofbauer et al., 2009), rabbit anti-HA (1:300; Cell Signaling Technology, Danvers, MA), rat anti-FLAG (1:200; Novus Biologicals, Littleton, CO), mouse anti-Drosophila ChAT (ChAT4B1; 1: 100; Developmental Studies Hybridoma Bank, Univ. Iowa) (Takagawa and Salvaterra, 1996), rabbit anti-GABA (1:500; A2052, Sigma-Aldrich, Switzerland), rabbit anti-5HT antiserum (1:1000; Sigma-Aldrich, catalog no. S-5545), mouse anti-tyrosine hydroxylase (LNC1, Millipore), rat anti-DDC (1:400; a gift from Dr J Hirsh) (Beall and Hirsh, 1987), mouse anti-beta galactosidase (1:200; Abcam, Cambridge, MA), mouse anti-V5-TAG (1:1000; AbD Serotec, UK), Dylight-549 conjugated mouse anti-V5 (1:500; AbD Serotec), rabbit anti-Drosophila GAD1 (1:1000; a gift from Dr. FR Jackson), rabbit anti-DvGluT (1:5000; a gift from Dr. A DiAntonio) as primary antibodies, and cross-adsorbed secondary antibodies to IgG (H+L): AlexaFluor-488 donkey anti-mouse (1:400; Jackson Labs), AlexaFluor-594 donkey anti-rabbit (1:500; Jackson Labs, Sacramento, CA), Cy3 donkey anti-rabbit (1:500; Jackson Labs), AlexaFluor-647 donkey anti-rat (1:300; Jackson Labs), AlexaFluor-488 goat anti-rabbit (1:800; Invitrogen A11034), and AlexaFluor-568 goat anti-mouse (1:400; Invitrogen A11031).
For determining the likely transmitter used by each MBON cell type, we immunolabeled brains from flies carrying the appropriate split-GAL4 drivers and pJFRC225-5xUAS-IVS-myr::smGFP-FLAG in VK00005. Tissues were first incubated in primary antibody against GABA, GAD1, or DvGluT for 2–3 days at 4°C, washed, incubated in secondary antibody for 2–3 days, and washed overnight. To visualize the MBON, we then incubated tissues in either rabbit anti-GFP or rat anti-FLAG (depending on the host species of other primary antibody) for 2–3 days, washed, and then incubated in secondary antibody for 2–3 days. Mixtures of 40–60 brains from 17 split-GAL4 drivers were stained in the same tube and mounted and imaged on the same glass slide to enable an unbiased comparison of immunoreactivity across MBONs. The expression pattern of the myr::smGFP-FLAG was used to genotype the brains. For other protocols, tissues were incubated in mixtures of multiple primary or secondary antibodies.
After immunohistochemistry, tissues were post-fixed with 4% PFA in PBS for 4 hr at RT followed by four, 15 min washes in PBT. To improve adhesion during mounting, tissue were washed in PBS (15 min) to remove the Triton and then placed on poly-L-lysine-coated cover slips to which they electrostatically adhere. Tissues were then dehydrated through a series of ethanol baths (30%, 50%, 75%, 95%, and 3 × 100%) for 10 min each and then 100% xylene three times for 5 min each in Coplin jars. Samples were embedded in a xylene-based mounting medium (DPX; Electron Microscopy Sciences, Hatfield, PA), and the DPX was allowed to dry for 2 days before imaging. For comparing expression intensities, up to 60 brains and VNCs were mounted on the same cover slip. Because tissues were attached to the flat surface of the cover slip in the same orientation, the same brain structures were located at the same depth during confocal imaging, facilitating a fair comparison of signal intensity across samples.
Imaging was done on an LSM710 confocal microscope (Zeiss). Brains and VNCs were imaged first at low-resolution using a Plan-Apochromat 20x/0.8 M27 objective (voxel size = 0.56 × 0.56 × 1.0 µm; 1024 × 1024 pixels per image plane). The region including the neurons of interest was then imaged at higher resolution by using a Plan-Apochromat 63x/1.40 oil immersion objective (voxel size = 0.19 × 0.19 × 0.38 µm; 1024 × 1024 pixels). For cell types too large to fit in a single image, regions of interests were scanned separately with multiple tiles (a maximum of five tiles was required to cover the entire brain and optic lobes) that were then stitched (Yu and Peng, 2011).
Confocal images were analyzed using the Janelia Workstation, a suite of tools for viewing and analyzing image data (S Murphy; K Rokicki; C Bruns; Y Yu; L Foster; E Trautman; D Olbris; T Wolff; A Nern; Y Aso; N Clack; P Davies; S Kravitz; T Safford, unpublished), ImageJ (http://imagej.nih.gov/ij/), Fiji (http://fiji.sc/), and Fluorender (http://www.sci.utah.edu/software/13-software/127-fluorender.html; [Wan et al., 2009]). Figure panels with black backgrounds are single slices or maximum intensity projections of confocal stacks or substacks. Figure panels with white or gray backgrounds show neurons that have been segmented using FluoRender.
Making an anatomical atlas from confocal images of many different split-GAL4 lines depends on being able to align data collected from individual specimens onto the framework of a standard brain. In order to minimize the amount of deformation required in the brain alignment process, we prepared a standard brain (JFRC2013) using fixation and dehydration steps identical to those used to prepare our experimental samples. After stitching (Yu and Peng, 2011) five tiles of high-resolution confocal image stacks covering the entire brain and optic lobes, debris on the brain surface were removed from the image using Fluorender. We then generated a downscaled (0.38 µm isotropic voxels) version of the JFRC2013 standard brain for use in alignment.
Because confocal imaging time was a limiting factor, we sought to develop a method that enabled alignment of a single 63× confocal image stack that covered only a portion of the brain to a standard model of the entire brain. Our alignment strategy is outlined in Figure 22. We imaged each brain twice: the entire brain at low-resolution (20× objective lens; voxel size, 0.56 × 0.56 × 1.0 µm) and a confocal stack (or stacks) covering the region of interest at high-resolution (63× objective lens; voxel size, 0.19 × 0.19 × 0.38 µm). The two images were scaled to have isotropic voxels; because the low-resolution images were obtained with an air objective, they were optically flattened compared to the high-resolution images and, thus, required scaling in the z-axis. We then aligned the high-resolution image tile to the low-resolution whole brain image, using the reference nc82 channel, by means of image stitching (Yu and Peng, 2011), which obtains translations through searching the maximum normalized cross correlation using the fast Fourier transform (FFT). Then affine registration of the low-resolution whole brain image to the JFRC2013 standard brain was used to provide a global alignment of the partial brain image to the standard brain. Finally, a non-linear transformation was applied to locally register the high-resolution tile image to the JFRC2013 standard brain using a symmetric diffeomorphic registration algorithm (Avants et al., 2008) with a combination of mutual information and normalized cross correlation as the similarity metric in the local alignment.
For generating the atlas, we used only ∼25% of the specimens that showed the best alignment to the standard brain based on the reference nc82 marker. In this way we were able to obtain very high-quality alignments as judged by two criteria. First, the alignments to the standard brain obtained from multiple specimens of the same GAL4 line were very similar (Figure 22E,F). Second, we observed the same relative arrangement of cell types when assessed using images from a single brain (Figure 11F) and by alignment of images of separate brains (Figure 22G). We routinely obtained alignment data of a given cell type from multiple brains to allow us to assess biological and alignment variability between samples.
Neuropil masks were generated in the JFRC2013 standard brain by alignment to the binary masks of the neuropils defined in Ito et al. (2014) , followed by manual editing of the neuropil borders guided by the nc82 staining of the JFRC2013 brain. To generate masks for each MB compartment, we averaged the registered intensities of MB extrinsic neurons projecting to the same compartment and samples representing the same KC cell type; we manually defined the borders between neighboring compartments after applying a Gaussian blur filter in 3D (sigma = 2).
After normalization of intensities between images, terminals and dendrites were segmented based on morphology and Syt::smGFP-HA distribution using FluoRender (Wan et al., 2009, 2012). To determine the distribution of projections in different brain regions shown in Figure 18E, signals from the cell types making up each of five groups—glutamatergic, GABAergic and cholinergic MBONs and the PPL1 and PAM DANs—were averaged within a group. Then the total signal within the volume of each neuropil mask for each of the five groups was divided by total signal observed in all neuropils. To calculate signal intensities shown in Figure 19C, the average of signals in 10 × 10 × 10 voxel volumes (3.8 µm in each dimension) were used.
To estimate degree of overlap between processes of MBONs and DANs in Figure 20, we used the method previously described by Cachero et al. (2010). We had multiple images for each cell type and we treated the two brain hemispheres separately, giving us on average 17.4 image pairs per cell type combination. We computed the overlap for each image pair separately; each cell in the matrices shown in Figure 20A–C represents the mean value.
The mushroom body of adult Drosophila characterized by GAL4 driversJournal of Neurogenetics 23:156–172.https://doi.org/10.1080/01677060802471718
Mushroom body output neurons encode valence and guide memory-based action selection in Drosophilae04580, eLife, 3, 10.7554/eLife.04580.
Specific dopaminergic neurons for the formation of labile aversive memoryCurrent Biology 20:1445–1451.https://doi.org/10.1016/j.cub.2010.06.048
Organization and postembryonic development of glial cells in the adult central brain of DrosophilaThe Journal of Neuroscience 28:13742–13753.https://doi.org/10.1523/JNEUROSCI.4844-08.2008
Regulation of the Drosophila dopa decarboxylase gene in neuronal and glial cellsGenes & Development 1:510–520.https://doi.org/10.1101/gad.1.5.510
Pattern recognition and machine learningNew York: Springer.
A map of octopaminergic neurons in the Drosophila brainThe Journal of Comparative Neurology 513:643–667.https://doi.org/10.1002/cne.21966
Cellular configuration of single octopamine neurons in DrosophilaThe Journal of Comparative Neurology 518:2355–2364.https://doi.org/10.1002/cne.22337
Different classes of input and output neurons reveal new features in microglomeruli of the adult Drosophila mushroom body calyxThe Journal of Comparative Neurology 520:2185–2201.https://doi.org/10.1002/cne.23037
Imaging a population code for odor identity in the Drosophila mushroom bodyJournal of Neuroscience 33:10568–10581.https://doi.org/10.1523/JNEUROSCI.0682-12.2013
Two functional but noncomplementing Drosophila tyrosine decarboxylase genesJournal of Biological Chemistry 280:14948–14955.https://doi.org/10.1074/jbc.M414197200
Tripartite mushroom body architecture revealed by antigenic markersLearning & Memory 5:38–51.
Drosophila learn opposing components of a compound food stimulusCurrent Biology 24:1723–1730.https://doi.org/10.1016/j.cub.2014.05.078
A unique mushroom body substructure common to basal cockroaches and to termitesThe Journal of Comparative Neurology 456:305–320.https://doi.org/10.1002/cne.10517
Stereotyped connectivity and computations in higher-order olfactory neuronsNature Neuroscience 17:280–288.https://doi.org/10.1038/nn.3613
Integration of the olfactory code across dendritic claws of single mushroom body neuronsNature Neuroscience 16:1821–1829.https://doi.org/10.1038/nn.3547
Neuronal architecture of the Central complex in Drosophila-melanogasterCell and Tissue Research 257:343–366.https://doi.org/10.1007/BF00261838
Drosophila mushroom body mutants are deficient in olfactory learningJournal of Neurogenetics 2:1–30.https://doi.org/10.3109/01677068509100140
The Wuerzburg hybridoma library against Drosophila brainJournal of Neurogenetics 23:78–91.https://doi.org/10.1080/01677060802471627
Cellular-resolution population imaging reveals robust sparse coding in the Drosophila mushroom bodyThe Journal of Neuroscience 31:11772–11785.https://doi.org/10.1523/JNEUROSCI.1099-11.2011
Learning classification in the olfactory system of insectsNeural Computation 16:1601–1640.https://doi.org/10.1162/089976604774201613
The Drosophila mushroom body is a quadruple structure of clonal units each of which contains a virtually identical set of neurones and glial cellsDevelopment 124:761–771.
The organization of extrinsic neurons and their implications in the functional roles of the mushroom bodies in Drosophila melanogaster MeigenLearning & Memory 5:52–77.
Intrinsic neurons of Drosophila mushroom bodies express short neuropeptide F: Relations to extrinsic neurons expressing different neurotransmittersJournal of Comparative Neurology 507:1479–1496.https://doi.org/10.1002/cne.21636
Three-dimensional reconstruction of the antennal lobe in Drosophila melanogasterThe Journal of Comparative Neurology 405:543–552.https://doi.org/10.1002/(SICI)1096-9861(19990322)405:43.0.CO;2-A
Olfactory network dynamics and the coding of multidimensional signalsNature Reviews Neuroscience 3:884–895.https://doi.org/10.1038/nrn964
Development of the Drosophila mushroom bodies: sequential generation of three distinct types of neurons from a neuroblastDevelopment 126:4065–4076.
Synaptic organization in the adult Drosophila mushroom body calyxJournal of Comparative Neurology 517:808–824.https://doi.org/10.1002/cne.22184
Drosophila ORB protein in two mushroom body output neurons is necessary for long-term memory formationProceedings of the National Academy of Sciences of USA 110:7898–7903.https://doi.org/10.1073/pnas.1216336110
Localization of the contacts between Kenyon cells and aminergic neurons in the Drosophila melanogaster brain using SplitGFP reconstitutionThe Journal of Comparative Neurology 521:3992–4026.https://doi.org/10.1002/cne.23388
Tools for neuroanatomy and neurogenetics in DrosophilaProceedings of the National Academy of Sciences of USA 105:9715–9720.https://doi.org/10.1073/pnas.0803697105
Punishment prediction by dopaminergic neurons in DrosophilaCurrent Biology 15:1953–1960.https://doi.org/10.1016/j.cub.2005.09.042
Appetitive and aversive visual learning in freely moving DrosophilaFrontiers in Behavioral Neuroscience 4:10.https://doi.org/10.3389/fnbeh.2010.00010
On the functional anatomy of the corpora pedunculata in insects (author's transl)Experimental Brain Research 19:406–432.https://doi.org/10.1007/BF00234464
Dopamine and octopamine differentiate between aversive and appetitive olfactory memories in DrosophilaJournal of Neuroscience 23:10495–10502.
Conditioned responses in courtship behavior of normal and mutant DrosophilaProceedings of the National Academy of Sciences of USA 76:3430–3434.https://doi.org/10.1073/pnas.76.7.3430
Neuronal architecture of the antennal lobe in Drosophila melanogasterCell and Tissue Research 262:9–34.https://doi.org/10.1007/BF00327741
Arthropod brains: evolution, functional elegance, and historical significanceCambridge, MA: Harvard University Press.
The mushroom bodies of Drosophila melanogaster: an immunocytological and golgi study of Kenyon cell organization in the calyces and lobesMicroscopy Research and Technique 62:151–169.https://doi.org/10.1002/jemt.10368
The central complex and the genetic dissection of locomotor behaviourCurrent Opinion in Neurobiology 12:633–638.https://doi.org/10.1016/S0959-4388(02)00385-9
Reinforcement learning: an introductionCambridge, Mass: MIT Press.
Organization of antennal lobe-associated neurons in adult Drosophila melanogaster brainThe Journal of Comparative Neurology 520:4067–4130.https://doi.org/10.1002/cne.23142
Neuronal assemblies of the Drosophila mushroom bodyThe Journal of Comparative Neurology 508:711–755.https://doi.org/10.1002/cne.21692
Dopaminergic neurons encode a distributed, asymmetric representation of temperature in DrosophilaThe Journal of Neuroscience 33:2166–2176a.https://doi.org/10.1523/JNEUROSCI.3933-12.2013
Classical conditioning and retention in normal and mutant Drosophila melanogasterJournal of Comparative Physiology. A, Sensory, Neural, and Behavioral Physiology 157:263–277.https://doi.org/10.1007/BF01350033
Olfactory representations by Drosophila mushroom body neuronsJournal of Neurophysiology 99:734–746.https://doi.org/10.1152/jn.01283.2007
SIFamide is a highly conserved neuropeptide: a comparative study in different insect speciesBiochemical and Biophysical Research Communications 320:334–341.https://doi.org/10.1016/j.bbrc.2004.05.173
Molecular architecture of smell and taste in DrosophilaAnnual Review of Neuroscience 30:505–533.https://doi.org/10.1146/annurev.neuro.30.051606.094306
Reinforcement signalling in Drosophila; dopamine does it all after allCurrent Opinion in Neurobiology 23:324–329.https://doi.org/10.1016/j.conb.2013.01.005
An interactive visualization tool for multi-channel confocal microscopy data in neurobiology researchIEEE Transactions on Visualization and Computer Graphics 15:1489–1496.https://doi.org/10.1109/TVCG.2009.118
FluoRender: an application of 2D image space methods for 3D and 4D confocal microscopy data visualization in neurobiology researchIEEE Pacific Visualization Symposium pp. 201–208.
Building the central complex in Drosophila: the generation and development of distinct neural subsetsThe Journal of Comparative Neurology 518:1525–1541.https://doi.org/10.1002/cne.22285
A complete developmental sequence of a Drosophila neuronal lineage as revealed by twin-spot MARCMPLOS Biology, 8, 10.1371/journal.pbio.1000461.
Automated high speed stitching of large 3d microscopic images2011 8th IEEE International Symposium on Biomedical Imaging: From Nano to Macro pp. 238–241.
Leslie C GriffithReviewing Editor; Brandeis University, United States
eLife posts the editorial decision letter and author response on a selection of the published articles (subject to the approval of the authors). An edited version of the letter sent to the authors after peer review is shown, indicating the substantive concerns or comments; minor concerns are not usually shown. Reviewers have the opportunity to discuss the decision before the letter is sent (see review process). Similarly, the author response typically shows only responses to the major concerns raised by the reviewers.
Thank you for sending your work entitled "The neuronal architecture of the mushroom body provides a logic for associative learning" for consideration at eLife. Your article has been favorably evaluated by K VijayRaghavan (Senior editor), Leslie Griffith (Reviewing editor), and 2 reviewers, one of whom is also a member of our Board of Reviewing Editors, and one of whom, Ann-Shyn Chiang, has agreed to reveal her identity.
The Reviewing editor and the reviewers discussed their comments before we reached this decision. Instead of summarizing the reviews, the Reviewing editor has appended all the comments from the two reviewers. As you see, the reviewers are impressed by the study, and most comments are meant to help further improve/clarify the presentation. These can be addressed by textual changes. We look forward to receiving the revised version of this paper.
Aso et al. reported stunning advances in the anatomical organization of the mushroom bodies, a key center for learning and memory in insects. Together with the accompanying behavioral analysis paper, these represent major breakthrough in understanding the structure and function of the mushroom bodies, their output, and their modulatory input. These studies have certainly laid down groundwork for years to decades of future investigations, and have already provided interesting insights on the logic of information processing principles through these intriguing structures. I am extremely enthusiastic in supporting the publication of these papers. Below I provide critiques to the anatomy paper mostly to ensure that the paper is accessible to as wide an audience as it can.
1) My strongest critique is the nomenclature, in particular the classes of MBONs. As far as I understand, most of these neurons are identified for the first time, at least to the resolution described, in this paper. So, the authors are in a position to name these neurons. I understand the rationales of the authors' choice using the connectivity with specific mushroom body lobes, and it may seem rational within this paper as it builds from Kenyon cell subtypes to MBONs. But even when reading the companion paper, only a slight distance away, it is already very difficult to remember these names (e.g., MBON-gamma4>gamma1gamma2, MBON-beta'2mp_bilateral, MBON-gamm1pedc>alpha/beta, etc.) or the rationale behind without constantly checking this paper. Consider many papers that will follow this paper in many years to come! Often complex names can turn off young scientists wishing to enter an otherwise very exciting field. In my view, these names should be greatly simplified. I have a specific suggestion. Since the authors found that MBONs neatly fall within three neurotransmitter phenotypes, and that MBONs that belong to a particular neurotransmitter category intersect with specific MB lobes at specific locations (both very interesting new findings!), why not name the MBONs after neurotransmitters (which is one of the most used convention for classifying neurons), and then within a neurotransmitter category, give a number, perhaps based on the proximal-distal axis of the mushroom body axons (e.g., MBON-ACh1, ACh2...; GABA1, GABA2...; Glu1, Glu2...). An alternative is to use consecutive numbers, for instance, MBON-ACh1-ACh8, GABA9-GABA12, GLU13-19, etc. With a modified Table 1, readers can quickly check which lobe MBON-GABA3 receives input from, etc. The authors may also consider renaming the DANs but this may have more historical constraint and DANs don't neatly match MBONs. Still, some simplification also seems to be useful.
2) Definition of cell type: I would add the following "We operationally define a cell type..." at the beginning, and "The screen was near saturation in our largeGAL4 collection."
3) I may have missed among the large amount of data, but can the authors describe MBON subtypes that are modulated by previously reported appetitive vs. aversive DANs?
4) A few critiques on Figure 21 (the summary figure). It is probably better to draw a more anatomically correct mushroom body schematics (with a bifurcation), with Glu, ACh, and GABA mapped to the correct anatomical loci (right now Glu is the most proximal to the cell body, which is misleading as in reality it is the most distal at the tip of medial lobes). Also, why is the lateral horn left out of the scheme since it is the target of some MBONs and is thought to mediate innate olfactory medicated behavior?
5) Discussion of what is known about the inputs to CRE, SMP, SIP, and SLP neuropils will be helpful for readers to appreciate how the US leads to behavior through the action of the DA-MBON loops.
6) The movies are very useful to educate readers about the 3D organization of neuronal composition and their projections. However, the names stay too transient for readers to learn/remember. The authors could color code the names for each subclass of neurons and leave them on the screen after the neuronal classes are introduced, at least during segments when different classes of neurons are presented at the same time, to help readers to learn the relative positions of different classes of neurons and to remember one thing or two after viewing the movies.
The study is a milestone in Drosophila neuroanatomy. Using state-of-art genetic tools, authors identified a comprehensive list of single neurons comprising the mushroom body and constructed a map of potential neural connections dividing MB lobes into 15 functional compartments. Moreover, they generated 85 split-GAL4 lines expressing exclusively in few MB related neurons that allow genetic manipulation of specific target neurons. Most major claims, except one which can be easily addressed (see below), are supported by thorough and convincing evidence. The study has great impact to the field and will transform the way of future study for understanding how genes and circuits control complex insect behaviors. I support the publication of the study and urge the authors make the generated tools immediately available to the entire field to further accelerate our understanding of brain functions.
It has been reported previously that "KC dendrites are segregated into 17 complementary domains according to their neuroblast clonal origins and birth orders (Lin et al., 2007). The authors stated that "The five KC types that receive olfactory information are each represented by hundreds of neurons per hemisphere and have their dendrites in the main calyx. Although their axons project to spatially segregated layers in the lobes, their dendritic arbors are intermingled in the calyx and KCs within a given cell type exhibit variable dendritic projection patterns (Figures 6 and 7). Moreover, the KCs receive input from an apparently random collection of glomeruli (Murthy et al., 2008, Gruntman and Turner, 2013, Caron et al., 2013)". The statement is vague and implies a random spatial distribution of KC dendrites in the main calyx. In fact, the main calyx is clearly subdivided into 4 paralleled divisions, each derived from dendrites of descendent neurons of one of the four neuroblasts (Technau and Heisenberg, 1982; Ito et al., 1997; Lee et al., 1999; Zhu et al., 2003, 2006; Lin et al., 2007). Within each division, dendritic arbors of each KC type segregate at specific and separate spatial domain (Lin et al., 2007). Authors agree with that the pioneer alpha/beta (the posterior alpha/beta) KCs have dendrites exclusively in the dorsal accessory calyx and show that the newly identified embryonic gamma KCs have dendrites exclusively in the ventral accessory calyx. Consistently, in the main calyx, dendritic arbors of 5 other KC types also appear to distribute differently from each other (Figures 6 and 7). It remains unclear how the segregated KC dendrites account for the random PN-KC connections. Authors should either clarify their statement and cite the previous report indicating segregated dendrite distribution of each KC type in the main calyx or provide additional evidences of detailed anatomical analysis if otherwise.https://doi.org/10.7554/eLife.04577.046
- Yoshinori Aso
- Daisuke Hattori
- Yang Yu
- Rebecca M Johnston
- Nirmala A Iyer
- Teri-TB Ngo
- Heather Dionne
- Richard Axel
- Gerald M Rubin
- Yoshinori Aso
- Hiromu Tanimoto
- LF Abbott
- LF Abbott
- Hiromu Tanimoto
- Hiromu Tanimoto
- Hiromu Tanimoto
- Hiromu Tanimoto
- Hiromu Tanimoto
- Daisuke Hattori
- LF Abbott
- LF Abbott
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
We thank Yichun Shuai (Cold Spring Harbor Laboratory) for pointing out the presence of MB extrinsic neurons in the GAL4 line VT999036, Barry Dickson for providing VT999036, and Christopher Potter (Johns Hopkins) for proving the QFrco vector and sharing unpublished information. F Rob Jackson (Tufts University), Aaron DiAntonio (Washington University), Jay Hirsh (University of Virginia), and the Developmental Studies Hybridoma Bank (created by the NICHD of the NIH and maintained at the University of Iowa) provided antibodies and Yong Wan and Hideo Otsuna (University of Utah) added new functions to Fluorender needed for this work. The Janelia Scientific Computing Software group (Sean Murphy, Konrad Rokicki, Christopher Bruns, Les Foster, Eric T Trautman, Donald J Olbris, Nathan Clack, Pete Davies, Saul Kravitz, Todd Safford, Charlotte Weaver, and Rob Svirskas) generated software tools, the Janelia Fly facility (Amanda Cavallaro, Todd Laverty, and others) helped in fly husbandry, and the FlyLight Project Team (Nick Abel, Gina DePasquale, Adrianne Enos, Joanna Hausenfluck, Phuson Hulamm, Reeham Motaher, Omatara Ogundeyi, Allison Sowell, Susana Tae, and Rebecca Vorimo) performed brain dissections, histological preparations, and confocal imaging. Margaret Bezrutczyk and Camille Rogine assisted in generating reagents for the PA-GFP experiments. Phyllis Kisloff assisted in manuscript preparation. We thank members of the Rubin lab (Aljoscha Nern, Chris Murphy, Barret D Pfeiffer, Tanya Wolff, Arnim Jenett, and Ming Wu) and Axel lab for materials and discussion. YA and HT initiated this work while participants in the Janelia Farm Visiting Scientist Program.
- Leslie C Griffith, Reviewing Editor, Brandeis University, United States
- Received: September 3, 2014
- Accepted: November 5, 2014
- Version of Record published: December 23, 2014 (version 1)
© 2014, Aso et al.
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Downloads (link to download the article as PDF)
Download citations (links to download the citations from this article in formats compatible with various reference manager tools)
Open citations (links to open the citations from this article in various online reference manager services)
Animals discriminate stimuli, learn their predictive value and use this knowledge to modify their behavior. In Drosophila, the mushroom body (MB) plays a key role in these processes. Sensory stimuli are sparsely represented by ∼2000 Kenyon cells, which converge onto 34 output neurons (MBONs) of 21 types. We studied the role of MBONs in several associative learning tasks and in sleep regulation, revealing the extent to which information flow is segregated into distinct channels and suggesting possible roles for the multi-layered MBON network. We also show that optogenetic activation of MBONs can, depending on cell type, induce repulsion or attraction in flies. The behavioral effects of MBON perturbation are combinatorial, suggesting that the MBON ensemble collectively represents valence. We propose that local, stimulus-specific dopaminergic modulation selectively alters the balance within the MBON network for those stimuli. Our results suggest that valence encoded by the MBON ensemble biases memory-based action selection.
A detailed map of the neurons that carry information away from the mushroom bodies in the brains of fruit flies has improved our understanding of the ways in which experiences can modify behaviour.