From stimulus to meaning

The world is full of sensory stimuli. In order to ensure survival and reproduction, animals must learn which cues will guide them to food sources and mating partners and which will predict potential threats. Fruit flies can learn to associate sensory stimuli, such as those of olfactory or visual nature, with a food source (reward) or an unpleasant stimulus (punishment). Decades of research have cemented a third order neuropil, the mushroom body, as the major center for associative learning in the insect brain (Heisenberg, 2003; Menzel, 2014). The fly mushroom body network consists of no more than 3000 neurons per brain hemisphere. Due to this relative numerical simplicity and the advanced genetic tool-box available, a comprehensive understanding of how memories are written at the level of the mushroom body network has surfaced in recent years.

The mushroom body principal cells, the Kenyon cells (KCs, approximately 2000 cells per hemisphere), hold a representation of the surrounding sensory world, particularly of olfactory cues. While airborne odors are perceived at the level of sensory receptor neurons, information is further computed at the next relay station (the antennal lobe), and then conveyed to KCs. A given olfactory cue specifically activates a sparse pattern of KCs; it is this pattern that codes odor identity within the mushroom body network (Stevens, 2015).

The cholinergic (Barnstedt et al., 2016) KCs extend parallel axon bundles that make up the mushroom body lobes. Within the lobes, they form excitatory en passant synapses with a small number of downstream partners, the mushroom body output neurons (MBONs; see Figure 1a). The dendritic fields of individual MBONs stereotypically tile the lobes of the mushroom body and mark distinct non-overlapping compartments (Tanaka et al., 2008; Aso et al., 2014b). Based on this anatomy, the thirty-five MBONs per hemisphere can be classified into twenty-one categories (types).

Interestingly, olfactory information gets categorized according to valence at this stage. The activity of specific types of MBONs is sufficient to promote odor driven approach while other types support odor driven avoidance behavior [Figure 1b; (Aso et al., 2014a; Owald et al., 2015a; Perisse et al., 2016)]. Indeed, when naïve flies are given the choice between a clean air stream and a repulsive odor, they predominantly choose to avoid the odor. Acutely blocking synaptic output from avoidance promoting MBONs flips this aversive behavior: flies now approach the repulsive odor. On the contrary, silencing approach-promoting MBONs during a choice situation increases avoidance behavior (Owald et al., 2015a; Perisse et al., 2016). This notion is further supported by optogenetic experiments (Aso et al., 2014a; Owald et al., 2015a): naïve flies expressing light-activatable cation channels (for example CsChrimson) were given the choice between an illuminated and a dark site. If CsChRimson was expressed in avoidance-promoting MBONs, flies would avoid, if the light-activatable channel was expressed in approach-promoting MBONs, flies would approach the light source. These experiments are in line with a model in which the identity of an odor is translated into a meaning at the synapse between KCs and MBONs (Heisenberg, 2003) and we will argue that this behavioral switch from avoidance to attraction behaviorally mimics observed cellular memory traces.

Fig. 1:

The major components of the mushroom body. a) Kenyon cells (KCs) shape the MBs (grey solid line). There are three major classes of KCs, the αβ, the α′β′ (in purple one example Kenyon cell) and the γ KCs which all receive input from projection neurons at their input region, the mushroom body calyx (black). The KC neurites project along the peduncle into the lobes (grey). Within the lobes KCs make en passant excitatory synapses (arrow heads) with downstream neurons, the mushroom body output neurons (MBONs). b) Distinct types of MBONs promote avoidance and approach behavior (purple and blue respectively). c) The dendritic innervation pattern of each class of MBONs (two examples, blue and purple) is matched by the innervation of a corresponding group of dopaminergic neurons (DANs, red and green). Together they define distinct compartments (dashed lines). The role of some MBONs has been investigated. For instance, the α′3 MBON has recently be shown to be involved in encoding novelty and familiarity (Hattori et al., 2017). However, the role of several remaining MBONs remains to be elucidated.

Assigning a value

As in mammals, associative learning in the fly depends on dopaminergic signaling. Interestingly, the compartmentalization of the mushroom body by MBON dendrites is perfectly matched by the presynaptic innervation pattern of two major clusters of dopaminergic neurons: a relatively small set of neurons (PPL1 cluster) provides the teaching signal for punishment and dopaminergic neurons of the larger PAM cluster convey rewarding information (Figure 1c). Notably, learning about different punishments, such as electric shock, extreme heat or bitter substances, all depend on the same small set of punishment-coding dopaminergic neurons [for review (Waddell, 2013)]. In contrast, forming associations linked to different rewarding events requires the activity of distinct sets of PAM dopaminergic neurons: the reward-related teaching signal provided by water recruits dopaminergic neurons that are different from those involved in learning about food rewards such as sugars (Burke et al., 2012; Liu et al., 2012; Lin et al., 2014). Reward provided by sugars can be even further distinguished on a neuronal level. Dopaminergic neurons signaling ‘sweet’ differ from those providing information on the nutritional content (Huetteroth et al., 2015; Yamagata et al., 2015). Because these distinct sets of reinforcing neurons innervate separate non-overlapping compartments, their activity affects distinct synapses between KCs and specific MBONs. Together, such an anatomical separation suggests that memories about punishment and different reward types are stored at specific KC-MBON synapses at defined sites of the respective compartments, thus providing the backbone for a distinguishable memory read and write system.

Reading and writing memories: synapses and networks

Punishment-coding dopamine neurons innervate compartments where approach promoting MBONs extend their dendrites and reward-related dopamine neurons innervate compartments covered by avoidance promoting MBONs. This pattern implies that during olfactory learning, compartment specific dopamine release induces a depression of synaptic connections between KCs coding for the trained odor and the downstream MBONs.

The first evidence for learning induced plasticity at the level of MBONs came from extracellular recordings in the honey bee (Menzel and Manz, 2005). However, studies from Drosophila have recently used genetically-encoded calcium indicators to measure neural activity in genetically-identified populations of neurons in trained and untrained flies [for review (Owald and Waddell, 2015)]. Utilizing such experimental strategies, it was recently demonstrated that the responses of specific glutamatergic avoidance MBONs (the so called M4/6 MBONs) downstream of sugar-reinforcing dopaminergic neurons were depressed for an odor that had previously been associated with a reward. Strikingly, the physiological depression observed here matches the behavioral switch from odor avoidance to approach observed when blocking these MBONs in the naïve fly (see above). Because interfering with these neurons during memory recall also abolishes learned behavior, depression of this KC to MBON connection most likely (a) is the site for appetitive memory storage and (b) is directly causal for the observed behavioral switch (Owald et al., 2015a).

Input to other sets of MBONs that promote approach and are downstream of punishment-coding dopaminergic neurons, is typically depressed after associating an odor with an electric shock (Séjourné et al., 2011; Hige et al., 2015; Perisse et al., 2016). Together these findings suggest synaptic depression as the major motif of synaptic plasticity in the mushroom body network. However, for some MBONs (for instance V2-MBONs, γ2α’1-MBONs) odor responses were potentiated after training (Bouzaiane et al., 2015; Owald et al., 2015a; Felsenberg et al., 2017). Indeed, some MBONs (like M4/6) show bidirectional plasticity: they are depressed after reward learning, but show an enhanced response for an aversively trained odor. The mechanisms underlying these plasticity traces, however, are not solely confined to local synaptic phenomena: enhanced responses after aversive conditioning more so seem to arise from a network effect (Owald et al., 2015a; Perisse et al., 2016). Depression of approach-promoting GABAergic MBONs (‘MVP2’) after aversive learning changes the inhibition these MBONs feed-forward onto avoidance MBONs. This disinhibition effectively leads to a potentiation of the latter class of neurons. Importantly, both sets of MBONs receive input from the odor-coding KCs, so that information of odor-identity can still be retrieved through such a network motif.

Fig. 2:

Olfactory memories are stored as a skew in the mushroom body output network. Olfactory input drives activity in a specific set of KCs (not shown), which in turn activate all classes of downstream MBONs. The approach promoting GABAergic MVP2 MBON (blue, also known as MBON γ1pedc>α/β) provides feed-forward inhibitory input to the avoidance promoting M4/6 MBONs (purple, also known as MBON β′2mp, MBON β2β′2a, MBON γ5β′2a). a) In the current skew model for a fly that is naïve for a presented odor activity of approach and avoidance promoting MBONs is balanced. This balance translates into no contribution of the mushroom body to odor driven behavior (arrows, blue for approach behavior and purple for avoidance behavior). b) During appetitive training an odor is presented coincidently with a reward. Thus, sugar reward activates dopaminergic neurons (DANs), which innervate compartments in which the odor driven KCs connect to avoidance promoting MBONs. c) During aversive training an odor coincides with punishment such as an electric shock, which drives dopaminergic neurons innervating compartments where KCs connect to approach MBONs. d–e) After training, synapses between the KCs activated by the trained odor and the respective MBONs are depressed. This learning induced synaptic plasticity skews the MBON network towards either approach after appetitive learning or towards avoidance after aversive learning when the fly encounters the trained odor again. This skew in the network leads to the expression of learned behavior.

Together, these findings give rise to an integrative model, which predicts that olfactory memories are manifested as a skew in the mushroom body output network [Figure 2; (Owald and Waddell, 2015)]. Odors with no assigned learned value drive approach and avoidance MBONs with equal strength. During associative learning an odor activates a specific set of KCs concurrent with the presence of a meaningful cue. The latter in turn drives the respective dopaminergic neurons innervating a specific compartment and induces synaptic plasticity at the underlying odor specific KC to MBON synapses. The change of odor drive to a particular group of MBONs biases the network either towards approach MBONs after reward learning, or towards avoidance MBONs after punishment learning. This skew then elicits the learned stimulus-driven behavior of the fly.

The anatomy of MBONs, however, suggests that they not only project out of the mushroom body to pre-motor areas, but also connect to the dopaminergic neurons that feed back to the mushroom body (Aso et al., 2014b; Lewis et al., 2015; Owald et al., 2015a; Eichler et al., 2017). Given that learning changes the drive of specific MBONs, the feedback to the dopaminergic systems is also changed when a fly encounters the learned odor again (Riemensperger et al., 2005). This motif of recurrent feedback loops turns out to be essential for the re-evaluation of learned behavior (Felsenberg et al., 2017).

Reconsidering what is true

In an ever-changing world, reliability of learned information must constantly be re-assessed to ensure adequate behavior. Thus, one of the most crucial aspects of memory to grant behavioral flexibility is that it is malleable. If an animal encounters a situation in which the stored information does not match the actual outcome (a mismatch condition which can be computed as the so called prediction error), learned information and previous memories need to be adjusted. In general, this can be achieved in two ways, either by updating the original memory or by the formation of a new opposing memory, an extinction memory. It seems to depend on the extent of the prediction error occurring during memory recall, which of the two processes is utilized. In Drosophila, an extinction memory can be formed by re-exposing flies to a previously rewarded odor in the absence of the expected food reward. Such an incongruity changes the behavior of flies; it nullifies the learned approach behavior (Tempel et al., 1983; Felsenberg et al., 2017). The new learning event, the extinction learning, depends on the teaching signal from punishment-coding dopaminergic neurons which are driven by the skew in the output network established during initial learning. Thus, the omission of reward is coded as a punishment and leads to the formation of a parallel aversive extinction memory which opposes the initial appetitive reward memory [Figure 3; (Felsenberg et al., 2017)].

Fig. 3:

Extinction of reward memory is driven by the punishment system. a) MBONs form recurrent feedback loops with dopaminergic neurons (DANs), which provide the teaching signal during associative learning: approach MBONs drive punishment-coding neurons and avoidance neurons are connected to reward-related dopaminergic neurons. b) After learning the response of the MBON network to the reward-associated odor is skewed towards approach MBONs. This skew translates into a relatively stronger drive of the punishment dopaminergic neurons during extinction learning. c) Rather than going back to the naïve state, a parallel, aversive memory for the trained odor is formed which annuls the skew in the MBON network (Felsenberg et al., 2017).

Expectance and reality do, however, sometimes match. If flies experience a reminder of their positive memories, which does not strongly conflict with the learned information, the memory undergoes a cycle of de- and re-stabilization. This process, called reconsolidation, is understood to be a well conserved memory update mechanism to integrate minor adjustments to the destabilized and therefore changeable memory (Nader, 2015). In flies, stabilized memories are insensitive to cooling-induced anesthesia. However, the application of a reminder renders the memory vulnerable, it is destabilized. If cooling is applied within a critical time window after the reminder, it interferes with the re-stabilization process of memory and hence erases the reward memory. Within this time window, temporally orchestrated activity of a specific MBON and distinct groups of dopaminergic neurons are required to ensure successful memory reconsolidation [(Felsenberg et al., 2017)].

Although processes underlying memory re-evaluation provide opportunities to alleviate problematic memories in humans, the mechanistic insight into these processes is sparse. Since the phenomena of extinction and reconsolidation are conserved across species (Eisenhardt, 2014; Nader, 2015) it might well be that the findings obtained in Drosophila represent coding principles that might help to unravel general mechanisms underlying memory reevaluation.

Outlook

The increasing knowledge of where and how memories are stored in the Drosophila brain, combined with genetic accessibility of the involved structures, provides a unique possibility to gain further mechanistic insights. Recent work has identified specific synapses as part of reward and punishment related memory traces. This offers the opportunity to investigate the molecular machinery and the cell specific genetic settings underlying learning-induced synaptic plasticity. Behavioral experiments linked with high resolution microscopy could determine the structural changes shaping synaptic efficacy. Precise knowledge of the involved network components will allow us to further tackle the coding principles of important processes involved in memory retrieval, consolidation and reevaluation.