Results

A tethered fruit fly, Drosophila, in the absence of sensory information, continuously changes its choice of flight direction [6]. Much as humans learn about the consequences of their actions, the fly's choices can be modulated by learning about the consequences of such decisions as when and where to turn (i.e., operant learning [7]). Procedurally, the task is not significantly altered by adding predictive stimuli to this task such that one (e.g., blue coloration of the environment) indicates which turning maneuvers (e.g., left turning) are punished with heat and the other (e.g., green coloration) indicates which decisions (e.g., right turning) are not punished; the decisions are still followed by the same consequences. However, this helpful indication of correct and incorrect choices drastically alters the biological processes underlying the task. Without the help of the colors, the flies require protein kinase C function, but not the rutabaga-encoded type I adenylyl cyclase, to learn to make the correct choice; with the colors, the results are reversed [1]. In order to further investigate this dominant effect of the classical colors on operant learning, wild-type and transgenic flies were first trained in situations with both operant and classical components present and were then tested for each component individually ( Figure 1).

Given the dominance of the colors in this paradigm, it may not come as a surprise that in a test without heat, after 8 min of such composite training, wild-type flies did not reveal any preference for left- or right-turning maneuvers if the helpful color filters were removed ( Figure 2A; i.e., the isolated operant component, situation 2 in Figure 1). Apparently, the colors inhibit operant learning. Why would the flies not learn an important predictor of punishment such as their own behavior? One hypothesis is that operant learning might lead to behavioral modifications, which in turn could potentially interfere with generalization of the classical color memory. Sensorimotor learning interfering with behavioral flexibility (“habit interference”) is a well-known phenomenon [8], and the balance between interference and transfer/generalization is a popular research topic [9]. We tested for the generalization of the classical memory by measuring color preference in a situation where straight flight (as opposed to constant turning) was required to reliably avoid the previously punished color (i.e., situation 3 in Figure 1, previously described in [10]; see Supplemental Experimental Procedures available online for details). After 8 min of composite training, wild-type flies successfully avoided the punished color via this orthogonal behavior ( Figure 2A), after a brief reminder training [10]. A commonly used experimental procedure to induce sensorimotor learning in other animals is overtraining [3] , [4] and [8] . According to the hypothesis above that learning of the operant behavior in flies may be analogous to sensorimotor learning leading to habit formation in mammals, extended training in flies should overcome the inhibition of operant learning and lead to a failure to generalize the isolated classical memory to the novel behavior. Consistent with this hypothesis, flies that were trained with equivalent operant and classical predictors for twice the regular amount of time showed significant performance indexes (PIs) in the control and in the operant test and no significant score in the generalization test ( Figure 2B). Observing the behavior of the flies, it was noticeable that, after extended training, some flies seemed to generate larger turning maneuvers toward the previously unpunished direction, compared with more symmetrical maneuvers from flies exposed to the regular amount of training ( Figure 2C). A quantitative evaluation of the behavioral data tended to confirm the qualitative observations, but the number of animals was too low to reach statistical significance (data not shown). Taken together, these results indicate that adjusting to the novel situation after extended composite training is difficult enough to disrupt performance in the generalization task (habit interference). Similar to a rodent pressing a lever for an aversive stimulus [2] , [3] and [4] , the fly, also only after extended training, keeps generating behaviors that interfere with avoiding the previously punished color.

In order to elucidate the neuronal substrates mediating these processes, specific neuronal ensembles in the fly's brain were silenced. Because previous evidence pointed toward the MBs being involved in specific generalization processes [11] , [12] , [13] and [14] , this neuropil was targeted with the UAS-GAL4 system to block synaptic output by expressing the bacterial tetanus toxin light chain [15]. The first P[GAL4] driver line was MB247, because this line has already seen widespread use as an MB-specific driver line [11] , [12] , [13] and [14] . MB247 drives expression in about 1600 of the ∼2000 Kenyon cells in all parts of the MB, except the prime lobes, and in some neurons of the central complex [16] and [17] . The heterozygous control crosses of driver and effector strains with Canton S wild-type strains reproduced wild-type behavior ( Figure 3A). Flies with impaired MB function can learn both the colors and how to modulate their turning movements [12] and [18] . Confirming these previous results, flies with tetanus toxin expression driven by line MB247 could master the composite learning task composed of these two predictors ( Figure 3B, situation 1). However, in a phenocopy of the wild-type flies after extended training, flies with such blocked MB output did not generalize the classical memory to a novel behavior and showed significant operant learning already after the regular 8 min of training ( Figure 3B, situations 3 and 2, respectively). Thus, with such manipulated MB function, flies appear to form habits prematurely.

Flies in which the P[GAL4] line c205 drives expression of a constitutively active G-Protein are defective in visual pattern discrimination learning [19]. Constitutive expression of tetanus toxin in the F5 neurons in the fan-shaped body of the central complex via the line c205 confirmed that the effects of tetanus toxin expression were specific to the MB: in contrast to the MB247 flies, these flies behaved similarly to wild-type and genetic control flies ( Figure 3C). In order to investigate which of the MB lobes are responsible for the inhibition of operant learning in such composite situations, transgenic flies with the P[GAL4] driver line 17D, which drives toxin expression mainly in the MB α and β lobes (core and surface) but not in the γ lobes [11] and [16] , were subjected to the same procedure. These flies show the same pattern of PIs as the MB247 flies: significant PIs in the control and in the purely operant test and no significant score in the generalization test ( Figure 3D), conclusively tying the inhibition of the operant component to MB neurons. Moreover, we can tentatively conclude that the MB γ lobes are probably not involved in this process.

Discussion

Spontaneous behavior has clear fitness benefits [6]. However, spontaneous behavioral variation may reduce efficiency by introducing mistakes. The success of an animal thus depends on finding the right balance between efficient exploitation of known resources through routine behavior and flexible exploration of possible new resources through novel behaviors (the exploitation-exploration dilemma [20] and [21] ). In a new situation, such as the operant paradigm used here, the animal explores the environment via spontaneous behaviors [6]. It learns about the stimuli in this environment and how they relate to each other primarily by engaging the classical learning system [1]. During this phase, the classical learning system inhibits the operant system via the MB, preventing direct modification of the behavior of the animal and keeping the memory flexible ( Figure 2 and Figure 3). After extended periods of time in this situation, the MB-mediated inhibition is overcome and the behavior is modified by the operant learning system, which may improve efficiency but also leads to inflexibility ( Figure 2B). The current data allow establishing a mechanistic model of how operant and classical learning systems may interact in composite learning situations and which biological substrates mediate these processes ( Figure 4). In this view, the Rutabaga adenylyl cyclase-dependent classical learning system inhibits the protein kinase C-dependent operant learning system via the MB. The operant learning system facilitates classical learning via still unknown, non-MB pathways (data not shown and [10]). This interaction leads to efficient learning, enables generalization, and prevents premature habit formation. In flies, it is not yet known whether the two learning systems are also separable anatomically. It is tempting to speculate that the interactions between the two learning systems are part of the mechanism achieving the balance between exploration and exploitation. In this hypothesis, the MBs provide the checks and balances to ensure that habits are formed only if their efficiency outweighs their disadvantage of being inflexible.

Such an MB function would be distinct from the one that the MBs are known to serve in olfactory classical conditioning. The current consensus is that the memory trace formed during this kind of learning lies within the MB Kenyon cells [22] , [23] and [24] . This is clearly not the case for visual learning, where the MBs are not essential [12] and [18] . Instead, specific features of the conditioned stimulus in visual learning appear to reside in distinct layers of the fan-shaped body of the central complex [19]. For visual learning, the MBs appear to keep classical memories flexible for use when the fly's situation changes. If the fly's sensory situation changes, this feature supports context generalization [11] and [12] and protects against sensory conflict [13] and [14] . If the fly's behavioral situation changes, this feature supports the form of generalization described here. From these accumulating recent results, it appears that the inhibitory function of the MB may be much more pervasive than previously thought. It is a tantalizing finding for all Drosophila learning and memory research that overtrained wild-type flies behave indistinguishably from flies with blocked MB output: whenever the neural substrate of a learning task is studied, the question of whether the training regime constitutes overtraining must now also be considered. This is reminiscent of vertebrate experiments, where the dorsal striatum and the hippocampus are viewed as competing learning systems with the dorsal striatum involved in skill-learning and the hippocampus in fact-learning [4] and [25] . Short training is primarily processed by the hippocampus, whereas prolonged training recruits the dorsal striatum. Interestingly, if the prelimbic medial prefrontal cortex is lesioned in rats, even short training leads to habit formation [26], reminiscent of the flies with blocked MB output. To my knowledge, habit formation has never been shown in any invertebrate model system before. This discovery entails that models for addiction and other compulsive disorders can now also be developed in the fly.

Combining the tools developed in the approach of localizing memory traces [17] and [19] with the experimental separation of operant and classical learning components [1], Drosophila has now entered the stage where we can start to unravel not only where memories are stored but also how and where basic neural subsystems interact to accomplish efficient learning in more ethologically relevant situations, without compromising generalization or prematurely engaging habit formation. Research on Drosophila has provided key insights into mechanisms of classical learning that are evolutionary conserved. The utility of this model system has now been extended to the study of complex learning situations comprising multiple, interacting learning systems on the behavioral, circuit, and genetic level. These studies expand a growing body of literature that simultaneously engaged memory systems can act both cooperatively and antagonistically.

Acknowledgments