For imaging DANs, we crossed females homozygous for both UAS-GCaMP3 (Tian et al., 2009) and TH-GAL4 (Friggi-Grelin et al., 2003) with males homozygous for mb247-DsRed; mb247-DsRed (Riemensperger et al., 2005) so that DsRed expression in the MBs could be used as a morphological landmark. We refer to the F1 flies as TH>GCaMP3. To drive GCaMP3 expression in the KCs we crossed homozygous male UAS-GCaMP3 flies with homozygous female OK107-GAL4 flies (Connolly et al., 1996). We refer to the F1 flies as OK107>GCaMP3. For imaging, we anesthetized a single fly on ice, fixed it in a holder, opened the fly head dorsally and covered the preparation with saline (Figure 2A).

We applied electric shocks (four 1.5 s long 90 V pulses) to the fly's legs by placing the fly on a custom-build copper grid (Figure 2A and Supplementary Figure 4B). We recorded the shock strength received by an individual fly using a bridge circuit (sampling rate: 16 kHz; Figure 3A, Supplementary Figure 4C, and Supplementary Table 2). We used 1-butanol (BUT) and 4-methylcyclohexanol (MCH) diluted in mineral oil (MO; BUT 1:500, MCH 1:1,000) as odorant stimuli, which we presented as 10 s long stimuli with a custom-build stimulator (Szyszka et al., 2011). We measured the dynamics of the odorant stimuli with a photo ionization detector (miniPID, Aurora Scientific Inc.). Rapid odorant stimulus termination (Supplementary Figure 4A) allowed us to distinguish between responses to the olfactory CS and to the electric shock US.

We measured the fluorescence of GCaMP3 and DsRed at a sampling rate of 5 Hz using a confocal laser scanning microscope, equipped with a 20 × water-immersion objective. Note that in some physiological paradigms animals are pre-exposed to the stimuli until the induced neuronal response strength becomes stable (Hige et al., 2015a). Since stimulus pre-exposure is not common in the behavioral odor—shock conditioning paradigm (Tully and Quinn, 1985), we did not pre-expose our flies to the applied stimuli.

First, we measured the DAN and KC responses to odorants and electric shock prior to training (Figure 1, trial 1–4, unpaired group). In DANs and KCs, odorants and electric shocks induced calcium responses across the MB compartments (Figure 2C). Both response dynamics and amplitudes differed between DANs and KCs, and across MB compartments (Figure 2D, Supplementary Figures 1, 2). Odorants induced DAN responses and KC responses in all MB compartments (Figure 2E). In general, responses were strongest to BUT (1-butanol), weaker to MCH (4-methylcyclohexanol) and weakest to MO (the solvent mineral oil). KCs in the β- and β'-lobe showed off-responses to the offset of odorants (Figure 2D and Supplementary Figure 2). KCs responded stronger to MO than DANs (Figure 2E). DANs responded stronger to electric shocks than KCs (Figures 2C–E), although both fly lines encountered equal shock strength (Supplementary Figure 4C). Note however, that the responses were normalized to the BUT responses, and that individual KCs could respond stronger to shock than to odorants. DAN responses to electric shock increased logarithmically with the electric current flow through the fly [linear regression line slopes ranged from 0.024 to 0.715 ΔF/F per log (nA) in α1/α'1 and γ1, respectively; Figure 3B and Supplementary Table 2], except in the MB-compartment α1/α'1, and were strongest in γ1 and γ2 (Figure 2E). There was a positive correlation between KC response strength and received current in the β'-lobe and in the γ3 compartment (Figure 3C and Supplementary Table 2).

To investigate the effect of odor—shock trace conditioning on DAN and KC responses to the CS and to the US, we combined odor—shock trace conditioning with calcium imaging (Figure 1). To this end, we adopted the trace conditioning protocol from Galili et al. (2011). As Galili and colleagues, we used BUT as the only CS, because trace conditioning with BUT yielded the best results in behavioral experiments. Flies of the paired group received six CS-US pairings, whereas flies of the unpaired control group received six unpaired CS and US presentations. After the training, we recorded responses to the CS, MCH, and the US alone in both groups. The unpaired group served as a control for non-biological effects of the experimental procedure (bleaching of GCaMP, changes in stimulus strength; Supplementary Figures 4, 5), as well as for non-associative effects of conditioning (sensitization, habituation, pseudo-conditioning; Tully, 1984). However, note that our experimental design does not allow us to differentiate between the proportional contribution of non-biological and non-associative effects to the measured neuronal response strength. As flies of the paired and the unpaired group received the same number of CS and US, but with a different stimulus timing, differences in DAN and KC responses between the paired and the unpaired group reflect associative plasticity.

DAN responses to the CS increased in the paired relative to the unpaired group (Figures 4A,B and Supplementary Figure 3A) in eight out of nine MB compartments (γ1–5, β'1–2, and junction), but neither in the ellipsoid body nor in the fan-shaped body (Figures 4, 6A, and Supplementary Figure 3A). This training-induced associative plasticity in DANs became visible after the first training trial (trial 5 in Figure 1). We did not detect a correlation between changes in the CS-induced response strength and the US strength that a fly received in each trial (Supplementary Figure 3E).

In addition to the relative increase of CS-induced DAN responses in the paired compared to the unpaired group, there was an increase in the absolute DAN response strength from the first training trial to the post training trial in three of the γ-lobe compartments [mixed-effect model for repeated-measures ANOVA, γ1: F_{(1, 20)} = 4.39, p < 0.05; γ4: F_{(1, 22)} = 6.23, p < 0.05; γ5: F_{(1, 23)} = 6.67, p < 0.05].

In contrast to DANs, KC responses to the CS decreased in the paired relative to the unpaired group, and this decrease occurred in three out of nine MB compartments (β'1–2, and junction; Figures 4A,B, 6A, and Supplementary Figure 3B). Training-induced associative plasticity in KCs became visible after the second training trial (trial 6 in Figure 1), one trial later than in DANs.

In the unpaired group, both DAN and KC responses to the CS decreased over trials in most compartments (Figure 4B and Supplementary Figures 3A,B). This response decrease could reflect non-associative plasticity due to repeated stimulus exposure, or it could be due to a decrease in CS concentration or bleaching of the calcium sensor (Supplementary Figure 4A). Indeed, additional control experiments revealed that DAN responses were sensitive to odorant concentration (100% compared to 60% of the initial odorant concentration): DANs generally responded stronger to the higher than to the lower odorant concentration (Supplementary Figure 5).

In both DANs and KCs, the responses to the control odorant MCH did not differ between the paired and unpaired group (Figure 4B and Supplementary Figures 3A,B), showing that the associative change in the DAN response was CS-specific and is not generalized to a different odor.

If Drosophila DANs were to encode the US prediction error similar to mammalian DANs (Schultz et al., 1997), then their US-induced responses should change in the course of the conditioning as the responses to the CS become stronger and the CS becomes predictive for the occurrence of the US. However, US-induced DAN (and KC) responses did not differ between the paired and unpaired group (Figure 4C, Supplementary Figures 1, 2, Supplementary Figure 3C,D). Moreover, unlike mammalian DANs, after conditioning Drosophila DANs did not change their activity when the US was omitted after presenting the CS (Figure 4A and Supplementary Figure 1). Therefore, the population of TH-GAL4-labeled DANs appears not to encode the US prediction error.

Taken together, associative plasticity differed between DANs and KCs in three ways: (1) Responses to the CS increased in DANs and decreased in KCs in the paired relative to the unpaired group. (2) Associative plasticity in the CS-induced responses occurred in DANs in eight out of nine compartments, whereas it occurred in KCs in three out of nine compartments (Figure 6A). (3) Associative plasticity in the response to the CS was visible after the 1st training trial in DANs and after the 2nd training trial in KCs. The lack of associative plasticity in US-induced responses suggests that DANs and KCs do not encode the US prediction error during trace conditioning.

It has been proposed that a conditioning-induced change in the cross-compartmental pattern of MBON activity induced by the CS reflects a change in the valence of the CS (Aso et al., 2014b; Owald and Waddell, 2015). Because DANs provide compartment-specific input to the MBs and drive the conditioning-induced changes of MBON responses (Cohn et al., 2015; Hige et al., 2015a), we asked whether and how odor—shock trace conditioning changes the cross-compartmental pattern of DAN activity (Figure 5A). To quantify the change in cross-compartmental activity patterns, we translated the recorded compartment responses for each fly into a vector that comprises the response strength of nine compartments. We quantified the dissimilarity between two activity patterns as the geometric angle between the two respective vectors in a 9-dimensional space. This yields a metric of pattern dissimilarity that is independent of response strength.

To analyze how the CS-induced spatial activity pattern changes during training, we compared the activity pattern for each trial with the pattern induced by the CS stimulation before the training (Figure 5B, “CS vs. 1st CS”). In both the paired and the unpaired group, the CS-induced activity patterns diverged from the initial pattern. This divergence was stronger in the unpaired than in the paired group (Figure 5B, “CS vs. 1st CS”). We next asked whether the US-induced activity patterns also change during conditioning. We therefore compared each US-induced activity pattern with the pattern induced by the first US stimulation (Figure 5B, “US vs. 1st US”). With repeated stimulation, the US-induced activity patterns diverged from the initial pattern. However, this divergence was less than in CS-induced activity patterns, and we found no difference in the effect between the paired and unpaired group (Figure 5B, “US vs. 1st US”). Thus, electric shock-induced activity patterns in DANs were unaffected by associative plasticity.

During odor—shock conditioning, the CS acquires the potential to elicit the conditioned response in Drosophila (Tully, 1984). Aversive conditioning experiments in Drosophila larvae revealed that the CS does not only become aversive, but that it actually gains predictive power (Schleyer et al., 2015). We therefore asked whether such an associative change in the predictive power of the CS could be reflected in the CS-induced spatial activity pattern of DANs. For example, does the CS-induced activity pattern become more similar to the US-induced activity pattern? To quantify whether the similarity between CS- and US-induced activity patterns changed during training, we compared each CS-induced activity pattern with the mean activity pattern induced by the US (Figure 5B, “CS vs. mean US”). In the paired group the mean angles between the CS- and US-induced activity patterns ranged between 0.6 and 1.0 rad (34–57°). In contrast, in the unpaired group the mean angles between the CS- and US-induced activity patterns diverged from 0.7 to 1.5 rad (40–86°). In the paired group the CS- and US-induced activity patterns were equally similar before and after training. However, in the unpaired group the CS- and US-induced activity patterns became less similar (Figure 5B, “CS vs. mean US”). In contrast to DANs, the cross-compartmental pattern of KC activity did not exhibit any associative changes (Supplementary Figure 6). In sum, associative plasticity preserved the degree of similarity between CS-induced and US-induced cross-compartmental activity patterns in DANs.

Previous studies suggested that DANs in the MB lobes respond strongly to electric shock and weakly to odorants (Riemensperger et al., 2005; Mao and Davis, 2009). The compartment-resolved analysis of our calcium imaging data refines this picture: We confirm that DANs of all imaged compartments respond to both electric shock and odorants, and we show that their relative response strength to odorants and electric shock differs across compartments. For example, DANs innervating γ1 responded stronger to electric shock than to odorants, while DANs innervating β'2 responded equally strong to odorants and electric shock (Figure 2E). We found the strongest DAN responses to electric shock in the compartments γ1 and γ2. These compartments receive input from PPL1-γ1pedc and PPL1-γ2α'1 DANs that mediate electric shock reinforcement (Aso et al., 2010, 2012). In all compartments, except in α1/α'1, the DAN response strength correlated positively with the current strength encountered by individual flies (Figure 3B). Thus, DANs are capable to encode the strength of the electric shock US (Mao and Davis, 2009), and this property may account for the positive dependence between electric shock strength and learning performance in flies (Tully and Quinn, 1985).

Calcium responses in KCs differ between MB lobes (Turner et al., 2008; Lin A. C. et al., 2014), and they differ between the compartments of a given lobe, possibly due to compartment-specific modulation by DANs and MBONs (Tanaka et al., 2008; Aso et al., 2014a; Cohn et al., 2015). KCs in γ2 and γ3 responded strongest to odorants (Figure 2A and Supplementary Figure 2), confirming the results of Cohn et al. (2015). KCs generally responded only weakly to electric shocks. Previously published strong KC responses to electric shock may be because electric shocks were applied to the flies' abdomen rather than to their legs, which might have resulted in a stronger stimulation (Akalal et al., 2010).

The associative strengthening of DAN responses to the olfactory CS (as compared to the unpaired control group; Figure 6A), confirms the previous report by Riemensperger et al. (2005). Associative plasticity occurred in those DANs that innervate the MBs (PPL1 and PAM cluster DANs; note that the used TH-GAL4 driver line covers only a small subpopulation of PAM neurons (Aso et al., 2012; Pech et al., 2013) but not in DANs that innervate the central complex (PPL1 and PPM3 cluster DANs). This is in line with the established role of MB innervating-DANs in associative memory formation, while central complex-innervating DANs are involved in behaviors such as locomotion (Kong et al., 2010), wakefulness (Liu et al., 2012), arousal (Ueno et al., 2012), and aggression (Alekseyenko et al., 2013), and are therefore not expected to show odor—shock conditioning-induced plasticity.

In contrast to previous studies (Wang et al., 2008; Akalal et al., 2010, 2011), we did not find an associative increase in KC calcium responses to the CS in the MB-lobes after odor—shock conditioning (Figure 6A). This may indicate either a difference between trace conditioning (this study) and standard conditioning (published data), or a difference in other experimental parameters that may also account for inconsistencies in the published effects of odor—shock conditioning (Zhang and Roman, 2013; Boto et al., 2014; Hige et al., 2015a).

The associative decrease in KC responses in the β'-lobe compartments (Figure 4B) is in line with previous studies that showed conditioning-induced depression of KC-to-MBON synapses (Cohn et al., 2015; Hige et al., 2015a). Therefore, we propose that the associative decrease in KC responses to the CS reflects a presynaptic depression at KC-to-MBON synapses in β'-lobe compartments (Figure 6B).

What is the site of neuronal plasticity that underlies the relative increase in DANs' responses to the olfactory CS? Riemensperger et al. (2005) proposed that DANs get odorant-driven excitatory input via a MBON feedback loop that is strengthened during odor—shock conditioning. However, the DAN population is composed of different neuron types that, to our knowledge, do not share a common input neither from MBONs nor from other neurons that could explain the global associative plasticity across MB compartments. Because KCs presumably provide the only common odor-driven input to all MB-innervating DANs, we suggest that the site of associative plasticity is located in a KC-to-DAN synapse. Indeed, KC-to-DAN synapses have recently been reported in Drosophila (Cervantes-Sandoval et al., 2017). Associative increase in CS-induced DAN responses occurred despite unaltered or decreased KC responses in the same compartment. This suggests that the associative plasticity occurs post-synaptic in DANs and is not inherited from KCs (Figure 6B). Note, that the associative changes in DANs' response strength could be influenced by lateral modulation via other compartments, as has been shown in the study of Cohn et al. (2015).

What is the neuronal substrate of CS-US coincidence detection in DANs and KCs? Drosophila trace conditioning depends on dopamine receptor-triggered signaling in KCs (Shuai et al., 2011), as is the case for standard conditioning (Kim et al., 2007; Qin et al., 2012). However, the CS-US coincidence detection mechanism in trace conditioning is unknown (Galili et al., 2011; Shuai et al., 2011; Dylla et al., 2013). In standard conditioning the CS-induced increase in KCs' calcium concentration coincides with the US-(dopamine)-induced second messengers, which is thought to synergistically activate the rutabaga adenylyl cyclase (Duerr and Quinn, 1982; Dudaí et al., 1983; Tomchik and Davis, 2009; Gervasi et al., 2010), and ultimately alters the strength of KC-to-MBON synapses (Dubnau et al., 2001; McGuire et al., 2001; Schwaerzel et al., 2003; Séjourné et al., 2011; Pai et al., 2013; Zhang and Roman, 2013; Aso et al., 2014b; Bouzaiane et al., 2015; Cohn et al., 2015; Hige et al., 2015a; Owald et al., 2015). This mechanism would not work for trace conditioning, because (1) at the time the US occurs, CS-induced increase in KCs' calcium concentration is back to baseline levels (Figure 2D, Supplementary Figure 2), and (2) trace conditioning does not involve the rutabaga adenylyl cyclase (Shuai et al., 2011). We therefore hypothesize that a non-rutabaga adenylyl cyclase (Adams et al., 2000) or a protein kinase C (Choi et al., 1991; Widmann et al., 2016) could serve as a molecular coincidence detector for the CS trace and the US. For example, the CS-induced calcium and dopamine signaling could lead to a sustained activation of an adenylyl cyclase or protein kinase C in KCs, which then would increase synergistically and drive synaptic plasticity during the US-induced dopamine signaling.

DAN responses to odorants and associative strengthening of DAN responses to the CS-odorant are not included in current models of associative learning in the MB (Busto et al., 2010; Owald and Waddell, 2015). However, associative plasticity is a common feature of US-mediating neurons, which occurs in mammalian and Drosophila DANs (Schultz et al., 1997; Riemensperger et al., 2005), and in an octopaminergic neuron in honey bees (Hammer, 1993).

What could be the function of odorant-induced responses and odor—shock conditioning-induced plasticity in DANs? MB-innervating DANs strengthened their response to the CS (as compared to the unpaired group) during odor—shock conditioning (Figure 4B), in line with Riemensperger et al. (2005). However, other than in monkey DANs (Montague et al., 1996; Schultz et al., 1997; Steinberg et al., 2013), we did not observe associative plasticity in DANs' response to the US (Figure 4C). Our data therefore support the idea that Drosophila DANs encode predictive power of the CS, e.g., US-prediction, but not the US-prediction error during classical conditioning (Riemensperger et al., 2005).

We found shock-induced responses and associative plasticity in DANs that are not involved in odor—shock conditioning, for example in DANs innervating β'1, γ3, γ4, and γ5 (Aso et al., 2010, 2012). This suggests that those DANs serve a function in aversive odor learning which is not captured by the commonly applied conditioning paradigms. For example, the relative strengthening of CS-induced responses could mediate reinforcement during second-order conditioning, in which a previously reinforced CS₁ can act as US in subsequent conditioning of a second CS₂ (Pavlov, 1927). As Drosophila is capable of second-order learning (Brembs and Heisenberg, 2001; Tabone and de Belle, 2011), this theory can be tested in behavioral experiments: if associative strengthening of DAN responses to the CS underlies CS₁-induced reinforcement in second-order conditioning, then preventing associative plasticity in DANs, or blocking their output during CS₂–CS₁ pairing should abolish second-order conditioning.

The occurrence of CS-induced responses and associative plasticity in most of the MB-innervating DANs suggests that the separation between the CS- and US-pathway and between different US-pathways is less strict than suggested in current models of associative learning in the MB. Associative plasticity in the spatial pattern of CS-induced DAN responses (Figure 5) makes them a potential neuronal substrate for encoding the US identity (Galili et al., 2011; Burke et al., 2012; Das et al., 2014; Lin S. et al., 2014; Cohn et al., 2015; Huetteroth et al., 2015) in CS-US memories and the predictive power of a CS.

Our data revealed similar response properties and plasticity rules across Drosophila DANs in the γ- and β'-lobe. This contrasts with their anatomical (Tanaka et al., 2008; Mao and Davis, 2009; Aso et al., 2014a) and functional heterogeneity (Krashes et al., 2009; Aso et al., 2010, 2012; Galili et al., 2011; Berry et al., 2012; Burke et al., 2012; Plaçais et al., 2012; Das et al., 2014; Cohn et al., 2015; Aso and Rubin, 2016; Felsenberg et al., 2017), which indicates yet undiscovered mechanisms and functions of DAN plasticity. Note, that we could not test whether the flies learned in the imaging setup, as currently no behavioral readout exists for odor—shock conditioning during physiological experiments. Nevertheless, since we used a conditioning protocol and stimulus application comparable to an established behavioral paradigm, we believe that the associative plasticity in neuronal responses that we found underlies behavioral associative plasticity. Therewith our data lay the foundations for causal studies on the function of associative plasticity in DANs.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.