Establishing the degree to which an individual larva seeks out or avoids an odorant requires repeated measurements of that larva’s response to the odor. We developed a Y-maze assay (Buchanan et al., 2015; Werkhoven et al., 2019) to repeatedly test an individual’s olfactory preference. The Y-mazes (Figure 1A) are constructed from agarose with channels slightly larger than the larvae, allowing free crawling only in a straight line (Heckscher et al., 2012; Sun and Heckscher, 2016). An individual larva travels down one channel and approaches the intersection with the other two branches of the maze. Here, the larva is presented with odorized air (or in this work, air containing CO₂) in one branch and pure air in the other. The larva then chooses and enters one of the two branches. This choice may be immediate or the result of a longer process in which the larva samples both channels and even reverses (Figure 1—video 2, Figure 1—video 3). When the larva reaches the end of its chosen channel, a circular chamber redirects it to return along the same channel to the intersection to make another choice. Custom computer vision software detects the motion of the larva while computer controlled valves manipulate the direction of airflow so that the larva is always presented with a fresh set of choices each time it approaches the intersection (Figure 1A, Figure 1—video 1).

We first sought to determine the suitability of this assay for measuring innate behavior. Drosophila larvae avoid carbon dioxide (CO₂) at all concentrations (Faucher et al., 2006; Jones et al., 2007; Kwon et al., 2007; Gershow et al., 2012). We presented larvae with a choice between humidified air and humidified air containing CO₂ each time they approached the central junction. At the 18% concentration used throughout this work, larvae with functional CO₂ receptors chose the CO₂-containing channel about 25% of the time. The probability of choosing the CO₂ containing channel increased as CO₂ concentration in that channel decreased (Figure 1F). Gr63a¹ (Jones et al., 2007) larvae lacking a functional CO₂ receptor were indifferent to the presence of CO₂ in the channel (Figure 1B), as were animals in which the CO₂ receptor neurons were silenced (Gr21a>Kir.21), indicating that larvae responded to the presence of CO₂ and not some other property of the CO₂ containing air stream. Silencing the Mushroom Body (OK107>Kir2.1) did not impair innate CO₂ avoidance.

Activation of the DAN-i1 pair of mushroom body input neurons has been shown to act as a reward for associative learning (Saumweber et al., 2018; Thum and Gerber, 2019; Schleyer et al., 2020; Weiglein et al., 2019; Eschbach et al., 2020b). In these experiments, the conditioned odor was innately attractive, but CO₂ is innately aversive. We wondered whether pairing DAN-i1 activation with CO₂ would lessen or even reverse the larva’s innate avoidance of CO₂.

To train larvae in the same Y-maze used to measure preference, we manipulated the valves so that the entire chamber was either filled with humidified air or with humidified air mixed with additional CO₂, independent of the position of the larva, which was not tracked during training. At the same time, we activated DAN-i1 neurons expressing CsChrimson using red LEDs built in to the apparatus. For some larvae, we activated DAN-i1 when CO₂ was present (paired, Figure 1D). For others, we activated the reward neurons when only air was present (reverse-paired, Figure 1D). Each training cycle consisted of one 15 s CO₂ presentation and one 15 s air presentation, with DAN-i1 activated for the entirety of the CO₂ (paired) or air (reverse-paired) presentation phase. The training protocols schematized in Figure 1D were repeated for 20 successive cycles. Thus, for instance, in the reverse-paired scheme CO₂ offset at t=15s coincided with reward onset, and the reward offset at t=30s coincided with CO₂ onset at t=0 of the subsequent cycle.

For each larva, we first measured naive preference and then preference following training. We found that in the paired group, larvae became indifferent to CO₂ presentation following 20 training cycles (Figure 1D, DANi1>CsChrimson, Paired). We did not observe any change in preference in the reverse-paired group (DANi1>CsChrimson, Reverse-Paired). Nor did we observe a preference change following paired training for genetically identical animals not fed all-trans-retinal (ATR), a necessary co-factor for CsChrimson function (DANi1>CsChrimson, Paired ATR-). Animals fed ATR but not exposed to red light failed to show a preference shift (DANi1>CsChrimson, No Training). Larvae of the parent strains fed ATR and given paired training showed no preference shift (Effector Control, Driver Control). To control for possible effects of DAN-i1 activation, we activated DAN-i1 in 15 s intervals without presenting CO₂ at all during the training (DANi1>CsChrimson, DAN w/o CO₂); these larvae showed no shift in preference.

Taken together these results show that the change in CO₂ preference requires activation of the DAN-i1 neurons and is not due to habituation (Twick et al., 2014; Das et al., 2011; Larkin et al., 2010), red light presentation, or other aspects of the training protocol. In particular, the paired and reverse-paired group experienced identical CO₂ presentations and DAN-i1 activations with the only difference the relative timing between CO₂ presentation and DAN-i1 activation.

Activation of DAN-i1 coincident with CO₂ presentation decreased larvae’s subsequent avoidance of CO₂. Formally, this admits two possibilities: the larva’s preference for CO₂ increased because CO₂ was presented at the same time as the reward or because CO₂ predicted the reward. To test whether learning was contingent on coincidence or prediction, we carried out an additional set of experiments. As before, we first tested innate preference, then presented 20 alternating cycles of 15 s of CO₂ followed by 15 s of air. However, this time during the conditioning phase, we either activated DAN-i1 7.5 s after CO₂ onset, in which case CO₂ predicted DAN-i1 activation, or 7.5 s before CO₂ onset, in which case CO₂ predicted withdrawal of DAN-i1 activation.

In both cases, DAN-i1 was activated in the presence of CO₂ for 7.5 s and in the presence of air alone for 7.5 s. If learning depended only on the coincidence between reward and CO₂ presentations, both should be equally effective at generating a change in preference. In fact, we only found an increase in CO₂ preference following training in which the CO₂ predicted the reward (Figure 1D).

Next, we asked whether reward prediction alone was sufficient to establish a memory, or if coincidence between CO₂ and DAN-i1 activation was also required. We altered the training protocol to present 15 s of CO₂ followed by 60 s of air. Some larvae were rewarded by activation of DAN-i1 in the 15 s immediately following CO₂ presentation (Forward Paired), while others were rewarded in the 15 s immediately prior to CO₂ presentation (Backwards Paired). For a third group of larvae, CO₂ was presented both before and after reward presentation (reward between CO₂ presentations). At no time was DAN-i1 activated in the presence of CO₂, but in the first group CO₂ predicted DAN-i1 activation while in the others it did not. We found an increased CO₂ preference for animals in this first group only (Figure 1E), indicating that reward prediction is both necessary and sufficient for learning in this assay.

In other associative conditioning experiments using DAN-i1 activation as a reward, decreased attraction to the odor was observed in the reverse-paired groups (Saumweber et al., 2018; Thum and Gerber, 2019; Schleyer et al., 2020). In our experiments, we did not see any evidence of increased aversion in the reverse-paired groups.

Untrained larvae avoided CO₂. After 20 cycles of paired, offset-after, or forward-paired training, larvae no longer avoided CO₂, but they also did not seek it out. We wondered whether it might be possible to train larvae to develop an attraction to the innately aversive CO₂. In other contexts, reward via activation of 3 DANs (DAN-i1, DAN-h1, DAN-j1 - whether DAN-h1 is present in second instar larvae, used in this study, is presently unreported) labeled by the 58E02-Gal4 line has been reported to produce strong learning scores (Saumweber et al., 2018; Rohwedder et al., 2016; Lyutova et al., 2019; Schleyer et al., 2020). We repeated the training protocol, substituting 58E02 activation for DAN-i1 activation alone, but did not see an increased preference following training compared to DAN-i1 activation alone (Figure 1D, 58E02>CsChrimson).

Next, we asked how varying the CO₂ concentration might affect animals’ performance in the assay. We presented lower concentrations of CO₂ both during the training and testing phases, and found that decreasing the CO₂ concentration decreased innate avoidance of CO₂. In all cases, following training, larvae lost avoidance to CO₂ but none showed statistically significant attraction (Figure 1F).

We investigated how change in preference for CO₂ following associative conditioning with DAN-i1 activation depended on the amount of training. As in the previous experiments, we first measured the innate preference, then trained each larva using repeated cycles alternating pure and CO₂ containing air, while activating DAN-i1 in concert with CO₂ presentation. In these experiments, however, we varied the number of training cycles an individual larva experienced. We found that as a group, larvae that had experienced more training chose the CO₂ containing channel more often (Figure 2A).

Our data showed that increasing the amount of training increased overall preference for CO₂ up to a saturation point. But what was the mechanism for this change? Did each cycle of training increase each larva’s preference for CO₂ by some small amount, with the effect accumulating over repeated training (graded learning)? Or did some larvae experience a dramatic preference change – from naive to fully trained – with each cycle of training, with the number of fully trained larvae increasing with training repetitions (quantized learning)?

Either quantized or graded learning can explain the shift of mean preference of a population (Gallistel et al., 2004); to differentiate between the modes of learning, we examined repeated decisions made by individual animals, measurements that were impossible in previous larval assays. For each larva, we quantified the change in CO₂ preference before and after training. Figure 2B shows a histogram of larva preference (the fraction of times an individual larva chose the CO₂ containing channel) after training, grouped by the number of cycles of training a larva received.

Larvae that received no training (0 cycles) formed a single population that chose CO₂ 27% of the time. Larvae that were trained to saturation (20 training cycles) also formed a single group centered around 52% probability of choosing CO₂. Both the graded and quantized learning models make the same predictions for these endpoints, but their predictions vary starkly for the intermediate cases. A graded learning model predicts that all larvae that received the same amount of training would form a single group whose mean preference for CO₂ would increase with increasing training. A quantized learning model predicts that larvae that have received the same amount of training will form two discrete groups (‘trained’ and ‘untrained’) with fixed centers whose means do not depend on the amount of training. With increased training an increasing fraction of larvae would be found in the trained group.

We fit the distributions of preference following conditioning to graded and quantized learning models. In the graded model, the preference was represented by a single Gaussian distribution whose mean and variance were a function of amount of training (orange, Figure 2). In the quantized model, the preference was represented by two Gaussian distributions; the fraction of larvae in each population was a function of the amount of training (teal, Figure 2).

We found that the data were better described by the quantized learning model (Table 4): larvae form two discrete groups, with the fraction in the trained group increasing with each cycle of additional training. The centers of the two groups do not vary with the amount of training, a point made most clear by considering the preference after training of all larvae taken together regardless of the amount of training received (Figure 2C), which shows two well defined and separated groups. From these data, we concluded that the effect of our associative conditioning on an individual larva is to either cause a discrete switch in preference or to leave the initial preference intact.

Next we asked what effect, if any, associative conditioning had on larvae that retained their innate preferences following training. Whether humans form associative memories gradually through repeated training or learn in an all-or-none manner has been the subject of debate in the Psychology literature (Roediger and Arnold, 2012); recent electrophysiological measurements in humans supports the all-or-none hypothesis (Ison et al., 2015). If learning is all-or-nothing, then if a larva has received training but has not yet expressed a behavioral switch, it is the same as if the larva has received no training at all. In this case, with every training cycle, regardless of past experience, every untrained larva will have the same probability of learning: ρ, and the effect of training can be described by a particularly simple equation(1)nu⁢(i+1)=nu⁢(i)-ρ⁢nu⁢(i)where nu⁢(i) are the number untrained larvae following i cycles of training. Note that ρ can depend on the training protocol or other external variables, but it does not depend on the past training experiences of the larvae, and can be considered a fixed constant for a given experimental condition. The solution to this equation is an exponentially decaying population of untrained larvae. For a given initial population nu⁢(0) of untrained larvae,(2)nu⁢(i)=(1-ρ)i⁢nu⁢(0)

Any so-called memoryless process like this produces an exponential decay of the initial population (Apostol, 1969). Meanwhile, processes with memory can produce other distributions. For example, if the training were cumulative, we would expect a threshold effect: as the number of cycles of training increased from 0, most larvae would initially remain untrained until a critical number of cycles (n_c) were reached and there would be a sudden shift to a mostly trained population. While a process with memory can also produce exponential decay (e.g. if each larva required a fixed n_c cycles of training to learn, and n_c was itself exponentially distributed), all memoryless processes must produce an exponential decay, and exponential decay is therefore an indicator of a memoryless (all-or-none) process.

Our fit to the quantized learning model produces an estimate of the fraction of larvae that remain untrained following training. We plotted the fraction of untrained larvae vs. number of training cycles and saw that the fraction of larvae in the untrained group exponentially decreased with increasing training (Figure 2E, note logarithmic y-axis on insert). We then fit the population distributions to an all-or-none quantized learning model in which the effect of a single training cycle was to train a fixed fraction of the remaining untrained larvae (purple, Figure 2). This model fit the data better than the graded learning model and almost as well as the original quantized learning model (in which the fraction of untrained larvae was fit separately to each group) despite having fewer parameters than either model. According to standard selection rules (BIC and AIC), the all-or-none quantized model best describes the data (Table 4).

Reversal learning, in which the reward contingency is reversed, and extinction, in which the conditioned stimulus is presented without reward, experiments explore cognitive flexibility. Previous experiments with both adult Drosophila (Tully et al., 1990; Ren et al., 2012; Wu et al., 2017; Vogt et al., 2015) and larval (Mancini et al., 2019) Drosophila demonstrated a reversal learning paradigm. Extinction has been demonstrated in adult flies (Felsenberg et al., 2017; Felsenberg et al., 2018; Schwaerzel et al., 2002) but not in larvae.

To test for extinction, we again first measured an individual larva’s CO₂ preference and then carried out associative conditioning for a given number (2-10) of training cycles. Next instead of immediately testing the larva’s new preference for CO₂, we exposed the larva to an extinction phase – 18 cycles of alternating CO₂ and air without any optogenetic reward. Following the extinction period, we tested larvae as usual to measure their changed preference for CO₂. As a control against the effects of increased CO₂ exposure, we also performed habituation experiments, which were the same as the extinction experiments, except the 18 unrewarded cycles were presented prior to the rewarded training cycles. The extinction and habituation protocols are schematized in Figure 3A.

When we compared the ‘habituated’ groups of larvae to larvae trained for the same number of cycles without habituation or extinction, we found that unrewarded CO₂ presentation prior to training had no effect on the eventual preference change (Figure 3B). This was unsurprising, as the initial testing period already offered a number of unrewarded CO₂ presentations. In contrast, unrewarded CO₂ presentations following training reversed the effect of training; for small (2 or 3 cycles) amounts of training, the reversal was almost complete (Figure 3B).

We previously observed that associative conditioning produced a discrete and quantized change in CO₂ preference. Here, we found that extinction following training greatly reversed the effects of conditioning. We wondered whether larvae that had been subject to both training and extinction reverted to their original CO₂ preference or to an intermediate state. In the former case, we would expect to see a bimodal distribution of preference change following training and extinction, while in the latter we would see a third group of larvae. This group would be most evident in experiments where two to four cycles of training were followed by extinction, as these had the largest deficit in the fraction of trained larvae compared to habituated larvae that received the same amount of training. We examined the preferences of all larvae following two to four cycles of training, grouped by whether they were normally trained, habituated, or subject to extinction (Figure 3C). In all cases, we observed two groups with the same central means and no evidence of a third intermediate group. We concluded that larvae subject to training then extinction reverted to their ‘untrained’ state.

We wondered whether larvae would still learn if they received additional training directly following extinction. As before, we measured the innate preference, presented three paired training cycles followed by the extinction phase. At this point, based on our previous results, larvae would have returned to their initial innate avoidance of CO₂. We then immediately presented three more paired training cycles before behavioral testing (Figure 3—figure supplement 1A). We found that following this training-extinction-training protocol, both the population preference for CO₂ (Figure 3—figure supplement 1B) and the fraction of larvae trained (Figure 3—figure supplement 1C) were comparable to larvae that had been trained three times without extinction cycles.

Given the relatively short duration of training and the ability of unrewarded CO₂ presentations to extinguish prior training, we wondered whether larvae might change their CO₂ preferences over the course of the hour-long post-training behavioral readout. In particular, might the apparent threshold of 50% attraction be an artifact due to a short period of attraction to CO₂ followed by a longer period of indifference or modest avoidance?

To test for a short period of increased attraction immediately following training, we reanalyzed the results of experiments with 2, 5, and 20 cycles of paired training. In each case, we compared the initial 10 min of the post-training choice assay to the final 50 min (Figure 3—figure supplement 2A) and found no significant difference between the initial and final periods for any of these training conditions. We then compared the mean preference over the first five choices (representing five unrewarded CO₂ presentations) made by each larva to the mean preference in the remainder of the experiment and again found no significant difference (Figure 3—figure supplement 2B). Breaking the behavioral readout into equal 15 min periods also reveals no strong temporal signal (Figure 3—figure supplement 2C–E).

Finally, we developed a new protocol to minimize the possible effects of extinction over the course of the behavioral readout, using the fact that training following extinction can re-establish a lost memory (Figure 3—figure supplement 1). We trained each larva with 10 paired cycles (5 min of training), then tested their preference for 15 min, then presented another 10 paired training cycles followed by another 15 min of testing, for a total of 4 training and testing blocks (Figure 3—figure supplement 3). The results were comparable to when we presented a single training block followed by an hour-long test period. Thus, we concluded that the apparent limit of 50% population preference to CO₂ following training was not due to the long time-scale of the behavioral readout.

Studies in adult (Tully et al., 1990; Yin et al., 1995; Margulies et al., 2005) and larval (Honjo and Furukubo-Tokunaga, 2005; Honjo and Furukubo-Tokunaga, 2009; Widmann et al., 2016; Khurana et al., 2009; Eschment et al., 2020; Aceves-Piña and Quinn, 1979) Drosophila have identified distinct memory phases: short-term memory (STM), middle-term memory (MTM), long-term memory (LTM), and anesthesia-resistant memory (ARM). LTM and ARM are consolidated forms of memory controlled by partially separate molecular and anatomical pathways (Isabel et al., 2004; Jacob and Waddell, 2020; Wu et al., 2012). ARM is resistant to anesthetic agents (Quinn et al., 1974); LTM requires cAMP response element-binding protein (CREB)-dependent transcription and de-novo protein synthesis, while ARM does not (Yin et al., 1995; Perazzona et al., 2004). Adults have been shown to retain memories for up to a week (Yin et al., 1995). Larvae trained to associate odor with electric shock form memories that persist for at least 8 hr (Khurana et al., 2009). Odor-salt memories have been shown to persist for at least 5 hr (Widmann et al., 2016; Eschment et al., 2020) and can be either ARM or LTM, depending on the initial feeding state of the larva.

We sought to determine whether we could create consolidated memories that would persist overnight, and if so, whether these memories represented ARM or LTM. As in previously described experiments, we first tested each larva’s individual preference in the Y-maze assay, trained it to associate CO₂ presentation with DAN-i1 activation, and then measured its individual preference again following training. After this second round of testing, we removed the larva from the apparatus and placed it on food (without ATR) overnight. The next day, we placed the larva back in the Y-maze and again tested its preference for CO₂, without any additional training.

We found that following 20 cycles of training, larvae became indifferent to CO₂ and this indifference persisted to the next day. Similarly, we found that most larvae switched preference following five cycles of training and retained that preference overnight. Larvae that received no training or 20 cycles of unpaired training had no change in CO₂ preference immediately following training or the next day (Figure 4B).

We had previously shown two cycles of training caused roughly half the larva to change preference immediately after training. We decided to use this partition to verify a correlation between immediate and long-term memories; we expected that larvae initially in the ‘trained’ group would also form a ‘trained’ group the following day. However, while we found that two cycles of training were sufficient to cause some larvae to become indifferent to CO₂ immediately following training, when we tested these larvae the next day, we found that all had reverted to their initial avoidance of CO₂.

There were two possible explanations for this reversion. Perhaps, two cycles of training were sufficient to form a short term memory, but more training was required to induce a long-term memory. Or perhaps the testing period, in which larvae were exposed repeatedly to CO₂ without reward, reversed the two-cycle training. To control for the latter, we modified the experimental protocol. We tested each larva’s innate preference, presented two training cycles, and then immediately removed the larva to food overnight, without any further testing. When we tested these larvae the next day, we found that they showed decreased avoidance of CO₂. This indicated that two cycles of training were sufficient to form a memory lasting overnight, but that immediate exposure to unrewarded CO₂ following this short training interval likely reversed the effects of training, an effect we observed in Figure 3. When larvae were trained for 20 cycles, omitting the testing had no effect on these larvae’s preferences the following day.

To confirm that extinction could explain the failure to form a persistent memory, we exposed larvae to three cycles of paired training, then 18 cycles of extinction (as in Figure 3) and then removed them to food overnight before testing their preferences the next day. As expected, these larvae avoided CO₂ as much the next day as they did prior to training (Figure 4B, Ext Post-Train).

We wondered whether memories that had consolidated overnight would be more resistant to extinction. We repeated the previous experiment with a single modification. As before, we tested the larva’s initial preference and trained it with three cycles of rewarded CO₂ presentation. This time, we immediately removed the larva to food following training. The next day, we returned the larva to the Y-maze and presented the extinction phase of 18 unrewarded CO₂ presentations prior to testing for CO₂ preference. We found that in this case, larvae still expressed an increased preference for CO₂ despite the extinction phase (Figure 4B, Ext Pre-Test). The only difference between the two experiments was whether we attempted extinction immediately after training or the next day. Thus, we concluded that overnight consolidation made memories more resistant to extinction.

ARM can be distinguished from LTM because the latter requires de novo protein synthesis and can be disrupted by ingestion of the translation-inhibitor cycloheximide (CXM). To incorporate CXM feeding, we modified our protocols. Instead of raising larvae on ATR supplemented food, we raised them on standard food and then fed them with ATR supplemented yeast paste for 4 hr prior to the experiment (ATR+/CXM-). For some larvae (ATR+/CXM+), we also added CXM to the yeast paste. In this way, we could be sure that if ATR+/CXM+ larvae ingested enough ATR to allow for CsChrimson activation of DAN-i1, they must have also ingested CXM as well. To further verify CXM ingestion, we placed ATR+/CXM+ and ATR+/CXM- larvae on clean food and allowed them to continue development. 95% of ATR+/CXM- larvae pupated, while only 45% of ATR+/CXM+ larvae pupated.

Following the 4 hr feeding period, ATR+/CXM+ and ATR+/CXM- larvae were treated identically. As in the previously described experiments, we first tested each larva’s individual preference in the Y-maze assay, trained the larva 20 times to associate CO₂ presentation with DAN-i1 activation, and then measured its individual preference again following training. After this second round of testing, we removed the larva from the apparatus and placed it on food (without ATR or CXM) overnight. The next day we placed the larva back in the Y-maze and again tested its preference for CO₂, without any additional training.

We found that performances tested immediately and 16 hr after training were both unaffected by CXM treatment. Following 20 cycles of training, larvae from both groups (ATR+/CXM+; ATR+/CXM-) became indifferent to CO₂ and this indifference persisted to the next day (Figure 4C). This suggests that the memory formation was independent of de novo protein synthesis.

In adult Drosophila, whether ARM or LTM is formed depends on the training protocol (Tully et al., 1990; Tully et al., 1994; Yin et al., 1995; Yu et al., 2006; Bouzaiane et al., 2015). ‘Massed’ training, in which all conditioning occurs in rapid sequence without rest intervals, results in ARM, while ‘spaced’ training, in which the conditioning occurs in blocks separated by intervals of time, produces LTM. Our training protocol more closely resembles massed training, so it seemed sensible that it would produce ARM. To see if we could instead develop LTM, we established a spaced training protocol. Larvae received two paired cycles of training, followed by a 15-min interval of air-presentation only; this sequence was repeated five times. To keep the total length of the experiment within a (covid-related) limited daily time window, we did not test the larvae immediately after training but only the next day.

Prior to spaced training, both ATR+/CXM- and ATR+/CXM+ larvae avoided CO₂ to the same degree. We found that 1 day following spaced training, ATR+/CXM+ larvae continued to avoid CO₂, while ATR+/CXM- larvae did not. This indicated that spaced training formed a memory whose retention was disrupted by CXM. To verify that spacing the trials was essential to forming a protein-synthesis dependent memory, we duplicated the experiments exactly, except we presented 10 cycles of training en masse, rather than spacing them. In this case, both ATR+/CXM- and ATR+/CXM- larvae expressed learned indifference to CO₂1 day following training (Figure 4C).

As an alternate to CXM feeding, LTM (but not ARM) formation can also be disrupted through use of hs-dCREB2-b, a heat-shock inducible dominant-negative repressor of transcription mediated by dCREB2-a. (Perazzona et al., 2004; Yin et al., 1995). Specifically, in adult flies expressing hs-dCREB2-b, memory retention is disrupted in a heat-shock-dependent manner following spaced, but not massed training (Yin et al., 1995). We therefore repeated our long-term memory experiments in larvae that in addition to expressing Chrimson in DAN-i1 neurons also carried the hs-dCREB2-b transgene. Massed and spaced training were carried out as previously described, using larvae raised on ATR supplemented food, except that some larvae (HS) received a 30 min heat-shock (at 37 C), followed by a 30 min recovery period (at 25 C) immediately prior to the beginning of the experiment (i.e. prior to the initial testing of naive preference). Preference for CO₂ was tested prior to training, immediately following training, and the next day, following an overnight rest on food without ATR.

We found that, congruent with our CXM experiments, the day after spaced training, heat-shocked larvae (Figure 4D,S 10x, HS) avoided CO₂ to the same extent they did prior to training, while larvae that were not heat-shocked (Figure 4D,S 10X, No HS) retained learned indifference; larvae that received massed training (Figure 4D,M 10x, HS and No HS) retained their learned indifference overnight, regardless of heat-shock. These results are consistent with similar experiments in adult flies (Yin et al., 1994; Yu et al., 2006).

Immediately following spaced training (80 min after the initiation of the spaced training protocol), heat-shocked larvae continued to avoid CO₂, showing that memory formation was impaired on a relatively short timescale. This is consistent with previous work in the larva, where dCREB2-b expression induced memory deficits beginning 30 min following a single 30 min training cycle (Honjo and Furukubo-Tokunaga, 2005), and immediately following 125 min of spaced training (Widmann et al., 2016). In those experiments, neither training protocol was shown to induce a persistent long-term memory.

We introduced a Y-maze assay capable of measuring the olfactory preferences of individual larval Drosophila and of in situ associative conditioning. We found that when larvae learn to associate CO₂ with reward neuron activation, the result is a switch from innate avoidance to learned indifference, with no intervening states. We demonstrated a protocol to form stable protein-synthesis dependent long term memories. This provides a strong starting point for ‘cracking’ a complete olfactory learning circuit.

We used SLA three-dimensional printing to create microfluidic masters for casting (Karagyozov et al., 2018; Chan et al., 2015). Masters were designed in Autodesk Inventor and printed on an Ember three-dimensonal printer (Autodesk) using black prototyping resin (Colorado Photopolymer Solutions). After printing, masters were washed in isopropyl alcohol, air-dried, and baked at 65C for 45 min to remove volatile additives and non-crosslinked resin. 4% agarose (Apex Quick Dissolve LE Agarose, Cat #20-102QD, Genesee Scientific) was poured over the masters and allowed to solidify; then mazes were removed from the mold. Agarose Y-mazes were stored in tap water before use.

The mazes are 1 mm in depth. Each channel is 1.818 mm in length and 0.4 mm in width, and ends in a circular chamber (radius = 1 mm) which redirects larva back to the intersection. An inlet channel (depth = 0.1 mm, length = 1.524 mm, width = 0.1 mm) to the circular chamber connects to tubing for our network of air, CO₂, and vacuum sources.

Individual larvae were selected for size and placed into a Y-maze using a paintbrush. The Y-maze was placed into a PDMS (Sylgard 184, 10:1 base:curing agent) base, where tubing was secured. The Y-maze and base were encased in a dark custom-built box. Larvae were monitored under 850 nm infrared illumination (Everlight Electronics Co Ltd, HIR11-21C/L11/TR8) using a Raspberry Pi NoIR camera (Adafruit, 3100), connected to a Raspberry Pi microcomputer (Raspberry Pi 3 Model B+, Adafruit, 3775). Experiments were recorded using the same camera, operating at 20 fps. Eight copies of the assay were built, to assay the behaviors of multiple larvae in parallel.

Pressure for air, CO₂, and vacuum were controlled at the sources (for vacuum regulation: 41585K43, McMaster-Carr; for pressure regulation: 43275K16, McMaster-Carr). CO₂ and air were humidified through a bubble humidifier. Vacuum, air, and CO₂ tubing to individual assays were separated through a block manifold after pressure control and humidification (BHH2-12, Clippard).

The CO₂ concentration was controlled by a resistive network of tubing connected to the air and CO₂ sources. This inexpensive alternative to a mass-flow controller produced a stable ratio of CO₂ to air that was consistent from day to day and independent of the overall flow rate. The direction of flow was controlled by solenoid pinch valves (NPV2-1C-03–12, Clippard), actuated by a custom circuit we designed.

Custom computer vision software detected the location of the larva in real time. Based on the larva’s location, computer controlled valves manipulated the direction of airflow so that the larva was always presented with a fresh set of choices each time it approached the intersection. The software randomly decided which channel would contain air and which contained air mixed with CO₂.

In each maze, one channel was selected to be the outlet for flow and the other two were inlets. An individual larva began in the outlet channel and approached the intersection of the Y-maze, then chose to enter either an inlet branch containing air with CO₂ or an inlet branch containing air only. When the larva’s full body entered the chosen channel, software recorded the larva’s choice of channel. When the larva reached the end of that channel and entered the circular chamber, valves switched to turn off CO₂ and to switch vacuum to the channel containing larva, making that channel the outlet. The CO₂ remained off (the larva experienced only pure air) until the larva exited the circular chamber. When the larva exited the circular chamber and proceeded towards the intersection, CO₂ was introduced to one randomly selected inlet channel.

Software recorded the location of the larva at every frame (approx 20 Hz); the direction of airflow in the maze (which channel(s) had air; which channel had CO₂ mixed with air, if any; and which channel had vacuum); and all decisions the larva made. We recorded when larvae entered or left a channel, and whether that channel presented CO₂. Larvae could take three actions as they approached the intersection: choose the channel containing air with CO₂ (scored as APPROACH); choose the channel containing pure air (scored as AVOID); or move backwards into their original channel before they reach the intersection. If a larva backed up and reentered the circular chamber it departed from before reaching the intersection, the software reset and presented the larva with a fresh set of choices when it next left the circle. We did not score backing up as a choice of either CO₂ or air.

Following an hour of testing, larvae were trained in the same Y-maze assay used to measure preference. During the training period, unless described otherwise, each 30 s training cycles alternated 15 s of CO₂ presentation, where both inlet channels contained a humidified mix of CO₂ and air; followed by 15 s of air presentation, where both inlet channels had humidified air alone. This cycle was repeated some number of times (specified for each experiment in the figures). Red LEDs (Sun LD, XZM2ACR55W-3) integrated into the setup were used to activate CsChrimson synchronously with CO₂ presentation (paired) or air presentation (reverse-paired).

The volume of the flow chamber was 11.68 mm³ and the volume of the tubing downstream of the valves is approximately 214 mm³. The flow rate exceeded 560 mm³/s, and the state of the chamber was taken to be the same as the state of valves.

Following training, larvae were tested for one hour in an identical scheme to that previously described for the naive measurement.

After larvae were placed into the Y-maze, larva were left in the maze in the dark for a minimum of 5 min. If a larva was not moving through the maze after 5 min, the larva was replaced before the experiment began. If larvae stopped moving through the maze during the first hour of testing, larvae were removed from the maze before training and results were discarded. This happened infrequently (approximately 5% of experiments).

For experiments in Figure 1D, reward presentation was offset from CO₂ onset. 30 s training cycles alternated 15 s of CO₂ presentation, where both channels contained a mix of CO₂ and air; followed by 15 s of air presentation, where neither channel had CO₂. Red LEDs are used to activate CsChrimson for 15 s. For some larvae, reward onset occurred 7.5 s after CO₂ presentation; for others, reward onset occurred 7.5 s before CO₂ presentation. For all experiments of this type, larvae were presented with 20 cycles of training.

For experiments in Figure 1E, 75-second training cycles alternated 15 s of CO₂ presentation, where both inlet channels contained a mix of CO₂ and air with 60 s of air presentation. For some larvae, reward presentation occurred immediately following CO₂ termination for 15 s. For others, reward presentation occurred 15 s prior to CO₂ onset, and reward presentation was terminated upon CO₂ presentation. For a third group of larvae, we rewarded larvae for 15 s between two CO₂ presentations. In this case, 15 s of CO₂ presentation was followed by 15 s of reward presentation in the absence of CO₂, followed by another 15 s of CO₂ presentation. After the second presentation, there was a 30 s air gap before the cycle repeated. For all experiments of these types, larvae were presented with 20 cycles of training.

For experiments in Figure 3, we used either an extinction or habituation protocol during training. For both types, larvae were tested for 1 hr to determine their innate CO₂ preference in the method described above.

For extinction experiments, larvae were trained in the same Y-maze used to measure preference. 30 s training cycles alternate 15 s of CO₂ presentation, where both channels contain a mix of CO₂ and air; followed by 15 s of air presentation, where neither channel had CO₂. Red LEDs were used to activate CsChrimson synchronously with CO₂ presentation. This training cycle was repeated some number of times (specified for each experiment above). Immediately after training, we presented the larva with 18 cycles of repeated CO₂/air exposure (15 s of CO₂ followed by 15 s of air; repeat) with no reward pairing. After these extinction cycles, larva preference for CO₂ was tested for one hour.

Habituation experiments were done exactly as for extinction experiments, except that the 18 unrewarded cycles of repeated CO₂/air exposure were presented immediately prior to the training cycles.

For experiments in Figure 4B, we tested each larva’s initial preference for one hour, then presented three rewarded paired training cycles. For some larvae (‘Extinction Post-Train’), we then immediately presented 18 extinction cycles, removed the larvae to food overnight as described above, and then tested their preferences for one hour the next day. For another set of larvae (’Extinction Pre-Test’), we removed the larvae to food immediately following training. The next day, after the larvae were cleaned and inserted into the Y-maze, they were exposed to 18 extinction cycles immediately prior to testing their CO₂ preferences for 1 hr.

For the memory retention experiments of Figure 4, testing and training followed identical procedures as above to establish larva preference. After the second round of testing testing, the larvae were removed from the Y-maze assay with a paintbrush and transferred to an individual 4% agar plate (30 mm, FB0875711YZ Fisher Scientific), with yeast paste added. Larvae were kept in the dark at 18 C for approximately 20 hr. Prior to experiments the next day, larva were removed from the agar plate and washed in water before being placed in a new Y-maze. Larvae were then tested for CO₂ preference for one hour as previously described. In all experiments in which larvae were removed from the apparatus and later retested, they were placed in the same apparatus but with a new agar Y-maze. Out of 443 larvae placed on agar plates to be tested the following day, 439 larvae were recovered and retested. The four lost larvae were excluded from analysis.

For specified experiments in section Figure 4, larva were raised on ATR- food plates at 25C for 48 hr. Second instar larvae were separated from food using 30% sucrose solution and washed in water. Four hours prior to experiments, larvae were transferred to an agar dish with yeast paste for feeding. Yeast paste was made with either a solution of 35 mM cycloheximide (CXM, Sigma Aldrich C7698) and 0.1 mM all-trans-retinal (ATR, Sigma Aldrich R2500) in 5% sucrose (ATR+/CXM+); or 0.1 mM ATR in 5% sucrose (ATR+/CXM-). To verify CXM ingestion, we placed ATR+/CXM+ and ATR+/CXM- larvae not selected for experiments back on clean food and allowed them to continue development. 95% of ATR+/CXM- larvae pupated, while only 45% of ATR+/CXM+ larvae pupated. Before the experiment, larvae were transferred to an empty petri dish and washed with tap water before being placed into a maze. Except where noted, the same experimental protocol was followed as for non-CXM overnight memory.

For the CXM experiments in section Figure 4, larvae were trained with either a massed or spaced training protocol. The 20x massed training protocol was as previously described for other experiments in Figure 4.

In the 10x spaced training protocol, larvae were first tested for 1 hr to determine their initial CO₂ preference. They then received two cycles of paired DAN-i1 activation with CO₂ presentation (15 s of CO₂ presentation paired with reward, followed by 15 s of air presentation), followed by 15 min of air presentation. This was then repeated five times (10 activations total). In these experiments, we did not test the larvae immediately following training but instead removed them to food and tested their preferences the next day only.

The 10x massed training protocol was identical to the 10x spaced training protocol, except training consisted of 10 sequential cycles of paired DAN-i1 activation with CO₂ presentation (15 s of CO₂ presentation paired with reward, followed by 15 s of air presentation, repeated 10 times). As in the spaced training experiments, larvae were removed to food immediately following training, and their preferences were tested the next day only.

Petri dishes with larvae were placed in an oven at 37°C for 30 min. The petri dish was placed in a water bath in the oven and covered to preserve humidity and to ensure ATR+ larvae were kept in the dark. The larvae were then removed to 25°C for a 30 min recovery period before experiments. Experiments began immediately after the recovery period. Larvae were kept in the dark at all times during this protocol.

Larvae were tested for CO₂ preference prior to training, immediately following training, and the next day, after being kept overnight on food without ATR. Some larvae in the spaced training groups were not active immediately following training. In the heat-shocked group, 11 out of 24 larvae made less than five decisions during the immediate test; in the not-heat-shocked group, 6 out of 22 larvae made less than five decisions during the immediate test. For this scheme, larvae were in the Y-maze for longer than previous experiments, as the spaced training protocol is 80 min (compared to approximately 10 min or less for the standard massed training). All larvae were retested the following day, even if the larva did not make many decisions when tested immediately following training. After the overnight rest period, all non-heat-shocked and 20/24 heat-shocked larvae were active in the final test period. Inactive larvae were included in the analysis and in the bootstrapping of error bars, but contributed little to the population measure because of the few decisions made.

There are a number of parameters that can be adjusted in our assay, including the identity and concentration of gas used as a CS, the concentration and timing of ATR feeding, the period, duty cycle, and number of CS and US presentations, and duration of behavioral readouts before and after training. We began with our normal protocol for optogenetic activation (Gepner et al., 2015; Gepner et al., 2018): eggs were laid on ATR supplemented food, and larvae were raised in the dark. We somewhat arbitrarily chose a 30 s, 50% duty cycle applied for 20 cycles as our standard for paired training presentation; we began with DANi1>CsChrimson based on previous work (Saumweber et al., 2018; Thum and Gerber, 2019; Schleyer et al., 2020; Weiglein et al., 2019; Eschbach et al., 2020b), and the fact that CsChrimson can be activated via red light without provoking a strong visual response. We then adjusted the CO₂ concentration to maximize the contrast between CO₂ preference before and after training. From this basic platform, we changed as little as possible while manipulating the parameter of interest, for example we maintained the 30 s 50% duty cycle paired training while changing the number of cycles, or we maintained 20 cycles while varying the temporal sequence of CS and US presentation, or we used exactly the same 30 s, 50% duty cycle, 20 cycle paired protocol while changing the driver to RF58E02.

The probability of choosing the CO₂ containing channel was scored for individual larvae and for populations as(3)ProbchooseCO2=#APPROACH#APPROACH+#AVOID

The population average was determined by dividing the total number of times any larva in the population chose the CO₂ containing channel by the total number of times any larva chose either channel. In other words, larvae that made more decisions contributed more heavily to the average.

The number of larva and total number of approach and avoid decisions made by larvae for each type of experiment is shown in Table 2. Error bars for 'probability choose CO₂’ data displays and all significance tests in the figures were generated by bootstrapping.

For each experimental set, we performed the bootstrapping as follows. If there were X larvae from that experiment, we selected X larvae with replacement from that set. Then, from each larvae selected, we selected with replacement from the decisions that larvae had made. For example, if the larvae had made Y ‘approach’ and Z ‘avoid’ decisions, we selected (Y+Z) decisions with replacement from that set to represent the larvae. We then calculated the population average from this generated set of animals. We generated 10,0000 numerical replicates using this bootstrapping method. Error bars were the standard deviation of these replicates. Note that in each replicate, the same animals were included in each (e.g. trained and untrained) group.

A p-value p<x indicates that at least x fraction of these replicates ended with the same ranking result (e.g. p<0.01 between trained and untrained would indicate that in at least 9900 out of 10,000 replicates, the trained group had a larger CO₂ preference than the untrained group or vice versa). These p-values are included in the ‘Hierarchical Bootstrap’ column of Table 3.

We also performed a non-hierarchical bootstrap, in which animals were resampled but their decisions were not, preserving any correlations between decisions. In this case, we generated 10,000 numerical replicates by selecting with replacement from that set of larvae; the actual sequence of decisions made by the resampled larvae was then used without further resampling. A p-value p<x indicates that at least x fraction of these replicates ended with the same ranking result. These p-values are included in the ‘Bootstrap Animal Only’ column of Table 3. In Table 3, we also show p-values for the Fisher’s exact test, which treats every decision as independent, and the Mann-Whitney u-test, which treats every larva in each group as a discrete measurement and does not account for differing numbers of decisions made by larvae.

To fit the data in Figure 2 to various models, we used a maximum-likelihood approach. First we grouped the larvae according to the number of cycles (n_c) of training they received. In each group, for each larva we quantified the number of decisions made following training. The number of decisions made by the jt⁢h larva that received n_c cycles of training was n⁢(nc,j) and the fraction of times the larva chose the CO₂ containing channel was p⁢(nc,j). Then we sought a set of parameters θ that maximized(4)∑nc∑jlog⁡P⁢(p⁢(nc,j)|n⁢(nc,j),θ)where P was the model specific-probability function. For instance, for the quantized learning (two-Gaussian shifting fraction) model:(5)P(p(nc,j)|n(nc,j),θ)=fu(nc)𝒩(p(nc,j),μu,σ~μu∗(1−μu)n(nc,j))+(1−fu(nc))𝒩(p(nc,j),μt,σ~μt∗(1−μt)n(nc,j))where 𝒩⁢(x,μ,σ)=12⁢π⁢σ2⁢exp-(x-μ)22⁢σ2 and the parameters θ are(6)θ={μu,μt,σ~,fu⁢(0),fu⁢(1),fu⁢(2),fu⁢(3),fu⁢(4),fu⁢(5),fu⁢(10),fu⁢(20)}

The parameter σ~ represents an adjustment to the expected variance due to counting statistics. If all larva chose randomly and independently from the two channels with a fixed probability p¯ of choosing CO₂, then we would expect that the number of times the CO₂ containing channel would be binomially distributed. For ease of computation, we approximated the binomial distribution as a normal distribution. In this case, the probability density of observing p⁢(nc,j) given n⁢(nc,j) would be normally distributed with mean p¯ and variance(7)σ2=(p¯)⁢(1-p¯)n⁢(nc,j)

In fact, we found that after choosing a CO₂ containing channel, both naive and trained larvae are less likely to choose the CO₂ containing channel the next time they approach the intersection. Because the choices are not independent, the variance of the mean of a series of choices is not given by Equation 7. Instead, we modeled the variance as(8)σ2=σ~2⁢(p¯)⁢(1-p¯)n⁢(nc,j)where σ~ was a global fit parameter in the shifting and exponential fraction models and in the shifting mean model a function of the amount training. This formulation preserves the properties that the variance should increase as the mean probability of choosing CO₂ approaches 50% and should be larger when fewer decisions are averaged together. However, if we instead just assume a single global σ, the results of our analysis (that the exponential fraction model is preferred) are unchanged.

In the graded learning (single Gaussian with shifting mean and variance) model, µ and σ were allowed to change as a function of training. The probability of an individual observation was(9)P⁢(p⁢(nc,j)|n⁢(nc,j),θ)=𝒩⁢(p⁢(nc,j),μ⁢(nc),σ⁢(nc)n⁢(nc,j))and the parameters were(10)θ={μ⁢(0),σ⁢(0),μ⁢(1),σ⁢(1),μ⁢(2),σ⁢(2),μ⁢(3),σ⁢(3),μ⁢(4),σ⁢(4),μ⁢(5),σ⁢(5),μ⁢(10),σ⁢(10),μ⁢(20),σ⁢(20)}

The exponential fraction model is identical to the quantized learning model, except that the fraction of untrained larvae is an exponentially decreasing function of the number of training cycles:(11)fu⁢(nc)=λncand the parameters were(12)θ={μu,μt,σ~,λ}

These models were then fit to the data by maximizing the log-likelihood of the observed data set using the MATLAB function fmincon. The predictions of these fits are shown in Figure 2. These results are summarized in Table 4, along with the Aikake and Bayes Information Criterion, AIC and BIC, which are used to compare models with different numbers of parameters. According to both AIC and BIC, the exponential fraction model is strongly favored.

Throughout the paper 'Fraction of larvae trained’ represents the best fit to the two Gaussian shifting fraction model. The error bars represent the uncertainty in the model fit. Specifically, they represent the range of f over which(13)log⁡P⁢(data|θ0,f)≥log⁡P⁢(data|θ0,f0)-12where f is the fraction of trained larvae, f₀, is the best fit fraction of trained larvae, and θ0 represents the best fit of the remainder of the parameters, which are not adjusted.