Serotonergic-dependent awareness is required for trace conditioning in honeybees Authors: Catherine Macri, Marco Paoli, Andrea Alamia, Michel Gho, Guillaume Isabel, Martin Giurfa Published: 2025 ABSTRACT SummaryIn humans, trace conditioning—where a temporal gap separates conditioned (CS) and unconditioned (US) stimuli—requires awareness and is sensitive to distraction, unlike delay conditioning. Here, we show that honeybees exhibit comparable features, suggesting cognitively demanding processes akin to awareness. Using the olfactory conditioning of the proboscis extension response, we found that bees learned in both paradigms, but trace conditioning resulted in weaker acquisition and memory. High-resolution video analysis revealed delayed response latencies and altered licking dynamics during trace learning. Strikingly, trace—but not delay—conditioning was impaired by visual and mechanosensory distractors. Neuropharmacological blockade of serotonin signaling selectively disrupted trace conditioning, implicating serotonergic circuits in bridging CS-US intervals. A computational model of the bee olfactory network supported this interpretation. Our findings reveal that bees engage higher-order cognitive mechanisms under temporal uncertainty, suggesting that core features of awareness may be evolutionarily conserved across distant taxa. Introduction Awareness is considered a facet of consciousness, often described as the specific experience of being cognizant of a particular event at a specific moment.1 It can be either focused (i.e., paying attention to a particular stimulus) or diffuse (such as being generally aware of one’s context without focusing on any specific detail).2 Awareness, therefore, represents active engagement with particular aspects of one’s mental state at any given moment. In humans, awareness of stimulus contingencies can be studied under controlled experimental conditions through participants’ reports of their subjective experiences.3 A classic example is found in studies that use distractors during trace and delay conditioning.4,5,6 These two forms of Pavlovian conditioning differ in the timing between the conditioned stimulus (CS) and the unconditioned stimulus (US). In delay conditioning, the CS is presented and remains active until the US is introduced, with both stimuli partially overlapping. In trace conditioning, a gap separates the CS and US presentations so that for conditioning to occur, the subject must maintain a memory “trace” of the CS during the gap. While this difference seems minor, it has substantial implications for learning: typically, trace conditioning proves more difficult to learn than delay conditioning7,8; yet, it can provide a significant advantage in scenarios where danger or reward signals are delayed. In addition, the two conditioning types depend on different neural substrates, as delay conditioning does not require an intact hippocampus and neocortex,9,10 while trace conditioning does.4,11,12,13,14 Comparative studies of delay and trace conditioning in humans and other mammals frequently use the conditioning of the eyeblink response.15,16 In this paradigm, a CS (e.g., a tone or a light) predicts an aversive air puff to the eye, eliciting an eyeblink to prevent the air puff. Repeated pairings result in the acquisition of a conditioned eyeblink response timed to the CS that predicts the air puff.17 Auditory versions of delay and trace eyeblink conditioning have been used to explore the role of awareness. In,6 during eyeblink conditioning, participants watched a silent movie and were later questioned about its content and their learning from the conditioning experience (contingency between tone/light and air puff). In trace conditioning, only participants who became aware of the CS-US relationship successfully learned the task, whereas awareness was not required for delay conditioning.6 To investigate this issue further, participants underwent trace conditioning while performing a distracting task that involved detecting a series of three odd digits in a sequence and pressing a button to report them. These participants failed to develop awareness and performed significantly worse than a control group that did not experience the distractive task, learning the association between tone and air puff. These findings suggest that awareness is essential for successful trace conditioning.6,18 The question of whether awareness is necessary for trace conditioning but not delay conditioning has not been investigated so far in the honeybee, a model organism for cognitive research among invertebrates.19,20,21 Honeybees can be conditioned in the laboratory using a Pavlovian protocol known as olfactory conditioning of the proboscis extension response (PER).22,23 In this protocol, harnessed bees learn to associate an odor (CS) with a sucrose solution (US), and the conditioned response is the proboscis extension to the odorant that predicts sucrose delivery. While delay conditioning is commonly used in these experiments,22,23,24 trace conditioning variants have been employed to study neural activity during the trace interval.25 However, it remains unclear whether distraction impairs trace conditioning in honeybees, which would indicate the need for an awareness-like state. In this study, we combined trace and delay conditioning with various types of distractors to assess their effects on learning success and memory retention in honeybees. Additionally, we developed a neural model to explain the success of trace conditioning and tested it using a neuropharmacological approach. Our findings show that distractors impair trace, but not delay, conditioning, and indicate that serotonergic signaling may play a crucial role in successful trace learning. These results imply that, similar to humans, honeybees may rely on an awareness of the CS-US contingency during trace conditioning, and that serotonergic signaling plays a crucial role in this process, suggesting that this cognitive state may extend to invertebrates. Results Delay and trace conditioning induce different learning and retention performances In this experiment, we aimed to compare the acquisition and retention performance of honeybees conditioned using delay and trace conditioning. During each trial, a single odorant (CS) was paired with a 50% (w/w) sucrose solution (US). Two odorants, 1-nonanol and 2-hexanol, were used as CS, with different groups of bees assigned to each. In delay conditioning, the odorant was presented for 6 s, followed by the sucrose solution for 3 s, with a 5 s interstimulus interval (ISI), resulting in a 1 s overlap between the stimuli (Figure 1A). In the trace conditioning protocol, the odorant was presented for 0.5 s, followed by the sucrose solution for 3 s, with the same ISI of 5 s as in delay conditioning, resulting in a CS-US gap of 4.5 s (Figure 1A). This trace conditioning protocol was adopted from a previous study that optimized stimulus timing for olfactory trace conditioning in honeybees.25 In both protocols, successful learning was demonstrated by a proboscis extension response (PER) to the conditioned odorant (Figure 1B). To prevent residual odorants from interfering with the trace conditioning paradigm, an air extractor was placed behind the bees. Its effectiveness was confirmed using a photoionization detector (Figure S1).Figure 1Honeybees learn odor-sucrose associations under both delay and trace conditioning(A) In delay conditioning, the conditioned stimulus (CS) and unconditioned stimulus (US) overlap in time (blue area). In trace conditioning, there is a stimulus-free gap between the offset of the CS and the onset of the US, known as the “trace” period (orange area). ISI: interstimulus interval, defined as the interval between the CS and US onsets.(B) Example of a harnessed bee displaying the proboscis extension reflex (PER). Image from Matsumoto et al.24(C) Percentage of bees that exhibited a conditioned response during conditioning trials. Shaded areas represent the 95% confidence intervals for performance in delay (blue) and trace (orange) conditioning. Sample sizes (N) are indicated. ∗∗∗, p < 0.001, GLMER, Group × Trial interaction. Honeybees learn odor-sucrose associations under both delay and trace conditioning (A) In delay conditioning, the conditioned stimulus (CS) and unconditioned stimulus (US) overlap in time (blue area). In trace conditioning, there is a stimulus-free gap between the offset of the CS and the onset of the US, known as the “trace” period (orange area). ISI: interstimulus interval, defined as the interval between the CS and US onsets. (B) Example of a harnessed bee displaying the proboscis extension reflex (PER). Image from Matsumoto et al.24 (C) Percentage of bees that exhibited a conditioned response during conditioning trials. Shaded areas represent the 95% confidence intervals for performance in delay (blue) and trace (orange) conditioning. Sample sizes (N) are indicated. ∗∗∗, p < 0.001, GLMER, Group × Trial interaction. After conditioning, retention performance was assessed by presenting the trained odorant and a novel odorant (to test memory specificity) in a random sequence without reward, 1 h (mid-term memory, MTM) and 24 h (long-term memory, LTM) after conditioning. MTM is known to be independent of protein synthesis, whereas LTM depends on this process.26,27 Separate groups of bees, conditioned in parallel, were used for each retention test. Learning (i.e., the percentage of bees responding to the CS) did not differ significantly among the eight trained groups (four for delay conditioning and four for trace conditioning), which varied by the odorant used as the CS (1-nonanol or 2-hexanol) and the retention test time (1 h or 24 h) (GLMER, Odor effect, z = −1.08, p = 0.28; Retention time effect, z = −0.63, p = 0.53). Therefore, the learning performances were pooled and represented as two learning curves, showing the percentage of PER to the CS across trials: one for delay conditioning and the other for trace conditioning (Figure 1C). Under both conditioning regimes, bees increased their responses to the CS across trials (Cochran’s Q test, delay conditioning: Q = 196, df: 4 p < 0.001; trace conditioning: Q = 117, df: 4, p < 0.001). However, acquisition levels were significantly higher in delay conditioning compared to trace conditioning (GLMER, Group × Trial Interaction, z = −4.26, p < 0.001), indicating that trace conditioning was more cognitively challenging than delay conditioning. Similar results were obtained when varying the duration of the CS in comparable trace conditioning experiments (Figure S2), demonstrating that this parameter did not influence our results and conclusions. Retention performance, evaluated both as the percentage of bees showing PER to the CS and the novel odorant (NOd) (Figure S3A) and as the percentage of bees exhibiting specific memory (i.e., responding to the CS but not the NOd; Figure S3B), confirmed the greater difficulty of trace conditioning compared to delay conditioning. Twenty-four hours after conditioning, trace-conditioned bees exhibited lower long-term retention levels (Figure S3A) and specific memory (Figure S3B) than delay-conditioned bees. Delay and trace conditioning induce different response latencies and licking frequency Having established significant differences in learning and retention between delay and trace conditioning, we next sought to determine whether these two conditioning methods result in qualitative differences in PER, reflecting differences in the timing of the CS-US interval.28 To address this question, we trained two groups of bees: one with delay (Delay) and the other with trace conditioning (Trace, Figure 2A). Since these two groups differed not only in the temporal relationship between CS and US but also in learning performance and CS duration, we conducted two additional experiments to determine if lower learning performance and shorter odorant duration could influence PER dynamics. In one experiment, we conditioned a group of bees using delay-conditioning with a lower-quality sucrose reward (20% w/w), which should reduce learning performance (Delay20, Figure 2A). In the other experiment, we conditioned bees using trace conditioning featuring a longer CS duration (Trace5s, Figure 2A). The performances of bees from the four groups trained with 2-hexanol or 1-nonanol as a CS were pooled because learning did not differ significantly between odorants (Figure 2B, GLMER, Odor effect, z = 0.71, p = 0.48).Figure 2Learning performances and features of the conditioned response (PER) to the odor during the memory test(A) Experimental designs of the four groups of bees included in the video analysis: delay conditioning (Delay), delay conditioning with 20% sucrose (Delay20), and trace conditioning (with CS durations of 0.5 s or 5 s, labeled Trace and Trace5s, respectively).(B) Percentage of bees that exhibited a conditioned response during conditioning trials for the groups Delay, Delay20, Trace, and Trace5s. The shaded area around the curves represents the 95% confidence interval. ∗∗, p < 0.01, Group Effect, pairwise comparisons.(C) Schematic of the frame-by-frame PER analysis. Odorant delivery onset was detected via illumination of an LED on the Arduino board. Proboscis extension latency was calculated by counting the number of frames until maximal response and converting it to seconds. The number of licking events—instances where the bee extended its glossa—was also recorded.(D) Latency of maximum proboscis extension for each group. The full line indicates the median and the dotted lines the interquartile range ∗, p < 0.05, Wilcoxon test.(E) Number of licking events for each group. The boxplots show the median and the interquartile range∗∗, p < 0.01; ∗∗∗, p < 0.001, Wilcoxon test. Learning performances and features of the conditioned response (PER) to the odor during the memory test (A) Experimental designs of the four groups of bees included in the video analysis: delay conditioning (Delay), delay conditioning with 20% sucrose (Delay20), and trace conditioning (with CS durations of 0.5 s or 5 s, labeled Trace and Trace5s, respectively). (B) Percentage of bees that exhibited a conditioned response during conditioning trials for the groups Delay, Delay20, Trace, and Trace5s. The shaded area around the curves represents the 95% confidence interval. ∗∗, p < 0.01, Group Effect, pairwise comparisons. (C) Schematic of the frame-by-frame PER analysis. Odorant delivery onset was detected via illumination of an LED on the Arduino board. Proboscis extension latency was calculated by counting the number of frames until maximal response and converting it to seconds. The number of licking events—instances where the bee extended its glossa—was also recorded. (D) Latency of maximum proboscis extension for each group. The full line indicates the median and the dotted lines the interquartile range ∗, p < 0.05, Wilcoxon test. (E) Number of licking events for each group. The boxplots show the median and the interquartile range∗∗, p < 0.01; ∗∗∗, p < 0.001, Wilcoxon test. In terms of learning performances (Figure 2B), the Delay group consistently exhibited higher conditioning levels than the Trace group, as observed in the previous experiment, and also outperformed the two additional groups (Delay20 and Trace5s) (post hoc analysis of the Group effect, Delay20 vs. Delay, z = 3.44, p = 0.003, Trace vs. Delay, z = 3.95, p < 0.001, Trace5s vs. Delay, z = 2.94, p = 0.018). The three groups with lower performance (Trace, Trace5s, and Delay20) did not differ significantly from one another (post hoc analysis of the Group effect, Delay20 vs. Trace, z = 0.39, p = 0.98, Delay20 vs. Trace 5s, z = −0.80, p = 0.86, Trace vs. Trace5s, z = −1.25, p = 0.59). These results confirm our previous findings that trace conditioning is more cognitively demanding than delay conditioning (Figure 1C) and that CS duration does not affect trace learning efficiency (Figure S2B). Furthermore, they demonstrate that a lower-quality US in the Delay20 group resulted in lower learning performances comparable to those observed in the trace conditioning variants. One hour after conditioning, we recorded PER responses to odorant delivery in all four groups using a video camera positioned above each bee’s head. These video recordings enabled us to quantify both the latency to the maximal proboscis extension and the number of licking events (glossa extensions) (Figure 2C). The latency to maximal PER differed between the Delay and the three other groups (Figure 2D). Bees from the Delay group extended their proboscis earlier than those from the Trace group (Wilcoxon test, W = 236.5, p = 0.04), and the within-group variability in latencies was significantly lower in the Delay group compared to the Trace group (one-tailed F test, F26,25 = 0.05, p < 0.001). While the latency to maximal PER in the Trace5s group was similar to that of the Trace group (Figure 2D, W = 438, p = 0.74), it did not differ significantly from the Delay group either (W = 342, p = 0.18). Overall, delay conditioning elicited an almost automatic response with reduced variability, whereas trace conditioning resulted in more variable latencies, suggesting hesitation about when to extend the proboscis. This higher variability in PER to the expected US in the trace-conditioned group could indicate that some bees attempted to time the CS-US interval; however, it may also reflect lower learning performance. The latency to full PER in the Delay20 group (with a lower-quality 20% sucrose reward) did not differ from that of the Trace group (W = 348.5, p = 0.80), but was significantly different from that of the Delay group trained with the 50% sucrose solution (Figure 2D, W = 261.5, p = 0.05). Overall, these results suggest that the high variability observed in the responses of the trace-conditioned bees does not necessarily reflect the timing of the CS-US interval, but may instead result from reduced learning efficiency associated with trace conditioning. Considering now the number of licking events (Figure 2E), bees from the Trace group showed significantly fewer licking attempts than delay-conditioned bees (Trace vs. Delay, W = 540, p < 0.001, Trace vs. Delay20, W = 633.5, p < 0.001). However, the Trace5s group, conditioned with a 5 s CS, exhibited a significantly higher number of licking events than the Trace group conditioned with a 0.5 s CS (Figure 2E, W = 233, p = 0.003). These findings suggest that the absence of licking events initially observed in the trace-conditioned group with a 0.5 s CS is likely attributable to the short duration of the odorant predicting the reward, rather than being an inherent behavioral feature of trace-conditioned PER. Distractors significantly impair trace conditioning, but not delay conditioning We next investigated whether, similar to humans, trace conditioning requires an awareness-like state of stimulus contingencies, which may be dispensable in delay conditioning. To explore this, we examined whether a distractor stimulus differentially affects trace and delay conditioning. Specifically, we compared the acquisition and retention performances of bees conditioned under the two regimes in the presence of a brief visual distractor (a 100 ms, 50 Hz white flash) delivered at a random time within the CS-US interval during each trial (Figure 3A). Bees were placed within an isolation chamber to minimize contextual cues and enhance distractor salience (Figure 3B).Figure 3Distractors impair learning performance in trace, but not delay, conditioning(A) Experimental designs of distractor delivery for both delay and trace conditioning. The distractor lasted 100 ms and was randomly delivered within a 3-s window centered around the midpoint of the 5-s interstimulus interval (ISI).(B) 3D-printed isolation chamber used to perform PER conditioning while minimizing contextual cues and enhancing distractor salience.(C) Example of a harnessed bee being stroked on the back with a feather, serving as a mechanosensory distractor. The feather was motor-controlled and applied for exactly 100 ms.(D and E) Learning curves (% of PER to the CS) in delay-conditioned (D) and trace-conditioned (E) bees from control groups (blue for delay, orange for trace) and visually distracted groups (black).(F and G) Learning curves (% of PER to the CS) in delay-conditioned (F) and trace-conditioned (G) bees from control groups (blue for delay, orange for trace) and mechanosensory distracted groups (black). Shaded areas around the curves represent 95% confidence intervals. ∗∗∗, p < 0.001, GLMER, Distraction effect. Distractors impair learning performance in trace, but not delay, conditioning (A) Experimental designs of distractor delivery for both delay and trace conditioning. The distractor lasted 100 ms and was randomly delivered within a 3-s window centered around the midpoint of the 5-s interstimulus interval (ISI). (B) 3D-printed isolation chamber used to perform PER conditioning while minimizing contextual cues and enhancing distractor salience. (C) Example of a harnessed bee being stroked on the back with a feather, serving as a mechanosensory distractor. The feather was motor-controlled and applied for exactly 100 ms. (D and E) Learning curves (% of PER to the CS) in delay-conditioned (D) and trace-conditioned (E) bees from control groups (blue for delay, orange for trace) and visually distracted groups (black). (F and G) Learning curves (% of PER to the CS) in delay-conditioned (F) and trace-conditioned (G) bees from control groups (blue for delay, orange for trace) and mechanosensory distracted groups (black). Shaded areas around the curves represent 95% confidence intervals. ∗∗∗, p < 0.001, GLMER, Distraction effect. As in the previous experiment, no differences in learning were observed based on the odorant (1-nonanol or 2-hexanol) used as the CS (GLMER on delay-conditioned bees, Odorant as CS effect, z = 0.01, p = 0.99, GLMER on trace-conditioned bees, Odorant as CS effect, z = −0.84, p = 0.40). Therefore, performances were presented as a single pooled curve for each conditioning type. One hour after conditioning, responses to the CS, the NOd (to assess memory specificity), and the distractor were measured in the absence of reinforcement to assess whether competition between the CS and the distractor for association with the US occurred during learning. Figure 3D shows the learning performance of delay-conditioned bees, comparing those not subjected to the visual distractor (controls) with the distracted bees. In both groups, bees increased their conditioned responses across trials (Cochran’s Q test, delay conditioning control: Q = 116, df = 3, p < 0.001; delay conditioning distraction: Q = 122, df = 3, p < 0.001). No significant difference was found between the performance of the two groups (GLMER on Delay conditioned bees, Distraction effect, z = 0.013, p = 0.99), indicating that the visual distractor had no impact on acquisition under the delay conditioning regime. Retention performance recorded 1 h after conditioning also showed no difference between control and distracted bees (Figure S4A; χ2 = 0, df = 1, p = 1). In contrast, the learning performance of control and distracted bees differed significantly under the trace conditioning regime (Figure 3E). While both groups of bees increased their response to the CS across trials (Cochran’s Q test, trace conditioning control: Q = 44.1, df = 3, p < 0.001; trace conditioning distractor: Q = 21.3, df = 3, p < 0.001), the visual distractor presented during the gap between the CS and US significantly impacted learning performance (Figure 3E). Distracted bees exhibited lower acquisition than control bees (GLMER on Trace conditioned bees, Distraction effect, z = −3.45, p < 0.001). A similar result was observed in a retention test performed 1 h after conditioning (Figure S4A; χ2 = 5.98, df = 1, p = 0.014). Overall, these results indicate that the visual distractor selectively interfered with trace conditioning but not with delay conditioning. A combined analysis further supported this conclusion (GLMER analysis on all bees, Distraction x Group effect, z = −3.266, p = 0.001). To evaluate the general validity of this conclusion, we repeated the previous experiments, this time using a mechanosensory distractor. We designed a device in which a feather delivered a 100 ms mechanosensory contact to the bee’s thorax during conditioning (Figure 3C). Then, we compared its effects on delay- and trace-conditioned bees. The odorant used as the conditioned stimulus (CS) did not significantly affect learning performance in either delay conditioning (GLMER on Delay conditioned bees, Odorant used as a CS effect, z = −0.49, p = 0.63) or trace conditioning (GLMER on Trace conditioned bees, Odorant used as a CS effect, z = −1.12, p = 0.26). Therefore, we combined the acquisition performance into a single learning curve for each conditioning type. Delay-conditioned control and mechanically distracted bees (Figure 3F) significantly increased their responses to the CS over the course of conditioning trials (Cochran’s Q test, delay conditioning control: Q = 78, df = 3, p < 0.001; delay conditioning distraction: Q = 81.1, df = 3, p < 0.001) regardless of the mechanosensory stimulation (GLMER on Delay conditioned bees, Distraction effect, z = 0.142, p = 0.89). Similar results were observed in a retention test conducted 1 h after conditioning (Figure S4B, control vs. distracted bees with specific memory: χ2 = 0.06, df = 1, p = 0.80). Thus, as with the visual distractor, the mechanosensory distractor had no detectable effect on either acquisition or retention performance in delay conditioning. A different result was observed in the case of trace conditioning. Although bees significantly increased their responses to the CS across conditioning trials (Cochran’s Q test, trace conditioning control: Q = 40.5, df = 3, p < 0.001; trace conditioning distractor: Q = 21.3, df = 3, p = 0.001), acquisition was impaired in the distracted group (Figure 3G) compared to the control, non-distracted group (GLMER on Trace conditioned bees, Distraction effect, z = −3.83, p < 0.001). Consistently, the retention performance of distracted bees was lower than that of control bees (Figure S4B; percentage of specific learners: χ2 = 7.72, df = 1, p = 0.005). In conclusion, both experiments show the selective interference of distractors, regardless of their modality, with trace but not delay conditioning. This conclusion is further supported by a combined analysis (GLMER on all bees, Distraction × Group effect: z = −3.33, p < 0.001). Since delay conditioning remained unaffected across conditions, we propose that distinct cognitive processes underlie trace and delay conditioning in honeybees, consistent with the idea that awareness of the stimulus contingencies is important for successful trace conditioning but is unnecessary for delay conditioning. Trace conditioning impairment by a distractor is not caused by stimulus competition or negative valence The impairment in learning performance caused by distractors during trace conditioning could be due not to a distortion of perceptual awareness but to a competition between the conditioned odorant and the distractor for associative strength based on their concurrent temporal relationship with the US. To test this hypothesis, we presented delay- and trace-conditioned bees with the distractor used during conditioning—either the 100 ms light flash or the 100 ms feather contact—one hour after the end of training, and quantified their PER scores. As shown in Table 1, there was a significant absence of appetitive responses to distractors after conditioning, confirming their neutral nature in our experiments. Thus, the reduction in CS responses during trace conditioning induced by distractors was attributable to their interference with awareness of stimulus contingencies rather than a shared associative strength with the CS.Table 1Distractors used are not learned in association with a sucrose rewardDistractorConditioning variantGroupBees showing PERVisual experimentWhite LED flashing at 50 Hz for 0.1 sDelayControl0/46Distraction0/46TraceControl0/56Distraction0/61Mechanosensory experimentFeather stroking the back of the bees for 0.1 sDelayControl0/42Distraction2/44TraceControl0/46Distraction0/45Number of bees exhibiting a proboscis extension response (PER) to the distracting stimuli. The dataset includes eight groups: four from the visual distractor experiment and four from the mechanosensory distractor experiment. Within each experiment, bees were trained using either delay or trace conditioning. For each conditioning variant, one group served as a control (no distractor), while the other was exposed to a distractor during the interstimulus interval. Responses to the distractor alone were tested after memory tests for the conditioned stimulus and the novel odorant. Table 1Distractors used are not learned in association with a sucrose rewardDistractorConditioning variantGroupBees showing PERVisual experimentWhite LED flashing at 50 Hz for 0.1 sDelayControl0/46Distraction0/46TraceControl0/56Distraction0/61Mechanosensory experimentFeather stroking the back of the bees for 0.1 sDelayControl0/42Distraction2/44TraceControl0/46Distraction0/45Number of bees exhibiting a proboscis extension response (PER) to the distracting stimuli. The dataset includes eight groups: four from the visual distractor experiment and four from the mechanosensory distractor experiment. Within each experiment, bees were trained using either delay or trace conditioning. For each conditioning variant, one group served as a control (no distractor), while the other was exposed to a distractor during the interstimulus interval. Responses to the distractor alone were tested after memory tests for the conditioned stimulus and the novel odorant. Distractors used are not learned in association with a sucrose reward Number of bees exhibiting a proboscis extension response (PER) to the distracting stimuli. The dataset includes eight groups: four from the visual distractor experiment and four from the mechanosensory distractor experiment. Within each experiment, bees were trained using either delay or trace conditioning. For each conditioning variant, one group served as a control (no distractor), while the other was exposed to a distractor during the interstimulus interval. Responses to the distractor alone were tested after memory tests for the conditioned stimulus and the novel odorant. An alternative hypothesis for the performance impairment observed during trace conditioning is that the distractor might have had an innate aversive valence, which interfered with the bee’s ability to learn the appetitive odor-sucrose association. To test this, we subjected two groups of bees to one trace or one delay conditioning trial in which bees experienced the odorant paired with sucrose, followed by a second trial (Figure 4A). In the second trial, control bees were exposed to the odorant alone (extinction trial), while distracted bees encountered the odorant followed by the distractor—the 100 ms light flash or the 100 ms feather contact. One hour after the second conditioning trial, bees were tested by presenting the CS alone and a novel odorant. If the distractor had an innate aversive valence, it should reduce CS responses in the distracted bees beyond the extinction effect observed in control bees exposed only to the odorant in the second trial.Figure 4Distracting stimuli do not negatively affect olfactory memory retention(A) Design of the extinction experiment used to assess whether the distractor carried neutral valence. The first trial was a standard conditioning trial (either trace or delay). The second trial was an extinction trial in which either the CS was presented alone (“Nothing”) or concurrently with a visual or mechanosensory distractor (“LED/Feather”). A memory test was conducted 1 h later.(B) Percentage of CS-specific responders for each of the six groups. Three groups were conditioned using delay conditioning (blue) and three using trace conditioning (orange). For each conditioning variant, one group underwent extinction without a distractor (control), and two groups were exposed to either a visual or mechanosensory distractor following the CS presentation. Memory was tested 1 h post-conditioning. N = 23–29 for each group. Error bars indicate 95% confidence intervals. Distracting stimuli do not negatively affect olfactory memory retention (A) Design of the extinction experiment used to assess whether the distractor carried neutral valence. The first trial was a standard conditioning trial (either trace or delay). The second trial was an extinction trial in which either the CS was presented alone (“Nothing”) or concurrently with a visual or mechanosensory distractor (“LED/Feather”). A memory test was conducted 1 h later. (B) Percentage of CS-specific responders for each of the six groups. Three groups were conditioned using delay conditioning (blue) and three using trace conditioning (orange). For each conditioning variant, one group underwent extinction without a distractor (control), and two groups were exposed to either a visual or mechanosensory distractor following the CS presentation. Memory was tested 1 h post-conditioning. N = 23–29 for each group. Error bars indicate 95% confidence intervals. In the test, the percentage of bees exhibiting specific memory—defined as those responding to the CS but not to the NOd—did not differ between the control group and either of the two distracted groups (feather group or flash-light group) (Figure 4B). Response rates were similar between control and distracted delay-conditioned bees (χ2 = 1.65, df = 2, p = 0.44) as well as between control and distracted trace-conditioned bees (χ2 = 0.01, df = 2, p = 0.99). These results rule out the possibility that the distractors had an innate aversive valence, which could have selectively impaired performance during trace conditioning. Instead, they support the interpretation that distractors interfered with the awareness of stimulus contingencies during trace, but not delay, conditioning. A neural network model reveals the importance of a timing element in keeping the trace After identifying the reported differences between delay and trace conditioning, we adopted a modeling approach to investigate whether different neural mechanisms underlie these two processes, as previously suggested by experimental studies in Drosophila.29 We hypothesized that the mushroom bodies (MBs), a multimodal brain structure involved in learning and memory,30 play a crucial role in olfactory trace conditioning. To test this hypothesis, we implemented a spiking neural network model of the MBs to identify the computational mechanisms required for trace conditioning tasks, and compared its performance to predictions under a delay conditioning regime (Figure 5A).Figure 5A spiking neural model accounts for learning performances in delay and trace conditioning(A) Schematic of the delay and trace conditioning protocols, showing the conditioned stimulus (CS; odorant, blue), unconditioned stimulus (US; sucrose, green), and distractor (DIS; red).(B) Graphical representation of the network architecture. PN: projection neurons (odorant input); LPN: lobula projection neurons (visual input); GAN: giant ascending neuron (mechanosensory input); KC: Kenyon cells; DPM: dorsal paired medial neurons; 5-HT: serotonin; MBON: mushroom body output neuron; OA: octopamine; US: unconditioned stimulus; PER: proboscis extension response.(C) Averaged acquisition curves from simulated data (% of PER to the CS) over 20 initializations for delay-conditioned (blue) and trace-conditioned (orange) neural networks. Solid lines represent control conditions; dotted lines indicate distracted conditions. Shaded areas around the curves represent the SD.(D) Average response levels of the model to the distractor stimulus.(E) Averaged acquisition curves of simulated data (% of PER to the CS) over 20 initializations for delay-conditioned (blue) and trace-conditioned (orange) neural networks with the DPM module excluded from the model. Shaded areas show the SD. A spiking neural model accounts for learning performances in delay and trace conditioning (A) Schematic of the delay and trace conditioning protocols, showing the conditioned stimulus (CS; odorant, blue), unconditioned stimulus (US; sucrose, green), and distractor (DIS; red). (B) Graphical representation of the network architecture. PN: projection neurons (odorant input); LPN: lobula projection neurons (visual input); GAN: giant ascending neuron (mechanosensory input); KC: Kenyon cells; DPM: dorsal paired medial neurons; 5-HT: serotonin; MBON: mushroom body output neuron; OA: octopamine; US: unconditioned stimulus; PER: proboscis extension response. (C) Averaged acquisition curves from simulated data (% of PER to the CS) over 20 initializations for delay-conditioned (blue) and trace-conditioned (orange) neural networks. Solid lines represent control conditions; dotted lines indicate distracted conditions. Shaded areas around the curves represent the SD. (D) Average response levels of the model to the distractor stimulus. (E) Averaged acquisition curves of simulated data (% of PER to the CS) over 20 initializations for delay-conditioned (blue) and trace-conditioned (orange) neural networks with the DPM module excluded from the model. Shaded areas show the SD. As shown in Figure 5B, the model is implemented based on the essential components of the honeybee brain required for associative olfactory learning. In blue, the olfactory input (CS) is relayed from the antennal lobe – the primary olfactory center – to the MB Kenyon cells (KCs) via projection neurons (PNs). In red, the distractor input is also conveyed to the MB KCs via other neural pathways, which vary depending on the distractor’s sensory modality. Importantly, associative learning occurs through spike-time-dependent-plasticity (STDP),31 enabling KCs to associate the olfactory CS with the appropriate motor response -specifically, the proboscis extension reflex to obtain the sucrose reward. This response is mediated by mushroom bodies output neurons (MBONs, green in Figure 5B; see Figure S5A for a rasterplot obtained for a representative trial in all experimental conditions). The model incorporates two neuromodulatory mechanisms that influence learning: an octopaminergic system (OA), which functions as an instructive reward signal32 and is essential for the STPD learning dynamic (see STAR Methods for details), and a simplified serotonergic system (5-HT), mediated by the dorsal paired medial neurons (DPMs). The latter is particularly relevant for trace conditioning, as it extends the coincidence detection window, thereby bridging the temporal gap between the CS and US, consistent with recent findings in Drosophila33. In our model, the serotonergic system acts as an attentional mechanism, filtering out irrelevant stimuli by accelerating the decay of their eligibility traces during the temporal gap between the CS and US. The eligibility trace serves as a temporary memory of an event, gradually decaying over time (i.e., through a given time constant τ; see STAR Methods for details). Specifically, the serotonergic system33 maintains the time constant τ of the CS during the CS-US interval, while reducing the time constant τ of the Distractor, thereby weakening the impact of irrelevant stimuli on learning (see STAR Methods for details). As mentioned above, honeybees learn the CS-US association in the delay-conditioning protocol, regardless of the distractor’s presence. However, in trace conditioning, learning was less efficient, and this effect was even more pronounced in the presence of a distractor (Figures 3D and 3F). Our simulations replicate these empirical observations (Figure 5C). Specifically, we simulated 12 trials of 1000-time steps (equivalent to 1000 ms) for each condition (delay/trace, with/without distractors) and repeated the training 20 times with different initializations (see Figure S5B to visualize the synaptic changes over 12 trials in one of the initializations). In each trial, the CS was presented after 100 ms and lasted for 500 ms. In delay conditioning, the US was delivered during the final 200 ms of the CS, whereas in trace conditioning, the US was presented 200 ms after the CS ended. If the distractor was included, it appeared randomly before the US, and in trace conditioning, it always occurred after the end of the CS. We then tested MBON activation by presenting only the CS or the distractor after each trial. The percentage of proboscis extension responses (%PER) shown in Figures 5C–5E is proportional to the number of spikes generated by the MBON (Mushroom Body Output Neuron) in each trial following training. As shown in Figure 5C, our results show that the model successfully learns the CS-US association in delay conditioning, but to a lesser extent in trace conditioning, mirroring our experimental data. In addition, as in the real data, the effect of the distractor was significant for trace, but not for delay, conditioning. Importantly, the model did not respond to the distractor stimuli (Figure 5D), demonstrating that the attentional mechanisms mediated by the DPM neurons effectively prevented the association of irrelevant stimuli. To further examine the role of this “awareness system,” we simulated a “lesion study” by suppressing the DPM module from the model, thereby eliminating its modulation of the time constant of the eligibility trace. Under these conditions, our simulations (Figure 5E) revealed a reduction of the response rate exclusively in trace conditioning. We thus aimed to test this prediction experimentally by pharmacologically blocking serotonin signaling. Testing the model: Serotonergic blockade impairs trace but not delay conditioning Our model, as well as empirical studies,33 suggests that serotonergic DPM neurons play a central role in successful trace conditioning. Genetic methods used to block 5-HT neurons are not feasible in a non-transgenic organism such as the honeybee. Therefore, we specifically aimed at inhibiting serotonergic signaling at the level of the mushroom bodies by injecting 5-HT antagonists via the ocellar nerve —a thick fiber bundle approximately 40 μm in diameter that extends medially and caudally from the dorsal margin of the head capsule to a depth of 300 μm into the protocerebrum.34 After carefully removing the lens of the median ocellus, we injected either phosphate-buffered saline (PBS) as a control, or a mixture of the two serotonin-receptor antagonists methiothepin mesylate and ketanserin tartrate dissolved in PBS,35,36 at concentrations of 1 mM or 10 mM. Following a 30-min rest period, bees were conditioned under either a trace or delay conditioning regime, and their memory retention was tested after 1 h. According to our model, serotonergic blockade should impair trace but not delay conditioning. Consistent with our previous experiments, learning performances did not differ based on the conditioned odorant (1-nonanol or 2-hexanol), as shown by the GLMER analysis (delay-conditioned bees: Odorant as CS effect, z = −0.94, p = 0.35; trace-conditioned bees: Odorant as CS effect, z = −1.31, p = 0.19). Consequently, performance data were pooled into a single curve for each conditioning type. The learning performance of three groups of bees (i.e., control bees, and bees injected with 1 mM or 10 mM of the serotonin antagonist cocktail) trained under a delay-conditioning regime is shown in Figure 6A. Post hoc comparisons revealed that learning performance did not differ significantly between PBS-injected bees and those injected with the serotonin antagonist cocktail, regardless of the concentration used in this case (GLMER on delay conditioned bees, post hoc comparisons for Treatment effect, PBS vs. Cocktail 1 mM, z = −0.07, p = 0.99, PBS vs. Cocktail 10 mM, z = -0.06, p = 0.99, Tukey-adjusted). Memory retention 1 h after conditioning was significant in all three groups, as each group showed a stronger response to the conditioned stimulus than to the novel odorant (Figure S6A; McNemar tests for each group; PBS: χ2 = 20.02, p < 0.001, Cocktail 1 mM: χ2 = 14.02, p < 0.001, Cocktail 10 mM: χ2 = 13.59, p < 0.001). Retention levels were similar across the three groups, as the percentage of bees exhibiting specific memory did not differ significantly (Figure S6C; χ2 = 0.53, df = 2, p = 0.77). Overall, these results demonstrate that serotonergic blockade did not affect olfactory delay conditioning nor the memory resulting from this form of conditioning.Figure 6Serotoninergic blockade specifically impairs learning performance in trace, but not in delay, conditioning(A) Acquisition curves showing the percentage of bees exhibiting a proboscis extension response (% of PER) to the conditioned stimulus across training trials in delay conditioning.(B) Same as in (A) for trace-conditioned bees. Control groups are shown in blue (Delay PBS) and orange (Trace PBS); pharmacologically treated groups received a serotonergic antagonist cocktail of methiothepin mesylate and ketanserin tartrate at 1 mM (pink) or 10 mM (purple). Shaded areas around the curves represent 95% confidence intervals. ∗∗, p < 0.01, GLMER, Treatment effect, pairwise comparisons. Serotoninergic blockade specifically impairs learning performance in trace, but not in delay, conditioning (A) Acquisition curves showing the percentage of bees exhibiting a proboscis extension response (% of PER) to the conditioned stimulus across training trials in delay conditioning. (B) Same as in (A) for trace-conditioned bees. Control groups are shown in blue (Delay PBS) and orange (Trace PBS); pharmacologically treated groups received a serotonergic antagonist cocktail of methiothepin mesylate and ketanserin tartrate at 1 mM (pink) or 10 mM (purple). Shaded areas around the curves represent 95% confidence intervals. ∗∗, p < 0.01, GLMER, Treatment effect, pairwise comparisons. The learning performance of the three groups of bees trained under a trace-conditioning regime is shown in Figure 6B. Post hoc comparisons revealed that bees subjected to serotonergic blockade at the higher concentration exhibited significantly lower learning compared to PBS-injected bees (GLMER on trace conditioned bees, post hoc comparisons for Treatment effect, PBS vs. Cocktail 10 mM, z = 3.02, p = 0.007). Bees injected with the lower concentration of the serotonin antagonist cocktail showed an intermediate performance, trending lower than PBS-injected bees but not significantly different from those receiving the higher concentration (GLMER on trace conditioned bees, post hoc comparisons for Treatment effect, PBS vs. Cocktail 1 mM, z = −2.22, p = 0.068, Cocktail 1 mM vs. Cocktail 10 mM, z = −0.89, p = 0.65, Tukey-adjusted). Memory retention, assessed 1 h after conditioning, followed a similar pattern. Only the PBS-injected bees displayed a higher response to the conditioned stimulus compared to the novel odorant, consistent with successful memory retention (Figure S6B; McNemar test: χ2 = 8.26, p = 0.004). In contrast, bees injected with either 1 mM or 10 mM of the antagonist cocktail did not show a higher response to the CS than to the NOd (McNemar test for Cocktail 1 mM: χ2 = 2.12, p = 0.15, McNemar test for Cocktail 10 mM: χ2 = 0.84, p = 0.36), indicating a failure to form a specific associative memory. However, specific memory was not different between control and treated bees (Figure S6C; χ2 = 3.99, p = 0.136), probably because memory levels were already low in the PBS-injected control bees. Thus, in line with the predictions of the model described above, the blockade of serotonergic signaling specifically impaired trace conditioning while leaving delay conditioning unaffected. Discussion Our study shows that honeybees can learn and memorize odor-sucrose associations both under a delay and a trace conditioning regime. Moreover, we have shown that the presence of distracting stimuli selectively impairs trace conditioning while leaving delay conditioning unaffected. Finally, both modeling and neuropharmacology highlight a central role of serotonergic signaling for trace conditioning. These findings suggest that, similar to humans, honeybees may rely on an awareness of the CS-US relationship during trace conditioning, but not during delay conditioning, thus extending the presence of this cognitive state to an invertebrate species. Distractors impair trace, but not delay conditioning Our results showed that despite the fact that bees learned the odor-sucrose association under both conditioning regimes, learning performance and memory retention were lower in trace conditioning than in delay conditioning. These results agree with previous findings on olfactory trace conditioning in bees25 and confirm that these two conditioning protocols differ in cognitive complexity. To successfully associate the CS and US in trace conditioning, the animal must retain a representation of the CS across the temporal gap separating it from the US. This process likely engages working memory, which allows for the active maintenance and manipulation of information over relatively long periods.37 The involvement of working memory in trace conditioning suggests a role for attentional processes, which can be categorized as either bottom-up (stimulus-driven, based on salience) or top-down (volitional, dependent on past experiences and goal-directed processing;38,39). Given that trace conditioning requires bridging a temporal gap, it is more likely mediated by top-down attentional control, whereby the animal actively maintains the CS representation while suppressing irrelevant or competing stimuli. This interpretation aligns with theories suggesting that cognitive abilities such as trace conditioning, rapid reversal learning, and cross-modal learning may be underpinned by forms of consciousness.40,41 Indeed, it has been proposed that failure to achieve trace conditioning in the presence of distractors – despite successful performance under undistracted conditions – could serve as one among several potential tests for animal consciousness.18 Our results demonstrate that introducing a distractor during the CS-US interval significantly impairs learning in trace-conditioned bees, but not in delay-conditioned bees. This result is consistent with findings reported for humans subjected to eyeblink trace and delay conditioning,4,5,6 which led to the conclusion that awareness of the stimulus contingencies is required for successful trace conditioning but not for delay conditioning. In that sense, our study did not seek to re-establish the known differences between trace and delay conditioning in honeybees, nor to simply reproduce findings on serotonergic modulation. Instead, our central objective was to examine how external distractors differentially affect trace and delay conditioning—a question that has never been addressed in honeybees. The ultimate goal was to determine if these two forms of learning differ in their requirement for awareness. This distinction is not peripheral, but rather central to our contribution. The distractors used in our experiments were neutral stimuli that did not compete with the CS for associative strength. Similar to experiments in rodents42 and in Drosophila,43 we used a neutral stimulus that selectively disrupted trace - but not delay - conditioning, remained non-associative, and exerted interference across modalities. Additionally, our distractors differed in their sensory modality from the olfactory CS to prevent the enhancement of associative learning. In fact, prior experiments25 showed that the presentation of an additional odorant during the gap increased trace-conditioning performances. Successful trace conditioning relies on the ability to maintain an internal representation of the CS until the US is presented. The fact that a distractor disrupts this process suggests that the bees must actively maintain a CS representation in time, a cognitive function potentially requiring attentional control or awareness.44 This idea is supported by two findings: first, learning is constrained by time as bees fail to associate the CS and US when the gap exceeds 6 s,25 suggesting a limit to how long they can sustain the CS representation. Second, prior studies indicate that bees can gradually learn to associate a CS and a US over increasingly long inter-stimulus intervals,25 a cognitive feature that resembles attentional learning mechanisms and is inconsistent with rigid, automatic mechanisms. This interpretation aligns with findings in Drosophila, where distractors have similarly been shown to impair trace visual learning,43 reinforcing the idea that attentional processes, and more specifically awareness of stimulus contingencies, play a key role in trace conditioning across species.45 Evidence for awareness without precise timing Our results suggest that bees exhibit awareness of stimulus contingencies during trace conditioning but do not encode precise timing between the CS and US. In contrast, studies in humans4,5,6 and mammals46 have shown that during trace conditioning, subjects exhibit the temporal tuning of their responses, with peak activity gradually shifting toward the expected US arrival time. In our study, the absence of such timing behavior may be attributed to the limited number of training trials, which could have constrained temporal encoding, or to a potential trade-off between learning speed and temporal specificity. Notably, a lack of precisely timed conditioned responses has been consistently reported in all insect trace conditioning studies to date.47,48 However, bees are capable of learning temporal relationships between events, as shown by experiments in bumble bees trained under a fixed-interval procedure.49 In this paradigm, bees received a sucrose reward for the first response (extending the proboscis through a hole in a container) that occurred after a fixed time interval had elapsed. Bumble bees trained with an initial reward and a variable gap (6 s, 12 s, or 36 s) adjusted their proboscis extension responses to approximately one-third of the interval, indicating an ability to encode temporal information under certain training conditions. The neural underpinnings of awareness of stimulus contingencies under trace conditioning in honeybees Since distractors were from a different sensory modality than the odorants used for conditioning, their effects on learning performance likely emerge at higher-order processing centers—where different sensory modalities converge—such as the mushroom bodies (MBs) or the central complex. In honeybees, the MBs are well known for their multimodal sensory integration capacities and are strong candidates for maintaining conditioned stimulus representations over time. Persistent odor representations—or “aftersmells”—have been observed at the level of projection neurons conveying olfactory information from the antennal lobe to the MBs.25,50,51 Modeling these responses at the level of Kenyon-cell (KC) connectivity supported the idea that the MBs may sustain neural activity across the CS-US gap.47 Theoretical models44 propose that such persistent activity could enable trace learning. This hypothesis is further supported by evidence from Drosophila, where MB-dependent trace conditioning has been demonstrated using olfactory stimuli.50,51 Conversely, Grover et al.43 recently showed that MBs are not required for visual trace conditioning in Drosophila, where the central complex appears to assume this role. These differences may reflect modality-specific circuit engagement: in flies, MBs receive sparse visual input,52 suggesting a reduced involvement in visual trace learning despite participation in simpler forms of visual associative learning.53,54 In contrast, the MBs of honeybees are deeply involved in both olfactory and visual learning and memory,19,26,30 making them strong candidates for supporting trace learning. Importantly, the Drosophila study did not include analyses of response timing43 and emphasized a role for dopaminergic signaling in trace conditioning. Our study highlighted the importance of serotonergic signaling for successful trace conditioning. In Drosophila, the dorsal paired medial (DPM) neuron—a serotonergic neuron—extends the time window for CS–US association,33 acting as a timing modulator. This finding was incorporated into our model and experimentally validated through pharmacological blockade of serotonergic signaling, which selectively impaired trace, but not delay, conditioning. Although a direct homologue of the Drosophila DPM neuron has yet to be identified in bees, our results suggest a conserved role for serotonergic signaling in attentional processes or in the maintenance of CS representations across the trace interval. While the previous study on Drosophila DPM serotonergic neurons showed that they can extend the time window for CS-US associations during trace conditioning,33 it did not investigate the impact of distractors on that process. In contrast, our work explores how serotonergic modulation relates specifically to the differential vulnerability of trace versus delay conditioning to distraction. This nuance is central to both our experimental results and the conceptual framework of our model, where serotonergic modulation is used not to explain trace conditioning itself and the length of the temporal gap, but rather the selective disruption of trace conditioning by distractors. Overall, our results indicate that trace learning may depend on species- and modality-specific neural circuitry, while also relying on shared neuromodulatory mechanisms such as serotonin-mediated temporal integration. These findings further support the role of the MBs as a central site for integrating and sustaining CS representations during the temporal gap in trace conditioning.44 Modeling olfactory learning performance with a minimal yet informative framework Other models have been developed to capture various aspects of insect—primarily Drosophila—olfactory learning. For instance, Bennett et al.55 proposed that dopaminergic neurons signal prediction errors rather than absolute reward by integrating feedback from mushroom body output neurons (MBONs), thereby modulating plasticity at Kenyon cell → MBON synapses. This approach contrasts with our model, which relies on a more classical associative mechanism without an explicit prediction-error component. However, the two frameworks are not mutually exclusive, and future work could explore hybrid models where even simple associative tasks are solved through error-driven learning. Springer and Nawrot56 introduced a model featuring parallel memory traces (appetitive and aversive), an idea conceptually related to our assumption of distinct traces for cues and distractors, reflected in the use of different time constants (τ). Their model also includes feedback from MBONs to dopaminergic neurons to generate reward predictions and adjust synaptic strength, adding a layer of dynamical plasticity not explicitly represented in our framework. Jürgensen et al.57 focused on second-order conditioning in Drosophila larvae, investigating how different circuit motifs—varying in feedback and feedforward connectivity—support higher-order associative learning. While their model incorporates dopaminergic signaling similarly to our use of octopaminergic neurons to signal reinforcement, their emphasis on second-order conditioning and task-specific performance optimization limits the generalizability of their findings to our paradigm of absolute conditioning. In summary, while these models incorporate greater biological detail—especially through feedback loops and reward prediction dynamics—our model maintains a simpler structure, which facilitates analysis and interpretation. It provides a minimal yet sufficient implementation for the associative learning task at hand, without the need to explicitly model feedback connections. Limitations of the study A key assumption of our study and modeling framework is the existence of serotonergic neurons in honeybees functionally analogous to the Dorsal Paired Medial (DPM) neurons described in Drosophila, which are known to extend the temporal window for CS-US associations during trace conditioning.33 While such neurons have not yet been identified in bees, our model posits that serotonergic signaling modulates stimulus processing by enhancing task-relevant representations and suppressing irrelevant or contextual ones. This functional asymmetry is implemented through biologically grounded mechanisms, including differential eligibility trace dynamics and spike-timing-dependent plasticity (STDP). Although direct experimental evidence for these neurons and their modulatory role in bees is currently lacking, our model remains neurobiologically plausible and offers specific, testable predictions. Future work can use these predictions to guide investigations into the neuroanatomy and neuromodulatory processes underlying trace learning in bees. Resource availability Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Prof. Dr. Martin Giurfa (martin.giurfa@sorbonne-universite.fr). Materials availability The materials used were purchased from standard chemical suppliers listed in the key resources table. No custom materials were created or used. Data and code availability •The datasets of this study are available in the Figshare repository: https://figshare.com/s/ee65ad8bc64e7920156f.•The code of the model can be found at: https://github.com/artipago/honeybee-spiking-network-model.•Any additional information required to reanalyze the data reported in this article is available from the lead contact upon request. •The datasets of this study are available in the Figshare repository: https://figshare.com/s/ee65ad8bc64e7920156f. The datasets of this study are available in the Figshare repository: https://figshare.com/s/ee65ad8bc64e7920156f. •The code of the model can be found at: https://github.com/artipago/honeybee-spiking-network-model. The code of the model can be found at: https://github.com/artipago/honeybee-spiking-network-model. •Any additional information required to reanalyze the data reported in this article is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this article is available from the lead contact upon request. Acknowledgments We thank two anonymous reviewers for their valuable and encouraging comments. We also thank Benjamin Paffhausen for designing, 3D-printing, and assisting in the implementation of the isolation chamber and the feather presentation using a fine motor. We also thank Nathan Gaubil for his implication in drafting the model’s code and Olivier Fernandez and Alexandre Catonné for managing the hives in the experimental apiaries of Toulouse and Paris, respectively. This work was supported by the 10.13039/501100000781European Research Council (ERC Advanced Grant COGNIBRAINS to M. Giurfa), the Toulouse Mind and Brain Institute (TMBI grant to A. Alamia and M. Giurfa), and by funding from the 10.13039/501100019125Sorbonne University, the French Research Council (CNRS), and the French National Institute of Health and Medical Research (INSERM). Author contributions C. M., M. P., G.I., and M.G. conceived and designed the study. C.M. conducted the experiments. M.Gh. helped with the experiments. A.A. designed and implemented the model. C.M. analyzed the data. All experiments were supervised by M.G. and G.I. Funding was obtained by M.G. C.M. and M.G. wrote the article. All authors reviewed and approved the final article. Declaration of interests The authors declare no competing interests. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work, the authors did not use AI assisted technologies. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERChemicals, peptides, and recombinant proteins2-hexanolSigma AldrichCas No. 626-93-71-nonanolSigma AldrichCas No. 143-08-8Methiothepin mesylateSigma AldrichCAS No. 74611-28-2Ketanserin tartrateSigma AldrichCAS No. 83846-83-7Deposited dataThe datasets generated and/or analyzed during the current study are available in FigshareThe authors of this articlehttps://figshare.com/s/ee65ad8bc64e7920156fThe code used in the modeling section is available in GithubThe authors of this articlehttps://github.com/artipago/honeybee-spiking-network-modelExperimental models: Organisms/strainsHoneybee workers Apis melliferaApiary of the University Paul Sabatier & apiary of Sorbonne university.N.A.Software and algorithmsR 4.3.2 Statistical softwareR Project for Statistical Computinghttps://cran.r-project.org/bin/windows/base/Graphpad Prism 8.0.2GraphPad Softwarewww.graphpad.comImageJ 1.54fImageJ Softwarehttps://imagej.net/ij/index.htmlArduino 2.3.6Arduino IDEhttps://www.arduino.cc/en/software/MATLAB 2024bThe MatWorks, Inc.https://www.mathworks.com/products/matlab.htmlOtherSaccharoseSigma-AldrichCas No. 57-50-1Phosphate buffered salineSigma AldrichMDL MFCD00131855 REAGENT or RESOURCESOURCEIDENTIFIERChemicals, peptides, and recombinant proteins2-hexanolSigma AldrichCas No. 626-93-71-nonanolSigma AldrichCas No. 143-08-8Methiothepin mesylateSigma AldrichCAS No. 74611-28-2Ketanserin tartrateSigma AldrichCAS No. 83846-83-7Deposited dataThe datasets generated and/or analyzed during the current study are available in FigshareThe authors of this articlehttps://figshare.com/s/ee65ad8bc64e7920156fThe code used in the modeling section is available in GithubThe authors of this articlehttps://github.com/artipago/honeybee-spiking-network-modelExperimental models: Organisms/strainsHoneybee workers Apis melliferaApiary of the University Paul Sabatier & apiary of Sorbonne university.N.A.Software and algorithmsR 4.3.2 Statistical softwareR Project for Statistical Computinghttps://cran.r-project.org/bin/windows/base/Graphpad Prism 8.0.2GraphPad Softwarewww.graphpad.comImageJ 1.54fImageJ Softwarehttps://imagej.net/ij/index.htmlArduino 2.3.6Arduino IDEhttps://www.arduino.cc/en/software/MATLAB 2024bThe MatWorks, Inc.https://www.mathworks.com/products/matlab.htmlOtherSaccharoseSigma-AldrichCas No. 57-50-1Phosphate buffered salineSigma AldrichMDL MFCD00131855 Experimental model and study participant details Insects Female worker honeybees (Apis mellifera) were collected from hives located in our university campus apiary. No institutional permission or ethical guidelines are required for research work on insects. Bees were captured in small glass vials and transported to the laboratory, located a few hundred meters away. Upon arrival, bees were briefly cooled on ice for 5 min and individually harnessed in plastic tubes (inner diameter ∼0.9 cm; wall thickness: 1–2 mm). The bees’ heads protruded from the tubes, allowing free movement of the antennae and mouthparts. Approximately 30 min after harnessing, the bees were fed 5 μL of 30% (w/w) sucrose solution to prevent excessive starvation and to standardize satiety level across individuals, and kept for 3 h in a dark, humid container maintained at 28°C. One hour before conditioning, proboscis extension reflex (PER) was assessed by touching the antennae with a toothpick dipped in 30% sucrose solution. Only bees showing a full PER were included in the experiments. Method details Olfactory stimulation Olfactory stimulation was controlled using a computer-controlled olfactometer that precisely timed the delivery of odorant pulses.58 To eliminate residual odorants and maintain the temporal integrity of the trace conditioning protocol, an air extractor was placed behind the bees to remove lingering volatiles. The absence of residual odorants during the interstimulus gap was verified (Figure S1) using a photoionization detector (miniPID, model 200a, Aurora Scientific Inc., Canada). Additionally, the airflow strength was carefully controlled to avoid unintended mechanical stimulation. Absolute conditioning All experiments employed absolute conditioning, in which a single conditioned stimulus (CS; odorant) predicted the delivery of an unconditioned stimulus (US; sucrose reward). Bees were trained with one of two odorants (CS: either 1-nonalol or 2-hexanol; Sigma-Aldrich, Saint-Quentin-Fallavier, France), while the other odorant was used as a novel odorant (NOd) during the memory retention tests to assess memory specificity. We followed a standardized conditioning protocol24 using 4 or 5 CS-US pairings with an inter-trial interval (ITI) of approximately 10 min. Each trial began with 15-s habituation period in the experimental setup to allow bees to acclimate to the airflow and context. Both delay and trace conditioning used the same inter-stimulus interval (ISI; interval between the CS and US onsets) of 5 s and the same US duration of 3 s, but differed in CS duration (Figure 1A). In delay conditioning, the CS lasted 6 s and overlapped with the US. In trace conditioning, the CS lasted 0.5 s and terminated 4.5 s before US onset, maintaining a constant ISI of 5 s across paradigms. In addition, we performed control experiments using longer CS durations in trace conditioning, which did not affect our conclusions (Figure S2). After each CS-US pairing, the bee remained in the experimental setup for 15 additional seconds, to minimize contextual learning. Each trial lasted 44 s. The presence (+) or absence (−) of the proboscis extension response (PER) was recorded during the 5-s interval between odorant onset and the onset of sucrose delivery (Figures 1A and 1B). Bees that responded spontaneously to the odorant in the first trial were excluded from the analysis24 since the CS must be neutral at baseline.7 Memory retention was tested either 1 h or 24 h after the last conditioning trial. Retention was assessed by sequentially presenting the CS and NOd in randomized order, with a 10-min inter-test interval. PER was recorded for each odorant presentation. Finally, the unconditioned response (PER to sucrose) was tested by applying sucrose directly to the antennae; bees that failed to exhibit PER were excluded from the analysis. To minimize extinction or fatigue effects, different groups of bees were conditioned and tested in parallel for each retention interval. Video analysis for PER latency and licking events Video-recordings were performed during memory tests in bees that underwent delay and trace conditioning procedures. In this way, we examined whether these conditioning regimes elicited qualitative differences in PER. Recordings were made using a video camera (SONY Handycam HDR-CX625, 50fps) and analyzed manually on a frame-by-frame basis with ImageJ. To determine the precise latency of PER onset, an Arduino-controlled LED was programmed to turn on at odor presentation and used as time reference. PER latency was defined as the time (in s) from LED onset to the first timepoint at which the proboscis reached its maximum extension (Figure 2C). Latency was calculated based on frame number and thus expressed as decimal values (e.g., 1.58 s). Additionally, we quantified the number of licking events, defined as discrete extensions of the glossa beyond the proboscis sheath. While the proboscis refers to the entire feeding organ used for lapping and sucking nectar, the glossa is its central, tongue-like component. Thus, a bee may extend its proboscis without necessarily engaging in licking behavior—i.e., without rhythmic glossa movements beyond the proboscis sheath. The video analysis was conducted by a single observer. To assess the reliability of this scoring and ensure that no observer bias affected the results, we asked a co-author to independently re-analyze a random subset of 40 videos (≈35% of the 113 total videos). The second observer was blind to the individual-level data extracted by the first observer, though both observers could not be blind to the conditioning group (delay vs. trace), since LED indicators on the Arduino marked CS and US onsets/offsets in the videos. These LED cues, necessary to define time zero (odor onset), also conveyed the temporal dynamics of the conditioning protocol, making full blinding to group assignment impossible. To quantify inter-observer reliability, we compared the measurements for latency and number of licking events. We used the following approaches: 1) Intraclass Correlation Coefficient (ICC): We used a two-way agreement model. The ICC was 98.3% (95% CI: 96.8–99.1) for licking events and 95.3% (95%CI: 90.9–97.5) for latency, indicating excellent agreement between observers. 2) Bland–Altman analysis: The mean difference between observers was close to zero (−0.09s for latency, −0.25 events for licking), and most points fell within ±1.96 SD, suggesting no systematic bias or drift between observers. These complementary analyses confirm a high degree of inter-observer reliability and support the robustness of the data collected. Distractor delivery To investigate the effect of distractive stimuli on learning under delay and trace conditioning protocols, we placed each bee inside a custom 3D-printed isolation chamber to minimize contextual cues and enhance distractor salience (Figure 3A). Two types of distractors were used: visual stimulation (light flashes from a white LED at 50 Hz) and mechanosensory stimulation (a feather that gently stroked the back of the bees) (Figure 3B). Distractors were delivered during the interstimulus interval (ISI) in both conditioning paradigms and lasted 100 ms. To prevent bees from predicting the distractor onset, its presentation was randomly timed within a 3-s window centered around the midpoint of the ISI. Distractors were applied in every conditioning trial. To determine whether the distractor stimuli were encoded as competing cues, we tested each bee’s behavioral response to the distractors following the memory retention tests. This assessment allowed us to evaluate whether distractors had acquired associative strength or remained neutral. Model The model architecture, shown in Figures 5A and 5B, was inspired by previous studies on insect olfactory learning.59 Importantly, it is not intended to replicate all anatomical connections in detail, but rather to provide a minimal yet informative framework capable of performing the task, as in.59 The network consists of two layers: the first layer comprises 78 projection neurons (PNs for odorant inputs, LPNs and GANs for visual and mechanosensory distractor input, respectively), which relay signals to the next layer, composed of 560 Kenyon cells (KCs). These KCs constitute the mushroom bodies and project onto a single motor output neuron (mushroom body output neuron, MBON), which governs the proboscis extension reflex, the behavioral response to the sucrose reward. Each neuron is modeled using the biologically plausible spiking dynamics proposed by Izhikevich et al.60 Specifically, the membrane potential v evolves according to the following equation:Cv˙=k(v−vr)(v−vt)−u+IWhere u represents the recovery currentu˙=α(β(v−vr)−u) If the membrane potential reaches the threshold vt, both the current and the membrane potential are set to v = γ and u = u+δ. Importantly, α, β, γ, δ and k are model parameters determining the neurons’ dynamics and firing rate. We set the neurons’ parameters to model the firing rate of honeybee KCs.61 Specifically, we set α = 0.01,β = −0.3,γ = −65,δ = 8,k = 0.015,C = 4,vt = −25 and vr = −85. Learning occurs via STDP mechanisms, as described by Izhikevich,31 mimicking the role of octopaminergic neuromodulation. In honeybees, octopamine serves as the functional equivalent of dopamine in flies, playing a crucial role in reward-based learning.32 In short, the synaptic strength s between pre- and post-synaptic neurons is modulated by the octopaminergic concentration d and an eligibility trace, which captures the temporal dynamics of synaptic modifications:s˙=cdWhere the eligibility trace c is defined asc˙=−cτc+STDP(τ)δ(t−tpre/post) Here, τ = tpost – tpre, where τ is the interspike interval. This variable, which reflects the relative timing between pre- and postsynaptic spikes, decays exponentially to zero with a time constant τc = 100 ms. As in previous work,31 the STDP function was defined as:STDP(tpost−tpre)={Ae(tpost−tpre)τSTDPiftpost−tpre<0Ae−(tpost−tpre)τSTDPiftpost−tpre≥0Where τSTDP was set to 20 ms The octopaminergic concentration is defined asd˙=−dτd+OA(t)where the octopaminergic time constant τd was set to 200 ms for all neurons, and OA(t) represents the pool of octopamine at any given moment. In particular, OA(t) is distributed to the KC neurons proportionally to the temporal delay between each spike and the delivery of the US (i.e., it was set to 0 in absence of the US) Intuitively, the eligibility trace functions as a memory signal that gradually decays over time, with a time-constant τc set to 100 ms for the neurons processing the cue signal. Additionally, we included the effect of serotonergic DPM neurons, which modulate the eligibility trace decay (i.e., τc) of KCs, acting as an attentional mechanism. Specifically, DPM modulation reduces the trace time-constant τc from 100 ms to 10 ms in neurons processing distractor stimuli, effectively limiting their influence on learning. For each condition, the model was trained twenty times over 12 trials, each trial consisting of 1000 time steps of 1 ms. In each trial, the conditioned stimulus (CS) was presented at t = 100 ms and lasted for 500 ms. In delay conditioning, the US was delivered during the final 200 ms of the CS presentation. In contrast, during trace conditioning, the US was delivered 200 ms after the CS had ended. Random stimuli were generated independently for both the CS and the distractors on each trial. When present, a distractor was introduced at a random time point preceding the US. In the trace conditioning condition, the distractor always occurred after the CS offset. For each condition, we trained the twenty models with different initializations for 12 trials. Pharmacological blockade of serotonergic signaling in the bee brain To inhibit serotonergic signaling at the level of the mushroom bodies, we injected 5-HT antagonists via the ocellar tract. A 10 μL microsyringe (World Precision Instruments, Friedberg, Germany) was used to inject 250 nL of drug solution into the brain of immobilized bees. The syringe tip was inserted approximately 50 μm through a small hole pierced in the cornea of the median ocellus, following the procedure described by Tedjakumala et al.34 This technique has been validated in previous studies, demonstrating that drugs injected through the ocellar tract diffuse into the brain, achieving homogeneous distribution within less than 5 min.62 Serotonergic neurons and circuits in the honeybee brain are primarily localized in the mushroom bodies and optic lobes, with additional staining observed in the central complex.63,64 Notably, both the optic lobes and the central complex are involved in visual processing, whereas our conditioning experiments were purely olfactory in nature. Therefore, any off-target effects of our serotonergic blockade beyond the mushroom bodies would likely have impacted primarily visual circuits, which are not relevant to our experimental design. To ensure sufficient effective pharmacological effect, injections were performed 30 min before conditioning, in accordance with previous works demonstrating that biogenic amine receptor antagonists reach peak efficacy during this time window.34,36,65 The treatment consisted of a serotonin antagonist-receptor cocktail containing methiothepin mesylate and ketanserin tartrate (both from Sigma Aldrich, Saint-Quentin-Fallavier, France). Two concentrations were tested: 1 mM and 10 mM, in phosphate-buffered saline (PBS). Control bees were injected with PBS alone. Quantification and statistical analyses The percentage of bees exhibiting a conditioned PER (%PER) in response to the odorant stimulation was calculated for each trial and test. Ninety-five percent confidence intervals were computed using a bootstrap method. Learning performance was analyzed using a Generalized Linear Mixed Effects Model (GLMER) for binomial family. Individual identity was included as a random factor, while fixed effects included conditioning type (delay vs. trace), memory test time (1 h vs. 24 h), conditioned stimulus (1-nonanol vs. 2-hexanol) and trial number. For the distraction experiment, distractor presence (yes/no) was added as an additional fixed factor. In the pharmacological experiment, treatment group (control, cocktail 1 mM, cocktail 10 mM) was included as a fixed factor. For both delay and trace conditioning, model selection was performed using an information theoretic approach based on Akaike Information Criterion (AIC). Z-values and p-values are reported throughout the manuscript. Within-group comparisons of CS vs. NOd responses were conducted using McNemar tests, appropriate for paired binomial data. Between-group comparisons (CS and NOd response levels across groups) and comparisons of the percentage of specific learners (bees responding to the CS but not the NOd) were analyzed using χ2 tests with reported degrees of freedom. Between-group comparisons of maximal PER latency and number of licking events were conducted using Wilcoxon rank-sum tests (W statistic reported). Variance in PER latency between Delay and Trace groups was compared using one-tailed F-tests (F value reported), reflecting our a priori expectation of a greater latency variability in the Trace group. All statistical analyses were conducted using the R 4.0.3 software (R project). GLMERs were performed using the lme4 package. Post-hoc analyses were done using the emmeans and lsmeans packages.