Worker honeybees were collected from outdoor colonies in the afternoon before the training day, immobilized on ice, restrained into small cylindrical metal tubes, and fed ad libitum with 30% sugar-water solution. They were then kept overnight in a dark, humid chamber (25°C and 60% humidity). In the morning of the following day, these bees were tested for their motivational state by touching the antennae with a 50% sucrose solution. Only bees which exhibited a clear PER were chosen for the experiment. Since recording electrodes were inserted, only one bee was trained at a time.

Animals were placed in front of an exhauster. Only one bee at a time was conditioned to differentiate between 1-Nonanol and 2-Hexanol. One odour was rewarded (CS+) and overlapped with an unconditioned stimulus (US; sucrose solution), the other one was unrewarded (CS-) without US (Experiment 1, Figure 1A). Eight bees received 1-Nonanol as CS+ and 2-Hexanol as CS-, whereas and eight bees were trained contrariwise. As both combinations were learned equally well, data were combined (Supplementary Figure S1). For olfactory stimulation, a Syntech CS-55 (Ockenfels Syntech GmbH, Kirchzarten, Germany) was used, generating a continuous air flow (1.0 L/min) in which the odour stimulus was injected by switching between an empty pipette (permanent stimulus) and a pipette prepared with a filter paper (2 cm²) soaked with 5 µL of the pure odorant. Switching between the two chambers avoids pressure loss and minimised physical artefacts. CS+ and CS- were presented 10 times each, in a pseudo-randomised order. Stimulation lasted 10 s with an ITI of 10 min. The US lasted 3 s and overlapped with the last 3 seconds of the CS+ (Figure 1A).

Animals were placed in a box (about 10 × 10 cm) which could be illuminated from above. Only one bee at a time was conditioned to differentiate between blue (435 nm) and green (488 nm) light, provided by a cold light lamp (Schott, KL1500 electronic) and monochromatic filters. One light stimulus was rewarded (CS+) and overlapped with an unconditioned stimulus (US; sucrose), the other one was unrewarded (CS-) without a US (Experiment 2, Figure 1A). Eight bees received green light as CS+ and blue light as CS-, whereas again eight bees were trained contrariwise. As both combinations were learned equally well, data were combined (Supplementary Figure S1). Stimulus duration was controlled using an electronical shutter driver (A Vincent Associates, UNIBLITZ Model VCM-D1, Rochester, NY, United States of America) and set to 10 s. The US lasted 3 s and overlapped with the last 3 seconds of light stimulation (Figure 1A). CS+ and CS- were presented 10 times each, in a pseudorandomized order with an ITI of 10 min. The experimental setup remained unchanged during both visual and olfactory conditioning (i.e., the lamp was kept above during olfactory conditioning and the pipette remained in front of the animal during visual conditioning) in order to keep a constant environment.

During compound conditioning we followed our recently established procedure (Becker et al., 2019). In short: two odours (Hexanal and Geraniol) were delivered as described above using a Syntech CS-55. To control and synchronise the light stimulation with the odour presentation, we used the TTL output of the Syntech CS-55. Two monochromatic LEDs (UV: 375 nm, intensity: 7.5×10¹³ photons cm⁻² s⁻¹; green: 525 nm, intensity: 3.93×10¹³ photons cm⁻² s⁻¹) were placed at the end of a transparent Plexiglass rod (diameter: 10 mm, length: 100 mm). Light transmission through the Plexiglass rod ensured diffusion before the light would reach the eye. Our custom build light device was positioned 3 cm above the bee’s head and orthogonal to the odour stimulation device (see: Becker et al., 2019). The experimenter could switch manually between the LEDs or set the light stimulus off, if an odour was presented alone. To present light alone, the stimulus pipette including the odour soaked filter paper was replaced by another empty pipette. To make sure that odour identity as well as light identity had no influence on the discrimination performance, each odour stimulus was combined with each light stimulus resulting in four different odour-light combinations. For each combination, 15 bees were trained to discriminate the unrewarded single elements light and odour (CS-) from the rewarded olfactory-visual compound stimulus (CS+). CS+ and CS- were presented 10 times each, in a pseudo-randomised fashion. Since there were 20 CS- trials (10 x odour, 10 x light), we shortened the experimental procedure to keep the motivation of the bees high by using an ITI of 5 minutes. Additionally, the CS+ and CS- presentation was shortened to 7 s, whereas the US presentation was same as before (Figure 1A). In all four combinations, bees learned to associate the olfactory-visual compound with the reward and to discriminate it from its single unrewarded elements (odour alone, light alone; Supplementary Figure S2).

In order to monitor the M17 muscle, we recorded differentially from two isolated silver wires (diameter: 20 µm) inserted at the approximate attachment point of the M-17 muscles between the ocelli and the compound eyes (Rehder, 1987; Smith and Menzel, 1989a). One wire was positioned on the left, the other one on the right hemisphere. To reduce electrostatic noise and get better visual access to the bees ‘neck’, the head was shaved with a razor blade beforehand. A third wire served as ground electrode and was attached to the metal tube. Recording signals were sampled at 15 kHz using an A/D converter (CED Micro1401, Cambridge Electronic Design, Cambridge, United Kingdom), and visualised using Spike2 software (Cambridge Electronic Design, Cambridge, United Kingdom). In parallel, we recorded a TTL signal of the stimulation devices to precisely analyse response latencies (Figure 1B). Please note, that the latency distribution analysis in Figure 5 includes only trials, in which bees responded to the CS + stimulus. In trials with no response, we were not able to calculate a latency. Temporal parameters like stimulation time and inter-trial interval (ITI) were controlled using the computer programme “TimingProtocol“ (Lichtenstein et al., 2018), which provided precise acoustical cues to the experimenter.

To compare CS- and CS + responses, we used a Chi²-test for the 10th trial and the memory test. To account for multiple comparisons in the multimodal conditioning experiments, were one CS+ and two CS- stimuli (three stimuli in total) were compared (Figure 2; Figure 3; Figure 4), we adjusted the alpha-level to p < 0.016 (0.05/3) after Bonferroni’s correction. The different latency distributions of the three conditions were compared using an ANOVA (p < 0.05 followed by post hoc Wilcoxon rank sum test with an adjusted alpha-level of p < 0.016 (0.05/3) to account for multiple comparisons after Bonferroni’s correction (Figure 5).

Since the different odour combinations were learned equally well (Supplementary Figure S1), we combined both data sets, resulting in a total of 16 animals. During the acquisition phase, bees learned to significantly differentiate between the rewarded (CS+) and the unrewarded (CS-) odour stimulus, statistically tested at the last conditioning trial (trial 10; Chi²-test; p < 0.001, Figure 2A) and could correctly recall that valence 1 hour later during a memory test (MEM; Chi²-test; p < 0.001, Figure 2A). In the course of the learning phase, the M17 response to the CS + became faster and finally reached a response latency of about 400 m (median) during the memory test (Figure 2B). This value is in the range of odour processing speed after operant conditioning (Ditzen et al., 2003) and classical conditioning (Strube-Bloss et al., 2011) reported in earlier studies.

Since both of the tested light combinations were learned equally well (Supplementary Figure S1), we combined the data of these experiments resulting in a total of 16 tested animals. During acquisition, bees learned to differentiate significantly between the rewarded (CS+) and the unrewarded (CS-) light stimulus (trial 10; Chi2-test; p < 0.001, Figure 3A) and remembered this stimulus valence also during a memory test (MEM; Chi2-test; p < 0.001, Figure 3A). Similar to the odour learning experiment, the time between stimulus onset and the activation of the M17 shortened with each trial (Figure 3B). However, in comparison to the response latencies observed during odour conditioning (Figure 2B), response time was much slower and reached 2.9 s (median) during the memory test (Figure 3B). Thus, although the acquisition pattern was similar during odour and visual learning (Figure 2A; Figure 3A), the time from stimulus onset to the final proboscis extension seemed strongly delayed for the light stimulus.

We tested all combinations of two odour and two light stimuli in a classical positive patterning experiment, in which the single elements light and odour were presented unrewarded (CS-) and their combination in an olfactory-visual compound was presented rewarded (CS+). Since all four combinations elicited a reliable discrimination between the CS- elements and their multimodal compound (CS+), which represents an olfactory-visual reward association (Supplementary Figure S2), we merged data of all CS+ and each CS- element (Figure 4A), resulting in an absolute number of 60 tested bees. To establish a reliable multimodal reward association, which can be discriminated from its single elements, at least eight acquisition trials of each CS kind were needed. Compared to unimodal discrimination learning, were one to two trials (olfaction, Figure 2) or three to four trials (light, Figure 3) were sufficient to reach a high level of correct responses, the complexity of the multimodal learning task caused a higher rate of false positives to the single elements, which seemed to be dominated by the odour element (Figure 4A, trial 1–5). From trial 7 onwards, however, both single elements were discriminated from their compound, which was highly significant at the last acquisition trial (trial 10: Chi²-test; p < 0.0003; after Bonferroni correction), resulting in a stable and significant discrimination after 1 h consolidation time (MEM: Chi²-test; p < 0.0003; after Bonferroni correction).

Similar to the unimodal learning (Figure 2; Figure 3) the response latencies showed a continuous decline over subsequent trials, indicating that the decision process to extend the proboscis accelerated with each trial (Figure 4B), resulting in a mean latency of 1.69 s (SD ± 1.04 s) at the 10th conditioning trial. Compared to the response latency, which was reached at the last conditioning trial during odour conditioning (mean = 0.85 s; SD ± 1.38 s), the processing of and responding to the CS + took about 0.84 s longer and was obviously delayed. However, when comparing the response latency of the last olfactory-visual CS + acquisition trial to the last visual CS + conditioning trial (mean = 3.86 s; SD ± 1.61 s), triggering the response behaviour (PER) was accelerated by about 2.17 s. Thus, the olfactory-visual compound seemed to be processed faster than the pure light stimulus, but slower than the pure odour stimulus.

In all three experiments, the learning performance for the reward-associated stimuli (CS+) reached about 80%–90% at the last conditioning trial and were significantly different from the responses towards the unrewarded (CS-) stimuli. Moreover, the stimulus values were correctly remembered in the memory tests of all experiments (Figure 2A; Figure 3A and Figure 4A). We therefore compared the M17 response latency distributions between the three CS + stimuli (light, odour and compound) for the last conditioning trial (Figure 5A) and the memory test (Figure 5B). The medians for the last conditioning trial (Figure 5A) were 4.0 s for light, 0.4 s for odour and 1.4 s for the olfactory-visual compound stimulus. Almost the same values were observed during the memory test (light: median = 2.9 s; odour: median = 0.4 s; compound: median = 1.1 s). The difference of the M17 response latencies among the three experimental groups was significant for the last conditioning trial (Figure 5A; ANOVA p < 0.001, post hoc Wilcoxon rank sum test; Bonferroni correction, p < 0.0003) as well as for the memory test (Figure 5B; ANOVA p < 0.001, post hoc Wilcoxon rank sum test; Bonferroni correction, p < 0.0003). Thereby, the slowest response behaviour of the proboscis was triggered by the visual stimulus and the fastest PER by the odour stimulus. Most strikingly, the PER elicited by the olfactory-visual compound stimulus was located intermediate, between the unimodal stimuli. Thus, our results indicate that the olfactory-visual compound is processed faster than the light stimulus, and therefore, accelerates the unimodal light response. However, as the compound response is processed slower than the odour stimulus, it follows that it also slows down the unimodal odour response.

To precisely monitor the response latency during classical differential conditioning in which animals had to discriminate between two odour stimuli (experiment 1), two light stimuli (experiment 2) and an olfactory-visual compound stimulus and its single elements (odour and light; experiment 3), we recorded a myogram of the muscle M17, which is activating the proboscis. The M17 muscle on each hemisphere connects the ligular arm of the proboscis with the inner wall of the head capsule exoskeleton, where it is attached dorsally between the compound eye and the outer ocelli (Snodgrass, 1956). Such recordings were successfully used to analyse temporal aspects of odour discrimination (Smith and Menzel, 1989a; Smith and Menzel, 1989b) and reward expectation (Gil et al., 2008). Here we extended the scope of application to better understand modality-specific response latencies after classical conditioning of uni- and multi-modal stimuli. To ensure that the animals establish an olfactory-visual reward association we largely followed the CS-US design for multimodal positive patterning (PP) experiments introduced by Mansur and others (2018). Since this procedure includes 10 more CS- trials, the probability to receive a reward decreased from 50% during unimodal learning experiments (10 x CS+, 10 x CS-) to only 33% (10 x multimodal CS+, 10 x CS- light, 10 x CS- odour). We therefore shortened the ITI to 5 min to keep the animals motivated. Since bees were able to successfully solve multimodal PP experiments with even shorter ITIs (Mansur et al., 2018) we were certain that the animals can build a solid olfactory-visual reward association.

As already observed for odours (Figure 2; see also Smith and Menzel, 1989b), the conditioning of a light stimulus also led to a decreased M17 response latency in the course of training (Figure 3). Whereas odour learning resulted in a median latency of about 0.4 s during the memory test (Figure 2), which is in line with previous studies (Smith and Menzel, 1989a; Smith and Menzel, 1989b; Gil et al., 2008; Strube-Bloss et al., 2011), visually induced response latencies were found to be much longer (2.9 s median in the memory test, Figure 3B). Until recently it was assumed that visual conditioning in restrained honeybees is only possible, if either the antennae of the bees were cut off completely (Hori et al., 2006) or at least the flagellar segments at the tip were removed (Niggebrügge et al., 2009) to force the bees to learn the visual modality. However, in more recent studies it was proven that bees with intact antennae can also be trained to light stimuli (Dobrin and Fahrbach, 2012; Lichtenstein et al., 2018; Lichtenstein et al., 2019). Furthermore, it was shown that the length of the conditioned stimulus seemed to play a critical role, and that the light has to be presented at least 7 s in total with 3 seconds of overlap with the reward to be successfully learned (Lichtenstein et al., 2018). Our experimental stimulation design followed these rules (Figure 1), which resulted in a reliable acquisition in all visual discrimination tasks (Supplementary Figure S1). We were therefore certain that the measured response latencies reflect a realistic difference in computing a visual reward association compared to an olfactory one. Moreover, our findings were supported by visual observations of PER activity in bumblebees, showing latencies of about 3 s in response to an unimodal colour stimulus after absolute conditioning (Riveros, 2023).

As with olfactory and visual stimuli (Figure 2; Figure 3), the conditioning of an olfactory-visual compound stimulus resulted in a decreasing response latency during the acquisition phase (Figure 4), indicating that the animals become faster to evaluate the rewarded stimulus correctly and extend their proboscis in expectation of the reward. The response time reaches a median of 1.4 s during the last conditioning trial (Figure 5A) and was almost stable 1 hour later during the memory test (median = 1.1 s; Figure 5B). During compound stimulation both modalities were presented for 7 s in a completely overlapping manner. In a recent study, in which such a compound stimulus was used in a conditioning experiment in bumblebees (Riveros, 2023), the author measured response latencies of ca. 2 s on average, which was only about 0.5 s slower compared to our findings in honeybees. This slight discrepancy might be related to the different methods used. Whereas we precisely recorded the muscle potentials activating the PER, Riveros (2023) visually observed the PER every 0.5 s triggered by a metronome sound. However, the study supports our findings, since the unimodal visual response time (∼3–4 s in both species) appears to be delayed by about 1–2 s in honeybees (Figure 3) and bumblebees (Riveros, 2023) compared to the olfactory-visual compound. Thus, our data support the efficacy-based hypotheses (Kulahci et al., 2008), namely, that multimodal stimuli enhance decision making, resulting in faster behavioural responses. On the other hand, when comparing our data to an olfactory stimulus, we found the opposite effect. In this case, the olfactory-visual compound stimulus delayed the response time, which might be the result of prolonged processing of the multimodal olfactory-visual reward association compared to a unimodal odour reward association. Although it seems disadvantageous for the animal at a first glance, such increased processing/sampling time was reported to improve the ability to recognise and differentiate low and very low odour concentrations (Wright et al., 2009) and thus increased the reliability of correct behavioural responses. Taken together, our results provide evidence for the basic assumption of multimodal integration, namely, the interaction of single modalities of a multimodal compound with each other by weakening and/or enhancing the processing of these single modalities (Calvert et al., 2004). In our case the response time to the olfactory modality was decelerated whereas the behavioural response time to the visual modality was accelerated compared to response time induced by the multimodal olfactory-visual compound stimulus (Figure 5).

Multimodal integration involves neural convergence of different sensory pathways. Olfactory projection neurons of the antennal lobes (AL) as well as visual projection neurons of the optic lobes innervate the calyces of the mushroom bodies (MB), where they converge on about 180.000 Kenyon cells, the MB principal neurons (for review see: (Fahrbach, 2006; Menzel, 2014; Groh and Rössler, 2020)). The information is further converged to a group of approximately 400 mushroom body output neurons (MBON) with long-range centrifugal connections to the lateral horn, protocerebral lobe, AL and to the contralateral brain side, and close-range feedback connections to the MB calyx (Rybak and Menzel, 1993; Strausfeld, 2002). Recent experiments have shown that a MBON subpopulation, which account for about 32% in the related study, respond to both visual and olfactory stimuli (Strube-Bloss and Rössler, 2018). The population activity in these multimodal MBONs separates olfactory from visual information, as well as unimodal information from their olfactory-visual compound, and thus is able to categorise incoming stimuli. An additional category of visual stimuli was shown to be computed by another subpopulation of MBONs, which separates UV-light from other visual stimuli (Schmalz et al., 2022).

Classical differential odour conditioning recruits insensitive MBONs to code for the reward-associated stimulus (Strube-Bloss et al., 2011; 2016). Therefore, the reward association can be a combination of complex stimulus features including odour identity and stimulation side encoded in the recruited MBON population activity. For example, an odour conditioned to the left antennae evoked a different activity pattern compared to the same odour presented to the right antennae (Strube-Bloss et al., 2016). Thus, the activity of recruited MBON integrates different stimulus features to a new representation of the reward associated compound stimulus. Such a compound was learned in our experiment 3 (Figure 4), in which the animals had to differentiate the single elements (odour and light) from their rewarded olfactory-visual compound. In this respect, it is most likely that our conditioning procedure was accompanied by the recruitment of initially non responding MBONs (ca. 17% of the population; (Strube-Bloss and Rössler, 2018)), which code after learning for the olfactory-visual compound. These MBONs may encode the behavioural decisions during sensory-motor transformation, as it was shown for MBONs in untrained cockroaches (Arican et al., 2022). In addition to a previously studied group of GABAergic MBONs, forming recurrent circuits within the MBs (Grünewald, 1999a; 1999b; Haehnel and Menzel, 2010; 2012), anatomical characterisation revealed MBONs with potential top-down recurrent connections to the AL (Rybak and Menzel, 1993; Kirschner et al., 2006). If we assume that such centrifugal-feedback neurons were recruited during the olfactory-visual conditioning they may allow the visual modality to influence odour processing at the AL-level, and might be responsible for the delayed response time we measured. Such cross-modal modulation of olfaction has been shown in the moth Spodoptera littoralis, in which the predator sound of a bat induced long-term sensitization of olfactory neurons in the AL (Anton et al., 2011). However, multimodal feedback might not be involved during olfactory learning, which might be the reason for the fast response time we observed. On the other hand, a similar type of centrifugal feedback connectivity like described above, might exist connecting the MB output with the optic lobes. In that case, the olfactory modality might modulate the visual processing and speed up its computation, resulting in a faster response compared to the unimodal visually triggered response. Knowledge about the influence and contributions of different MBON subpopulations will be addressed in future experiments, in which we will combine olfactory-visual compound conditioning with extracellular long term recordings of MBONs.