Statistical olfactory learning in honey bees

Summary

Statistical learning is a cognitive process for detecting regularities in sensory inputs typically presented as strings of sounds, shapes, or objects, enabling species to predict future events, thereby guiding decision-making and behavior. Such an adaptive trait has been demonstrated in vertebrates, including human and non-human primates, birds, and dogs. It remains unclear whether invertebrates, which possess smaller and simpler neural systems than vertebrates, can extract statistical information from sensory inputs. Here, we show for the first time that honey bees are able to learn and recall the temporal (statistical) structure of an olfactory stimulus. These results suggest that statistical learning is a fundamental component of a conserved cognitive toolkit present, even in invertebrates.

Graphical abstract

Highlights

• •Honey bees can learn and recall the temporal sequence of two odorants (AB vs. BA)
• •Bees encode the full sequence pattern, not just the position of individual odorants
• •Both odorants are learned, but the one near the reward dominates
• •Statistical learning is a conserved cognitive capacity present also in invertebrates

Zoology; Entomology; Neuroscience

Introduction

Our surroundings are rich in regularities distributed across space and time, occurring in different sensory modalities. The ability to detect such regularities is crucial for a species’ successful adaptation to the environment. Statistical learning is a fundamental mechanism for identifying regularities in sensory inputs, supporting core functions such as visual processing, auditory discrimination, and early language-learning.^1,2 Tracking regularities organized temporally, as in a string of sounds or odorants, allows species to anticipate future events and guide their behavior.

Comparative studies have focused on vertebrate species so far (e.g., human and non-human primates,^3,4,5 rats,⁶ birds,⁷ and dogs⁸ to name a few), implementing regularities in sensory modalities that are relevant to each species and experimental protocols that best adapt to their ecology.^2,9 Baboons (Papio papio) were trained to discriminate patterns of visual objects organized sequentially according to ABA or ABB patterns, with A and B being two arbitrary objects, and generalize such patterns to new strings composed of different objects (e.g., CDC vs. CDD).⁴ Baboons learned patterns with adjacent repetitions (e.g., ABB) more easily. One-to-three-day-old human newborns learned the order of appearance of pairs of shapes presented sequentially, based on transitional probabilities (e.g., the square always followed the triangle, and the circle always followed the x shape, whereas the probability between pair was 0.5).³ This experimental paradigm was adapted to investigate visual statistical learning in newborn chicks.⁷ Right after hatching, visually naive chicks (Gallus gallus) were familiarized with a similar sequence of shapes organized according to the same statistical structure. Chicks were then tested with two visual displays: one showing the familiar (structured) sequence and the other showing a random presentation of the same shapes. The chicks were able to discriminate between the structured and the random sequence, in line with the results obtained in human newborns. More relevant for the present study is the evidence on rats (Rattus norvegicus), which were trained to differentiate strings of sounds organized in opposite orders, such as AB or BA (A and B being two arbitrary sounds). Rats successfully learned the temporal regularities and discriminated between the two string types.⁶

To this day, it remains unclear whether invertebrate species, equipped with a simpler brain, are capable of extracting statistical regularities from the environment. We tackled this issue by investigating the domestic honey bee (Apis mellifera), an established model in cognitive neuroscience.^10,11 We used an experimental paradigm for appetitive associative learning: the olfactory conditioning of the proboscis extension reflex (PER).^12,13 This is a robust protocol that is ecologically relevant for honey bees and allows controlled investigations of olfactory learning, memory, and perception. In this paradigm, bees associate a neutral odorant with a sugar reward whose application to the antennas triggers a proboscis extension reflex (see Video S1). Once the association is established, the odorant alone can trigger the proboscis response, providing an effective readout for exploring the mechanisms of appetitive learning in controlled conditions. This approach was recently explored in a preliminary study investigating the ability of bees to learn and generalize the pattern of three-odorant strings.¹⁴ The results, while showing only partial success, offer potentially useful observations regarding the constraints of such complex learning tasks. The learning phenomena observed in bees (e.g., trace conditioning,¹⁵ second-order conditioning,¹⁶ and extinction learning¹⁷) mirror those found in vertebrates, suggesting that the underlying neuronal mechanisms may be based on common principles.¹⁸ These parallels with learning in vertebrates suggest that, despite their simpler cognitive systems, honey bees are capable of complex forms of learning. Given their well-characterized neurobiology, robust behavioral paradigms, and the evolutionary distance from vertebrates, honey bees provide a powerful comparative model for investigating whether statistical learning is a conserved feature of animal cognition.

Video S1. Exemplary video recording of a memory test tracked with DeepLabCut software

The video shows a restrained honeybee facing the outlet of the olfactometer. Colored dots indicate: the joints of right and left antennas (scapus, proximal, and distal extremities of the flagellum); the initial, medial and terminal portion of the proboscis; the tip of the left and right mandibulae; one of the LEDs on the Arduino, signaling the onset of the olfactory sequence, which was used as the t = 0 s reference point in our analysis.

Here, we investigate bees’ ability to learn, discriminate, and recall short strings of odorants based on the temporal order of the composing odorants rather than on their chemical composition. Bees were exposed to two-element odor strings comprising the same odorants but presented in opposite order (e.g., AB vs. BA). The odorants used to implement the strings were easily identifiable and distinguishable by bees¹⁹ (see STAR Methods).

Results and discussion

Honey bees can learn the temporal sequence of two odorants

We trained bees to discriminate between odor strings comprising two elements, A and B, presented in opposite orders: AB vs. BA (Figure 1A). The rewarded/conditioned string (CS+) (e.g., AB) was paired with an appetitive sugar solution, while the reversed string (CS-) (e.g., BA) was unrewarded. The differential conditioning protocol comprised six stimulations with the CS+ and six stimulations with the CS-, delivered pseudo-randomly with a 10-min inter-trial-interval. To control for odor-related biases, AB and BA were alternated as the rewarded stimulus (CS+). Two-odorant sets were used as A and B, i.e., (±)-2-hexanol and nonanal (n = 223 bees) and 1-hexanol and (±)-2-nonanol (n = 202 bees). Bees did not show any a priori preference for the odorants, which were selected for being well distinguished by the animals¹⁹ (see STAR Methods). For clarity, throughout the whole manuscript, we will refer to AB as the CS+ and BA as the CS-.

Figure 1.

Differential sequence learning and 1-h memory test in honey bees

(A) Scheme of the differential appetitive conditioning protocol: six exposures to the conditioned sequence (CS+, e.g., AB), each followed by a 3-s sucrose reward with a 1-s overlap with the second element of the sequence, and six exposures to the unrewarded reversed sequence (CS-, e.g., BA). One hour after the last learning trial, bees were divided into four groups and subjected to different memory retention tests. Abbreviations: groups 1–4, Gr. 1–4; unconditioned stimulus, US.

(B) Learning curve for the differential conditioning phase for all individuals independently of the odorant strings used as conditioned stimuli (n = 425). Data represent the % of bees showing PER during the 10-s odor sequence, with 95% confidence intervals. GLMM analysis shows a significant response difference from trial three (T3) onward (∗, p ∼0). Abbreviations: proboscis extension response, PER. See also Figure S1 and S2, and Table S1.

(C) Each conditioned bee underwent two 1-h memory tests to evaluate the relevance of different features of the conditioned stimulus for memory recall. Data represent the % of bees showing PER during the test sequence, with 95% confidence intervals. The number of individuals (n) and p-values from McNemar’s test comparing response rates to the tested stimuli are displayed above each plot.

(D) Memory tests were video recorded and proboscis extension was tracked to quantify response dynamics during olfactory stimulation (see also Video S1). Differences in proboscis responses probabilities [%] to the different stimuli within each pattern were tested using a Wilcoxon signed-rank test. p values corrected for false discovery rate (FDR) are shown above each plot.

For both sets of odorants, and independently from which string was used as CS+, bees discriminated the strings from the third training trial onwards (generalized linear mixed model [GLMM], trial/stimulus interaction effect, from T3 to T6 p ∼0.000; Figure 1B; Table S1; Figures S1 and S2). Discrimination improved across trials, indicating progressive task learning.

Because four groups of bees were conditioned with different odorant string pairings, we specifically tested whether the rewarded configuration (i.e., which odorant string served as CS+) influenced learning performance. For this, we fitted a GLMM including stimulus type (CS+/CS-), learning trial, and conditioning group as fixed factors. The three-way interaction (trial × stimulus × conditioning group) was not statistically significant (F_[15,10150] = 0.79, p = 0.69), suggesting that learning progressed similarly regardless of the rewarded configuration (effect sizes [odds ratio] of individual three-way interactions ranged from 0.44 to 1.46, with 95% confidence intervals all including 1, indicating negligible effects). In other words, bees trained with any of the four CS+/CS- pairings exhibited comparable levels of associative learning (Figure S2). Based on this, all individuals were pooled for the main analysis (Figures 1B–1D).

Analysis of odor string memory: Which elements guide recognition?

Since both odorants predicted the presence or absence of food reward, recognition of the correct string must have occurred based on the temporal order of appearance of the odorants. One hour after conditioning, we split the bees into four independent groups for memory tests, each probing which features of the odor strings were used by bees to learn the temporal pattern.

Group 1

Bees were presented with the conditioned and reversed odorant strings (i.e., AB and BA) and showed significant recognition of the CS+ and the CS- (McNemar’s test, n = 79, p = 9.1∗10⁻⁹, odds ratio = 71; Figure 1C). These results indicate that the previously acquired information on the odorant string was retained for at least 1 h.

Group 2

Bees were tested with individual odorants, A or B alone, to assess whether they memorized the strings as whole entities or whether they focused on each odorant. Bees showed a significantly higher PER probability to the second odorant of the previously rewarded string (n = 98, p = 1.1∗10⁻⁸, odds ratio = 0.10; Figure 1C). These findings suggest that, in addition to discriminating strings based on the position of the odorants, bees responded significantly more to the second odorant (i.e., B). It is thus plausible that bees assigned greater value to B for its adjacency to the reward in half of the learning trials.

Group 3

Bees were tested with novel combinations of odorants, in which the second odorant of rewarded and unrewarded strings was replaced by a novel, neutral odorant C (also well-perceived by bees¹⁹), i.e., AB→AC; BA→BC. Bees trained with the first odorant set ((±)-2-hexanol/nonanal) received (±)-2-octanol as the novel odorant C, while bees trained with the second odorant set (1-hexanol/(±)-2-nonanol), received 2-hexanone. This manipulation allowed us to test if the first odorant acquired predictive value, leading to higher PER to C in the AC string but not in the BC string. Results showed that bees responded significantly more to BC, which included B, the odorant adjacent to the reward during training (n = 88, p = 2.0∗10⁻⁷, odds ratio = 0.02; Figure 1C).

Group 4

Bees were tested with three-odorant strings, in which the novel odorant C was added at the beginning of the rewarded and the unrewarded strings, i.e., CAB and CBA, respectively. The goal of this test was to assess whether the positional information of odorants served as cue for memory retrieval. During differential conditioning with AB (CS+) and BA (CS-), B was rewarded when it occurred in second position. If bees rely on position, they should exhibit a higher proboscis-extension response probability to CBA, where B occupies the second slot. Conversely, if bees encoded the full pattern, they should respond more to CAB, which preserves the rewarded string AB. We found that bees conditioned with AB responded significantly more to CAB than to CBA (n = 78, p = 3.6∗10⁻⁶, odds ratio = 9), indicating that they encoded the full (AB) pattern rather than just the position of individual odorants (Figure 1C).

All memory tests were video-recorded and analyzed to track PER dynamics during olfactory stimulation (Figure 1D; see also Video S1). The mean response probability across bees was higher when the second odorant of the rewarded string (i.e., B) appeared. This result suggests that bees relied on the second odorant for memory retrieval.

Interestingly, the response probability to the novel odorant C was similar when bees were stimulated with AC or BC sequences (group 3). However, the response to C increased significantly when C followed A (n = 58, p = 0.021, rank-based effect size = 0.32) and decreased significantly when C followed B (n = 53, p = 0.021, rank-based effect size = 0.30). Importantly, C alone did not elicit a response, as shown by the low-response probability when C was presented at the start of a sequence (Figure 1D, group 4). In addition, both A and B, when presented individually, were followed by a decline in response rate (Figure 1D, group 2). Thus, the unexpected increase in response probability to C after A during AC stimulation could suggest that odorant A may have acquired predictive value, influencing responses to the otherwise neutral odorant C.

Conclusive remarks

Overall, the present study shows that bees encode temporal relationships that define sequentially organized odorants and use this information to guide their behavior. Bees learned and recalled a rewarded odor string, either when presented alone or embedded within a more complex olfactory context (e.g., three-element string). Furthermore, bees assigned greater importance to the odorant adjacent to the reward. The increased response to a neutral odorant following the first odorant of the rewarded string (Figure 1D, group 3) highlights bees’ sensitivity to the temporal structure of the olfactory input.

Statistical learning allows tracking of regularities from sensory inputs in vertebrate species with different evolutionary histories^1,2 serving both communicative^2,20 and non-communicative functions.^21,22 Here, we showed that an invertebrate species, the honey bee, can learn regularities from olfactory inputs presented sequentially, and detect changes in those regularities. This ability may assist bees in spatial orientation and foraging by signaling the proximity of landmarks or food sources, offering an evolutionary advantage for navigating complex olfactory environments.

Limitations of the study

This study shows that honey bees are capable of learning the order of two odorants presented sequentially in a string, indicating that they extract statistical information on the relative temporal occurrence of the odorants. While several control test conditions were performed to assess which features of the odor strings were relevant for learning, each bee was not exposed to more than two unrewarded stimuli. This was critical to avoid the extinction of the response to the conditioned stimulus. Therefore, bees tested with AB vs. BA (group 1) were not the same individuals later tested with A vs. B (group 2). Unexpectedly, the response probability to BA was lower than to B alone. This difference could reflect the variability between the two groups of bees in learning strength or memory retention, rather than an effect of the stimuli themselves.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Marco Paoli (marco.paoli@ube.fr).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Data resulting from manual annotation as well as from DeepLabCut video-tracking have been deposited on GitHub and are publicly available (https://github.com/mp599/2025_honeybee_statistical_learning).
• •All MATLAB codes developed for data analysis, statistics, and visualization have been deposited on GitHub and are publicly available (https://github.com/mp599/2025_honeybee_statistical_learning).
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

C.M, M.P., and M.G. thank the CNRS, Inserm and Sorbonne University for support. M.G was supported by an ERC Advanced Grant (COGNIBRAINS, 835032). C.S. was supported by an ERC Starting grant “GALA” (101115991).

Author contributions

Conceptualization, M.P., M.G., and C.S.; investigation, C.M. and M.P.; methodology, C.M., M.P., M.G., and C.S.; data curation, M.P.; visualization, M.P.; software, M.P.; supervision, M.P., M.G., and C.S.; project administration, M.P., M.G., and C.S.; funding acquisition, M.G.; writing – original draft, M.P. and C.M.; writing – review and editing, C.M., M.P., M.G., and C.S.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
(±)-2-hexanol	Sigma-Aldrich	CAS number 626-93-7
nonanal	Sigma-Aldrich	CAS number 124-19-6
hexanol	Sigma-Aldrich	CAS number 111-27-3
(±)-2-nonanol	Sigma-Aldrich	CAS number 628-99-9
(±)-2-octanol	Sigma-Aldrich	CAS number 123-96-6
2-hexanone	Sigma-Aldrich	CAS number 591-78-6
Deposited data
Repository data	GitHub	https://github.com/mp599/2025_honeybee_statistical_learning
Experimental models: Organisms/strains
Honey Bee workers Apis mellifera	N/A	N/A
Software and algorithms
MatLab R2018a	The MathWorks, Inc.	N/A
DeepLabCut	https://github.com/DeepLabCut/DeepLabCut	N/A
Arduino	https://www.arduino.cc	N/A

Experimental model and subject details

Insects

Honey bee foragers from colonies of the Neuro-SU laboratory, located on the Pierre et Marie Curie Campus (Sorbonne Université, 75005, Paris), were collected in the afternoon and housed in custom-printed plastic cages in groups of 16 individuals. Forager bees were collected from a feeder containing a sugar/water solution located approx. 10 m from the hives. Each group was provided with 240 microliters of 50% sugar/water solution (an average of 15 μl/individual) and maintained overnight in a dark and humid incubator at 30°C. The following morning, the bees were individually harnessed in plastic tubes, secured with tape, and fed 3 μl of sugar solution. After a three-hour rest period, they underwent the olfactory conditioning protocol.²³

Method details

Differential appetitive conditioning

Honey bees were subject to a differential conditioning protocol in which one odorant sequence (e.g., AB) was paired with a sugar reward, while the reversed sequence (e.g., BA) was not. Two sets of odorants were used for this experiment. One group of bees (n=223) was trained to discriminate between sequences of (±)-2-hexanol (CAS number 626-93-7) and nonanal (CAS number 124-19-6). Half of these bees were trained with (±)-2-hexanol followed by nonanal as the conditioned sequence (CS+), and with the reversed sequence (nonanal followed by (±)-2-hexanol) as the non-conditioned stimulus (CS-). The remaining bees were trained with the opposite sequence assignments. A second group of bees (n=202) was trained with sequences composed of hexanol (CAS number 111-27-3) and (±)-2-nonanol (CAS number 628-99-9). Half were trained with hexanol followed by (±)-2-nonanol as CS+, while the rest were trained with the opposite sequence assignments. For the olfactory stimulation, undiluted odorants (1 mL) were placed in 20 mL glass vials and the headspace was delivered to the bees using an Arduino Uno-controlled odorant delivery device as previously described.^24,25 All selected odorants are common volatile organic compounds found in flowers. Importantly, they do not possess innate valence and are perceptually distinct to bees.¹⁹

In the differential conditioning protocol, harnessed bees were positioned in front of the odorant delivery device. Each trial began with a 20-second exposure to odorless air (familiarization phase) followed by a 5-second presentation of the first odorant, a 5-second presentation of the second odorant, and a final 20-second exposure to odorless air. For the conditioned stimulus (CS+), a 50% sugar reward was provided for 3 seconds, beginning during the last second of the second odorant presentation. Each bee underwent six rewarded (CS+) and six unrewarded (CS-) trials presented in a pseudorandom order and with a 10-minute inter-trial interval. Bees that failed to extend the proboscis upon sucrose stimulation either before the start of the conditioning protocol or after the final memory test were excluded from the analysis.

Memory retention test

One hour after the last conditioning trial, bees were tested for memory retention using one of the following configurations of olfactory stimuli: (1) the conditioned sequences (e.g., AB and BA; Group-1); (2) the elemental components of the sequence (e.g., A and B; Group-2); (3) a novel sequence composed of the first element of the conditioned sequences followed by a novel odorant (e.g., AC and BC; Group-3); (4) the conditioned sequences preceded by a novel odorant (e.g., CAB and CBA; Group-4).

The tests occurred 10 minutes apart in a pseudorandom order. For bees conditioned with the first odorant set (i.e., (±)-2-hexanol and nonanal), (±)-2-octanol (CAS number 123-96-6) was adopted as the novel odorant (C). For bees conditioned with the second odorant set (i.e., 1-hexanol and (±)-2-nonanol), 2-hexanone (CAS number 591-78-6) served as the novel odorant. All odorants were selected because of their perceptual difference for honey bees¹⁹ and were purchased from Sigma-Aldrich (St. Louis, MO, USA).

Quantification and statistical analysis

Video recording and tracking of proboscis extension

Memory retention tests were recorded at 50 fps using a Sony HandyCam HDR-CX625 video camera. Proboscis extension was tracked using the DeepLabCut software.²⁶ A total of 400 frames were extracted from videos of 20 individuals (20 frames/video) and manually labelled to generate the training dataset. For this study, the base, midpoint, and terminal tip of the proboscis were tracked when visible in the frame). After the first network training session (ResNet50, 500000 iterations), frames with low-likelihood labels were manually corrected to refine the model. The refined network was then used to track proboscis movement across all videos. All videos. A representative example can be found in Video S1.

Data analysis

Data analysis was conducted using MatLab R2024b (The MathWorks Inc.). The trained network reliably identified the proboscis tip during extension, while its absence was consistently associated with low likelihood scores. Hence, a likelihood threshold of 0.8 was applied to detect proboscis extension. Bees that exhibited proboscis extension before stimulus onset were excluded from the analysis, as these instances prevented meaningful analysis of the relationship between olfactory stimulation and appetitive learning. Overall, 579 videos were acquired, of which 84 were discarded because of proboscis movement before stimulus onset, leaving 495 videos for analysis.

Statistical tests

Learning curves (Figure 1B) were analyzed using a generalized linear mixed model (GLMM) using the fitgmle MatLab function. Trial, stimulus, and individual were included as categorical variables. The model assessed the effects of trial and stimulus, as well as their interaction (trial:stimulus), to evaluate learning progression and stimulus-dependent differences. A McNemar’s test was used to compare proboscis extension response scores for the two olfactory sequences presented during the 1-hour memory retention test (Figure 1C). A Wilcoxon signed-rank test was used to compare the average proboscis response activity between the first and second elements of the olfactory sequence. For each odorant, the mean proboscis response across individuals was calculated in the middle of the olfactory stimulation, i.e., between 2 and 3 seconds after odorant onset (Figure 1D). A post-hoc false discovery rate (FDR) correction was applied to the resulting p-values.

Published: September 27, 2025