Glyphosate impairs aversive learning in bumblebees

Agrochemicals represent prominent anthropogenic stressors contributing to the ongoing global insect decline. While their impact is generally assessed in terms of mortality rates, non-lethal effects on fitness are equally important to insect conservation. Glyphosate, a commonly used herbicide, is toxic to many animal species, and thought to impact a range of physiological functions. In this study, we investigate the impact of long-term exposure to glyphosate on locomotion, phototaxis and learning abilities in bumblebees, using a fully automated high-throughput assay. We find that glyphosate exposure had a very slight and transient impact on locomotion, while leaving the phototactic drive unaffected. Glyphosate exposure also reduced attraction towards UV light when blue was given as an alternative and, most strikingly, impaired learning of aversive stimuli. Thus, glyphosate had specific actions on sensory and cognitive processes. These non-lethal perceptual and cognitive impairments likely represent a significant obstacle to foraging and predator avoidance for wild bumblebees exposed to glyphosate. Similar effects in other species could contribute to a widespread reduction in foraging efficiency across ecosystems, driven by the large-scale application of this herbicide. The high-throughput paradigm presented in this study can be adapted to investigate sublethal effects of other agrochemicals on bumblebees or other important pollinator species, opening up a critical new avenue for the study of anthropogenic stressors.


Introduction
The term "Anthropocene" has been proposed to describe the current geological epoch, thus emphasizing the impact of our species on ecosystems worldwide (Lewis and Maslin, 2015). The global decline in insect diversity and abundancy (Dirzo et al., 2014;van Klink et al., 2020;Hallmann et al., 2017;Seibold et al., 2019), including essential pollinators (Potts et al., 2010;Powney et al., 2019), is believed to be part of this heavy human footprint. Agrochemicals figure prominently among the multiple interacting anthropogenic stressors that contribute to this crisis (Goulson et al., 2015;Sánchez-Bayo and Wyckhuys, 2019). Indeed, it is increasingly clear that the impacts of these chemicals on non-target species have been wildly underestimated (Daam et al., 2019;Topping et al., 2020;Uhl and Bruhl, 2019). Current risk assessments consider only mortality rates, ignoring the many non-lethal effects that can reduce an individual's chances at reproduction and survival (Uhl and Bruhl, 2019;Barascou et al., 2021). In addition, they fail to account for the variety of environmental conditions in which a chemical might be used, or for interactions between stressors (Goulson et al., 2015; Sánchez-Bayo and Wyckhuys, 2019; Daam et al., 2019;Topping et al., 2020;Uhl and Bruhl, 2019).
Glyphosate is the active component of many herbicide formulations, and as such it is both the most widely used agrochemical worldwide (Benbrook, 2016;Maggi et al., 2020) and one of the most controversial. While its potential carcinogenic effect on humans has been re-attributed to the surfactants used in some formulations (Kudsk and Mathiassen, 2020;Meftaul et al., 2020), studies have now revealed toxic effects of glyphosate or its formulations on aquatic invertebrates, amphibians and fish [reviewed in (Matozzo et al., 2020;Tresnakova et al., 2021)], but also on terrestrial insects and spiders (Battisti et al., 2021;Abraham et al., 2018;Evans et al., 2010;Tahir et al., 2019;Smith et al., 2021). In particular, a range of non-lethal but detrimental effects have been reported in bees exposed to glyphosate or glyphosate-based herbicides at field-realistic concentrations (Maggi et al., 2020;Giesy et al., 2000;Thompson et al., 2014;Rubio et al., 2014;Zioga et al., 2022;Thompson et al., 2022;Karise et al., 2017). This includes delayed or abnormal larval development (Tome et al., 2020;Vazquez et al., 2018), disturbed sleeping patterns (Vazquez et al., 2020) and impaired appetitive behaviour (Herbert et al., 2014;Luo et al., 2021;Mengoni Gonalons and Farina, 2018), navigation (Balbuena et al., 2015), vision (Helander et al., 2023) and memory formation (Hernández et al., 2021). At the collective level, glyphosate-exposed bumblebee colonies were unable to maintain their brood at the optimal temperature when confronted with food shortage (Weidenmuller et al., 2022). It is important to note that while most of these reported changes have a limited breadth and effect size, taken together they likely have a significant impact on colony function and survival.
Glyphosate kills plants by inhibiting the Shikimate pathway, an essential step in the production of aromatic amino acids. This pathway is not found in animals, but it is used by some commensal micro-organisms (Leino et al., 2021). Indeed, glyphosate ingestion has been shown to impact the composition and function of the gut microbiota in both honeybees (Motta et al., 2018) and bumblebees (Cullen et al., 2023), thus potentially altering the metabolism of affected bees. In addition, glyphosate exposure increases the susceptibility of insects to infection by reducing immune responses, in particular the melanization pathway (Smith et al., 2021;Motta et al., 2020;Motta et al., 2022). How these mechanisms of action result in the pleiotropic effects of glyphosate on the behaviour and physiology of bees, however, remains elusive, especially with respect to cognitive functions.
In order to assess the true risks that agrochemicals pose for pollinators, current risk assessment procedures need to go beyond survival analysis. Standardized, high throughput bioassays are necessary for such testing. In this study, we use such an automated, high throughput training device, the yAPIS (Nouvian and Galizia, 2019), to investigate the impact of long-term exposure to glyphosate. We tested over 400 bumblebee workers and could extract information about their locomotion, phototaxis and learning abilities from our protocol.

Colony rearing and glyphosate treatment
In total, 14 Bombus terrestris colonies contributed to our experiments. Two of these colonies (8 and 9) were produced from wild-caught queens (collecting permit A.W., 4.4.2019; Landratsamt Konstanz), the others were obtained from a commercial breeder (STB control, Aarbergen, Germany). Upon arrival, all workers were removed using forceps under red light and marked with a white dot on their thorax to later identify newly emerged bumblebees. The nest and brood were carefully split in two halves and transferred into two adjacent brood chambers separated by a wire mesh (split-colony design, Fig. 1A, for more details see (Weidenmuller et al., 2022)). Workers were then distributed evenly between the two sides of a colony. They could smell and touch each other through the mesh and the queen was moved between the two sides of their colony every 24 h, providing brood of all stages in both sides. This method allowed us to compare individuals within one colony when testing for sub-lethal effects of pesticides, accounting for the often considerable inter-colony variability (Raine et al., 2006a).
Colonies were fed daily with sugar water (sucrose 50 % w/w) and pollen (control side), or with sugar water laced with glyphosate (at the nominal concentration of 5 mg/l) and pollen (treated side). Very few studies have measured glyphosate residuals in the environment, but this concentration is within the range of what has been reported on plant materials and in bee products (Giesy et al., 2000;Thompson et al., 2014;Zioga et al., 2022;Thompson et al., 2022) and is in the middle of the range of glyphosate concentrations used in published studies on honey bees. This treatment lasted between 30 and 64 days, depending on the colony (Suppl.

Fig. 1. Methods
A. Split-colony with the brood chambers at the bottom (separated by a wire mesh), and the foraging chambers on top. The bees were fed with glyphosate-laced sugar water or pure sugar water in the petri dishes placed in the foraging chamber (A and B). The experimenter was blind as to which of the A or B solution contained the glyphosate. B. The y-APIS apparatus, with each arm illuminated by one of the three types of LEDs available. C. Training protocols for each of the 2 groups (paired or unpaired), taking the blue light as an example CS+. For half of the bees, the shocks were paired to green instead (not represented). In each group, half of the bees were trained with the sequence shown (GBBGBGGB), the other half with its opposite (BGGBGBBG, not represented). All bees were tested once for each of the 3 test configurations, and test order was balanced across bees within each group.
dissolving glyphosate (analytical standard, purity ≥98 %, Sigma Aldrich, USA. Lot: BCBW9283) in distilled water. Both sugar solutions were prepared and labelled by another lab member, such that the experimenters were blind as to which side of the colony received the glyphosate treatment. An overview of sample sizes for each colony and each treatment is given in Suppl. Table 1. Note the roughly equal sampling of individuals from both sides of each colony. Given this even distribution between treatments and because of the extremely high variability in the number of bumblebees that could be obtained from each colony, all colonies were pooled and "colony of origin" was not further included as a variable during analysis.
The colonies involved in this study are the same as the ones tested for collective thermoregulation in (Weidenmuller et al., 2022). This should have no impact on the results presented here, as none of the bumblebees participated in any learning task, visual or otherwise, before the start of our experiments. The time taken to start treatment and complete these other experiments is responsible for the variation in exposure durations between our colonies.

Bumblebees
For several weeks after the start of the experiment, newly emerged workers were removed daily and tagged with individually-numbered disks (Opalith Plättchen) before being returned to their colony (and side). Their eclosion date was noted so that their age was known when they were later trained and tested. We tested as many age marked bumblebees as we could find (57.8 % of tested bumblebees), but also included unmarked (younger) worker bumblebees.
The bumblebees participating in the experiments were removed from their colony using forceps, under red light. They were introduced inside the behavioural training and testing chamber yAPIS (see below) immediately after being caught. Because of the limited space inside the apparatus -originally designed for honeybeesthe largest bumblebees could not be tested (maximum thorax width of tested bumblebees: 5.23 mm, average: 3.89 ± 0.51 mm).
To avoid repeated capture of the same individuals, the bumblebees were not returned to their colony following testing but were frozen directly after being removed from the yAPIS. Thorax width was later measured using a digital caliper for 77.7 % of the bumblebees tested (following a miscommunication, 22.3 % of the bumblebees were sent for chemical analysis before they could be measured). Suppl. Table 1 presents a comprehensive overview of the information available for each individual bumblebee.

yAPIS apparatus
Bumblebees were individually trained in an automated y-maze consisting in 3 arms of equal length (14 cm) at 120 • from each other, the yAPIS (Fig. 1B). This apparatus has been described in detail before (Nouvian and Galizia, 2019). Briefly, the bumblebee was tracked in real time within the y-maze so that lights placed underneath the transparent floor could be switched on relative to its position. We used three types of LEDs, spanning different wavelengths: human green (λ = 520 nm), human blue (λ = 465 nm) and ultraviolet (λ = 375 nm), that we refer to as green, blue and UV for simplicity. The light intensity settings were based on our previous work with honeybees (Nouvian and Galizia, 2019). The lights were set at 64 %, 44 % and 2.4 % of the maximum intensity, corresponding to a total photon counts of 898.10 12 , 1964.10 12 and 31.10 12 quanta/cm 2 /s, for green, blue and UV respectively. The apparatus can also deliver electric shocks from a grid placed on the floor and ceiling of each arm, that were used as unconditioned aversive stimuli (trains of 200 ms, 10 V shocks delivered at 2 Hz).

Training protocols
All bumblebees were first given 5 min in the dark to habituate and explore after being placed in the apparatus. The training phase then started, and consisted of a sequence of 8 light presentations (4 times green, 4 times blue in a pseudo-random order), given for 10 s and with a 30 s inter-stimulus interval (Fig. 1C). During these presentations a single wavelength was displayed throughout the apparatus, thus the bumblebee was passively exposed. Within each treatment (control and glyphosate), the bumblebees were randomly assigned to either the paired group or the unpaired group. For the paired group, the electric shocks (unconditioned stimulus, US) were paired with one of the wavelengths (CS+), but not with the other (CS-). Half of the bumblebees were trained with green as CS+ and the other half with blue, so that any bias in preference between these stimuli could be removed by pooling the data. The unpaired group also experienced the shocks and lights, but not in close temporal association (the shocks were delivered during the inter-stimulus interval). After a 4 min resting period, all bumblebees were then tested for their light preference in 3 tests, the order of which was balanced across bumblebees, again with a 30s inter-stimulus interval. During each test, the bumblebee was given a choice between 2 wavelengths (blue vs. green, green vs. UV and blue vs. UV). The UV light was a novel stimulus to assess if the bumblebees had learned the shocked colour, the safe colour or both. The lights were switched on depending on the bumblebee's position, such that the bumblebee always started in the remaining dark arm. Note that no shocks were given during the test phase. During the 1st round of experiments (colonies 1, 2 and 3), we did not include an unpaired group and the tests lasted 20 s. This test duration was based on previous experiments with honeybees (Nouvian and Galizia, 2019), however bumblebees walked much slower (14 mm/s on average, against 30 mm/s for honeybees) so they often didn't reach the decision point before the end of the test ( Fig. 2A, panel "Td = 20 s"). This problem was partly solved by increasing the test duration to 30 s for all subsequent experiments ( Fig. 2A, panel "Td = 30 s"). Given that the results were qualitatively identical and that we always sampled each side of a colony equally to ensure colony of origin could not be a confounding factor when assessing the effect of the glyphosate treatment, we pooled all the data for later analysis.

Quantification and statistical analysis
The yAPIS records a time series of the bumblebee's position within the apparatus, as well as of any event triggered by the training protocol (lights on/off, shocks delivered). This raw data was first analyzed using custom-written Python scripts, to extract relevant information such as speed, the percentages of time that the bumblebee spent inside each of the three light conditions (dark, wavelength 1, wavelength 2) during the tests and the first light being entered (if any). Further statistical analysis was then performed in Matlab R2020a. A χ 2 test was performed to compare the percentages of bumblebees that first entered each of the two test wavelengths, against the null hypothesis that bumblebees chose randomly. Preference (unpaired group) or learning (paired groups) were further assessed by comparing the percentages of time spent within the two wavelengths of interest, using a paired Wilcoxon signed rank test. In addition, for each bumblebee we calculated a learning index LI that was defined as: with CS + and CS − the percentages of time spent into the light previously associated or not with the shocks, respectively. Thus, a learning index of 1 characterizes perfect learning (the bumblebee only visited the previously safe stimulus), 0 an equal preference between the lights and − 1 a fully incorrect choice (the bumblebee only visited the previously shocked stimulus). About 25 % of the bumblebees did not visit either of the two lights during the test, as a consequence no learning index could be calculated for them. This proportion was not affected by the treatment with glyphosate (see results). We tested whether the learning index distribution had a median of zero (no learning) using a one-sample Wilcoxon signed rank test. To assess how bumblebees responded to the electric shocks, we calculated the difference in average walking speed (Δspeed) between the 40 s of CS+ exposure (light + shocks) and the 40 s of CS-exposure (light only) in the training phase. We tested for correlations by using Spearman's rho. For representation purposes, the linear relationship between the ranks of each variable was shown by predicting the value of the ordinate variable, relative to its median and range, using the empirical cumulative density function of the abscissa variable. In all cases in which multiple comparisons were performed, the pvalues were corrected using a False Discovery Rate (FDR) procedure, to control for both type I and type II errors (Verhoeven et al., 2005).
Generalised linear models (ordered beta regression (Kubinec, 2022)) for the effects of age, shock reaction (Δspeed), glyphosate treatment and their interactions on learning index were fitted to the full dataset, accounting for observations both between (− 1,1) and at the bounds [− 1,1] of possible learning index observations. Where exact data on age were missing, values drawn from a uniform distribution between the upper and lower limits of possible ages for each individual were imputed (van Buuren, 2012), generating 100 datasets with plausible estimates of age for all individuals. Draws from models fitted to the imputed datasets were pooled to provide plausible ranges for each parameter estimate. Model comparison was performed using leave-one-out cross-validation (Vehtari et al., 2016) to identify the set of model parameters with the highest predictive performance (lowest LOO information criterion). All models were constructed and estimated in the Bayesian programming language Stan (Carpenter et al., 2017) via the brms (Bürkner, 2017) and ordbetareg (Kubinec, 2022) packages in R v4.2.2 (R Core Team, 2022). Chain convergence was assessed for each modelled parameter via trace plots and R calculation (Vehtari et al., 2021). Models converged well, producing R values in the range 0.9981-1.0093 for all parameters. The modelled cutoffs at learning indices of − 0.64 and 0.38 suggest that individuals learning below and above these levels, respectively, had a high probability of visiting only one arm for the duration of the trial, resulting in learning indices of − 1 or 1.
All statistical information can be found in the figure legends and results section. Suppl. Table 1 presents an overview of the sample sizes. Suppl. Tables 2 & 3, as well as Suppl. Fig. 2, show the details of the generalised linear modelling.

Results & discussion
We tested the effect of long-term glyphosate exposure on bumblebee locomotion, phototaxis and cognition. A total of 435 bumblebees were tested in the yAPIS, among which only 13 were excluded due to small Fig. 2. Phototactic behaviour A. Percentage of bees staying in the dark arm during at least one of the 3 tests, and hence not making a choice between the coloured lights. Glyphosate exposure did not change the proportion of bees not entering the lights, but the training protocol (paired vs unpaired) did for control bees. Fisher tests, sample sizes indicated within each bar. B. Walking speed of the bees during the test phase, for bees always entering the lights (dark grey) and for the bees staying in the dark during at least one of the 3 tests (light grey). The bees staying in the dark were much slower on average than the bees that made choices, and glyphosate exposure did not modify this pattern. 2-way ANOVA, same sample sizes as in panel A. C. In the unpaired groups, percentage of time (mean ± s.e.m.) that the bees spent within each stimulus during the 3 test situations (blue vs green, green vs UV, blue vs UV). The bees always started in the dark arm, and for each configuration, only bees that entered a coloured light are considered (% no choice indicated at the bottom). Control bees prefer the UV rather than the blue light, but not glyphosate-exposed bees. Wilcoxon signed rank tests comparing the two lights, corrected with FDR, n c = n g = 55 bees. Td: test duration, C: control bees, G: glyphosate-treated bees, *: p < 0.05, ***: p < 0.001, #: p = 0.056, n.s.: not significant. technical issues, resulting in a sample size of 422 individuals. All bumblebees went through a circa 19 min protocol consisting of a habituation phase, a training (or pseudo-training) phase and a testing phase.

Glyphosate does not impair locomotion or phototaxis during the testing phase
We detected a slight decrease in walking speed in bumblebees exposed to glyphosate, but only during the initial habituation phase (Suppl. Fig. 1; Wilcoxon signed rank test, habituation: T = 47,711, z = 2.124, p = 0.034). During the training and testing phases, the walking speeds of glyphosate-exposed bumblebees were similar to those of control bumblebees (training: T = 45,919, z = 0.693, p = 0.488; test: T = 45,809, z = 0.606, p = 0.545), thus locomotor effects could not explain differences in behaviour during the tests.
Some bumblebees did not choose between the two coloured lights during the tests, but instead remained in the dark arm where they started. The proportion of bumblebees that stayed in the dark during at least one of the three tests did not vary between glyphosate-exposed and control bumblebees ( Fig. 2A, Fisher exact test, Td = 20 s: p = 0.679; Td = 30 s, unpaired group: p = 0.094, paired group: p = 0.623). Thus, the phototactic drive was not generally impaired in glyphosate-exposed bumblebees. The bumblebees not making a choice were significantly slower than those that entered the lights irrespective of glyphosate exposure, suggesting that they were in poor condition (Fig. 2B, 2-way ANOVA, choice vs no choice: F(1,418) = 179.63, p < 0.001; interaction term: F(1,418) = 1.67, p = 0.197; main effect of glyphosate treatment: F(1,418) = 0.56, p = 0.453). Note that we purposedly did not set an exclusion criterion in our dataset in case glyphosate exposure had subtle effects that we did not account for. However, since learning performance could not be assessed when the bumblebees did not make a choice, these bumblebees were excluded from most analysis.
Interestingly, control bumblebeeswhich had not been exposed to glyphosatefrom the unpaired group stayed in the dark arm less frequently than those from the paired group ( Fig. 2A, Fisher exact test, odds ratio 2.90, confidence interval 1.19-7.07, p = 0.020). This may be a result of the shocks happening in the dark for the unpaired group (thus increasing the bumblebees' motivation to leave this stimulus), or of the shocks happening in the lights for the paired group, or both. This behaviour was not observed in glyphosate-exposed bumblebees (odds ratio 1.33, confidence interval 0.65-2.73, p = 0.478), thus providing a first evidence of learning deficit in these bumblebees.

Glyphosate changes the preference between blue and UV lights
Before evaluating the effect of glyphosate exposure on aversive learning, we first checked for a possible difference in the innate response to the light stimuli used as CS between glyphosate-exposed and control bumblebees. To do so, we analyzed the behaviour of the bumblebees subjected to the unpaired protocol, as these bumblebees had no opportunity to form an associative memory between the shocks and a specific wavelength.
We found that control bumblebees had no preference between blue and green lights ( Fig. 2C; Wilcoxon signed rank test corrected with FDR, T = 743, z = 1.298, p = 0.194), and neither did glyphosate-treated bumblebees (T = 652, z = 1.518, p = 0.193). Control bumblebees had a tendency to prefer the UV light over the green one, albeit this was only close to significance after FDR correction (T = 461, z = − 2.076, p = 0.057). Glyphosate-exposed bumblebees had a similar preference (T = 320, z = − 2.409, p = 0.048). Given that our analysis is slightly underpowered (0.748 for the control group and 0.620 for the glyphosate group), and that the effect sizes have similar magnitudes (r = 0.280 for the control group, and r = 0.325 for the glyphosate group), this likely doesn't reflect a true difference between groups. Note also that the light intensities were calibrated previously for equal preference in honeybees (Nouvian and Galizia, 2019), hence these results may indicate that bumblebees are more sensitive or more attracted to UV light. Alternatively, previous exposure to blue and green (during the unpaired protocol) may have decreased the phototactic attraction towards these lights in comparison with the novel UV stimulus, as observed in honeybees (Nouvian and Galizia, 2020).
While control bumblebees also preferred UV to blue (T = 361, z = − 2.502, p = 0.037), glyphosate-treated bees did not exhibit such preference (T = 457, z = − 0.683, p = 0.495). That glyphosate could impact bumblebee vision is supported by the results of a recent study, demonstrating that bumblebees receiving a single dose of a glyphosate-based herbicide were not able to distinguish between similar colours in a learning task (Helander et al., 2023). Modifications of the bumblebees' perception of UV light is particularly worrying because this range of the spectrum is implicated in the detection of polarized light (Meyer-Rochow, 1981). This ability is important for the bees to navigate via detection of the "sky compass", especially in dim light conditions (Patel et al., 2022;Rossel and Wehner, 1986;Zeil et al., 2014;Jacobs-Jessen, 1959;Wellington, 1974). Acute glyphosate exposure negatively affected navigation abilities in honeybees, an effect that was attributed to impaired foraging motivation and appetitive learning abilities (Balbuena et al., 2015). Our results raise the question of whether an impairment in the detection of the sky compass could also have played a role. Furthermore, UV reflectance and UV patterns are important parameters of flower coloration, strongly influencing the foraging efficiency and flower choices of bees (Chittka et al., 1994;Lunau et al., 1996;Papiorek et al., 2016;Foster et al., 2014). To sum up, even a slight shift in UV sensitivity could have broad implications for these pollinators.
As a more direct consequence for our study, both the biased preference for UV in control bumblebees and its disruption after glyphosate exposure made it difficult to interpret the blue/UV and green/UV tests when evaluating learning performance. Hence, in the following sections we consider only the results from the stable (and most relevant) blue/ green test.

Glyphosate impairs aversive learning
Control bumblebees decreased the time spent in the CS+ relative to the CS-( Fig. 3A; Wilcoxon signed rank test, T = 2591, z = − 2.844, p = 0.004), and exhibited a bias in their first choice in favour of the CS-( Fig. 3B; χ 2 test against a balanced choice between the two stimuli, χ 2 (1158) = 4.037, p = 0.045). We calculated a learning index LI by dividing the difference in percentage of time spent within each CS by the total percentage of time in these lights. A learning index of 1 characterizes perfect learning (the bumblebee only visited the CS-), 0 an equal preference between the lights and − 1 a fully incorrect choice (the bumblebee only visited the CS+). The median learning index of control bumblebees was significantly above 0 (Fig. 3C, one-sample Wilcoxon signed rank test, T = 5014, z = 3.491, p < 0.001).
By contrast, glyphosate-exposed bumblebees did not learn the association: they did not avoid the CS+, neither by reducing the time spent in this light ( Fig. 3A; Wilcoxon signed rank test, T = 3127, z = − 0.581, p = 0.562) nor when making their first choice ( Fig. 3B; χ 2 test against a balanced choice between the two stimuli, χ 2 (1154) = 0.109, p = 0.741).
The distribution of learning indexes was centred around 0 in this group, again confirming the absence of learning (Fig. 3C, one-sample Wilcoxon signed rank test, T = 3806, z = 1.335, p = 0.182).
The ability to associate a noxious stimulus with particular cues is a fundamental pre-requisite for survival. Through this adaptive behaviour, animals have a better chance of avoiding encounters with poisons, predators and parasites. Hence the learning impairment that we demonstrate here, caused by exposure to glyphosate, could substantially increase the mortality rate of foragers. Such depletion of the workforce would have an obvious impact on colony success, although this remains to be confirmed experimentally. A small impairment of learning abilities after acute or chronic glyphosate exposure had already been reported in honeybees (Herbert et al., 2014;Luo et al., 2021;Mengoni Gonalons and Farina, 2018). These studies used an appetitive, olfactory task (sugar paired to an odour), and also reported a decrease in sucrose responsiveness after glyphosate exposure, which likely contributed to the altered learning performance. Free-flying bumblebees trained after acute exposure to glyphosate performed similarly to controls in a 10odour foraging task (Helander et al., 2023). The same study reports that visual learning was not affected when the discrimination task was restricted to two easily differentiable colours, but a 10-colour task revealed that fine-colour discrimination was impaired in treated bumblebees (Helander et al., 2023). Thus, the impacts of glyphosate on learning seem to arise mainly from modifications in the sensory systems of bees, such as taste and vision. A single study reported a purely cognitive effect, namely a decay in memory retention, in honeybees treated with glyphosate (Hernández et al., 2021). Did glyphosate exposure impair sensory perception in our experiment? On the CS side, the wavelengths we used for our task (465 and 520 nm) should be very easy to discriminate for bumblebees (Peitsch et al., 1992;Dyer et al., 2008). It seems unlikely that colour processing could be affected to such an extent, but we cannot rule it out. On the US side, we compared the walking speeds of the bumblebees in the CS+ versus the CS-during the training phase (i.e. when they experienced a coloured light with the shocks versus only a coloured light), which we called Δspeed. We found that bumblebees walked faster when they received the shocks, both in the control group ( Fig. 3D; one-sided Wilcoxon signed rank test, T = 7289, z = 1.750, p = 0.040) and the glyphosate-treated group (T = 7375.5, z = 2.539, p = 0.006). Furthermore, this acceleration was correlated with learning performance -estimated by the learning indexin control bumblebees (Spearman's ρ, ρ = 0.288, p = 0.001) but not in glyphosate-treated bumblebees (ρ = 0.079, p = 0.401). Generalised linear modelling (Fig. 3E, see Methods) confirmed that glyphosate treatment had a negative effect on the relationship between shock reaction and learning index (odds ratio 1.27 times reduction, 95 % credible interval 0.94-1.72), effectively negating the higher learning indices . The bees always started in the dark arm. The percentage of bees not represented because they did not make a choice is indicated at the bottom. Control bees spent more time in the CS-than the CS+, but not glyphosate-exposed bees. Wilcoxon signed rank test. B. Percentage of bees first entering each stimulus during the test (CS+ in red, CS-in yellow). The first choice of control bees was biased in favour of the CS-, but not the first choice of glyphosate-exposed bees. Chi square test against random choice. C. Distribution of learning index values in each treatment group. Values above 0 indicate learning (the bees spent more time in the CS-than the CS+). One-sample Wilcoxon signed rank test. D. Distribution of the difference in walking speed (Δspeed) between CS+ (shocks) and CS-(no shock) during the training phase. One-sided Wilcoxon signed rank test. The bees react to the electric shocks by accelerating. n c = 159 bees; n g = 156 bees. E. Correlation between Δspeed (reaction to shocks) and learning index. In control bees, sensitivity to shock is correlated with an increase in learning performance, but this is not the case for glyphosate-exposed bees. C: control bees, G: glyphosate-treated bees, *: p < 0.05, **: p < 0.01, ***: p < 0.001, n.s.: not significant. Bees making a choice (all panels except D): n c = 121 bees; n g = 115 bees. for greater sensitivities in control bumblebees (odds ratio 1.31 times per Δspeed, 95 % credible interval 1.08-1.60). Taken together, these results demonstrate that glyphosate exposure did not impair pain perception in our experiment. Rather, this herbicide likely affected cognitive processes.

Glyphosate prevents an age-correlated increase in performance
Learning abilities may vary between individuals depending on a number of factors. Two commonly tested factors in bumblebees are age and size, although their impact is not consistent across studies (Raine et al., 2006b;Riveros and Gronenberg, 2009;Merling et al., 2020;Frasnelli et al., 2021). The age and size distributions of our bumblebees were similar in the control-and glyphosate-treated groups (Mann-Whitney U tests; age: Fig. 4A, U = 14,514, z = − 0.674, p = 0.500; size: Suppl. Fig. 3A, U = 27,155, z = 0.791, p = 0.429). According to our data, learning performance was not correlated with size (Suppl. Fig. 3C; Spearman's ρ; control: ρ = 0.086, p = 0.427; glyphosate: ρ = − 0.160, p = 0.139) -although our size range was limited by space-constraints within the yAPIS system, which may have impacted our results.
Indeed, the largest bumblebees (thorax width > 5 mm) did not fit inside the apparatus, which is intentionally limited in height to prevent flying (see Methods).
On the other hand, learning performance did correlate with age. Only a subset of bumblebees could be included in this analysis, namely age-marked bumblebees which entered at least one of the lit arms during the test (61 bumblebees in the control group, 66 in the glyphosateexposed group), nonetheless the overall inhibitory effect of glyphosate on learning was evident also in this restricted dataset ( Fig. 4B; onesample Wilcoxon signed rank test; control: T = 1215, z = 1.986, p = 0.047; glyphosate: T = 1231, z = 0.817, p = 0.414). In the control group, older bumblebees performed better than their younger siblings ( Fig. 4C; Spearman's ρ, ρ = 0.347, p = 0.006). This age effect completely disappeared in the glyphosate-treated group ( Fig. 4C; ρ = 0.113, p = 0.364). However, after accounting for the effect of Δspeed, generalised linear modelling did not find age to be an independent predictor of learning performance (ELPD difference − 1.9, s.e. 2.6). We hypothesize that the relationship between age and learning performance may be indirect, thus explaining why age is a correlated but not a causal factor.

Fig. 4.
Glyphosate prevents an agecorrelated increase in performance A. Age distribution within each group of bees. There is no difference between the control and glyphosate-exposed groups. Wilcoxon signed rank test. n c = 122 bees; n g = 121 bees. B. Distribution of learning index values in each treatment group, for the subset of bees considered. One-sample Wilcoxon signed rank test. n c = 61 bees; n g = 66 bees. C. Correlation between age and learning index. Control bees (in black) perform better as they age but not glyphosate-exposed bees (in grey). Spearman rho. n c = 61 bees; n g = 66 bees. C: control bees, G: glyphosate-treated bees, *: p < 0.05, **: p < 0.01, n.s.: not significant.

Conclusions
In this study, we demonstrate that chronic exposure to glyphosate reduces the ability of bumblebees to associate an aversive stimulus with a visual cue during a differential learning task. This inability to learn about danger cues may put bumblebees at a higher risk of falling prey to predators and parasites while foraging. This study thus adds to the growing body of work concerned with the negative impact of glyphosate on bee health (Weidenmuller et al., 2022) and especially on bee cognition (Vazquez et al., 2020;Farina et al., 2019). Glyphosate exposure impacted bumblebee physiology and nervous system function in several ways, from sensory perception to cognition. This could result from a broad disruption of brain maturation or function. Further research will be needed to elucidate glyphosate's mechanism of action on insect cognition, as well as to evaluate if this effect is temporary or permanent.
Furthermore, we present a proof of concept that the yAPIS, a fully automated, high throughput assay, can be used for testing the sublethal effect of chemicals under controlled laboratory conditions. This assay was developed for honeybees but could easily be adapted to bumblebees, suggesting that any species willing to walk inside the apparatus could be tested in the same way. This could include wild and solitary bees which have been largely understudied so far. The yAPIS protocol also has the advantages of not requiring restraining, nor a lengthy pretraining (e.g., to get the bees to fly in a foraging arena), and measures multiple parameters simultaneously.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability
Data will be made available on request.