Cognitive scores of bees exposed to various environmental stressors

Environmental stressors have sublethal consequences on animals, often affecting the mean of phenotypic traits in populations. However, effects on inter-individual variability are poorly understood. Since phenotypic variability is the basis for adaptation, any change due to stressors may have important implications for population resilience. 

Here we explored this possibility in bees by analysing raw datasets from 23 studies (5,618 bees) in which individuals were first exposed to stressors and then tested for cognitive tasks.

While all types of stressors decreased the mean cognitive performance of bees, they increased cognitive variability. Focusing on 14 pesticide studies, we found that the mode of exposure to stressors and the dose were critical. Mean cognitive performance was more affected by a chronic exposure than by an acute exposure. Yet, cognitive variability increased with increasing doses following both exposure durations.

Policy implications: Current guidelines for the authorization of plant protection products on the European market prioritize acute over chronic toxicity assessments on non-target organisms. By overlooking the consequences of a chronic exposure, regulatory authorities may register new products or doses that are harmful to bee populations. Our findings call for more research on stress-induced phenotypic variation and its incorporation to policy guidelines to help identify levels and modes of exposure animals can cope with.

Methods
Search and selection of datasets

The search for datasets in scientific publications falling within the scope of our research question was performed in July 2020 using the PubMed database. The words used for the search were (“Stressor” OR “Pesticide” OR “Parasite”) AND (“Cognition” OR “Learning”) AND (“Bees”). This search was not restricted to any section of the manuscripts and automatically extended to similar terms intended under the MeSH hierarchy of the database. A total of 240 studies were found, of which 18 met our inclusion criteria regarding the cognitive task and the type of stressor (see below). The search terms under which each study was found are available in the Supplementary Table 1. Five datasets belonging to the authors of this study were also included as they filled the inclusion criteria. These studies measured the impact of stressors on the cognitive performance of bees. The list of the 23 selected studies is available in Table 1.

Cognitive tasks: We focused on cognitive data from bees exposed to stressors during their adult life. The effect of stressors on larvae could not be analysed due to the lack of data available (two studies). In all the selected studies, cognitive performance was assessed using associative learning paradigms testing the ability of bees to associate an olfactory or/and a visual stimulus with an appetitive or aversive reinforcement (Giurfa, 2007). Olfactory learning was tested in 18 out of the 23 studies. These studies used learning protocols based on the appetitive conditioning of the proboscis extension response (PER; 16 studies) or the aversive conditioning of the sting extension response (SER; 2 studies). Either response was conditioned by presenting bees a conditioned stimulus (an odour) reinforced with an unconditioned stimulus (sucrose solution or electric shock), for 3-15 trials in appetitive assays and 5-6 trials in aversive assays. Trainings included absolute learning (the odour is reinforced) and differential learning (an odour is reinforced; the other is not). Visual learning was tested in 5 out of the 23 studies. These studies used appetitive conditioning protocols in a Y-maze or on artificial flowers (i.e. feeders), or aversive conditioning protocols with electric shocks. One of these studies applied a multimodal appetitive conditioning combining both odour and colour cues to be learnt by bees in an array of artificial flowers (Muth et al. 2019). Here again bees were tested for differential learning.

Stressors: Stressor types covered different pesticides, parasites, predator odours, alarm pheromones, and heavy metal pollutants. Experiments performed with pesticides whose median lethal dose (LD50; i.e. dose that kills 50% of the population) could not be identified in the literature were excluded from our final selection. 

Exposure duration: In all the studies, stressors were applied before the cognitive tests, except in one study in which it was used as the conditioning stimulus to be learned (i.e. alarm and predator pheromones (Wang et al. 2016)). We categorised the duration of exposure using the common dichotomy between acute and chronic exposures. An acute exposure was characterized by a single administration of the pesticide to each individual bee. When bees were exposed to the pesticide more than once, either as a substance present in their environment or as a food directly offered to each individual, the exposure type was considered chronic.

Bees: The bee species studied in the selected publications were the honey bees Apis cerana and Apis mellifera, and the bumblebees Bombus impatiens and Bombus terrestris. These species were not selected purposefully, but rather emerged as the species most represented in our dataset from the refinement obtained with other inclusion criteria. We considered bee genus (Apis or Bombus) for the analyses. 

Table 1: Summary of the 23 studies used. 

Stressor

Bee genus

Exposure type

Reference

Pesticide

Apis

Acute

(Ludicke and Nieh, 2020)

Pesticide

Apis

Acute

(Hesselbach and Scheiner, 2018)

Pesticide

Apis 

Acute

(Urlacher et al., 2016)

Pesticide

Apis

Acute 

(Tan et al., 2015)

Pesticide

Apis

Chronic

(Mustard et al., 2020)

Pesticide

Apis

Chronic

(Tan et al., 2017)

Pesticide

Apis, Bombus

Acute

(Siviter et al., 2019)

Pesticide

Bombus

Acute

(Muth et al., 2019)

Pesticide

Bombus

Acute, chronic

(Stanley, Smith and Raine, 2015)

Pesticide

Bombus

Chronic

(Smith et al., 2020)

Pesticide

Bombus

Chronic

(Lämsä et al., 2018)

Pesticide

Bombus

Chronic

(Phelps et al., 2018)

Pesticide, coexposure

Apis

Chronic

(Colin, Plath, et al., 2020)

Parasite

Bombus

Acute

(Gomez-Moracho et al., 2021)

Parasite

Bombus

Acute

(Martin, Fountain and Brown, 2018)

Pollution

Apis

Acute

(Monchanin et al. unpublished)

Pollution

Apis

Acute

(Monchanin, Drujont, et al., 2021)

Pollution

Apis

Chronic

(Monchanin, Blanc-brude, et al., 2021)

Other

Apis

Acute

(Wang et al., 2016)

Other

Apis

Acute

(Shepherd et al., 2018)

Other

Apis

Chronic

(Shepherd et al., 2019)

Coexposure

Apis, Bombus

Acute/Chronic

(Piiroinen and Goulson, 2016)

Coexposure

Bombus

Acute/Chronic

(Piiroinen et al., 2016)

Dataset organisation and normalisation of variables 

All but three raw datasets were available online with the published material. Those three datasets were kindly provided by their authors, i.e. Dara Stanley and Ken Tan. The raw data were downloaded and saved as .csv files. A new dataset was created, which combined information on the species, the cognitive task studied, the type of stressor, the type of exposure (acute/chronic), and, in the case of pesticide studies, the dose (µg/bee) or concentration (ppb). The dose (acute exposure) and concentration (chronic exposure) were normalized as the percentage of the LD50. When learning performance was measured as a binary response (e.g. success vs. failure) across multiple trials, the raw data were used to calculate a learning score for each individual corresponding to the number of successful trials. This was required because the variance in binary variables can be mathematically predicted from the mean and sample size and does not reflect biological variance (Supplementary Fig. 1). Each study provided individual cognitive scores for at least one experimental treatment and control group. There was a total of 73 experimental treatments across the 23 studies.

To compare the mean cognitive performance and the cognitive variability across studies, we used a standardized method for the meta-analysis of variation (Nakagawa et al., 2015; Senior, Viechtbauer and Nakagawa, 2020). This method controls for the mean – variance linear relationship that may exist in a dataset by using unbiased effect size statistics of the mean and variability, i.e. the natural logarithm of the ratio between the means (lnRR) and the natural logarithm of the ratio between the coefficients of variation (lnCVR) of treated and control groups, respectively. Changes in lnCVR are not an indirect consequence of changes in lnRR, as would have been the case had we analysed the variance and the mean, but they rather reflect changes in variability per se.

The two pre-requisites for this method are (i) to use log scale data and (ii) to observe a mean-variance linear relationship. Studies for which negative cognitive scores were present were transformed to log-scale data by adding the minimum score to all individuals. The mean and standard deviation of the cognitive scores, as well as sample sizes, were calculated for each experimental treatment and control group. A linear relationship and positive correlation were found between the log sample mean and standard deviation in our dataset (Supplementary Fig. 2). All pre-requisites being met, we then calculated the lnRR and lnCVR for each experimental treatment and control group (i.e. 73 effect sizes) as well as their sampling (error) variance using equations corrected for the sample size described in (Senior, Viechtbauer and Nakagawa, 2020). Individual bees in control and treated groups in all study designs were considered independent.

Data analyses

All analyses were conducted in R Studio v.1.2.5033 (RStudio Team 2015). The package metafor (Viechtbauer, 2010)was used to compute multilevel meta-analytic models (MLMA), multilevel meta-regression models (MLMR) and to generate forest plots. The study and experiment identifier were always set as nested random effects. For MLMR and depending on the question, the type of stressor, the genus, the type of task, the exposure duration and/or the percentage of LD50 were defined as fixed effects. The restricted maximum likelihood approach (REML) was used to estimate the parameters of the meta- analysis models. Forest plots were used to show the effect size estimates lnRR and lnCVR and their 95% confidence interval (CI). In this approach, positive effect sizes reflect higher means (lnRR) or coefficients of variation (lnCVR) in the treated group compared to controls. Negative effect sizes reflect lower means (lnRR) or coefficients of variation (lnCVR) in the treated group compared to controls. Effects are significant when the 95%CI do not span across the zero line.