Developing and Evaluating a New Method and Apparatus for Examining Bumble Bee Foraging Behavior

A key question in pollination biology is that of how pollinators identify and choose foraging patches. Several approaches have been employed for this, including field studies and large greenhouse flight chambers. Most methods used to date are limited, though, by reliance on a combination of artificial flowers, large spatial scales, or lack of spatially distinct floral patches. To address this issue, we designed and tested a y-maze flight arena and tested it using the bumblebee Bombus impatiens and canola plants. Our results indicate that the system is not biased by environmental conditions, or by an innate “handedness” of Bombus impatiens. We found that bees made all the expected patch choices when presented with soil, plants without flowers, or plants with flowers. This new method is important and useful as it allows researchers to ask questions of both plant health and insect behavior and the chamber system is modular allowing for simple changes to the setup to focus on different questions. • Y-maze flight arena was designed to evaluate foraging behavior on intact plants. • No evidence of side preference in individual bees. • Behaviors observed in the y-maze appear to correspond to behaviors observed in other settings.


Introduction
A principal issue in understanding pollinator foraging ecology is determining how pollinators locate suitable patches of forage. Is a patch located by randomly sampling the area? Following those around you? Spatial memory? Or, by a cue such as a smell or color? Approximately 75% of plants globally are insect pollinated, and as such, insect pollinators play an important role in ecosystem functioning in both natural and agricultural environments (Losey and Vaughan 2006;National Research Council 2007;James and Pitts-Singer 2008;Potts et al. 2010). Plants often attract pollinators with floral displays across a variety of modalities composed of visual, olfactory, tactile, and taste stimuli (Raguso 2004). The composition and quality of these cues are directly related to plant health, which is mediated by a variety of biotic and abiotic factors (Strauss et al. 2004;Adler et al. 2006;Wink 2017). Understanding how plants communicate with insect pollinators and how insect pollinators interpret plant cues is fundamental to understanding ecosystem function and stability.
A significant body of research evaluates how foragers locate rewarding flowers and distinguish rewarding flowers from unrewarding ones (Leonard et al. 2011;Kaczorowski et al. 2012;Rivest et al. 2017). This work has elucidated many complex behaviors that highlight the importance of plant signaling uniformity in a bee's ability to locate suitable forage (reviewed in Chittka 2017). Previous work assessing pollinator behavior on larger spatial scales has used large greenhouse spaces (Lefebvre et al. 2007) or in-field assessments (Galen and Green 1987;Cartar and Real 1997). Both are similarly limited by the ability to track an individual bee over the course of a foraging bout. Recent work has been conducted with radio tracking of bees in the field (Hagen et al. 2011;Woodgate et al. 2016) but linking plant traits with choice is difficult as you must account for variability across the environment on any given day and between days (Comba 1999;Chittka et al. 2013;Muth et al. 2017). Further, most studies at the patch level focus on plant species as a determining factor, while controlled studies often focus on individual flowers or artificial flowers (Shaw et al. 2020). Collectively, these factors limit our knowledge of how a plant's physiological state influences pollinator foraging behavior, particularly at the patch level (Jacobsen and Raguso 2018). This is an important obstacle to evaluating and describing the roles plants play in patch level choices and is also pivotal for describing factors that drive pollinator composition within ecosystems.
The ability to fully understand behavior largely depends on our ability to narrow the focus of observations without losing the forest to the trees. That is, to reduce the number of potentially confounding factors, and to break behaviors down into more simply observed and quantified patterns of choices. Thus, to examine foraging choices and behaviors in many animals, including numerous insect species, researchers often employ choice assays which reduce the complexity associated with experimental design. Choice assays allow the reduction of experimental designs to a system of simple choices to assess the effects of specific stimuli on behavior (Dunbar 1995). Choice assays are ideal as they present a subject with an array of conditions and allow the subject to make real world navigation decisions with the stimuli presented (Cartar and Dill 1990;Abada et al. 2009;Jaworski et al. 2015). Notable experimental setups such as flight arenas and Free-Moving Proboscis Extension Response (FMPER) have traditionally been used for assessing pollinator choices and foraging behaviors (Muth et al. 2018;Lawson et al. 2018;Avarguès-Weber et al. 2018). For most other animals, including terrestrial insects, we employ simple mazes such as T or Y mazes for measuring choice preferences (Mammals: Conrad et al. 1996;Pajor et al. 2003, Gastropods: Manríquez et al. 2014, Cephalopods: Lee 1992, Sharks: Zahuranec 1983, Insects: Prager et al. 2014). These systems present binary choices to the organism, allowing for in-depth examination of specific choice behaviors (i.e., will it lay eggs on X over Y, or will it feed on Y over Z). Few studies, however, have used simple maze designs with whole intact plants when examining pollinators (summarized in Table 1). Instead, the focus has been on cue learning and its relation to dynamic foraging choices (Lawson et al. 2018;Avarguès-Weber et al. 2018), or response to flower-like structures (pseudo-flowers) produced by plant pathogens (Schiestl et al. 2006). Most of the behavioral studies on live intact plants rely on large group observational data, mainly because accounting for the dynamic nature of the field is difficult to do within a laboratory or greenhouse setting (Cibula and Zimmerman 1987;Cartar and Dill 1990;Lefebvre et al. 2007;Klatt et al. 2013).
To address some of the limitations associated with existing methods for evaluating complex bumble bee foraging behavior, we designed an apparatus for examining bee foraging that allowed for bees to make patch level choices on intact plants. The apparatus and associated analytical methods are designed as a method for evaluating patch level choices within a controlled (greenhouse) environment using a modified Y-maze system. The purpose of this design is to provide a modular and flexible system that can be used by researchers to assess how plant health, or the Artificial flower species composition of flower patches might impact patch level foraging choices. To assess the viability of this maze system for these types of patch level pollinator choice assays, we considered three main questions. First, we examined whether environmental conditions within the system played a significant role in pollinator choice. Second, we tested whether individuals exhibited an inherent directional bias that may limit the Y-maze's applicability. Finally, we tested that the system was suitable and effective for evaluating bee preferences and movement.

Cage Design
All experiments were performed in a custom designed and constructed Y-maze chamber (Figs. 1 and 2). The chamber design was selected based on the successful models from both Dyer et al. (2008) and Jaworski et al. (2015) and failed attempts at using both "U" and "T" shaped maze designs. The success of the Y-maze is based on the simple maneuvering and spatial needs for a bee to successfully move through the maze (Conrad et al. 1996). The cage was constructed using untreated spruce lumber (2"x4"), plywood, and plexiglass. Rear walls of both treatment arms were made of high-density fiberboard (from here on, referred to as hardboard or HDF) and were painted white. A single vent with a 10.16 cm opening made of PVC fitting was placed at  Fig. 2 Diagram of the Y-Maze. The hive was kept in the hive chamber (A) and remained there. Individuals were released from the hive chamber after being isolated from the hive. Right and Left were conferred in reference to forager's left and right when exiting the hive. Each terminal treatment arm is approximately 0.898m3 and constructed of two plexiglass sides, a plexiglass top, a plywood side and bottom, and an HDF rear wall. The hive arm and neutral zone were constructed using plywood sides and bottom with a plexiglass top. The hive chamber measured ~ 0.177m3 the rear of each treatment arm (Fig. 1A), to regulate the internal temperature of the chamber. These vents can be used for regulation of airflow, but in this experiment were left open to assess the system at its most basic level. Each terminal treatment arm was approximately 0.898 m 3 and constructed of two UV transparent plexiglass sides, a UV transparent plexiglass top, a plywood side and bottom, and an HDF rear wall. The hive arm and neutral zone (Fig. 2) were constructed using plywood sides and bottom with a plexiglass top. The hive arm measured ~ 0.177 m 3 . The Y-maze was equipped with iButton probes (Thermochron, Baulkham, Australia) (Fig. 1), to measure temperature and humidity. The system overall was housed in a temperature-controlled greenhouse and was oriented so that sunlight and greenhouse lights illuminated both treatment arms with approximately equal intensity. Light intensity was measured using a Flower Power Sensor (Parrot Drones SAS, Paris, France). These measurements were taken to make sure the micro-climate across the chambers was not a factor influencing choice.
For clarity, the following terms are used throughout to refer to the components and segments of the Y-maze: "Hive arm" refers to the arm of the Y-maze that houses the hive. "Neutral zone" describes the space between the hive arm and the decision line for either terminal arm (left or right). "Right and left with respect to arms of the Y-maze" refers to the arm direction oriented for a bee leaving the hive arm. "Y-maze" refers to the entire apparatus. Collectively, the right and left arms are known as the "terminal arms" (Fig. 2).

Bees
All experiments were performed using the common eastern bumblebee (Hymenoptera, Apidae, Bombus impatiens) and canola (Brassica napus, details below). Bombus impatiens is a commercially available pollinator and has been documented in the prairie region of Canada (Palmier and Sheffield 2019) which makes it a potential pollinator of canola. Colonies of B. impatiens were commercially sourced from BioBest Canada (Ontario, Canada) and were maintained, trained, and observed within the Y-maze. The chamber was maintained at 26-30 o C and 40-60% relative humidity and was monitored by iButton probes (Fig. 2). Daily photoperiod within the chamber was 16:8 h Light: Dark. Bees had unlimited access to a commercial sugar solution (BioGluc, Biobest Canada) within the hive but were not provided pollen resources unless foraging on flowers. Nectar and pollen resources were offered Ad libitum between observation and training days and were only restricted 12 h prior to observation sessions.

Plants
All training and experiments were performed on canola (Brassica napus, cv. AC Excel). AC Excel is an older variety (Rakow 1993) and can be acquired without seed treatments, removing potential confounding factors that might limit foraging such as insecticides (e.g., neonicotinoids) (Whitehorn et al. 2012). Plants were grown in greenhouse conditions without the use of systemic pesticides. Plants that displayed signs of feeding damage by pest insects or had pest insects on them were disposed of, and all remaining plants in the cohort were treated with foliar applied non-systemic insecticidal soap (Safer Brand, Woodstream Corporation, USA). Plants were grown in cohorts of 60-80 plants for 5-8 weeks. Following a 7-12-day period after flowering, flowering plants were used for training and experimental observations. Nonflowering plants were grown until bud production, but prior to flowering (around 4-6 weeks). New plants were introduced in the Y-maze for the observation period and were replaced after each observation or if they had not been used in for at least 24 h.

Training
Prior to their use in assays, bees were trained to maneuver the Y-maze by placing twelve flowering canola plants in each terminal treatment arm of the Y-maze and providing the bees free access to explore the Y-maze chamber. This was accomplished by placing an entire hive within the Y-maze with the entrance and exit open such that they could freely forage. Individuals were not able to be marked, and as such we considered training at a colony level, by observing approximately ½ of the reported colony members foraging at any given time. During the training period, bees were able to freely fly, forage and return to the hive. Training took place the day before each observation day, for a total of 12 h. . Training is important as it helps the individuals spatially localize and orient themselves, their hive, and potential forage patches (Woodgate et al. 2016).

Choice Bioassays
To evaluate the Y-maze apparatus, we conducted a series of assays using simple trivial conditions with obvious expected outcomes. These were intended to answer two questions (1) Do the bees display and inherent handedness (side preference) in their choices within the arena and (2) do bees display the choice of chamber arm that would be expected given a particular treatment condition. To do this, trained bees were used to conduct choice-bioassays within in the Y-maze. A single worker bee was released into the Y-maze from the hive arm ( Fig. 2) and allowed to visit treatment arms ad-libitum for 30 min, or until the individual returned to the hive or remained inactive for 5 min. During the choice-bioassays, the individual bee was continuously observed, and its behaviors were recorded using BORIS behavioral software (Friard and Gamba 2016). The chamber was cleaned using a 70% ethanol solution between bioassays, to remove any potential chemical residue from previous foragers (Pearce et al. 2017).
Nine treatment combinations were examined (Table 2)  Ten bees were observed per treatment per day, with each treatment having both a morning (0800-1200) and afternoon (1300-1600) observation session, for a total of 20 bees per treatment. Following the observation period, bees were euthanized using liquid nitrogen and stored in an ultracold freezer at -80°C. A total of 180 B. impatiens workers from five hives were used in these experiments. Each hive was randomly assigned to only two treatment combinations to maintain hive health and due to constraints associated with plant viability.

Behavior Coding
All behaviors were coded and recorded in sequence using the BORIS software program (Friard and Gamba 2016). The following behaviors were recorded: the bee's location within the Y-maze (right terminal arm, left terminal arm, neutral zone, hive arm; Fig. 2) and the number of landings on plants or flowers. Right and left were designated based on forager orientation when leaving the hive arm (Fig. 2). No attempt was made to determine whether the bee successfully foraged or not. We defined bees as having visited a plant if their legs contacted it. Location choices were defined as individuals crossing the decision line. This line marked the boundary between the "neutral zone" and either the right or left terminal arm. It is also important to note that within the ethogram, entering the hive arm was not assessed as making a hive choice. Hive choice was only assessed if the individual attempted to enter the hive. This distinction was made because of the use of the hive as

Analysis of Bumble Bee Positional Preference and Foraging Strategy
All analyses were performed in R Studio V 3.6.0. Individual forager preference in the Y-maze was analyzed as an individual's first choice, their time to make a choice (latent time), the number of flower/ plant interactions, their time spent in each chamber, and foraging strategy (sequential movement through the Y-maze). Foragers were expected to alter preference depending on the position of plants and plant status (Raine and Chittka 2008). Mixed effects models were initially fit to data using the lme4 package (Bates et al. 2015), with subject, replication, and time (morning or afternoon) defined as random factors, and were assessed using environmental conditions (temperature, light and humidity) as fixed variables in conjunction with treatment. Previous work has used hive as a random blocking effect, but hive was not considered here when comparing across all treatments due to constraints of the experimental set-up.
To account for this, treatment combinations were randomly assigned to hive, and the hive factor was not considered in overall models but for each treatment, as each treatment had two hives associated with it that could be treated as blocks.
Overall, mixed effects models of treatment with subject and time as random factors were fitted and compared with generalized linear models (no random effects). For all response variables, the generalized linear models had AIC values that were lower (< 100) than the mixed effects models, which suggested that generalized linear models were a better fit than mixed effects models and that individuals within each treatment could be treated as being independent of each other. This simplified the analysis a response of the number of bees that visited each chamber and a single treatment factor applied to each chamber. Practically speaking once we have a single factor and a count response there is a little benefit to mixed model or generalized model over a traditional Chisquare analysis. Consequently, choice was assessed with Pearson's Chi-squares instead and total time and flower interactions were assessed using Welch's two sample t-test. All three choice options, hive, left and right, were compared to each other. Expected values of hive choices (E H ) were calculated by taking the proportion of total hive choices (C H ) across all treatments.
This was done to adjust for the foraging motivation of individuals leaving the hive and assumed that individuals leaving the hive were doing so with the intent to forage. Right-left choice expected values (E LR ) were calculated by assuming all non-hive choices should be equal.
Within each directional treatment, hive could be used as a random blocking factor, as multiple hives were sampled for individual bees. This analysis was performed with the Lmer function in the lme4 package, subject was treated as a nested random factor within hive and time was dropped as a random factor since hive and time random effects cannot be separated. A likelihood-chi square test was used to assess the role each factor, position, and plant status played in predicting choice.
Foragers make repeated decisions both between floral patches and within floral patches. To examine these choices beyond the first choice, we constructed Markov state diagrams of observed proportions of choices and analyzed probabilities using Markov Chain Monte Carlo (MCMC) via the R package bayespref (Fordyce et al. 2011;Francis et al. 2016). MCMC analysis generates estimates of the probabilities of making any given choice, telling us which choices are made and how often. On the other hand, the Markov diagrams show us the specific transitions between states in the Y-maze, essentially when choices are made. Thus, paring these two methods together allows us to examine both the frequency of when a choice is made and the conditions under which that choice is made. This system allows us to do this analysis without marking or radio tracking each individual bee. Markov diagrams assessed the first 4 choices of a forager within the maze. This number was chosen because foragers on average made 4.36 choices per observation period. MCMC models were run for 100,000 generations with a 10,000-burn in.

Do Environmental Factors Influence Choice?
Of the environmental factors measured (humidity, light, and temperature), none were shown to be significant predictors of choice, foraging time, or flower/ plant interactions in any treatment combination. This was true for both generalized linear modelling and mixed effects modelling. Mixed effects models considering environmental conditions (temperature, light, and humidity) as random effects showed no improvement over generalized linear models that did not consider environmental conditions (AIC 152.02 and 150.765 respectively). Further analysis using Welch's two sample t-test indicated that neither chamber significantly differed from the other across all meas- Is There Side Preference Bias?
In assessing the validity of the Y-maze for evaluating forager choices, the potential that bees have an inherent "side preference" needs to be considered as a possible confounding factor. We defined side preference to be a "preference for choosing, remaining, or foraging in one terminal arm versus the other, when both stimuli were equal." Essentially side preference would be an innate directional preference independent of any stimulus or environmental condition. A forager's hive of origin had no significant effect on first choice (t 4 = -0.426, P = 0.671), number of flower interactions (t 4 = -0.472, P = 0.637), or time spent in either arm (t 4 = 1.412, P = 0.1602). Further, first choices did not significantly differ between left and right overall (Table 3, X 2 1 = 0.028, P = 0.8673). Plant status was a significant predictor a foragers' first choice within the Y-maze (Table 3, X 2 1 = 7.582, P = 0.0059). When choice was examined as left, right, or return to hive, only the choice to return to the hive was significant (Table 3, X 2 1 = 8.591, P = 0.0034). Hive was chosen less frequently than left or right. This is expected if bees are leaving the hive with the intention to forage, as they would be expected to fly to a flower rather than to return the hive.
Neither the total time spent in each chamber (Table 4, t 188.85 = -1.134, P = 0.2577) nor the total number of flower interactions (Table 4, t 263.08 = -0.437, P = 0.6626) differed significantly between chambers. The only treatment combinations that yielded significant differences between left vs. right were the ones which included FP:NFP or NP (Table 4).

Do Bees Choose Patches and Treatments as Expected?
The foraging preferences of bees under different conditions were examined by constructing Markov diagrams and employing MCMC. In treatment FP:FP, MCMC analysis showed a slight preference for the left arm (~ 39%) over the right arm (~ 36%) (Fig. 3A). However, the credible intervals highly overlap at (Left: lower = 0.272, upper = 0.506; Right: lower = 0.253, upper = 0.486), which suggests no significant difference between arms. This contrasts with the preference individuals displayed when presented with unequal floral cues in terminal arms preferring flowering plants (0.484) over terminal arms housing non-flowering plants or no plants (0.280) (Fig. 3B. This is expected as there is no inherent difference in the patches presented and therefore bees should choose forage equally between both arms. When considering treatments where both terminal arms contained matched plant status (FP:FP, NFP:NFP, and NP:NP), a similar pattern to FP:FP was observed (Fig. 3A, C, D).
When flowers were absent from both chambers, individuals were more likely to move between terminal arms than when flowers were present in both chambers (Fig. 4A). This might also be expected if bees are searching for resources (flowers). When flowers were present in both arms, individuals were more likely to return to the hive after visiting one arm than to move between arms (Fig. 4B). This suggests some amount of satiation of the foraging drive when visiting one arm. When non-flowering plants were present in both arms, foragers were equally as likely to return to the hive without visiting either arm as they were to visit either arm (Fig. 4C). When no plants were in both arms this pattern was not observed, individuals were more likely to move between terminal arms for the observed period. Finally, when one treatment arm contained flowering plants (FP) and the other arm contained either NFP or NP, foragers were more likely to select the chamber with flowers (Fig. 4D).

Discussion
In this study, we developed and evaluated a new apparatus and statistical analysis methodology for examining the foraging behavior, and specifically the patch choices of bumble bees. Our results suggest that the system is effective and free from any obvious biasing factors. Environmental factors are known to influence the host selection and foraging success of pollinators (Comba 1999;Chittka et al. 2013;Muth et al. 2017). Accordingly, we specifically designed and oriented the system within our greenhouse to mitigate the effects of any of these factors. Within the system, environmental conditions did not play a significant role in choices to visit either terminal arm, in the time spent in each arm, or in the number of flower/plant interactions exhibited by individuals. Another potential confounding factor in choice bioassays is side preference. That is, an inherent bias within individuals towards one direction (AKA handedness). Previous research has indicated that there may be slight directional biases within Y-maze chambers using intact plants (Jaworski et al. 2015). There was no evidence of side preference in this study, nor was their evidence of side preference reported by Dyer et al. (2008) who used a variation of a Y-maze. The evenness of the micro-climate across the chambers (e.g., temperature controlled with approximately equal light intensity across the apparatus) in our design may have contributed to the lack of directional side preference displayed by the bees. It is also notable that this research was conducted with a different bee species Table 4 Analysis of difference between left and right for total time foraging in either arm, number of flower interactions, and first choice. Total time and flower interactions were analyzed using a Welch's t-test. Choice was analyzed using chi-squares and results are reported in Table 2 Trial forage. In this study, plant status drove the differences observed between preference for the left and right terminal arms, thus, we conclude that the y-maze apparatus described here is well suited for analyzing foraging choices mediated by plant signaling. Previous, research on the foraging behavior of Bombus spp. has focused on intra-patch choices and learning strategies (Gumbert 2000;Raine and Chittka 2008;Riveros and Gronenberg 2012;Groen et al. 2016;Riveros et al. 2020). Understanding and quantifying these behaviors is vitally important for understanding the implications of unequal or uncertain floral resources within a floral patch (Lefebvre et al. 2007;Carvell et al. 2011;Minahan and Brunet 2018). This focus is one of necessity as understanding patch level choices necessarily involves field work that can be limited in which used Bombus terrestris (Hymenoptera, Apidae, L.). These two differences could explain the behaviors observed in this study. The non-significant difference in total time spent in each terminal arm and the number of flower interactions suggests that in addition to a lack of choice or preference, there is also no preference for remaining in either arm. Within the treatment combinations significant differences were only observed when flowers were paired with non-flowering plants or no plants, which is further supported by previous studies on plant-pollinator interactions (Odell et al. 1999;Leonard and Masek 2014;Rivest et al. 2017;Chittka 2017). Bees are expected to respond to signals from flowering plants and to spend more time with flowering plants on which they can scope by the variability over large spatial scales (Cartar and Abrahams 1996;Hagen et al. 2011;Woodgate et al. 2016). In our study, we observed that bumblebees presented with flowering plants in both arms of the Y-maze had a higher probability of returning to the hive after making a choice, than continuing to move between the arms. This recidivism to the hive suggests that each arm contains enough floral resources to satiate a foraging motivation, which indicates that this system could serve as a proxy for patch level foraging choices (Waage 1979;Cartar and Abrahams 1996;Goulson 2000;Lefebvre et al. 2007).
When presented with non-flowering plants, a third of the observed bees returned to the hive without making a choice, while when being presented with no plant, only 15% returned to the hive without making a choice. This suggests that there is some cue from the non-flowering plants that the bees can recognize and associate with non-flowering plants. Further, when no plants were present, foragers spent the observed time primarily traveling between terminal arms. This suggests that the absence of floral cues may induce search responses where the bees are looking for floral patches (Waage 1979;Goulson 2000). This is further supported by the observation that individuals who visit an arm containing soil or non-flowering plants subsequently move to the arm containing flowering plants 2/3rds of the time. The difference in observed behaviors implies that bees can distinguish suitable patches, unsuitable patches, and non-existent patches which would be expected (Amaya Márquez 2009;Chittka 2017). More importantly, it confirms that the terminal arms in our apparatus likely function as distinct floral patches and can be used to simulate differing patch conditions.

Conclusion
The results of this study indicate that our apparatus is suitable for addressing questions about patch level forager decisions with respect to plant traits. Specifically, the smaller spatial scale allows for more in-depth analysis of the behavior of individual foragers that may not be practical in a field setting. The choices made by foraging individuals in this study followed predictions of forager behavior laid out in previous literature (Lefebvre et al. 2007;Leonard et al. 2011;Kaczorowski et al. 2012;Austin et al. 2019). This design allows for easy manipulation of simulated patches, facilitating studies on foraging and pollination behavior that incorporate plant condition. The statistical methods also present a simple solution for looking at foraging choices over set periods with the MCMC modeling and Markov designs to look over longer foraging periods. Additionally, it facilitates reciprocal measurements of plant trait effects on bee foraging behavior and of bee foraging behavior on plant traits. This could prove particularly useful for examining pollinator enhancement from seed mixes in strips, insecticide effects on pollination, and plant mediated effects on pollinator behaviors and colony health.