Study system

Medicago sativa L. (Fabaceae), known as alfalfa or lucerne, is a tetraploid, self-compatible, perennial legume, widely cultivated worldwide. Flowers are organized into clusters or racemes and range in color from pale to dark purple [10]. Alfalfa relies mainly on bees for pollination and subsequent seed set, and a tripping mechanism enables pollen transfer. When a pollinator presses the two lower petals that form the keel, the enclosed stamens and pistil are forcefully released, and pollen is deposited on the pollinator’s body in a process called tripping. At the same time, pollen already on the pollinator’s body gets deposited on the stigma. Flowers that are not tripped do not set or sire seeds.

While alfalfa does not have sexually compatible wild relatives in the United States, feral alfalfa populations are relatively common near alfalfa seed production areas [16]. Moreover, the GR gene has been detected in a significant number of these feral populations and the gene can persist over time [16, 24]. The GR gene has also been identified in feral alfalfa populations in areas where only hay is produced [Brunet (unpublished data)]. Therefore, the GR gene can move not only among cultivated fields but also to feral populations and it does persist in the wild.

Managed pollinators in alfalfa seed production in the United States include the alfalfa leafcutting bee, M. rotundata and the European honey bee, A. mellifera. The strict habitat requirement of the alkali bee, Nomia melanderi Cockerell, restricts its use to the Walla Walla valley in Washington [23]. Wild bees, including the common eastern bumble bee, B. impatiens, also visit alfalfa flowers [25]. Wild bees have been shown to contribute to alfalfa pollination in Argentina [26] although this has not been reported for the United States. While bumble bees are used as pollinators in alfalfa breeding programs, they are not used in alfalfa seed production in the United States. These distinct bee species trip flowers at different frequencies (i.e. tripping rates, or the proportion of visited flowers that are tripped by a pollinator) [6, 27], and a higher tripping rate is linked to a greater seed set per bee visit [10]. While alfalfa leafcutting bees have high tripping rates and are efficient pollinators, honey bees learn to avoid the tripping mechanism and often become nectar robbers by drawing nectar from the side of the flower.

Experimental design

The experiments were carried out inside cages set up in a greenhouse room at the University of Wisconsin Walnut Street Greenhouse in Madison, WI, USA. A smaller 2.0 x 2.0 x 1.8 m cage adjoined a larger 4.0 x 2.0 x 1.8 m cage (Fig 1). The frames of the cages were made of PVC tubing and were covered with mosquito netting (Skeeta, Inc., Bradenton, FL, USA) which was also used to separate the two adjacent cages. Glyphosate-resistant (GR) alfalfa plants, also known as Roundup Ready (RR) alfalfa, were placed in the small cage and used as pollen donors in the experiment. Conventional alfalfa plants were placed in the large cage and used as pollen recipients.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Experimental set up.

A bee hive, or nesting board with bees, was kept in a small cage (A) adjacent to a large cage (B). Prior to trials, bees were trained to forage on conventional alfalfa plants in the small cage. During a trial, glyphosate-resistant (GR) donor plants were presented to a bee inside the small cage. After a bee tripped five flowers on the donor plants, it was moved to the large cage to forage on the conventional plants.

https://doi.org/10.1371/journal.pone.0271780.g001

The GR plants used as pollen donors had a single copy of the GR gene, and at least three copies of the GR allele. Because alfalfa is a tetraploid, all pollen grains of plants with at least three copies of the GR allele will carry at least one GR allele. Because the GR allele is dominant, one copy of the gene suffices for the expression of glyphosate resistance in the seeds and seedlings. While the GR gene is hemizygous when inserted, crosses are made between plants to increase the number of copies of the GR allele in plants that carried only one inserted copy of the GR gene. The number of gene and allele copies is verified by the proportion of GR seeds produced in crosses and via quantitative PCR.

Trials were run for one bee species at a time. Prior to starting a set of trials, bees were trained to visit alfalfa plants in the small cage, using several blooming conventional alfalfa plants. For bumble bees, we used a hive containing 75–100 workers (Koppert Biological Systems, Howell, MI, USA) and for honey bees, a four-frame nucleus honey bee colony (Gentle Breeze Honey, Inc., Mount Horeb, WI, USA). For leafcutting bees, 30–50 bees, incubated from commercially available cocoons, were released into the training cage every two days for the duration of the experiment. Both male and female leafcutting bees were released in the small cage to allow mating, because mated females collect pollen and nectar to provision their eggs. Only female bees were used in the trials. A 30 x 50 cm nesting board was placed in the small cage to permit leafcutting bees to nest. The hive or nesting board was set on the side of the small cage opposite the opening to the large cage (Fig 1).

Trials began after bees foraged consistently on alfalfa in the training cage. Prior to starting a trial, alfalfa plants used for training were removed from the small cage and bees were not allowed to visit any plant for at least 15 hours. We could close the bumble bee hive the evening prior to a trial, but honey bees and leafcutting bees remained free inside the cage. However, in the absence of plants, most honey bees would stay inside the hive and leafcutting bees would lay still on the netting. To set up a trial, two flowering GR alfalfa plants, used as pollen donors, were placed in the small cage. In addition, 12 conventional flowering alfalfa plants, used as pollen recipients, were placed in a linear array in the larger cage, each plant separated by 15 cm (Fig 1). Floral display size on recipient plants was standardized to 5–6 racemes per plant and 4–6 untripped flowers per raceme. A thin thread of a different color was used to distinguish the five racemes on each recipient plant, and each flower within a raceme was marked with a small dot of a distinct colored paint. Based on our observations, these markings did not affect bee behavior.

Bee observations

Only one individual bee was tested in the large cage at any given time. For bumble bee, one female bee was released from the colony into the small cage. After the bumble bee tripped five GR flowers from the donor plants in the small cage, the screen separating the cages was lifted and the bee was moved to the larger cage. For honey bees and leafcutting bees, the first female bee to start foraging on the donor plants in the cage was tracked until it tripped five flowers. The bee was then caught in a centrifuge tube and moved to the large cage. Tripping five flowers required visiting, on average, between six and 14 flowers, depending on the bee species.

As soon as a bee started foraging on the conventional plants in the large cage, we used a voice recorder to note each plant, raceme, and flower visited in succession by the bee, for the duration of the foraging bout. A foraging bout could start from any plant position and revisits to plants and flowers were permitted. For the social bees, a foraging bout ended when the bee started flying back towards the hive, and for the leafcutting bees a bout ended when the bee landed for more than two minutes without grooming. Prior experience indicated that leafcutting bees were unlikely to keep foraging after this two-minute rest period.

Flower order, distance traveled between successive flowers, and fruit collection

Each flower visited in succession during a foraging bout was given a flower order, corresponding to the order in which the flower was visited by the bee. At the end of each trial, we measured the distance between all pairs of flowers visited in succession by the bee. Each tripped flower was marked with a numbered tag and periodically checked for pod (fruit) development. Untripped flowers do not form pods because they remain closed, with no pollen being removed from the anthers or deposited on the stigmas. Pods were collected 4–6 weeks following a trial, when seeds were mature. Each individual pod with the associated tag was kept in a separate coin paper envelope and transferred to a refrigerator until pods could be hand-threshed to extract the seeds, and seeds could be counted and tested for the presence of the glyphosate resistance gene.

Detection of the glyphosate resistance gene

Seeds collected from the bumble bee and leafcutting bee trials were tested for the presence of the GR gene using a seedling phenotypic assay, combined with test strips [28]. The phenotypic assay is a modification of a protocol originally developed by [29] to distinguish between conventional and GR seedlings. The test strips, lateral flow immunoassay AgraStrip^® RUR Seed and Leaf TraitChek™ (Romer Labs Inc., Union, MO, USA), detect the CP4 EPSP protein. Because the incubator needed to grow the seedlings was not accessible during the COVID-19 stay at home period, seeds collected from the honey bee trials were all individually tested using the test strips.

For the phenotypic assay, seeds were placed in petri dishes lined with filter paper and watered with a low-concentration (80 ppm) glyphosate solution. The petri dishes were placed in an incubator set at 20°C, under a 16h light: 8h dark regime for 14 days. Three days following the initial set-up, seedlings were watered again with the low-concentration glyphosate solution. Thereafter, seedlings were watered with a 0.1% Plant Preservative Mixture (PPM, Caisson Laboratories, Inc., Smithfield, UT) deionized water solution to prevent fungal growth. At the end of the 14-day period, seedlings were evaluated for traits associated with glyphosate resistance, such as increased root length, smaller root width, and development of secondary roots or root hair [28, 29]. When a seedling was categorized as positive using the phenotypic assay, the presence of the GR gene in the seedling was confirmed using a test strip [16]. Seedlings scored as conventional based on the phenotypic assay were bulk-tested in groups of twenty with the test strip. When a positive result was detected in a bulked sample, each seedling was individually tested with a test strip using its remaining cotyledons.

Model selection

We used Generalized linear mixed models (GLMM) with a binomial error distribution and a logit link function to examine the relationship between the logit of the probability of getting GR seeds in a pod (log [p/(1-p)]) with increasing order of flower visited or cumulative distance traveled by a bee in a foraging bout. The binomial error distribution reflected the data structure which included binomially distributed constrained counts (presence or absence of a GR seed). The logit transformation limited the values of p between 0 and 1. Here, p is the probability of getting a GR seed in a pod or the modelled proportion of GR seeds in a pod and (1-p) is the probability of getting a conventional (non-GR) seed in a pod. The origin for distance was the first flower visited and we added incrementally the distances between pairs of flowers visited in succession during a foraging bout.

The first model fitted to the data assumed a linear relationship between the logit of the dependent variable and each predictor (independent) variable (order of flower visited or cumulative distance traveled in a foraging bout). Because the logit of the probability of getting GR seeds in a pod may exhibit a different rate of decline as a function of the independent variable, we evaluated two additional models. The second model used the natural log of flower order or distance traveled, and the third model utilized the square root of flower order or distance traveled. In these two models, the relationship between the logit of the dependent variable and the independent variable is no longer linear. These models are abbreviated in the remainder of the manuscript, as the (1) “standard”, (2) “log”, and (3) “sqrt” models. Each model was fitted to the combined runs for each bee species using the “glmer” function in the lme4 package [30] in R [31].

To take into account the variation among runs (foraging bouts), for each of the three types of models just described–standard, log or sqrt–we fitted a mixed effect model with either random intercept or both random intercept and random slope. Thus, for each bee species and each predictor variable (flower order or cumulative distance), we fitted six different models. Each model has a fixed (population means) intercept and slope. In addition, the random intercept corresponds to a random run effect and represents variation among individual foraging bouts in the logit of the probability of getting GR seeds set in the first flower visited. The random slope considers variation among runs in the linear decline in the logit of the probability of getting GR seeds in a pod with increasing flowers visited or distances traveled. Because individual bees were not marked in this study, we are examining variation among foraging bouts, but in a study with marked bees and a single foraging bout per bee these variables would describe the bee-to-bee variation.

The six models per predictor variable per bee species were compared using the Akaike Information Criterion (AIC). The model with the lowest AIC was considered the best fit to the data and models were considered distinct if the difference in AIC between models was greater than two [32]. We examined a total of 36 models, two predictor variables times three bee species times the six models described above. The figures, presented for the best models, illustrate the back transformed logit, i. e. p or the probability of getting GR seeds in a pod. Logits were back transformed using the following equation: p = exp (a + b predictor variable) / [1 + exp (a + b predictor variable)]. Individual data points on the figures represent the observed proportion of GR seeds in a pod, for each pod within each of the foraging bouts, where a pod is a fruit developed following bee visits to a flower.

Analyses were performed on tripped flowers because untripped flowers do not set seeds. The order of visit of tripped flowers was determined and cumulative distance calculated between tripped flowers visited in succession. The cumulative distance was converted from centimetres to metres to facilitate model fitting. Because a flower is only tripped once, if a flower was visited more than once, we assumed it got tripped during the first bee visit. A previous study indicated no increase in the number of pollen grains deposited over repeated visits [33].

We obtained the best model to explain the data (proportion of GR seeds in a pod) for each bee species separately. We can only compare the model parameters between bee species when the same type of model best explains the data for these bee species. When these conditions are met, models can be compared between bee species by adding bee species and the interaction between bee species and either flower or distance traveled as factors in the model. This approach fits the same model type to each bee species and compares their slopes and intercepts. A statistically significant species effect indicates distinct intercepts between bee species, while a statistically significant interaction illustrates different slopes. However, when different model types best fit the data for each bee species, this statistical model cannot be used to compare parameters between bee species and the intercepts and slopes are contrasted visually.

Comparing foraging metrics among bee species

We compared different foraging metrics among the three bee species using one-way analysis of variance (ANOVA), followed by multiple pairwise-comparisons with Tukey HSD tests in R [31]. These foraging metrics included the total number of flowers visited per foraging bout, the proportion of flowers visited in a foraging bout that were tripped, foraging bout duration, and the cumulative distance traveled per foraging bout. A foraging bout starts with the first flower a bee visits in the array, and ends when the bee leaves the array. We also determined the proportion of flowers tripped in a foraging bout that set seeds, and the total number of seeds set and of GR seeds set per pod, and per foraging bout.

Bumble bee models

We had 12 foraging bouts (runs) and 322 observations (pods with seeds). The logistic regression analysis only considers the pods that set seeds. For both flower order and cumulative distance, the log-odds ratio or logit of getting GR seeds in a pod was best described by the log model with random intercept, log_e (p/(1-p) = intercept—slope log_e flower order (or distance) (Table 1). The intercept and slope with standard errors for the fixed effects of the best model for flower and for distance are presented in Table 2. These standard errors (SE) illustrate the accuracy of the parameters. For the flower model, the standard deviation (SD) associated with the random effect for the intercept was 1.41 and illustrates variation among foraging bouts in the logit of getting GR seeds in the first flower visited in the foraging bout. For distance, the standard deviation for the random intercept was 1.50. The probability of getting GR seeds in a pod decreased with successive flowers visited for flower order (Fig 2A), and a similar trend was observed for the cumulative distance traveled by a bee (Fig 2B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Probability of getting GR seeds in a pod with increasing flower order or distance traveled (m) by a bee.

Probabilities represent back transformed logits (see text for details). For bumble bee, the log model with random intercept was the best model for both (A) flower order and (B) cumulative distance; for honey bee, the square root (solid line) and the standard (dashed line) models with random intercept were equally good candidates for (C) flower order while the square root model with random intercept was the best model for (D) distance; for leafcutting bee, the square root model with random slope and random intercept was the best model for both (E) flower order and (F) distance.

https://doi.org/10.1371/journal.pone.0271780.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Akaike information criterion (AIC) values for flower and distance models for three bee species.

https://doi.org/10.1371/journal.pone.0271780.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Fixed effect estimates for the best flower and distance model(s) for bumble bees (BB), honey bees (HB), or leafcutting bees (LCB).

https://doi.org/10.1371/journal.pone.0271780.t002

Honey bee models

The data for honey bees included 14 foraging bouts and 268 observations. We obtained two best models for flower order, the square root model with random intercept, log_e (p/(1-p)) = intercept—slope sqrt flower order, and the standard model with random intercept, log_e (p/(1-p)) = intercept—slope flower order (ΔAIC < 2) (Table 1). The intercept and slope with standard error for the fixed effects of the best model are presented in Table 2 The standard deviation associated with the random intercept was 0.75. For distance, the square root model with random intercept was the best model (Table 1), with parameters for the fixed effects of the model presented in Table 2. The standard deviation associated with the random intercept was 0.82. In all cases, the probability of getting GR seeds in a pod decreased as more flowers were visited in succession by a bee (Fig 2C), or longer cumulative distances were traveled (Fig 2D).

Leafcutting bee models

We obtained 27 foraging bouts and 202 observations for leafcutting bees. The square root model with random slope and random intercept best explained the logit of the probability of getting GR seeds in a pod, for both successive flowers and for cumulative distance (Table 1). The parameters for the fixed effects of the model (intercept and slope) with respective standard errors are presented in Table 2. For flower order, the standard deviations were 2.12 for the random intercept, and 0.89 for the random slope. For distance, the standard deviation associated with the random intercept was 1.51 and with the random slope 1.90. The probability of getting GR seeds in a pod decreased with successive flowers visited (Fig 2E), and with longer cumulative distance traveled by a bee (Fig 2F). For leafcutting bees, in addition to the random intercept, the random slope variable, i. e. considering among foraging bout variation in slope, improved the fit of the model to the data.

Comparing models among bee species

Because the type of the best model (standard, log or sqrt), for either flower order or distance, differed among bee species (Table 2), we could not apply the same model type to statistically compare the intercepts and slopes among bee species. Therefore, to address potential differences among bee species, we visually contrasted the graphs illustrating the probability of finding a GR seed in a pod with successive flowers visited and cumulative distance traveled (Fig 2). Bumble bees had a fairly steep rate of decay in the probability of getting a GR seed in a pod (slope), but the probability levelled off after visiting approximately 30 flowers or traveling five meters, creating a distribution with a long tail (Fig 2A and 2B). When bumble bees moved pollen from flower to flower, GR seeds could still be found with a low probability in a pod after a bee visited 250 flowers or traveled over 40 meters. Honey bees were intermediate between bumble bees and leafcutting bees, with a negligible probability of getting a GR seed in a pod after a honey bee visited approximately 100 flowers or traveled around 15 meters (Fig 2C and 2D). The rate of change in the probability of getting a GR seed in a pod with increasing flower order or distance was steep when pollen was moved by leafcutting bees (Fig 2E and 2F). The probability of getting a GR seed in a pod was negligible after a leafcutting bee visited 50 + flowers or traveled five + meters (Fig 2E and 2F).

Foraging metrics

All foraging metrics, except the proportion of tripped flowers that set seeds, varied among bee species (Table 3). Bumblebees visited more flowers, spent more time foraging, and traveled greater cumulative distance in a foraging bout relative to honey bees or leafcutting bees (Table 3). Leafcutting bees tripped the most flowers and set a greater proportion of GR seeds per foraging bout relative to honey bees and bumble bees (Table 3). Honey bees set the most mature seeds per pod and were otherwise similar to either bumble bees or leafcutting bees, depending on the variable examined (Table 3). Although leafcutting bees set a greater proportion of GR seeds per foraging bout, they set fewer seeds, and the lowest number of GR seeds per foraging bout (Mean ± SE) (9.22 ± 1.31), relative to bumble bees (31.08 ± 11.87), and honey bees (17.07 ± 3.49). However, per pod, leafcutting bees set slightly more GR seeds (1.47 ± 0.14) relative to honey bees (1.03 ± 0.20) and bumble bees (0.98 ± 0.31).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Differences in foraging metrics (per foraging bout) among bee species with respective means and standard errors.

https://doi.org/10.1371/journal.pone.0271780.t003