Methods to identify pollinator visitation patterns include passive sampling techniques (which sample bees and other insects that are present at a particular site, and not necessarily visiting the focal flowering plant species), hand/net collection of pollinators on focal plants, and direct observations of pollinator interaction with the flowers on the focal plant species. Though these methods may be used with all bee species, they vary in benefits and disadvantages, which include biases toward different taxa, time constraints, and false detection.

The most commonly used passive sampling tools in pollination research are bee bowls and blue vane traps. Bee bowls, or pan traps, are containers of soapy water colored in blue, yellow, or white to attract and collect bees (Droege et al., 2017). This method is inexpensive, requires no training to use, and can sample continuously over time. However, bowl traps are biased in that they are more likely to attract pollinators when the surrounding floral community is less dense (Kuhlman et al., 2021), are less likely to capture larger bees (Rhoades et al., 2017), and their efficacy can vary depending on bowl size and the degree of ultraviolet reflection, which can vary among manufacturers (Prendergast et al., 2020). Like bee bowls, blue vane traps (BVTs) are also used as a passive sampling method, where a base is attached to a UV reflective blue vane top to attract and trap bees. BVTs are particularly useful for broadly sampling the insect community (Gibbs et al., 2017), are more successful in capturing larger bees and attract less non-bee bycatch species than bee bowls, but are less likely to capture small, parasitic species such as Nomada and Sphecodes (Joshi et al., 2015). Because BVTs are less likely to be disturbed by wildlife scavenge for water, researchers can use diluted ethylene glycol as a killing and preserving fluid rather than soapy water. Preservation with ethylene glycol rather than soapy water prevents bee degradation, which allows researchers to only check traps periodically rather than daily. However, though BVTs have many benefits, it has been suggested that these traps may oversample areas and negatively impact pollinator populations (Gibbs et al., 2017). Regardless of the method used, passive sampling still has limitations in determining focal pollinators and the background pollinator community. For example, a bee species may not be detected if more attractive flowers are nearby (Baum and Wallen, 2011; Kuhlman et al., 2021). In addition, passive trapping methods under-sample certain bee species, such as cleptoparasitic species (Mackert, 2019), and trap other insect species, such as beetles, ants, and flies (Prendergast et al., 2020; Montgomery et al., 2021). Passive sampling methods also cannot confirm that captured species are actually visiting focal plant species near traps. Therefore, these methods should be used with caution and used in supplement with other detection methods.

Active methods include direct observation and targeted collection directly from flowers. Researchers can capture bees visiting flowers by sweep net, insect-vacuum, or by hand with vials. Collection alone does not indicate that bees are visiting the focal plant, as sweep netting and insect-vacuums can collect individuals that are near flowers but not collecting nectar or pollen. Direct observations are required to determine if pollinators are collecting nectar and/or pollen from a focal flowering plant. A benefit of active methods is that unlike passive sampling, these methods confirm that a particular bee species actually visited the focal plant. However, active methods take more time and trained collectors or observers, are biased toward slow-flying bees, and lack the potential to continuously sample as passive traps do (Wilson et al., 2008; Prendergast et al., 2020). Observations may also be biased depending on the level of expertise of the observer, and identification of more cryptic genera such as Andrena or Lasioglossum is not possible without immobile specimens and a microscope (Biddinger et al., 2013b; Sutherland et al., 2015). Time is also a variable that must be taken into consideration. Studies often only collect or observe bees at each focal plant for a few minutes before moving to the next, potentially missing important visitors. For example, Vaudo et al. (2014) found that certain plants had high visitation rates at particular times of the day, which could have been missed if visitation had not been consistently observed in a controlled setting (this study took place in a greenhouse). Some crops such as pumpkin also have long bloom periods and may be dependent on different pollinator taxa throughout this period. As such, preliminary work should be conducted to determine bee foraging schedules and floral bloom periods before designing optimal sampling or observation strategies.

Factors other than nutrition also impact which bees may visit a focal plant, as nutrition is not the only driver of bee populations in a study area. Certain species require particular nesting materials to persist in a given area (e.g., small carpenter bees Ceratina require hollow, pithy stems, and mason bees require a choice of soil textures] (Pinilla-Gallego et al., 2018; Rehan, 2020). Management practices such as pesticide application can also negatively affect bee populations, with differential pesticide sensitivities across species (Biddinger et al., 2013a; Woodcock et al., 2016; Manjon et al., 2018). Tilling can also reduce populations of ground nesting bees (Ullmann et al., 2016). Therefore, a bee species that is expected to be present in a study system may not be due to lack of nesting resources or other stressors, such as pesticide exposure, and not due to the nutritional content of the focal plant. Floral rate of visitation may also be confounded by the surrounding plant-pollinator community (Essenberg et al., 2012; Valido et al., 2019). Attractive plants nearby or competition among bee species may alter visitation rates to the focal plant species. Accounting for variation bee populations, species richness, and visitation rate due to variance in other resources, environmental stressors, and the background community of plants and pollinators requires studies with diverse background communities in different habitats or geographic regions, which can be time consuming and costly. Determining visitation of certain plants may also be difficult due to challenges in sampling from or observing focal plant species. For example, tall trees such as black cherry (Prunus serotina) have blooms that are several meters above the ground, which require a lift to allow researchers to access (McLaughlin et al., 2021).

Even if care is taken to prevent bias from collection methodology, observers, timing, and landscape factors, there are still significant constraints when using the above methods. First, there is the significant challenge of insect identifications, as it may be quite time-consuming, and requires highly trained scientists who can identify the hundreds of bee species that may be present in a given region (Jamieson et al., 2019; Kilpatrick et al., 2020; Engel et al., 2021). When trained scientists are not available, bees may be identified to morpho-taxa, or easily identified species and groups of bees with common morphological traits. Common groups used include: honey bees, bumble bees, large carpenter bees, hairy legged bees, small black bees, large black bees, and small green bees, but these classifications differ (and may be more specific) based on research goals and the geographic area of interest (Ullman et al., 2008). Identification to morpho-taxa, while easier, can only give information on very broad ranges of pollinators and does not provide detailed taxonomic classifications that aid in practical application of results. Sorting Hymenoptera by morpho-taxa can also underestimate species richness by 12%, and non-specialists have been found to correctly group only 44% of taxonomic species (Derraik et al., 2010).

For Approach 1, researchers have traditionally identified and evaluated the key floral visitors of focal plants by using passive and active collection or observation methods. Though these methods can be applied to all bee species, there have been limitations in extensively sampling bees across time and landscapes. Properly identifying wild bees that visit focal plants has also constrained these studies. Below, we discuss how automated monitoring and identification technologies will allow researchers to address questions of floral visitation at much larger spatio-temporal scales than has previously been feasible.

Cameras, automated monitoring systems, and automated identification technologies have rapidly advanced since their inception (Devillers, 2014; Preti et al., 2021). Using cameras to monitor insect behavior is not a new concept and was first conducted by Kondo et al. in 1994 (Reviewed in Preti et al., 2021). However, using cameras to monitor insect plant-pollinator interactions was not demonstrated until 2009 (Marten-Rodriguez and Fenster, 2008). Then, in Lihoreau et al. (2012), motion-sensitive cameras were used for the first time to track pollinator movement. Researchers have also furthered the field by developing automated data processing and artificial intelligence programs that present unique opportunities to study plant-pollinator interactions (Høye et al., 2021).

The development of automated monitoring and identification systems will allow researchers to study plant-pollinator interactions across greater spatio-temporal scales than has previously been feasible. Installation of cameras at focal plants can allow researchers to collect exceptionally large datasets to answer questions that would be nearly impossible, or extremely time consuming, to answer by human observation or sampling alone (Høye et al., 2021). Plant-pollinator interactions may be recorded continuously across multiple days by cameras, and cameras may also be placed in many geographic regions or landscape types to monitor all sites simultaneously. Klomberg et al. (2022) were able to simultaneously monitor plant-pollinator interactions at an elevational gradient along 6 different transects. Similar studies conducted by human observations can only capture the community of insect visitors during a snapshot of time, which can miss key pollinators, such as squash bees that forage very early in the morning and retreat to below-ground nests in the evening. Frequency and richness of insect visitation is also dependent upon weather and time of day, and studies with human observation alone could be biased by observation period (Goodwin et al., 2021). Finally, plant-pollinator community dynamics change throughout seasons, between seasons, and vary by year (Mora et al., 2020). Even if constant human observation is possible, sampler fatigue interferes with observation after a short period, and samplers may be biased to notice larger, showy species more than smaller, fast-moving species (Biddinger et al., 2013b). Continuous, reliable monitoring by cameras across seasons and years can uncover more fine-scale nuances in a changing pollinator community.

However, the collection of these large data sets is only beneficial if coupled with artificial-intelligence powered computer-vision based automated identification systems. On a basic level, computer vision and deep learning systems can identify when a pollinator is present at the focal flower and only record these small bursts of activity, thereby drastically decreasing data storage and battery power of field equipment (Pegoraro et al., 2020). These short bursts of video or photos can then be viewed by a researcher to identify the pollinator and behavior exhibited. For example, Steen (2017) monitored 24-h forager activity rhythms using automated camera monitoring and data processing. This study would have been extremely labor intensive without automated systems, as researchers would have to parse through data manually to identify useful photos. This technology may also be used in conjunction with automated identification systems to determine which insects are present (Raitoharju et al., 2018). These systems are still limited in that they require training with large image libraries before they can successfully identify subjects, which involves hours of researcher observation and outlining of pollinators in photos. While the development of large, shared image libraries can facilitate the development of automated identification systems, often these systems require images that are taken in similar contexts as the functional camera systems. In addition, certain bee taxa are easier to distinguish between than others. For example, Spiesman et al. (2021) used convolutional neural network classification models to distinguish between 36 different species of bumble bees in North America with a ~92% accuracy rate, but many species have different phenotypes by region, such as Bombus rufocinctus, and were often falsely identified as other species. Therefore, current automated identification systems require distinct physical characteristics that are visible to cameras and easily measured, which may not be possible with more cryptic bee genera that are distinguished by minute details, such as Andrena, Lassioglossum, or Melissodes. However, cameras could be combined with other monitoring tools, such as radar to monitor wing beat frequency, which has been used to identify moths (Wang et al., 2017). Bees that may not be identified by computer systems may also have the potential to be identified by experts, depending on freeze-frame image quality. This may be particularly useful for fast-moving species that are difficult to tell apart by observing by eye alone, and visual attractants could be used to pull bees to a single camera. In Box 2, we briefly discuss how bee identification may be used in parallel with molecular methods in Approach 2. Artificial intelligence systems can also track bees as they move throughout the camera frame, allowing more complex behaviors to be quantified. Ratnayake et al. (2021) utilized this image-based tracking software to track unmarked honey bee foragers in the field. This software can account for multiple floral choices by the same insect rather than counting each bee as a single visitation event. Previous image-based tracking methods have required barcoded tags, which require a large time investment to apply to many individuals (Greenwald et al., 2015; Gernat et al., 2018; Jones et al., 2020).

Though these technologies present intriguing possibilities, there are still challenges to applying them broadly across bee species. First, automated identification will still have limitations at the genus and species level for many taxa regardless of how vast the reference library is, as some minute details may only be viewed under a dissecting microscope or discerned by an expert (e.g., wing venation, scopal hair structure, facial fovea, etc.). Thus, investment in improving current online bee identification keys (e.g., Discover Life, https://www.discoverlife.org/mp/20q?guide=Bee_genera), which are more geared toward those already trained in taxonomy, should occur concurrently, as it will be necessary for anyone using automated identification technology to assess the accuracy of their results. Future integration of algorithm and camera systems that allow for recording when movement is detected, automated identification of insects, and behavioral tracking of individual bees will expand the spatio-temporal scales at which plant-pollinator interactions can be studied.

One of the key methods used to assess plant or pollen preference of a focal bee species is by observation of visitation. Researchers may test particular plants in an open field setting to see which plants bee species prefer (Nicholls and Hempel de Ibarra, 2017), but this type of method is more often used with plants as the species of interest rather than bees, as described Approach 1. Observations may also be conducted in an enclosed or controlled setting, such as a hoophouse, foraging arena, or flight cage. These methods have been used with honey bees (Hendriksma et al., 2019a), bumble bees (Russell et al., 2016), mason bees (Boff et al., 2021), and wild bees (Dukas and Real, 1991; Howard et al., 2021). Hoophouses or flight cages can be quite large (>10 ft in width or length) to conduct studies where many bees are permitted to forage, while foraging arenas tend to be smaller sized and maintained with the lab for foraging trials with individual bees. Foraging arenas typically consist of a box with a plexiglass wall for viewing and a forager entrance. Preferences can be assessed by offering flowers, dishes of pollen, or even artificial diets to understand bee nutritional requirements (Vaudo et al., 2016). Real flowers offer more realistic scenarios and can be used for floral trait choice rather than nutritional preference alone, but also come with confounding variables such as floral humidity and volatiles that play a role in attraction (Arx et al., 2012; Burkle and Runyon, 2019). Imitation flowers, pollen, and artificial diets do not have these confounding variables, but are less field-realistic. Any foraging choices should have randomized locations within the enclosure that are spatially rearranged between trials to prevent bias from forager learning and trap lining (Saleh and Chittka, 2007; Ohashi and Thomson, 2009). Choices should also be regularly replaced to prevent chemical scent deposition from interfering with future trials (Saleh et al., 2006).

Using enclosed spaces to observe bee foraging patterns allows for more control than field observations. In enclosed settings, individual bee responses, and thus, variation in responses within species, population or colony, can be observed. Foraging arenas also make it possible to observe individuals as “naive foragers,” which allows researchers to understand if preferences are innate or learned (Raine and Chittka, 2007). In addition, extraneous variables are removed from the study system. For example, there is no interference from focal bees visiting other flowers, or from other pollinators competing with the focal bee species (Russell et al., 2017; Adler et al., 2020). Furthermore, the timing of foraging can be controlled such that observations are initiated as bee foraging initiates (thus, the bees are not visiting flowers outside of the observation period) and the bees are allowed to forage when all flowers are open (see Vaudo et al., 2016) and so are choosing among all flowers. The difficulty associated with enclosed foraging studies is that they are primarily amenable to bee species that can be managed in an indoor or semi-field setting. However, it is possible to collect wild bees from a natural setting and transfer them to a foraging arena for a short period, though there are few examples of this method in the literature (see Howard et al., 2021 for Lassioglossum and Dukas and Real, 1991 for Xylocopa). Even managed bees do not always forage in an enclosed setting, and it is likely that researchers will require some understanding of how a focal bee species perceives its environment before arriving at a successful arena design. For example, UV ray intensity has been shown to impact indoor bee foraging, positively or negatively, depending upon species (Kendall et al., 2021). Even with previously used managed bees, each flight cage or foraging arena design may also require additional troubleshooting during preliminary data collection. Bee behavior can be influenced by variables such as light source, light intensity, and color, and each experimental protocol will require adjustments that are appropriate for the questions and surrounding environmental context (Chittka and Thomson, 2001; Erickson and Russell, pers. comm.). Regardless of optimization, there will be cases where a forager will not complete a foraging bout, so many trials may be necessary to meet statistical significance. As such, these studies can take a significant amount of time to complete.

Bee floral preferences may also be determined by examining whole bee pollen provisions sampled from bees foraging under natural conditions. Because these samples can be collected from bees foraging in the field, it is possible to assess preferences across the broader plant community at the field or landscape scale, versus providing bees with a small number of focal plants and assessing visitation through direct observation. The plant species bees visit, and the proportion of visits to each species, can be determined by examining pollen via microscopy (Campos et al., 2021). Pollen can be removed from bees collected from field sites, as they forage on flowers or return to the nest, or pollen can be collected from stores and larval provisions within nests. Pollen provisions have previously been collected from nests in wood, stems, soil, and even empty snail shells, and thus these methods should be amenable to use with any wild bee species (Müller et al., 2006). Pollen microscopy has been used to identify pollen sources in provisions from honey bees (Lau et al., 2019), bumble bees (Whittington et al., 2004), mason bees (Kraemer and Favi, 2005), and wild bees (Larkin et al., 2008). Pollen samples can be used if samples are dry, frozen (Pinilla-Gallego and Isaacs, 2018), suspended in ethanol or in insect tissues (Jones, 2012, 2014), and thus this method is feasible to also use with museum or stored samples, and is not impacted by differences in amounts of nectar or water mixed into larval pollen provisions (e.g., dry, doughy provisions of Osmia cornifrons vs. the soupy provisions of Osmia pumilla) (Wood et al., 2019). The benefit to these methods is that the scale of studies can be drastically increased (from a set of focal plants to the landscape scale, and across sites), and foraging preferences of individual bees can be evaluated, rather than averaging across individuals from a set of observations. This also prevents personnel from having to follow bees as they forage in the field, which can be quite difficult. Determining the plants visited in a landscape can also be used to examine entire plant-pollinator networks to understand how communities are interacting, and if these community networks change between seasons due to plant or bee cycles within particular species (Arceo-Gómez et al., 2020). Examining whole pollen collections of many individuals from the same species can also show trends in visitation that observations alone may not detect (Pornon et al., 2017), or individual variation within a colony or species (Klein et al., 2017; Hendriksma et al., 2019a). However, though examining whole pollen provisions is a useful strategy to employ, it is important to consider what choices bees made in the landscape based on available flowering plant species. Conducting community plant surveys to determine what plants were most and least common can help researchers understand if bees are collecting what happens to be most abundant or if they are seeking out rare species. Determining broad trends in plant preference can lead to choice tests between individual plant species and cultivars to narrow down the best planting choices for developing pollinator forage habitat, or to eliminate competitive species around a crop that is the target of pollination.

Researchers have traditionally used microscopy to view bee-collected pollen and identify these samples to plant taxa the sample was derived from Campos et al. (2021). Pollen acetolysis, a process wherein pollen is treated with powerful acids to remove cellular content, is often used to make the morphological features of pollen grains more visible (Jones, 2014; Halbritter et al., 2018). Staining with safranin O during slide mounting also improves contrast of features. One of the main benefits of using microscopy over other methods is that pollen species may be identified even if the available quantity is quite low, which is necessary when recovering pollen from bee tissues. However, pollen slide preparation requires a significant time investment and is not a high-throughput method. Proper pollen identification also requires experience and training. Identification keys are freely available from the Global Pollen Project database or natural history museums, but the vast majority of plants are not documented in these libraries, and museums often have fragmented collections based on geographic location (Martin and Harvey, 2017). Automated pollen identification programs are also available (Cao et al., 2020; Aperio ImageScope - Pathology Slide Viewing Software, 2021). These programs can make identification less time consuming, as though slide mounting pollen and taking photos of pollen grains is necessary, extensive prior knowledge of pollen morphological traits is not needed. This technique is 98% effective at accurately identifying pollen to species, including species previously considered indistinguishable from one another by palynologists (species in Fagaceae and Asteraceae) (Sevillano et al., 2020). One caveat to this system is that though it has a high success rate, the authors were only able to train and test the system to discern between 46 different species, and training the system to identify pollen from all the flowering plants in the world (~450,000) would take a significant amount of time (Pimm and Joppa, 2015).

For Approach 2, studies evaluating foraging choices in enclosed settings are feasible primarily with managed bee species that forage under such conditions. However, pollen loads of foragers of any bee species can be readily assessed, though it is only possible to collect pollen samples used to feed developing larvae in bees where nests can be managed or easily identified and harvested. Below, we detail how DNA metabarcoding has made examining pollen provisions more high throughput, allowing for additional comparisons across time, bee species, and landscapes.

The plant species that pollen has been derived from may also be identified by using DNA metabarcoding. This method has widened the breath of species that may be investigated, which include honey bees (Milla et al., 2021), bumble bees (Leidenfrost et al., 2020), mason bees (Vaudo et al., 2020b), and several wild bee species (Gous et al., 2019; Potter et al., 2019; Casanelles-Abella et al., 2021). This technique involves amplifying loci in pollen that are common across plant taxa, but vary in sequence across species, genera, or families. Loci are amplified through multiple rounds of PCR in a nested design (PCR1 amplifies target regions with generic primers, PCR2 attaches next-generation sequencing read-priming oligonucleotides, ad PCR3 adds dual multiplex indices) which can then be sequenced and compared to a reference database such as MetaClassifier (Richardson et al., 2019; Sponsler et al., 2020). This nested design is not required, but initial amplification in PCR improves the robustness of the assay by minimizing any amplification biases and mistagging events (Richardson et al., 2019). Plant-pollinator networks created from DNA metabarcoding are more complete, more accurate, and show fewer specialization linkages than those created with observations alone (Wood et al., 2016; Arstingstall et al., 2021). This technique is high throughput and allows for larger sampling efforts across years, which can help account for variation in plant blooming phenology due to differences in growing degree day accumulation each year. In Box 2, we briefly discuss how DNA metabarcoding may also be applied to bee identification.

DNA metabarcoding can be a powerful approach for screening pollen samples, but it is not as effective in determining the amount of each pollen type as microscopy (Richardson et al., 2015) and is limited in its taxonomic resolution. In most studies, identification is limited to the genus level, though some biomarkers (ITS) perform better than others (rbcL, matK) (Kolter and Gemeinholzer, 2021). ITS2 can provide much better resolution for plant classification than ITS1, but ITS2 use may result in fungal co-amplification (Cheng et al., 2016; Gous et al., 2019). Additional loci that are commonly used are trnL and psbA-trnH (Dormontt et al., 2018). Recent studies have demonstrated that species-level resolution can be improved if the genomic reference libraries are expanded to include sequences from more species. Kolter and Gemeinholzer (2021) demonstrated that species identification was improved by 19% (from 43 to 62%) for the ITS biomarker when the number of ITS sequences/species in the reference library was increased from 1 to 7. Furthermore, the inclusion of additional loci to the analysis can improve taxa resolution, but the identification of these loci can be challenging, as they must be conserved enough to allow binding from a conserved primer set but variable enough to allow for species sequence differences (Kadlec et al., 2017). Johnson et al. (2019) identified 353 single-copy protein coding genes from the One Thousand Plant Transcriptome Initiative and developed a set of 75,151 probes that target these genes that function over 600 angiosperms; this provides tremendous resolution for species identification and phylogenetic analyses. Alternatively, sequencing of organelle (chloroplast) genomes or whole genome skimming using high throughput sequencing can provide a substantial amount of sequence data, but analysis requires that the reference library contains the full sequence for the target species (Johnson et al., 2019).

Though DNA metabarcoding is a high-throughput method, there are still significant limitations in processing the substantial number of samples that are typically necessary for these studies, which can be particularly large if the goal is to create plant-pollinator networks, compare across species, or compare across different land use patterns, which can require hundreds of samples (Cullen et al., 2021). Conducting nested PCR on individual samples also takes a significant amount of time (and thus higher personnel costs) and material required. Microfluidics approaches can significantly reduce handling time and allow for more high-throughput parallel processing but require access to expensive equipment (Gostel et al., 2020). Alternatively, pull-down approaches (where sheared DNA from the samples binds to sequence-specific probes) can allow for multi-sequence amplifications within a single reaction, which can be especially useful if larger numbers of loci are used (Hale et al., 2020). Nonetheless, in all cases, adequate representation of the target species in the reference library is necessary to bioinformatically identify the samples. With efforts to sequence the genomes of plants and animals expanding (Lewin et al., 2018), reference libraries are also expanding, providing new opportunities for improving approaches to study the nutritional ecology of wild bees.

Though bees may visit certain plants, plant visitation cannot necessarily be equated to quality nutrition. For example, honey bees will often prioritize pollen quantity over quality (Leonhardt and Blüthgen, 2012). In addition, solitary bees may sacrifice pollen quality for quantity in poor landscapes, as they must make many small trips to collect resources while conserving energy rather than flying long distances (Ne'eman et al., 2006; Greenleaf et al., 2007; Biddinger et al., 2013a). Therefore, visitation rates, determined by foraging observations or examining pollen provisions, should be carefully considered with choice assays for particular nutrient levels or plant traits.

Nutritional preferences may be determined by measuring nutrient levels of bee-collected pollen, or of pollen from plant species that bees prefer. Methods have been developed to assess the concentrations of different micro- and macronutrients (namely, protein and lipids) in pollen, which are described in detail in Box 3. This approach is useful because it allows the investigator to understand the broad nutritional needs of a bee species rather than plant visitation rates alone. For example, recent studies have evaluated whether protein and lipid content are associated with different foraging preferences in bumble bees and found a preference for plants providing pollen with a specific protein: lipid ratio (Vaudo et al., 2016). These dietary preferences also shift when bumble bee colonies have brood to provision (Kraus et al., 2019). Honey bee collected pollen has also been screened for nutritional content (Ghosh et al., 2020), but few studies have assessed the nutritional composition of bee-collected pollen, or the nutritional composition of plants preferred by bees, in species outside of honey bees, bumble bees, and mason bees (Weiner et al., 2010; Belina-Aldemita Ma et al., 2019).

On a broader scale, evaluation of 82 plant species demonstrated that there is a wide range of pollen protein: lipid ratios, and pollen collected from three bee species (honey bees, bumble bees, mason bees) exhibited distinct protein: lipid ratios (Vaudo et al., 2020a). This study then suggests that different bee species may have different nutritional preferences, and thus forage preferentially on different plant species in the field, thereby effectively niche partitioning according to pollen nutrition. However, bees may be able to mix pollen from different plant species to achieve their preferred macronutrient ratios, and thus it remains to be determined if, for habitat restoration practices, it is preferable to provide plants that provide pollen that match a target bee species' protein: lipid ratio or provide plants that span a broader nutritional range, so many species can forage to meet their targets. Ruedenauer et al. (2019) assessed 387 different plant species belonging to 229 genera and 75 different families, and found that pollen protein concentration and protein: lipid ratios vary by phylogeny, suggesting that pollinator preference over time has impacted pollen nutritional content. However, the relationships between pollen nutrition, floral traits, and bee preference have yet to be fully investigated.

Rather than simply evaluating the nutritional content of pollen collected from foraging bees or in the plant species that bees prefer to forage from, it is also possible to conduct detailed bioassays using synthetic or altered pollen diets to evaluate the nutritional preferences for different quantities and ratios of key nutrients, including protein and lipids. The Geometric Framework for Nutrition provides a conceptual and technical approach for these studies (Simpson and Raubenheimer, 2012). For these studies, animals are offered diets that vary in the ratios of two macro- or micronutrients. If the animal eats from only one of these diets, its nutritional intake would fall along the “rail” (represented as a line plotted in a graph representing nutritional space) of that diet, where the ratio is kept constant as the quantity of both micro or macronutrients increases. However, if the animal feeds selectively from more than one diet, its consumption will not fall along the rails for these diets, fall somewhere in between. This preferred ratio is considered the “intake target” of the animal. This intake target can vary depending on the development or physiological state of the animal and can vary between species. In addition, determining which nutrients an animal will compromise in favor of obtaining the correct amounts of another can give insight as to which macro- or micronutrient is most important (Simpson and Raubenheimer, 2012). Such studies in honey bees have demonstrated that pollen protein: carbohydrate ratios (Hendriksma et al., 2019a), protein: lipid ratios (Stabler et al., 2021), and amino acid content (Hendriksma and Shafir, 2016) impact dietary preference, and that micronutrient requirements vary by caste (Filipiak et al., 2017). In addition to the previously mentioned studies on bumble bees, studies also indicate that bumble bees prioritize carbohydrates over protein (Hendriksma et al., 2019a), prefer diets with particular amino acid concentrations and sources (Stabler et al., 2015), and use lipids as a nutritional cue when assessing diet quality (Ruedenauer et al., 2020). Fewer studies have been conducted with solitary bees, but it has been demonstrated that carbohydrates are prioritized over protein in mason bees (Austin and Gilbert, 2021).

The approaches above examine bees' nutritional preferences with their assumption that these preferences reflect the optimal diets, but to fully test this hypothesis it is necessary to conduct studies in which bees are fed diets with different nutritional content, and their development, health, and fitness are measured. These studies can be conducted within controlled laboratory settings or in the field. Most field studies that involve habitat restoration by introducing planting schemes are limited in that they often show that managed sites have higher abundance of wild bees, but do not measure background population levels before and after site restoration. The increase in abundance of bees to the flowers or study site may reflect increased attraction to the site, and not necessarily an overall increase in the background population (Kleijn et al., 2011).

While many different measurements of success can be used in laboratory settings to assess the effect of nutrition on wild bee health and fitness outcomes, most wild bee species have yet to be successfully captured and maintained (able to feed on lab diets for an extended time period), reared (raised from immature stage to adult), or bred (raised from immature stage to adult with reproduction) in a lab setting. Bee species used in these studies thus far include honey bees (Hendriksma et al., 2019b), bumble bees (Treanore et al., 2019), mason bees (Filipiak and Filipiak, 2020), small carpenter bees (Lawson et al., 2020), and sweat bees (Roulston and Cane, 2002). Measures of success used in a nutritional assay also depend on life history traits of the focal bee species. For instance, social species can be monitored for colony growth and reproduction (e.g., wax production, resource allocation, and brood levels) (Treanore et al., 2019; Ruedenauer et al., 2020). Cavity-nesting, solitary species do not produce wax or store long-term resources for a colony but can be monitored in trap nests for offspring production and sex ratio, which is a key indicator of fitness and environmental quality in solitary bees (Eckhardt et al., 2014; Filipiak and Filipiak, 2020). Offspring production and sex ratios of ground nesting species can be assessed by taking soil cores of nests, or by pouring talcum powder in tunnels to make removing excess soil around nests easier (Leonard and Harmon-Threatt, 2019). Other measures of success that can be used for any bee species include measuring bee body size (e.g., intertegular distance, head capsule width, body mass, etc.) (Grab et al., 2019), developmental time (Lawson et al., 2020), and survival (Cane, 1987; Praz et al., 2008; Bramke et al., 2019; Fortuin et al., 2021). These physiological methods are beneficial in that they can be determined by examining individuals externally and do not require expensive equipment, but they do not provide evidence as to the underlying nutritional qualities that make a diet good or bad for bee health.

For Approach 3, the nutritional content of collected pollen can be assessed in any bee species in which foragers collect substantial pollen loads (most commonly used methods require ~6 mg of pollen for macronutrient analysis, Vaudo et al., 2020a) or where nests can be managed or easily identified and harvested. Studies of bee preferences for synthetic diets or studies of diet impacts on bee health, development and reproduction can only be completed in bee species that are amenable to some level of management and rearing under laboratory or semi-field conditions, which we detail in Table 1. Box 4 discusses the benefits and constraints of using different species for nutritional assays. Below, we discuss emerging molecular tools to assess the impact of diets on physiology, and how these methods may be applied to additional wild bee species.

Molecular methods may be used to measure bee development, health and reproduction, and may also provide insights into the mechanisms underlying the impact of nutritionally varied diets on bees. These methods can be used with any bee species (Grozinger and Zayed, 2020), but to our knowledge have thus far only been used to study the impacts of variation in pollen-based nutrition with honey bees (Crone and Grozinger, 2021) and bumble bees (Stabler et al., 2015). Expression levels of genes can be evaluated across the genome, using transcriptomics/RNA-seq, or for candidate genes using quantitative PCR. These studies can monitor changes in genes involved in detoxification pathways (Colgan et al., 2019), growth and development (Castelli et al., 2020), or immune function (Brunner et al., 2014; Tritschler et al., 2017). Gene expression may also be assessed when pairing diet with a stressor to understand how diet and stress interact to impact overall health. For example, Brunner et al. (2014) measured how different diets impacted bumble bee response to the gut parasite Crithidia bombi and found that lack of dietary protein impacted immune response, resulting in failure to produce antimicrobial peptides.

The lack of information currently available on wild bee gene expression patterns and genome characterization has likely been a limiting factor of using molecular and physiological methods in wild bee studies. These studies require that the genes underlying physiological processes are well-described, and thus a particular expression pattern is known to be associated with variation in levels of nutritional resources, detoxification, immunocompetence, or reproductive potential. While the function of some genes are well-conserved across insects (and arthropods in general), such as those involved in immune response, functional characterization is typically necessary to obtain baseline information about how genes in a particular species and at a particular developmental or physiological stage are responding to a given nutritional stressor (Palmer and Jiggins, 2015). Efforts are increasing to characterize genomes equally across insect taxa, and this will facilitate more in-depth studies concerning wild bee nutrition (Hotaling et al., 2021).

Rather than measuring expression of specific genes, it is also possible to measure levels of nutrient stores (proteins, lipids, and glycogen) in the bee's body, which can be measured through hemolymph (as a measure of readily available nutrients) (Stabler et al., 2015), or through body tissues (as a measure of long term nutrient storage) (Amsalem et al., 2015; Woodard et al., 2020). Measuring body nutrient levels can indicate the level of nutrient that was stored for use, and the excesses that may have been excreted (Lee, 2007).

Molecular methods may also be used to measure genetic diversity in bee populations by evaluating microsatellites, or non-coding regions of DNA that vary based on relatedness between individuals (Maebe et al., 2015). This technique would be useful when measuring a change in bee population numbers and diversity at field sites where supplemental forage is provided. Population dynamics can be measured within foraging distance of these resources and the surrounding areas to determine if bee populations are increasing due to more resources or if more bees are simply attracted to these enhanced habitats. Using this method with wild bee species would first require sequencing and development of microsatellite targets (see Khidr et al., 2014 for an example), so evaluating many bee species at once could be quite costly and time consuming.