Fertilizer and herbicide alter nectar and pollen quality with consequences for pollinator floral choices

Flower-visiting insects in agroecosystems forage on weeds exposed to agrochemicals that may compromise the quality of their floral resources. We conducted complementary field and greenhouse experiments to evaluate: 1) the effect of low concentrations of agrochemical exposure on nectar and pollen quality and 2) the relationship between floral resource quality and insect visitation. We found pollen amino acid concentrations were lower in plants exposed to low concentrations of herbicide, and pollen fatty acid concentrations were lower in plants exposed to low concentrations of fertilizer, while nectar amino acids were higher in plants exposed to low concentrations of either fertilizer or herbicide. Exposure to low fertilizer concentrations also increased the quantity of pollen and nectar produced per flower. The responses of plants exposed to the experimental treatments in the greenhouse helped explain insect visitation in the field study. The insect visitation rate correlated with nectar amino acids, pollen amino acids, and pollen fatty acids. An interaction between pollen protein and floral display suggested pollen amino acid concentrations drove insect preference among plant species when floral display sizes were large. We show that floral resource quality is sensitive to agrochemical exposure and that flower-visiting insects are sensitive to variation in floral resource quality.

Methods
Greenhouse Study Design
We selected seven plant species for our experiment (Cirsium vulgare, (Savi) Ten. Asteraceae, Epilobium hirsutum, L. Onagraceae, Filipendula ulmaria, (L.) Maxim. Rosaceae, Hypochaeris radicata, L. Asteraceae, Origanum vulgare, L. Lamiaceae, Phacelia tanacetifolia, Benth. Boraginaceae, and Plantago lanceolata, L. Plantaginaceae). These comprised six native perennial and one non-native annual herbaceous species (P. tanacetifolia), selected as pollinator-attractive (Clifford, 1962; Russo et al., 2022), and likely to be found on agricultural field edges in Europe. They represent a diverse group of plant families with regard to floral resource quality (Ruedenauer et al., 2019b; Vaudo et al., 2020; Zu et al., 2021).

To collect sufficient quantities of pollen and nectar for nutritional analyses from each species in each treatment, and to avoid interrupting natural patterns of insect visitation in the field, we conducted a concurrent greenhouse study. We collected 20 wild individuals of each perennial species in the spring of 2017 and planted them in pots with field soil in the greenhouse. The annual species (P. tanacetifolia) was planted in potting media in the greenhouse with 20 seeds (purchased from a regional seed supplier: QuickCrop Ireland ©) to each pot.

After the plants were established, we randomly assigned five individuals of each species to each treatment (see below). Treatments were applied with a watering can holding 10 litres of water, applied across the five individuals of each of the seven species once a week. The plants were also treated with an insecticidal/fungicidal product (SB Plant Invigorator ©) once a week to control pest outbreaks. The insecticidal treatment was applied evenly across all plant species and treatments.

The four experimental treatments were designed to simulate non-target agrochemical exposure on field edges: (1) control (20 L water), (2) run-off concentration of NPK fertilizer (in 10 L water plus 10 L untreated water), (3) low concentrations of herbicide (glyphosate in 10 L water plus 10 L untreated water), or (4) a combination treatment (same low concentrations of NPK in 10 L water and glyphosate in 10 L water). The treatments were mixed with 10 litres of water used for a foliar application once a week for three months. The first four weeks of application were the highest concentration, the second four weeks lower, and the last four weeks the lowest (Table 1). These applications were based on estimates of field-edge exposure; there is commonly a high concentration spring application of chemical fertilizer and herbicide, followed by decreasing exposure later in the growing season. Concentrations were selected using published studies of fertilizer run-off (Korsaeth & Eltun, 2000; Bertol et al., 2007; Craig & Mannix, 2009; Russo et al., 2020). Because glyphosate is not mobile in the groundwater, we based our highest glyphosate application on the US EPA’s Maximum Contaminant Level (MCL) for safe drinking water (United States Environmental Protection Agency, 2003). The highest concentration we applied was less than half the maximum level detected in (Silva et al., 2019), or roughly 7.6 % of a standard annual field application (1440 g/ha) (Dupont, Strandberg & Damgaard, 2018). Outside of this treatment regime, the plants in our experiments received only water.

In the greenhouse, we collected pollen and nectar daily between the hours 0600–1000. We collected sufficient quantities of pollen for nutritional analyses (8 to 16 mg per species in each treatment) from six of the seven species (all except O. vulgare, which produced very little pollen in the greenhouse), and sufficient quantities of nectar for nutritional analyses from three of the seven species. Cirsium vulgare, H. radicata, F. ulmaria, and P. lanceolata either did not produce nectar or had small inflorescences from which we were not able to obtain sufficient quantities of nectar. We collected nectar and pollen from greenhouse plants to avoid interrupting normal insect foraging behaviour in the field.

We counted the inflorescences from which we collected pollen and nectar on each collection day. Pollen was collected from dehisced anthers directly into Eppendorf tubes and transferred immediately to a -20°C freezer. For F. ulmaria, E. hirsutum, and P. tanacetifolia we collected whole anthers; while for H. radicata, C. vulgare, and P. lanceolata, we collected fresh pollen. We collected nectar with microcapillary tubes and measured the filled volume before transferring them to a -20°C freezer. We calculated the average amount of nectar per inflorescence. Because pollen was collected fresh and later dried for analysis, we measured the total dry weight of pollen divided by the total flowers sampled for each species and treatment at the end of the season (Table S1).

Field Study Design
We conducted a field experiment to measure the effects of non-target agrochemical exposure on plant growth and pollinator visitation from 2017–2018 in Dublin, Ireland (Russo et al., 2020). The study consisted of four experimental treatment plots (2 x 2m) replicated across eight sites over two years (four sites in 2017 and four different sites in 2018). The sites were located in urban Dublin and selected based on space availability in collaboration with businesses and research entities, as well as the absence of outside exposure to herbicide or fertilizer. Each plot contained the same plant community with equal densities of individuals of the same seven plant species as the greenhouse experiment (above). The same experimental treatments as described above for the greenhouse experiment, with the same concentrations of fertilizer and herbicide, were used in the field experiment in both years of the study (Table 1). Treatments were randomly assigned to plots within a site at the beginning of the season. For the purposes of this study, we were primarily interested in the pollinator visitation from the field experiment.

Once the plants in the field began to flower, we sampled insects that came in contact with the reproductive parts of the inflorescences for at least 1 s. On each sample day at each site, we collected flower-visiting insects on each flowering plant species at each plot for five minutes using an insect vacuum (total of 96 sample days, 623 date-plot-samples, or 2036 five-minute samples (approximately 170 h)). Each site was visited between 12–14 times for collections in both 2017 and 2018; the number of site collections varied due to variation in the timing of flowering between different sites. We sampled between the hours of 0700 and 1800 (84 % of the samples were collected between from 1000 to 1600). The order in which we visited sites, plots within sites, and species within plots was randomized during each sampling event. We also recorded the number of inflorescences of each species during each sampling event. Insect species that could be identified in the field (specifically Apis mellifera, Linneaus Apidae, Episyrphus balteatus, De Geer Syrphidae, Bombus pascuorum, Scopoli Apidae, B. lapidarius, Linnaeus, and B. pratorum, Linnaeus) were released alive at the end of the sampling event. Collected specimens were transferred to a freezer and identified at the end of the field season [53,54]. Bee identifications were verified by Dr. Úna Fitzpatrick of the National Biodiversity Data Centre (Waterford, Ireland), while hoverfly specimens were identified by Dr. Martin Speight (Trinity College Dublin, Ireland).

Chemical Analyses
We quantified amino acids in 3-6mg of pollen of each of six plant species and four treatments and analysed three subsamples of each pollen sample for amino acids (72 samples). We used high-performance liquid chromatography (HPLC) and a spectrum analyser to identify the peaks of the individual amino acids, and the area under the curve of the spectra corresponded to the quantity of individual amino acids (full description in Supplemental Materials). 

We quantified fatty acids in 5–10mg of pollen of each of six plant species and four treatments and analysed two subsamples of each pollen sample (48 samples) (Trinkl et al., 2020). The fatty acids were analysed via gas chromatography/mass spectrometry (GCMS, Supplemental Materials).

We quantified the amino acids and sugars of the nectar from three plant species and four treatments (12 samples) as described in Venjakob et al. (2020) at the University of Freiburg in Freiburg, Germany. The analysis of the nectar amino acids and sugars was carried out chromatographically with an HPLC system (Agilent Technologies 1260 Series; Agilent, Böblingen, Germany, Supplemental Materials).

Data Analysis
Our ultimate goal was to determine whether the treatments in the greenhouse resulted in changes in pollen and nectar quality and whether these changes corresponded to the changes in pollinator visitation we observed in the field. As such, we aggregated the field visitation data over time to each plant species in each treatment.

First, we tested for differences between treatments among plant species in terms of the concentrations of (a) pollen total amino acids (summed concentrations of all amino acids), (b) pollen total fatty acids (summed concentrations of all fatty acids), (c) number of different pollen fatty acids, (d) pollen production per flower, (e) nectar total amino acids, (f) nectar total sugars, and (g) nectar production per flower. We used generalized linear mixed effect models (GLMMs, R package “lme4”) with treatment as a fixed effect and subsample nested within plant species as the random effect (Bates et al., 2014). Note these are not true replicates because we pooled pollen and nectar across individuals of a species within treatments from the greenhouse to have a sufficient quantity to analyse. Instead, these numbers represent variation within and between samples relative to variation within our subsamples. We provide results from among species comparisons in the supplement (Table S2).

Next, we tested for a correlation between the pollen or nectar attributes, or between the pollen and nectar attributes and flower visitation in the subgroups laid out above for data aggregated at the species and treatment level. We used visitation rate (abundance of visitors in a given sample divided by the size of the floral display (inflorescence size*number)) as a normalized measure for comparing visitation among plant species with variable floral displays. When visitation rate increases, it indicates a per floral unit preference (Russo et al., 2019b, 2020). We also tested the relationship between visitation rate and the ratio of proteins:lipids in the pollen (here pollen amino acids vs pollen fatty acids).

We separately evaluated the following groups of flower-visiting insects: (1) pollen-collecting bees (females of non-parasitic species, 1320 observations), (2) all bees (1755 observations), (3) bumblebees (1178 observations), (4) honeybees (386 observations), (5) hoverflies (Syrphidae, 677 observations), and (6) all flower-visiting insects (2567 observations). We hypothesized pollen-collecting bees would be most sensitive to pollen quality because they are provisioning offspring, and that bumblebees would be sensitive to protein:lipid ratios in the pollen as found in previous studies (e.g. (Vaudo et al., 2016a; Russo et al., 2019b)).

Next, we tested whether any of the attributes of pollen or nectar significantly improved the fit of the visitation data in the field, compared to published models of pollinator visitation (Russo et al., 2020). These tested whether pollinator visitation was influenced by pollen/nectar quality beyond previously established variables. We used a model selection process, choosing the model with the lowest AICc (function dredge in the package “MuMin” (Barton, 2009)). The site and plant species were treated as random effects, while the floral display and experimental treatment were fixed effects (Table 2 for full model structures). We then tested for interactions between the fixed effects in the model with the lowest AICc, and removed fixed effects that were not significant. We reported the marginal and conditional R2 for all models. Because the field visitation data were zero-inflated, we ran two sets of models. First, we ran a model with a binary presence/absence response variable. Next, we ran a model using only samples where flower-visiting insects were recorded, with insect abundance as the response variable.

农业生态系统中访花昆虫会取食暴露于农用化学品（agrochemicals）下的杂草，而这些农用化学品可能会损害其花类资源的质量。本研究开展了互补的野外与温室实验，以评估两个研究目标：1）低浓度农用化学品暴露对花蜜与花粉质量的影响；2）花类资源质量与昆虫访花行为之间的关联。我们发现，暴露于低浓度除草剂的植物其花粉氨基酸浓度更低，暴露于低浓度化肥的植物其花粉脂肪酸浓度更低，而暴露于低浓度化肥或除草剂的植物其花蜜氨基酸浓度更高。低浓度化肥暴露还会提升每朵花产生的花粉与花蜜量。温室实验中植物对实验处理的响应，能够帮助解释野外研究中观察到的昆虫访花行为。昆虫访花速率与花蜜氨基酸、花粉氨基酸及花粉脂肪酸均存在相关性。花粉蛋白与花展示量之间的交互作用表明，当花展示规模较大时，花粉氨基酸浓度是驱动不同植物物种间昆虫偏好的关键因素。本研究证实，花类资源质量对农用化学品暴露敏感，且访花昆虫对花类资源质量的变化同样敏感。

### 方法
#### 温室实验设计
本研究选取7种植物开展实验：蓟（*Cirsium vulgare*, (Savi) Ten. 菊科Asteraceae）、柳叶菜（*Epilobium hirsutum*, L. 柳叶菜科Onagraceae）、绣线菊（*Filipendula ulmaria*, (L.) Maxim. 蔷薇科Rosaceae）、猫耳菊（*Hypochaeris radicata*, L. 菊科Asteraceae）、牛至（*Origanum vulgare*, L. 唇形科Lamiaceae）、天蓝紫草（*Phacelia tanacetifolia*, Benth. 紫草科Boraginaceae）与长叶车前（*Plantago lanceolata*, L. 车前科Plantaginaceae）。其中包含6种本土多年生草本植物与1种非本土一年生草本植物（*P. tanacetifolia*），这些植物均被证实可吸引传粉者（Clifford, 1962; Russo et al., 2022），且常见于欧洲农田边缘。它们在花类资源质量方面涵盖了多个不同的植物科属（Ruedenauer et al., 2019b; Vaudo et al., 2020; Zu et al., 2021）。

为了从每个处理组的各物种中获取足够数量的花粉与花蜜用于营养分析，同时避免干扰野外昆虫的自然访花行为，我们同步开展了温室实验。2017年春季，我们采集了每个多年生物种的20株野生个体，种植于温室中的田间土壤盆钵内。对于一年生物种*P. tanacetifolia*，我们向每个盆钵中播种20粒种子（购自爱尔兰本地种子供应商QuickCrop Ireland ©），种植于温室育苗基质中。

植株定植后，我们将每个物种的5个个体随机分配至各处理组（详见下文）。使用装有10L水的喷壶，每周对7个物种的每组5株植株施加一次处理液。同时，每周喷施一次杀虫/杀菌剂（SB Plant Invigorator ©）以控制虫害，该杀虫处理均匀施加于所有植物物种与处理组。

本研究的4种实验处理旨在模拟农田边缘的非靶标农用化学品暴露：(1) 对照组（仅20L水）；(2) NPK化肥径流浓度处理（10L水中添加化肥，再加入10L去离子水）；(3) 低浓度除草剂处理（10L水中添加草甘膦，再加入10L去离子水）；(4) 复合处理（10L水中添加低浓度NPK化肥，同时10L水中添加低浓度草甘膦）。每周以叶面喷施的方式施加10L处理液，持续3个月。处理浓度分为三个阶段：前4周为最高浓度，中间4周为中等浓度，最后4周为最低浓度（详见表1）。该施加方案基于农田边缘暴露场景的估算：农田通常在春季施用高浓度化肥与除草剂，生长季后期暴露浓度逐渐降低。浓度设置参考了已发表的化肥径流相关研究（Korsaeth & Eltun, 2000; Bertol et al., 2007; Craig & Mannix, 2009; Russo et al., 2020）。由于草甘膦在地下水迁移性较弱，本研究最高草甘膦施用量参考了美国环境保护署（US EPA）规定的饮用水安全最大污染物浓度（Maximum Contaminant Level, MCL）（United States Environmental Protection Agency, 2003）。本研究施加的最高浓度不到Silva等人（2019）检测到的最高浓度的一半，约为标准年度农田施用量（1440 g/ha）的7.6%（Dupont, Strandberg & Damgaard, 2018）。除上述处理方案外，实验植株仅接受清水浇灌。

在温室中，我们于每日06:00–10:00采集花粉与花蜜。我们从7个物种中的6个（除*O. vulgare*外，该物种在温室中花粉产量极低）采集到足够用于营养分析的花粉量（每个处理组的每个物种需8–16 mg），并从7个物种中的3个采集到足够的花蜜样本。*Cirsium vulgare*、*H. radicata*、*F. ulmaria*与*P. lanceolata*要么不产花蜜，要么花序过小无法获取足够的花蜜样本。我们从温室植株中采集花粉与花蜜，以避免干扰野外昆虫的正常访食行为。

我们记录了每个采样日采集花粉与花蜜的花序数量。花粉直接从开裂的花药中采集至Eppendorf管中，并立即转移至-20℃冰箱保存。对于*F. ulmaria*、*E. hirsutum*与*P. tanacetifolia*，我们采集整个花药；对于*H. radicata*、*C. vulgare*与*P. lanceolata*，我们直接采集新鲜花粉。我们使用毛细管采集花蜜，在转移至-20℃冰箱前先测量花蜜充盈体积，并计算每个花序的平均花蜜量。由于花粉为新鲜采集后干燥用于分析，我们在实验季末计算了每个物种与处理组的总花粉干重除以采样总花数（详见表S1）。

#### 野外实验设计
2017–2018年，我们在爱尔兰都柏林开展了野外实验，以评估非靶标农用化学品暴露对植物生长与传粉者访花行为的影响（Russo et al., 2020）。本研究包含4个2×2m的实验处理样地，在两年间于8个样点重复开展（2017年4个样点，2018年更换为另外4个样点）。样点位于都柏林市区，基于空间可用性、与企业及研究机构的合作情况，以及无外源除草剂或化肥暴露的条件进行选取。每个样地包含与温室实验完全一致的植物群落，7个物种的个体密度相同。两年的野外实验均使用与温室实验相同的实验处理与化肥、除草剂浓度（详见表1）。实验季初期，我们在每个样点内将处理随机分配至各样地。本研究主要关注野外实验中的传粉者访花数据。

当野外植株开始开花后，我们采集与花序生殖部位接触至少1秒的昆虫。在每个样点的每个采样日，我们使用昆虫真空采集器对每个样地中每个开花物种的访花昆虫进行5分钟的采样（总计96个采样日，623个日期-样地样本，或2036个5分钟采样样本，累计约170小时）。2017与2018年，每个样点的采样次数为12–14次，采样次数因不同样点的开花时间差异而有所变化。我们的采样时间为07:00–18:00（其中84%的样本采集于10:00–16:00）。每次采样事件中，我们对样点、样点内的样地以及样地内的物种的访问顺序均进行随机化处理。我们同时记录了每次采样事件中每个物种的花序数量。在野外可直接识别的昆虫物种（具体包括西方蜜蜂*Apis mellifera*, Linnaeus 蜜蜂科Apidae、黑带食蚜蝇*Episyrphus balteatus*, De Geer 食蚜蝇科Syrphidae、红三叶草熊蜂*Bombus pascuorum*, Scopoli 蜜蜂科Apidae、钝尾熊蜂*B. lapidarius*, Linnaeus与小峰熊蜂*B. pratorum*, Linnaeus）在采样结束后被原地释放。采集的标本被转移至冰箱保存，并在野外季末完成鉴定[53,54]。蜂类鉴定由爱尔兰国家生物多样性数据中心的Úna Fitzpatrick博士完成，食蚜蝇标本则由都柏林圣三一学院的Martin Speight博士鉴定。

#### 化学分析
我们对6个植物物种的4个处理组的3–6mg花粉样本进行氨基酸定量分析，并对每个花粉样本的3个子样本进行氨基酸检测（共72个样本）。我们使用高效液相色谱（high-performance liquid chromatography, HPLC）与光谱分析仪识别单个氨基酸的峰，光谱曲线下的面积对应单个氨基酸的含量（详细方法见补充材料）。

我们对6个植物物种的4个处理组的5–10mg花粉样本进行脂肪酸定量分析，并对每个花粉样本的2个子样本进行检测（共48个样本）（Trinkl et al., 2020）。脂肪酸通过气相色谱/质谱联用仪（gas chromatography/mass spectrometry, GCMS）进行分析（详见补充材料）。

我们参考Venjakob等人（2020）的方法，在德国弗莱堡大学弗莱堡分校对3个植物物种的4个处理组的花蜜样本中的氨基酸与糖分进行定量分析（共12个样本）。花蜜氨基酸与糖分的分析通过HPLC系统（Agilent Technologies 1260 Series; Agilent, Böblingen, Germany）完成（详见补充材料）。

#### 数据分析
本研究的最终目标是明确温室实验中的处理是否会导致花粉与花蜜质量发生变化，以及这些变化是否与我们在野外观察到的传粉者访花变化相对应。因此，我们将野外访花数据按时间汇总至每个处理组的每个植物物种。

首先，我们针对以下指标检验不同处理组间植物物种的差异：(a) 花粉总氨基酸浓度（所有氨基酸浓度之和）、(b) 花粉总脂肪酸浓度（所有脂肪酸浓度之和）、(c) 花粉中不同脂肪酸的种类数、(d) 每朵花的花粉产量、(e) 花蜜总氨基酸浓度、(f) 花蜜总糖分浓度、(g) 每朵花的花蜜产量。我们使用广义线性混合效应模型（generalized linear mixed effect models, GLMMs，R包"lme4"），以处理为固定效应，以植物物种内嵌套的子样本为随机效应（Bates et al., 2014）。请注意，这些并非真正的重复样本，因为我们将温室实验中同一处理组内同一物种的多个个体的花粉与花蜜混合，以获取足够的分析样本量。上述数值仅反映了子样本内部以及样本间的变异。我们在补充材料中提供了物种间比较的结果（表S2）。

接下来，我们在上述按物种与处理组汇总的数据子集中，检验花粉或花蜜属性之间，或花粉/花蜜属性与花访率之间的相关性。我们使用访花速率（给定样本中的访花者数量除以花展示规模（花序大小×数量））作为标准化指标，以比较不同花展示规模的植物物种间的访花情况。当访花速率升高时，表明该物种的单位花偏好性更强（Russo et al., 2019b, 2020）。我们同时检验了访花速率与花粉中蛋白质:脂质比值（即花粉氨基酸浓度与花粉脂肪酸浓度之比）之间的关系。

我们分别对以下六类访花昆虫进行了分析：(1) 采粉蜂类（非寄生性物种的雌性个体，共1320次观测）、(2) 所有蜂类（共1755次观测）、(3) 熊蜂属昆虫（共1178次观测）、(4) 西方蜜蜂（共386次观测）、(5) 食蚜蝇科昆虫（Syrphidae，共677次观测）、(6) 所有访花昆虫（共2567次观测）。我们假设采粉蜂类对花粉质量最为敏感，因为它们需要为后代储备食物；同时熊蜂属昆虫对花粉的蛋白质:脂质比值较为敏感，这与此前的研究结果一致（例如Vaudo et al., 2016a; Russo et al., 2019b）。

接下来，我们检验相较于已发表的传粉者访花模型（Russo et al., 2020），花粉或花蜜的任何属性是否能够显著提升野外访花数据的模型拟合度。这一检验旨在明确传粉者访花是否会受到花/花蜜质量的额外影响，而不仅仅受已确立的变量影响。我们使用模型选择流程，选择具有最低AICc值的模型（R包"MuMIn"中的dredge函数（Barton, 2009））。样点与植物物种作为随机效应，花展示规模与实验处理作为固定效应（完整模型结构详见表2）。随后，我们对AICc值最低的模型中的固定效应间的交互作用进行检验，并移除无显著性的固定效应。我们报告了所有模型的边际R²与条件R²。由于野外访花数据存在零膨胀问题，我们运行了两组模型：首先，我们构建了以存在/缺席为二元响应变量的模型；其次，我们仅使用记录到访花昆虫的样本，以昆虫数量作为响应变量构建模型。