技术创新为内容建设添翼

1. The objectives of studies employing pollination syndromes

In the 346 papers reviewed, pollination syndromes were used in a broad variety of contexts and questions. I scored the major objectives of each study (each study may have multiple objectives; hence percentages of categories do not sum to 100%). The majority of studies aimed at generally describing the study system’s pollination syndrome (52.6%), testing the concept (33.8%) and experimentally assessing, for example, the breeding system or relative effectiveness of different pollinators (30.3%). Here 21.7% of studies investigated whether pollination syndromes correlated with other factors such as the availability of flowering resources and 17.3% used syndromes to analyse whether and how pollinators affect/correlate with clade diversification. Pollination syndromes were also used to address conservation issues (7.8%), modelling and statistical prediction of pollinators (3.8% and 2.3%, respectively); 5.8% sought to better understand the evolution of pollination syndromes at the genomic level (e.g. evolution of colour genes).

2. The problem of scale

Traditional pollination syndromes were developed as a broad, functional–ecological classification system that would be applicable across angiosperms. At this broadest possible taxonomic scale, it follows that more general categorisations (e.g. widely distributed floral traits, which can be scored for all taxa) are used. It is not surprising that researchers have found mismatches between these idealised, traditional syndromes and the trait combinations realised in actual flowers (Ollerton et al., 2009). Furthermore, it has become clear that pollination syndromes differ in different parts of the world (e.g. the ‘bat’ and ‘bird’ syndrome in the Old and New world tropics are characterised by different trait combinations, Fleming et al., 2009). Unless the aim of a study is to test whether a (group of) plant species fits into ‘traditional’ syndromes, the overall conclusion was that studies at lower taxonomic scales (e.g. within families) should focus on taxon-specific traits to accurately describe syndromes in the group and region under study (Ollerton et al., 2009).

The 346 papers screened indeed addressed pollination syndromes at various taxonomic scales (Fig. 1a). The majority, 59.8%, worked at small taxonomic units (single or few species), 21.4% addressed questions at generic, tribal or family level, 15.9% worked at the community level and the remaining 2.9% worked at other/higher scales. In total, 244 studies (70.5%) included floral traits. The majority of these (77.5%) recorded one or several of the traits included in traditional syndromes (as listed by Rosas-Guerrero et al., 2014; Table S2; Fig. 1b); 90.6% included one or several additional traits such as quantitative measurements of nectar, scent, colour or flower morphology, pollen grain features, pollen to ovule ratios or shape of nectaries. Of the c. 10% of studies that did not include additional traits, the majority (37.8%) were community-level studies including many different taxa, followed by studies on single taxa or studies at the generic/tribal/family level (29.7% each).

These results suggested that particularly researchers working at small taxonomic scales have heeded the call to use more refined trait datasets to evaluate pollination syndromes. Ultimately, the lack of a single ‘scale’ at which pollination syndromes operate may turn out to be a strength rather than a weakness. As long as authors are clear about the scale at which they work, it offers the necessary flexibility to adjust character coding to functionally meaningful traits of the specific study system. In the long run, this may even prove valuable in understanding universal patterns of pollinator-mediated selection and angiosperm evolution (and hence angiosperm-wide pollination syndromes) as new traits and pollinator combinations hitherto not considered may emerge as important (also see section II.7 ‘There is more to pollination syndromes than a handful of floral traits’).

3. The predominant focus on (few) specialised systems

Obviously, pollination syndromes were conceived for lineages comprising taxa with markedly diverging flower traits specialised on different functional pollinator groups. This focus on specialised systems came in for criticism as a large number of angiosperms indeed is generalised on being pollinated by several functional groups (Waser et al., 1996).

In the literature reviewed, species from 76 different plant families were considered (Fig. 2; Table S1; note that I excluded community-level studies and studies at higher taxonomic or nonsystematic scales from this count). This represents 18.9% of the 416 angiosperm families currently recognised (The Angiosperm Phylogeny Group et al., 2016). The plant families studied were mostly groups well known for enigmatic, specialised pollination syndromes, such as Orchidaceae (24 studies, Monocots), Gesneriaceae (17 studies), Solanaceae (i.e. Petunia, 17 studies), Plantaginaceae (i.e. Penstemon, 14 studies; all Lamiids) and Fabaceae (13 studies, Fabids). Some of the largest families, like Asteraceae, were only represented by a few (two) studies, and families known for generalised pollination systems such as Apiaceae or Brassicaceae were not included at all.

Quantifying our potential perception bias arising from the focus on (few) lineages with ‘exciting’ pollination biology will only be possible in future once we know more about hitherto neglected lineages and systems with seemingly ‘boring’ (often generalised) flowers (Vasconcelos et al., 2019). Albeit not included in the literature reviewed, recent studies on generalised systems do exist (e.g. Gómez et al., 2014, 2016; Zych et al., 2019). These have documented geographic patterning in floral traits, and could partially link it with adaptation to the local pollinator community. Studies performed under experimental conditions could further demonstrate strong selection on scent mediated by a subset of the pollinator community even in seemingly generalised flowers (Gervasi & Schiestl, 2017). Investigating how, why and when generalised systems, and perhaps a ‘generalist syndrome’, are maintained despite specialisation on the local pollinator community will be one important approach to acquiring a more holistic understanding of the mechanisms underlying syndrome formation (Aigner, 2006). Furthermore, studying large plant groups with phenotypically ‘boring’ flowers is essential to understand potential intrinsic (system specific) and extrinsic (e.g. abiotic and biotic environment) constraints which may have prevented floral diversification despite species diversification. Finally, the investigation of ‘new’ plant groups whose pollination biology is understudied may lead to exciting discoveries of novel pollination systems and associated floral trait changes such as pollination by marine invertebrates (van Tussenbroek et al., 2016) or pollinator shifts in functionally highly specialised systems (Dellinger et al., 2019a).

4. Pollinator observations and classifying pollinators into ‘functional groups’

Different pollinator species are pooled into ‘functional groups’ based on how they interact with flowers (Fenster et al., 2004). Such groups may include a single species or many different species (Armbruster, 2017). Again, while it may be useful to work with the traditional categories (sensu Rosas-Guerrero et al., 2014) at high taxonomic levels, the majority of studies investigates syndromes at lower taxonomic levels (Fig. 1a) and may find a more refined categorisation useful (Murúa & Espíndola, 2015). Only empirical observations allow us to correctly understand flower–pollinator interactions and hence devise meaningful functional categories (if necessary), disregarding taxonomic affinities (Fenster et al., 2009a). How many current studies are indeed based on empirical pollinator observations and how have they classified pollinators?

In total, 315 of the 346 studies used functional groups for analyses. Of these, 54.9% had performed pollinator observations in the specific study and 32.4% were based on pollinator observations from the literature (Fig. 1c). Only 7% relied on inferences made solely based on floral traits (syndromes) and the remaining 5.7% used combinations of observations, literature reports and syndrome inference. Here, 83.3% of the studies based on direct pollinator observations (a total of 173) worked at low taxonomic levels (one or a few species) and primarily aimed at describing and testing/verifying pollination syndromes (73.3%). Most of these studies summarised pollinators into traditional functional groups despite identifying pollinators at the lowest taxonomic level possible (often species). Only 4.4% of studies created unusual functional groups based on observed or presumed behavioural or morphological differences between pollinators, such as phytophagous vs dung beetles (Schiestl & D?tterl, 2012), 15 insect groups (Koski et al., 2015) or straight-bill and sickle-bill hummingbirds (Lagomarsino et al., 2017).

These results showed that a large proportion of (particularly small-scale) studies indeed verified syndromes through empirical pollinator observations and only a small fraction of studies worked without empirical data. This does not mean, however, that we may reduce efforts to document pollinators. Several of the large-scale studies included in this review suffered from deficient pollinator observations and hence could not systematically analyse flower character evolution in relation to pollination (e.g. Reginato & Michelangeli, 2016; Heckenhauer et al., 2018; Vasconcelos et al., 2019). This observation is particularly important in the light of modern sequencing techniques and online search engines, which allow for the fast collation of large datasets, but leave us uninformed of the functional–ecological implications of flower mode (Ollerton et al., 2015).

Regarding classification of functional groups, most studies adhered to traditional or taxonomic groups to summarise observed patterns. It is unclear whether a different grouping would have led to different conclusions. However, the fact that many authors have identified pollinators to species is promising, as it allows more refined assessments of flower–pollinator relationships (Murúa & Espíndola, 2015; Pichler et al., 2020). In the past, only a few authors have made use of the flexibility of the functional-group idea in that it allows (and actually demands!) grouping based on the interaction differences observed in the study system, irrespective of taxonomic affinities of the pollinators. For example, Wilson et al. (2004) grouped individuals of the same bee species into nectar foragers vs pollen foragers and Armbruster (1993) differentiated fragrance-collecting male bees from resin-collecting female euglossine bees. At the other perspective, pollinators that we intuitively perceive as different may sometimes function as one group (e.g. nectar foraging hummingbirds, bats, rodents, flower piercers; Dellinger et al., 2019a).

Generally, we have to be aware of the fact that any typological categorisation, such as functional groups and syndromes, leaves us at the boundaries of subjective interpretation of continuous, dynamic natural systems. While functional groups have been shown to be highly useful to summarise patterns, I urge that we continue to move on from typology to the (objective) quantification of syndromes when possible. I propose two approaches towards delineating functional pollinator groups and making classification less typological. The first approach starts with the pollinators. Quantifying the contribution of (different) flower visitors to pollination is the most powerful tool for understanding their importance as pollinators (see section II.5 ‘Visitation by multiple functional groups and selection by the most efficient pollinator’). Pollination efficiency may be combined with measurements of pollinator morphology (e.g. length of mouthparts). These matrices may be combined with floral trait matrices to search for convergence patterns (Guzmán et al., 2017, Pichler et al., 2020) and may then be evaluated in the light of known or observed sensory/behavioural differences of pollinators to devise functional groups. Manipulative experiments may be conducted subsequently to identify the factors which render a specific set of taxa a ‘functional’ group and when and how this group may be narrowed or widened (e.g. through trait manipulation restricting pollen transfer). The second approach starts from the flowers. Particularly when working at macroevolutionary scales, authors may not want to classify flowers a priori but rather use statistical methods (Table 2) to detect patterns of floral trait convergence, and only then relate these to pollinator data (Pichler et al., 2020).

5. Visitation by multiple functional groups and selection by the ‘most efficient pollinator’

When formulating the ‘most effective pollinator principle’, Stebbins (1970) emphasised that ‘character syndromes’ (pollination syndromes) do not preclude the existence of secondary, potentially less efficient pollinators. Indeed, most flowers are visited by multiple functional pollinator groups (e.g. Waser et al., 1996; Ollerton et al., 2007). One major advancement that pollination-syndrome research has seen was to acknowledge this fact and to start measuring the relative contribution to pollination (pollination efficiency) of (diverse) pollinator assemblages instead of only documenting their flower visitation (Fenster et al., 2004; Rosas-Guerrero et al., 2014; Santiago-Hernández et al., 2019). In specialised, and even in many generalised, flowers (which are easily accessible to various different pollinator groups), only a subset of all visitors was found to be effective as pollinators (Barto? et al., 2005; Rosas-Guerrero et al., 2014). Recording pollinator efficiency requires more time and potentially laboratory equipment than assessing visitation only. How many authors have employed more refined measures to identifying ‘most efficient pollinators’ and what have they found? Does visitation by multiple visitors make syndrome predictions less accurate?

Of the 173 studies based on direct observations, the majority (63.2%) included both an evaluation of visitation rate and pollen transfer efficiency (Fig. 1d). In total, 18.1% only documented visitation rates, 11.1% only documented pollen transfer and 7.6% only reported types of pollinators. As expected, the majority of studies (61.2%) found more than two functional groups visiting; 13.4% reported visitation by two groups and 25.4% documented only one functional group visiting (Fig. 1e). The vast majority of studies documenting two or more functional groups as visitors reported only a subset of these visitors as pollinators (68.8% and 93.8%, respectively); 71.4% of studies reported the most frequent visitor also being the most important pollinator (but also see King et al., 2013); 122 studies formulated a hypothesis about an expected pollinator/visitor, based either on previous observations or expectations derived from floral syndrome traits (see below). Here, 85.2% could at least partially confirm these expectations, while 13.9% found other pollinators/visitors than expected and one study did not find any legitimate pollinator. Syndrome expectations were confirmed or contradicted equally often, regardless of whether multiple functional groups or only one was visiting (83.5% at least partly confirmed syndrome expectation with multiple functional groups visiting and 90% when one functional group was visiting).

These results underline findings reported previously: visitation by multiple functional pollinator groups is common but does not negatively impact the predictive power of pollination syndromes. It will be interesting, however, to further evaluate the impact of secondary/less efficient pollinators on floral trait evolution. Generally, pollination by multiple functional pollinator groups is believed to be disadvantageous, as it may induce strong fitness costs through adaptive trade-offs (Muchhala, 2008; Pady?áková et al., 2013; Strelin et al., 2017). It seems, however, that not all pollinator combinations are equally costly and trade-offs may sometimes remain small (Aigner, 2006; Ashworth et al., 2015). An exciting and unanswered question in this context is whether some pollinator combinations (e.g. hummingbirds and bats, see Lagomarsino & Muchhala, 2019; Dellinger et al., 2019b) are more likely to evolve into stable bimodal systems (i.e. systems with intermediate adaptations to two distinct functional pollinator groups sensu Goldblatt and Manning, 2006) than others (e.g. hummingbirds and bees, Castellanos et al., 2004)? Why and how is bimodality maintained and under which conditions does generalisation evolve?

Furthermore, despite the predominance of pollinators as selective agents on flowers (Caruso et al., 2018), other biotic interactions (e.g. herbivores, competition) are also important (Jogesh et al., 2017). In the reviewed literature, only eight studies addressed and quantified other agents of selection than pollinators: six studied floral herbivory and two addressed reward robbing. Particularly when pollinators and antagonists are attracted through the same traits (e.g. large, colourful displays, Brothers & Atwell, 2014), this may induce conflicting selection on these traits and potentially lead to the evolution of defence/deterrence traits (Doubleday et al., 2013). Incorporating antagonistic interactions (i.e. through fitness estimates) and potential defence traits into studies on pollination syndromes is important since defence traits may promote (preadaptations) or constrain shifts between pollination syndromes (Hanley et al., 2009) and may also explain mismatches between traditional and observed syndromes (see section II.8 ‘Not all floral traits are equally informative’).

6. Prediction of pollinators

Particularly in large plant clades in which time, logistic and budgetary constraints limit the number of species for which pollinators can be observed empirically, pollination syndromes are used for predicting pollinators (Lagomarsino et al., 2017; Dellinger et al., 2019a). Several authors have argued, however, that such predictions are risky, particularly when no (or few) pollinator observations exist for the respective plant group, when multiple functional groups visit flowers or when only very few floral traits are considered (Ollerton et al., 2009; Fenster et al., 2009a). In small-scale studies, pollination syndromes are often used to formulate a hypothesis about a species’ most likely pollinator that is then evaluated through measurements of pollinator efficiency. I extracted this information from the literature to assess how often authors verified/falsified their hypothesis and to understand which floral traits led them to their specific hypothesis.

In total, 64 studies hypothesised which functional group would pollinate the study species based on syndrome traits. Of those, 27.8% expressed doubt on classification into clear-cut syndromes because they observed floral trait combinations commonly assigned to different syndromes (Fig. 3a). Of those, 85% correctly predicted visitation by mixed assemblages, while 15% found different pollinators than expected from syndrome traits (Fig. 3b,c). Studies that did not express uncertainty about syndrome classification could (at least partially) confirm predictions in 86.3% of cases and 13.6% of studies reported other pollinators than expected (Fig. 3b,c). In this regard, 58 studies substantiated their syndrome expectations by naming one to three (traditional) syndrome traits in the introduction or materials and methods sections: colour (55.2%), reward (34.5%), scent (31.0%) and flower shape (25.9%) being the most commonly used traits (Fig. 4; Table S2).

These results demonstrated that authors could predict pollinators with high accuracy even when basing predictions on only a few traits. Importantly, these traits also allowed authors to correctly identify species that were likely to be pollinated by multiple functional groups. How can these results be extrapolated to larger scales for which empirical verification of syndrome assumptions is not possible? Promising advancements have been made, particularly in the field of Machine Learning (Pichler et al., 2020). A series of powerful and mathematically flexible algorithms exists that may first be trained and evaluated on species for which empirical pollinator observations exist, and then be used to predict pollinators for (closely) related species that lack pollinator observations (Table 1; Dellinger et al., 2019a; Pichler et al., 2020).

7. There is more to pollination syndromes than a handful of floral traits

Several authors have criticised traditional pollination syndromes in being limited to a relatively narrow set of traits that are easy and fast to code (Fenster et al., 2009a; Abrahamczyk et al., 2017). This is particularly problematic in lineages in which pollination syndromes are best characterised by traits not included in traditional syndromes (Table 1; Dellinger et al., 2019a). Furthermore, quantitative traits such as colour, scent or shape are difficult to discretise and may lead to subjective, erroneous categorisation (Abrahamczyk et al., 2017). How many and which floral traits are generally included in pollination-syndrome studies? And how many of those are quantitative?

Of the 244 studies that included floral traits, the majority (77.2%) was indeed based on a relatively small set of traits (1–10; Fig. 1b). Here, 66.8% of studies included quantitative measurements of reward or attraction traits (nectar: 36.4%, scent 19.7%, colour 13.9%); 22.1% measured corolla dimensions quantitatively and 20.5% measured both sterile and reproductive floral organs quantitatively. Only 3.3% of studies were based on morphometric assessments of flower shape and only two studies quantified morphometric relationships between both corolla and reproductive organs.

These results confirmed that pollination-syndrome studies are often limited to narrow, qualitative, rather than quantitative, sets of traits. While pollinators could obviously be successfully predicted even from only a few (qualitative) traits, I argue that pollination syndromes should not be limited to pollinator prediction. By definition, pollination syndromes are multivariate in nature and (may) carry information on complex, multitrait phenotypes (Junker & Parachnowitsch, 2015). The long-lasting focus on the same narrow set of traits prevents us from exploring new areas of phenotypic evolution and gaining a more profound understanding of how flowers function. Adding new traits to studies pays off, however. For example, the relatively recent addition of quantitative assessments of floral scent through headspace techniques and gas chromatography-mass spectrometry analyses has opened up a new ‘unseen’ world (Schiestl & D?tterl, 2012). As studies in different systems are compiled, we start to see patterns of scent compounds that are potentially relevant in the communication with different functional pollinator groups and floral antagonists (Junker & Parachnowitsch, 2015). Another fruitful potential addition to pollination syndromes could be petal cell shape. Several authors have reported conical cells associated with bee pollination and flat cells found in bird-pollinated plants (Papiorek et al., 2014; Dellinger et al., 2019a). Such findings merit further investigation to evaluate whether these patterns are system specific or possibly applicable across broad taxonomic scales.

The choice of appropriate methods and metrics for the collection of phenotypic data is crucial in all biological disciplines and often limited by time constraints, equipment or available (floral) material (Table 1). Coding for traits qualitatively, as is common in pollination-syndrome research, and is fast and informative particularly when analysing large numbers of disparate species. The discretisation of biological traits always bears risks, however (Wilson et al., 2007; Abrahamczyk et al., 2017). First, discretisation may be highly subjective and circular in that categories are established for patterns (e.g. syndromes) one expects to find. Second, while character coding may be accurate when performed by the same observer, it may be difficult to repeat for a different observer. And third, trait categorisation misses the recording of within-taxon variation, which is essential to understand the adaptive potential of populations or direction/strength of selection. Thus, authors should critically evaluate whether and which traits they can record quantitatively in order to arrive with objective, comparable and repeatable datasets. Quantitative trait measurements are nowadays increasingly used for advertisement traits (Table 1; Fig. 1b) but still relatively uncommon when studying flower–pollinator fit traits. Morphometrics provides a useful approach to quantifying fit traits (Pérez et al., 2007; Kaczorowski et al., 2012; Pérez-Barrales et al., 2014; Strelin et al., 2017; Wilson et al., 2017; Smith & Kriebel, 2018). To date, most authors have considered flowers in only two dimensions and focused on corolla shape only (e.g. Kaczorowski et al., 2012; Gómez et al., 2016; Smith & Kriebel, 2018; Benítez-Vieyra et al., 2019). It is high time, however, that we start studying flowers (and their pollinators) as what they are: highly three-dimensional structures. The relative position of the corolla and the reproductive organs is decisive in mediating fit with the pollinators and assuring pollen transfer. Recent advances in obtaining detailed 3D-flower models through high-resolution X-ray computed tomography (HRXCT) now provide the necessary basis for studying flower shape variation through landmark-based geometric morphometrics without losing information on the spatial arrangement of floral organs (van der Niet et al., 2010; Sedeek et al., 2014; Wang et al., 2015; Wilson et al., 2017; Dellinger et al., 2019c; Kellenberger et al., 2019; Reich et al., 2020). A wide array of methods is available and constantly being expanded, allowing for statistical analyses of multivariate quantitative trait datasets (Table 2). If, finally, authors choose to rather work with qualitative traits, they need to establish objective criteria for trait discretisation which may also be employed by future researchers working on the same system (Wilson et al., 2017; Dellinger et al., 2019a). To this end, categorisation may outperform quantification when working with morphologically or functionally complex composite traits which cannot be expressed by simple measurements (e.g. pollen-expulsion mechanism, Dellinger et al., 2019a).

8. Not all floral traits are equally informative

Not all floral traits are affected equally by (pollinator-mediated) selection (Pérez et al., 2007; Pérez-Barrales et al., 2014; Fenster et al., 2015, Caruso et al., 2018), and not all traits are equally important in each pollination syndrome (Johnson, 2013; Zung et al., 2015; Dellinger et al., 2019a). Despite this knowledge, traditional pollination syndromes often lump together all traits without further analysing differences in specificity and function. Are there (traditional syndrome) traits that are commonly recorded but bear little predictive power?

In total, 77 of the 97 studies estimating pollinator efficiency recorded floral traits. The traits most commonly assessed were (mostly nectar) reward (62.3%), timing or duration of anthesis/sexual phase (48.1%), scent (35.0%), corolla length (32.5%), corolla width (29.9%) and colour (29.8%; Fig. 5a; Table S3). Interestingly, 56 of these studies reported some traits as being particularly informative in characterising the specific pollination syndrome or in discriminating visitors from pollinators (Fig. 5b; Table S4). Nectar (39.3%), colour (30.4%), scent (26.8%), flower size (19.6%) and anthesis patterns (16.1%) were reported as important traits most often. In addition, 25 studies stated which traits were uninformative or misleading in syndrome classification (Fig. 5c; Table S5), with colour (52.0%) being most uninformative/confounding (i.e. contradicting traditional syndrome expectations), followed by scent (32.0%), shape (24.0%) and reward (16.0%).

Clearly, these results have to be treated with care, as traits have not been recorded/analysed the same way. However, it is noteworthy that nectar rewards are assessed most often and reported as highly informative, while colour is assessed only half as often but reported as most uninformative/misleading (Fig. 5d). Flower colour undoubtedly is highly important in many systems (Dudash et al., 2011; Johnson, 2013; Rodrigues et al., 2018), but not in all. Selection studies have shown, for example, that flower colour is not selected by pollinators, while display size and plant height are (Silene, Brothers & Atwell, 2014; Fenster et al., 2015). Indeed, as an advertisement trait, colour may also underlie selection by antagonists (avoidance of herbivores, Camargo et al., 2018) or pollinator competition (Muchhala et al., 2014). It would be too early (and most likely wrong) to dismiss any trait from pollination syndromes. Instead, I argue that we should focus on traits which do and do not match syndrome expectations alike. Trait mismatch may come from complex biotic interactions, but also be a mere consequence of a species’ phylogenetic background (developmental/genetic constraints) and the limited number of trait combinations that could evolve (O’Meara et al., 2016). By systematically evaluating the relative importance of a large number of traits and quantifying (mis)match with traditional syndromes, we may better pinpoint the traits important in adaptation to pollinators in each study system and also understand how adaptation can be achieved through new trait combinations (Dellinger et al., 2014, Pichler et al., 2020).

9. Analysing floral evolution and diversity in the context of pollination syndromes

To date, our understanding of how flowers evolve and adapt to pollinators remains fragmentary. From a functional perspective, flowers are highly synorganised (integrated) structures in which different organs co-function to achieve flower protection, pollinator attraction and rewarding as well as pollen transfer, fertilisation and seed development (Armbruster, 2014; Diggle, 2014). Despite this overall (functional) integration, the above-mentioned processes may influence different parts of the flower differently, potentially leading to different (functional) modules within the flower (Armbruster, 2014; Esteve-Altava, 2017). Recent macroevolutionary analyses have indeed demonstrated that such functional modules may evolve independently of each other, that is along different trajectories and at different rates, in response to pollinator-mediated selection (Opedal, 2019; Dellinger et al., 2019c; Reich et al., 2020). This relative evolutionary independence of functional modules has been proposed to facilitate (and hence explain the frequent occurrence of) shifts between pollination syndromes in closely related plant species (Armbruster, 2014). Several authors have suggested that a single mutation of large effect in advertisement traits (i.e. colour or scent) suffices to induce a pollinator shift, as it may lead to the attraction of a different pollinator (Bradshaw & Schemske, 2003; Wessinger et al., 2014; Wessinger & Rausher, 2015; Sheehan et al., 2016 there were only two studies that calculated disparity). Alternatively, scenarios have been proposed in which strong (extrinsic) alterations in the pollinator community (i.e. environmental perturbation, colonisation of new habitat) render a new pollinator more efficient and flowers gradually adapt to this new pollinator, that is through optimisation of efficiency traits (Thomson & Wilson, 2008; Kay et al., 2019). What is the likelihood of either of these scenarios? Do advertisement traits consistently evolve faster than and change before efficiency traits? How important are changes in advertisement traits in mediating reproductive isolation in sympatry compared to allopatry? And to what extent are such trait changes independent (modular) or integrated?

Molecular developmental studies, experimental manipulations, selection studies and QTL screening have the capacity to get to the core of these questions (Bolstad et al., 2014; Wessinger et al., 2014). In recent years, particular focus has been placed on studying the advertisement traits colour and scent (each trait included in 33% of studies in the literature review, but also see Kostyun et al., 2019). In several pollination-syndrome model systems (e.g. Antirrhinum, Mimulus, Penstemon, Petunia), these traits are controlled by few genes and trait modification have had large effects on pollinator attraction. Through the creation of floral phenotypes intermediate between bee and moth pollination, for example, Dell’Olivo & Kuhlemeier (2013) could show that colour change significantly increased pollination by the new pollinator, supporting the idea that single mutations of large effect could trigger pollinator shifts. Furthermore, changes in colour have also brought about a change in other functional traits such as nectar guides (affecting the correct positioning of pollinators) or petal cell shapes (reviewed by Yuan et al., 2014). Such observed correlations have been attributed to pleiotropy (e.g. reward and corolla shape in Penstemon, Wessinger et al., 2014; flower size and nectar in Jaltomata, Kostyun et al., 2019) or to linked, but independent, gene clusters (e.g. colour, scent and reproductive organ length in Petunia, Hermann et al., 2013; Kostyun et al., 2019). Both processes allow the rapid adaptation of multiple traits to changing pollinator selection regimes.

The above-mentioned studies linking phenotypic to genotypic variation are restricted to a limited number of model systems that can be propagated and crossed in great numbers (Smith et al., 2020). The vast majority of biological systems and plant–pollinator interactions is not part of this group, however. Phylogenetic comparative methods offer a powerful tool to analysing tempo, mode and mechanisms of floral trait evolution in these systems (Smith & Kriebel, 2018; Wessinger et al., 2019; Dellinger et al., 2019c). In recent years, authors have often tested the importance of a single or a few isolated traits in explaining diversification (Lagomarsino et al., 2016). For understanding mechanisms and patterns of flower evolution and diversification in relation to pollination, however, multivariate trait datasets are needed. (Geometric) morphometrics are particularly promising for testing different hypotheses of shape evolution in flowers. While such tests have a long tradition in anthropology and zoology (Esteve-Altava, 2017), this is not the case in plants. This is somewhat surprising as monophyletic plant lineages with repeated independent shifts into different pollination syndromes offer the optimal set-up to testing the role of developmental constraints and functional adaptation under divergent (pollinator) selection regimes. How do pollinator shifts affect correlation structures within the flower? Is there developmental, functional or evolutionary linkage between floral organs/modules? Are certain intrafloral modules more integrated than others and (how) do they differ in their rate of evolution? Are these patterns different between different pollination syndromes? Although testing of these questions at macroevolutionary scales in plants is in its infancy, first patterns start arising. First, efficiency traits involved in the direct physical interaction with pollinators show particularly strong signals of selection and correlated (integrated) evolution (Opedal, 2018, 2019; Benítez-Vieyra et al., 2019; Dellinger et al., 2019c; Reich et al., 2020; compared with Caruso et al., 2018, also see Fig. 5d). Second, pollinator shifts could disrupt both type and strength of correlation structures within flowers and change functional modules (Dellinger et al., 2019c; Reich et al., 2020). And third, the modularity of specialised flowers was found to be structured by function, while modularity of generalised flowers was structured by developmental affinities of floral organs (Reich et al., 2020). It will be interesting to relate these results to advertisement traits to understand more deeply their interplay during pollinator shifts and, potentially, speciation.

Another concept that ties in here is the concept of evolvability. The standing genetic variation of a trait under selection determines its potential to adapt, and hence to constrain or facilitate shifts in pollination syndrome (Opedal, 2019). Rarely have pollination biologists aimed at quantifying this variation and associated evolutionary potential (Armbruster, 2014). What is the relationship between floral phenotypic variation/diversity (flower disparity, Chartier et al., 2017), evolvability and functional pollinator groups? Are flowers adapted to certain functional pollinator groups more disparate than flowers adapted to other groups? Are ancestral pollination syndromes (estimated from extant taxa) more disparate than shifted syndromes? And is disparity (as a measure of evolvability) in the ancestral pollination syndrome a prerequisite for pollinator shifts (Opedal, 2019)? Although disparity could be calculated and analysed easily from multivariate trait datasets (31.1% of studies in the literature reviewed contained multivariate analyses), there were only two studies that calculated disparity (Dellinger et al., 2019a; Vasconcelos et al., 2019). These studies found high disparity in two large bee-pollinated clades (Melastomataceae, Myrtaceae), but this led to pollinator shifts in only one clade (Dellinger et al., 2019a). Disparity of the shifted syndromes was significantly lower (Dellinger et al., 2019a).

Earlier this year, the extension of comparative methods to genotype-to-phenotype mapping approaches through the development of new software tools has been proposed (Smith et al., 2020). This extension holds tremendous and exciting potential for pollination syndromes to more deeply investigate the questions and patterns outlined above in nonmodel organisms (Smith et al., 2020).