Adaptation is the driving force behind the amazing diversity of life on this planet, and the past several years have seen an explosion of experimental studies seeking to identify the genetic and molecular bases of adaptive traits (Hoekstra et al., 2006; Rebeiz et al., 2009; Wittkopp et al., 2009; Chan et al., 2010). An important question that emerges from these studies is whether the genetic changes contributing to phenotypic evolution are predictable. In other words, does natural selection preferentially target specific genes—or certain types of mutations—during adaptation? Such predictability suggests that phenotypic change occurs via non-random sampling of potential genetic variation. This may indicate substantial constraint on the process of adaptation, as the epistatic and pleiotropic structure of developmental regulatory networks limits the mutations that result in evolutionary transitions (Stern, 2011). As a consequence, the question of predictability has been of particular interest in the evo-devo literature where multiple surveys have been conducted across broad suites of traits in plants and animals in an attempt to determine whether certain types of evolutionarily relevant mutations are more common than others (Stern, 2000; Carroll, 2005; Hoekstra and Coyne, 2007; Prud'homme et al., 2007; Wray, 2007; Carroll, 2008; Lynch and Wagner, 2008; Stern and Orgogozo, 2008; Wagner and Lynch, 2008; Stern and Orgogozo, 2009).

The independent evolution of the same trait among populations or species is known as phenotypic convergence. Convergence has been shown to occur at different phylogenetic distances. For example, closely related populations within the species Arabidopsis thaliana have independently evolved early-flowering phenotypes (Johanson et al., 2000), while across much broader time scales, at least nine different plant families have independently evolved succulence in arid environments (e.g., Cactaceae and Euphorbiaceae) (Mauseth, 2004). In many of these examples, compelling evidence for adaptation exists because convergence accompanies similar environmental pressures, suggesting that natural selection independently drove the evolution of similar phenotypes. Moreover, from a mechanistic perspective, multiple origins of the same trait offer naturally occurring replicates of the evolutionary process and provide us with an opportunity to explore the types of constraints that dictate the course of phenotypic evolution (Kopp, 2009).

When the genetic basis for this phenomenon is characterized, it has been found in a variety of systems that changes in the same gene, sometimes involving the same substitution, are responsible for the independent origins of the same character. However, in other systems, independent phenotypic transitions have distinct genetic and molecular mechanisms (Arendt and Reznick, 2008; Martin and Orgogozo, 2013). The presence of genetic convergence suggests that evolution can at times be predictable under certain environmental, genetic, and phylogenetic contexts. This can occur due to constraints in the available variation upon which natural selection can act, which can lead to predictable evolutionary outcomes. Therefore, there is substantial interest in the factors that determine how and why convergence at the genetic level accompanies convergence at the phenotypic level (Manceau et al., 2010).

Plant evo-devo has focused primarily on the evolution of the regulatory networks involved in transitions in floral organ identity (Kramer and Irish, 1999; Becker and Theissen, 2003), floral symmetry (Preston and Hileman, 2009), and flowering time (Caicedo et al., 2004). Plant pigmentation (especially of flowers) frequently has been investigated from an evolutionary genetics perspective (Rausher, 2008), with several excellent examples where the genetic and molecular basis of evolutionary transitions in flower color has been characterized. However, this trait has yet to be widely adopted by plant evolutionary developmental biologists. This is perhaps surprising given the prominence of pigmentation traits in the study of animal evo-devo (e.g., Gompel et al., 2005), as many of the big questions in evo-devo are approachable by examining general patterns among these case studies. These include the roles of developmental constraint, redundancy in the genome, the origin of novelty, and the relationship between micro- and macroevolution (Jenner and Wills, 2007). Transitions in floral color are common among the angiosperms, with sister species frequently showing variation in pigment intensity and hue (Rausher, 2008). The anthocyanin pigments are the most common floral pigments and are produced via highly conserved structural and regulatory components (Grotewold, 2006). Differences in flower color among populations or closely related species are often suspected to be directly related to plant fitness, as mediated through pollinator preferences (Bradshaw and Schemske, 2003; Streisfeld et al., 2013). Therefore, the phylogenetically conserved anthocyanin pathway and its associated adaptive phenotypes represent excellent opportunities to merge evo-devo and ecological/evolutionary genetics to investigate the predictability of genetic changes associated with flower color transitions.

Flower color transitions can involve gains or losses of pigment, but how the directionality of flower color change impacts predictability is largely unexplored. Therefore, our primary goal in this review is to explore how this phylogenetic context affects the predictability of flower color change. While flower color transitions may occur via random genetic drift or indirect selection, we focus the review on instances where flower color is presumed to be a direct target of selection via pollinator visitation, physiological traits, or pleiotropic consequences. We also examine the role that population frequency plays on predictability, as flower color changes can occur in nature both as within population polymorphisms and as fixed differences between populations and species. Based on our knowledge of the biochemistry and regulation of anthocyanins, we present experimental predictions for the types of mutations expected to underlie different evolutionary transitions in flower color. We then review the fit of experimental genetic data to our predictions and examine possible mechanisms that may account for the current patterns. For situations where no empirical genetic data exist, we provide suggestions for future work that will make formal tests of these hypotheses possible. We stress that many of the outstanding questions in this field can only be assessed by characterizing multiple examples of the genetic basis of this ecologically important trait.

As introduced above, we define genetic change to be predictable when mutations of one type occur more frequently than others to confer the same phenotypic transition. The term “substitution bias” has been used to reflect this type of predictability for fixed differences between populations (Streisfeld and Rausher, 2011). In particular, when the same phenotype fixes repeatedly across lineages, a substitution bias exists if one type of genetic change is responsible for this convergence more frequently than other types. It is generally accepted that there are two processes that can generate substitution bias: a particular type of mutation has a higher mutation rate (“mutation bias”) or a higher probability of fixation once the mutation arises (“fixation bias”).

To understand how differences in mutation and/or fixation biases influence substitution bias, imagine a situation where multiple, independent lineages have evolved the derived phenotype “B,” and mutations in multiple categories (1, 2, 3,…., n) can generate this phenotype (see Streisfeld and Rausher, 2011 for formal description). A mutation bias exists if mutations in one or more categories occur more frequently than others, as can occur via differences in mutational target size (e.g., Ossowski et al., 2010). By contrast, a fixation bias occurs when there are different probabilities of fixation among these categories. Fixation probabilities are proportional to the net selection coefficient (s), which consists of the difference between the strength of selection favoring the advantageous phenotype (s_a) and any deleterious pleiotropic consequences of the same mutation (s_p; in other words: s = s_a − s_p). Because selection operates on phenotypes and not genotypes, s_a will be similar for all mutations causing phenotype B in a given environment. However, some mutations will have fewer negative pleiotropic consequences and thus a higher net selection coefficient. Those mutations with the highest s are thus the most likely to be fixed. This implies that the predictability of genetic evolution is closely tied to the extent of deleterious pleiotropy associated with the mutations capable of generating phenotype “B.”

While this framework exists for understanding and assessing whether genetic change is predictable, direct calculation of mutation and fixation biases are exceedingly difficult, primarily because they require great effort to monitor rates of spontaneous mutations and the subsequent fixation rate of those mutations in nature. Therefore, previous attempts at examining predictability have focused mostly on determining whether substitution biases exist for different types of mutations (e.g., coding vs. cis-regulatory mutations) across a broad spectrum of traits from both plants and animals (Carroll, 2005; Hoekstra and Coyne, 2007; Wray, 2007; Stern and Orgogozo, 2008). However, substitution bias is most effectively estimated by focusing on the repeated evolution of a single phenotypic trait, because each trait has a unique mutational spectrum (Kopp, 2009; Streisfeld and Rausher, 2011; Conte et al., 2012; Martin and Orgogozo, 2013). Indeed, the repeated evolution of similar flower color phenotypes has allowed for a rigorous analysis of not only substitution biases, but also the mutation and fixation biases that contribute to substitution bias (Streisfeld and Rausher, 2011).

Two underappreciated factors that can influence the likelihood of predictability are the directionality of the phenotypic transition and the population frequency in which polymorphisms exist. Because losses of floral anthocyanins are believed to occur more frequently than gains, previous work has focused exclusively on losses (Streisfeld and Rausher, 2011). However, several examples now document gains in floral pigmentation between populations, which may impact the spectrum of mutations capable of generating the phenotype. Similarly, the literature on predictability is based largely on differences that occur between species. However, it has been argued that the types of mutations involved in phenotypic change can vary due to the efficacy of natural selection across different evolutionary timescales (Stern and Orgogozo, 2008). Ultimately, those mutations that are universally advantageous in a given environment have the greatest potential to become fixed and persist between populations or species. However, examples of convergence in floral pigmentation are common both as within population polymorphisms as well as fixed differences between populations. Nevertheless, little attention has been paid to how the types of mutations that may be responsible for these convergent phenotypes will vary depending on the frequency of the derived flower color. Thus, the primary goal of this review is to examine how directionality and population frequency influence the predictability of the genetic changes associated with evolutionary transitions in flower color.

We begin with a brief introduction of the genetics, biochemistry, and regulation of the anthocyanin pigmentation pathway. Our goal here is to provide context for the sections that follow as opposed to presenting an exhaustive review of the field. For more details on these topics, we refer the reader to several excellent, recent reviews on the subject (Holton and Cornish, 1995; Grotewold, 2006; Feller et al., 2011; Hichri et al., 2011; Davies et al., 2012).

Anthocyanins are common floral pigments that give rise to blue, purple, and red colors, and the flavonoid biosynthetic pathway responsible for anthocyanin pigmentation is highly conserved (Holton and Cornish, 1995; Winkel-Shirley, 2002). Flower color appears to be an evolutionarily labile trait, and transitions in floral anthocyanin pigmentation typically are of one of two types: (1) changes in pigment intensity, which are generally determined by the concentration of pigment in floral tissue, and (2) changes in floral hue, which are often determined by the types of pigments and co-pigments present in a flower. In some circumstances, shifts in both intensity and hue have been reported (Hopkins and Rausher, 2011). Because the predictability of floral hue transitions have been the focus of previous work (see Streisfeld and Rausher, 2011; Wessinger and Rausher, 2012), we focus here on transitions in pigment intensity.

With a few exceptions, all anthocyanin pigments are derived from one of three anthocyanin precursors: pelargonidin, cyanidin, and delphinidin (Grotewold, 2006). These precursors are produced by alternative branches of the pathway (Figure 1). At least six different enzymatic reactions must occur to produce pigments, and the genes coding for these enzymes have been characterized in several species (Figure 1; Holton and Cornish, 1995). These enzymes also serve as branching points for the non-pigmented flavonoids, including lignins, flavonols, and other phenolic compounds such as tannins that play important physiological roles in plant stress responses (Winkel-Shirley, 2002).

In addition to the structural backbone of the pathway, much is known about the regulation of these enzymes. In all species that have been examined, coordinated gene regulation of multiple pathway enzymes occurs at the level of transcription via the interaction of cis-regulatory elements in pathway genes and their transcription factor complexes (Quattrocchio et al., 2006). Detailed investigation from Arabidopsis, maize, and Petunia has provided a nearly complete description of an interacting transcription factor complex that is made up of proteins encoded by members from three large gene families: the R2R3-MYB, basic helix-loop-helix (bHLH) and WD40-repeat (WDR) families (Broun, 2005; Koes et al., 2005; Ramsay and Glover, 2005). Hereafter, we refer to this complex as the MBW complex, which has become a paradigm for our understanding of coordinated gene regulation in plants (Feller et al., 2011). Moreover, MYB-domain proteins lacking one of the canonical repeat domains associated with DNA binding (known as R3-MYBs), have been shown to repress anthocyanins and other flavonoids in vegetative tissues (Quattrocchio et al., 2006; Matsui et al., 2008; Dubos et al., 2010; Albert et al., 2011). However, because the full range of factors that can repress the pathway is not fully characterized, we focus primarily on the role of the MBW complex in regulating differences in anthocyanin pigment intensity.

In addition to its conserved role in regulating the expression of anthocyanin pigmentation, the MBW complex also controls alternative developmental pathways that affect various aspects of epidermal cell fate, including the initiation of trichomes, alteration of vacuolar pH, production of root hairs, seed coat mucilage, and proanthocyanidin synthesis in seeds (Broun, 2005; Koes et al., 2005; Ramsay and Glover, 2005). As has been demonstrated in Arabidopsis, a single WDR protein (TTG-1) and a limited number of bHLH proteins (e.g., GL3, EGL3, TT8) are responsible for regulating all of these functions (Broun, 2005). The specificity of each function appears to be the result of the particular R2R3-MYB protein that joins the complex (e.g., TT2, PAP1, PAP2, WER, GL1; Ramsay and Glover, 2005; Schwinn et al., 2006; Dubos et al., 2010).

R2R3-MYBs represent one of the largest families of plant transcriptional regulators, and they appear to have undergone a rapid expansion during the evolution of land plants (Feller et al., 2011). For example, different copies of R2R3-MYB proteins are known to confer tissue- and cell-specific anthocyanin regulation in several species, whereas other copies are responsible for trichome initiation and proanthocyanidin synthesis in seeds (Broun, 2005; Koes et al., 2005; Morita et al., 2006; Quattrocchio et al., 2006; Schwinn et al., 2006; Gonzalez et al., 2008). This pattern suggests that duplication of R2R3-MYB proteins has allowed different copies to become specialized in their expression and developmental control of specific traits. By contrast, this specialization does not appear to have occurred for WDR or bHLH proteins (Koes et al., 2005; Ramsay and Glover, 2005; Morita et al., 2006), which has implications for the predictability of the genetic changes responsible for flower color evolution (see below).

Based on this understanding of the structural and regulatory components of this pathway, we can specify the set of possible mutation types that can give rise to repeated transitions in floral pigment intensity. Specifically, to obtain a change in pigment intensity, the amount of flux through the pathway must be altered. In principle, such alteration may be caused in any of four ways: (i) by cis-regulatory changes to enzyme-coding genes that directly affect enzymatic expression; (ii) by coding-sequence mutations in enzyme-coding genes; (iii) by cis-regulatory changes to anthocyanin transcription factors; or (iv) by coding-sequence mutations to anthocyanin transcription factors (Streisfeld and Rausher, 2011). Previous investigation has shown that functional deactivation or reduced expression of any of the pathway enzymes or transcription factors can result in a loss of pigmentation (Grotewold, 2006), though as we discuss below, this set of mutations is reduced if the transition involves a gain of floral pigmentation.

It is important to recognize that loss vs. gain of function can occur at either the phenotypic or genetic level, and no correspondence between the two is necessary. For example, loss of pigment production can be achieved either through functional inactivation of enzymes or transcription factors (loss of function at the genetic level) or through recruitment of an inhibitor of anthocyanin production (gain of function). We focus this review at the phenotypic level, with gains and losses referring to the intensity of floral anthocyanin pigmentation. The direction of these transitions is typically determined based on phylogenetic comparative methods. Sophisticated tools that jointly estimate the transition and diversification rates have been developed recently for this purpose (Maddison et al., 2007; Fitzjohn et al., 2009). However, for those groups where characterization of the genetic basis for evolutionary transitions in floral pigment intensity exists, these methods have unfortunately not yet been employed. Therefore, acknowledging the potential limitations, we define “losses” and “gains” in this review largely in a parsimony framework, where we consider a “gain” when an anthocyanic flower occurs in a clade dominated by unpigmented flowers, and a “loss” when flowers completely lack or have significant reductions in floral anthocyanins within a clade of mostly pigmented flowers.

Gains of novel traits are expected to occur less frequently than losses because there are typically more ways to break a pathway than to restore it. This argument essentially states that the mutation spectrum for losses will be greater than the mutation spectrum for gains. While this may often be true (see below), some traits may be easier to gain than others. For example, with the notable exception of the Caryophyllales, nearly all higher plants produce anthocyanin in vegetative tissue (Brockington et al., 2011), where it is involved in a variety of physiological functions (Winkel-Shirley, 2002; Grotewold, 2006; Rausher, 2008). As a result, the ABP and its regulation are conserved through large spans of evolutionary time, regardless of whether pigment is produced in flowers. Consequently, gaining floral anthocyanins in a lineage usually does not require the restoration of a non-functioning pathway, but instead a simple shift in regulation of the ABP genes into this expression domain. Therefore, gains of floral anthocyanin pigmentation may occur more readily than gains in other phenotypic traits.

Although a mutation in any of the four categories listed above is capable of generating a loss of floral anthocyanins, the deployment of ABP gene regulation into the flower will likely be restricted to mutations in transcription factors (Martin et al., 2010). In lineages where anthocyanins were previously absent, the novel recruitment of the ABP in flowers can be accomplished via activating mutations in members of the MBW complex. Because of their functional specificity and limited deleterious pleiotropic effects, cis-regulatory mutations in R2R3-MYBs are the most likely mutations to produce these co-option events. Alternatively, increased pigment intensity can occur in lineages where anthocyanins are already present. This can occur either by removing the repression of a transcription factor that inhibits proper expression of the ABP genes in floral tissue or by quantitatively increasing expression of ABP genes in flowers.

As with losses, the deleterious pleiotropic effects of the various mutations capable of increasing floral pigmentation are expected to be the major factor determining which mutations are targeted by natural selection. However, because the range of possible mutations that can contribute to a gain is less than a loss, it is unlikely that natural selection will have the opportunity to simultaneously sample multiple mutations and target the one with the fewest pleiotropic effects. Instead, the predictions for gains are more contingent on the relative balance between the strength of selection favoring the novel flower color and the negative pleiotropic effects of that mutation. Therefore, in the case of anthocyanin gains, it may be common for a mutation with deleterious effects to fix in a population if the strength of natural selection favoring a novel flower color outweighs the negative pleiotropic effects of that mutation.

Flower color evolution traditionally has been viewed as driven primarily by variation in pollinators (Fenster et al., 2004). However, it is now appreciated that non-pollinator agents of selection can also play significant roles (Armbruster, 2002; Strauss and Whittall, 2006). For example, as a consequence of the multifunctional ABP, flavonoids that play important physiological roles for the plant have the potential to result in indirect changes in floral anthocyanin pigment intensity. Therefore, even though multiple agents of selection may be responsible for differences in flower color, in attempting to make the connection between flower color transitions and the spectrum of negative pleiotropy that can arise, we find it illustrative to first consider cases in which gains in anthocyanin are being directly favored (by whatever agent). Thus, for simplicity, we present the following scenarios as if pollinators alone impose direct selection on flower color transitions, with negative pleiotropic consequences arising as a result of altered vegetative physiological traits under the control of the same genes. We first describe the various mechanisms that can lead to evolutionary gains in floral pigmentation and then determine how the strength of selection on this phenotype affects the types of mutations that result in gains in flower color in nature.

Flower color has been the subject of analysis since the origin of the field of genetics (Mendel, 1866). The ease with which it can be measured, its potential impact on plant fitness, and the characterization of the associated molecular pathways have continued to make this trait an attractive focus of research (Rausher, 2008). Ecological and evolutionary genetic studies of this trait have provided answers to long-standing questions regarding the molecular basis of adaptive traits. Moreover, using a phylogenetic context to examine the genetic basis of repeated phenotypic transitions makes it possible to address central goals of evo-devo research, such as whether genetic evolution is predictable. We have shown that the directionality of evolutionary transitions in flower color, is very likely to impact the precise nature of genetic evolution by altering the mutation spectrum available for gains and losses of floral pigmentation (Table 1). Nevertheless, all but one of the fixed differences in flower color between populations involve cis-regulatory or coding mutations in R2R3-MYB proteins. This remarkable pattern of genetic convergence occurs both for losses and gains of floral anthocyanins and can be explained by the amazing diversity and specialization of R2R3-MYB proteins present in the genomes of most flowering plants. Moreover, the fact that a broader set of mutations is evident among segregating polymorphisms further supports the notion that natural selection preferentially and predictably targets those mutations with the fewest deleterious pleiotropic effects.

Although we realize that sample sizes are small and that patterns may change as more examples are characterized, current results are consistent with the predictions that we generated simply based on our knowledge of the ABP and its regulation. However, it is important to note that many of our predictions rest on the assumption that differences in pleiotropy accompany the various mutations capable of altering flower colors. While a reasonable assumption, empirical studies that characterize the extent of pleiotropy of individual mutations and their associated fitness consequences are important next steps for determining the mechanisms that lead to this pattern of genetic convergence (e.g., Von Wettberg et al., 2010).

In conclusion, this review demonstrates how focusing on the repeated evolution of single traits allows us to make mechanistic predictions of the genetic targets of selection responsible for adaptive shifts in phenotype. In addition, because the ABP has conserved physiological roles in vegetative tissue, recurrent gains of flower color are observed in many plant families. Therefore, the evolution of R2R3-MYB transcription factors also can be used to study how conserved developmental networks are co-opted for novel functions, providing invaluable insight into an important source of phenotypic diversity. Finally, even though the long history of studying the anthocyanin pathway and its conserved nature among the angiosperms make transitions in flower color an excellent model for addressing these questions, if repeated transitions in other traits are studied in a similar way, it may be possible to identify general and replicated patterns about the predictability of genetic evolution. Therefore, this study underscores the need to continue dissecting the genetic basis of convergent evolution across phylogenetic scales, including the further development of flower color transitions as an essential model trait in evo-devo.