A Similar Genetic Architecture Underlies the Convergent Evolution of the Selfing Syndrome in Capsella

Independent evolutionary reductions of petal size in the genus Capsella are associated with convergent evolution of gene expression and underlain by similar genetic architectures.

Abstract

Whether, and to what extent, phenotypic evolution follows predictable genetic paths remains an important question in evolutionary biology. Convergent evolution of similar characters provides a unique opportunity to address this question. The transition to selfing and the associated changes in flower morphology are among the most prominent examples of repeated evolution in plants. In this study, we take advantage of the independent transitions to self-fertilization in the genus Capsella to compare the similarities between parallel modifications of floral traits and test for genetic and developmental constraints imposed on flower evolution in the context of the selfing syndrome. Capsella rubella and Capsella orientalis emerged independently but evolved almost identical flower characters. Not only is the evolutionary outcome identical but the same developmental strategies underlie the convergent reduction of flower size. This has been associated with convergent evolution of gene expression changes. The transcriptomic changes common to both selfing lineages are enriched in genes with low network connectivity and with organ-specific expression patterns. Comparative genetic mapping also suggests that, at least in the case of petal size evolution, these similarities have a similar genetic basis. Based on these results, we hypothesize that the limited availability of low-pleiotropy paths predetermines closely related species to similar evolutionary outcomes.

INTRODUCTION

Phenotypic changes evolved in response to changes in selective pressures or as a result of genetic drift. The influence of stochastic, deterministic, and contingent factors in defining the route of phenotypic evolution is still debated (Blount et al., 2018). Yet, in nature, there are many examples in which different populations have evolved identical features when confronted with similar ecological challenges (Losos, 2010; McGhee, 2011). These similarities are especially intriguing when they do not merely reflect the phylogenetic relatedness of individuals. One explanation that has been proposed is that these features represent optimal, or maybe even the only possible, adaptations to particular environmental constraints (McGhee, 2016). This does not, however, appear to be a universal feature, since many species have also adapted to the same environment through different mechanisms (Blount et al., 2018). Other factors, such as the genetic and developmental constitutions of species, may therefore also influence the repeatability of phenotypic evolution. In agreement with a role of evolutionary history in predisposing adaptive paths, meta-analyses of described events of repeated evolution have indeed suggested that convergent morphological adaptations seem to be more frequent in closely related lineages (Ord and Summers, 2015). Because species that share ancestry have accumulated the same mutations and evolved the same developmental programs on which future adaptations will be built, they may be expected to respond similarly to new selective pressures. Additionally, the existence of shared standing variation in potentially adaptive traits may also underlie the repeated evolution of similar phenotypes (Colosimo, 2005). Indeed, in such cases, repeated selective events of the same genetic variant may have occurred in independent lineages, as they provide a fitness advantage in the face of a new environmental constraint. Yet, the extent to which these different factors explain the convergent adaptations of closely related species is not well understood.

Phenotypes are controlled by a series of genes that regulate each other through transcription or posttranscriptional interactions (Barabási and Oltvai, 2004). The gene regulatory networks (GRNs) controlling different traits are themselves interconnected, and these connections will be modified during ontogenesis to ensure the successful orchestration of developmental programs. The complex structure of these networks may itself limit the genetic solutions to adaptation and predetermine evolutionary outcomes (Shubin et al., 2009; Brakefield, 2011). For instance, highly connected hub genes play an essential role in network connectivity, and thus, any perturbations in their activities can have drastic pleiotropic consequences and strongly impair the fitness of organisms (Jeong et al., 2001; Proulx et al., 2005; Batada et al., 2006). For this reason, only variations in a few genes may provide enough adaptive advantage in the face of new conditions because they allow specific changes in phenotypes and limit the pleiotropic consequences (Stern, 2013). If such genes are limited, they are likely to constitute evolutionary hotspots, explaining the similarities observed in independent convergent evolution. In support of this prediction, a number of studies have now succeeded in identifying genes repeatedly involved in independent evolution of the same phenotype (Shindo et al., 2005; Stern and Orgogozo, 2008; Stern, 2013; Sicard et al., 2014; Esfeld et al., 2018). What these studies have also indicated is that the connectivity within GRNs changes during ontogenesis, and thus, the target of selection within a given GRN is highly dependent on the developmental context (Kittelmann et al., 2018). Also, contrary to the above prediction, the perturbation of highly connected regulatory hubs has been shown to provide rapid evolutionary advantages by promoting phenotypic diversity (Koubkova-Yu et al., 2018). Furthermore, not all networks appear to have a limited capacity to evolve. In freshwater stickleback (Gasterosteus aculeatus), for instance, pharyngeal tooth number has evolved independently through different means (Ellis et al., 2015). Identifying GRNs involved in convergent evolution may, therefore, provide important insights on how network structure can constrain phenotype evolvability. In particular, the hotspot genes previously identified have mostly been isolated from studies of genetically simple traits with limited numbers of genes involved, such as flower color or animal pigmentation (Stern, 2013). By contrast, much less is known about the genetic basis of convergent evolution of highly polygenic traits, with organ size being one prominent example.

The transition from outcrossing to selfing in plants offers a unique opportunity to investigate the genetic bases of convergent evolution of polygenic aspects of morphology. Self-fertilization is believed to be selected when compatible mates are rare and, thus, when its benefits in terms of reproductive assurance outweigh the cost of inbreeding depression (Busch and Delph, 2012). This transition has occurred independently hundreds of times during plant evolution and, in many cases of animal-pollinated species, it has been followed by very similar changes in flower morphology and function (Stebbins, 1957). These similarities are such that these changes have been termed the selfing syndrome (Sicard and Lenhard, 2011). They include, in the predominantly selfing lineages, a strong reduction of flower size, a reduced pollen-to-ovule ratio, and a decrease in the production of nectar and scent. The frequencies of these events and the resulting similarities in terms of phenotypic evolution allow challenging the above hypotheses regarding the influence of phylogenetic distances on the repeatability of evolutionary paths. Yet, we lack studies directly comparing the extent of similarities in the independent evolution of the selfing syndrome.

The genus Capsella has emerged as an ideal model in which to study the phenotypic consequences of the transition toward self-fertilization (Figure 1; Sicard et al., 2011). In this genus, two independent transitions to selfing have occurred, both presumably through the breakdown of the self-incompatibility system (Hurka et al., 2012). In a western lineage, the predominantly self-fertilizing Capsella rubella (Cr) evolved from the outbreeding ancestor Capsella grandiflora (Cg) within the last 200,000 years (Foxe et al., 2009; Guo et al., 2009; Koenig et al., 2018). An earlier and independent event in an eastern lineage gave rise to Capsella orientalis (Co) from a presumed Cg-like ancestor within the last 2 million years (Hurka et al., 2012; Bachmann et al., 2019). The genetic basis of selfing syndrome evolution has been extensively studied in Cr (Sicard et al., 2011; Slotte et al., 2012). In this species, a complex genetic basis underlies the evolution of flower size and pollen-to-ovule ratio. At least six mutations have contributed to reducing the size of the petals. Yet, most of the worldwide-distributed accessions appear to share these mutations, suggesting that the evolution of such traits occurred early during the history of Cr, or at least before its geographical spread (Sicard et al., 2011; Slotte et al., 2012). So far, little is known about the genetic basis of selfing syndrome evolution in Co and, more particularly, on whether or not these two instances of convergent evolution rely on similar molecular mechanisms.

Figure 1.

Convergent Evolution of the Flower Morphology after the Transition to Selfing in the Genus Capsella.

(A) Photographs of an inflorescence and an individual flower of Cg, Cr, and Co are shown at the same magnification. Bars = 3 mm.

(B) Quantification of leaf size and flower organ size.

(C) Quantification of sexual allocation. Pollen and ovule numbers were measured and used to calculate the pollen-to-ovule ratio.

Letters indicate significant differences as determined by one-way ANOVA (Supplemental File) with Tukey’s HSD test. The number under each box plot represents the number of independent accessions used for measurement (mean over individual means from each accession was taken; individual means are shown as dots). See also Supplemental Figure 1 and Supplemental Table 1.

In this study, we compared the developmental, transcriptomic, and genetic bases of the selfing syndrome in Co and Cr. In particular, we asked whether, due to their close relatedness, the phenotypic changes in these two species resulted from similar developmental and molecular mechanisms. We identified common transcriptomic changes between these independent instances of selfing syndrome evolution and studied the pleiotropic properties to determine if their repeated involvement may have been predicted from network structures.

RESULTS

Convergent Evolution of Flower Morphology after the Transition to Selfing in the Capsella Genus

To determine the extent of similarities in the phenotypic changes that have occurred in both selfing lineages, we compared the flower morphology in Cg, Cr, and Co. The sizes of different flower organs, as well as of vegetative organs, were quantified in five representative accessions for each species (Supplemental Table 1). Petal area was about 5 times smaller in both selfers (Co and Cr) compared with the outcrosser (Cg; Figure 1B). This reduction was consistently observed between the different accessions of the three species analyzed (Supplemental Figure 1). Average leaf size was only slightly reduced in Cr and Co (12 and 20% smaller, respectively). Furthermore, the leaf area was highly variable within each species, and the leaves from some of the selfing accessions were as large as those of Cg (Supplemental Figure 1). Within the flower, organ size reduction was not limited to petals but also affected other organs (Figure 1; Supplemental Figure 1). Sepals were 15 and 21% shorter in Cr and Co, respectively, compared with Cg, and the petals for both selfers were 2 times shorter than Cg’s; moreover, anther and carpel length were 25% smaller in the selfers compared with the outcrosser. In both Cr and Co, the evolution of self-fertilization has therefore been accompanied by a reduction of floral organ size and, in particular, a strong decrease in petal dimensions. Strikingly, the size of all flower organs was not significantly different between the two selfers.

In many cases, the transition to selfing is associated with a shift in the allocation of sexual resources (Sicard and Lenhard, 2011). We, therefore, compared the number of pollen grains and ovules in the three Capsella species. The pollen-to-ovule ratio was reduced to a similar extent in both Co and Cr (Figure 1C). This reduction is explained by a similar decrease, approximately fourfold, in the number of pollen grains formed by each flower. As previously reported, we observed a slight increase, ∼1.2-fold, in ovule number in Cr (Sicard et al., 2011). This increase was, however, not observed in Co, which developed a similar number of ovules to Cg. The increase in ovule number was relatively weak compared with the decrease in pollen number and, therefore, had a limited influence on the pollen-to-ovule ratio. As a result, the pollen-to-ovule ratio did not differ between the selfers.

Overall, these results indicate that despite their independent evolution Co and Cr have evolved very similar flower phenotypes. Furthermore, based on the comparison of leaf area between the three species, the changes in organ size seem to be largely restricted to flowers and most prominently to petals, indicating that organ-specific mechanisms are likely underlying these evolutionary changes.

The Same Developmental Mechanism Underlies the Reduction of Petal Size in Cr and Co

To determine whether similar development processes were responsible for the reduction of petal size in Cr and Co, we compared petal growth in each Capsella species using representative accessions (Figure 2; Supplemental Figure 2).

Figure 2.

Similar Developmental Mechanisms Underlie Petal Size Reduction in Both Cr and Co.

(A) PCA on EFDs of the petal outlines. The PC1 to PC7 morphospace is shown on the left, where each dot represents the value of a single petal. The comparison of the PC1 values between each species is shown in a box plot on the right. Letters indicate significant differences as determined by one-way ANOVA (Supplemental File) with Tukey’s HSD test. The number of samples used is indicated under each box plot.

(B) Developmental series of petal growth in Cg, Cr, and Co (n = 4).

(C) Average cell size (left) and cell number (right) along the petal longitudinal axis for the three Capsella species. The values are means of five replicates ± se. Fitted curves are displayed, and the dark-gray area around each line represents the 95% point-wise confidence interval.

(D) Total cell numbers in the three Capsella species. Letters indicate significance as determined by one-way ANOVA (Supplemental File) with Tukey’s HSD test. The number of samples is indicated under each box plot. Individual measurements are shown as dots. See also Supplemental Figure 2.

Because the final organ geometry reflects the sum of different growth patterns that occurred during development, it would be expected that, if petals have been reduced through different mechanisms in the two selfing lineages, their overall shape should also differ. We, therefore, compared the petal geometry of the different Capsella species using principal component analysis (PCA) on elliptic Fourier descriptors (EDFs) of the petal outlines. This analysis indicated that 74.3% of the total variance could be explained by a single principal component, PC1 (Figure 2A). This PC reflected a proportional variation in the overall petal size, with a stronger modification along the transversal axis (Figure 2A). PC1 clearly discriminated Cg from Cr and Co (Kruskal-Wallis test, P = 2.2 × 10⁻¹⁶; Figure 2A) but failed to separate Co from Cr. Only PC7 and PC8, which each only explained 0.1% of the total variance, were significantly different between Cr and Co. However, these two PCs reflect only minor shape variation of the outline curvature at the junction between the petal limb and claw and were not significantly different between each of the selfers and the outcrosser (Supplemental Figure 2). The difference in PC7 and PC8 between Cr and Co is, therefore, more likely to be due to phenotypic variation in the ancestral outcrossing population than to independent adaptation to self-fertilization. Overall, these results indicate that the two selfers have evolved similar petal size and shape, suggesting analogous underlying developmental processes.

In plants, the final organ size is mostly determined by cell proliferation and cell elongation. Any variations in the rate or duration of these processes will, therefore, influence the final organ dimension. Following the growth of petal primordia over time in a representative accession for each of the species did not reveal any difference in the initial petal growth between Cg, Cr, and Co (Figure 2B). However, the petals ceased growing earlier in Cr and Co compared with Cg, indicating that the duration rather than the rate of growth has been modified in the selfers. No difference in average cell size per segment along the longitudinal axis of the petal was observed between the three species (Figure 2C). The numbers of cells in each section, however, strongly differed between the selfers and the outcrosser (Figure 2C). Cg petals consisted of ∼4.5 times more cells than those of Cr and Co (Figure 2D). The total cell number in both selfers was not significantly different, and the number of cells followed an identical pattern along the longitudinal axis. The two selfers have therefore evolved petals constituted by the same number of cells.

Despite the independent evolution of Cr and Co lineages, the same developmental mechanism underlies the reduction of the petal size (i.e., a reduced cell number due to shortening of the cell proliferation period).

Convergent Evolution of Gene Expression in Cr and Co

In many cases, the morphology of organisms evolved as a result of changes in the expression and/or mRNA abundance of genes regulating developmental programs (Stern, 2000; Carroll, 2008; Stern and Orgogozo, 2008). The reduction of flower size in Cr is no exception, and the underlying polymorphisms identified so far affect the mRNA quantity of key growth regulators (Sicard et al., 2016; Fujikura et al., 2018). Given the developmental similarities between Cr and Co, we hypothesized that their morphological adaptations to selfing may have been caused by mutations affecting the same GRNs. If this was the case, it would be expected that the convergent morphological evolution in these lineages has also been accompanied by similar changes in gene expression.

To test this hypothesis, high-throughput RNA-seq was performed at different developmental stages in the three diploid Capsella species. These stages included 10-d-old seedlings and two stages of flower development: young flower buds undergoing intense cell proliferation (hereafter called young flowers) as well as old expanding and maturing flower buds (hereafter named old flowers). PCA using variance-stabilizing transformation (vst)-normalized read counts as well as hierarchical clustering based on Euclidean distances showed a clear grouping of the different replicates, or samples, according to their tissue of origin, their differentiation levels, and their evolutionary distance. The PCA further distinguished the samples according to their mating system (Figure 3A; Supplemental Figure 3A). PC1, which explained 21% of the total variance, separated the samples according to their tissue of origin. Flower samples clustered together at low PC1 values, while seedling samples showed higher scores. PC2 discriminated the samples according to their level of differentiation, with actively dividing tissues, such as young flowers, having higher PC2 values than the mature tissues, old flowers, and seedlings, which have a lower proportion of dividing cells. PC3 separated the samples according to their phylogenetic relationships, with Co being farther apart from both Cr and Cg. PC4, however, distinguished the samples according to their mating system, with Cr and Co having lower PC4 values than Cg. This separation suggests that similar transcriptomic changes have occurred in both selfing lineages.

Figure 3.

Convergent Evolution of Gene Expression after the Transition to Selfing in the Genus Capsella.

(A) PCA factorial map illustrating the distribution of the species and tissue transcriptomes along the largest components of variance. Each symbol represents a biological replicate. The symbol shape and color legend is indicated on the top. The percentage of variance explained by each component is indicated on the axis titles.

(B) The genes differentially expressed between Cr and Cg are compared with those differentially expressed between Co and Cg. The significance of the overlap was calculated using a hypergeometric test. The P value is indicated under the Venn diagram.

(C) Correlation in gene expression changes between the Cr and Co DEGs. The Spearman’s correlation coefficient r is indicated on the graph.

(D) Genes called differentially expressed in the same direction in both selfers (coDEGs) at each developmental stage are compared between tissue types.

(E) Neighbor-joining tree based on Euclidean distances calculated using vst-normalized read counts. Note that the branch separating the selfers from Cg is longer when the species comparison uses the flower transcriptome. See also Supplemental Figures 3 to 5.

To better quantify the extent of overlap between the transcriptomes of the two selfers, we performed a differential gene expression analysis comparing first each selfer with the outcrosser before overlapping the lists of differentially expressed genes (DEGs). Out of 26,521 genes, we found 8242 DEGs (3097 upregulated and 5613 downregulated genes in at least one developmental stage) and 10,281 DEGs (4047 upregulated and 7021 downregulated genes in at least one developmental stage) in Cr and Co, respectively (Figure 3B; Supplemental Figure 3). Among them, 6205 genes were called differentially expressed in both selfers, of which 4173 (∼70%) genes show changes in the same direction relative to the outcrosser in at least one developmental stage. We later refer to these genes as coDEGs.

To determine whether this convergent evolution was also associated with correlated changes in gene expression levels, we compared the fold changes in transcript abundance of the coDEGs in Cr and Co (Figure 3C). Gene expression changes were globally and significantly correlated (Spearman’s r = 0.76, P < 2.2e-16), suggesting that the expression of the same genes has also evolved to a similar magnitude in both selfing lineages. Consistent with the phenotypic changes that occurred in selfers, coDEGs were enriched for many different Gene Ontology terms, including developmental processes involved in reproduction or metabolic processes (Supplemental Figure 4).

Because the phenotypic changes in selfers are mostly apparent in flowers, we next tested whether the convergent evolution of gene expression was mostly restricted to the flower transcriptomes. Interspecific distances based on pairwise gene expression differences revealed a longer gene expression branch length between the selfers and the outcrosser in the flowers compared with seedlings, indicating that the transcriptome of the selfers has diverged to a larger extent from the outcrosser in flowers (Figure 3E). Consistently, we observed that ∼67% of the coDEGs were codifferentially expressed only in the flowers, 10% only in seedlings, and 23% in all developmental stages (Figure 3D; Supplemental Figures 5A and 5B). The fold change in gene expression level was, however, similarly correlated for the coDEGs in both seedling and flower developmental stages (Spearman’s r = 0.69 and 0.68, respectively; Supplemental Figures 5C and 5D).

The repeated evolution of flower morphology has, therefore, been accompanied by a convergent evolution of gene expression, most of which occurred within the flowers, suggesting an accelerated evolution of the flower transcriptome after the transition to selfing.

Low-Pleiotropic Organ-Specific GRNs Underlie the Convergent Evolution of the Selfing Syndrome

Gene “reuse,” and/or convergent evolution in gene expression, in repeated phenotypic evolution has often been proposed to be explained by the fact that only a few genes may have a specialized function that allows optimizing the phenotype of specific functions while avoiding pleiotropic effects (Stern, 2013). We, therefore, sought to determine whether the coDEGs were characterized by a reduced level of pleiotropy. We first analyzed their expression patterns to determine whether these genes tend to be expressed, and therefore have a function, in specific developmental stages. The correlation between gene expression patterns across the developmental stages between selfers and outcrossers was lower for the coDEGs compared with all other genes (Figure 4A; Supplemental Figures 6A and 6B). This was consistent with the observation that most coDEGs were called differentially expressed in a specific tissue and, furthermore, indicated that the expression of these genes has been modified at specific developmental stages.

Figure 4.

Low-Pleiotropy GRNs Underlie the Convergent Evolution of Flower Morphology.

(A) Heat map illustrating scaled expression values for DEGs going in the same direction in Cr and Co at one or more developmental stages.

(B) Percentage of genes expressed preferentially at one developmental stage within the genes differentially expressed in both selfers compared with non coDEGs.

For (A) and (B), YF indicates the proportion of genes that are specifically expressed in young flowers, OF in old flowers, AF in all flower samples, and SD in seedlings. The proportion of genes with equal expression in all samples is shown in beige (Constitutive).

(C) Box plots showing the sums of network connectivity within a Cg population (as determined by Josephs et al. [2017], with 12,896 quantified genes) for coDEGs and all other genes.

(D) Box plots showing coefficients of variation of gene expression levels by expression group within the same Cg population.

For (C) and (D), expression categories for coDEGs (blue) and all other genes (gray) are based on averages over all samples, [−Inf,1) corresponds to genes with tags per million (TPM) up to 1, (1,10) corresponds to genes with TPM between 1 and 10, and (10,Inf] indicates genes with TPM above 10. Categories roughly contain the same number of genes. Asterisks indicate significance (P < 0.05) determined by a Wilcoxon test. See also Supplemental Figure 6 and the Supplemental File.

To determine whether this was due to organ-specific changes in the expression of constitutively expressed genes or changes in genes expressed in a specific developmental context, we determined the proportion of coDEGs and the non-coDEGs (here called others) expressed in an organ-specific manner, which we defined as the genes whose expression is increased by at least fivefold in a given developmental stage (Figure 4B). The coDEGs were enriched in stage-specific genes (63%) compared with the non-coDEGs (31%), with a majority of flower-specific expression (47% of all DEGs). Consistent with this, the list of genes called codifferentially expressed only in flower stages contained a larger proportion (56%) of genes with a stronger expression in flower, whereas the seedling-specific coDEGs were further enriched in genes expressed mostly in seedlings (31%). The majority of the convergent changes in gene expression appears, therefore, to have affected preferentially genes whose expression is enriched at a specific developmental stage and, therefore, likely to have reduced pleiotropic effects.

The strength of connections between genes, also termed network connectivity, can be used to estimate gene pleiotropy from transcriptome data sets, as a gene having higher connectivity is more likely to have a pleiotropic effect (Josephs et al., 2017). We, therefore, used the measures of connectivity of Josephs et al. (2017) extracted from the genome-wide analysis of gene expression in ∼150 Cg individuals to test whether coDEGs tend to have low pleiotropy. The coDEGs show a significant reduction in the sum of connectivity when compared with the non-coDEGs (Figure 4C). This pattern was, however, not observed in 2000 randomly selected genes, suggesting that the coDEGs are particularly enriched in genes with low pleiotropy (Supplemental Figure 6C).

Consistent with previous results indicating a negative association between network connectivity and nonsynonymous divergence, the coDEGs showed a significant increase in the ratio of substitution rates at nonsynonymous and synonymous sites that was not observed for a list of 2000 randomly selected genes (Supplemental Figures 6E and 6F; Josephs et al., 2017). The coefficient of variation calculated from genome-wide expression data of Cg was also significantly higher for the coDEGs, especially for genes with high expression values (Figure 4D; Supplemental Figures 6D and 6G). No change in coefficient of variation was observed in 2000 randomly selected genes. Thus, these observations suggest that the two independent transitions to selfing have reused genes under weak functional constraints, with low pleiotropy and variable expression in the ancestral population.

A Similar Genetic Basis to the Independent Reduction of Petal Size in Capsella

The similarities in the phenotypic and molecular evolution of the two selfing lineages raise the question of whether they may be caused by recurrent mutations in the same genes. To test this hypothesis, we generated Co × Cg and Co × Cr interspecific hybrids using ovule rescue. From the F1 populations, we selected a single individual for each cross type and harvested F2 seeds. We therefore first sought to analyze the segregation of selfing syndrome traits in these populations to test for a common genetic basis. Indeed, if the same genes underlie the independent evolution of the selfing syndrome, the two selfers would carry mutations at the same genomic positions and, thus, selfing syndrome traits should not segregate in the progeny of Co × Cr F1 hybrids. In such a case, the phenotypic distributions of such a population should overlap with those of the two parents. If, however, the change in flower morphology was caused by mutations in different loci, their phenotypic distributions in the same F2 populations would be expected to transgress beyond the parental values.

We therefore analyzed the segregation of petal size as well as ovule and pollen numbers in the Co × Cr F2 populations (Figure 5; Supplemental Figure 7). The petal size distribution was narrower (coefficient of variation = 0.2) in Co × Cr F2 hybrids compared with F2 progeny of crosses between either of the selfers with the outcrosser (coefficient of variation = 0.34 and 0.27 for Cr × Cg and Co × Cg, respectively) and had a mean centered between those of the two selfer parents (Supplemental Figure 7). We detected transgressive segregation beyond the higher parental values but not on the lower side of the distribution. More importantly, no large flower phenotypes were observed in the Co × Cr F2 population. We next compared the frequency distribution of petal size observed in this F2 population with frequency distributions simulated with different quantitative trait locus (QTL) models having an increasing number of contributing loci (Supplemental Figure 8). The observed phenotypic distribution was not statistically different from those simulated with the models using one or no segregating petal size QTL. The two distributions became different only when the model used at least three segregating loci. From this point, the spread of the distribution was higher in the simulated data compared with the observed data (coefficient of variation = 0.3 against 0.2). This is consistent with only a few loci influencing petal size in the Co × Cr F2 population and suggests that either part of the genetic basis underlying the reduction of petal size is shared between the two selfers or that the mutations fixed within each of these lineages may act nonadditively.

Figure 5.

Genetic Basis of the Reduction of Petal Size after the Transition to Selfing in Capsella.

(A) Petal area frequency distribution in Cg × Cr, Co × Cr, and Co × Cg F2 populations. The vertical lines indicate the average values of Cr (green dash-dotted line), Co (red dashed line), Cg (blue dotted line), and F2 population (black dashed line).

(B) Comparison of the genetic basis underlying the reduction of petal size in Cr and Co. The positions of QTLs influencing petal size are indicated on the physical map. The QTLs identified in this study in the Co × Cg population are indicated with a green line and in the Co × Cr with a blue dash-dotted line. Those previously identified in the Cr × Cg recombinant inbred line population (Sicard et al., 2011) are shown with a dotted red line. The LOD score peaks are indicated by horizontal lines, and the vertical lines represent the 1.5-LOD score confidence interval. The width of the vertical lines illustrates the strength of the QTL. See also Supplemental Figures 7 to 11 and Supplemental Tables 2 to 4.

The frequency distribution of ovule and pollen numbers in the Co × Cr F2 population was, however, much wider (Supplemental Figure 7). We found significant transgressive segregation for both of these traits. The segregation of ovule numbers was best explained by a model with at least 12 segregating QTLs, and the pollen number distribution fits more accurately the distribution from models with at least 16 segregating QTLs. These results suggest therefore that a large number of loci influencing these traits are segregating in the Co × Cr F2 population, suggesting a different genetic basis in both selfers.

To confirm these observations and determine whether the limited transgressive segregation observed for petal size was likely due to a shared genetic basis, we used the populations generated above to map the loci involved in the reduction of petal size in Co. A total of 462 F2 individuals of the Co × Cg and 381 F2 individuals of the Co × Cr populations were genotyped using double-digest restriction site-associated DNA (ddRAD) sequencing and phenotyped for vegetative and reproductive traits. The genotypes obtained were then used to establish genetic maps using the Kosambi mapping function. As expected from previous studies, this resulted, for both populations, in eight linkage groups that were in good agreement with the positions of the markers on the genome sequence (Supplemental Figures 9 and 10; Slotte et al., 2013). ddRAD markers were globally evenly distributed throughout the genome, and only weak segregation distortion was observed along the eight linkage groups. We therefore used this genotype information to map QTLs influencing the measured traits (Figure 5; Supplemental Figure 11). A multiple-QTL approach was used to identify the most informative QTL models, which were then used to position the QTLs within the genome (Supplemental Tables 2 and 3). The positions of the QTLs identified were then compared with previous QTL mapping experiments in Cr × Cg F2 and recombinant inbred line populations (Figure 5; Supplemental Figure 11; Sicard et al., 2011; Slotte et al., 2012).

Only one QTL, SIQTL1, influencing the self-compatibility in Co was identified. This QTL overlaps with the Capsella S-locus, which contains the S-locus reporter kinase and the linked S-locus Cys-rich proteins and which has been reported to underlie the loss of self-incompatibility in Cr (Supplemental Figure 11; Guo et al., 2009; Sicard et al., 2011; Slotte et al., 2012). Consistently, self-incompatibility was not segregating in the Co × Cr F2 population. This further supports the observation that mutations at the same locus seem to underlie the loss of self-incompatibility in Cr and Co (Bachmann et al., 2019).

For petal area, we identified eight QTLs that together explained 53% of the total phenotypic variance in the Co × Cg population (Figure 5). Most of these QTLs act additively, as we only detected a significant interaction between PAQTL5 and PAQTL8 (Supplemental Table 2). Notably, the confidence intervals of three of these QTLs, which together explain more that 33% of the total phenotypic variance, overlaps with those of QTLs previously identified in two other independent Cr × Cg populations (Figure 5; Supplemental Figure 11; Sicard et al., 2011; Slotte et al., 2012). In particular, the QTL on scaffold 2, which here explained by itself 23% of the total phenotypic variance in the Co × Cg population, was also found to have a major influence on Cr petal size. To estimate the likelihood of identifying a similar genetic architecture by chance in such a comparative QTL experiment, we estimated the probability of identifying three overlapping QTLs to 0.04 considering the average window size of 3 Mb and eight QTLs located on different linkage groups of an average size of 27 Mb. To account for the conditions of our experimental design and for the fact that additional undetected QTLs may also influence petal size, we simulated 1000 comparative QTL mapping experiments using an estimated true number of QTLs of 13. Only ∼4.6% of these simulations led to an overlap of three QTLs, further suggesting that the overlap we observed is significant and may reflect the recurrent recruitment of similar genomic regions in independent reductions of petal size (Supplemental Figure 11B). From the remaining QTLs, only at PAQTL4, PAQTL6, and PAQTL7, which together explained 11% of the total phenotypic variance, the Co alleles reduced the petal size and are therefore likely to underlie lineage-specific modifications after the transition to selfing (Supplemental Table 3). Consistent with the above segregation analysis, this QTL mapping experiment suggests that the loci with stronger effects may have contributed to the reduction of petal size in both of the selfing lineages. PAQTL1 to PAQTL3 affected both the length and width of the petals, but none of them influences leaf size, indicating that, as in Cr, these loci act in an organ-specific manner. As could be expected, none of these QTLs were detected in the Co × Cr population (Figure 5; Supplemental Figure 11).

We identified five and three QTLs influencing ovule number within the Co × Cg and Co × Cr F2 populations, respectively. Two of these QTLs, OQTL5 in Co × Cg and OQTL2 in Co × Cr, were overlapping with the QTLs identified in Cr × Cg populations (Supplemental Figure 11). The QTL models identified, however, explained only 20% of the total phenotypic variance, suggesting that many other loci with weak effects were also segregating in this population (Supplemental Table 2). Similarly, no significant QTLs influencing pollen number were detected in these populations, suggesting, here as well, the contribution of a large number of small-effect mutations. This scenario is consistent with the above segregation analysis, which revealed that the observed phenotypic distribution could be explained by a complex genetic basis.

DISCUSSION

We sought to analyze the extent of similarities in the phenotypic evolution that have followed independent transitions to selfing in two closely related species. Remarkably, our results indicate that the two selfers have evolved almost identical flower size and shape. This has occurred through the same developmental mechanism, a shortening of the duration of the cell proliferation period. As a result, these two species have evolved petals consisting of the same number of cells. The similarities were not only restricted to flower size, but both of these species also showed a similar decrease in the number of pollen grains. The number of ovules was, however, only increased in Cr. The level of similarities in the phenotypic changes of the two selfers is intriguing and raises the question of what are the underlying causes.

Ecological Value of Selfing Syndrome Traits

These similarities could be explained by the fact that these specific trait values confer a strong phenotypic advantage in a selfing context. Strong reduction in both petal size and pollen numbers are prominent recurring features of the selfing syndrome (Sicard et al., 2011). Theoretical modeling and the rapid evolution of selfing syndrome observed in recently emerged selfer lineages support the idea that these phenotypic changes could evolve as a result of positive selection rather than the relaxation of constraints imposed by the needs of attracting pollinators (Foxe et al., 2009; Guo et al., 2009; Busch et al., 2011; Glémin and Ronfort, 2013).

In both Mimulus and Capsella, flower size has been shown to positively correlate with the distance between anthers and stigma, also known as herkogamy (Fishman et al., 2002; Sicard et al., 2011). Variation in herkogamy negatively affects the ability of plants to self-fertilize, most likely because it facilitates the deposition of self-pollen onto the stigma (Lande and Schemske, 1985; Takebayashi et al., 2006; Luo and Widmer, 2013; Griffin and Willi, 2014; Toräng et al., 2017). A reduced flower size may therefore act as mating system modifiers by facilitating self-pollination and, thus, the establishment of selfing lineages (Barrett et al., 2014; Wozniak and Sicard, 2018). In agreement with this idea, the introgression of a nonfunctional Cr S-locus into Cg was not sufficient to produce highly efficient self-pollinating plants but resulted in a selfing efficiency of about half that observed for Cr plants, which argued that the changes in flower morphology that occurred in the latter have improved autogamous selfing (Sicard et al., 2011). Flower size has been shown to be reduced through different mechanisms and to different extents in different selfing species (Sicard and Lenhard, 2011). However, here both selfers have evolved the same petal morphology. Moreover, they have done so through the same developmental strategy, and our genetic and transcriptomic analyses suggest that they have, at least in part, used similar genetic paths. The large extent of similarities may therefore not only be driven by the adaptive value associated with petal size but also because a limited number of genetic solutions exist to reduce flower size without having detrimental consequences on plant fitness.

Compared with flower size, it is less obvious how a reduction in pollen number would act as a mating system modifier by improving autogamous selfing. It is more likely that the reduction in pollen number reflects the reallocation of the resources invested in male function in outcrossing species (Barrett et al., 2014; Wozniak and Sicard, 2018). Here, our genetic analyses indicate a different genetic basis in the two selfing lineages with the contributions of several small effect mutations. This would be consistent with this trait evolving as a consequence rather than a cause of selfing evolution, in which case mutations with small effects refining the fitness optimum would be expected to contribute (Barrett et al., 2014). That said, why would these two species evolve the same number of pollen grains? The pollen-to-ovule ratio has been shown to correlate with mating systems and ecology (Cruden, 1977; Jürgens et al., 2002; Sicard and Lenhard, 2011). It is therefore conceivable that an optimal ratio between pollen grains and ovules exists and that because Cr and Co share the same ancestry and life history characters, their phenotypes would converge toward the same values.

The Genetic Constraints Imposed on Flower Size Evolution

The convergent evolution of gene expression together with the similarities in the developmental changes after independent transition to selfing in Capsella suggest that, to a large extent, the two selfing lineages share the molecular mechanisms underlying the reduction of petal size. Different growth regulators controlling organ cell number have been shown to regulate different sets of genes in Arabidopsis (Arabidopsis thaliana; Gonzalez et al., 2010). Therefore, at least with respect to organ size, the convergent evolution of gene expression observed between Cr and Co is unlikely to reflect the fact that the developmental outcomes are highly constrained and controlled by a limited number of genes. It rather suggests that mutations affecting the same gene network have been independently recruited in the two lineages.

These observations together with the significant overlap between the genetic bases of these independent events lead us to further hypothesize that they may even have been caused by mutations within the same genes. This raises the question of why, considering the large amount of growth regulators identified in plants, the same genes would be involved in independent evolution of organ size (Czesnick and Lenhard, 2015). Such similitudes in the evolution of the two selfing lineages could be expected if, for instance, the adaptation to selfing in one species would have been helped by the introgression of selfing alleles from the oldest selfing lineage. This seems, however, unlikely for several reasons. First, the lineage in which Co has evolved diverged from the Cg/Cr lineage ∼1 to 2 million years ago (Hurka et al., 2012; Douglas et al., 2015). Furthermore, the species occupy nonoverlapping geographical ranges, which prevent any gene flow between modern populations (Hurka et al., 2012). Finally, no signs of recent introgression were identified at close proximity of shared polymorphisms between the two Capsella selfing lineages (Koenig et al., 2018).

An alternative hypothesis could be that the shared genetic basis of the reduction of petal size in the two selfer lineages reflects a recent evolution of flower size in the modern Cg population used to generate our mapping populations. This seems however unlikely because we found an overlap between three QTLs explaining a large part of the interspecific difference in petal size. If these QTLs reflected intraspecific variation with the outcrossers, we would expect to observe phenotypic differentiation among Cg populations, yet these populations do not differ significantly in any of the traits analyzed here (Supplemental Figure 1). Additionally, the petal size QTLs identified in Co here also overlap with the QTLs identified by Slotte et al. (2012), who used a different Cg parent to generate the Cg × Cr hybrids. Furthermore, at least two of the overlapping QTLs appear to have contributed to a reduction of petal size to different extents in the two selfers, suggesting that different mutations at these loci have been fixed in each of these lineages (Supplemental Figures 11C and 11D). The similarities in these two instances of selfing syndrome evolution may therefore suggest that specific regulatory nodes in the gene network controlling flower size are more suitable to adapt plant phenotypes to selfing.

Although a large number of pleiotropic growth regulators have been identified in plants, it is still unclear how many of them can be modified to change the size of a specific organ. In Capsella, the transition to selfing has been associated with a strong reduction of flower dimensions without strong effects on the overall body size. Mutations affecting petal-specific regulatory elements have been shown to play an important role in the Cr selfing syndrome (Sicard et al., 2016). It is therefore plausible that only the genes with similar regulatory architecture could specifically reduce the size of the flowers. Consistent with the recurrent involvement of mutations having an organ-specific effect, we have observed that the gene expression changes common to the two selfers tend to correspond to genes mostly expressed in flowers. Yet, it is still unclear how many of the plant growth regulators could lead to organ-specific phenotypic evolution.

The capture of standing genetic variants from the ancestral outcrossing population has contributed to the evolution of the selfing syndrome in Cr (Sicard et al., 2016). Furthermore, long-term balancing selection has been shown to maintain polymorphisms over several million years in Cg (Koenig et al., 2018). The availability of genetic variants having an adaptive value for selfers in outcrosser populations could therefore predetermine the path to selfing syndrome evolution. Indeed, if, for instance, purifying selection on small-flower alleles was less efficient at specific loci, it could explain why such loci would underlie independent evolutionary events. Cases of convergent evolution due to shared standing variation have been described before (Colosimo, 2005). One of the three overlapping QTLs had a similar effect on the size of the petals in Co and Cr, which could suggest that the same genetic variants have been fixed independently in the two selfers at this locus and could further support the availability of genetic variants as a factor determining the direction of the evolutionary path (Supplemental Figures 11C and 11D). Considering the diversity of factors that could influence the estimation of these effect sizes, a clear answer to this question will nevertheless require the identification of the underlying genetic variants. We also observed that the expression of coDEGs tends to be more variable in the outcrossing population. The similarities in the independent evolution of the selfing syndrome may therefore originate from the GRNs under weak functional constraints that may provide the necessary variation for a rapid track to phenotypic evolution.

Pleiotropy of the Genes Involved in Independent Evolution of the Selfing Syndrome

Genes with low pleiotropy but strong effects on phenotypes are expected to be more suitable for phenotypic adaptation (Stern, 2013). The convergent evolution in gene expression in the two Capsella lineages appears to affect genes with low pleiotropy. These genes are on average less connected compared with other DEGs, and most of them are expressed predominantly at a particular developmental stage or in a particular organ type. Furthermore, these genes also show higher amino acid divergence, and among them, genes having a high expression level also show higher coefficients of variation in gene expression in Cg. These observations are consistent with previous findings indicating that genes with lower connectivity are under weaker functional constraints (Josephs et al., 2017). Altogether, these observations are in agreement with the idea that genes with lower pleiotropy are more likely to underlie repeated instances of phenotypic evolution. Although it may be difficult to decipher the cause-consequence relationship, these results also suggest that the recurrent involvement of the same genes is also determined by the selective pressures imposed by the network structure. By having a more variable expression, the genes with low connectivity could provide genetic variation upon which evolutionary processes could act. The genetic variation caused by the reduced negative selection on low-pleiotropy genes could therefore by itself predetermine the paths of evolution.

In summary, our results are in line with the predictions that independent phenotypic changes in response to the same ecological challenge are likely to be similar between closely related species that share ancestry (Blount et al., 2018). Nonetheless, our study not only revealed similar evolutionary outcomes but also suggested that very similar evolutionary paths have been followed by the two lineages in response to selfing evolution. Determining whether the shared genetic basis of the two independent reductions of petal size reflects the recurrent recruitment of mutations within the same genes, and thus the existence of genetic constraints imposed on flower evolution, will require further studies aiming to identify the genes involved. This will further allow determining whether such constraints are caused by limited genetic solutions to the evolution of specific phenotypes or by the availability of genetic variants to feed evolutionary processes. By identifying loci involved in independent selfing syndrome evolution, this work offers the means to answer these questions.

METHODS

Biological Materials and Growing Conditions

The geographical origins of the different Capsella rubella, Capsella orientalis, and Capsella grandiflora accessions used in this study are detailed in Supplemental Table 2.1. Cr1504, Co1983, and Cg926 were used for the analysis of petal development and the comparative transcriptomics experiment. Interspecific hybrids were generated by crossing Co1983 with Cr4.23 and Co1983 with Cg926 followed by ovule rescue as previously described (Nasrallah et al., 2000; Sicard et al., 2015). F1 plants were allowed to self for each cross type to generate the F2 populations. The phenotyping analysis in the F2 population was conducted using the progeny of one F1 plant per cross type. Note that Cr4.23 was used instead of Cr1504 in the genetic mapping experiment because genetic incompatibility between Co1983 and Cr1504 strongly impaired hybrid performance (Sicard et al., 2015). Nevertheless, because genetic diversity is relatively low within Cr and the evolution of selfing syndrome occurred before its geographical spread, both Cr4.23 and Cr1504 shared the same mutations and molecular mechanisms underlying the evolution of selfing syndrome (Sicard et al., 2011, 2015; Slotte et al., 2012; Fujikura et al., 2018).

Seeds were surface-sterilized and plated on half-strength Murashige and Skoog medium. Plants were grown in a 3:3:1 mixture of soil (Einheitserde Typ P:Einheitserde Typ T:vermiculite; Kausek) in a growth chamber at the University of Potsdam. All the plants were grown under long-day conditions (16 h of light/8 h of dark) at 70% humidity, with a temperature cycle of 22°C during the day and 16°C during the night, and with a light level of 150 μmol m⁻² s⁻¹ provided by fluorescent lamps (Philips Master TL-D58 W 840 Reflex).

Morphological Measurements

To compare the species-wide phenotypes of Cr, Co, and Cg, five accessions for each species were analyzed as listed in Supplemental Table 2.1. For all these accessions and the F2 interspecific populations, the following characters were measured.

Leaf size was quantified from the fully expanded 12th leaves. Flower organs were measured from the 15th and 16th fully opened flowers on the main inflorescence stem. Three petals, three sepals, and three stamens were measured for each plant. Carpels were collected from the mature 17th and 18th flowers before fertilization. Dissected flower organs were flattened and scanned at a resolution of 3200 dots per inch, while leaf images were digitalized at 300 dots per inch (HP ScanJet 4370). Area, length, and width of the different organs were measured from these digitalized images using ImageJ (http://rsbweb.nih.gov/ij/).

The number of pollen grains and ovules produced by each flower were estimated from two flowers per individual. To estimate the number of ovules, the carpels from the 15th and 16th flowers were incubated in a clearing solution (0.2 g of chloral hydrate, 0.02 g of glycerol, and 0.05 mL of distilled water) overnight. They were then mounted in the same solution on microscope slides, and the number of ovules in each carpel were counted using a light microscope (Olympus BX51). Pollen numbers were estimated from the 19th and 20th flowers just before anther opening. After desiccation at 37°C overnight, the pollen contained in the anthers was released in 5% (v/v) Tween 20 by sonication. The number of pollen grains per flower was then determined using a hemocytometer and a light microscope (Olympus BX51).

The morphometric analysis of petal shape was performed using EDFs (Kuhl and Giardina, 1982). First, the digital images were converted into binary images using ImageJ (Schneider et al., 2012). Coordinates of the petal outlines were extracted using the bwboundaries function in Matlab. Outlines were then Fourier transformed (Kume, 2010) using the base of the petal as reference. Petal shapes were then compared by performing a PCA on EDF coefficients using the R function prcomp (R Core Team, 2018). The effect of the principal components was illustrated by reconstituting the petal outlines through inverse elliptical Fourier transformations (Kume, 2010) and using the maximum and minimum principal component scores. Principal components discriminating the Capsella species were identified using a Kruskal-Wallis test with the formula principal component score ∼ species.

The kinetic analysis of petal growth was performed by manually dissecting two developing petals from each flower bud, starting from the oldest unopened flower and extending toward the youngest buds until the petals could not be dissected. Petal images were acquired using an Olympus BX51 microscope, and areas were quantified as described above. Average petal area was then plotted against time based on the average plastochron of each genotype, which was itself estimated from the number of flowers produced over a 7-d period. To investigate the cellular basis of differences in organ size, dried-gel agarose prints (Horiguchi et al., 2006) of whole petals were generated from five plants for each species’ representative accession. Cell outlines were imaged under differential interference contrast on a BX51 microscope (Olympus) using an AxioCam ICc3 camera (Zeiss). The Python module scikit-image was used to process the resulting images. Cell outlines were segmented by adaptive thresholding. Binarized cell borders were then dilated, skeletonized, and curated by overlaying them with the initial images in GIMP (https://www.gimp.org/). Cell areas and centroid coordinates were extracted and used to determine the cell numbers and average area per section along the petal longitudinal axis. The comparison of cellular patterns in each genotype was performed using R (R Core Team, 2018).

Transcriptome Analysis

Total RNA was extracted using the RNeasy Plant Mini Kit (Qiagen) from 10-d-old seedlings, dividing flower buds, and expanding/maturing flower buds. Buds were considered in an active cell division period when they were younger than stage 10 according to the Arabidopsis (Arabidopsis thaliana) nomenclature of developmental stages (Bowman et al., 1991). Older flower buds were pooled and considered to correspond to the maturation/expanding phase. Three biological replicates for each development stage and for each species were used. TruSeq RNA libraries were generated according to the manufacturer’s instructions (Illumina) and sequenced using the Illumina HiSeq2000 instrument (1 × 50 cycles). RNA-seq data were processed using Trimmomatic (Bolger et al., 2014) to remove adapter sequences. Quality control was done using FastQC (http://www.bioinformatics.bbsrc.ac.uk/projects/fastqc). Reads for all samples were mapped against the Cr reference genome (Cru_183 from phytozome.org) using TopHat2 (Kim et al., 2013). Quantification for Cr predicted genes was done using HTSeq-count (Anders et al., 2015). Data analyses and illustrations were done using R (R Core Team, 2018). DEGs between each selfer and the outcrosser at each developmental stage were obtained using DESeq2 (Love et al., 2014). Genes were considered to be significantly changed when adjusted P values were below 0.05 and log₂ fold change was above 1. Reads counts were vst-normalized for clustering and PCA using DESeq2. PCA was done using the R prcomp function. Neighbor-joining trees were constructed with the R ape package (Paradis et al., 2004) using pairwise Euclidean distances based on vst-normalized read counts. Analyses with gene connectivity and divergence measurements were done using publicly available data sets (Josephs et al., 2017). Expression values to calculate the coefficients of variation in a Cg population were obtained using Kallisto (Bray et al., 2016), and raw data were downloaded from the National Center for Biotechnology Information Sequence Read Archive, project number PRJNA275635. The identification of the gene categories overrepresented in the list of DEGs was performed using Gene Ontology enrichment analysis (Ashburner et al., 2000; Mi et al., 2017; Carbon et al., 2019).

Phenotypic Segregation Analysis

Transgressive segregation was estimated as previously described by Rodríguez et al. (2005). χ² tests were used to compare the number of individuals expected to exceed the means of the parent by at least 2 sd (calculated based on the number of individuals above this threshold in the parental populations scaled to the size of the F2 population), with the observed number of individuals exceeding those thresholds in the F2 population, where sd was the pooled sd of parental populations. Frequency distributions were simulated with various QTL models using the sim.cross function of the R/QTL package add-ons implemented in R (Broman et al., 2003; Arends et al., 2010; R Core Team, 2018). The effect and position of the QTL, as well as the genetic map used in the simulation, were based on previous QTL mapping experiments in Cr (Sicard et al., 2011; Slotte et al., 2012). The sim.cross function was modified to center the distribution around the mean between the two selfers average using previously estimated residual variance for each of the traits investigated (Sicard et al., 2011; Slotte et al., 2012). We assumed no dominance deviations, since previous studies found overall weak dominance deviations compared with additive QTL effects (Slotte et al., 2012). Modeled distributions were then compared with observed frequencies using the Kolmogorov-Smirnov test.

Genotyping by Sequencing

DNA was extracted from lyophilized leaves for each F2 individual as previously described by (Li et al. 2013). DNA concentrations were quantified using the QuantiFluor dsDNA System kit according to the manufacturer’s instructions (Promega) in the Roche LightCycler 480 system. ddRAD libraries were prepared as previously described by (Peterson et al. 2012). A total of 500 ng of high-quality genomic DNA per individual was digested with EcoRI and MspI (New England Biolabs). Fifty nanograms of each digestion reaction was ligated to barcoded adapters and pooled in 48 sample groups. Size selection on each pool was performed via gel extraction using NucleoSpin Gel and PCR Clean-up kit (Macherey-Nagel) targeting fragment sizes between 250 and 500 bp. Distinct pools were combined at equimolar ratios and sequenced using the Illumina NextSeq 500 instrument (2 × 75 cycles). Reads were demultiplexed by Illumina index and ddRAD barcode. They were then mapped at a per sample level against the Cr reference genome (Cru_183 from phytozome.org) using BWA-MEM (Li, 2013). Variant calling was performed using SAMtools (Li et al., 2009). Those calls were then further processed using R (R Core Team, 2018) to code the genotypes relative to their parents and select variable sites.

QTL Mapping

Linkage maps were constructed using R/ASMap (Taylor and Butler, 2017) with the Kosambi function. Individuals with more than 80% of missing genotype information were excluded, and individuals with more than 90% identical genotypes were fused. P values were adjusted until most of the markers were assigned to a linkage group. In total, 1420 and 379 markers were used to reconstitute the Co × Cg and Co × Cr genetic maps, respectively.

QTLs influencing reproductive and vegetative traits were mapped using R/QTL (Broman et al., 2003). The Shapiro-Wilk and Kolmogorov-Smirnov normality tests were used to determine whether the frequency distribution of the data analyzed followed a normal distribution (Supplemental Table 4). In Co × Cg petal length, width, leaf area, and pollen number were not normal and were transformed to a normal distribution using the Box-Cox transformation. In Co × Cr petal length, ovule number and pollen number were not normal. Only petal length could be transformed to a normal distribution using the Box-Cox transformation. The transformed data were then used in the QTL analysis.

For scored traits such as self-incompatibility, we used QTL mapping function with phenotype model adapted to binary traits. An initial single-QTL genome scan was conducted using Haley-Knott regression to calculate the logarithm of the odds (LOD) scores. A genome-wide permutation (n = 1000) test was then conducted to assess the LOD significance threshold (5%). Interactions between QTLs were determined using the scantwo function with a Haley-Knott regression, and significance thresholds (5%) were tested as above, using a genome-wide permutation test (n = 1000). Significant QTLs and their interactions were then used to define QTL models that were assessed using the fitqtl function and the drop-one-term analysis. A multiple-QTL analysis was then used to refine the QTL model until the best fit was identified (Supplemental Table 2). For all significant QTLs, estimates of the additive effect, dominance deviation, as well as the percentage of species variation were obtained (Supplemental Table 2). A 1.5-LOD support interval was used to estimate the position of each QTL.

To estimate the significance of the overlap between the genetic basis of petal size reduction in the two Capsella selfing lineages, we followed the equations described by Blankers et al. (2019). We first estimated the probabilities of observing overlap between the support intervals of three QTLs by performing 100,000 iterations, where at each iteration two segments of 3 Mb, corresponding to the average support interval of the QTLs detected in this study and those detected by Sicard et al. (2011), were randomly drawn from each of the eight linkage groups, each having a length of 27 Mb (average chromosome length calculated from the estimated nuclear size of Cr; Slotte et al., 2013).

We next simulated 1000 comparative QTL experiments using realistic experimental conditions to estimate the proportion of experiments that could by chance detect an overlap between the 1.5-LOD support interval of zero to seven pairs of QTLs. To this end, we first estimated the true number of contributing loci to 13 QTLs based on our estimates of QTL additive effects and the number of QTLs detected in this study following the equations of Otto and Jones (2000). In each QTL analysis for each iteration, we simulated a linkage map with eight linkage groups having a length identical to each assembled Cr scaffold (Slotte et al., 2013) and containing randomly distributed markers averaging a spacing of 0.1 Mb. Thirteen QTLs were randomly placed with these linkage groups, and their effects were drawn from a normal distribution fitted to the density distribution of QTL effects estimated in this and previous QTL studies (Sicard et al., 2011). Using the simulated map and phenotypes for 500 F2 individuals, we performed a genome scan using the randomly generated multiple QTL model and an automated selection of cofactors. Each time a 1.5-LOD support interval was calculated for each retained QTL. The number of overlapping support interval windows between the independent QTL experiments was determined for each iteration.

Statistical Analysis

The phenotypic distributions for each genotype or species are presented with box plots. In these plots, the middle lines correspond to the median and the first and third quartiles are represented by the lower and upper hinges. The whiskers extend until the maximum or minimum values comprised with the 1.5 interquartile ranges. Values beyond this range were considered as outliers and indicated as dots.

Statistical analyses were conducted in R (R Core Team, 2018). The multiple comparisons of phenotypic means were assessed by using Tukey’s HSD posthoc test using the agricolae package add-ons implemented in R (R Core Team, 2018) . For two-sample comparisons, we used a two-tailed Student’s t test assuming unequal variances. The null hypothesis was rejected at P < 0.05. For the kinetic analysis of petal development, the data are presented as means ± se.

Accession Numbers

RNA-seq and ddRAD-seq data were deposited in the National Center for Biotechnology Information Sequence Read Archive database (http://www.ncbi.nlm.nih.gov/sra) under accession number PRJNA589850.

Supplemental Data

• Supplemental Figure 1. Convergent evolution of the flower morphology after the transition to selfing in the genus Capsella.

• Supplemental Figure 2. Similar developmental mechanisms underlie petal size reduction in both Cr and Co.

• Supplemental Figure 3. Convergent evolution of gene expression after the transition to selfing in the genus Capsella.

• Supplemental Figure 4. Gene Ontology enrichment analysis of the coDEGs.

• Supplemental Figure 5. Higher divergence in the flower transcriptome after the transition to selfing.

• Supplemental Figure 6. Low-pleiotropic GRNs underlie the convergent evolution of morphology.

• Supplemental Figure 7. Segregation analysis of selfing syndrome traits.

• Supplemental Figure 8. Simulations of selfing syndrome trait segregations under different QTL models.

• Supplemental Figure 9. Genetic map of the Co × Cg population.

• Supplemental Figure 10. Genetic map of the Co × Cr population.

• Supplemental Figure 11. Genetic basis of selfing syndrome traits in Capsella.

• Supplemental Table 1. Geographic origin of accessions used in this study.

• Supplemental Table 2. Multiple-QTL models.

• Supplemental Table 3. QTL analysis.

• Supplemental Table 4. Normality tests for distribution of phenotypes in the F2 and parental populations.

• Supplemental File. Summary statistics.

DIVE Curated Terms

The following phenotypic, genotypic, and functional terms are of significance to the work described in this paper:

• flower AmiGo: PO:0009046

• petal AmiGo: PO:0009032

• pollen AmiGo: PO:0025281