Phasing subgenomes

The disomic inheritance of C. bursa-pastoris [36, 37] allowed us to successfully phase most of the heterozygous sites in the 31 samples analyzed in this study (Fig 1A, S1 Table). Out of 7.1 million high confidence SNPs, our phasing procedure produced an alignment of 5.4 million phased polymorphic sites across the 31 accessions of C. bursa-pastoris. Scaling these phased SNPs to the whole genome resulted in the alignment of 80.6 Mb that had the same level of heterozygosity as the unphased data. The alignment of these whole genome sequences of C. bursa-pastoris with 13 sequences of C. grandiflora, 10 sequences of C. orientalis, one sequence of C. rubella (the reference), and one sequence of N. paniculata used here as an outgroup, yielded 12.8 million polymorphic sites that we used in all analyses. The information for each accession is provided in S1 Table.

We analyzed the structure of the phased data with phylogenetic analyses. The separation of the two subgenomes was strongly supported in the reconstructed whole genome tree (Fig 1B). The grouping with a corresponding parental species was also maintained in the phylogenetic analyses of each subgenome separately (S1 Fig) as expected given assumptions of the phasing approach. The tree consisted of two highly supported (100% bootstrap) major clades grouping C. grandiflora and the C. grandiflora / rubella lineage descended subgenome of C. bursa-pastoris (hereafter the Cbp_Cg subgenome), on the one hand, and C. orientalis and the C. orientalis lineage descended subgenome of C. bursa-pastoris (hereafter the Cbp_Co subgenome), on the other hand. We also analyzed phylogenetic signals at a finer genomic scale using a sliding window approach with 100-kb window size (Fig 1C). Exclusive monophyly of C. orientalis with the Cbp_Co subgenome, and C. grandiflora and C. rubella with Cbp_Cg subgenome was detected in 95% and 83% of trees, respectively (S2 Fig).

Comparing linkage disequilibrium (LD) decay within and across subgenomes—e.g. comparing r² between SNP1 and SNP2 in Cbp_Cg and r² between SNP1 in Cbp_Cg and SNP2 in Cbp_Co—suggested largely consistent phasing within each subgenome after accounting for population history (S3 Fig). Within homeologues (particularly Cbp_Cg) we observed high LD in the Middle Eastern and European samples with gradual decay as genetic distance increases. LD decayed more rapidly in the Asian samples than in the other populations (possibly due to less population structure and expansion) but also showed the expected decay within both subgenomes. In contrast, r² with SNPs from the other subgenome or randomly selected from either subgenome was substantially lower and did not show any decay within 10 KB, particularly in the Middle Eastern and European samples. High LD and gradual decay within, and not between, subgenomes suggested phasing was largely effective in separating the two subgenomes and preserving signals of their demographic history.

The genomic data analyzed in this study were phased using the computational phasing of reads mapped to the C. rubella reference (HapCUT method) and was validated with mapping the same reads to the recent assembly of C. bursa-pastoris [37] (HomeoRoq method, see Material and methods). The final alignment between the alternatively phased dataset comprised 800 K SNPs. This alignment was relatively small because mapping to two subgenomes resulted in 2x smaller coverage and subsequent genotype calling, filtering, phasing, and alignment cumulatively further reduced this size (S4 Fig). Phylogenetic trees reconstructed for 10 K sliding windows showed 100% split between Cbp_Cg and Cbp_Co subgenomes (S5 Fig). Using a smaller window of 1 K resulted in 83% of trees with mutual monophyly of Cbp_Cg and Cbp_Co subgenomes (S6 Fig). The other 17% still showed a split between Cbp_Cg and Cbp_Co for the majority of samples, but some single samples were not consistent with the mutual monophyly of the two subgenomes. This could be either due to incomplete lineage sorting or potential local phasing errors. Nevertheless, the major clustering into two subgenomes and three populations was largely maintained in the 17% of trees (S7 Fig). In all trees, the same samples phased with two alternative methods (HapCUT and HomeoRoq) clustered together for most of the European samples. The Middle Eastern cluster showed slightly less consistent clustering, and the Asian samples showed clustering into two groups according to the phasing method. The cause of these discrepancies is difficult to define because errors are possible in both datasets. To localize the problematic regions, we also compared the phased dataset analyzed in this study (HapCUT phased) with the assembly sequences of C. bursa-pastoris [37]. This comparison covered almost all positions of the dataset analyzed in this study (10.9Mb out of 12.7Mb). Most of the discrepancies were located in the vicinity of pericentromeric regions (S8 Fig)), where assembling reads is problematic. In summary, there was a high concordance of the results obtained with the different phasing methods and those that differed are unlikely to have affected the main results of the study.

A geographically structured differentiation between subgenomes

For both subgenomes, the three C. bursa-pastoris populations, Asia (ASI), Europe (EUR) and Middle East (ME), constituted well-defined phylogenetic clusters (Fig 1B and 1C). However, the relationships of each subgenome with its parental species differed. The Cbp_Cg subgenome formed a monophyletic clade with C. grandiflora at its base. In contrast, the Cbp_Co subgenome was paraphyletic with C. orientalis that clustered within the ASI group instead of being outside of all C. bursa-pastoris Cbp_Co subgenomes (this was observed in both phasing methods S9 Fig). This clustering was unexpected and suggested potential gene flow between the ASI group and C. orientalis or multiple origins of the Cbp_Co subgenome. Nucleotide diversity was higher on the Cbp_Cg subgenome than on the Cbp_Co subgenome for both EUR and ME (Fig 2, S10 Fig, S2 Table), though the difference was significant only for EUR (p-values: 0.005 and 0.154 for EUR and ME, respectively). The opposite pattern was observed for ASI (Fig 2): there the nucleotide diversity in the Cbp_Co subgenome was significantly higher than in the Cbp_Cg subgenome (p-value < 0.0001). Interestingly, the diversity of the Cbp_Co subgenome in all populations was significantly higher than the diversity of its parental species, C. orientalis (p-value < 0.0001).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Variation in nucleotide diversity (π) between populations of C. bursa-pastoris and parental species.

This boxplot shows the distributions of π estimated along the genome using 100 Kb sliding windows. Co and Cg indicate C. orientalis and C. grandiflora / rubella descendant subgenomes, respectively. ASI, EUR ME, CO and CG correspond to Asian, European and Middle Eastern populations of C. bursa-pastoris, C. orientalis, and C. grandiflora, respectively.

https://doi.org/10.1371/journal.pgen.1007949.g002

A 10-fold or greater decrease in N_e in all selfers but not in C. grandiflora

To reconstruct the changes in effective population size (N_e) over time in the three C. bursa-pastoris populations and the two ancestral species, we used a pairwise sequentially Markovian coalescent model (PSMC). First, we reconstructed the demographic histories of C. orientalis and C. grandiflora (Fig 3). In C. grandiflora, N_e was mostly constant with some slight decrease in the recent past, but the N_e of C. orientalis decreased continuously towards the present. In C. bursa-pastoris, despite a simultaneous rapid range expansion, N_e of EUR and ME populations also gradually decreased starting from around 100-200 kya. The ASI population showed a similar pattern but with population size recovery around 5-10 kya and a subsequent decrease to the same N_e as in EUR and ME. The N_e patterns of the two subgenomes were similar within each population. Overall, the N_e history of C. bursa-pastoris was most similar to that of its selfing ancestor, C. orientalis. We also verified these PSMC results with SMC++, which can consider more than two haploid genomes and incorporates linkage disequilibrium (LD) in coalescent hidden Markov models [39]. The general trend was globally the same but the recent decline of C. orientalis was sharper and fluctuations in N_e were more pronounced (S11 Fig). In summary, the overall pattern of N_e change over time was mostly the same between the two subgenomes and between the three populations of C. bursa-pastoris and it was largely similar to the pattern observed for the diploid selfer C. orientalis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Population size histories of C. bursa-pastoris and its parental species.

Effective population sizes were inferred with PSMC using whole-genome sequences from a pair of haplotypes per population (thick lines) and 100 bootstrap replicates (thin lines). The estimates for different pairs were similar and shown in S11 Fig. Co and Cg specify the Cbp_Co and Cbp_Cg subgenomes of C. bursa-pastoris and corresponding parental species in the CO & CG plot. ASI, EUR, ME, CO & CG indicate Asian, European and Middle Eastern populations of C. bursa-pastoris, and C. orientalis and C. grandiflora, respectively. The axes are in log scale and the most recent times where PSMC is less reliable were excluded.

https://doi.org/10.1371/journal.pgen.1007949.g003

Phylogenetic relationships between C. bursa-pastoris and its parental species vary among populations

To quantify the relationships between populations of C. bursa-pastoris and the two parental species, we applied a topology weighting method that calculates the contribution of each individual group topology to a full tree [40]. We looked at the topologies joining each subgenome of C. bursa-pastoris and a corresponding parental lineage. There are 15 possible topologies for three populations of C. bursa-pastoris, a parental species, and the root. We grouped these topologies into five main groups: species trees—topologies that place a parental lineage as a basal branch to C. bursa-pastoris; three groups that join one of the populations of C. bursa-pastoris with a parental lineage and potentially signifies admixture; and all other trees that place a parental lineage within C. bursa-pastoris but do not relate it with a particular population of C. bursa-pastoris (Fig 4A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Topology weighting of the three populations of C. bursa-pastoris and parental species.

(A) Fifteen possible rooted topologies for the three groups of C. bursa-pastoris in one subgenome and the corresponding parental species. The topologies are grouped into five main groups. The Cbp_Co and Cbp_Cg subgenomes are marked with Co and Cg. ASI, EUR ME, CO, CR, N indicate Asian, European and Middle Eastern populations of C. bursa-pastoris, C. orientalis, C. rubella, and N. paniculata, respectively. (B) Topology weightings for 100 SNP windows plotted along 8 main scaffolds with loess smoothing (span = 1Mb). The tentative centromeric regions are shaded and only eight major scaffolds (chromosomes) are shown. (C) Average weighting for the five main topology groups. The topology groups are in the same order (left-right and bottom-up) and colors in all three plots.

https://doi.org/10.1371/journal.pgen.1007949.g004

These topology weightings varied along the subgenomes and illustrated distinct patterns between the two subgenomes (Fig 4B). In the Cbp_Co subgenome, the largest average weighting was for the topology grouping the ASI population of C. bursa-pastoris with C. orientalis (Fig 4C), and the species topology had the second largest average weighting. The difference between the average weighting in these two topology groups was statistically significant (S3 Table). In contrast, the species topologies weighting dominated in the Cbp_Cg subgenome, regardless if C. rubella or C. grandiflora were used as a parental lineage (Fig 4C, S12 Fig, S4 and S5 Tables). The topology uniting the Cbp_Cg subgenome of the EUR population with C. rubella had the largest topology weighting among the topologies indicating admixture between these clusters (Fig 4C). Thus, the two subgenomes differed substantially in the pattern of topology weighting and there were signs of a potential admixture of EUR and ASI with C. rubella and C. orientalis, respectively.

C. bursa-pastoris subgenomes show signs of admixture with diploid relatives

The phylogenetic grouping of C. orientalis with the Asian Cbp_Co subgenome, together with topology weighting results and the relatively elevated nucleotide diversity in this subgenome, suggested the possibility of gene flow between C. orientalis and C. bursa-pastoris in the ASI population. To test this hypothesis, and at the same time to check for possibilities of gene exchange between C. bursa-pastoris and other Capsella species, we conducted two complementary tests of admixture.

We first used the ABBA-BABA test, a coalescent-based method that relies on the assumption that alleles under incomplete lineage sorting are expected to be equally frequent in two descendant populations in the absence of admixture between any of them and a third population that diverged earlier on from the same common ancestor [41, 42]. The deviation from equal frequency is measured with the D-statistic, which ranges between 0 and 1, with 0 indicating no gene flow and 1 meaning complete admixture. The ABBA-BABA test also provides an estimate of the fraction of the genome that is admixed by comparing the observed difference in ABBA-BABA with the difference expected under a scenario of complete admixture (f-statistic). We estimated D and f for triplets including one diploid species and two populations of C. bursa-pastoris represented by the most related subgenome to that species (Table 1). N. paniculata was the outgroup in all tests. The D-statistic were significantly different from 0 in most of the tests, so we considered all three combinations per test group (see Table 1) to determine the pairs that were the most likely to be admixed. The largest fraction of admixture was identified for the pair of the ASI Cbp_Co subgenome and C. orientalis with an estimate of f indicating that at least 14% of the ASI Cbp_Co subgenome is admixed. The second largest proportion of admixture was detected between C. rubella and the EUR Cbp_Cg subgenome with f estimate of at least 8%. The estimates for tests with C. grandiflora were the smallest but similar to those obtained for C. rubella. The latter may reflect the strong genetic similarity between these two species rather than real gene flow between C. grandiflora and C. bursa-pastoris which, based on crosses, seems unlikely. Indeed, we crossed individuals from the three populations of C. bursa-pastoris with their three diploid relatives to test for the presence of reproductive barriers. Regardless of the direction of the crosses, all crosses between C. rubella and the three populations of C. bursa-pastoris produced viable F1 seeds. In contrast, crosses involving C. grandiflora mostly failed and the abortion rate approached 100%. Details on these crosses are provided in S1 Appendix. The admixture with C. rubella in EUR and C. orientalis in ASI was also supported by the unphased and complete (i.e. no missing genotypes) genomic data (S6 and S7 Tables). Finally, it should be pointed out that given that evidence for C. bursa-pastoris monophyly is weak, it is also possible that the signals of admixture with the parental species that we are detecting here actually reflect introgression from an independently-arisen C. bursa-pastoris into either Cbp_Co or Cbp_Cg subgenomes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Results of the ABBA-BABA tests assessing admixture between C. bursa-pastoris and C. orientalis, C. grandiflora and C. rubella.

https://doi.org/10.1371/journal.pgen.1007949.t001

We then used HAPMIX, a haplotype-based method, which enables capture of both large and fine-scale admixture, as well as an absolute estimate of the proportion of the genome that was admixed. For the analysis of the Cbp_Cg subgenome of C. bursa-pastoris, the highest levels of admixture were found consistently across regions to be with the diploid C. rubella. In Europe, genomic regions showed an average 18% of admixture with C. rubella, followed by 11% in the Middle East, and just 2% in Asia (S8 Table, S13A Fig). All three populations also showed signs of admixture with C. grandiflora but to a reduced extent compared to C. rubella (7% in Europe, 6% in the Middle East, 0.2% in Asia). C. rubella is highly similar to C. grandiflora, and as noted above, we expect difficulties in discerning between the two, suggesting that much of the signal of admixture with C. grandiflora could in fact be due to admixture with C. rubella. Of the regions putatively admixed with C. grandiflora, 78%-96% of sites called as admixed overlapped with those from C. rubella, none of which occurred in unique regions for C. grandiflora. Because of this, and in combination with the reduced genome-wide probability of admixture with the diploid C. grandiflora compared to C. rubella (e.g. 0.11 compared to 0.24 in Europe), we argue that the signals of admixture with the diploid C. grandiflora were likely an artifact of its similarity with the regions of C. rubella admixture. These findings in accord with the ABBA-BABA results imply that the Cbp_Cg subgenome has experienced significant admixture with C. rubella in Europe, and to a lesser extent in the Middle East.

For the analysis of the Cbp_Co subgenome, signals of admixture with the diploid C. orientalis were present in all three populations. In the ME population, genomic regions showed an average 18-21% admixture depending on the reference populations used (S8 Table, S13B Fig). Using the Middle East population for the analysis of the Cbp_Co subgenomes of EUR and ASI, since it was the least admixed in the HAPMIX results, yielded 15% C. orientalis admixture in Asia, and 14% in Europe. These findings suggest admixture of the diploid C. orientalis with the Cbp_Co subgenome across all three geographic regions. Assuming these levels of admixture accurately reflect reality, we do not have a non-admixed reference population to use for HAPMIX, and are thus violating a key assumption of the method. HAPMIX inferences for the Cbp_Co subgenome should therefore be taken with caution but we note that the results for ASI and ME are generally congruent with the admixture pattern obtained with ABBA-BABA.

Selection and gene expression

Deleterious mutation accumulation in subgenomes reflects parental legacy.

We first estimated the nucleotide diversity at 0-fold (π₀) and 4-fold (π₄) degenerate sites and then the ratio π₀/π₄ as a measure of purifying selection. Low values of π₀/π₄ would indicate more efficient purifying selection [43]. As expected, π₀/π₄ was much lower in C. grandiflora than in C. orientalis. In C. bursa-pastoris, purifying selection was more efficient in the Cbp_Cg subgenome than in the Cbp_Co subgenome in both EUR and ME. However, the opposite was observed in the ASI population. For both subgenomes, the ASI population had the highest value of π₀/π₄ even if compared with C. orientalis (S14 Fig).

We then investigated the differences in deleterious mutations among subgenomes and populations by classifying nonsynonymous mutations with the SIFT4G algorithm that uses site conservation across species to predict the selective effect of nonsynonymous changes [44]. In order to control for possible biases due to the unequal genetic distance between the different genomes and C. rubella, we used both C. rubella and A. thaliana SIFT4G annotation databases. Because we were interested in the number of deleterious mutations that accumulated after speciation of C. bursa-pastoris, we polarized the mutations of all three species with the reconstructed ancestral sequences of the common ancestors (see Material and methods).

Regardless of the SIFT4G database (C. rubella or A. thaliana), the proportion of deleterious nonsynonymous sites among derived mutations was always significantly higher in C. orientalis and the Cbp_Co subgenomes than in C. grandiflora and the Cbp_Cg subgenomes (Fig 5, S9 and S10 Tables). Within C. bursa-pastoris, the proportion of deleterious mutations depended on the population considered with the highest value in the ASI population and the smallest in EUR. It is also noteworthy that the proportion of deleterious nonsynonymous sites of the Cbp_Co subgenome in EUR and ME was significantly smaller than that of C. orientalis suggesting that a higher effective population size in the Cbp_Co subgenome than in its ancestor led to more efficient purifying selection in these two populations. On the other hand, the proportion of deleterious nonsynonymous sites in the ASI Cbp_Co subgenome was larger than in C. orientalis, but this difference was only significant for the A. thaliana database. The Cbp_Cg subgenome also had a significantly higher proportion of deleterious sites in ASI than in EUR and ME in all comparisons. In conclusion, the proportion of deleterious sites in the two subgenomes of extant C. bursa-pastoris still reflected the differences between the parental species and the efficacy of purifying selection in the different C. bursa-pastoris populations was associated to their synonymous nucleotide diversity or, equivalently, to their effective population size.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Genetic load in the subgenomes of C. bursa-pastoris and its parental species.

The proportion of deleterious nonsynonymous changes was estimated with SIFT4G on derived alleles, i.e. alleles accumulated after the speciation of C. bursa-pastoris. The left plot shows the results obtained with the C. rubella database and the right plot those obtained with A. thaliana database. Co and Cg are the two subgenomes of C. bursa-pastoris. ASI, EUR, ME, CO, CG indicate Asian, European and Middle Eastern populations of C. bursa-pastoris, and parental species C. orientalis and C. grandiflora, respectively.

https://doi.org/10.1371/journal.pgen.1007949.g005

Selective sweeps differ between subgenomes among populations of C. bursa-pastoris.

The three populations of C. bursa-pastoris also differ in patterns of positive selection. Selective sweeps were more significant on the Cbp_Cg subgenome than on the Cbp_Co subgenome in EUR and ME, whereas the opposite was true in the ASI population (Fig 6). The regions harboring significant sweep signals were also larger on the Cbp_Cg subgenome than on the Cbp_Co subgenome in EUR and ME (total length 42 Mb, 50 Mb vs 9 Mb, 3 Mb), whereas in ASI the sweep regions were larger on Cbp_Co than on Cbp_Cg (total length 4 Mb vs 830 Kb). Although the locations of the Cbp_Cg sweep signals in EUR and ME largely overlap, the patterns differed between the two populations. For example, the strongest sweep signal in EUR was located on scaffold 1, whereas the strongest sweep signal in ME was on scaffold 6. In addition, EUR had many sweeps in Cbp_Co subgenome (109 in EUR_Cg, 128 in EUR_Co), but they all were small and hardly above the significance threshold (Fig 6). In the ME population, the sweep signals in the Cbp_Cg subgenome were prevailing both in size and numbers (101 in ME_Cg, 22 in ME_Co). The ASI population differed strongly from both EUR and ME not only because most of its sweep signals were on the Cbp_Co subgenome but also because these sweeps regions were narrower and less pronounced (Fig 6). We also verified these selection signals with the most stringent filtering of phased fragments (phasing state of all SNPs was supported by fixed differences between parental species) and the major differences between population were maintained (S15 Fig). Thus, all three populations of C. bursa-pastoris were distinct in their selective sweeps patterns with the ASI population being the one least affected by positive selection.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Selective sweep differences between populations of C. bursa-pastoris.

Selective sweeps are detected with the composite likelihood ratio statistics (CLR) along the Cbp_Co and Cbp_Cg subgenomes in Asian (ASI), European (EUR) and Middle Eastern (ME) populations. The 0.01 significance level defined with data simulated under a standard neutral model (green dashed line). Pericentromeric regions are removed. Only eight major scaffolds (chromosomes) are shown.

https://doi.org/10.1371/journal.pgen.1007949.g006

Lack of a strong expression dominance of one subgenome.

One can expect differences in selection patterns between subgenomes and populations of C. bursa-pastoris to also affect the relative expression of the two subgenomes, or homeologue-specific expression (HSE). In particular, biased adaptation towards one subgenome may select for decreased expression of the other subgenome. Hence, given the observed selection pattern, one would expect the Cbp_Cg subgenome to be overexpressed in EUR and ME, and the Cbp_Co subgenome to be overexpressed in ASI.

To assess HSE, we analyzed the RNA-Seq data of 24 accessions representing all three populations of C. bursa-pastoris in a hierarchical Bayesian model that integrates information from both RNA and DNA data [45]. Overall, in agreement with Douglas et al. [10], one subgenome did not dominate the other in the 24 accessions considered together, though a few genes demonstrated a slight expression shift toward the Cbp_Cg subgenome. On average, we assessed HSE in 13,589 genes per accession (range 12,808-15,340) and 18% of them showed significant HSE (posterior probability of HSE > 0.99). The expression ratios between subgenomes (defined here as Cbp_Co / Total) across all assayed genes in the DNA data were close to equal (mean = 0.495). Thus, there was no strong mapping bias.

We found some HSE variation among populations. The mean expression ratios for all genes were 0.494, 0.489, and 0.489 in the ASI, EUR, and ME accessions, respectively, and these mean ratios for genes with significant HSE were 0.487, 0.465, 0.468. The difference in mean ratio between EUR and ME was not significant, but both EUR and ME were significantly different from ASI (S11 Table). In addition, the distribution of expression ratios between the two subgenomes was right-skewed in EUR and ME, whereas in the ASI population, the distribution was more symmetrical (Fig 7). The difference between the populations was particularly evident in the grand mean values (Fig 7). Thus, the shift towards higher expression of the Cbp_Cg subgenome was more prominent in Europe and the Middle East than in Asia.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Distributions of expression ratios between the two subgenomes of C. bursa-pastoris.

The homeologue specific expression (HSE) is estimated by the fraction the Cbp_Co subgenome relative to the total expression level. The upper part presents the distributions for DNA counts, the middle plots show the expression for all assayed gene and the lower plot shows only the distribution for genes with significant HSE. The histograms present the distribution of the allelic ratio, whereas the boxplots summarize these results with the grand mean for every sample. ASI, EUR, and ME indicate Asian, European and Middle Eastern populations, respectively.

https://doi.org/10.1371/journal.pgen.1007949.g007

The effect of parental legacy is detectable even after ∼100,000 years of co-evolution of two initially different subgenomes

The genetic composition of parental species obviously has a strong impact on the evolution of subgenomes in allopolyploids. For example, the allopolyploid Arabidopsis kamchatica originated from the outcrossing diploids Arabidopsis halleri and Arabidopsis lyrata [46] that have rather similar effective population sizes and the two subgenomes of A. kamchatica have thus similar effective population sizes and levels of negative selection [47]. In contrast, the effective population size of the diploid outcrossing ancestor of C. bursa-pastoris, C. grandiflora, is ten times larger than that of its selfing ancestor C. orientalis [10]. Any analysis of the difference in effective population size between the subgenomes of C. bursa-pastoris or of their evolutionary trajectories must therefore account for this initial difference. After the bottleneck associated with the origin of C. bursa-pastoris and the reduction in N_e due to the shift to selfing [48], the effective population sizes of the two subgenomes are expected to progressively converge and decrease along the same trajectory.

While this was indeed the observed overall pattern, the trajectories followed by the two subgenomes in the three populations differed: in Europe the initial level was similar to that in the Middle East but higher than in Asia and the decline of N_e of the Cbp_Cg subgenome was delayed compared to the sudden decline experienced by the Cbp_Co subgenome. In contrast, in Asia the two subgenomes initially followed similar downwards trajectories but N_e increased again in both subgenomes at around 40,000 ya. In the diploid C. orientalis, there was a period of stasis followed by a steeper decline than in the tetraploid. The difference in demography across the three regions could indicate multiple origins of C. bursa-pastoris as suggested by Douglas et al. [10] and the difference between the diploid and the tetraploid could reflect a mixture of the population expansion experienced by the tetraploid and the buffering effect of tetraploidy against deleterious mutations.

There was a clearly noticeable difference between the two subgenomes in the number of accumulated deleterious mutations. Based on the strong differences in N_e, one would expect the efficacy of selection to be much higher in C. grandiflora than in C. orientalis that has a much smaller N_e [49]. In the analysis of the genetic load, we indeed observed that C. orientalis had a higher proportion of deleterious mutations than C. grandiflora. Hence, the amount of genetic load most likely was different between the Cbp_Cg and Cbp_Co subgenomes of C. bursa-pastoris at the time of the species emergence. Interestingly, hundreds of thousands of generations of selfing did not totally erase the differences between the two subgenomes and, today, the Cbp_Co subgenome still accumulates more deleterious mutations than the Cbp_Cg subgenome. Despite being statistically significant, the difference between subgenomes is smaller than between C. orientalis and C. grandiflora suggesting a convergence of the two subgenomes, which is also confirmed by analyses of genetic load dispersion and gene expression regulation [50].

Everything else being equal, genomic redundancy in polyploids is expected to lead to relaxation of purifying selection, and to a higher rate of accumulation of deleterious mutations than in diploids. Douglas and coworkers [10], using forward simulations, concluded that indeed the Cbp_Cg subgenome showed an excess of deleterious mutations compared to what would have been expected from mating system shift alone and ascribed this excess of deleterious mutations to genomic redundancy. Our study also suggests that both effective population size and genomic redundancy contribute to the observed pattern. The proportion of derived deleterious mutations in both subgenomes is, with the exception of the Cbp_Co subgenome in the Asian population, lower than the same proportion in C. orientalis but much higher than in C. grandiflora where it is extremely low (Fig 5). Furthermore purifying selection is weaker on the Cbp_Co subgenome than on the Cbp_Cg subgenome. These data can be explained as follows. First, purifying selection on the Cbp_Co subgenome is stronger than in C. orientalis because the latter has a smaller effective population size than C. bursa-pastoris. Second, purifying selection is weaker on the Cbp_Co subgenome than on the Cbp_Cg because of genomic redundancy. Or, stated differently, because of its highest initial genetic quality the Cbp_Cg may be more sheltered from deleterious mutations than the Cbp_Co subgenome.

Nucleotide diversity also demonstrated the effect of parental legacy. The Cbp_Cg subgenome inherited from the more variable outcrosser C. grandiflora was still more diverse in all populations except the Asian one. The maintenance of part of the parental legacy in both cases suggests that, in spite of their initial differences, both subgenomes have experienced similar levels of fixation since the creation of the species. The Asian population is an exception in this regard because it was affected by secondary gene flow as discussed below. Variation in nucleotide diversity in the coding part of the genome also demonstrated similarity in the efficacy of purifying selection between the two subgenomes and their corresponding parental lineages, though the pattern in the ASI population was the reverse of that observed in the parental lineages. The effect of parental legacy in C. bursa-pastoris was also noted in Douglas et al. [10]. Thus, the effect of the genetic background of hybridizing species may not be as overwhelming as the effect of mating system but it still impacts the fate of the two subgenomes long after the species arose.

Admixture with diploid species and/or multiple origins

Based on coalescent simulations and the amount of shared variation between C. bursa-pastoris and its parental species Douglas et al. [10] ruled out a single founder but noted that it would be difficult to estimate the exact number of founding lineages. Douglas et al. [10] did not consider hybridization but an earlier study detected gene flow from C. rubella to the European C. bursa-pastoris using 12 nuclear loci and a coalescent-based isolation-with-migration model [51], a result which is in agreement with the general occurrence of abundant trans-specific polymorphism in the Arabidopsis genus [52]. The present study adds two new twists to the story. First, our results indicate that shared polymorphisms were not symmetrical: namely, the Cbp_Cg subgenome showed signs of admixture with C. rubella in the EUR and ME populations, whereas the Cbp_Co subgenome was admixed with C. orientalis in ASI. Second, in both the whole genome and density trees, C. orientalis appears as derived from the C. bursa-pastoris Cbp_Co subgenome rather than the converse as one would have expected. No such anomaly was observed for C. grandiflora that, as expected, grouped at the root of the C. bursa-pastoris Cbp_Cg subgenome. These results could be explained by a mixture of multiple origins and a recent admixture. Multiple origins seem to be common in allotetraploids [27, 53, 54] and interploidy gene flow has already been inferred for Capsella [51] and other plant genera [55–57].

Our crossing results did not reject the possibility of ongoing gene flow between C. bursa-pastoris and parental lineages in both Europe and Asia. The distribution of European and Asian populations of C. bursa-pastoris partially overlap with the distribution ranges of C. rubella and C. orientalis, respectively (Fig 1A). The exact proportion of admixture remains unclear at this stage. Taken at face value, the strongest admixture was between the ASI Cbp_Co subgenome and C. orientalis. Considering the overlapping estimates of f-statistic and HAPMIX, the proportion of admixture between the ASI Cbp_Co subgenomes and C. orientalis was around 14%-23%. The admixture between the EUR Cbp_Cg subgenome and C. rubella was also strong, being around 8-20%. There were also signs of minor admixture in the ME population with both C. orientalis and C. rubella. This lack of a non-admixed population posed a problem of correct estimation of the proportion of admixture for both the ABBA-BABA and HAPMIX approaches.

In the ABBA-BABA test, departures from the assumptions can lead to incorrect interpretation of the results. We assumed monophyly of the three populations of C. bursa-pastoris, which may be wrong if these populations were of multiple origins. Thus, the observed shared polymorphism might be due to closer relatedness of our C. orientalis samples with the parent of the ASI population than with the parent of EUR and ME populations, but not admixture. Departures from the assumptions of the ABBA-BABA test can lead to under- or overestimated admixture. In the present case, some proportion of the variation shared between P₃ and both P₁ and P₂ populations could be due to gene flow and not due to incomplete lineage sorting and this would lead to underestimating the amount of admixture. On the other hand, small N_e and recent divergence of the populations used in the test can inflate estimates of D [58]. Further, the behavior of D in tests involving both selfing and outcrossing species has not been assessed yet. The D statistics were significantly different from zero in all our comparisons suggesting that admixture did indeed occur in all populations of C. bursa-pastoris. The f statistic is considered less prone to be affected by these factors [58], and it was more reliable in our tests too. Its values were close to zero in the alternative combinations for the ABBA-BABA tests where we did not expect to find admixture, while D had high estimates (S12 Table). Thus, the f values are the closest to the real proportion of admixture we could obtain.

In HAPMIX, if the reference populations have non-negligible levels of admixture with each other, such that they have few differences, it will be difficult for HAPMIX to distinguish with which reference population the focal population is more likely to share ancestry, driving admixture probabilities to intermediate values. Therefore, we observed a discrepancy between the results of HAPMIX and ABBA-BABA in the estimates of admixture between the EUR Cbp_Co subgenome and C. orientalis. However, the results for the Cbp_Cg subgenome largely agreed between HAPMIX and ABBA-BABA and, together with the results by Slotte et al. [51] and our crossing experiment bolsters the hypothesis of admixture between C. rubella and C. bursa-pastoris in Europe. On balance, a scenario with a single origin of C. bursa-pastoris with later rampant admixture with C. orientalis in Asia and less extensive admixture with C. rubella in Europe is consistent with our data.

On the other hand, our results could also be obtained under a scenario of multiple origins. Such a scenario seems particularly likely if one looks at Fig 4, where the histories of the Cbp_Co and Cbp_Cg subgenomes are strikingly different. If we assume that C. orientalis and C. grandiflora are indeed parental lineages and there was no unknown parental lineage that went extinct, this picture can be only explained by a separate and more recent origin of the ASI population (Fig 8). However, the scenario of multiple origins and post-speciation admixture are not mutually exclusive. The signs of gene flow between EUR and C. rubella are still best explained by post-speciation admixture. The weak signs of admixture between C. bursa-pastoris and C. orientalis in EUR and ME are also difficult to fit into a scenario involving only multiple origins. A possibility is that these signs of admixture resulted from gene flow from ASI to EUR and ME within C. bursa-pastoris. The ASI population is more related to C. orientalis and the presence of its alleles in EUR and ME could be spuriously recognized as introgressed from ASI. Regardless of whether a single or a multiple origins scenario is the true one, our results demonstrate that the history of C. bursa-pastoris is far more complex than previously imagined.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. A tentative scenario of multiple origins of C. bursa-pastoris.

The Asian population originated separately from other C. bursa-pastoris populations. There is gene flow between the European C. bursa-pastoris and C. rubella (solid arrow), and there may be gene flow between the Asian population and C. orientalis (dashed arrow). ASI, EUR, and ME indicate Asian, European and Middle Eastern populations of C. bursa-pastoris, respectively.

https://doi.org/10.1371/journal.pgen.1007949.g008

Weak subgenome-specific expression differences

Many allopolyploid species show subgenome expression bias, where one subgenome tends to be overexpressed relative to the other one [59–62]. This expression dominance is often observed in synthetic allopolyploids [63–66] and thus the major part of such preferential subgenome dominance is probably established immediately after allopolyploidization. The subgenome expression dominance is also suggested to be largely defined by parental expression differences [67, 68]. Contradictory results on patterns of subgenome specific expression in C. bursa-pastoris have been obtained so far. Douglas et al. [10] concluded that there is no strong homeologue expression bias and those few genes showing HSE could be explained by parental expression differences. However, genes with HSE do show a slight bias towards over-expression of the Cbp_Cg subgenome inherited from C. grandiflora / rubella lineage on the Figure 3B in Douglas et al. [10]. In contrast, Steige et al. [69] reported higher expression of the Cbp_Co subgenome inherited from C. orientalis in three accessions, and Cbp_Cg over-expression in a fourth one (CbpGR). Steige et al. [69] hypothesized that the over-expression of the Cbp_Co subgenome might be related to a higher number of transposable elements in this subgenome, but they did not find any evidence of this and could not explain the down-regulation of the Cbp_Co subgenome in the CbpGR accession and in the artificial hybrid between C. rubella and C. orientalis.

Considering the population histories of C. bursa-pastoris sheds some light on these discrepancies. The results of Douglas et al. [10] and Steige et al. [69] are consistent with the hypothesis that cis-regulatory differences between the C. orientalis and C. grandiflora / rubella genomes result in over-expression of the Cbp_Cg subgenome in a hybrid comprising both genomes. Thus, in the absence of other factors, the slight over-expression of the Cbp_Cg subgenome would be the default HSE pattern in C. bursa-pastoris. In accordance with this, we observed over-expression of the Cbp_Cg subgenome in the ME and EUR populations that are most likely the closest to the region of origin of C. bursa-pastoris [34]. The accessions that show over-expression of the Cbp_Cg subgenome in Douglas et al. [10] (SE14 from Sweden) and in Steige et al. [69] (CbpGR from Greece), as we now know belong to the EUR population [34]. Hence, their results are consistent with ours and expected if the HSE is defined primarily by the differences between the parental lineages. On the other hand, we observed that genes with HSE in the ASI population showed equal expression between the two subgenomes. The accessions showing over-expression of the Cbp_Co subgenome in Steige et al. [69] also mostly belong to the ASI population (CbpKMB and CbpGY, though not CbpDE that putatively originates from Germany). Thus, the Asian accessions show a HSE that differs from the default pattern. This difference can be caused by the selection preference for the Cbp_Co subgenome and/or by admixture with C. orientalis that enhanced the cis-regulatory elements of the Cbp_Co subgenome. The ASI population experienced a strong population bottleneck, so genetic drift played some role as well. These explanations need to be confirmed because HSE can be influenced by many factors (e.g. trans-regulatory elements, gene methylation, transposable elements), but it is clear that there are different directions of HSE in populations of C. bursa-pastoris and they are caused by the different evolutionary histories of those populations.

The reason we observed an equal expression between subgenomes in ASI, whereas Steige et al. [69] detected expression bias of the Cbp_Co subgenome for Asian samples, could also be due to different approaches in our analyses. First, we extracted RNA from seedlings, whereas Steige et al. [69] obtained RNA from leaves and flower buds. Variation in HSE for different tissues of C. bursa-pastoris is not characterized yet, so the Cbp_Co expression in seedlings may not be apparent yet. Second, we mapped reads to the C. rubella reference with masked polymorphism, whereas Steige et al. [69] used the reconstructed reference of an F1 hybrid between C. orientalis and C. rubella. The bias in our DNA data was not stronger than in Steige et al. [69], so which method is more appropriate remains to be found out.

Conclusion

Three salient, and sometimes unexpected, features of the evolution of the tetraploid shepherd’s purse that emerged from the present study, are its complex origin and possible admixture with diploid relatives, the long-lasting effects of the difference between its two parental species, and the importance of demography in shaping its current genomic diversity. Hence, the present study suggests that understanding the evolution of tetraploid species without paying due attention to the historical and ecological backgrounds under which it occurred could be misleading.

Sequence data

We obtained the whole genome sequences of 31 accessions of C. bursa-pastoris and the transcriptomes of 24 of these accessions. Transcriptome data used in this study was generated previously from seedling growth in the same growth chamber [35]. Whole genome DNA data consisted of 10 accessions sequenced previously [10] and 21 accessions sequenced in this study. New DNA samples were sequenced using the same technology as the downloaded ones (100-bp paired-end reads, Illumina HiSeq 2000 platform, SciLife, Stockholm, Sweden). The mean genomic coverage of C. bursa-pastoris samples was 47x. We also used the previously generated genomic data of 10 C. orientalis and 13 C. grandiflora samples [10]. For the analysis requiring an out-group, we used the whole genome assembly of Neslia paniculata [70]. Detailed information on the samples is provided in S1 Table.

Genotype calling and phasing

DNA reads from each individual of C. bursa-pastoris were mapped to the Capsella rubella reference genome [70] and subsequently computationally phased the two subgenomes. We favored this approach, which has been already successfully implemented in [10], over mapping to two alternative genomes or read-sorting genotyping algorithms because the two subgenomes of C. bursa-pastoris are quite similar to each other (∼1-3% divergence), and alternative approaches would have large regions where reads cannot be assigned to the parent of origin [21, 54, 71, 72]. To account for the divergence from the reference and ensure minimal read-mapping bias between the two subgenomes, we performed tolerant read mapping using Stampy v1.0.22 [73] with the substitution rate set to 0.025. Potential PCR duplicates were marked using Picard Tools 1.115 (http://picard.sourceforge.net) and were ignored during genotyping. Genotypes were called using HaplotypeCaller from the Genome Analysis Tool Kit (GATK) v3.5 in the GVCF diploid mode and heterozygosity set to 0.015 [74]. Genotypes were filtered for depth between 6 and 100 reads (the 5th and 99th coverage percentiles, respectively) to remove low confidence genotypes due to low coverage or due to their location in repetitive regions and paralogs, which usually have abnormally high coverage. This approach produced a VCF file containing all called sites. This VCF was used in the analyses requiring both polymorphic and monomorphic sites for correct estimates. To obtain a set of SNPs with the highest confidence possible, we generated another VCF file that contained only polymorphic sites and applied more stringent filtering. We set to no-call all sites that met the following criteria: MQ < 30, SOR > 4, QD < 2, FS > 60, MQRankSum < -20, ReadPosRankSum < -10, ReadPosRankSum > 10. These filtering criteria were defined following GATK Best Practices [75] with some adjustment guided by the obtained distributions of the GATK annotation scores (S16 Fig).

To phase the C. bursa-pastoris subgenomes, we run HapCUT version 0.7 [76] on each sample from the VCF with the stringently filtered SNPs. The phased haplotype fragments were then concatenated into two sequences descended from C. grandiflora and C. orientalis (S17 Fig). The origin of haplotypes in HapCUT fragments was defined using sites with fixed heterozygotes in C. bursa-pastoris and fixed differences between the parental lineages. The fixed differences were defined as fixed between 10 C. orientalis and 13 C. grandiflora with maximum 20% of missing data per position. Fragments that had small (< 2 sites) or no overlap with variation in C. grandiflora and C. orientalis as well as those that looked chimeric (prevailing phasing state was supported by less than 90% of sites) were set to missing data (S18 Fig). Additionally, we also set to missing the sites that were defined as not real variants or not heterozygous by HapCUT (flagged with FV). We checked that the distribution of the length of phased haplotype blocks and the proportion of introduced missing data across samples were not strongly different (S19 Fig). HapCUT phasing produced the alignment that had only heterozygous sites and removed all the sites that were non-variant within but variable between individuals. We restored this inter-individual variation by introducing the same proportion of missing data into non-variant sites as it was introduced to heterozygous sites during the phasing. Similarly, we merged the phased SNPs dataset with non-polymorphic sites from the whole genome data to keep the level of heterozygosity as in the unphased data.

To ensure that there was no sign of a general bias towards the Cbp_Cg subgenome, which is less divergent from the reference genome than Cbp_Co, we checked that there was no strong bias towards the reference allele in the VCF file. The bias was only ∼4% and it should not affect alleles calling in heterozygous sites even at the lowest coverage. We also surveyed the level of heterozygosity and coverage along the genome and showed that there was no regional dropout of Cbp_Co haplotypes (S20 Fig). In addition, the accuracy of phasing was assessed with the linkage disequilibrium (LD) variation between and within the subgenomes. From each subgenome, we randomly sampled 1000 SNPs on each chromosome (the principal SNPs) and calculated r² between them and the next 1000 SNPs. These SNPs could originate from the same genome as the principal SNPs or from the opposite subgenome. We only considered SNP pairs where at least 5 samples had genotypes for both SNPs. As a baseline, we also calculated LD in the same way for a genome containing a random mixture of SNPs from both subgenomes. LD gradually decayed with increasing distance between SNPs within homeologues but not between them or when SNPs from both subgenomes were randomly mixed (S3D Fig), further indicating that there was no major phasing error in our data.

To verify our phasing procedure with an alternative phasing method, we utilized the HomeoRoq pipeline [77] that was successfully used to phase transcriptomic and genomic data in allopolyploid Arabidopsis kamchatica [78, 79]. While HapCUT relies on the read-back phasing of the reads mapped to one reference, HomeoRoq classifies the reads mapped to two references corresponding to the subgenomes. As the two references, we used the recently published assembly of C. bursa-pastoris [37]. We independently mapped our reads to the each assembled subgenome. Then we classified reads with the HomeoRoq to keep only unique and common reads, and to remove unclassified reads from SAM files. These SAM files were processed the same way as described above to obtain SNPs. Next, we aligned the main dataset with the HomeoRoq phased data, according to the alignment between the C. rubella reference and the C. bursa-pastoris assembly. The latter was performed by aligning each subgenome of C. bursa-pastoris to C. rubella with the LASTZ program [80] following the procedure described in [10]. Finally, we analyzed the clustering of the samples phased with the two different methods using a sliding window phylogenetic analysis as described below.

The sequences of C. grandiflora and C. orientalis were created using the GVCF files produced by Douglas et al. [10]. The variants were called as described above with additional filtering for fixed differences between the two species. For some of the analyses, where the software was not able to treat heterozygous genotypes properly, we pseudo-phased the sequences of C. grandiflora and C. orientalis by randomizing alleles in heterozygous genotypes.

The data-sets in all the analyses comprised the alignment of phased C. bursa-pastoris sequences, C. grandiflora, C. orientalis, C. rubella (the reference sequence) and N. paniculata. This alignment was filtered for missing data such that genomic positions with more than 80% of missing genotypes were removed. We also removed the repetitive sequences as annotated in Slotte et al. [70] and pericentromeric regions that we delineated based on the density of repetitive regions and missing data. The final data-set had even proportion of missing data in the three populations of C. bursa-pastoris and diploid species (S21 Fig).

Reconstruction of the ancestral sequences

Several analyses presented in this paper required polarized sequence data. The most common approach to polarizing the alleles is to use an outgroup. However, the alignment of Capsella species and N. paniculata, the nearest outgroup with a whole genome sequence available, resulted in substantial reduction of the dataset due to missing data. To overcome this drawback, as well as to track mutations’ origin on the phylogenetic branches, we reconstructed ancestral sequences for major phylogenetic splits. The reconstruction was performed on the tree that was assumed to represent a true history of the Capsella species (S22 Fig) using the empirical Bayes joint reconstruction method implemented in PAML v4.6 [81].

Population differentiation

To assess the degree of differentiation among populations for the two subgenomes, we estimated absolute divergence (Dxy) and nucleotide diversity (π) of the phased genomes using a sliding window approach. The estimates were calculated on non-overlapping 100 Kb windows using the EggLib Python module [82]. The p-values for the difference in mean values were estimated using 10,000 bootstrap resamples from 100 Kb windows.

Temporal change in N_e

We reconstructed changes of N_e over time with both PSMC [83] and SMC++ [39]. We first masked potential CpG islands and all nonsynonymous sites in the genome to avoid bias caused by variation in mutation rates or selective effects. We randomly paired haplotypes for estimation in C. orientalis and did the same for estimations based on the two subgenomes of C. bursa-pastoris. SMC++ was run on all samples from a population, with default parameter settings. For PSMC runs, we set parameters to “-N25 -t15 -r5 -p 4+25*2+4+6”. Variation in N_e was estimated using 100 bootstrap replicates and three different pairs. We chose a mutation rate equal to the mutation rate of A. thaliana, μ = 7 × 10⁻⁹ per site per generation [84] and a generation time of 1 year for all Capsella species.

Phylogenomic analyses

We reconstructed a whole genome phylogeny to explore the relationship between the phased subgenomes of the three populations of C. bursa-pastoris as well as its parental species. To investigate the local phylogenetic relationships along the genome, we also conducted a sliding window phylogenetic analysis using non-overlapping 100 Kb windows. In both analyses, phylogenetic trees were reconstructed using the neighbor-joining algorithm and absolute genetic distance in R package ape [85]. Additionally, a whole genome phylogenetic tree was also reconstructed using the maximum-likelihood approach with the GTRGAMMA model and 100 bootstrap replicates in RAxML v8.2.4 [86] (S23 Fig). The trees from the sliding window analysis were described by counting the frequency of monophyly of different groups with the Newick Utilities [87]. The variation in topology across the genome was also described using topology weighting implemented in TWISST [40]. The weighting was estimated for 100 SNPs windows where each sample was genotyped for at least 50 SNPs. To test for the difference in mean topology weighting, we fitted the generalized linear model with a binomial distribution and performed multiple comparisons for the contrasts of interest with the glht function from the multcomp library in R [88].

Tests for admixture

To evaluate the presence of admixture between the parental species and C. bursa-pastoris, we calculated the ABBA-BABA based statistics, D, an estimate of departure from incomplete lineage sorting, and f, an estimate of admixture proportion [41, 42]. These statistics and their significance, which was estimated with a 1Mb block jackknife method, were calculated from population allele frequencies with scripts from Martin et al. [89]. We also used HAPMIX [90] to infer haplotype blocks of admixture with the diploid C. grandiflora, C. rubella, and C. orientalis into the three populations of C. bursa-pastoris for each phased subgenome. We removed sites with more than 20% missing data for each population. The remaining missing data was imputed for the parental populations used in each analysis. As this method determines the probability of ancestry from a diploid progenitor population relative to a non-admixed C. bursa-pastoris subgenome population, we defined regions of the subgenomes as putatively admixed if the probability of ancestry from the progenitor diploid was greater than 50%.

Selection tests

To search for selective sweeps, we used SweepFinder2 [91]. SweepFinder2 was run on the data-set that besides polarized SNPs also included fixed derived alleles. This enables accounting for variation in mutation rate along the genome and increases power to detect sweeps [92]. The critical composite likelihood ratio (CLR) values were determined using a 1% cut-off of the CLR values estimated in 100 simulations under a standard neutral model. The simulations were performed with fastsimcoal2 [93]. We assumed a mutation rate of 7 × 10⁻⁹ per site per generation, the population effective sizes for every population and subgenome were inferred from the θ values approximated by genetic diversity (π), and the average recombination rate was estimated using LDhelmet v1.7 [94]. In addition, we estimated the ratio between nucleotide diversity at 0-fold (π₀) and 4-fold degenerate sites (π₄) in 5-6 samples with the lowest amount of missing data in each group.

Genetic load estimation

To identify differences in genetic load between populations of C. bursa-pastoris (as well as to assess the effect of selfing on accumulation of deleterious mutations), we classified mutations into tolerated and deleterious ones using SIFT4G [44]. We built the SIFT4G Capsella rubella reference partition database and used it to annotate our SNPs dataset. Then we analyzed the frequencies of tolerated and deleterious mutations. We also verified this analysis by using A. thaliana SIFT4G database and annotating C. bursa-pastoris according to the alignment between the two species. This verification was performed to make sure that the observed results were not due to a reference bias because C. rubella is closer to C. grandiflora than to C. orientalis. To get only the annotation of the mutations that occurred after speciation of C. bursa-pastoris, we polarized the mutations with the reconstructed ancestral sequences (see above) and analyzed only derived mutations. We verified this polarization by analyzing only species(subgenome)-specific mutation (e.g. mutations unique to C. bursa-pastoris Cbp_Co subgenome, C. bursa-pastoris Cbp_Cg subgenome, C. orientalis, C. grandiflora, and C. rubella) (S24 Fig). All the counts were presented relative to the total number of annotated sites to avoid bias caused by variation in missing data between samples. The means of the genetic load were compared using the generalized linear model as we did for the topology weighting except that here we used a quasibinomial distribution due to overdispersion.

Homeolog-specific expression analyses

Mapping of RNA-Seq reads to the C. rubella reference genome was conducted similarly to the mapping of DNA data using Stampy v1.0.22 [73] with the substitution rate set to 0.025. Although potential PCR duplicates are usually not removed from RNA-Seq data, for the allele-specific expression analysis removing duplicates is recommended [95]. We marked duplicates with Picard Tools 1.115 and did not use them during the genotyping and homeolog-specific expression assessment. Variants were called using HaplotypeCaller (GATK) with heterozygosity set to 0.015, and minimum Phred-scaled call confidence of 20.0, and minimum Phred-scaled emit confidence of 20.0 as recommended for RNA-Seq data in GATK Best Practices [75]. Among the obtained polymorphic sites those that had MQ < 30.00, QD < 2.00, FS > 30.000 were filtered out. Calls with coverage of fewer than 10 reads were also excluded. Alleles counting was carried out using ASEReadCounter from GATK.

Homeolog-specific expression was assessed within the statistical framework developed by Skelly et al. [45]. This framework uses a Markov chain Monte Carlo (MCMC) method for parameter estimation and incorporates information from both RNA and DNA data to exclude highly biased SNPs and calibrate for the noise in read counts due to statistical sampling and technical variability. First, we used DNA data to identify and remove SNPs that strongly deviated from the 0.5 mapping ratio. Second, we estimated the variation in allele counts using unbiased SNPs in the DNA data. Next, we fitted an RNA model using parameter estimated from DNA data in the previous step. Finally, we calculated a Bayesian analog of false discovery rate (FDR) with a posterior probability of homeologue specific expression (HSE) > 0.99 and defined genes with significant HSE given the estimated FDR. All inferences were performed using 200,000 MCMC iterations with burn-in of 20,000 and thin interval of 100. Each model was run three times with different starting parameters to verify convergence.

Data access

DNA sequence data generated for 21 accessions of C. bursa-pastoris is submitted to the NCBI database under the sequence read archive number SRP126886. Previously generated DNA sequence data for 10 accessions of C. bursa-pastoris, 10 accessions C. orientalis and 13 accessions C. grandiflora is available in the NCBI (SRP050328, SRP041585, SRP044121). RNA-Seq data is also available in the NCBI (SRA320558). Both phased and unphased SNPs, phylogenetic trees, reconstructed ancestral sequences, estimates of π and Dxy with sliding window approach, results of PSMC and SMC++, SIFT4G annotations, CLR estimates of sweepFinder2, TWISST and HAPMIX outputs, homeologue-specific gene expression values are deposited to the Open Science Framework Repository (DOI: 10.17605/OSF.IO/5VC34) [96].