The Evolution of Selfing Is Accompanied by Reduced Efficacy of Selection and Purging of Deleterious Mutations

Abstract

The transition from outcrossing to selfing is predicted to reduce the genome-wide efficacy of selection because of the lower effective population size (N_e) that accompanies this change in mating system. However, strongly recessive deleterious mutations exposed in the homozygous backgrounds of selfers should be under strong purifying selection. Here, we examine estimates of the distribution of fitness effects (DFE) and changes in the magnitude of effective selection coefficients (N_es) acting on mutations during the transition from outcrossing to selfing. Using forward simulations, we investigated the ability of a DFE inference approach to detect the joint influence of mating system and the dominance of deleterious mutations on selection efficacy. We investigated predictions from our simulations in the annual plant Eichhornia paniculata, in which selfing has evolved from outcrossing on multiple occasions. We used range-wide sampling to generate population genomic datasets and identified nonsynonymous and synonymous polymorphisms segregating in outcrossing and selfing populations. We found that the transition to selfing was accompanied by a change in the DFE, with a larger fraction of effectively neutral sites (N_es < 1), a result consistent with the effects of reduced N_e in selfers. Moreover, an increased proportion of sites in selfers were under strong purifying selection (N_es > 100), and simulations suggest that this is due to the exposure of recessive deleterious mutations. We conclude that the transition to selfing has been accompanied by the genome-wide influences of reduced N_e and strong purifying selection against deleterious recessive mutations, an example of purging at the molecular level.

MATING system transitions provide important opportunities to investigate the influence of genetic drift and natural selection on plant genomes. The evolution of predominant self-fertilization (selfing) from cross-fertilization (outcrossing) is recognized as the most frequent evolutionary transition involving the reproductive systems of flowering plants (Stebbins 1957). Although selfing may be favored in the short term, due to the transmission advantage of selfing variants and ability of individuals to set seed in pollen-limited conditions, selfing species represent only 10–15% of angiosperms and predominant selfing is often viewed as an evolutionary dead end (reviewed in Igic and Busch 2013; Wright et al. 2013; Barrett et al. 2014). The ephemeral nature of selfing lineages may be primarily due to the accumulation of deleterious mutations (Charlesworth et al. 1993a). Also, a reduced rate of fixation of beneficial mutations can limit the ability of selfing populations to adapt to changing environments (Glémin and Ronfort 2013). Nevertheless, empirical support for selfing as an evolutionary dead end is mixed (reviewed in Takebayashi and Morrell 2001; Igic and Busch 2013). Characterizing the frequency of mutations and selection pressures acting on them in selfing populations should provide insight into why some populations persist while others go extinct.

Theoretical models predict the efficacy of selection should be lower in genomes of selfing compared to outcrossing populations, because all else being equal, the effective population size (N_e) is reduced in selfers (Charlesworth et al. 1993b; Charlesworth and Wright 2001; Wright et al. 2008). In selfers, N_e is reduced owing to restricted recombination but also by demographic factors because single individuals have the ability to found colonies resulting in population bottlenecks (Baker 1955; Lloyd 1980; Charlesworth and Wright 2001; Pannell and Fields 2014). With a genome-wide reduction in the N_e of selfers, genetic drift will become more prevalent, reducing genetic variation and limiting the efficacy of selection. Moreover, greater linkage among weakly selected sites under opposing selective forces may interfere with selection efficiency (McVean and Charlesworth 2000; Comeron et al. 2008). Thus, both demographic and genetic processes in selfing populations can interact to reduce the efficacy of natural selection.

Early investigations of the genomic consequences of transitions from outcrossing to selfing based on divergence among related outcrossing and selfing species provided limited support for theoretical predictions of a reduced efficacy of selection in selfing populations [e.g., Arabidopsis (Wright et al. 2002), Mimulus (Sweigart and Willis 2003), Caenorhabditis (Cutter et al. 2006), and Triticeae (Haudry et al. 2008; Escobar et al. 2010)]. More recent findings of elevated levels of deleterious polymorphisms or a greater frequency of unpreferred codons in selfing Arabidopsis (Cao et al. 2011; Qiu et al. 2011), Eichhornia (Ness et al. 2012), Capsella (Qiu et al. 2011; Brandvain et al. 2013; Slotte et al. 2013), Collinsia (Hazzouri et al. 2013), and Neurospora (Gioti et al. 2013) are consistent with the hypothesis of relaxed selection in selfing populations.

The efficacy of selection in selfing populations will also be affected by how strongly recessive mutations are masked by heterozygosity. A negative relation between the dominance of mutations and their deleterious effects has been hypothesized (Simmons and Crow 1977; Crow and Simmons 1983). Since deleterious mutations are on average partially recessive (h ∼ 0.2–0.25; Agrawal and Whitlock 2011; Manna et al. 2011), the more homozygous backgrounds of selfing populations offer fewer chances for the masking of mutations, and this should result in more effective selection against deleterious mutations (Pollak 1987; Caballero and Hill 1992; Charlesworth 1992; Glémin 2007). Indeed, theory predicts that strongly deleterious, highly recessive mutations are more likely to be purged by selection after recurrent inbreeding (Hedrick 1994; Wang et al. 1999), even though both strongly and mildly deleterious mutations contribute to inbreeding depression (Charlesworth and Charlesworth 1999; Wang et al. 1999; Charlesworth and Willis 2009). If a large proportion of deleterious mutations are strongly recessive, selfing populations could persist if deleterious mutation load is purged (Barrett and Charlesworth 1991; Glémin and Ronfort 2013). Purging has been documented across mammals, insects, mollusks, and plants, based on assays of fitness traits (Crnokrak and Barrett 2002). More recently, Szövényi et al. (2014) using divergence-based metrics found genomic evidence for stronger selective pressures acting on a moss species with haploid-dominant life cycle undergoing intragametophytic selfing. However, they found no evidence for the role of reduced N_e.

One reason for the weak empirical support for theoretical expectations on changes in selection efficacy could be that the effects of reduced N_e and increased homozygosity counteract one another in selfers. A reduction in N_e is expected to lower selection efficacy, whereas increased homozygosity should make selection more efficient by exposing recessive mutations. Glémin (2007) predicted that relaxed selection should be detected under a range of mutation and population size parameters in spite of the countervailing effect of homozygosity. Moreover, he found that divergence-based measures of selection were more likely to reveal differences compared to those based on polymorphism data. However, recently derived selfing lineages, which may be most common, are unlikely to have had sufficient time to accumulate many substitutions.

An approach for investigating the counteracting effects of reduced N_e and increased homozygosity is to characterize the underlying distribution of fitness effects of mutations using polymorphism data. In particular, the method of Keightley and Eyre-Walker (2007), jointly estimates selective and demographic parameters to infer the distribution of fitness effects (hereafter DFE) of new nonsynonymous mutations. This method allows a single population size change making it more robust to violations of assumptions due to demographic history and summarizes the effective strength of selection (N_es) acting on new nonsynonymous mutations. In selfing populations with reduced N_e, we expect a shift in the DFE such that there are a higher proportion of effectively neutral mutations (N_es < 1). More efficient purging of recessive deleterious mutations may shift the DFE in the opposite direction, resulting in a higher fraction of strongly deleterious mutations. However, the DFE approach assumes all mutations are codominant (h = 0.5). It is therefore unclear if this approach is capable of uncovering the joint effects of the dominance of mutations and genetic drift on patterns of selection.

We had two main objectives in this study. The first was to determine the extent to which the DFE inference approach may be useful for characterizing changes in selection efficacy following the transition from outcrossing to selfing, under a range of dominance parameters. To do this, we used forward population genetic simulations exploring how the estimated DFE shifts when the rate of selfing in a population increases. Our second objective was to investigate empirical support for theoretical predictions on the genomic consequences of the transition to selfing in the diploid aquatic flowering plant Eichhornia paniculata (Pontederiaceae) using the DFE inference approach.

Eichhornia paniculata populations possess a wide range of mating systems ranging from outcrossing to predominant selfing (Barrett and Husband 1990). This variation is associated with the evolutionary breakdown of tristyly, a floral polymorphism in which outbreeding populations are composed of three floral morphs maintained by negative frequency-dependent mating. All tristylous populations occur in the arid caatinga region of Northeast Brazil, where the species occurs in ephemeral ponds and ditches. Stochastic forces associated with small population size have destabilized tristyly on multiple occasions resulting in the spread and fixation of selfing variants (Barrett et al. 1989; Husband and Barrett 1992, 1993). This fixation of these variants has accompanied transitions in morph structure from stylar trimorphism through dimorphism to monomorphism. Independent transitions to selfing associated with long-distance dispersal have given rise to selfing populations in the Caribbean and Central America (Barrett et al. 2009). Mid-styled selfing variants occur commonly on Jamaica and Cuba, whereas long-styled selfing variants are restricted to a few small isolated populations in Nicaragua and Mexico. Roughly 60% of the variation in outcrossing rates in E. paniculata can be explained by the morph structure of populations and the frequency of selfing variants (Barrett and Husband 1990). Because trimorphic populations are predominantly outcrossing and monomorphic populations exhibit high rates of selfing, we use morph structure as a proxy for outcrossing rate in our study. Studies of allozyme variation and nucleotide diversity indicate that monomorphic populations of E. paniculata have low levels of heterozygosity (Glover and Barrett 1987; Barrett and Husband 1990; Husband and Barrett 1993; Ness et al. 2010), consistent with high selfing and/or colonization bottlenecks associated with long-distance dispersal. Purging of genetic load has been documented in E. paniculata based on a study on fitness traits (Barrett and Charlesworth 1991), but as yet there is no molecular genetic evidence for this phenomenon.

Our investigations address the following questions: (1) Is the DFE inference approach of Keightley and Eyre-Walker (2007) able to separate the contrasting genomic signals of reduced selection efficacy and stronger purifying selection against recessive mutations accompanying the transition to selfing? (2) Is there empirical support for these contrasting genomic signals based on the transition to selfing in E. paniculata? Our findings demonstrate that the DFE approach provides a valuable tool for investigating the genomic consequences of selfing, and that in E. paniculata transitions to selfing are accompanied by genome-wide influences of reduced N_e and the purging of recessive deleterious mutations.

Materials and Methods

Genome structure and content of simulated datasets

Using forward simulations, we investigated how patterns of selection change following the shift to selfing. We conducted forward simulations using the software SLiM (Messer 2013), which implements a Wright–Fisher model with selection and nonoverlapping generations. We started with a completely outcrossing population (t = 1) composed of 1000 individuals (census size, N) with 100-Mbp genomes and constant genome-wide mutation (μ = 7 × 10⁻⁹ per site per generation) and recombination (r = 5 × 10⁻⁸ per site per generation) rates. Genomes were composed of alternating 800 bp of noncoding (NC) and 200 bp of coding (C) DNA. All NC sites and 25% of C sites were neutral. The remaining 75% of sites in coding regions had selection coefficients drawn from a gamma distribution with shape parameter (β) of 0.3 and Ns of 0.5, 5, 15, 25, 35, 45, 55, 65, 75, 85, or 95. The proportion of sites with a given Ns is shown in Supporting Information, Figure S1. We assumed there were no beneficial mutations. To isolate the effect of the dominance of mutations, we used fixed h of 0.2, 0.5, or 0.8 across the entire genome.

Varying outcrossing rate, N, or r

We conducted three sets of simulations varying the outcrossing rate (t), N, or r parameters one at a time. In the first, we ran simulations for 10N generations and then introduced a split leading to a second population. The population resulting from the split had the same N as the ancestral outcrossing population (t = 1) but we decreased t to 0.02, effectively simulating a shift to selfing. From this set of simulations we estimated the realized N_e for selfing population by calculating the reduction in synonymous diversity. We ran a second set of simulations in which the population resulting from the split had the same N as the realized N_e of selfers from the first set of simulations, but with the outcrossing rate unchanged (t = 1). From the first set of simulations, we also calculated the effective recombination rate (r_e) of the selfing population using the equation r_e = r_outcrosser × (1 − F_is), where F_is is the coefficient of inbreeding (Nordborg 2000). As we were unable to change r midway during the simulations, we ran a third set of simulations where the ancestral outcrossing population had the same r as the r_e of selfers from the first set of simulations. For this set of simulations, we did not create a population split. We ran all simulations for another 6N generations to allow populations to reach equilibrium, as expected under coalescent predictions. During this period, we randomly sampled eight individuals from each population at each 1N interval, and calculated the number of nonsynonymous and synonymous mutations that had accumulated independently. We use these values to generate folded allele frequency spectra (hereafter AFS). We estimated the DFEs for all simulated datasets using the approach of Keightley and Eyre-Walker (2007), as described in the section below. For each N_es category in the DFE, we generated confidence intervals (C.I.s) based on 120 independent runs. We performed two-sample t-tests to compare the DFEs of simulated outcrossing and selfing populations using R (R Development Core Team 2011). We also reported the expected DFE for the outcrossing population represented as a gamma distribution using the β and Ns parameters. As a distribution of Ns was used, we calculated the DFE for each Ns value and scaled the individual DFEs by their proportional contribution to the genome (Figure S1). To generate the expected DFE for the selfing population under an additive model, we first scaled the Ns parameter by the observed reduction in synonymous diversity accompanying the shift to selfing but leaving β unchanged. Further, we multiplied Ns by 1.96 to account for the homozygosity of selfing genomes (1 + F_is, see Caballero and Hill 1992).

Jointly varying h and Ns

We ran separate simulations to assess the joint impact of the dominance level of mutations and strength of purifying selection acting on them. In these additional simulations, we assumed h = 0.2 and Ns = 95 for 31.5% of the sites in the C region to simulate strongly recessive deleterious sites. The remaining 43.5% of the sites in the C region had h = 0.5 and Ns ranging from 0.5 to 85 to simulate partially codominant weakly and strongly deleterious sites. After simulating a split to a predominant selfing population (t = 0.02), we calculated the average fitness and number of neutral and deleterious mutations that accumulated in eight randomly sampled individuals. As SLiM outputs haploid chromosomes, we randomly paired two genomes to construct a diploid individual. For a given site in each of these individuals, the fitness effect of nonsynonymous mutations that were present once was 1 − hs. If the same mutation occurred in both chromosome copies, its fitness effect was 1 − s as they would be homozygous. The fitness of an individual was the multiplicative product of the fitness effect of all mutations in their genome. Note, because highly selfing populations are mostly homozygous, this procedure does not simulate the true genotypic composition of the population, but allows for a more direct comparison with outcrossers of fitness and deleterious mutation accumulation. We estimated the DFEs for the simulated datasets using the approach of Keightley and Eyre-Walker (2007), as described in the section below. We generated C.I.s based on 120 independent runs of simulations and performed a two-way analysis of variance with mating system and time as factors using R (R Development Core Team 2011).

Estimating selective and demographic parameters

We inferred the DFE for each dataset using the source code version of the DFE-α software available from Peter Keightley’s website (http://lanner.cap.ed.ac.uk/~eang33/est-dfe-files.tar.gz). To ensure our parameter estimates from this maximum likelihood method reached global optima, we randomized the starting values for mean selection against deleterious mutations (N_eE(s)), β, and t₂ parameters and ran the program 10 times. From the output estimates across the runs, we chose the parameters that resulted in the highest maximum likelihood. We report the proportion of mutations falling into a given N_es range of the DFE (<1, 1–10, 10–100, and >100).

Sampling of Eichhornia paniculata populations

We sampled open-pollinated seeds from 20 populations of E. paniculata across the species range, including 10 outcrossing (tristylous) and 10 selfing populations (Table S1). All outcrossing populations were from Northeast Brazil. Ness et al. (2010) found that the outcrossing populations from this region clustered in the same genetic structure group. We studied a subsample of the genotypes used by Ness et al. (2010). The selfing populations originated from Cuba (n = 3), Jamaica (n = 5), Mexico (n = 1), and Nicaragua (n = 1). We grew plants under uniform glasshouse conditions at the University of Toronto. To maximize our sampling effort, we selected one individual per population following the scattered sampling approach (Wakeley and Lessard 2003; Städler et al. 2009), which assumes that alleles coalesce faster within compared to between demes, thus maximizing the number of unique alleles represented.

Nucleic acid extraction and sequencing

We extracted both RNA and DNA from floral buds of E. paniculata. We chose floral buds as they contain both gametophytic and sporophytic genes increasing our sampling breadth. We extracted RNA using the Spectrum Plant Total RNA kit (Sigma-Aldrich). The extracted RNA samples were used to make Illumina TruSeq RNA libraries that were sequenced using the 100-bp paired end protocol on Illumina HiSequation 2000 at the McGill University and Génome Québec Innovation Centre. The samples were sequenced across two lanes with 10 samples multiplexed in each lane, with outcrossing and selfing samples evenly distributed across the lanes. We further extracted genomic DNA from the floral buds of the Mexican sample using a modified variant of the CTAB extraction protocol (Doyle and Doyle 1987, 1990; Edwards et al. 1991) treating the lysed cells with ribonuclease A to remove contaminant RNA. An Illumina TruSeq DNA library was prepared from this sample and sequenced on a separate lane on the Illumina HiSequation 2000. After sequencing, we removed reads of <50 bp and reads with >10% of “N” bases using a custom Perl script, retaining ∼92% of the original data. The raw sequence data are available under accession no. SRP049636 at the Sequence Read Archive (http://www.ncbi.nlm.nih.gov/sra), and the associated BioProject accession no. is PRJNA266681 (http://www.ncbi.nlm.nih.gov/bioproject/).

Assembly and identification of coding regions

Although Ness et al. (2011) generated a transcriptome reference for E. paniculata, we conducted an independent assembly as the greater number of genotypes, longer reads, and new assembly software facilitated more accurate construction of contigs. Nevertheless, we followed a similar approach as in Ness et al. (2011). We chose selfing genotypes to generate the consensus assembly under the expectation that their more homozygous genomes would limit the problem caused by alleles that appear heterozygous but actually belong to paralogous loci. We generated a de novo assembly from the Jamaican transcriptome samples using the programs Velvet 1.2.08 (Zerbino and Birney 2008) and Oases 0.2.08 (Schulz et al. 2012). We found the best parameters for the assembly using VelvetOptimser 2.2.4 (Zerbino 2010) that indicated that k-mers with a length of 75–85 were optimal. The assembled reference transcriptome had a total size of 65.53 Mbp with an N₅₀ of 2.2 kb (Figure S2). For the contigs in the assembled reference, we predicted the location of coding regions through BLAST searches to known proteins in plant databases and removed contigs without any matches. We did this using a combination of tBLASTx (Altschul et al. 1990) to Viridiplantae database and GeneWise 2.4.1 (Birney et al. 2004). After identifying coding regions in assembled contigs, we trimmed bases before start and after stop codons using custom Perl scripts retaining 29,336 loci at the end of this step.

Read mapping

After generating the reference assembly, we mapped all reads from E. paniculata transcriptome samples and the genomic DNA sample to the assembly. First, we mapped short reads with Burrows-Wheeler Aligner (BWA, v0.6.2-r126) using default parameters (Li and Durbin 2009). Further, we used the BWA “sampe” command to combine the paired end read mapping results together and Samtools “view” command (Samtools v0.1.18 r982:295; Li et al. 2009) to convert the mapping results into a binary alignment format (BAM format). After the first round of read mapping, we used Stampy 1.0.20 software with default parameters (Lunter and Goodson 2011) to map more divergent reads, as well as to identify insertions and deletions (indels). Approximately 92% of filtered reads for each sample mapped successfully. Next, we processed the read mapping output into a format required for variant calling software using four programs (SamFormatConverter, ReorderSam, AddOrReplaceReadGroups, and BuildBamIndex) that were part of the Picard tools package 1.100 using default settings (http://picard.sourceforge.net). The processed BAM files are available under accession no. SRP049636 at the Sequence Read Archive (http://www.ncbi.nlm.nih.gov/sra).

Variant calling and filtering

After read mapping, we identified single nucleotide polymorphisms (SNPs) segregating among E. paniculata transcriptome samples using a set of programs from the Genome Analysis Toolkit (GATK) v2.7-4-g6f46d11 (Depristo et al. 2011). We first used RealignerTargetCreator and IndelRealigner with default parameters and identified and realigned sequences within ∼3 kb of an indel where erroneous mismatches were more likely to have occurred. Further, we ran UnifiedGenotyper with the BadCigar read filter to call invariants sites, SNPs, and indels from all samples. For a given variant, we allowed for a maximum of six alternate alleles to be genotyped. We performed a number of filtering steps to minimize spurious SNP calls. From the UnifiedGenotyper output, we only retained sites for which Phred scaled quality score (QUAL) was >60 and depth in each individual sequenced was >20. We only retained a SNP if the Phred scaled genotype quality (GQ) for all samples was >60. We also excluded 5 bp on either side of an indel where spurious SNP calls are likely to be made. For the SNPs identified in the outcrossing genotypes, we performed a test for Hardy–Weinberg equilibrium (Wigginton et al. 2005) as implemented in VCFtools (v0.1.11) (Danecek et al. 2011). We validated SNP calls by comparisons to a range-wide polymorphism dataset for E. paniculata generated by Barrett et al. (2009) and Ness et al. (2010). These studies extensively sampled 225 E. paniculata individuals from 25 populations across Brazil, the Caribbean, and Central America. We performed a custom BLAST search comparing the 10 EST-derived nuclear loci investigated by these authors to our reference using the makeblastdb and blastn programs, part of the NCBI BLAST+ toolkit, version 2.2.26 (Camacho et al. 2009). Seven nuclear loci matched contigs in our dataset that we had removed during the filtering stages. We aligned the remaining three nuclear loci and sequences for the best matching contig in our study using MUSCLE (Edgar 2004) as implemented in MEGA6 (Tamura et al. 2013) excluding sites with gaps. Only the alignment between EP0314 and the best matching locus from our study (Locus_13533) identified polymorphic sites segregating in both sets of sequences and the region-specific polymorphisms segregating in our samples matched those identified in the previous studies (Table S2).

Filtering paralogous sequences

If duplicated regions were assembled into a single contig, differences between paralogs might erroneously be called as a polymorphic difference at a single site. To address this we first removed loci containing sites heterozygous in all 10 selfing genotypes identified with the aid of VCFtools (v0.1.11) (Danecek et al. 2011). Such variants are unlikely to be real, given the largely homozygous backgrounds of selfing populations and the occurrence of at least two independent transitions to selfing (Barrett et al. 2009). Second, we mapped genomic reads from a selfing genotype from Mexico to the transcriptome reference. We used the Samtools “depth command” (Samtools v0.1.18 r982:295; Li et al. 2009) to find depth per site and used a custom Perl script to calculate the mean across a given contig. We removed loci with <15× or >60× genomic coverage (Figure S3) based on the assumption that coverage should be even across the entire genome.

Removing shared variant sites

To focus on contemporary selection pressures, we used only sites that were unique to each lineage in comparisons of selection efficacy between outcrossing and selfing populations. We used a custom Perl script to convert the UnifiedGenotyper VCF output into a FASTA format. Next, we identified polymorphic sites that were shared between outcrossing and selfing populations using a custom Perl script. We substituted bases at those sites in our FASTA files to “N” using a script from http://raven.iab.alaska.edu/~ntakebay/teaching/programming/perl-scripts/perl-scripts.html (selectSites.pl). We retained a total of 16,416 transcripts at the end of the filtering stages. Note that our simulations used a similar approach, as only lineage-specific mutations were analyzed while inferring the DFE.

Comparing outcrossing and selfing populations of Eichhornia paniculata

We first processed the output from the SNP filtering steps and generated the necessary variables so we could compare the efficacy of selection in outcrossing and selfing populations. Because selfers are more highly homozygous than outcrossers, they have effectively half the number of chromosomes. Therefore, we generated a haploid copy of the diploid chromosome for each individual and randomly chose one of the bases at a given heterozygous site using a custom Perl script. As we only had two samples from Central America, we were unable to characterize the AFS and therefore we did not include the Central American populations in most comparisons. To keep the number of chromosomes analyzed the same while performing comparisons, 8 of the 10 outcrossing genotypes from Brazil were randomly selected.

Estimating the strength of selection acting on mutations in Eichhornia paniculata populations

We used the Polymorphorama script (Andolfatto 2007; Haddrill et al. 2008) to generate locus-specific nonsynonymous and synonymous folded AFS and number of invariant sites for the outcrossing and selfing populations. This program includes twofold degenerate sites while calculating the AFS. Twofold degenerate sites might inflate the number of nonsynonymous compared to synonymous polymorphisms resulting in biased conclusions on estimates of selection. However, Williamson et al. (2014) reported that DFE inferences were unaffected by the use of such twofold sites. We generated 200 bootstrap replicate E. paniculata datasets, after resampling randomly across the loci using R (R Development Core Team 2011), and generated the sums and means for each dataset. We inferred the DFE for each replicate dataset as described earlier in this section. For the resulting N_es categories in the DFE, we generated mean values and 95% C.I.s across the 200 bootstrap replicates. To generate the C.I.s, we excluded the top and bottom 2.5% of bootstrap replicates and used the smallest and largest values from the remainder to represent the lower and upper limits of the C.I., respectively. We determined whether the N_es categories from the DFE differed between outcrossing and selfing populations using the randomization test following Keightley and Eyre-Walker (2007). A difference in one N_es category of the DFE will not be independent of differences in the others. Finally, we used the β and N_eE(s) parameters estimated by the DFE-α software to plot the cumulative proportion of nonsynonymous mutations. This allowed us to have more confidence in our inferences from the DFE, even if mutations of extremely large deleterious effect (N_es >> 100) were absent in the samples analyzed.

Results

Inferring the DFE of simulated outcrossing and selfing populations

Fixed dominance coefficients across the genome:

Using forward simulations, we investigated changes in selection efficacy in outcrossing and selfing populations after a split from a common outcrossing ancestor. After 6N generations, there was ∼75% reduction in synonymous diversity following the shift to selfing. As the level of dominance across the entire genome was increased from 0.2 to 0.8, the DFE of outcrossing populations shifted such that a larger fraction of sites was inferred to be under stronger levels of selection (Figure 1). The DFE of outcrossers under h = 0.5 matched the expectations based on the given Ns and β parameters. Consistent with expectations, there was a deficit of sites in the N_es > 100 category of the DFE of the outcrossing population under h = 0.2, when compared to the level expected under an additive model. In contrast, the DFE for selfing populations remained largely the same as the level of dominance increased. Under all three dominance levels, there was an excess of sites under stronger levels of selection than expected due to the N_e reduction and increased homozygosity accompanying the shift to selfing. Under h = 0.2, π_n/π_s was smaller in selfers compared to outcrossers (Table S3), a significantly larger fraction of sites was in the N_es > 100 category, and a significantly smaller fraction of sites was in the N_es 10–100 categories of the estimated DFE of selfers (Figure 1, Table S4), consistent with the effects of purging. In contrast, under h = 0.8, selfers had a larger π_n/π_s (Table S3) and a significantly larger and smaller fraction of sites in the N_es < 1 and N_es 10–100 categories, respectively (Figure 1, Table S4). These results indicate that with increasingly recessive mutations there is power to detect a purging effect in selfing populations, whereas with increasing dominance there is evidence for a greater proportion of effectively neutral mutations compared with outcrossing populations.

Figure 1.

Distribution of fitness effects (DFE) of new nonsynonymous mutations for simulated outcrossing (t = 1.00) and selfing (t = 0.02) populations under various fixed dominance coefficients (h = 0.2, 0.5 or 0.8). N_es is the product of N_e and the selection coefficient (s). Shown are (A) the expected DFEs, (B) h = 0.5, (C) h = 0.2, and (D) h = 0.8. Simulations illustrate the change in DFE for both populations after 6N generations after a split from the common outcrossing ancestor. The coding regions were under various selection coefficients (Ns = 0.5–95) all sampled from a gamma distribution with shape parameter (β) of 0.3. We generated the expected DFE for the outcrossing population represented as a gamma distribution using the β and Ns parameters. We generated the expected DFE for the selfing population by scaling the Ns parameter by the observed 75% reduction in synonymous diversity in selfers and multiplying it by 1.96 to account for effective dominance levels of mutations in selfing genomes while leaving β unchanged. We generated the observed DFEs by randomly sampling and generating allele frequency spectra using eight genomes from populations of 1000 in size. Shown are the mean proportions of sites for each N_es category and their respective confidence intervals based on 120 simulations.

Nonequilibrium demographic factors:

The difference between the expected and estimated proportion of strongly selected sites in selfing populations may reflect a failure to fully control for nonequilibrium demographic factors caused by the transition to selfing and the effects of Hill–Robertson interference. Under all dominance levels, the synonymous AFS for outcrossing and selfing populations were skewed toward rare variants, compared to expectations under neutral equilibrium (Figure S4, A and B). However, such a skew was not observed in the outcrossing population when we included mutations that were shared between both populations (Figure S4C). The skew was likely due to the use of unique mutations to estimate selection pressures, which will be rare, given their recent origin. In contrast, selfers still showed a skewed AFS even when all mutations were included (Figure S4D). The inclusion of shared mutations did not affect inferences about the effect of the dominance of mutations on selection efficacy (Figure S5). However, there was a larger difference between the observed and expected number of mutations in the N_es > 100 category under h = 0.5 and h = 0.8. Additionally, as the simulated genomes had a much larger fraction of noncoding compared to coding regions (Figure S1), our analysis may be biased by the fact that a large proportion of neutral sites would have been far from selected sites. Using synonymous mutations occurring only within coding regions did not affect the observed differences between the DFEs of outcrossing and selfing populations (Figure S6), although the estimated C.I.s were larger.

Reduced N_e and stronger background selection:

To disentangle how reduced N_e and r_e interact with the dominance of mutations, we varied these parameters one at a time, simulating outcrossing populations with equivalent reductions in effective size and recombination to our simulated selfing population. There was an increase in the fraction of sites in the lower N_es categories of the observed DFE under N = 250, equivalent to the realized neutral N_e of the previously simulated selfing population, when compared to N = 1000 (Figure 2). The observed DFEs for N = 250 under all dominance levels harbored a greater fraction of sites in the larger N_es categories when compared to the expected levels. However, simulating this reduction in N_e did not show the purging signal we observed in selfing populations; there was no evidence for a significant increase in the proportion of strongly selected mutations in simulations with recessive mutations (compare Figure 2C with Figure 1). The DFE of the outcrossing population with r = 1.96 × 10⁻⁹ per site per generation, equivalent to the realized r_e of previously simulated selfing populations, had a larger fraction of mutations in smaller N_es categories when compared to the DFE of the outcrossing population with r = 5 × 10⁻⁸ per site per generation under h = 0.5 and under h = 0.8 (Figure 1 and Figure S7). However, this change was not as large as the level expected due to the N_e reduction accompanying the shift to selfing. As with the simulations of outcrossing populations subjected to reduced N_e, we found no evidence of the purging signal of a larger fraction of strongly deleterious mutations compared with the ancestral outcrosser, again showing that this signal is specific to the transition to selfing. Thus, these simulations indicate that the signal of increased proportions of strongly selected sites is a unique effect of the transition to selfing and driven by the purging effect of increased homozygosity.

Figure 2.

Distribution of fitness effects (DFE) of new nonsynonymous mutations for simulated outcrossing populations of two census sizes (N) under various fixed dominance coefficients (h = 0.2, 0.5, or 0.8). Shown are (A) the expected DFEs, (B) h = 0.5, (C) h = 0.2, and (D) h = 0.8. Simulations illustrate the change in DFE for both populations after 6N generations after the split of a population of N = 250 from the common ancestral population of N = 1000. The coding regions were under various selection coefficients (Ns = 0.5–95) all sampled from a gamma distribution with shape parameter (β) of 0.3. We generated the expected DFEs for the populations represented as a gamma distribution using the β and Ns parameters. We generated the observed DFEs by randomly sampling and generating allele frequency spectra using eight genomes from populations of 1000 in size. Shown are the mean proportions of sites for each N_es category and their respective confidence intervals based on 120 simulations.

Mixed dominance coefficients across genome:

When both dominance and selection coefficients were jointly varied, selfers accumulated more nonsynonymous mutations compared to outcrossers, even though both accumulated the same number of synonymous mutations (Figure 3A). There was a greater decline in the fitness of selfers compared to outcrossers over 6N generations (Figure 3B). We further investigated how such mutation accumulation influenced the selection efficacy for outcrossing and selfing populations. Mean π_n for simulated outcrossers and selfers was 1.98 × 10⁻⁵ and 3.64 × 10⁻⁶, respectively; mean π_s for simulated outcrossers and selfers was 4.94 × 10⁻⁵ and 8.53 × 10⁻⁶, respectively. After the split from the common outcrossing ancestor, the estimated DFE of the selfing population shifted such that a significantly larger fraction of sites were in the N_es < 1 category at 1N and 3N generations (Figure 4, Table S5). Also, there was a significantly larger fraction of sites in the N_es > 100 category of the selfing population during the first 4N generations. The interaction between mating system and time was a significant factor influencing the proportion of mutations in the N_es < 1 and N_es > 100 categories. Moreover, the number of sites in each N_es category of the DFE fluctuated over generations. When there were both weakly selected additive sites and strongly selected recessive sites in the simulated genome, the DFE approach inferred the occurrence of more very weak and very strong deleterious mutations in selfers, although such effects varied with the time since the transition to selfing. Thus, with a model with varying dominance coefficients, it is possible to detect signals of both relaxed purifying selection and purging, although the relaxed selection signal is less prevalent, despite an observed decline in fitness in selfing populations.

Figure 3.

The type and effects of mutations accumulated by four simulated outcrossing (black) and four simulated selfing (gray) individuals over 6N generations after a split from a common outcrossing ancestor when both dominance and selection coefficients were jointly varied. We randomly sampled eight haploid genomes from populations of 1000 in size. The genomes were randomly paired to create four diploid individuals. Shown are (A) the number of nonsynonymous and synonymous mutations that accumulated in the genomes and (B) mean and 95% confidence intervals for the fitness of individuals.

Figure 4.

Distribution of fitness effects (DFE) of new nonsynonymous mutations for simulated outcrossing (t = 1.00) and selfing (t = 0.02) populations when both dominance and selection coefficients were jointly varied. Within the coding region of the genome in the simulations, 31.5% of sites had h = 0.2 and Ns = 95, 43.5% of the sites had h = 0.5 and Ns ranging from 0.5 to 85, and the remaining had h = 0.5 and Ns = 0. Simulations illustrate the change in DFE for both populations over 6N generations after a split from the common outcrossing ancestor. We generated DFEs by randomly sampling and generating allele frequency spectra using eight genomes from population sizes of 1000. N_es is the product of N_e and the selection coefficient (s). Shown are the mean proportions of sites for each N_es category and their respective confidence intervals based on 120 simulations.

Inferring the DFE of outcrossing and selfing populations of Eichhornia paniculata

Selection and demographic parameters:

Of the 16,416 loci, 4485 were polymorphic in outcrossers and 1586 were polymorphic in selfers, after excluding variants shared between outcrossing and selfing populations. Polymorphic loci in outcrossers and selfers had between 1–40 and 1–20 segregating sites, respectively, with the distribution of polymorphisms per locus right skewed (Figure S8). Mean π_n for outcrossers and selfers was 4.93 × 10⁻⁴ and 8.99 × 10⁻⁵, respectively; mean π_s for outcrossers and selfers was 3.09 × 10⁻³ and 5.62 × 10⁻⁴, respectively. We identified 20,388 nonsynonymous and 32,384 synonymous polymorphisms in outcrossers and 3613 nonsynonymous and 5730 synonymous polymorphisms in selfers (Table S6). There was an excess of rare nonsynonymous compared to synonymous polymorphisms in both outcrossing and selfing populations (Figure 5). Selfers had a slight deficit of rare synonymous polymorphisms compared to outcrossers. Randomization tests, as implemented in Keightley and Eyre-Walker (2007), indicated there were significantly fewer sites in the N_es: 1–10 and 10–100 categories and a significantly greater proportion of sites in the N_es > 100 in selfers compared to outcrossers (Figure 6). Although we did not detect a significant difference for the N_es < 1 category at the 5% level, there was a general trend for a larger number of mutations in this category for selfing populations. Our plots of the cumulative proportion of nonsynonymous mutations indicate that selfers had a marginally larger fraction of mutations with N_es < 3 (Figure S9).

Figure 5.

Folded nonsynonymous and synonymous allele frequency spectra for outcrossing and selfing populations of Eichhornia paniculata. Haploid chromosomes from one individual from each of eight outcrossing populations from Northeast Brazil and eight selfing populations from the Caribbean were used to generate the frequency spectra. Eight outcrossing individuals from the 10 that were sequenced were randomly selected to keep the number of chromosomes sampled the same while performing the comparison.

Figure 6.

Distribution of fitness effects (DFE) of new nonsynonymous mutations for outcrossing and selfing Eichhornia paniculata. N_es is the product of N_e and the selection coefficient (s). Eight Caribbean selfing and eight outcrossing individuals were used to generate the DFEs. Eight outcrossing samples from the 10 that were sequenced were randomly selected to keep the number of chromosomes sampled the same while performing the comparisons. Error bars on top of each N_es category are 95% confidence intervals from 200 bootstrap replicates generated by resampling over loci. We used a randomization test (see Keightley and Eyre-Walker 2007) to compare outcrossing and selfing populations and to assess significance at 0.5% level (indicated by *).

Data filtering:

As inferring the DFE is dependent on the AFS being compared, we investigated how different quality cut-offs and filtering affected the patterns observed. A single site quality cut-off for both invariant and variant sites could disproportionally exclude invariant sites. However, in our case, reducing the quality cut-off for both sites did not affect the DFE inference for outcrossing and selfing populations (Figure S10). We also excluded an additional 245 loci that contained at least one site heterozygous in two or more selfers, a more stringent filter for potential paralogs (Table S7). Removing such loci from the comparisons did not alter our inferences (Figure S11). Less than 3% of polymorphic sites across outcrossing genotypes departed from Hardy–Weinberg equilibrium at the 5% level and these sites were localized to 73 loci. Again, removal of these loci did not affect our inferences on differences between outcrossing and selfing genotypes (Figure S12). Finally, we investigated how the relative proportions of nonsynonymous and synonymous polymorphisms per loci influenced our results. As the distribution of π_n/π_s was right skewed (Figure S13), we excluded 135 loci that had π_n/π_s > 1 in either outcrossers or selfers and repeated the DFE analyses. Although we still observed an excess of sites in the N_es > 100 category in selfers, there was substantial overlap in the proportion of sites in the N_es < 1 category between outcrossers and selfers (Figure S14 and Table S8).

Pooling independent shifts to selfing:

We investigated whether pooling additional samples from an independent transition to selfing influenced the AFS and the inferred DFE. As the inclusion of the two additional samples from Central America resulted in a total of 10 selfing samples, we repeated the analyses using all 10 outcrossing genotypes to keep the number of chromosomes compared the same. With the inclusion of the additional selfing genotypes, we still observed an excess of rare nonsynonymous compared to synonymous polymorphisms in both outcrossing and selfing populations (Figure S15). In the synonymous AFS, selfing populations had a deficit of singletons and an excess of doubletons compared to outcrossing populations. Moreover, we found that using the randomization tests, the underlying DFE and all discrete N_es categories (Figure S16) were significantly different between outcrossing and selfing populations at the 0.5% level.

Discussion

Our forward population-genetic simulations demonstrate that reduced N_e and increased purging, accompanying the transition to selfing, can be detected by estimating the underlying distribution of deleterious mutational effects. Our simulations show that this is not expected under a single dominance coefficient for nonsynonymous mutations, but can be explained if there is a mixture of strongly selected recessive mutations that experience effective purging in selfing populations, and more weakly selected additive mutations subject to relaxed selection. Our empirical study of E. paniculata revealed a small increase in the proportion of effectively neutral nonsynonymous mutations and a significant increase in the proportion of strongly selected sites in selfing compared to outcrossing populations. The spread in the distribution of selective coefficients observed in both the simulated and empirical datasets is consistent with the effects of both purging and a reduced efficacy of selection in selfing populations.

Consequences of reduced N_e in selfing populations

Our simulations revealed an increased number of sites that were under weaker purifying selection following the transition to selfing as a result of reduced N_e and r_e. Further, relaxed selection was driven largely by partially dominant rather than recessive mutations. The observed pattern is consistent with theoretical predictions that selection should be less efficient in selfers (reviewed in Charlesworth and Wright 2001). In both the simulated and empirical results, the estimated magnitude of relaxed selection appears to be small (Figure 4 and Figure 6), and there may be several reasons for this. First, it is possible that even a large reduction in N_e does not substantially alter the DFE in selfing populations if a significant proportion of mutations are strongly selected (N_es >>1). Second, the relatively recent evolutionary transition to selfing in E. paniculata (see Ness et al. 2010) may not have allowed for sufficient time for the effects of reduced N_e to be detected at genome-wide scales. However, our simulations examining the shift to selfing over 6N generations did not find a very large reduction in the N_es < 1 category of the DFE and patterns of relaxed selection disappeared after 3N generations. Although reduced N_e may only moderately reduce genome-wide selection efficacy, our simulation results imply that such subtle shifts may be associated with early and cumulative declines in fitness in selfing populations (Figure 3).

The smaller than expected reduction in the efficacy of selection following the shift to selfing in the forward simulations might suggest that the DFE inference approach underestimates the effects of reduced N_e. If the method underestimated the magnitude of relaxed selection, it could explain why the observed shift in the empirical data toward effectively neutral sites in the DFE of selfers was relatively weak. An alternate possibility is that even genome-wide polymorphism datasets have limited power to detect relaxed selection (and see Glémin 2007), particularly when there are multiple and contrasting forces acting on the genome. Consistent with this suggestion, the difference between π_n/π_s in simulated outcrossers and selfers under an additive model incorporating mixed dominance and selection coefficients was small. Even so, the empirical pattern of relaxed selection became less apparent when we excluded loci with π_n/π_s > 1, indicating they might be contributing to genome-wide patterns.

Our empirical results demonstrate demographic effects associated with the colonization history of selfing populations of E. paniculata. Pooling genotypes from independent shifts to selfing had a significant influence on the AFS, as indicated by the excess of doubletons in selfers. The Central American genotypes were fixed for alternate alleles compared to samples from the Caribbean, probably explaining this result. Eichhornia paniculata occurs in ephemeral aquatic habitats and populations experience striking annual variation in size, including frequent local extinctions (Barrett and Husband 1997; Husband and Barrett 1998). Recent studies suggest that nonsynonymous mutations reach equilibrium faster than synonymous mutations after bottlenecks; therefore comparisons of the two could generate spurious signals of relaxed selection (Pennings et al. 2014; Simons et al. 2014). Because E. paniculata only colonized the Caribbean ∼125,000 years ago (Ness et al. 2010), populations from that region might not have had sufficient time to recover from founding events. If the fit of demographic models for selfing populations was less accurate compared to outcrossing populations, our signal of relaxed selection could, in part, reflect this effect. However, we found that even when the synonymous spectra did not meet neutral expectations, the inference of the DFE was similar to the case when the spectra did match such expectations. Although the demographic correction implemented in the DFE inference approach should account for distortions of the AFS away from equilibrium, it is unclear if the marginal trend of relaxed selection in the empirical data, which was significant when samples from separate origins were pooled, was influenced by violations of neutral expectations. Our simulations gave us qualitatively similar differences in the DFE to the empirical data and provided evidence for a fitness decline and the accumulation of deleterious mutations in selfing populations (Figure 3). These results suggest that the weak trend observed in our empirical results reflects a true signal of the reduced efficacy of selection in selfing populations.

Exposure of recessive deleterious mutations

We observed a large proportion of strongly deleterious sites in the DFEs obtained from both simulated and empirical data. By using forward simulations, we found the DFE approach was able to infer the effect of varying dominance levels of mutations on patterns of selection. As deleterious mutations became more dominant, the effective strength of selection acting against them in outcrossing populations increased. In contrast, the dominance level of mutations had little effect in selfing populations, where the largely homozygous background of genotypes exposes the deleterious effects of recessive and dominant mutations. Further, our simulations indicated the DFE approach was able to identify how both partially dominant and strongly recessive deleterious mutations shape selection efficacy experienced by selfers and outcrossers. Linkage between neutral, weakly, and strongly deleterious mutations, under variable dominance coefficients, may have complicated attempts to disentangle their individual effects on the DFE but did not lead to any directional biases. Overall, our simulation results indicated the empirical patterns of increased accumulation of effectively neutral and strongly deleterious mutations occur across a wide range of selection and dominance coefficients, after transitions to selfing. Previous experimental studies in E. paniculata, involving multigenerational fitness comparisons of phenotypic traits following selfing and outcrossing, were unable to detect significant inbreeding depression in selfing populations from the Caribbean and the authors proposed that this was largely a result of purging of genetic load (Barrett and Charlesworth 1991). Our findings are consistent with this earlier study in providing molecular evidence to support the purging hypothesis.

A key question is whether the observed excess of strongly deleterious mutations in selfing populations was, in part, due to the methods used to infer selection efficacy. The approach of Keightley and Eyre-Walker (2007) is known to imprecisely infer the DFE when the distribution is multimodal (Kousathanas and Keightley 2013) and overestimates the strength of purifying selection in the face of linked selection (Messer and Petrov 2013). Also, the DFE estimation approach infers the proportion of mutations that are strongly deleterious (N_es > 100) from the β parameter, even if no such mutations segregate in samples. However, plotting the DFE as a continuous distribution indicated there was a greater spread of selection coefficients in selfers. Furthermore, the magnitude of the shift toward effectively neutral sites in simulated selfing populations was lower than expected due to reduced N_e, or the level experienced by simulated outcrossing population with a census size equivalent to the realized N_e of the selfers. Both of these results suggest that countervailing homozygosity effects stifle a unidirectional change in the DFE. Therefore, the observed patterns are consistent with an effect of both purging and reduced efficacy of selection in selfing populations, even if estimates of the magnitude of the effect may be imprecise.

At this stage we are unable to assess if an assumption of the DFE inference approach, namely that all sites experienced independent selection pressures, was violated in our study. It seems probable that this assumption would be violated in selfing populations, with many neutral sites linked to deleterious ones. In this case, the DFE approach might overestimate the proportion of deleterious sites. However, over short distances in the genome our bootstrapping by locus approach addresses some of the uncertainty due to linkage blocks, as some bootstrap replicates may not contain all of the genes within a given block. Furthermore, our simulations suggest the patterns we observed are expected under a model that incorporates both strongly recessive deleterious mutations and slightly deleterious, more additive mutations, consistent with a role for purging of harmful recessive alleles and an accumulation of weakly deleterious mutations.

Finally, we have not considered the effects of beneficial mutations in interpreting the results of our study. Recessive beneficial mutations may also be exposed in selfing populations leading to selective sweeps. As beneficial mutations are expected to be fixed rapidly after they appear in selfing populations, they are unlikely to have been segregating in the small number of E. paniculata samples used in our study. Linkage between recessive beneficial and deleterious mutations should increase the time taken for sweeps to occur (Hartfield and Glémin 2014). Hence, future studies estimating the rate of adaptation in selfing populations should consider the influence of strongly deleterious mutations, as our results indicate that they can influence genome-wide patterns of selection.