Genomic implications of the repeated shift to self-fertilization across a species’ geographic distribution

Abstract

The ability to self-fertilize often varies among closely related hermaphroditic plant species, though, variation can also exist within species. In the North American Arabidopsis lyrata, the shift from self-incompatibility (SI) to selfing established in multiple regions independently, mostly since recent postglacial range expansion. This has made the species an ideal model for the investigation of the genomic basis of the breakdown of SI and its population genetic consequences. By comparing nearby selfing and outcrossing populations across the entire species’ geographic distribution, we investigated variation at the self-incompatibility (S-)locus and across the genome. Furthermore, a diallel crossing experiment on one mixed-mating population was performed to gain insight into the inheritance of mating system variation. We confirmed that the breakdown of SI had evolved in several S-locus backgrounds. The diallel suggested the involvement of biparental contributions with dominance relations. Though, the population-level genome-wide association study did not single out clear-cut candidate genes but several regions with one near the S-locus. On the implication side, selfing as compared to outcrossing populations had less than half of the genomic diversity, while the number and length of runs of homozygosity (ROHs) scaled with the degree of inbreeding. Selfing populations with a history of long expansion had the longest ROHs. The results highlight that mating system shift to selfing, its genetic underpinning and the likely negative genomic consequences for evolutionary potential can be strongly interlinked with past range dynamics.

Introduction

Shifts in the mating system, often from outcrossing to self-fertilization (selfing), have evolved repeatedly in many flowering plant families (Igic et al. 2008), affecting about 10%–15% of all extant species (Goodwillie et al. 2005). Mating system may also be variable within a species, whereby selfing is maintained in generally outcrossing species or populations at low frequencies (Goodwillie et al. 2005; Willi and Määttänen 2010). Why mating system variation exists and is maintained has been a long-standing debate (Darwin 1877; Stebbins 1957; Wright et al. 2013). While selfing may be beneficial under certain conditions, for example, providing reproduction assurance when only few individuals colonize new or remote habitats (Pannell et al. 2015), it leads to inbreeding depression (Charlesworth and Willis 2009) and is predicted to become an evolutionary dead-end in the long run (Stebbins 1957; Goldberg et al. 2010). However, there are also conflicting theoretical findings, suggesting that selfing taxa should establish larger ranges (Eriksson and Rafajlović 2021), speaking against a consistent genetic burden of selfing. A refined understanding requires more detailed knowledge on the genetic underpinning of mating system variation on the one hand, and the genomic consequences on the other hand.

Mating system shifts can occur rapidly, especially in cases where outcrossing by self-incompatibility (SI) is mandated by only few genes (Durand et al. 2020). The breakdown of SI may then involve a loss-of-function mutation in a gene of the self-incompatibility (S)-locus characterized by low recombination. In Brassicaceae, the S-locus includes a receptor kinase (SRK) expressed in the papillae of stigmas and a cysteine-rich (SCR) protein expressed in pollen development (Abhinandan et al. 2022). If SCR on the pollen coat successfully binds with its haplotypic SRK partner on the papillae membrane, an incompatibility cascade is launched. The expression of SCR is sporophytic (SSI), as compared to gametophytic (GSI) in other plant taxa. An alternative for the breakdown of SI may involve modifiers of S-gene expression, with S-allele-specific phenotypic effect (Mable et al. 2017; Li et al. 2023). Further possibilities include the loss of function of genes acting downstream from the initial recognition expressed in the pistil, changes in the expression or turnover of the S-allele products, and modifications of the pistil environment (Good-Avila et al. 2008).

The genomic basis of the breakdown of SI is often difficult to detect, because self-compatibility with selfing increases genetic drift (Pollak 1987). Drift is a stochastic process and can lead to the random fixation of alleles across the genome (Whitlock and Bürger 2004). Furthermore, drift and low levels of effective recombination increase linkage and the hitchhiking of variants (Hartfield and Glémin 2014). Together, they make the detection of causal mutations hard. Unsurprisingly, the majority of reportings involve a nonfunctional S-locus in several extant selfing species (e.g. Capsella rubella (Guo et al. 2009), C. orientalis (Bachmann et al. 2019), A. thaliana (Boggs et al. 2009), A. kamchatica (Tsuchimatsu et al. 2012), and Leavenworthia alabamica (Busch et al. 2011)). Given that genomic differentiation is expected to increase with time due to drift and selection (Ravinet et al. 2017), studies on populations or species that have only recently undergone mating system shifts could provide deeper insight into variation in the breakdown. An example is Arabidopsis lyrata subsp. lyrata (in the following abbreviated as A. lyrata) with a distribution in mostly temperate eastern North America. Regions with predominantly self-compatible (SC), selfing populations adjacent to predominantly SI, outcrossing populations exist. In a recent study, Li et al. (2023) showed in a crossing experiment involving both types of populations from central parts of the Great Lakes that selfing was likely due to the existence of modifiers specific to certain S-alleles, but unlinked to the S-locus itself. Furthermore, the modifiers seem to act dominantly. Genomic analyses have so far been unsuccessful in pinpointing the causal components (Mable et al. 2017).

A shift to selfing may have various implications on the genetics of populations and species. Theory predicts a reduction in the standing genetic variation (Glémin and Ronfort 2012), increased genome-wide linkage disequilibrium (LD; Hartfield and Glémin 2014), and the increase in the frequency of slightly deleterious alleles by genetic drift (Wright 1931). The predicted long-term evolutionary consequences are more conflicting. On the one hand, selfing may constrain the adaptive potential of populations or species due to the reduced levels of standing genetic variation and the lower efficacy of selection (Glémin and Ronfort 2012; Hartfield and Glémin 2016; Hodgins and Yeaman 2019). On the other hand, selfing taxa with large population sizes may purge deleterious alleles more effectively (Glémin 2003) and show accelerated adaptive evolutionary change involving recessive mutations if their beneficial effects are considerable (Hartfield and Glémin 2016). Positive consequences may result in founder populations of predominantly selfing individuals having higher persistence in a new environment than their outcrossing counterparts (Sachdeva 2019). Coupled with the demographic advantage of reproductive assurance during further spread, this may explain why selfing taxa generally have larger niche breadths and larger ranges as compared to related outcrossing taxa (Grant and Kalisz 2020).

Empirical studies on the population genetic consequences of selfing in comparison to outcrossing have indeed documented lower genetic diversity, higher homozygosity, and higher linkage (e.g. Mable and Adam 2007; Huang et al. 2020; Teterina et al. 2023). Though in their details, results often seem to differ from theoretical expectations (e.g. Willi and Määttänen 2010). An important factor may be that selfing is associated with other demographic factors such as isolation after successful colonization or bottlenecks following the spread of a selfing taxon (Wright et al. 2013). It is therefore important to include such demographic aspects in the study of the genetic and evolutionary implications of mating system variation. In A. lyrata, most shifts to selfing indeed occurred during or since postglacial range expansion. A population relatedness tree suggested that after expanding out of the Driftless Area of Wisconsin, two shifts to mixed-mating/selfing occurred during or after northward expansion to Lake Superior, and at least three shifts during the subsequent southeastward expansion to Lake Erie and Lake Huron (Mable et al. 2017; Willi et al. 2018). From the second refuge area in eastern Pennsylvania, colonization happened in a star-like manner, with at least two shifts, one in coastal Virginia and one in the Appalachians of North Carolina and Tennessee. Finally, the eighth region is in an old rear-edge part of the species’ distribution, in the Ozarks of Missouri. Former studies showed that the shift to selfing in A. lyrata was associated with lowered genomic variation and increased mutation load (Willi et al. 2018), a reduced signature of positive selection (Willi et al. 2020), and increased LD that extended over longer genomic distances (Lucek and Willi 2021). Similarly, it was shown in Arabis alpina that postglacial mating system shift resulted in a strong reduction in genetic diversity and an increase in mutation load (Laenen et al. 2018), though interactions between mating system shift and expansion have not been investigated in either of the systems.

In this study, we first attempted to evaluate the likely role of the S-locus in the breakdown of SI in the different parts of the range of A. lyrata. As across the regions no association with specific S-alleles was detected, a diallel crossing experiment was performed on a—phylogenetically speaking—relatively ancestral mixed-mating population, to estimate the contribution of the maternal and paternal genotype to mating system variation. As results indicated a biparentally inherited basis, we performed a genome-wide association study (GWAS) on replicate populations varying in mating system in search of shared causal genomic regions. Finally, we estimated the population genetic implications of the shifts to SC and selfing.

Methods

Sample collection and sequencing

We selected nine A. lyrata populations that are known to be predominantly selfing or have a mixed-mating system (Griffin and Willi 2014) as well as nine nearby and closely related, predominantly outcrossing populations with a shared history of either rear-edge isolation (MO, USA) or some expansion (Willi et al. 2018; Supplementary Table 1, Fig. 1B). Samples were collected during the reproductive seasons in 2007, 2011, and 2014 (Supplementary Table 1). We had two sequence datasets available. The first consisted of Illumina HiSeq 2000 paired-end reads of a length of 150 bps and ~10× coverage of two individuals of the 18 populations (Willi et al. 2022; European Nucleotide Archive: PRJEB30473 and PRJEB23202). The second consisted of Illumina HiSeq 2000 paired-end reads of a length of 100 bps and ~120× coverage of population pools of (equimolar DNA of) 25 individuals of each of the 18 populations (Fracassetti et al. 2015; Willi et al. 2018; European Nucleotide Archive: PRJEB8335).

Fig. 1.

Diversity at the female specificity determinant S-receptor kinase (SRK). A) Depiction of SRK with the respective subdomains. B) Location of the studied populations. The color of the triangles represents the mating system, approximated by the population inbreeding coefficient F_IS, and those of population abbreviations the mating system in categories: outcrossing in black, mixed-mating in pink, and selfing in red. Population abbreviations reflect the respective US state or Canadian province followed by a number. C) Marginal effect (slope; with 95% confidence interval) of a generalized linear mixed-effects model with the predicted probability for being heterozygous for SRK in relationship to F_IS. D) Phylogenetic relationship among populations based on two resequenced genomes per population (left) compared with the phylogenetic relationship of SRK haplotypes. The SRK tree was collapsed to haplotype groups. Lines connect the SRK allele(s) of a respective individual with the detected haplotype group. Gray dots on the phylogenies indicate nodes with >95% bootstrap support. The color of the triangles and the lines connecting the two phylogenies represent F_IS. Indicated in the SRK tree are the number of unique haplotypes found for each group, the dominance class, and the closest matching haplotype from Mable et al. (2017), the latter being italized.

De-novo assembly of the S-locus gene SRK

Because the S-allele diversity and the divergence among alleles is commonly high within and across populations (Goubet et al. 2012), we used a pipeline developed by Genete et al. (2020) and performed a local de-novo assembly using the individual-level re-sequence data. We focused on a ~1,000 bps long exon section of the SRK sequence that encodes the extracellular part of the protein with its three domains (Fig. 1A) and is directly involved in the SI response (Nasrallah and Nasrallah 2014); SCR was not considered as reference sequences are generally limited. We downloaded all available SRK sequences for A. lyrata including their paralogs from GenBank on 18 November 2020 and used these to perform a kmer filtering, extracting similar sequences as SRK with a script from Genete et al. (2020). Then we performed de-novo assemblies for the extracted reads of each individual with SPAdes 3.13.0 (Bankevich et al. 2012) using the following parameters: -k21, 41, 81, and -careful. To identify the SRK genes, we then aligned all obtained contigs against the SRK references used for the kmer filtering with YASS 1.14 (Noe and Kucherov 2005) and only retained contigs with more than 100 bps that aligned, also removing assembled paralogs. Finally, we identified the contigs that had all three protein domains fully assembled by annotating all contigs with Interproscan 5.47-82.0 (Jones et al. 2014). The retained sequences were aligned with MAFFT (Katoh 2002), and we established the phylogenetic relationship with RAxML 8.2.11 (Stamatakis 2014) under the GTRCAT model with 1,000 bootstrap replicates.

We contrasted the SRK allele tree with a population-phylogenetic tree using the individual-level sequence data. We aligned reads against the A. lyrata reference genome and used BCFtools 1.10.2 (Danecek and McCarthy 2017) to call variants. With VCFtools 0.1.16, (Danecek et al. 2011), we subsequently filtered out: indels, rare variants with minor allele frequency (MAF) <0.05, >10% missing data, and spatial overrepresentation by considering only single nucleotide polymorphisms (SNPs) that were at least 100 bps apart. This resulted in 293,098 SNPs, which we used to establish a phylogenetic relationship in RAxML with 1,000 bootstrap replicates. For the latter, we included an ascertainment bias correction under the GTRGAMMA model to account for the fact that we only included polymorphic SNP positions. The resulting tree was compared with the one based on SRK haplotypes. To identify the phylogenetic group to which an SRK haplotype might belong, we used BLAST+ (Camacho et al. 2009) and compared each against the S-locus sequences from Goubet et al. (2012) and Mable et al. (2017).

We investigated the impact of mating system shifts from outcrossing to selfing on heterozygosity at the S-locus. For this, we employed a generalized linear mixed-effects model with a binomial distribution with the package lme4 (Bates et al. 2015) in R 3.5.1 (R Core Team 2021). The response variable was whether SRK was heterozygous (1) or not (0). Mating system was the fixed effect and population the random effect. Mating system was approximated by the population inbreeding coefficient, F_IS, calculated from microsatellite data mostly based on the same plant material as was used for pool-seq (Willi and Määttänen 2010; Griffin and Willi 2014). An F_IS of about 0.45 was associated with a multilocus outcrossing rate t_m of 0.2, which is typically the cutoff for calling a taxon selfing.

Crossing experiment

To gain more insight into the inheritance of mating system variation in A. lyrata, we performed a diallel crossing experiment on a mixed-mating population of the Keweenawa peninsula, MI, USA collected in 2007 (MI5, outbreeding coefficient of 0.574 estimated by progeny array; Willi and Määttänen 2010). The diallel allows the variance partitioning into sex-of-parent in-/dependent inheritance. The diallel was established over three generations under climatized greenhouse conditions. In the first generation, we raised one seedling of each of 30 field-collected mother plants, which we then crossed in hap-hazardly arranged pairs. In the second generation, one seedling of each of the 15 crosses was randomly assigned to one of three blocks of five. Plants of a block were crossed with each other reciprocally in a diallel without selfings. Sixty crosses were planned in total, but fewer were produced because one plant did not flower and others did not produce enough flowers. In these two rounds of crossing, we performed hand-pollinations on emasculated flower buds to avoid cross-contamination. In the third generation, 4 (rarely 5 or 6) offspring plants per cross, split over 4 spatial blocks, were raised and assessed for their ability to self-fertilize.

Once plants started flowering, they were self-pollinated on six flowers if possible. We estimated SC for each of these plants (N = 143) as the fraction of carpels that developed into a fruit, distinguishing: fully developed siliques scored as 1, not or hardly developed siliques as 0, and partial fruit development as 0.5 (sometimes containing 1 seed). As data were strongly bimodal, we transformed it to binary, with values of 0 or, when the fraction of developed carpels was equal or larger than 0.15, 1. The cutoff of 0.15 was chosen such that borderline cases with partial fruit development only once were still considered SI. The fraction of SI genotypes was 0.56 and was close to the observed outcrossing rate in that population in nature.

To partition variance in mating system (SI/SC) in the diallel-cross experiment, we employed the model of Cockerham and Weir (1977; Fry 2004). Random effects were block (of parental plants) and sporophyte-by-pollen donor interaction within the block. The covariance structure of the R matrix was general linear with five parameters, allowing the partitioning of variance into 1) genotype independent of male or female sex, 2) interaction between genotypes independent of sex, 3) maternal genotype, 4) paternal genotype, and 5) interactions between genotypes depending on whether one acted as a male or a female. Analysis was done with proc glimmix in SAS 9.4 (SAS Institute, Cary NC) and the testing of variance components with likelihood ratio tests.

GWAS and outlier analysis

To identify genomic regions that might be associated with a shift in the mating system, either directly or indirectly, for example, through the selfing syndrome, we performed a GWAS on the pool-seq allele frequency data. We first trimmed the raw sequences separately for each population and independent Illumina run using PoPoolation 1.2.2 (Kofler et al. 2011), setting a minimal base quality threshold of 20 and keeping only reads ≥84 bp. We mapped all retained reads against the A. lyrata genome v1.0 (Hu et al. 2011) with BWA-MEM 0.7.15 (Li 2013). We masked the centromeric regions as well as two regions on scaffold 2 (position ranges: 8,746,475–8,835,273 and 9,128,838–9,212,301), which share very high similarity with the A. thaliana chloroplast genome. We retained reads that mapped to scaffolds I–VIII, representing the eight chromosomes of A. lyrata. We used Picard 2.1.1 (http://broadinstitute.github.io/picard) to remove duplicate reads and SAMtools 1.10 (Li et al. 2009) to retain only properly paired reads with mapping quality over 20. Using SAMtools, we generated a mpileup file comprising all populations and called SNPs with VarScan 2.4.1 (Koboldt et al. 2012), requiring per population a minimal read depth of 50 at a given position and a minimal variant allele frequency of 0.03. We removed previously identified repeat sites in the A. lyrata genome (Fracassetti et al. 2015) with BEDtools 2.27.1 (Quinlan and Hall 2010). With VCFtools and on the level of the population, we removed indels and kept only biallelic SNP positions that had a depth of 50–500 and a minimal genotype quality of 28. Across populations, a MAF of 0.03 and ≤25% missing data were required. Lastly, SNPs with a strand bias of more than 90% in a population were filtered out. This procedure resulted in a dataset comprising 1,935,698 polymorphic SNPs.

GWAS on mating system variation was performed with BayPass 2.1 (Gautier 2015). Mating system was approximated by F_IS. BayPass extends the approach by Coop et al. (2010) and Günther and Coop (2013) by estimating and accounting for the hierarchical structure of populations using the (scaled) covariance matrix of population allele frequencies (Gautier 2015). We used the auxiliary covariate model with default parameters and 5,000 burn-in iterations in the MCMC chain, followed by 25,000 iterations. To reduce artifacts due to potential variability between runs, we performed 10 independent BayPass runs. We then calculated the average Bayes Factor (BF), expressed in deciban units (dB), for each SNP as a quantification of the degree of relationship between mating system and the standardized allele frequency. To identify outlier SNPs, we employed a Hidden Markov model (HMM) following Lucek et al. (2019). We modeled two states (low vs. high) on BF with the R package HiddenMarkov (Harte 2021). To identify the biological processes of the genes with outlier SNPs, we performed a gene ontology (GO) enrichment analysis based on 10,000 randomization steps in R using SNP2GO (Szkiba et al. 2014), implementing a False Discovery Rate (FDR) correction. We used the A. lyrata annotation of Rawat et al. (2015) and restricted our analysis to biological processes. To test if outlier SNPs were clustered on one chromosome, we performed a Grubbs’ test on the frequencies of outliers for each chromosome.

Genome-wide genetic diversity, homozygosity, and recombination

We tested for the effect of the mating system and range expansion on genetic variation and homozygosity. For each population, expansion distance was the distance (in km) of a population along the map-projected population tree back to the presumable glacial refuge area, or the great circle distance for rear-edge populations (Willi et al. 2018). A first dependent variable was Watterson’s θ for intergenic regions of individual-level sequence data. It was estimated with mlRho 2.9 (Haubold et al. 2010), following the protocol of Lucek and Willi (2021). A Bayesian mixed-effects model analysis was employed with the package blme (Chung et al. 2013). Fixed effects were F_IS, log₁₀-transformed expansion distance, and their interaction; the random effect was population.

We next identified runs of homozygosity (ROHs) across each individual genome with PLINK 1.90 (Purcell et al. 2007). We first filtered the VCF file, setting across populations a minimal depth of 6, an MAF of 0.05, and allowing up to 10% missing data. In PLINK, we set a minimum of 50 SNPs for a scanned window, allowing both up to 3 heterozygous calls and 3 missing sites. The minimal length of an ROH to be considered was 250 kb and a homozygous density of 1,000. We fitted Bayesian linear mixed-effects models similar to the one specified above, but with the additional fixed effect of average genome-wide sequence depth. For the length of ROHs, we used the same model but with chromosome as an additional random effect.

We estimated genome-wide (effective) recombination using the software ReLERNN (Adrion et al. 2020). While classic software tools require many individually sequenced specimens per population to estimate recombination, ReLERNN implements a machine-learning algorithm that can also handle pool-seq data. We ran ReLERNN following the suggested best practice for pool-seq data (Adrion et al. 2020), with few specific adjustments. For the initial coalescent simulation step, ReLERNN implements msprime (Kelleher et al. 2016) and simulated 100,000 training datasets and 1,000 validation and test datasets based on the observed SNP diversity in a population. Additional input was an empirical estimate of the per-base mutation rate from A. thaliana (Ossowski et al. 2010). In a next step, ReLERNN uses the simulated dataset to train a recurrent neural network, for which we used the standard settings as suggested by (Adrion et al. 2020), applying also a population-level MAF of 0.03. With the trained network, the original data is then assessed including a bootstrap approach to estimate the 95% confidence interval. Importantly, ReLERNN only estimates recombination within self-set bins, whose sizes are optimized based on the SNP density of the initial datasets, separately for each chromosome. Bins therefore differed in lengths (mean: 230,604 bps; 95% CI: 71,933–747,020; Supplementary Table 2). We thus applied a sliding window approach and calculated the average recombination rate across 100 kb windows for each population across the genome. Finally, the recombination rate for each 100 kb window across all chromosomes and populations was analyzed with a Bayesian linear mixed-effects model. The six fixed effects were: phylogenetic cluster, F_IS, log₁₀-transformed expansion distance, the presence or absence of GWAS outlier SNPs, the average log₁₀-transformed distance of each 100 kb window from the respective centromere as well as the interaction between F_IS and expansion distance. While the former three were predictors on the level of the population, the latter two were predictors on the level of the genomic window. Phylogenetic cluster assignment of populations into west or east (east: NC and VA in Fig. 1B) was based on previous analyses (Willi et al. 2018). Random effects included population crossed with scaffold. Significance was estimated with type III Wald χ² testing.

Results

Diversity at the S-locus

Individual-level sequence data allowed the assembly of 52 unique SRK haplotypes across the 72 haploid genomes. These fell into 9 haplotype groups in the allele tree (Fig. 1D). The allele tree contained three of the four phylogenetic clusters that were formerly described for A. lyrata (Goubet et al. 2012). Most common in our dataset was (grouped) haplotype Al01 (S1), which together with Ah3 (S3) belongs to recessive haplotype classes. Four haplotypes were less common, the dominance class IV haplotypes Al39 (S39) and Ah29 (S13), S23, and S19, while we found three novel haplotypes. Contrasting the population tree with the SRK allele tree revealed that selfing and mixed-mating populations had often different SRK alleles (Fig. 1D). Selfing populations had alleles of the groups Al01, Al39 and S19, as well as of two of the three novel haplotypes. All of these haplotype groups except S19 were also represented in outcrossing populations. Shifts to selfing tended to be associated with reduced heterozygosity at SRK (Fig. 1C; χ²₁ = 3.12, P = 0.077). In detail, 44.4% of the 18 individuals from outcrossing populations were heterozygous, when for selfing (with mixed-mating) populations only 21.4% of the 18 individuals were heterozygous. Importantly, results indicate that selfing evolved in several SRK backgrounds that are shared with outcrossing populations.

Crossing experiment

By experimentally assessing the inheritance of mating system variation in a mixed-mating population based on diallel crosses, we revealed significant nuclear interaction effects (Supplementary Table 3). The model attributed 73% of the variance to such dominance effects (Vnn, P < 0.05), the rest—though not significant—to interactions between genotypes depending on whether one acted as a male or a female (7%). Residual variance contributed with 20% to the total variance. Results suggest that one or several genes with a strong dominance hierarchy greatly determine SC.

GWAS and outlier analysis

To identify genomic regions associated with a shift to selfing, we performed a GWAS on population-based data. The mating system was approximated by F_IS. In total, 4,048 Hidden-Markov outliers across all populations were detected (0.21% of all investigated SNPs). Outliers were enriched on chromosome 7 (Grubbs’ test: G = 2.44, P < 0.001), which comprised 49.8% of all outlier SNPs (Fig. 2A). Of these, 989 outliers occurred within a 210 kb region around the S-locus (Fig. 2B).

Fig. 2.

Outcome of the genome-wide association study (GWAS) for population-based F_IS. A) Number of HMM outlier SNPs for each chromosome. B) Manhattan plot depicting the Bayes Factor for each SNP included in the GWAS within a 210 kb region around the S-locus on chromosome 7. HMM outliers are depicted in blue.

A total of 215 genes contained outlier SNPs and were associated with 341 unique GO terms, of which 38 were significantly enriched (Supplementary Table 4). For chromosome 7, outlier SNPs occurred in genes associated with various functions, including seed and embryo development and responses to stress factors such as temperature or bacteria. Outlier SNPs of other chromosomes occurred in genes that were similarly linked, among other functions, to defence, embryo development, or flower development, suggesting that mating system shifts are associated with genes with similar phenotypic effects. However, none of the 215 genes with outlier SNPs had been listed as relevant in the cascade of SI or downstream events (Abhinandan et al. 2022).

Genome-wide genetic diversity, homozygosity, and recombination

Shifts to selfing were significantly associated with an overall reduction in genetic diversity (Fig. 3A; χ²₁ = 39.40, P < 0.0001; for the entire model: marginal R² = 0.685, conditional R² = 0.972), reducing θ by 0.0026 on average. Here, the interaction between the mating system and range expansion was not significant (χ²₁ = 0.16, P = 0.6905). Consistently, the length of ROHs (P = 0.0027; entire model: marginal R² = 0.069, conditional R² = 0.104) and their number (P = 0.0012; entire model: marginal R² = 0.569, conditional R² = 0.812) increased with selfing (Fig. 3B and C; Supplementary Table 5). For the length of ROHs, the interaction with range expansion was positive and significant (P = 0.0318), when for the number of ROHs a trend occurred (P = 0.0510), suggesting that selfing populations with a longer expansion history had higher degrees of inbreeding overall, resulting in longer and somewhat more ROHs (Supplementary Fig. 1). For most selfing populations, we found that ROHs on chromosome 7 overlapped at least with one window of reduced heterozygosity reported for a pool of SC as compared to SI F2 offspring of a cross between a selfing and an outcrossing population (Mable et al. 2017). Exceptions were the selfing population ON 11 and the mixed-mating population VA2, which together with most outcrossing populations had no ROHs in these windows (Supplementary Fig. 2).

Fig. 3.

The relationship between mating system shift and genetic diversity. A) Boxplot showing the difference in average genome-wide genetic diversity (Watterson’s θ) between outcrossing and selfing/mixed-mating populations. Marginal effects (with 95% confidence intervals) of Bayesian linear mixed-effects models with B) the number of runs of homozygosity (ROHs) or C) the length of ROHs as the dependent variable.

We detected differences in (effective) recombination rates in 100 kb windows among the populations (Fig. 4A). With the overall statistical model, we found rates of recombination to differ between populations from the two phylogenetic clusters (entire model: marginal R² = 0.113, conditional R² = 0.268; Fig. 4B; Supplementary Table 6). While recombination rates were slightly higher in windows that contained GWAS outlier loci (Fig. 4C), they were significantly lower in populations with increased levels of inbreeding/selfing (Fig. 4D), but not with longer range expansion. Furthermore, the interaction between levels of inbreeding and range expansion was not significant. Across the genome, recombination increased with increasing distance from the centromeres (Fig. 4E).

Fig. 4.

Mating system shifts and recombination across the genome of A. lyrata. A) Density plot depicting the distribution of recombination, estimated as crossovers per base and generation within 100 kb windows from pool-seq data using the software ReLERNN. Line colors depict population-level F_IS (see Fig. 1B). Marginal effects (with 95% confidence intervals) of a linear mixed-effects model with recombination as the dependent variable and the following fixed factors: B) phylogenetic cluster, C) the presence or absence of GWAS outlier SNPs, D) mating system (F_IS), and E) the distance of each window from the centromere (bps, log scale).

Discussion

Breakdown of SI and the S-locus

The breakdown of SI in Brassicaceae is often attributed to loss-of-function mutations in at least one of two genes that constitute the initial SI response and are part of the S-locus (Tsuchimatsu et al. 2010; Goubet et al. 2012; Abhinandan et al. 2022; Kolesnikova et al. 2023). Therefore, in order to understand the genomic basis of SC in Arabidopsis lyrata, we started with the study of the S-locus. Following Genete et al. (2020), we assembled the extracellular coding regions of the female specificity determinant SRK (Fig. 1). Several S-locus haplotypes in Arabidopsis have been described so far (Schierup et al. 2001; Prigoda et al. 2005; Mable and Adam 2007; Goubet et al. 2012; Mable et al. 2017). In our study, we found S-haplotypes S1, S3, S13, S19, S23, S39, and three additional new haplotypes, closely related to S1, S3, and S13, respectively (Fig. 1D).

S-alleles differ in dominance, corresponding to different phylogenetic groups, and particularly recessive haplotypes occur at higher frequencies (Castric et al. 2010; Goubet et al. 2012). Consistently, we found the recessive haplotype Al01 (S1) to be the most common. It occurred in selfing and outcrossing populations and was about equally often homozygous and heterozygous in the two types of populations. Second most common was the new haplotype that is closely related to S3, which belongs to the same dominance class as S1 (Schierup et al. 2001). Also, this new haplotype was found to occur in both selfing and outcrossing populations. Third most common was the new haplotype related to recessive S1. It was homozygous in the selfing north-shore Lake Superior populations MI6 and ON11 and occurred as heterozygote in outcrossing VA1. Overall, data revealed only a trend of lower SRK heterozygosity in selfing populations (Fig. 1C).

Another important insight is that individuals of selfing populations had SRK alleles of five of the nine detected haplotype groups, which speaks against selfing having evolved in a particular S-haplotype background. In line, a recent study on North American A. lyrata suggested that a modifier causes SC in some populations (Li et al. 2023). Li et al. (2023) argued that the modifier may specifically inactivate S1, similar to the mechanism that regulates the dominance hierarchy among different S-alleles by silencing each other via sRNA (Durand et al. 2020). Li et al. (2023) suggested that the modifier was maintained at a low frequency in outcrossing populations, but had probably been driven to fixation in the selfing populations. In accordance, we found homozygotes for the S1 allele in populations with high and low levels of inbreeding (e.g. ON4 and PA4, respectively; Fig. 1D). In our study, the modifier hypothesis could explain the breakdown of SI in two populations, MO2 and ON4, as individuals of these selfing populations were homozygous for S1 (ON4 is RON in Li et al. 2023).

Further evidence for the unlikely global role of SRK or SCR in the shift to selfing was that the expected heterozygosity for SRK only marginally decreased with mating system shifts, but the sample size was limited (Fig. 1C). The diallel-cross experiment on one mixed-mating population, MI 5, revealed that variation in mating system was inherited biparentally (Supplementary Table 3). The diallel highlighted that only a small and insignificant amount of variance was due to interactions between genotypes depending on whether one acted as a male or a female in a cross. Instead, variation in mating system (SI/SC) was—to a very high extent—explained by nuclear genetic components with dominance. However, to which degree this pattern upholds across the species’ distribution requires further investigation.

Finally, we performed an association study across all populations to pinpoint common genes and genomic regions associated with a shift to selfing if there was convergence in this evolutionary transition. We found only a very small fraction of loci to be outliers despite our dense genomic sampling, many of which clustered near the S-locus (Fig. 2). These outliers overlapped with a region previously found to be homozygous in a pool of SC offspring but heterozygous in SI offspring; F2 offspring were from a cross between a selfing and an outcrossing A. lyrata population (Mable et al. 2017). Consistently, we found that ROHs of several individuals particularly of selfing populations overlapped with this region (Supplementary Fig. 2). While Mable et al. (2017) detected patterns in the cross with selfing ON4, the overlap in results with our study that had a wider geographic scope may be taken as a hint that the modifier hypothesis could apply to SC in A. lyrata more broadly.

Though, associations could have other sources. Association studies are often prone to identifying outliers that are only indirectly associated with the actual positions under selection due to increased linkage in such regions (Bush and Moore 2012). A shift to selfing is coupled with increased linkage in A. lyrata as a consequence of demographic processes and drift (Lucek and Willi 2021), but can also result in increased clustering of outliers in putative regions under selection (Lucek et al. 2019). The observed clustering of outliers on chromosome 7 and especially close to the S-locus is likely to be at least in part attributed to linkage to the S-locus, potentially as a result of balancing selection (Roux et al. 2013). Focusing on three populations of North American A. lyrata, a recent study found that linkage to the S-locus distorted the phylogenetic signal within a 10–15 kb region up and downstream of the S-locus (Le Veve et al. 2024). Using more populations and a denser genomic sampling, we identified a more extended region upstream but not downstream of the S-locus, which could indicate that the downstream region is less conserved and maybe less linked across our populations.

Associations may also be the by-product of the evolution of SC and selfing. Several genes in the upstream region containing outlier loci were associated with stress response, defence, or embryo and flower development (Supplementary Table 4), the latter being aspects of the selfing syndrome in A. lyrata (Carleial et al. 2017). The selfing syndrome captures some common phenotypic convergences in selfing taxa, typically involving traits linked to flower size or pollen production (Eckert et al. 2009; Shimizu and Tsuchimatsu 2015). In A. lyrata, it was shown that pollen production is significantly lower in selfing as compared to outcrossing populations (Willi 2013).

The population genetic implications of selfing

Increased selfing is predicted to come with population genetic consequences. For instance, reduced genetic exchange among individuals of the same population is expected to increase homozygosity and to decrease genetic diversity and effective recombination (Pollak 1987; Glémin and Ronfort 2012; Hartfield and Glémin 2016). These effects might become stronger with range expansion, as selfing populations at the range edge might undergo more severe bottlenecks and experience stronger drift (Excoffier et al. 2009). Here, we observed a significant decrease in standing genetic variation among selfing populations (Fig. 3A), which was, however, independent of how far the populations had expanded. The number and length of ROHs were increased in selfing populations (Fig. 3B, C). Especially the length was increased even more in selfing populations far from the origin of expansions, suggesting that selfing with range expansion comes with demographic settings that lead to the additional buildup of homozygosity (Supplementary Fig. 1). Extended ROHs are indicators of recent inbreeding, whereas short ROHs are indicators of inbreeding in the more distant past (Keller et al. 2011; Curik et al. 2014). Our finding of longer ROHs in selfing populations in general is thus consistent with the recent, that is, the postglacial mating system shifts in our studied populations (Griffin and Willi 2014).

Increased ROHs have been shown to be positively correlated with LD and negatively with recombination (Herrero-Medrano et al. 2013). In line with theoretical predictions (Nordborg 2000; Hartfield and Glémin 2016), we detected a decrease in (effective) recombination rate in selfing populations, though the decrease was independent of expansion distance (Fig. 4D). This is consistent with the overall reduced genetic diversity at the range edges, where levels of LD reach high levels in selfing and outcrossing populations (Lucek and Willi 2021). Reduced recombination may result in increased LD and genetic hitchhiking (Charlesworth and Wright 2001). As a consequence, background selection and selective sweeps can then reduce genomic variation across a larger part of the genome, and the overall effectiveness of selection is constrained by the Hill-Robertson effect (Lucek and Willi 2021). Similar reductions in recombination have recently been demonstrated between the holocentric selfing nematode Caenorhabditis elegans and the outcrossing C. remanei, where recombination was generally reduced in the central parts of the chromosomes, but species-specific regions of increased recombination occurred, likely as a result of different selection regimes (Teterina et al. 2023).

Although we estimated recombination with a deep-learning approach that had been shown to produce reliable estimates for pool-seq data, demographic processes such as population size changes could affect these estimates (Adrion et al. 2020). A considerable fraction of variation in recombination could be explained by factors other than mating system. Genomic regions with common outlier SNPs had higher recombination rates (Fig. 4C). A historical contingency came with the two distinct genetic clusters with different glacial refugia; recombination was significantly lower in populations of the western as compared to the eastern cluster (Fig. 4B). This is consistent with the previous finding that standing genetic variation is higher in the eastern cluster (Walden et al. 2020). Finally, the rate of recombination also varied across the genome, with recombination increasing with increasing distance from the centromeres (Fig. 4E). This is a common observation among species and has been attributed to different chromatin structures (Stapley et al. 2017).

Conclusions

Investigating repeated mating system shifts within North American A. lyrata, we show that the genomic patterns generally corroborate theoretical predictions and former studies in A. lyrata involving fewer populations and/or genetic markers (Mable et al. 2017; Li et al. 2023; Le Veve et al. 2024). The shift to selfing is not linked to an S- or SRK allele, suggesting that mating system shift is likely caused by a modifier outside of the S-locus. A potential candidate region is just upstream of the S-locus, which overlaps with large ROHs detected in selfing populations and a previously identified homozygous region in SC offspring of an F2 cross between SC and SI plants (Mable et al. 2017). With a different cross, we showed that SC is inherited biparentally and nonadditively, at least in the one studied population. Across populations, mating system shifts result in reduced genetic variation, increased runs of homozygosity, and lower effective recombination. Selfing with long expansion had led to particularly long ROHs. Many of the population genetic changes have been suggested to also contribute to species’ range limits, for instance in cases where reduced recombination rates between adaptive loci would slow down range expansions because of poor purging of locally deleterious alleles at the expansion front (Eriksson and Rafajlović 2021). Selfing populations of A. lyrata perform indeed relatively poorly in experiments (Willi 2013) and may therefore have contributed to the range limits in this species.

Supplementary Material

Supplementary material is available at Journal of Heredity online.

Funding

This work was supported by the Swiss National Science Foundation (31003A_166322 and 310030_184763 to YW). KL was further supported by the Swiss National Science Foundation (PCEFP3_202869).

Conflict of interest statement. None declared.

Author contributions

Kay Lucek (Conceptualization, Data curation, Formal analysis, Investigation, Writing - original draft), Jana M. Flury (Formal analysis, Methodology, Writing - review & editing), and Yvonne Willi (Conceptualization, Formal analysis, Funding acquisition, Investigation, Writing - review & editing)

Data availability

Genomic data is available from NCBI BioProjects: PRJEB8335, PRJEB30473. Scripts and output data are deposited on Zenodo (10.5281/zenodo.13893011).