Hybridization is a recurrent evolutionary stimulus in wild yeast speciation

Abstract

Hybridization can result in reproductively isolated and phenotypically distinct lineages that evolve as independent hybrid species. How frequently hybridization leads to speciation remains largely unknown. Here we examine the potential recurrence of hybrid speciation in the wild yeast Saccharomyces paradoxus in North America, which comprises two endemic lineages SpB and SpC, and an incipient hybrid species, SpC*. Using whole-genome sequences from more than 300 strains, we uncover the hybrid origin of another group, SpD, that emerged from hybridization between SpC* and one of its parental species, the widespread SpB. We show that SpD has the potential to evolve as a novel hybrid species, because it displays phenotypic novelties that include an intermediate transcriptome profile, and partial reproductive isolation with its most abundant sympatric parental species, SpB. Our findings show that repetitive cycles of divergence and hybridization quickly generate diversity and reproductive isolation, providing the raw material for speciation by hybridization.

Hybridization can contribute to diversity from the genomic to the species level. Here, Eberlein, Hénault et al. investigate genomic, transcriptomic and phenotypic variation among wild lineages of the yeast Saccharomyces paradoxus and suggest that an incipient species has formed by recurrent hybridization.

Introduction

Hybridization among species is considered as a stimulating force in evolution¹, but its consequences are difficult to predict. On one hand, it can lead to the collapse of species barriers², and on the other, to the formation of new species³. Hybrids can benefit from extreme phenotypes that allow them to exploit novel ecological niches⁴ or simply new combinations of traits that offer a strong advantage over the parents⁵. The frequency at which species arise from hybridization and under which circumstances this process takes place remain to be fully understood^6,7.

Hybridization among eukaryotic microbes has been shown to be a powerful mean to create genomic and phenotypic diversity^8,9, however its contribution to the evolution of natural populations still has to be fully investigated. Studies of fungi in close association with humans suggest that hybridization could be a key driver of microbial diversity. For instance, animal and plant fungal pathogens¹⁰ rely on hybridization to colonize new hosts and acquire virulence traits. In addition, budding yeasts of the genus Saccharomyces have drawn diversity from inter-specific crosses, with tens of domesticated strains that adapted to industrial environments at least partly, thanks to their hybrid origin¹¹. Population genomics studies have also shown that inter-species introgression could contribute to yeast genomic diversity in nature¹², although these studies involve at least one partially domesticated species in most instances. Documenting the evolutionary and ecological conditions in which hybridization events take place and their consequences in completely natural systems is imperative because most of the microbial diversity present today has evolved in the absence of human influence.

A recent population genomics study of Saccharomyces paradoxus, a budding yeast found worldwide on the bark of deciduous trees and their associated soils^13,14, showed that a novel North American species evolved through hybridization about 10,000 years ago¹⁵. This hybrid species (SpC*) originated from the secondary contact between the two most abundant species, SpB that occupies a large fraction of the continent, and SpC¹⁵, which is found almost exclusively so far in the north east. SpC* shows a unique profile of growth phenotypes¹⁵, occurs mostly in the zone of sympatry between its two parental species and shows reproductive isolation with both of them, which is caused at least partially by genome rearrangements. These findings revealed that hybridization occurred between two incipient species (SpB and SpC) that originated a little more than 100,000 years ago and that it led to the formation of SpC*. A recent study by Xia et al.¹⁶ identified a novel group of strains, SpD (originally defined as “Clade d” and then mistakenly assigned to the SpC* group), which exhibit signatures of genomic admixture, potentially involving the same parental species as SpC*. Analyses by Hénault et al.¹⁷ suggested that SpD could have arisen from a second hybridization between SpB and SpC, indicating that hybridization could have occurred multiple times in different locations^16,17.

With additional sampling, genome sequencing and systematic phenotyping, we find that SpD strains are recent hybrids between the hybrid species SpC* and the most abundant North American lineage, SpB. This backcross between a hybrid and one of its parental species generated novel growth phenotypes, intermediate transcriptional profiles, a novel genome architecture and partial reproductive isolation. These results highlight that speciation in yeast may result from repeated cycles of divergence and hybridization.

Results

A new hybrid lineage in North America

We examined whether the origin of SpD can be traced back to a hybridization event between the previously described endemic S. paradoxus species SpB and SpC. We assessed the population structure and genetic relationship from fully sequenced genomes of 316 S. paradoxus strains, which included 38 newly sampled strains (2016), 34 strains previously sampled, 91 genomes from Xia et al.¹⁶, and 153 genomes from Leducq et al.¹⁵ (Fig. 1, Supplementary Figure 1, Supplementary Data 1-2). In agreement with previous phylogenetic analyses^16,17, we found that the strains form five clusters corresponding to SpA, SpB, SpC, SpC*, and SpD (Fig. 1b, c, Supplementary Figure 2). The two main lineages SpB and SpC exhibit nucleotide divergence of 2.2% on average (Supplementary Figure 3). SpB shows high nucleotide diversity (0.42%), which likely results from its large geographic distribution and sub-population structure (Supplementary Figure 4). SpA exhibits little diversity (0.1%), which is consistent with its recent introduction from Europe¹⁵. The SpD group lies in between the two main lineages SpB and SpC, next to the hybrid species SpC* and exhibits the highest level of within-group nucleotide diversity (0.6%) (Supplementary Figure 3).

Fig. 1.

Population structure of Saccharomyces paradoxus in North America. a Sampling locations (circles) of 316 whole-genome sequenced strains from five distinct groups: SpA, SpB, SpC, SpC*, and SpD^15–17. The map was drawn with R v3.3.3⁸² using the package maps (version 3.3.0). Detailed strain information can be found in Supplementary Data 1. b Genetic relationship among isolates from a maximum likelihood phylogenetic tree built from 25,280 variable positions. SpD splits into the two sub-clades SpD1 and SpD2 as previously suggested by Hénault et al.¹⁷. c Grouping of strains by principal component analysis showing the relationship among lineages. PC1 to PC3 separate SpA, SpB, SpC, SpC*, and SpD as independent genetic clusters (For color code, see Fig. 1a). The spread of strains (median and the 20th to 80th percentile range) in each PC (x-axis) is shown. d The two models (M01, M02) from the f4 statistics explain admixed ancestry of SpC* and SpD. Dotted arrows indicate admixture events. e Ranking of the 15 models based on their fit to the data. Filled squares on the top indicate if the models assume admixed ancestry of SpD and SpC*. The five best models indicate hybrid origin of SpD. The two best models matching all f4 statistics (filled dots) suggest a hybrid origin for SpD resulting from a cross between SpB and the hybrid SpC* (model M01), or hybrid origin of SpD resulting from a cross between SpB and SpC, occurring before the origin of a SpC* hybrid (model M02)

We used population genetics analysis (f4 statistics) to test different hypotheses about admixture. Detecting admixture between lineages would explain their evolutionary relationships and, if so, decipher the SpD ancestry relative to SpB, SpC, and SpC*. The genomic composition of SpD is best supported by two models involving admixture: either SpD arose from admixture between SpB and the hybrid species SpC* (M01) or between SpB and SpC before the origin of SpC* (M02; Fig. 1d, e, Supplementary Figures 5-9). These results confirm that SpD is a second hybrid lineage related to SpC*, but they do not allow us to disentangle which of SpC* or SpD emerged first.

Hybridization between a hybrid species and its parent

The geographical distributions of North American endemic S. paradoxus lineages support the first of the two admixture models (M01), suggesting that SpD originated from hybridization between SpB and SpC*. All of the SpD strains were isolated from a small area near Toronto, Canada¹⁶, which is part of the wide distribution of SpB. In contrast, neither SpC nor SpC* had been isolated from this area prior to the current study, in spite of intensive sampling efforts¹⁶. Our most recent sampling and genotyping of S. paradoxus isolates revealed a single SpC* strain isolated approximately 2 km away from SpD sampling site (Fig. 1a, Supplementary Data 1). This suggests that although rare, SpC* is present in this SpB-dominated area and makes natural hybridization scenarios between SpB and SpC* possible.

The recent descent of SpD from the hybridization of SpB with the hybrid species SpC* (M01) is supported by two largely independent data sets: (i) structural genomic variation and (ii) discrete introgressed regions in the hybrid genomes. We examined structural genomic variation by comparing six new de novo genome assemblies produced from Oxford Nanopore long reads (Supplementary Table 1) and one published assembly (YPS138, SpB) produced from PacBio long reads¹⁸. We confirmed that chromosomal inversions and translocations segregate among SpA, SpB, SpC, and SpC*^15,19 (Fig. 2b). We also found that some of these rearrangements segregate within SpD. The SpD lineage was previously suggested to comprise two subclades¹⁷ and its within-lineage nucleotide diversity was the highest of all lineages, demonstrating the role of hybridization in generating genomic diversity (Supplementary Figure 3). We thus separated the SpD strains according to this classification, into two sub-groups named SpD1 and SpD2. In the two most parsimonious tree topologies based on chromosomal inversions and translocations (Fig. 2a, b; Supplementary Figure 10), SpD1 and SpD2 formed a monophyletic group with SpB and SpC*, while SpC was more divergent. Inversion i3 on chromosome V and translocation t1 between chromosomes VI and XIII (corresponding to VItXIII described in Leducq et al.¹⁵) are shared by SpC* and SpD, while inversion i5 on chromosome VI is shared by SpB, SpC*, and SpD (Fig. 2c). In addition, inversion i1 on chromosome X is common to SpB and SpD2. However, some rearrangements (e.g. i2) conflict with these most parsimonious topologies. These conflicts may originate from hybridization, although incomplete lineage sorting or independent occurrence due to inversion or translocation hotspots could explain them as well. Nevertheless, the most parsimonious scenarios clearly show that both SpD1 and SpD2 share more structural genomic similarity with SpC* than SpC, supporting its SpB-SpC* origin.

Fig. 2.

Genome rearrangements support that SpD results from the backcross of the hybrid species SpC* with its parent SpB. a The average pairwise distance (in number of rearrangements) between genomes for each of the 945 tree topologies tested. Topologies are ranked according to a maximum parsimony criterion, the best topologies involving the least rearrangements. The five best topologies (blue shaded area and inset) are highlighted. b Two bifurcating tree topologies (T300 and T284) best describe the evolutionary relationships among North American S. paradoxus according to the maximum parsimony criterion from a. Both show that SpD is more closely related to SpC* and SpB than to SpC. Edges are labelled with the corresponding rearrangements, “iX” denoting inversions and “tX” denoting translocations. Red dotted lines connect conflicting rearrangements, i.e. rearrangements that occur on two or more branches. c Pairwise synteny across seven North American S. paradoxus genomes using SpA as a reference. Insets highlight examples of genomic rearrangements. Inversion i3 on chromosome V and translocation t1 (VItXIII) support the SpC*-SpD close relationship; inversion i5 on chromosome VI supports the SpB-SpC*-SpD relationship; inversion i1 on chromosome X supports the SpB-SpD relationship; inversion i4 on chromosome X is specific to SpC; and inversion i2 on chromosome X is an example of conflicting rearrangement, being shared by SpA, one SpB strain and one SpD strain. Red segments represent inversions. Alignments longer than 1 kb are shown

Having the strongest support for the model under which SpD originated from a cross between SpB and SpC*, we mapped the regions of genomic admixture of SpD strains to reference genomes from SpB and SpC*, which enabled to examine the patterns and size of parental genomic regions in the SpD strains (Fig. 3a). SpD strains have about 50% SpB and 50% SpC* of genomic origin (Figure 3Aii). The patterns of SpB- and SpC*-like regions allowed us to distinguish between the two SpD sub-groups, as seen from the genome-wide phylogenies¹⁷. Meiotic recombination and asymmetrical parental backcrossing affect the pattern of parental genomic blocks in hybrid genomes. Through meiotic recombination, uniparental backcrossing enlarges the size of blocks that share ancestry with the backcrossing parent²⁰. In the hybrid species SpC*, our analysis detects ~3% introgression from SpB (minor parent). The minor proportion of SpB elements shows that crossing-overs from meiotic recombination and backcrosses with SpC have reduced the genomic blocks of SpB ancestry to a few regions.

Fig. 3.

Genome-wide pattern of introgression in the young and old hybrids. a Shared ancestry in 17 SpC* and 13 SpD strains. Genomic positions are colored according to parental genotypes. (i) 51 genes contain partially or completely introgressed genes from SpB (white bar) and are fixed in SpC*. (ii) Fragments of SpB or SpC* ancestry in 13 SpD strains. 12 SpD strains show large ancestral fragments shared with SpB or SpC* (ring 1–2; orientation: from inside to outside). Strain WX21 (ring 13) has many heterozygous regions across the genome, where the separation between SpB and SpC*-like ancestral was not possible (blank elements) (Supplementary Figure 12). Rendered by white bars (chrV, VI, and XII) are nine genes that SpC* inherited from SpB and that are shared between all SpD strains with SpC*. b Length distribution of introgressions from SpB shows shorter blocks on average in SpC* than in SpD, supporting a more recent origin of the latter (median of 1.1 kb in SpC* and 12.2 kb in SpD respectively; Mann-Whitney U-test, P-value < 2.2e−16). The genome-wide proportion of regions with SpB-ancestry in SpC* is 3.1% on average, while it is 44.7% on average for SpD. c Relative time of origin of SpD and SpC* (T₁) estimated from the divergence of SpB-like regions in each admixed lineage (SpD or SpC*) with corresponding regions in SpB. T₂—divergence time of SpB-like regions in admixed lineages with corresponding regions in SpC. Times were estimated for one strain of SpC* and all 13 strains of SpD from the same sampling location. Diamonds depict mean estimates of divergence time across all contiguous SpB-like regions with more than 1000 informative sites. In all SpD strains, SpB-like regions are significantly younger than in SpC* (mean T₁, Mann-Whitney test, P-value < 2.3E−05), but divergence time of SpB-like regions with SpC remains the same in two admixed lineages (mean T₂, Mann-Whitney U-test, P-value > 0.23). d Dating the hybridization events. The lineage SpD emerged only recently (~2500 years ago) from the crosses between SpB and SpC*. Arrows indicate initial backcrosses between SpC* and SpC

These introgressed regions could play an important role in the ecological divergence of SpC*, as fixed SpB introgressions are marginally enriched for genes involved in the response to amino acids¹⁷ (Supplementary Figure 11, Supplementary Table 2–3). We compared the size of SpB parental introgression blocks between SpC* and SpD, as longer blocks in SpD could confirm its most recent origin. The length of SpB-like blocks in SpD is significantly greater than in SpC* (median of 12.2 kb and 1.1 kb respectively; Mann-Whitney U-test, P-value < 2.2e−16, Fig. 3b). Since the fragments of both SpB and SpC* origin represent about equal proportions in SpD genomes, it seems unlikely that SpD experienced extensive backcrossing. Indeed, the block size distributions of SpB and SpC* elements in SpD genomes were not significantly different (median of 12.2 kb and 10.4 kb, respectively; Mann-Whitney U-test, P-value 0.99). We thus hypothesize that SpD strains are early generation hybrids from crosses between SpB and SpC*. The early-generation hybrid hypothesis is further supported by the number of apparent crossing-overs per chromosome (average 10.2), which is in the reported range of 2–10 crossing-overs per meiosis for S. cerevisiae²¹. However, recombination rate was recently shown to be about 40% lower in S. paradoxus compared to S. cerevisiae, which would push the origin of these strains a little further back in time²². These observations support a recent hybrid origin for the SpD strains, which have likely undergone only few rounds of meiosis.

In the SpD strain WX21, many regions could not be assigned to either SpB or SpC* ancestry because they appeared to be heterozygous, with haplotypes from both parental lineages (last outer ring, Fig. 3a). This is the only noticeable case of heterozygosity across all the North American strains reported so far. This suggests that although rare, outcrossing does occur in these lineages. Low levels of heterozygosity in other SpD strains are not in contradiction with the early hybridization hypothesis, since Saccharomyces haploids can readily autodiploidize through mating-type switching. Thus, spores from first-generation SpD hybrids could have become homozygous after a single meiosis. The breakpoints between heterozygous haplotypes indicate that WX21 likely originated from a cross between two strains of the SpD2 sub-group (Supplementary Figure 12). This result shows that admixture is still ongoing in this group of strains and generates genomic diversity.

Recent origin of the hybrid SpD group

We estimated the divergence time of both hybridization events (SpC*, SpD) by comparing the genetic divergence of SpB-like regions that are SpD-specific and SpC*-specific relative to SpB. We found that the mean divergence of SpD from SpB is significantly lower than the mean divergence between SpC* and SpB (Fig. 3c; T1, Mann-Whitney U-test, P-value < 2.3e−05). In addition, we used a Bayesian approach²³ to estimate the timing of the two hybridization events (Fig. 3d). Using the concatenated sequences of nine genes that were admixed from SpB into SpC* and from SpC* into SpD (Supplementary Tables 4-5), we estimated the divergence of SpC* from SpB to be 18,500 years ago (±5,000). This is in the same range as the dating initially performed on different sets of loci¹⁵. The divergence of SpD from SpB was estimated to be of 2600 years (±750) (Supplementary Table 6). Further, we estimated that diversification within the SpD lineage could have occurred very recently, about 1800±750 years ago, with most of the strains not being older than a few hundred years. The young age of the SpD lineage estimated by these methods is consistent with its very limited geographic distribution (Fig. 1a).

Impact of hybridization on molecular and growth phenotypes

The persistence of a hybrid lineage and its development into a separate species depend on several factors, including hybrid-specific phenotypes and reproductive isolation from parental species^4,24. The hybrid species SpC* is phenotypically distinct and partially reproductively isolated from its two parental species SpB and SpC¹⁵. The very recent hybridization of SpC* with SpB to form the SpD group allowed to investigate two important aspects of inter-species hybridization: the emergence of new phenotypes and reproductive isolation. By assuming that the current state of SpD is a good approximation for the early stage that follows hybridization, comparing SpD and SpC* allows to examine the phenotypic evolution of a typical S. paradoxus hybrid lineage over two time frames separated by about 10,000 years of evolution.

To characterize the phenotypic profile of the SpD group relative to other lineages, we measured the growth of 229 strains, including 12 SpD strains, in 24 growth conditions that probe a diversity of potential metabolic performances. These conditions comprise various carbon sources, nitrogen sources, chemical compounds or incubation temperatures (Supplementary Data 3). We computed two measures of colony growth based on highly resolved growth curves: integrated growth through time (Area Under the Curve, AUC) and maximal growth rate (Maximum Slope, MS; Supplementary Figure 13). Hierarchical clustering of the strains for these two traits across conditions indicated that most SpD strains are phenotypically similar to each other and distinct from the other lineages (Fig. 4a). Linear discriminant analysis further suggested a linear combination of conditions that distinguishes SpD from the other lineages (Supplementary Figures 14–16).

Fig. 4.

Phenotypic divergence and reproductive isolation of SpD. a Hierarchical clustering of normalized colony growth profiles in 24 conditions based on AUC or MS. The heatmaps correspond to z-score normalization of the rows (conditions) excluding outliers. CSM: complete synthetic medium, Me-α-DGP: methyl-α-D-glucopyranoside, MSM: minimal synthetic medium. b Value normalization used to compare hybrids with their parental lineages across conditions. The difference d of a given strain’s value (white star) to the mid-parent value (green star) is divided by the half-difference a between the parental distributions medians. c Distributions of normalized AUC and MS for the three hybrid lineages SpC*, SpD1, and SpD2 and their parental lineages. P-values of one-sided Mann-Whitney U-tests are detailed in the tables. d PCA on genome-wide gene expression levels. The two first dimensions cumulatively explain 27.8% of the total variance. The 20 genes most contributing to these dimensions are shown. SpC and SpC* are grouping together, while SpD strains show an intermediate transcription profile between SpC/SpC* and SpB. e Reproductive isolation measured from spore viability tests of intra- and inter-lineage crosses from four lineages (SpD is separated into SpD1 and SpD2). High spore viability was detected for all intra-lineage crosses. Inter-lineage crosses generally suffer from low viability. However, inter-lineage crosses between recently emerged or close related lineages (SpC*-SpD) indicate higher fertility than crosses between older lineages (SpB-SpC). The number of crosses in each case is shown in brackets. f Conceptual framework adapted from Mallet³⁷ to illustrate how hybrids may explore genotypic and phenotypic spaces and fitness peaks. In this framework, SpD may be a group of individuals that need to exploit further genotypic and phenotypic space to gain higher fitness and to potentially occupy novel ecological niches. The color gradient corresponds to variation in fitness with darker colors representing higher fitness. This shows several potential unoccupied and potential niches (dark gray) with fitness peaks that could lead, if reached, to the persistence for SpD

Since outcrossing is rare in yeast²⁵, the long-term persistence of a new lineage likely depends on how it initially performs in the local environment as opposed to how reproductively isolated it is. We therefore compared the performance of the hybrid lineages SpC* and SpD (respectively SpD1 and SpD2) with their respective parental lineages across the 24 conditions (Fig. 4b). Despite their genetic similarity to SpC, SpC* strains showed an overall stronger growth than SpC strains and weaker than SpB strains (Fig. 4c), as previously observed by Leducq et al.¹⁵. SpC* outperformed both parental lineages when grown on mannose for the MS trait but not in any other condition. This suggests that SpC* could show hybrid superiority in specific conditions, which may have contributed to its long-term success. The SpD1 clade performed significantly worse than SpB and SpC*, while SpD2 was worse than SpB but not significantly different from SpC* (Fig. 4c). Neither SpD1 nor SpD2 significantly outperformed both of their parental lineages in any condition tested (Supplementary Figure 17). SpD exhibited phenotypic novelty but has an apparent lack of overall advantage compared to its parental lineages. This suggests that SpD might be at a disadvantage across a wide range of growth conditions and thus may reflect the extrinsic reproductive isolation between its parental lineages. Since SpD is mostly allopatric to SpC* and SpC, its lack of performance relative to the sympatric SpB may be the most consequential. We cannot exclude that the SpD group could be more competitive in conditions that we did not test as it is difficult to determine what the fitness determinants of each lineage are in their respective environments. Overall, these results clearly show that SpD hybrid strains have unique growth profiles and that hybridization has large consequences on complex phenotypes.

To further examine the role of hybridization in generating new phenotypes in the two hybrid lineages SpC* and SpD, we performed gene expression profiling on a subset of 24 strains in two replicates each (Supplementary Tables 7-8, Supplementary Data 4, Supplementary Figures 18-19). The genome-wide gene expression profile of SpD showed an intermediate phenotype in the first principal component, consistent with its almost 50% SpB and 50% SpC* genome composition (Fig. 4d, Supplementary Figures 20-21). Gene Ontology (GO) enrichment analysis from pairwise differential expression (Supplementary Figure 22) revealed that the SpD lineage is particularly distinguished from both SpB and SpC* in terms of functions related to amino acid biosynthetic/catabolic processes, peptidyl-serine dephosphorylation (SpD1 to SpB and SpC*) and ribosome biogenesis (SpD2 to SpB and SpC*) (Supplementary Figure 22; Supplementary Data 5). SpC* clustered strongly with its closest parent SpC in the two first components, as expected from their shared ancestry of ~97%. It appears that transcription profiles therefore largely reflect the genomic composition of the admixed strains, more than the phenotypic profiles, potentially because most transcriptional changes result from neutral evolution and thus reflect genetic divergence²⁶ (Fig. 1). However, these transcriptional changes could also play a causal role in the poor performance of hybrid strains in specific conditions as those are associated with core metabolic (amino acids) and growth (ribosome biogenesis) processes.

SpD shows partial reproductive isolation with other lineages

The persistence of SpD as a genetically distinct group requires that it is reproductively isolated from its parental species. Liti et al.²⁷ observed a positive correlation between nucleotide divergence and reproductive isolation in Saccharomyces sensu stricto yeasts, showing that reproductive isolation accumulates with time. This is also the case in our study system^15,19. However, crosses between the parental lineage SpB or SpC and the hybrid species SpC* resulted in similar degrees of spore survival (38 and 49% respectively) even though SpC* has higher sequence identity with SpC (see also: Leducq et al.¹⁵,Charron et al.¹⁹). Chromosomal rearrangements and genetic incompatibilities can accelerate the onset of reproductive isolation between lineages²⁸. The isolation of SpC and SpC* was indeed previously shown to result at least partly from chromosomal rearrangements, explaining the deviation from the general trend observed within the genus²⁷. SpD could also benefit from such rearrangements that cause partial isolation from its parents.

We thus sought to measure the degree of reproductive isolation of SpD and observed high fertility among SpD2 strains (mean = 94%; n = 3; Fig. 4e). However, SpD1s showed a decreased fertility when crossed with each other (mean = 65%, n = 3). The same degree in fertility was also observed after the direct sporulation of wild SpD1 and SpD2 homothallic isolates (Supplementary Data 6). Since SpD1 also exhibited weaker overall growth in the phenotypic screen, these strains may bear an excess of deleterious alleles or allele combinations, which could lower both spore viability and colony growth measured in various environmental conditions.

We found that SpD1 and SpD2 show relatively high fertility when crossed with the young hybrid species SpC* (Fig. 4e and Supplementary Data 6). Fertility dropped when these SpD strains were crossed with more diverged lineages, such as SpC. Surprisingly, backcrosses of SpD with the parental lineage SpB also show very low spore survival (SpD2, mean = 28% (4–47%), n = 6; SpD1, mean = 38% (24–48%), n = 6), similar to what we observe in crosses between the older lineages SpB and SpC. This partial reproductive isolation between SpB and SpD could enable the persistence of both lineages in sympatry on the long term.

One notable exception are crosses between SpD strains and strains of a rare group (~1%) of SpB strains, called SpBf, which harbor an important translocation between chromosomes VI and XIII (VItXIII). These crosses showed a spore survival (Supplementary Data 6) similar to what is observed for crosses with SpC*. Previous data showed that SpBf strains are the closest SpB relatives to SpC*, because they share the VItXIII translocation (t1, Fig. 2) and this translocation was shown to be correlated with spore inviability in crosses between SpC* and SpC¹⁵. Our results, however, show that the higher fertility of SpD-SpBf crosses may not be due to the presence of the VItXIII translocation¹⁵. Indeed, we detected the VItXIII translocation in SpD2, SpC* and SpBf strains but not in SpD1 (Supplementary Figure 23). Therefore, the presence of the translocation likely does not explain SpBf’s higher fertility with SpD1 than with SpD2. Other genomic rearrangements, detected or not detected in the structural analysis (Fig. 2c), could play important roles.

Overall, SpD strains show highly reduced fertility with the two most abundant lineages in North America, SpB and SpC. This applies particularly to crosses with SpB strains, which occur in the region where SpD is found. Despite its young origin, SpD appears to have achieved levels of reproductive isolation with SpB similar to those between SpB and SpC^29,30, providing an opportunity to examine how hybridization may lead to rapid isolation. The sufficient reduction in fertility could therefore enable the persistence of the SpD lineage and further its independent and divergent evolution. The relatively high fertility with its other parental species SpC* may compromise this independent evolution. However, since SpC* is extremely rare outside of the SpB-SpC sympatric region in the St-Lawrence valley and given the low frequency of sexual reproduction in S. paradoxus, SpD-SpC* mating remains highly unlikely. Geographical barriers could therefore play the major role in this case.

Discussion

In this study, we present a case of recurring hybridization in a eukaryotic microorganism. We show that a new group of diverse hybrid strains has evolved from a backcross between a young hybrid species and its parental lineage. Our findings highlight that recurring hybridization events in nature contribute to genomic, phenotypic and potentially species diversity.

The evolutionary conditions that make recurrent hybridization and the emergence of new lineages possible may be very restrictive and are not completely known. However, Blanckaert and Bank³⁰ recently showed that reproductive isolation could arise very rapidly through incompatibilities, where specific combinations of parental haplotypes in the F₂ generation cause reproductive isolation with the parental lineages. This could be accelerated in yeast through local inbreeding or selfing allowed by mating type switching within a few cell divisions²⁵, rapidly making incompatibility loci homozygous and thus eliminating the costly phase of segregation of incompatibilities. The nature of these incompatibilities remains to be identified and might involve mechanisms including elevated  mutation rate³¹, mitochondrial-nuclear interactions^32,33 and chromosomal rearrangements¹⁵.

Rapid hybrid speciation within a few generations based on hybrid phenotypes was recently shown to occur by Lamichhaney et al.³⁴. Our findings further support that this could be a fast process and reveal that repeated hybridization could occur within a time frame of a few generations of sexual reproductions. The SpD lineage displays novel growth and transcriptome phenotypes, along with partial reproductive isolation with other lineages. These conditions could be sufficient for the emergence of another hybrid species. One challenging question is whether and how the young SpD group will persist through time given its apparent lack of growth advantage compared to the parental species. This observation contrasts with many studies showing that F₁ yeast hybrids usually show hybrid vigor across a broad range of conditions^35,36. The overall poor performance of the young SpD hybrids relative to the expected hybrid vigor and compared to the older SpC*, could represent a transient state, in which gene combinations are being sorted by natural selection. Following a scenario proposed by Mallet³⁷, SpD could represent an early hybridization event with individuals that are distant from a phenotypic optimum, a stage of homoploid hybrid speciation that could be inevitable. These hopeful monsters could take advantage of new gene combinations in future generations that may eventually make reaching an optimum possible (Fig. 4f). Long term persistence would therefore depend on the sorting of genetic diversity currently found in this group of strains. Also, the spatial density of Saccharomyces strains has been shown to be very low³⁸ such that growth competition may be limited. Other factors than growth rate, such as cell survival to specific stresses or the exploitation of specific resources, may be greater determinants of long-term maintenance. It remains challenging to study the ecological and genomic parameters that contribute to microbial evolution in a natural setting. A key element for future investigations will therefore be the study of fitness components in natural conditions, for instance through the implementation of common garden-experiments³⁹ and the ability to follow the evolution of genotypes through time¹⁶. The use of the model system S. paradoxus may be one of the few systems to offer this opportunity.

Methods

Strain collection

The 74 novel S. paradoxus strains sequenced in this study were sampled between 2014 and 2016 and collected from different substrates and locations (Supplementary Data 1-2). Samples were isolated and enriched for Saccharomyces strains with the method described in Sniegowski et al.¹⁴ and confirmed to belong to S. paradoxus by sequencing a ~750 bp genomic region including the loci ITS1 and ITS2 (Oligo forward: GTTTCCGTAGGTGAACCTGC; Oligo reverse: ATATGCTTAAGTTCAGCGGGT)⁴⁰, following the protocol described in Leducq et al.⁴¹.

Whole-genome sequencing by Illumina

Genomic DNA was extracted from independently grown colonies on solid rich media (YPD, Yeast extract, Peptone, Dextrose) with DNeasy Blood & Tissue Kit (Qiagen, Hiden, Germany). Libraries were prepared with the Illumina Nextera XT kit (Illumina, San Diego, USA) following the manufacture’s protocol. Libraries and input DNA were quantified with AccuClear Ultra High Sensitivity dsDNA Quantitation Kit (Biotium, Fremont, USA) using the Fusion Optics (SPARK, TECAN, Männedorf, Switzerland). Pooled libraries were sequenced on a single lane of HiSeqX (150PE, Illumina, San Diego, USA) at the Genome Quebec Innovation center (Montréal, Canada). The 74 strains were sequenced with an average genome-wide coverage of ×45(Supplementary Figure 1, Supplementary Data 2). Raw sequences are accessible at NCBI (bio project ID PRJNA479851).

Read mapping and variant calling

Genome sequences from the 244 previously sequenced S. paradoxus strains from North America were downloaded and added to our newly sequencing strains^15,16 (NCBI accession: PRJNA277692, PRJNA324830; Supplementary Data 1). Independently sequenced libraries of the same strains from the Leducq et al.¹⁵ study were combined. Raw reads were mapped on the reference genome of CBS432¹⁸ using Bowtie2 v2.1.0⁴² with default settings and the local alignment option. Strains 16_053B and 16_187A were removed because 16_053B showed poor sequencing results and 16_187A revealed low alignment rates (coverage < 1) due to a contamination. Sequences were sorted and indexed with Samtools v1.8⁴³.

Variants were flagged using default variant quality filters implemented in the Platypus variant caller v0.8.1⁴⁴. The program Vcftools v0.1.15⁴⁵ was used to remove variants that showed either of these features: (1) low quality per depth ratio (QD filter); (2) in regions with low root-mean-square mapping quality (MQ filter); (3) had quality below 20 Phred scores (Q20 filter); (4) failed the strand bias filter (strandBias filter); (5) fell in the low-complexity regions (SC filter); (6) were supported by reads with bad quality bases or were present on one strand only (badReads filter); (7) fell in regions supported by too many haplotypes (HapScore filter) or (8) showed lower than expected frequencies (alleleBias filter). Additionally, genotypes with Phred scores below 20 were masked. INDEL variants were removed using a custom python2.7 script. Further, we used vcflib v1.0.0 with the vcfallelicprimitives function to split Multiple Nucleotide Polymorphisms (MNPs) into Single Nucleotide Polymorphisms (SNPs), and Bcftools v1.6⁴³ with the norm function to remove duplicated positions. Finally, only biallelic variants were retained with two alleles across all strains. The overall alignment rate to the European SpA reference genome CBS432¹⁸ was 82.4 %, with an estimated average coverage of 32.6 over the 316 genomes (Supplementary Figure 1).

Global phylogeny and principal component analysis

A subset of 25,280 variants regularly spaced across the genome was selected from a total of 205,206 variants across the 316 S. paradoxus strains. A maximum likelihood phylogenetic tree was built with RAxML v8.2.9⁴⁶. The tree was estimated with 1000 rapid bootstraps, under generalized time-reversible (GTR) model with gamma distribution of substitution rates. The program Adegenet v2.1.0⁴⁷ was used to perform a Principal Component Analysis (PCA) in which we calculated the median values of each principal component for each lineage and the 20th and 80th percentile to identify the PCs that split closely related lineages (e.g. SpC and SpC*; Supplementary Figure 2).

Inference of population structure

The population genetic structure in the SpB lineage was estimated using the Bayesian clustering method implemented in Structure v2.3.4⁴⁸ (Supplementary Figure 4). The filtered set of 25,280 SNPs was used and only sites that were variable in the SpB lineage were selected. Additionally, adjacent SNPs in strong linkage disequilibrium (LD) were removed using Plink v1.90b4.4⁴⁹. Parameters—indep-pairwise 50 5 0.5 were used to remove one SNP from a pair in a 50 SNP window, with a step of 5 SNPs, if the estimate of LD (r2) for the pair was higher than 0.5. The list of the remaining SNPs in vcf format was converted into Structure format using PGDSpider v2.1.1.0⁵⁰. Structure analysis was performed for several number of clusters, respectively ranging from k = 2–8, with 10 replicates per each cluster. Each replicate analysis was run for 30,000 steps with the first 10,000 steps excluded as a burn-in. Delta K statistic was calculated (Supplementary Figure 4B) to identify the most likely number of clusters with Harvester v.0.6.94⁵¹. Multiple replicate analyses of each cluster were then aligned using CLUMPP v1.1.2⁵².

Test of admixture scenarios

In the PCA (Supplementary Figure 2), strains from the lineage SpC* group closely with SpC, whereas SpD is intermediate between SpB and SpC/SpC* for the first two PCs. To confirm that SpD is a hybrid lineage and to identify signatures of shared ancestry with other lineages, we tested 15 different models using genome-wide polymorphism data. The models assume admixed ancestry of SpD and/or admixed ancestry of SpC* or no admixture events (Supplementary Figure 5). Models were compared with the data by means of f4 statistics calculated in the qpDstat program from AdmixTools v1.0.1 package⁵³. Their fit to the data was evaluated using admixture graph fitting implemented in admixturegraph R package v1.0.2⁵⁴. In admixturegraph, the models are represented as admixture graphs that generate expectations of f4 statistics based on graph edge lengths and admixture proportions, and the parameters of the graphs are optimized using a cost function that minimizes differences between expected and observed f4 statistics. The fitted f4 statistics for each of 15 models are shown in Supplementary Figure 6. Prior to f4 calculations, only variants present in at least 4 strains in every lineage (SpA, SpB, SpC, SpC*, and SpD) were selected. Models were ranked according to the minimal error, which is an estimate of a cost function for the given model for the best fitting values of f4 (Fig. 1e). When running value optimization 100 times, model M01 yielded fitted values within 3 standard errors of observed statistics in 100% of runs, whereas model M02 yielded similar values only in 58% of runs. Other models never yielded a perfect fit with the data. To make sure that the two recognized groups of SpD (SpD1 and SpD2) follow the same admixture history, we separately tested all 15 models using only SpD strains from the SpD1 and SpD2 group. We found the same models supporting admixture history of these groups (Supplementary Figures 7-9). Model M01 yielded fitted values within 3 standard errors of observed statistics in 100% of runs for SpD1 and SpD2, whereas model M02 yielded similar values only in 86 and 65% of runs for SpD1, and SpD2, respectively. Other models never yielded a perfect fit with the data.

Long-read sequencing

Six strains from the North American lineages were sequenced with the Oxford Nanopore technology using two FLO-MIN107 flow cells (R9.5). The subset of strains passed on each run is shown in Supplementary Table 1. The 1D Native barcoding genomic DNA protocol with kits SQK-LSK108 and EXP-NBD103 was used. Genomic DNA was extracted with two successive phenol-chloroform extractions followed by ethanol precipitation. One microgram (run 1) or 1.5 µg (run 2) of genomic DNA per strain was fragmented using Covaris g-TUBEs at 6000 rpm (run 1) or 4200 rpm (run 2) in an Eppendorf 5424 (Hamburg, Germany) centrifuge. Fragmentation was assessed by agarose gel electrophoresis. No FFPE DNA repair step was performed. For run 1, all DNA recovered after adapter ligation was pooled and loaded on the flow cell due to low yield, while for run 2 the amounts of DNA were adjusted to achieve approximately equal representation of the six strains in the two combined libraries. Sequencing was performed using the default script provided with the MinKNOW software v1.7.14.

De-novo assembly of long-read sequencing data

The data from the two independent sequencing runs were basecalled, demultiplexed, and trimmed separately, after which they were combined. Reads that passed the MinKNOW quality filter (pass directory) in fast5 format were basecalled and demultiplexed using Albacore v2.0.2 with the following parameters: −f FLO-MIN107 −k SQK-LSK108 −o fastq −−barcoding. Adapter trimming was performed using Porechop v0.2.2 (github.com/rrwick/Porechop) with parameter −b set for demultiplexing. Only the reads assigned to the expected barcode by both Albacore and Porechop were kept. De novo genome assembly was performed using SMARTdenovo with parameters −J 1000 −c 1.

The following polishing steps were performed on the resulting draft assemblies. First, the raw assemblies were polished using Nanopolish v0.9.0⁵⁵ with the trimmed Nanopore reads. The basecalled sequences of trimmed nanopore reads in fasta format were mapped on the raw assemblies using BWA MEM v0.7.16a-r1181⁵⁶ with parameter −x ont2d. The Nanopolish variants program was run with parameters −p 2 −min-candidate-frequency 0.1 to yield polished consensus assembly sequences. Second, Illumina reads for the same strains (PE 100 bp or 150 bp) were used to perform additional correction on the Nanopolish-processed assemblies using Pilon v1.22⁵⁷. Adapter and barcode sequences for the Illumina TruSeq sequencing kit were retrieved and combined in a custom reference library (TruSeq_custom_retrieved.fa). Then Illumina reads were trimmed using Trimmomatic v0.36⁵⁸ with parameters PE −phred33 ILLUMINACLIP:TruSeq_custom_retrieved.fa:2:30:10 TRAILING:3 SLIDINGWINDOW: 4:15 MINLEN:36. The trimmed reads were mapped on the Nanopolish-processed assemblies using BWA MEM with default parameters. Optical and PCR duplicates were marked using Picard MarkDuplicates (broadinstitute.github.io/picard) with default parameters and the alignment files were filtered using Samtools view v1.5⁴³ with parameter -F  set to 0 × 704, after which Pilon was run with the –diploid parameter. The corrected contigs were aligned with the PacBio assembly of a North American S. paradoxus strain (YPS138) produced by Yue et al.¹⁸ using Mauve v2.4.0⁵⁹ with the progressiveMauve algorithm and default parameters. Contigs were manually assigned to chromosomes and were re-ordered using a custom Python v3.6.3 script to yield the final assemblies. Assembly sizes and N50 are presented in Supplementary Table 1.

Evolutionary analysis of genomic rearrangements

Genomic rearrangements were identified by performing a multiple genome alignment with rearrangements using Mauve. The progressiveMauve algorithm was used with default parameters to align the six SMARTdenovo assemblies along with the PacBio assembly of one SpB North American strain (YPS138; Kuehne et al.⁶⁰) produced by Yue et al.¹⁸. Unless otherwise stated, all further analyses were performed using custom Python v3.6.3 scripts. Genomic blocks of local collinearity were extracted from the .backbone file produced by Mauve and filtered to exclude blocks smaller than 500 bp, or which were not conserved among the seven genomes. Block permutations were exported in the GRIMM⁶¹ format, manually excluding the contig boundaries that corresponded to chromosome splits as those were unlikely to correspond to real splits. The comprehensive set of rooted bifurcating tree topologies with six labelled leaves was generated using the NetworkX package⁶² in Python and exported to Newick format. MGRA v2.2.1⁶³ was used to reconstruct ancestral genomic block permutations for each tree topology independently, allowing to map the number of rearrangements that occurred on each branch. The tree topologies implying the smallest total number of rearrangements were selected as the ones most likely describing the evolutionary relationships among the genomes.

Detection of introgressed regions and genes

Genome-wide fasta-alignments of the 316 S. paradoxus strains were prepared to identify introgressed regions. Samtools v1.8 and Bcftools v1.8⁴³ were used to calculate the genotype likelihood from the bam-formatted alignment files, to call variants and to create single fasta files for each individual strain. Here, the S. paradoxus SpA strain CBS432¹⁸ was used as a reference genome. A custom Python script was used to combine the single fasta files into chromosome-separated alignment files. Nucleotide diversity within and divergence between lineages were calculated from genome-wide fasta files (model:JC69) in Pangorn v.2.3.1⁶⁴ using all 316 individuals.

The detection of introgressed fragments of SpB-origin in the SpC* strains was based on two approaches. First, for each gene, we calculated the genetic distance (model: JC69) of each SpC* strain to each SpB and SpC strain. We then determined if each SpC* strain was genetically closer to either SpB or SpC at that locus. We marked a gene as being fixed and introgressed from SpB in the SpC* lineage, when all SpC* strains had a smaller genetic distance to all SpB strains than to any SpC strain. As input data, we used the 17 SpC* strains (as identified by PCA, the global phylogeny and in Leducq et al.¹⁵) and compared them to the 213 SpB strains and 50 SpC strains from our current data set of 316 strains.

The program HybridCheck v1.0⁶⁵ was used to perform a Triplet-Test, where the strains UWOPS-79140 and LL2012_027 represented the lineages SpB and SpC to precisely delimit regions of shared ancestry in the 17 SpC* strains. The initial window size was set to 1 bp and was further extended in 1 bp steps along the chromosomes to identify introgressed blocks. Fixed introgressed regions were defined by identifying sequences that showed shared ancestry with SpB in all SpC* strains (Fig. 3a). The detection of introgression in the 13 SpD strains was done as described above, with the SpB strain UWOPS-79140 and the SpC* strain LL2012_016 to distinguish regions of different ancestry. This data was plotted using the R library ggBio⁶⁶.

The average coverage and frequency of heterozygous sites within 1 kb windows along the genome was calculated for strains that showed large regions of missing information in HybridCheck (Supplementary Figure 12). This analysis was computed in Samtools v.1.8⁴³ with the command depth using sorted bam-files. Heterozygous sites where filtered from the variant file (vcf format), which was used in the first place to construct the global phylogeny.

Gene ontology enrichment for the introgressed genes was performed in GOrilla⁶⁷ using the complete list of genes present in our S. paradoxus strains as the reference list (Supplementary Figure 11, Supplementary Table 3).

Fixed introgressed genes of SpB-origin in SpC*

Previous analysis of Hénault et al.¹⁷ using the data from Leducq et al.¹⁵, identified 105 introgressed genes in H0 regions (= fixed and introgressed regions of SpB-origin) of SpC*. The number of introgressed genes in this study however is lower than the number detected by Hénault et al.¹⁷ (Supplementary Table 2). The difference is likely linked to the two distinct detection methods for introgressions and the different sets of strains (11 strains in Leducq et al.¹⁵ vs. 17 strains in this study). The detection method in Leducq et al.¹⁵ was based on a 5 kb window approach (non-sliding). This method likely increased the number of potentially introgressed genes by including all complete or partial, not introgressed genes found within 5 kb windows showing overall signature of introgression. The sliding window-based approach in HybridCheck was based on a 1 bp step size with an initial window size of 1 bp and therefore, identified shorter introgressed fragments with higher confidence, which led to a precise detection of introgressed genes in SpC* strains.

Distribution of introgressed fragments in SpC* and SpD

The size of fragments that showed SpB ancestry along the genomes of the 17 SpC* and 13 SpD strains was calculated using the data from HybridCheck. Fragments of shared ancestry that were within 5 kb distance were concatenated (Fig. 3b).

Dating of the origin of SpD lineage

To determine which of the two admixed lineages SpD or SpC* is more recent, the approximate age of the hybrid lineage SpC* and SpD was assessed from pairwise nucleotide divergence (Fig. 3c). This was done by estimating nucleotide divergence of SpB-like regions between SpD/SpC* and SpB (T₁). To make sure that the differences are not caused by mutation rate variation between lineages, divergence time of the same SpB-like regions between SpD/SpC* and SpC was estimated (T₂). To date the emergence of SpD, SpB-like regions present in SpD but absent in SpC* were extracted. T₁ was calculated using sequence alignments of 5 lineages with topology (((SpB, SpD),(SpC*,SpC)),SpA), where SpA is an outgroup. T₁ corresponds to split between SpB and SpD, and T₂ corresponds to split between (SpB, SpD) and (SpC*, SpC). T was estimated by counting private derived variants accumulated on tree branches of 5 lineage topologies in the following way:

$$T_1=\frac{1}{N}\times \frac{BAAAA+ABAAA}{2}$$ 1

$$T_2=\frac{1}{N}\times \frac{\frac{BAAAA+ABAAA}{2}+BBAAA+\frac{AABAA+AAABA}{2}+AABBA}{2}$$ 2

where N is the number of sites in the alignment, A are the ancestral and B are the derived variants. To date the emergence of SpC*, SpB-like regions present in SpC* were extracted. For sequence alignments of 4 lineages with the topology ((SpB, SpC*),SpC),SpA), an analogous estimation of T₁ and T₂ was performed, where T₁ corresponds to split between SpB and SpC*, and T₂ corresponds to split between (SpB,SpC*) and SpC. SpB-like regions in SpD and SpC* were determined using HybridCheck, with SpB and SpC strains as parental strains (see: analysis on the detection of introgression). Only regions having at least 1000 informative sites in the alignment were considered. After excluding positions with missing information, we obtained ancestral and derived variant patterns for each region with the fasta2dfoil program in the dfoil package⁶⁸. Times were estimated for all available 13 SpD strains, SpC* strain from the same location (16_199Ci), and one strain per SpB, SpC, and SpA.

Further, the Bayesian method implemented in Beast 2.0²³ was used to estimate the divergence time of SpD from other lineages (Fig. 3d). Two independent dataset were prepared: (1) Nine genes of SpB ancestry in SpC* (fixed in SpC*) and in SpD (fixed in all SpD; SpC* ancestry) were concatenated (Fig. 3a, Supplementary Table 4). (2) In addition, a second dataset comprised of concatenated sequence of 4 fragments from chromosomes V (position: 50–130 kb), XII (position: 860–1000 kb), XIV (position: 510–650 kb) and XVI (position: 510–780 kb) for a total size of 630 kb was selected. The fragments of the second dataset (2) were identified in HybridCheck as being inherited from SpC to SpC* and subsequently from SpC* to all the SpD strains. For both datasets, the following analysis was limited to 66 strains that were randomly chosen as representatives of all 5 lineages, but included all SpC* and SpD genomes (Supplementary Table 5). In Beast, a substitution rate of 1.67e−10 was chosen from estimates in Zhu et al.⁶⁹ for S. cerevisiae and previous studies that estimated the divergence time for SpC* in Leducq et al.¹⁵. A fixed clock that assumed equal substitution rates among branches was used with a calibration time of 100,000 years ago (±10,000 years) for the split of the lineages SpB and SpC, which corresponds to the onset of the last glaciation (~100,000 years ago). The analysis was performed using 15 × 10⁶ MCMC iterations with 30% of burn-in and statistical sampling every 10,000 iteration (Supplementary Table 6). We used FigTree v1.4.2⁷⁰ to visualize the trees.

High-throughput yeast colony growth

A total of 229 sequenced North American S. paradoxus strains available in our collection were arranged into four random arrays of 1536 positions on OmniTray plates (Thermo Fisher Scientific, Waltham, USA) of YPD solid media (1% yeast extract, 2% tryptone, 2% glucose, 2% agar) using a BM5-BC-48 colony processing robot (S&P Robotics Inc., Toronto, Canada). Each strain was represented in 12 replicates randomly distributed across the four arrays. Strains 16_236B, 16_277B and UWOPS-80-13 were excluded because their colony texture did not allow proper handling by the robotic pin tool, and strain 16_187A was excluded after the identification of a contamination in the sequencing data. The two outer rows and columns of the arrays were filled with wild S. cerevisiae colonies to avoid plate border effects on growth. The 1536 arrays were replicated on 25 different media (Supplementary Data 3) and incubated at 25 °C (unless otherwise stated): complex media (1% yeast extract, 2% tryptone, 2% agar) with 2% fructose, glucose, galactose, maltose, mannose, methyl α-D-glucopyranoside, sucrose or 3% glycerol; complex media with 2% glucose incubated at 4 °C, 15 °C, 30 °C or 37 °C; synthetic media (2% glucose, 0.174% yeast nitrogen base, 2% agar) with 0.5% asparagine, glycine, glutamine, histidine, isoleucine, lysine, proline, tyrosine or ammonium sulfate (AS); synthetic sucrose allantoin medium (2% sucrose, 0.174% yeast nitrogen base, 2% agar, 0.005% allantoin); synthetic complete media (2% glucose, 0.174% yeast nitrogen base, 2% agar, 0.5% AS, 0.134% complete drop-out) alone or supplemented with 50 μM menadione or 200 ng/μl rapamycin. After the colonies reached replicable size (48–96 h, depending on the condition), they were replicated on the same media. For the media incubated at 25 °C, plates were placed into an automated incubator-imager (S&P Robotics Inc., Toronto, Canada) and photographed at 2 h intervals. For the other temperatures, the plates were incubated in separate incubators and taken out briefly to be photographed on the colony-processing robot, first at 2 h intervals (6 h for plates at 4 °C) and at increasing time intervals afterwards.

Parameters imputation and dataset filtering

Plate images were processed using the gitter R package v1.1.1⁷¹ using a standard plate image as value for the argument ref.image.file and setting plate.format to 1536. All the downstream analyses were performed using custom Python v3.6.3 scripts. Time values assigned to each picture were corrected to attribute a single time value to all pictures taken during a same imaging round. Colony entries that had a size of 0 or flags for abnormal circularity or overlapping boundaries were removed. The filtered dataset contained between one and 12 replicates for each of the 315 333 (strain × condition × imaging time point) combinations, with an average of 11.74 replicates per combination and 0.012% of the combinations having less than three replicates. The median of the 12 or less replicate colony sizes for any given strain was transformed in log₂ and normalized by subtracting the initial median log₂ colony size to yield growth curves. For complex media with fructose, glucose, galactose, mannose, methyl α-D-glucopyranoside and sucrose at 25 °C, growth curves were truncated at 80 h because colonies on these media had achieved stationary phase. Each growth curve contained between 184 and 852 individual data points, corresponding to between 29.72 and 100% of the total expected data points (between 73.66 and 100% when excluding synthetic medium with histidine). Three metrics were extracted from these curves: area under the curve (AUC), maximum slope (MS) and endpoint colony size (ECS). AUC was computed by fitting a cubic spline to the raw data points using the UnivariateSpline function from the scipy Python package v1.1.0 and by integrating the resulting function using the spline’s integral method. The MS was computed as the 98th percentile of the set of linear regression slopes fitted in 5-timepoints wide overlapping sliding windows with a correlation coefficient r > 0.8. ECS was the colony size at the last time point of each growth curve. Only AUC and MS were considered due to the high correlation between AUC and ECS (Pearson’s r = 0.98). Although AUC and MS yield overall similar patterns in multiple factors analysis (MFA), both were considered separately for further analysis due to their lower correlation (Pearson’s r = 0.86). Growth on synthetic medium with lysine was extremely slow for all strains, thus this condition was excluded from the analysis. Strain MSH-1S11 was excluded due to visible contamination of the colonies. Additionally, strains UCD 62-186, UCD 62-268 and YPS695 were excluded because their AUC, MS or ECS value was 0 in at least one condition. A preliminary PCA analysis on the AUC data identified 8 strains that grouped closely on PC1 (43% of total variance explained): 14_164C, 14_169C, 14_177C, 15_005C, 16_033B, LL2012_004, R23 and yHKS267. These strains had very poor growth on all the conditions tested and were thus excluded from the analysis.

Colony growth data analysis by dimensionality reduction

Several dimensionality reduction methods were applied on AUC and MS data. PCA was performed using the decomposition.PCA function from the Python package scikit-learn v0.19.1 with argument svd_solver = ‘full’. LDA was performed using the discriminant_analysis.LinearDiscriminantAnalysis function from scikit-learn with argument solver = ‘eigen’. Additionally, the general agreement between AUC and MS data was assessed by applying multiple factor analysis (MFA), using AUC and MS values as groups of variables. MFA was performed using the MFA function from the Python package Prince v0.4.0 with arguments rescale_with_mean = False, rescale_with_std = False, n_components = 48, n_iter = 10, copy = True, engine = ‘auto’, random_state = 42.

Hierarchical clustering of growth values

Strains were clustered according to their AUC or MS values across conditions. Hierarchical clustering was performed using the function cluster.hierarchy.linkage from scipy, with complete linkage as the method and Spearman correlation as the distance metric. Heatmaps were generated using the clustermap function from the Python package Seaborn v0.8.1⁷², with values normalized by z-score transformation within conditions. Heatmaps were plotted by setting clustermap’s argument robust = true to restrict color mapping to the values falling between the 2nd and 98th percentiles of the data.

Growth comparison between SpC*, SpD, and parental lineages

AUC and MS values for the hybrid lineages SpC* and SpD (respectively SpD sub-group 1 (SpD1) and SpD subgroup 2 (SpD2)) were compared to that of their respective parental populations. To include all conditions in a global comparison, we normalized the growth values of all relevant strains in a given condition to center them around the intermediate point between parental distributions and to scale them according to the amplitude of the difference between the parental distirbutions (Fig. 4b). For each strain (hybrid or parental), we computed the scaled difference d between its value and the midpoint between the median values of the two parental lineages for each condition tested:

$$d=\frac{x_s-\frac{1}{2}\left[median\left(X_{P1}\right)+median\left(X_{P2}\right)\right]}{\frac{1}{2}∣median\left(X_{P1}\right)-median\left(X_{P2}\right)∣}$$ 3

where x_s is the value of a given strain (hybrid or parental), and X_P1, X_P2 are the distributions of values for the first and second parental lineages, respectively. The differences between distributions of d values for the lineages SpC* and SpD and parental populations was tested using one-sided Mann-Whitney U-tests.

Strain material and cell cultures for RNAseq analysis

A subset of 24 diploid strains that were representatives of all five S. paradoxus lineages were used for gene expression profiling (Supplementary Table 7, Supplementary Data 4). Two biological replicates per strain were grown at 25 °C in synthetic media supplemented with allantoin (0.174 % Yeast Nitrogen Base, 2% Sucrose, 0.005% allantoin). Allantoin was used because it was shown to best represent variation in growth on Maple sap, one of the natural environment in which S. paradoxus is found in North American deciduous forests⁷³. Two hundred and fifty milliliter cultures at 0.03 initial OD_600/ml were prepared from pre-cultures in exponential phase and cultures grew until reaching OD_600/ml of 0.6–0.7, were centrifuged at 4 °C to collect cell pellets and were immediately frozen in liquid nitrogen (stored at −80 °C).

Cell pellets were re-suspended in 2.5 ml lysis buffer (10 mM Tris-HCL pH 7.4, 100 mM NaCl, 30 mM MgCl₂) and cell droplets were prepared in liquid nitrogen. Frozen droplets were grinded in a RETSCH Mixer MM400 (Retsch, Haan, Germany) for 2 min and the chambers were place in liquid nitrogen to keep the cell extracts frozen. This procedure was repeated 15 times. Afterwards, samples were centrifuged for 10 min (13,000 g) at 4 °C. The supernatants (lysate) were transferred to new tubes, frozen in liquid nitrogen and stored at −80 °C until further processing.

RNA extraction, RNA library preparation and sequencing

Two hundred microliter of lysate for each sample was purified using the RNA Clean & Conentrator-25 kit (R1018; Zymo Research, Irvine, USA) following the manufacturer’s instructions. DNA contamination was eliminated by performing the DNase I In-Column treatment (E1010; Zymo Research). Samples were eluted in 25 µl DNase/RNase-free water. The Quantseq 3′ mRNA kit (Lexogen, Vienna, Austria) was used for library preparation⁷⁴. Fourty eight libraries were sequenced on 4 Ion Torrent chips with an expected read yield of >60 million reads per chip. The final average coverage over the 6000 genes was more than 800 reads per gene. Adapters, sequences of less than 30 bp in length, Poly-A tails and sequences with quality score lower than 15 were removed using the program Cutadapt⁷⁵. Read quality statistics were retrieved from the program FastQC⁷⁶. Raw sequences can be downloaded under the NCBI bio project ID PRJNA480398.

Genomes de-novo assembly and annotation

Six reference genomes were prepared (Supplementary Figure 18), respectively one for the lineages SpA (LL2012_026A), SpB (UWOPS-79140), SpC (LL2012_027), SpC* (LL2012_016) and two for the lineage SpD (SpD1 = R24, SpD2 = WX19), as the latter was shown to cluster in two distinct groups¹⁷. De-novo assemblies were performed on the raw sequencing reads from Leducq et al.¹⁵ and Xia et al.¹⁶. Megahit v1.1.2⁷⁷ was used to assemble reads into large contigs with the k-mer option set to 41, 51, 71, 91, 111, 131, 159, 187, and 215, which led to 706–1323 contigs, with a mean N50 of 137,674 (122,691 to 154,133). Contigs were assembled into chromosomes using Chromosomer v.0.1.4⁷⁸ using the CBS432¹⁸ sequence as reference. The program AUGUSTUS v. 3.2.3⁷⁹ predicted between 5398 and 5438 genes for the six reference genomes. A custom Python script was used to identify orthologous genes between the six reference genomes. Briefly, this script combined a syntenic approach with the detection of orthologs based on blastp⁸⁰ results from all query proteins against all reference proteins. The position of a given gene and the sequence similarity to ORFs of the reference was used to identify orthologous genes in the six reference genomes. Gene prediction was further extended to non-annotated ORFs detected from transcription data and increased the annotations of ORFs to 5315-5355 genes per reference genome (Supplementary Table 8).

Read counts and differential expression

The program BWA v.0.7.17⁸¹ was used to align the 48 transcriptomes to the six reference genomes of the same lineage (lineage-specific, Supplementary Data 4). Since the library preparation and subsequent sequencing produced only one fragment per gene that was located downstream in the 3′ UTR region, a custom script counted the number of reads per gene. Briefly, a first window of 400 bp (100 bp in 3’ end plus 300 bp of 3′ UTR region for each gene) was generated. In this first window, the shape of the distribution of reads was detected with the aim to identify the highest peak representing the expression of a gene. The window was further extended in 100 bp steps in both directions until the distribution was rendered by a valley of low read counts (<10 reads or <10% of reads from the highest detected peak in the extended window).

The analysis of differential expression was performed in R⁸² using the DeSeq⁸³ pipeline (Supplementary Figures 19-22). In DeSeq, raw read counts were normalized using the rlog function, which transformed the raw read count data into log₂ scale and corrected for variation coming from different library sizes. We performed PCA on the genes that were common to all individuals of the six lineages (in DeSeq; Fig. 4d, Supplementary Figures 20-21). GO-enrichment analyses of pairwise expression comparison (Supplementary Figure 22, Supplementary Data 5) were performed using Gorilla⁶⁷.

Heterothallic strain construction and spore viability

The mating type switching locus (HO) of five different SpD strains, respectively M2 and R22 of SpD1, and WX20, WX21, R21 of SpD2¹⁶ was deleted using the procedure and oligonucleotides described in Charron et al.¹⁹. Hygromycin (HPH), Nourseothricin (NAT) and Kanamycin (KAN) deletion cassettes were amplified with primers CLOP40-E9 (ACATCCTTATAGGCAGCAATCAATTCCATCTAAACTTTAACCAGCTGAAGCTTCGTACGC; forward) and CLOP48-C9 (TTTATTACATACAACTTTTTTTTAATAATATACATATTGCATAGGCCACTAGTGGATCTG; reverse) or CLOP40-E8 (TTAATTACATAACAATTTTTTTTTATAATATACATATTGCATAGGCCACTAGTGGATCTG; reverse). After transformation, successful deletions of HO were verified by PCR using primers CLOP48-C11 (ACAGAAGCTTGTTGAAGCGC; forward) and HPH_B (GTCGCGGTGAGTTCAGGCTT; reverse - HPH cassette) or KAN_B_R (CTGCAGCGAGGAGCCGTAAT; reverse - KAN cassette) or NAT_B (CGGTAAGCCGTGTCGTCAAG; reverse - NAT cassette). Diploids were sporulated and segregants were tested for mating type using primers verifMATa_F (ACTCCACTTCAAGTAAGAGTTTG; forward - MATa) or verifMATalpha_F (GCACGGAATATGGGACTACTTCG; forward - MATα) and verifMATa/alpha_R (AGTCACATCAAGATCGTTTATGG; reverse).

Crosses among the five SpD strains and with 11 additional strains from other lineages (Supplementary Data 6) were performed following the protocol in Charron et al.¹⁹. Spore viability of crosses was measured from the proportion of visible colonies from 24 tetrads per cross. Spores that form visible colonies after a 72-h incubation period at 30 °C were considered viable. In addition, fertility of wild, diploid SpD strains was quantified (without any genetic modification) by sporulation and subsequent dissection (Supplementary Data 6, Fig. 4e).

Detection of the chromosomal translocation VItXIII in SpD

To detect the presence of the translocation between chromosome VI and the right arm of chromosome XIII (VItXIII)¹⁵ in SpD strains, the scaffold TA04_6134 from Leducq et al.¹⁵ was used. This scaffold was generated from sequencing data from the SpC* strain 2012_018 and has a length of ~172 kb. It harbors a translocation from chromosome XIII that fused to the right arm of chromosome VI in the SpC*. This translocation was shown to be unique to all SpC* individuals and present in a few SpB individuals denoted as SpB-fusion (SpBf). To assess the presence of this translocation in the SpD strains, a subset of 42 strains from all five lineages was generated (including the three SpBf, all 17 SpC* and the 13 SpD strains) and raw sequencing reads were mapped to the scaffold TA04_6134 using Bowtie2 v2.1.0⁴². The coverage per single position was calculated with depth in Samtools v1.8⁴³ and averaged in 1 kb windows along the scaffold. The presence of reads at the junction (~20 kb) that connect the two fused sequences was used as evidence for the existence of the translocation (Supplementary Figure 23). From the alignment files (bam), a variant (vcf) file with the same parameters as above was prepared using Platypus v0.8.1⁴⁴ and vcftools v0.1.15⁴⁵. Phylogenetic relationships where assessed from variants in the scaffold region of 36 kb to 57 kb in R⁸² with the packages Ape⁸⁴ and Phangorn⁶⁴.

Code availability

All custom code will be available from the corresponding authors upon request.

Reporting summary

Further information on experimental design is available in the Nature Research Reporting Summary linked to this article.

Author contributions

C.E. and C.R.L. planned this study. C.E., A.F., M.H., and M.B. performed the population genomics analysis. C.E. and M.B. performed the transcriptomic analysis. M.H. performed the phenotypic screen and the long-read sequencing analysis. G.C. performed the reproductive isolation experiments. C.E. wrote the paper with contributions from M.H., A.F., G.C., M.B., L.M.K., J.B.A. and C.R.L.

Data availability

The data that support the findings of this study are available from the corresponding authors upon request. All sequencing data generated in this study are available in the NCBI BioProject and Sequence Read Archive (SRA) databases under the accession numbers PRJNA479851 (short reads for 74 genomes), PRJNA514804 (long reads for 6 genomes) and PRJNA480398 (48 transcriptomes). Sequences from previous studies can be found under the accession numbers PRJNA277692 (Leducq et al.¹⁵) and PRJNA324830 (Xia et al.¹⁶).

Competing interests

The authors declare no competing interests.

Electronic supplementary material

Supplementary Information accompanies this paper at 10.1038/s41467-019-08809-7.