Abstract
A large fraction of human complex trait heritability is due to a high number of variants with small marginal effects and their interactions with genotype and environment. Such alleles are more easily studied in model organisms, where environment, genetic makeup, and allele frequencies can be controlled. Here, we examine the effect of natural genetic variation on heritable traits in a very large pool of baker’s yeast from a multiparent 12th generation intercross. We selected four representative founder strains to produce the Saccharomyces Genome Resequencing Project (SGRP)-4X mapping population and sequenced 192 segregants to generate an accurate genetic map. Using these individuals, we mapped 25 loci linked to growth traits under heat stress, arsenite, and paraquat, the majority of which were best explained by a diverging phenotype caused by a single allele in one condition. By sequencing pooled DNA from millions of segregants grown under heat stress, we further identified 34 and 39 regions selected in haploid and diploid pools, respectively, with most of the selection against a single allele. While the most parsimonious model for the majority of loci mapped using either approach was the effect of an allele private to one founder, we could validate examples of pleiotropic effects and complex allelic series at a locus. SGRP-4X is a deeply characterized resource that provides a framework for powerful and high-resolution genetic analysis of yeast phenotypes and serves as a test bed for testing avenues to attack human complex traits.
THE strong tendency for progeny to closely resemble their parents has turned out to be difficult to understand in detail. Nearly all traits, including lifetime risk for many common diseases, have a complex genetic basis that is determined by multiple quantitative trait loci (QTL) (Donnelly 2008; Manolio et al. 2009). The first step toward accurate models of trait variability, and a prerequisite for predicting and modulating them, is characterization of the underlying genetic factors in the context of rest of the genome and their external environment. Research in model systems has led the way in this effort and produced powerful experimental and computational approaches for genetic mapping (Nordborg and Weigel 2008; Flint and Mackay 2009; Mackay et al. 2009).
A traditional, well-controlled approach for finding the QTL underlying natural phenotypic variation is to analyze a large number of progenies from two-parent crosses (Brem et al. 2002; Simon et al. 2008). Studies using this design have improved our understanding of complex traits and provided concrete evidence of natural segregating variants (Mackay et al. 2009), but have been limited in their scope with regard to the extent of genetic variation between the two parents. Mapping populations of popular model organisms ranging from fruit flies (King et al. 2012) and mice (Churchill et al. 2004; Durrant et al. 2011) to plants (Kover et al. 2009; Gan et al. 2011; Huang et al. 2011) has recently expanded the genetic and phenotypic diversity available to study by incorporating a wider repertoire of founder lines. These panels are the forefront of complex trait research and bear close resemblance to natural populations in multiple ways. Beyond increased variation, multiple founders introduce the possibility of more than two independent alleles at a locus and a larger space of potential epistatic interactions (Huang et al. 2012), while additional outcrossing rounds break linkage to further mix alleles. Nevertheless, lack of complete genotype information and the extent of remaining linkage in these recombinant lines have limited the power to detect small-effect QTL and identify the causative loci.
A multiparent mapping population has notably been missing in the budding yeast Saccharomyces cerevisiae, perhaps the most powerful eukaryotic model organism (Liti and Louis 2012). Yeast is a powerhouse of quantitative genetics due to its small genome size and very high recombination rate and the potential for accurate quantitative phenotyping, ease of obtaining and maintaining large mapping populations, and the ability to manipulate the genome at a single-base resolution. So far, nearly all yeast recombinant panels have been constructed by crossing the reference laboratory strain S288c (or one of its derivatives) with a wild isolate. However, laboratory strains poorly recapitulate the properties of natural populations, often represent phenotypic outliers when compared to the rest of the species (Warringer et al. 2011), and contain artificial auxotrophic markers that confound mapping experiments (Perlstein et al. 2007). To overcome this problem, we previously picked four natural yeast isolates sequenced in the Saccharomyces Genome Resequencing Project (SGRP) (Liti et al. 2009a) and released first-generation recombinant lines for each of the six pairwise crosses (Cubillos et al. 2011).
Here, we present SGRP-4X, a yeast mapping population obtained from outcrossing four wild founders representative of the main S. cerevisiae lineages for 12 generations. SGRP-4X contains >10 million segregants with fine-grained mosaic genomes and greatly reduced linkage, while retaining the phenotypic diversity of the parental strains. We demonstrate the power and resolution of QTL mapping in this population by both traditional linkage analysis on 179 genotyped and phenotyped individuals and a recently developed approach of sequencing the entire population under selection. SGRP-4X is a powerful, deeply characterized resource for high-resolution mapping of complex traits.
Materials and Methods
Generating the SGRP-4X advanced intercross lines
Two biological replicates of the intercrossed population were made. For replicate 1, parental strains YPS128 [North American (“NA”): MATa, ho::HygMX, ura3::KanMX] and DBVPG6044 [West African (“WA”): MATα, ho::HygMX, ura3::KanMX] were crossed and grown overnight in complete media (YPDA) to generate diploid F1 hybrids. In parallel, strains Y12 [Sake (“SA”): MATa, ho::HygMX, ura3::KanMX, lys2::URA3] and DBVPG6765 [Wine/European (“WE”): MATα, ho::HygMX, ura3::KanMX, lys2::URA3] were similarly crossed. To confirm successful crosses, we isolated individual colonies and performed mating tests, using tester strains Y55-2369 (MATa, hoΔ, ura2-1, tyr1-1) and Y55-2370 (MATa, hoΔ, ura2-1, tyr1-1) as well as diagnostic PCR for the MAT locus (Huxley et al. 1990). For replicate 2, we repeated the procedure described above with inverted combination of mating types (NA MATα × WA MATa and SA MATα × WE MATa). F1 hybrids and haploid segregants were treated and generated as previously described (Parts et al. 2011). Briefly, F1 hybrids were grown overnight and full plates were replicated onto KAc at 23° for sporulation during 10 days. Cells were collected in water, treated with an equal amount of ether, and vortexed for 10 min to kill unsporulated cells. After the cells were washed in water, they were resuspended in 900 ml of sterile water and treated with 100 ml of Zymolase (10 mg/ml) to remove the ascus. Cell mixtures were vortexed for 5 min to increase spore dispersion. In both replicates, haploid cells derived from both F1 hybrids (NA × WA and SA × WE) were mixed in equal amounts, vortexed for 5 min, plated onto complete YPDA media, and grown overnight. Full plates were replica plated onto minimal media (MIN) to select for diploid F2 hybrids containing genetic contributions from all four founders followed by a replica plating on YPDA. This procedure was repeated 11 times to create the F12 population, which we term SGRP-4X. Finally, F12 diploid hybrids were sporulated for 10 days in KAc at 23° and tetrads were dissected as previously described (Naumov et al. 1994). Viable spores with correct 2:2 segregations for the MAT locus and ura3 and lys2 auxotrophies were selected. We picked a total of 192 segregants (some from the same tetrad) from replicate 1 and stored them in glycerol stocks at −80°. We estimate that the pool goes through 10–15 mitotic generations during the high cell density replating cycles (YPDA/MIN/YPDA), between each sexual generation. This gives a point estimate of 150 cell divisions (12.5 generations × 12 intercross cycles) for SGRP-4X individuals starting from the founder strains, with a range from 120 to 180.
Sequencing and read mapping
The four founder strains, 192 isolated segregants, and large F12 segregant pools (see below) were sequenced using the Illumina HiSeq and Illumina GAII platforms, with 2 × 108-bp paired end libraries prepared according to standard protocol (Kozarewa et al. 2009; Quail et al. 2012). We used the Sanger Centre sequencing pipelines for base calling and alignment. BWA version 0.5.8c (r1536) (Li and Durbin 2009) was used to map reads to the S. cerevisiae reference genome with options “aln -q 15 -t 2”, and we further filtered mappings to have quality scores of at least 30. Parental genome assembly and SNP calling will be reported in a separate article (A. Bergström, J. T. Simpson, F. Salinas, L. Parts, A. Zia, A. N. Nguyen Ba, A. M. Moses, E. J. Louis, V. Mustonen, J. Warringer, R. Durbin, and G. Liti, unpublished results) and can be obtained from the web resource http://www.moseslab.csb.utoronto.ca/sgrp/.
Calling segregant genotypes
We called genotypes at each segregating site for each segregant, using a simple four-state hidden Markov model (HMM). The hidden state of the model for sample s and locus l is a 1-of-4 random variable xs,l = (xs,l,1, … , xs,l,4) that emits base i corresponding to parent p with probability es,l,i^(1 − xs,l,p)(1 − es,l,i)^xs,l,p, given the error probability of the base call es,l,i calculated from the base quality. The transition probabilities are derived from approximate recombination rates assuming uniformly distributed events. We used p(xs,l = xs,l+1) = exp(−r dl,l+1), where r is the per-base recombination rate fixed at 1.2 × 10−5, and dl,l+1 is the distance between the two consecutive loci. We used standard methods to fit the HMM. In practice, as individual base calls were confident and the sequencing was of sufficiently high coverage, this approach made little difference compared to using maximum-likelihood estimate calls for private alleles and joining consecutive calls into haplotype blocks.
Phenotyping segregants and reciprocal hemyzygous strains
The individual sequenced segregants and reciprocal hemizygous strains were subjected to precise growth phenotyping under three conditions (40° heat, 1.5 mM arsenite, and 100 µg/ml paraquat) in two technical replicates, using high-resolution microcultivation instruments (Bioscreen C; Oy Growth Curves, Raisio, Finland) for quantitative growth as previously described (Liti et al. 2009a; Cubillos et al. 2011; Warringer et al. 2011). Briefly, growth curves were dissected into three fitness variables, growth rate, growth efficiency, and growth lag, which were extracted using an automated procedure (Warringer and Blomberg 2003). Growth rate was calculated as population doubling time (hours) from the slope of the exponential phase, growth lag (hours) as the time until the start of detectable exponential phase, and growth efficiency (optical density units) as the optical density of cultures in stationary phase. The phenotypic variance for each F1 cross and for F12 outbred crosses was estimated as the sum of squares of deviations of single measurements from the mean. The number of transgressive segregants was calculated as previously described (Marullo et al. 2006), considering the fittest and weakest founders as phenotypic value boundaries.
Linkage mapping
We discarded segregants with evidence for diploid karyotype. For each phenotyped trait, we calculated the mean across technical replicates of the three fitness variables calculated as described above. We separately fitted the effect of each parental allele at every segregating site for each fitness trait in a standard linear model. We used the genotype at the mating-type locus and auxotrophic markers LYS2 and URA3 as covariates, as we found these to have a detectable effect on the growth profiles. We considered only sites where genotypes were called for at least 100 segregants and each parental allele was observed at least four times. The linkage LOD scores (base 10) were used to call peaks with strongest signal of at least LOD 4 and their support regions within 3 LOD units of it. We repeated this procedure 1000 times with permuted genotypes, keeping the correlation structure in the phenotypes due to the covariates. The nominal P-values for the peaks were calculated as the average number of peaks detected in permutations with the same or a stronger maximum linkage signal. We retained peaks with empirical false discovery rate (FDR) < 0.5 and also calculated their q-values as the expected fraction of false positive calls when using the peak P-value as cutoff. While the cutoff of 0.5 is lenient, it has good reproducibility properties, and we have validated several of these moderate signals.
Pleiotropy calculation
We calculated linkage LOD scores as described above and performed permutations in the same way, recording for each peak the strongest linkage LOD score for the other traits of any parental allele for any site in a 40-kb window centered on the peak. We calculated the nominal P-value of a pleiotropic effect as the average number of peaks per permutation with a LOD score of at least 4 and a secondary LOD score corresponding to another phenotype stronger than the observed peak’s best secondary LOD score in the same window. We called a pleiotropic signal if the nominal P-value was <0.1, which corresponds to an FDR < 30% (9 of 25 pleiotropic QTL, 2.5 expected).
Reproducibility calculation
We took the regions previously mapped for the same growth traits in F1 segregants (Cubillos et al. 2011) and calculated the strongest LOD score for the corresponding trait in any parental background in a 100-kb window centered on the locus. We calculated the nominal P-value as the fraction of all 100-kb windows in the genome that had an equal or stronger signal and calculated the P-values and their q-values in a standard way.
Pooled selection
Pools of millions of segregants were subjected to selection for high-temperature growth and nonselective media growth in duplicates as previously described (Parts et al. 2011). Briefly, two replicate pools of 10–100 million cells were collected from sporulation media for haploid and diploid F12 populations and treated with ether and zymolase. Spores were plated on YPDA and incubated at 40° until full growth was obtained. Each plate was incubated for 48 hr and then resuspended in distilled water. Ten percent of the cells were used for the next replating and the rest for DNA extraction. In total, 10 time points (T0, … , T10, corresponding to plates 0–10) for both control and heat-stress conditions were sampled, from which T0, T4, T8, and T10 were sequenced to an average genome coverage of 210× (Supporting Information, Table S6). We have previously argued that contributions of genetic drift and de novo mutations are negligible for detectable allele frequency changes in these experiments, due to the limited number of generations during the selection process (Parts et al. 2011; Illingworth et al. 2012).
Estimating allele frequencies
As a first pass, we estimated posterior allele frequencies for each parent separately. We considered all sites private to parent p. For each locus l, we inferred the posterior distribution of allele frequency fc,r,t,l,p ∼ Beta(ac,r,t,l,p, bc,r,t,l,p) for experimental condition c, replicate r, and time-point t following Parts et al. (2011). In short, prob(fc,r,t,l,p|D) ∼ prodl′ fc,r,t,l,p^(xc,r,t,l′,p×rl,l′)(1 − fc,r,t,l,p)^(xc,r,t,l′,∼p×rl,l′), where xc,r,t,l,p and xc,r,t,l,∼p are the number of alleles of parent p and alleles not of parent p observed in condition c, replica r, time t, and locus l, and rl,l′ is the recombination rate between loci l and l′. Loci l′ were chosen to be at least 100 bp apart to ensure the sampled alleles come from independent reads. rl,l′ was set to 0 if it was <0.9, so that only strongly linked sites were used in the allele frequency calculation. After the initial pass, we filtered out sites for which the posterior allele frequency mean ac,r,t,l,p/(ac,r,t,l,p + bc,r,t,l,p) was at least 0.2 away from the empirical mean xc,r,t,l,p/(xc,r,t,l,p + xc,r,t,l,∼p) at the site, as in the examined cases these outliers were due to alignment issues. We then also included sites with a 2:2 segregation between parents and recalculated the posterior allele frequency estimates from the first round of posterior estimates with an expectation-maximization-like approach. To compare allele frequencies in the sequenced segregants to those of sequenced pool, we split the genome into nonoverlapping windows of 500 segregating sites, calculated Pearson’s r for each parental allele frequency in each window separately, and used the median of all the calculated r’s as a summary statistic.
Locating QTL in pools
To identify selected regions, we estimated allele frequencies for each time point as outlined above, quantified their changes between sampled time points, compared the changes during selection to changes during a control experiment (pool grown under optimal conditions at 23°), and combined the comparisons across time points and replicates. For replica r of selection condition c, and every measured time point t > 0 (T4, T8, and T10), we calculated the posterior distribution of the difference in allele frequency dc,r,t,l,p at each locus l and parent p. We approximated the allele frequency posteriors with a normal distribution with mean mc,r,t,l,p and variance vc,r,t,l,p, which gives a normal distribution for the difference distribution, dc,r,t,l,p ∼ N(mc,r,t,l,p − mc,r,t−1,l,p, vc,r,t,l,p + vc,r,t−1,l,p). We then calculated P-values of increasing and decreasing allele frequency for each time point at each locus and any ploidy, by numerically estimating Pc,r,t,l,p+ = prob(dc,r,t,l,p > 0) and Pc,r,t,l,p,− = prob(dc,r,t,l,p < 0). We combined these P-values across time points, using Fisher’s method, calculating Pc,r,l,p,s = χ2(−2 log(prodt(Pc,r,t,l,p,s)); 2*Tr), where Tr is the number of time points sequenced for replicate r, and s is one of “−” for decrease and “+” for increase in allele frequency. We then calculated q-values from the P-values by considering the null distribution of all control experiment (c = 0) P-values {P0,r,l,p,s}r,l,p calculated as above, according to the method of Storey and Tibshirani (2003). We expect allele frequencies not to be subject to selection during the control experiment, and thus any changes should be attributed to drift or other unknown confounding factors. We called QTL as regions where the minimum q-value across directions and parents was <0.05, combining consecutive sites with minimum q-values below the threshold into a single QTL. Ties (e.g., in cases of stretch of q-values = 0) were ordered by minr,t,p,s Pc,r,t,l,p,s to prioritize loci.
We considered QTL to be shared between the previously described intercrossed F12 WA × NA population (Parts et al. 2011) and the SGRP-4X when the distance between mapped intervals was <10 kb.
QTL model selection
To compare goodness-of-fit of biallelic (1 vs. 3, 2 vs. 2) and more complicated (1 vs. 1 vs. 2, 1 vs. 1 vs. 1 vs. 1) linkage models, we used the Akaike information criterion, 2k − 2 log prob(D|x). The log-likelihood was calculated from the assumed underlying normal distribution in the standard way, and k is the number of free parameters in the model. We calculated the Akaike information criterion (AIC) for all possible allele configurations and picked the one with the smallest score as the QTL model.
For similar comparison of the selection QTL models, we calculated the AIC as above, with a different likelihood model. We posited the existence of one, two, three, or four different driver alleles. For each model and replicate, we estimated a separate fitness parameter for each driver allele from its frequency changes between day 0 and day 8. We then calculated the expected frequencies of the other alleles under this model and calculated the probability of observing each allele frequency from a normal distribution N(expected change; observed change, sp2), where sp2 is the variance in allele frequencies estimated from changes during the control experiment c separately for each allele, sp2 = varl,r(mc,r,1,l,p − mc,r,0,l,p). We picked the model with the smallest average AIC across the two replicas, if it was substantially better supported (AIC difference at least 1.0) than the null model of no change.
Results and Discussion
Genetic background of founder strains
We selected four founder strains of distinct geographic and ecological origins (Figure 1A) that are segregating at 64% of the S. cerevisiae SNPs previously described (Liti et al. 2009a). We used the DBVPG6765 strain as representative of the WE lineage, YPS128 of the NA lineage, Y12 of the SA lineage, and DBVPG6044 of the WA one. These strains have been extensively characterized at the genomic (Liti et al. 2009a) and phenotypic (Warringer et al. 2011) levels and successfully used for QTL mapping in two-parent crosses (Cubillos et al. 2011; Parts et al. 2011; Salinas et al. 2012). To further improve the reference genomes, we resequenced the founders at high coverage and generated high-quality assemblies, containing >95% of the sequence for each strain in one large scaffold per chromosome (A. Bergström, J. T. Simpson, F. Salinas, L. Parts, A. Zia, A. N. Nguyen Ba, A. M. Moses, E. J. Louis, V. Mustonen, J. Warringer, R. Durbin, and G. Liti, unpublished results). With the exception of the subtelomeric regions, we found no large structural variants that could prevent meiotic recombination and cause significant allelic distortions, making these strains highly suitable for genetic crosses.
Genetic diversity and breeding of the founder strains. (A) The four founder strains sample a large fraction of the genetic variation in the species. The phylogenetic tree includes all the strains sequenced in the SGRP project (Liti et al. 2009a), with major geographic clusters highlighted. (B) Distribution and summary statistics of single-nucleotide variants between the founders. The Venn diagram shows the sharing of derived alleles between parents, after rooting each variant using S. paradoxus as the outgroup. An additional 20,405 SNPs (19% of the total) are not included as they could not be rooted. Numbers in sectors give the count and fraction of nonsynonymous polymorphisms and fraction of nonsynonymous polymorphisms predicted to have a deleterious effect on protein function from SIFT analysis (Ng and Henikoff 2003). (C) Chromosome II sequence similarity between founders (logarithmic scale, full-genome plots given in Figure S1). The pairwise similarity between founder strain genome sequences was computed in windows of size 10 kb as N/(d + 1), where N is the number of sites in the window with high-quality alignments and d is the number of single-nucleotide differences between the strains within these sites. Values corresponding to different levels of sequence divergence are indicated on the vertical axis. Strain colors are the same as in A. The bottom two similarity plots show the same statistic for S288c/RM11 and S288c/YJM789 crosses previously used in yeast QTL mapping, with an uneven distribution of sequence variation. (D) Cross design of SGRP-4X (see also Figure S2).
We identified a total of 109,701 SNPs in the parents, most (86%) of which are private to a single lineage (Figure 1B). In comparison, a commonly used BY × RM cross has less than half the number (∼47,000) of segregating sites (Nagarajan et al. 2010). The majority, 68,727 SNPs, map within nondubious ORFs (average of 11.8 SNPs per gene) and nearly one-third, 21,467 SNPs, alter the protein sequence (average of 3.7 changes per protein). Of the 5793 nondubious ORFs in the genome, 5414 (93%) have at least one SNP and 4542 (78%) have at least one amino acid difference between the parents. Further, 16% of all nonsynonymous changes are predicted to be deleterious by assessing the conservation of the residue in homologous proteins with the SIFT score (Kumar et al. 2009; A. Bergström, J. T. Simpson, F. Salinas, L. Parts, A. Zia, A. N. Nguyen Ba, A. M. Moses, E. J. Louis, V. Mustonen, J. Warringer, R. Durbin, and G. Liti, unpublished results). Interestingly, derived alleles (as determined from comparisons with other Saccharomyces species) private to a single lineage are predicted to be deleterious more frequently than those present in multiple lineages (18% vs. 9%) (Figure 1B and A. Bergström, J. T. Simpson, F. Salinas, L. Parts, A. Zia, A. N. Nguyen Ba, A. M. Moses, E. J. Louis, V. Mustonen, J. Warringer, R. Durbin, and G. Liti, unpublished results).
Spatial statistics of genetic variation are important determinants of the mapping population quality. Nearly all previous yeast QTL mapping studies have utilized laboratory strains with mosaic genomes of variable local ancestry (Liti et al. 2009a). In contrast, genetic distances between our founder strains estimated in 10-kb windows are narrowly centered on the genome average, and thus the polymorphic sites are evenly distributed across the genome (Figure 1C and Figure S1). Therefore, we avoid problems arising from admixture that otherwise affect the nature and spatial distribution of identified QTL and interactions between coevolved alleles.
Genomic landscape of the mapping population
We generated the SGRP-4X mapping population using a funnel design similar to the mouse collaborative cross (Churchill et al. 2004) (Figure 1D and Figure S2). Initially, we crossed haploid segregants from two F1 crosses containing complementary auxotrophies and selected for diploids on minimal media (Materials and Methods, Figure S2). We then applied 11 rounds of random intercrosses as previously described (Parts et al. 2011). In addition to establishing the >10 million-segregant population for pooled studies, we isolated 192 haploid individuals, sequenced them to a median of 11× coverage, and called genotypes (Materials and Methods). Thirteen segregants and 11 individual chromosomes were heterozygous for a large fraction of polymorphic sites, suggesting contamination or aneuploidy (Figure S3), and were filtered from further analyses.
The majority of the genome had all four parental backgrounds segregating in the pool, with median allele frequencies of 0.26 (WE), 0.21 (WA), 0.26 (NA), and 0.26 (SA) and parental allele frequencies between 0.1 and 0.5 for 89% of the sites (Figure 2A). The lower allele frequency of the WA genome is consistent with slightly slower growth of the WA parent on minimal medium and less efficient sporulation, resulting in negative selection against the loci harboring responsible WA alleles. There were 9 regions of at least 40 kb with more extreme allele frequencies (<10% or >40%, Table S1A) and 54 (<15% and >35%) after the intercrosses (Table S1B, Figure S4), some of which are analyzed further below. Allele frequencies estimated by deep sequencing total DNA from the entire pool were strongly correlated with estimates from segregants (median Pearson’s r in 500 site blocks >0.75, Materials and Methods), indicating that the isolated individuals are representative of the population.
Genetic and phenotypic diversity of SGRP-4X. (A) Distribution of the fraction of the genome contributed by each founder to the sequenced segregants. (B) Fine mosaic structure of SGRP-4X individuals. Chromosome IIs from 10 individual segregants (y-axis) are painted according to the parental background at each locus, using the colors from Figure 1A. (C) Inferred scaled linkage disequilibrium measure D′ for chromosome II, evaluated in 5-kb windows in the WA × NA cross, SGRP-4X, and the S288c × YJM789 cross from Mancera et al. (2008). Loci with minor allele private to one founder were used for the SGRP-4X inference. Blue crosses reflect the absence of measurement in regions that did not contain markers. (D) Distribution of normalized growth rate of individual isolated segregants under nonstress, heat, arsenite, and paraquat conditions. Average values for founder strains from multiple replicas are indicated.
De novo SNP calling in the 11.2-Mb mappable genome of the 179 haploid individuals yielded 100 SNPs. This corresponds to a mutation rate of 3.4 × 10−10 per site per cell division (assuming 150 generations during the intercross process, Materials and Methods), very close to the commonly cited rate of 3.3 × 10−10 (Lynch et al. 2008). There were five de novo SNPs in the 591 kb of the genome subject to selection during the intercross (5.2 expected, Table S1A), and none of them were predicted to be deleterious by SIFT. Thus, we have no evidence supporting their contribution to formation of allele frequencies in SGRP-4X. A full description of these variants will be reported in a separate article (A. Bergström, J. T. Simpson, F. Salinas, L. Parts, A. Zia, A. N. Nguyen Ba, A. M. Moses, E. J. Louis, V. Mustonen, J. Warringer, R. Durbin, and G. Liti, unpublished results).
Whole-genome sequencing of the segregants allows accurate estimation of the genetic map. The outcrossing divided each genome into haplotype blocks identical by descent to one of the parental strains (Figure 2B), with an average segregant having 374 such blocks of median length 23 kb (Figure S5A). Using a recently developed inference method that leverages linkage disequilibrium between genotyped sites (Illingworth et al. 2013), we estimated an average genome-wide per-base per-generation recombination rate of 3.2 × 10−6, giving a total number of recombination events approximately fourfold higher than that observed for F1 segregants (Mancera et al. 2008). Surprisingly, the recombination rate in SGRP-4X is 80% higher than estimated in the F12 WA × NA cross (1.7 × 10−6, Figure S5B). We observed recombination hot and cold spots in concordance with previous reports, with centromeres exhibiting a lower than average recombination rate (Figure 2C), and a set of short haplotype blocks likely corresponding to noncrossovers (Figure S5A). A full description of the inference and analysis of recombination rates is given elsewhere (Illingworth et al. 2013).
Overall, parental contributions to the SGRP-4X population are close to the 25% expected under selectively neutral random outcrossing. This compares favorably with smaller multiparent populations in other model organisms (Kover et al. 2009; Philip et al. 2011; King et al. 2012), some of which have an unequal representation of founder strains due to selection and drift during breeding. While a handful of loci were selected in the SGRP-4X during intercross (Table S1), loss of genetic complexity by drift was negligible due to the extremely large population size (>10 million) maintained. The even contribution from founders coupled with the very high number of recombination events makes SGRP-4X well suited for QTL mapping.
Phenotypic diversity of SGRP-4X
Outcrossing a multifounder population over many generations breaks linkage disequilibrium and increases haplotype diversity. As a result, allele structures that have coevolved in parental lineages over millions of years are disrupted or rearranged in new combinations, which can alter the distribution of phenotypes. To assess whether traits were affected by the genetic churning of the intercross, we quantified the mitotic growth properties (rate, lag, and efficiency) of the retained 179 haploid segregants in basal conditions (nonstress) and exposure to three stresses: a toxic natural variant of arsenic [arsenite, As(III)], an oxidative agent generating superoxide radicals (paraquat), and high temperature (heat, 40°) (Materials and Methods, Table S2).
Overall, segregants were remarkably fit in nonstress conditions, with almost all of them (170/179) exceeding the growth rate of the notoriously slowly proliferating reference strain BY4741 (Figure 2D). Thus, the breakup of coevolved alleles had little overall effect on fitness, in contrast to severe growth defects detected in crosses between highly diverged lineages of the sister species S. paradoxus (Liti et al. 2009b). We observed a generally poor performance of the segregants in paraquat, to the extent that 35 F12 recombinants were unable to proliferate in the concentration previously used for F1 progeny (Cubillos et al. 2011). We hypothesize that this reflects the presence of epistatic interactions that have evolved in the parental populations for pathways underlying the oxidative stress response. These interactions are broken up in the multiparent cross due to the more rare cosegregation of compatible alleles and the reduction in physical linkage, which will target linked QTL.
Many crossing rounds can reduce phenotypic complexity in the population due to stabilizing and directional selection for particular trait values or outbreeding depression. The distributions of measured growth traits were unimodal and lacked skew in the nonstress condition, which would have been expected if substantial directional selection had been acting during the intercross (Figure 2D). To assess the evolution of trait complexity across generations, we compared the SGRP-4X phenotype distributions to those measured in the pairwise F1 crosses between the four parents (Cubillos et al. 2011). The SGRP-4X segregants were 1.6 and 2.7 times more variable in the absence of stress and in high-temperature environments, whereas they were only half as variable in the presence of As(III) (Figure S6). These trends were also reflected in the degree of transgression, i.e., the frequency of segregants with more extreme trait values than those of the parents (Figure S7), with more variable traits showing higher transgression. The markedly negative skew of paraquat growth phenotypes in SGRP-4X was evident in the extreme levels of negative transgression (47 transgressive segregants in paraquat vs. 3 and 1 in heat and arsenite, respectively), even under substantially milder paraquat exposure (Figure S7).
Taken together, the data from individually phenotyped segregants show that increasing genetic diversity by including multiple parents and performing many rounds of intercross did not substantially reduce fitness, had mostly limited impact on trait complexity, and imposed no strong directional selection during the process.
QTL mapping by linkage analysis
Availability of whole-genome genotype and phenotype data for the 179 segregants allowed us to assess the ability to detect QTL for the growth traits described above. While the segregants were sequenced primarily with the goal of assessing the recombination landscape (Illingworth et al. 2013) and we were powered only to detect large effects, it is informative to assess the QTL mapping resolution and reproducibility properties when using individual segregants. We identified 6, 16, and 3 significant QTL for mitotic growth properties under exposure to heat, arsenite, and paraquat, respectively (FDR < 0.5, Materials and Methods, Table S3 and Figure S8A), with a median interval size of 27.8 kb. Five of the 7 previously detected QTL with LOD > 9 from the six pairwise F1 crosses (Cubillos et al. 2011) were replicated in SGRP-4X at a FDR = 0.17 (Materials and Methods, Table S4A). This number increased to 16 of 24 QTL with a more lenient cutoff of LOD > 3 for F1 crosses (Cubillos et al. 2011) and FDR = 0.5 for this study (Table S4B). The rest of the previously mapped QTL could be undetected due to the complexity of the genetic interactions, overestimates of the true effect size in a smaller sample (Beavis effect), or false positives in the previous screen.
In multiple cases, the QTL regions included known causal or strong candidate genes. One replicated QTL at chromosome IV (1505–1524 kb) included the polymorphic ARR gene cluster present in the SA background (Cubillos et al. 2011), which is known to be essential for arsenite tolerance (Bobrowicz et al. 1997). The arsenite growth rate QTL at chromosome XV (932–957 kb) contains the candidate gene VMA4, a subunit of the vacuolar H+-ATPase known to be involved in As(III) tolerance in the reference strain (Zhou et al. 2009), with the peak of strongest linkage only 340 bp away (Figure 3A).
Linkage analysis in 179 SGRP-4X segregants. (A) Strength of linkage to arsenite resistance for a QTL region mapped on chromosome XV. For each parental allele, −log10(P) values of linkage (y-axis) are given across a chromosome XV region (x-axis). Colors for alleles are as in the bottom panel and Figure 1. Bottom panel shows growth rate in arsenite (y-axis) for segregants with different genotypes at the site of strongest linkage at the QTL locus (x-axis, jitter added). Average values for each parent are given under the corresponding segregants. (B) Number and pleiotropy of linkage QTL. The number of genome-wide significant QTL found using linkage mapping for the three conditions is displayed on the x-axis, with colors indicating pleiotropic QTL significant in another condition at a lower threshold (Materials and Methods).
We next asked whether some QTL were present in more than one condition. We defined a QTL as pleiotropic if the strongest linkage to growth in another environment within a 20-kb window was significant (Materials and Methods) and identified 22 trait-specific QTL and 3 pleiotropic QTL (Figure 3B, Table S5, and Figure S8B). For example, the arsenite QTL at chromosome XI 68–82 kb also had a strong linkage signal for heat growth efficiency at chromosome XI 48 kb. This region contains the gene PEX1, which is an AAA-peroxin whose deletion allele in the reference background has a fitness effect under both heat stress (Sinha et al. 2008) and arsenite (Pan et al. 2010).
A segregant population allows detection of antagonistic alleles that increase fitness in spite of coming from an unfit strain or vice versa (Rieseberg et al. 2003). For example, although the NA strain has superior resistance to heat, we found NA alleles that are detrimental for this phenotype (Table S3). Similarly, five of the QTL mapped in the arsenite-resistant SA and WA strains were antagonistic and contributed negatively to mitotic growth in the presence of arsenite. Such alleles may have arisen in yeast due to antagonistic pleiotropy (Orr 1998; Qian et al. 2012) or linkage to selected alleles (Liti and Louis 2012). Once released from their original genetic context and placed in an environment with a single stressor, they are free to exhibit their positive effect on growth.
The 12 rounds of intercrosses break linkage between nearby sites and improve QTL mapping resolution. Whereas we would expect a 3-LOD support interval of 64 kb in the 96-segregant F1 populations (Cubillos et al. 2011) if recombination is uniform, it narrows down to 6 kb in the 179 F12 segregants from SGRP-4X. In practice, we observed longer support intervals (median 27.8 kb), as the heterogeneity in recombination rate results in the majority of the genome having stronger than average linkage, and some regions likely contain linked QTL (Liti and Louis 2012) that we merged into a single peak. While with 179 sequenced segregants we did not have enough power to map the abundant weaker-effect alleles that make up a large fraction of the narrow-sense heritability (Bloom et al. 2013), we did recover the strong QTL from our earlier work. Overall, we mapped 25 narrow QTL regions that replicated previous results and contained strong candidates for follow-up, demonstrating the utility of SGRP-4X for complex trait mapping.
Allele frequency dynamics in SGRP-4X under heat selection
QTL can be efficiently mapped by bulk genotyping selected populations (Brauer et al. 2006; Segre et al. 2006; Ehrenreich et al. 2010; Wenger et al. 2010), which is most powerful when selection operates for a prolonged period (Parts et al. 2011). To apply this approach in SGRP-4X, we used the 40° heat exposure environment, where two of our founder strains (NA and SA) are fit, while the other two are moderately unfit (WE) or strongly unfit (WA) (Figure 2D, Materials and Methods). In addition, growth under high temperature is a classical complex trait, for which many genes are known to be involved, and a large number of causal loci have been characterized (Steinmetz et al. 2002; Sinha et al. 2006, 2008; Cubillos et al. 2011; Parts et al. 2011).
We identified 34 and 39 regions with significant allele frequency changes in haploid and diploid pools, respectively, and designated them as QTL (FDR 1%, Materials and Methods, Table S7, Figure 4A, and Figure S9). Nineteen of these overlap between ploidies, suggesting that heat-resistance QTL operate largely independently of ploidy, indicating dominant or additive effects and giving additional evidence for the reproducibility of our calls. The mapped regions have a median size of 4.8 kb, a resolution comparable to that expected under linkage mapping and finer than that of bulk segregant analyses in F1 crosses (Ehrenreich et al. 2012).
QTL mapping from allele frequency changes upon heat selection. (A) Allele frequencies at IRA2 region during heat selection. Top four panels represent allele frequencies (y-axis) of each parental allele in the pool at different time points [day 0 (T0), black; day 8 (T4) haploid, green; day 16 (T8) diploid, red; day 20 (T10) diploid, blue] around the IRA2 locus (x-axis). The bottom panel represents genes around the interval and combined signal for allele frequency change (y-axis). Evidence for increase [−log10(P)] is given in the top part, while evidence for decrease is in the bottom (Materials and Methods). (B) Overlap of QTL mapped using selection. Venn diagram depicts QTL found previously in the NA × WA cross (purple) and in this study using haploid (gray) and diploid (orange) pools. (C) Decay of linkage in a region under selection for two cross designs. Thick blue lines show the LD′ profile (y-axis) up- and downstream from the driver location of chromosome XV (IRA2, x-axis) for the WA × NA cross (left) and SGRP-4X (right). The fraction of observed WA alleles at each site is plotted for each segregating site with a black circle for pool after control experiment (WA × NA cross, replica 2, T4; SGRP-4X, replica 2, T4), and a red circle for pool after heat selection (WA × NA cross, replica 2, T6; SGRP-4X, replica 2, T10). Only the sites where WA is the derived allele are plotted for SGRP-4X. The LD′ profiles are computed from the segregants’ genotype data (Illingworth et al. 2013) with no knowledge of the selection data. The LD′ curves demonstrate complete linkage at the driver locus and gradual decline as a function of the distance to the driver, which is faster in SGRP-4X and visually matches the difference in size and shape of the selective sweeps upon heat selection.
The number of QTL found is a substantial increase compared to the 21 mapped using the same methodology in a F12 WA × NA cross (Parts et al. 2011). Seven and 10 of these previously identified loci were detected in the haploid and diploid outbred populations at 1% FDR, respectively (Figure 4B), which increased to 13 and 17 at a more lenient 5% FDR (Table S7). A further 3 were within 30 kb of a QTL, but outside its mapped region, suggesting either linked QTL or low resolution in one of the two studies. For three loci (IRA1, IRA2, and the subtelomeric region of chromosome XIII), we previously observed fixation of the beneficial allele upon selection (Parts et al. 2011; Illingworth et al. 2012). In contrast, while the allele frequencies in these regions changed by >15% in the current study, no allele was fixed or completely removed from the SGRP-4X pool even after long-term selection (Figure 4A).
Both linkage and selection mapping approaches benefit from additional recombination events to give narrower QTL intervals. We found that the allele frequency changes in the IRA2 region selected under heat stress were markedly different between the previous WA × NA cross and SGRP-4X (Figure 4C). The higher recombination rate in SGRP-4X resulted in a QTL centered more precisely on the validated IRA2 gene. Also, a large upstream region was no longer selected for in SGRP-4X, which could be explained either by a linked upstream QTL or by differences in the local recombination rate between the two crosses. The large decrease in linkage disequilibrium (LD) in the region (Figure 4C) strongly supports the latter explanation, showing how recombination reshapes the landscape of allele frequency changes under selection.
Surprisingly, all but one of the alleles increasing in frequency (7/8 and 6/6 in haploids and diploids) were from the only intermediate heat-resistant WE background, thus showing abundant advantageous variants in one parent. We found no positively selected alleles from the NA strain, which is most heat resistant. All the detrimental alleles were from heat-intermediate and heat-sensitive strains, with 17 and 25 from WE and 14 and 10 from WA in haploids and diploids, respectively (Table S7). The abundance of detrimental variants supports the recently found excess of loss-of-function alleles in the unfit WA strain (Warringer et al. 2011; Zorgo et al. 2012), likely due to adaptation to a very specific niche that does not include a strong positive influence of high temperature. Overall, the large number of regions selected in the SGRP-4X population under high temperature deepens the characterization of the highly complex heat-resistance phenotype, with different wild strains harboring alleles that modulate growth and survival.
Complexity of genotype–phenotype relationship in the SGRP-4X
A key goal of quantitative genetic studies is to estimate the phenotypic contribution of different alleles at each QTL (King et al. 2012). In the SGRP-4X and other multiparent populations (Churchill et al. 2004; Kover et al. 2009; King et al. 2012), it is possible to distinguish between biallelic QTL, where there are only two distinct effect sizes that segregate either in 1:3 or in 2:2 ratio (Figure 5A), and multiallelic ones, where more than two alleles have an independent fitness effect (segregating 1:1:2 or 1:1:1:1).
Allelic heterogeneity of QTL. (A) Number of QTL for different segregation models. The biallelic (1:3 or 2:2) and multiallelic (1:1:2 or 1:1:1:1) models are defined based on phenotypic effects. Colored circles depict the four alleles distributed in all possible fitness-segregating scenarios, “>” and “<” indicate greater or lower fitness. Numbers refer to QTL fitting each model for linkage analysis, intercross, and heat selection in haploid and diploid populations. (B) Allele frequencies in three regions (chromosomes VIII, XII, and V) after intercross rounds, but before selection experiments. Chromosome VIII region was inferred to have a 1:3, chromosome XII region a 2:2, and chromosome V region a complex 1:2:1 allele segregation. Bottom panel of each plot highlights individual genes around the interval, with potential causal candidates in green.
Comparison of linkage models with varying numbers of free parameters (Material and Methods) revealed that most linkage QTL (5/6 for heat, 2/3 for paraquat, and 10/16 for arsenite) are best explained by the effect of a single allele (Table S8 and Figure 5A). The more complex 1:1:2 patterns, where backgrounds of two founders have an independent beneficial or deleterious effect, explained the remaining QTL, while the 2:2 segregation pattern, with the causal allele shared between two founders, was never observed. These results suggest that most strong genetic contributors to the phenotype in SGRP-4X are alleles private to a single founder. Abundant rare coding variants specific to individual human populations recently reported (Abecasis et al. 2012; Fu et al. 2012; Keinan and Clark 2012; Tennessen et al. 2012) could similarly underlie a substantial proportion of heritable trait variation.
Next, we used statistical model selection in QTL regions of the intercross (Table S1) and heat stress (Table S7) to infer which alleles have independent fitness effects. For example, we previously detected a chromosome V region strongly selected during consecutive rounds of crossing of the WA and NA strains (Parts et al. 2011; Illingworth et al. 2012), but we were unable to distinguish positive selection for one parent from negative selection against the other. Allele frequencies in SGRP-4X indicated at least three different fitness values for this locus, with strong selection for the NA parent, two neutral alleles (SA and WE), and negative selection against the WA background (Figure 5B). In general, we found almost all regions strongly selected during the crossing to have more than one driving allele (7/9 regions, Table S8B). For heat selection after the crossing, we could confidently identify 29 and 34 biallelic QTL and 5 and 5 multiallelic QTL in haploids and diploids, respectively (Table S8, C and D). For instance, we observed strong selection for the NA version of IRA1 on chromosome II, weaker selection for the WE one, and negative selection against the SA and WA alleles.
Understanding the effects of multiple variants at a single locus has so far been limited by mapping precision and sensitivity, experimental variance in phenotyping, and low throughput of candidate validation (Mackay et al. 2009; Trontin et al. 2011). We have overcome some of these constraints in yeast by sensitively identifying allelic contributions in a pool under selection (Parts et al. 2011). Here, we observed a predominance of single alleles underlying QTL, consistent with a large fraction of heritable phenotypic variation being due to alleles private to individual founders. However, more complex allelic series for QTL in selection pools were also evident. Our results are consistent with previous findings of excess biallelic QTL compared to other allele distributions in F1 crosses (Cubillos et al. 2011; Ehrenreich et al. 2012), although this could be affected by the choice of founder strains, crossing design, and mapping approach.
Narrow mapping intervals allow rapid causal gene identification
QTL mapping in SGRP-4X resulted in narrow regions ripe for following up with single-gene studies. First, we used reciprocal hemizygosity (Steinmetz et al. 2002) to test the effects of different IRA1 and IRA2 alleles on growth under heat stress. The IRA genes are key regulators of the RAS signaling pathway, whose defects result in hyperactive RAS. This in turn leads to high levels of cyclic AMP (cAMP) and high PKA activity that inhibits the heat-stress response. We have previously shown that WA alleles of IRA1 and IRA2 have a reduced fitness at high temperature (Parts et al. 2011) and mapped these loci again in SGRP-4X, where the frequency of IRA2WA decreased from 0.47 to 0.32 under selection, while IRA2NA and IRA2SA frequencies increased by 6% and 27%, respectively.
Quantitative growth curves and plate spot dilutions of reciprocal hemizygous strains support the model where IRA2WA is heat sensitive, while IRA2NA and IRA2WE have nearly identical and superior fitness (Figure 6A and Figure S10). Surprisingly, although IRA2SA showed the greatest frequency increase, we did not find differences with any other allele in the reciprocal hemizygosity assay (Figure S10), even when compared to the weak IRA2WA. In contrast, at the IRA1 locus, allele frequency changes suggest that IRA1SA and IRA1WA are both deleterious (Table S8). We replicated this result by reciprocal hemizygosity in the WA × SA cross, where the IRA1SA allele did not outperform the known deleterious IRA1WA version (Figure 6B). These results suggest that in the SA background heat resistance might not be mediated through the cAMP/PKA pathway and instead be triggered by alternative mechanisms and complex genetic interactions.
Phenotypes of strains with reciprocally hemizygous genotypes for causal genes. (A) Lower fitness of the IRA2WA allele. Shown is the average of each fitness component measurement (n = 2) for all combinations of reciprocal hemizygote pairs of IRA2 alleles. Individual alleles were deleted from the heterozygous diploid strain and assessed for their individual contributions to high-temperature growth. (B) Fitness averages for the IRA1 reciprocal hemizygotes in the WA × SA cross. (C) Mitotic growth curves for arsenite and high-temperature growth in reciprocal hemizygotes for the VPS53 gene derived from the WE × SA and WA × SA crosses.
Finally, we validated the pleiotropic effect of candidate gene VPS53 in heat and arsenite in the WE × SA and WE × WA hybrids. While the VPS53 region LOD score was below the genome-wide significance cutoff in QTL mapping, it had strong support in the early phenotyping data, which reduced to moderate linkage support in both conditions after observing all the replicates. Hybrids containing the WE allele had a substantial fitness advantage under both stresses, while no differences were observed for any of the other alleles (Figure 6C and Figure S11A). Interestingly, the effect manifested primarily on the lag time in arsenite, a compound mainly acting via a lag time increase, whereas it manifested principally as an efficiency effect when encountering heat stress (Figure S11B). Thus, the mapped QTL is strain dependent (Cubillos et al. 2011) with VPS53WE being the fittest allele with a fully penetrant effect regardless of the cross combination.
The narrow mapping intervals in SGRP-4X allowed us to rapidly follow up promising candidate genes and validate the complexity of fitness contributions of a single locus. Full understanding of these complex architectures requires substantial follow-up work involving repetition of the selection experiments with fixation of different alleles to distinguish between interacting partners and independent fitness. The ability to quantify the effects of individual alleles and the narrow mapping intervals help in designing these studies and in constructing informative allele combinations.
Prospects of mapping complex traits in model organisms
The goal of using model organisms and controlled crosses for quantitative genetics research is to closely mimic either idealized population genetic models or natural populations in terms of genetic and phenotypic variation. The SGRP-4X population presented here has nearly equal contributions from all founders, uniform spatial distribution of segregating sites, little effect from drift on allele frequencies, and low linkage, making it a close model for an idealized multiparent advanced intercross. Any bias in the initial allele frequencies due to selection and associated hitchhiking of linked alleles during the intercross will not affect the usability of the population for mapping traits during pooled selection experiments so long as the initial frequencies remain substantially >0. Having a control experiment and several time points enabled us to fully factor in the initial allele frequencies, and QTL are called only if during the selection phase there is a significant change against that baseline.
Even more genetic and trait diversity can be incorporated into the mapping population by crossing more distant parents. For example, several extremely divergent, but not reproductively isolated, strains have recently been recovered in China (Wang et al. 2012). To obtain model populations closer to that of humans, other natural isolates and mosaic lab strains could be crossed in bulk to produce a population with variable allele frequencies. However, as the mapping approaches used here rely on at least moderate frequencies of causal alleles for sensitive detection, they have to be adapted to scenarios with rarer alleles.
In general, model populations remain fruitful for the study of isolated individual genetic mechanisms, be it rare variants, epistasis, or common small-effect alleles, and ideally, specialized controlled populations would be constructed for each. SGRP-4X is useful in a variety of settings, as we have shown for the study of weaker alleles and of multiple contributors at a locus and following up allelic series and interactions between loci. With rapid advances in highly multiplexed whole-genome sequencing, an even larger number of new individuals derived from the SGRP4-X could be genotyped in the future to approach nearly full power to map trait loci (Bloom et al. 2013; Ludlow et al. 2013; Wilkening et al. 2013).
We have harnessed the awesome power of baker’s yeast for quantitative genetics studies. Full-genome sequences containing a large fraction of natural variation in the species, high-resolution phenotypes, an extremely large number of individuals, and low linkage disequilibrium have produced a very powerful mapping population. This resource can be further utilized for linkage or selection mapping of more phenotypes, creating the outbred diploid lines for extremely large-scale mapping, crossed to various deletion, overexpression, and tag collections for understanding phenotypes underlying the complex trait alleles and modified by standard tools for validating and following up small-effect alleles.
Acknowledgments
We thank all the members of the Sanger Institute sequencing production teams for generating the sequence data. We thank Alex Mott and Agnes Llored for technical help. This work was funded by grants from Atip-Avenir, Association pour la Recherche Contre le Cancer (SFI20111203947) (to G.L.), and The Wellcome Trust (grant WT077192/Z/05/Z) (to R.D.). We further acknowledge the Wellcome Trust for support under grant 098051. This work was also supported in part by Region Midi Pyrénées (France) under grant 09005247 and was carried out in the frame of the European Cooperation in Science and Technology Action (FA0907-Bioflavour) under the European Union’s Seventh Framework Programme for Research (FP7). F.A.C. is supported by Conicyt–Programa de Atracción e Inserción/Concurso Nacional de Apoyo al retorno de investigadores/as desde el extranjero grant 82130010. L.P. is supported by a fellowship from the Canadian Institute for Advanced Research and a Marie Curie International Outgoing Fellowship.
Footnotes
Communicating editor: M. Johnston
- Received July 25, 2013.
- Accepted September 8, 2013.
- Copyright © 2013 by the Genetics Society of America
