Adaptation to High Ethanol Reveals Complex Evolutionary Pathways

Tolerance to high levels of ethanol is an ecologically and industrially relevant phenotype of microbes, but the molecular mechanisms underlying this complex trait remain largely unknown. Here, we use long-term experimental evolution of isogenic yeast populations of different initial ploidy to study adaptation to increasing levels of ethanol. Whole-genome sequencing of more than 30 evolved populations and over 100 adapted clones isolated throughout this two-year evolution experiment revealed how a complex interplay of de novo single nucleotide mutations, copy number variation, ploidy changes, mutator phenotypes, and clonal interference led to a significant increase in ethanol tolerance. Although the specific mutations differ between different evolved lineages, application of a novel computational pipeline, PheNetic, revealed that many mutations target functional modules involved in stress response, cell cycle regulation, DNA repair and respiration. Measuring the fitness effects of selected mutations introduced in non-evolved ethanol-sensitive cells revealed several adaptive mutations that had previously not been implicated in ethanol tolerance, including mutations in PRT1, VPS70 and MEX67. Interestingly, variation in VPS70 was recently identified as a QTL for ethanol tolerance in an industrial bio-ethanol strain. Taken together, our results show how, in contrast to adaptation to some other stresses, adaptation to a continuous complex and severe stress involves interplay of different evolutionary mechanisms. In addition, our study reveals functional modules involved in ethanol resistance and identifies several mutations that could help to improve the ethanol tolerance of industrial yeasts.


Introduction
The ability to survive and proliferate in high levels of ethanol is an ecologically important and industrially relevant trait of yeast cells. The ethanol produced by yeast cells slows down growth of competing microbes, but at higher concentrations, it causes stress for the yeast cells themselves. Different yeast strains show significant differences in their ability to grow in the presence of ethanol, with the more ethanol-tolerant ones likely having a fitness advantage over non-tolerant strains [1][2][3]. Moreover, ethanol tolerance is a key trait of industrial yeasts that often encounter very high ethanol concentrations, for example during beer and wine making and industrial bio-ethanol production.
Because of its industrial importance, there is great interest in fully understanding the genetic underpinnings of ethanol tolerance in microbes. Different experimental approaches have been used, including screening of (deletion) mutants for increased ethanol tolerance, transcriptome analysis of ethanol stressed cells and QTL analyses aimed at identifying mutations that cause differences in ethanol tolerance between different yeast strains [4][5][6][7][8][9][10]. Together, these studies have linked multiple different genetic loci to ethanol tolerance and identified hundreds of genes, involved in a multitude of cellular processes [11][12][13][14]. While it becomes increasingly clear that ethanol is a complex stress that acts on several different processes including increasing fluidity and permeability of cellular membranes, changing activity and solubility of membrane-bound and cytosolic proteins and interfering with the proton motive force (for review, see [15][16][17], the exact molecular mechanisms and genetic architecture underlying ethanol tolerance are still largely unknown. Experimental evolution to study adaptation to increased ethanol levels could provide more insight into the molecular mechanisms underlying ethanol tolerance since such experiments could reveal different mutational paths that make a sensitive strain more tolerant. Only a handful of studies have looked at adaptation to ethanol in originally non-ethanol tolerant microbes exposed to gradually increasing levels of ethanol [6,[18][19][20][21][22]. These have mostly focused on the physiological adaptations found in the evolved cells and have not performed an extensive analysis of the mechanisms and genetic changes underlying this adaptation. Hence, a comprehensive analysis of the type and number of mutations a non-ethanol tolerant strain can (or needs to) acquire to become more ethanol tolerant is still lacking. Experimental evolution has proven to be a valuable tool to investigate the different mechanisms and pathways important for cells to adapt to specific selective conditions. Seminal papers have increased our understanding of the molecular basis of adaptation to specific stresses, such as heat stress, nutrient limitation and antibiotic treatment [23][24][25][26][27][28]. Recent advances in DNA sequencing technologies allow affordable and fast sequencing of complete genomes of clones and populations. While sequencing clones yields information on individual lineages within the experiment, population data provides information on the heterogeneity of adaptation. Additionally, sequencing samples isolated at different time points during the evolution experiment makes it possible to capture evolution in action. This has provided valuable information on the rate and types of mutations underlying adaptation, the genetic basis of 'novel' phenotypes and the existence of parallel pathways to establish comparable phenotypic outcomes [29][30][31][32].
For example, a common strategy observed in populations evolving under nutrient limitation is the amplification of genetic regions encoding transporters responsible for the uptake of the limiting nutrient [24,33]. Other studies using multiple replicate populations have discovered a high degree of parallelism in the adaptive solutions found by different populations. However, in other cases, the evolutionary pathways can be more complex. For example, clonal interference, the competition between lineages carrying different beneficial mutations, is a commonly observed phenomenon in evolution of asexually propagating populations that can increase complexity of mutational dynamics as well as impede the spread of beneficial mutations in a population. Clonal interference is typically expected to be prevalent in large populations that are adapting to a complex stress, where multiple adaptive mutations can occur [32,34,35].
To unravel the molecular mechanisms of adaptation to a specific condition, most studies have used isogenic replicate populations, with all cells having the same initial genome size. Genome size can significantly change during evolution; with both small-scale changes (chromosomal deletions and amplifications) and large-scale changes (increase or decrease in ploidy). Moreover, ploidy shifts have been reported in the evolutionary history of many organisms, including Saccharomyces cerevisiae and as a response to selective pressure [36][37][38]. Conversely, genome size has also been reported to affect evolution rate: polyploidy has been suggested to increase adaptability [39,40]. Multiplying the amount of DNA increases the genetic material available for evolution to tinker with and can alter gene expression [41][42][43]. These polyploid genomes can be unstable, resulting in loss of chromosomes and thus aneuploid cells [44][45][46]. Although studies have looked at adaptation of lineages of different ploidy, none have followed the mutational dynamics in these evolving populations over time in detail [36,[47][48][49].
In this study, we use experimental evolution to dissect the adaptive mechanisms underlying ethanol tolerance in yeast. Six isogenic Saccharomyces cerevisiae populations of different ploidy (haploid, diploid and tetraploid) were asexually propagated in a turbidostat over a two-year period, with ethanol levels gradually increasing during the experiment. This step-wise increase in exogenous ethanol levels resulted in a constant selective pressure for our cells. Wholegenome sequencing of evolved populations and isolated, ethanol-tolerant clones at different times during the experiment allowed us to paint a detailed picture of the mutational dynamics in the different populations. Several common themes in the type of adaptations and evolutionary mechanisms emerge, with all lines showing extensive clonal interference as well as copy number variations. Additionally, the haploid and tetraploid lines showed rapid convergence towards a diploid state. Despite these common themes, we find multiple lineage-specific adaptations, with little overlap in the mutated genes between the different populations. By applying a novel computational pipeline to identify affected pathways, we were able to reveal overlap between the functional modules affected in the different adapted populations, with both novel and previously established pathways and genes contributing to ethanol tolerance. Importantly, introduction of specific mutated alleles present in adapted populations into the ancestral strain significantly increased its ethanol tolerance, demonstrating the adaptive nature of these mutations and the potential of using experimental evolution to unravel and improve a complex phenotype.

Evolved cells are more ethanol tolerant
To study the mutational dynamics underlying increased ethanol tolerance, six prototrophic, isogenic S. cerevisiae strains of different ploidy (two haploid, two diploid and two tetraploid lines-with each line of the same initial ploidy started from the same preculture (VK111, VK145 and VK202 respectively, see also S4 Table) were subjected to increasing levels of ethanol. Specifically, ethanol levels were gradually increased from 6% (v/v) to 12% over a two-year period in a continuous turbidostat with glucose (4% (w/v)) as a carbon source. Samples were taken at regular time intervals and subjected to whole-genome sequencing analysis. Chloramphenical was added to the medium of our long-term evolution experiment to prevent bacterial growth. Chloramphenicol does not affect yeast growth (S1 Fig). In addition to sequencing each of the 6 evolving populations, we also sequenced the genomes of three clones isolated from each of the population samples, resulting in a total of 34 population samples and 102 clonal samples that were sequenced. Fig 1 depicts the experimental set-up as well as the time points (number of generations) for which whole-genome sequencing was performed.
We determined the relative fitness of the evolved populations, isolated at 40 and 200 generations, and generally observed increases in fitness in high (9% v/v) EtOH (Fig 2 & Materials and Methods). In general, clones taken from the same population at the same time point had similar fitness; with some notable exceptions suggesting considerable heterogeneity within populations (S2 Fig). These fitness measurements were performed in 9% ethanol because the ancestral reference strains did not grow at higher ethanol levels. It is important to note that the actual increase in fitness of our evolved strains in higher ethanol levels (above 9% (v/v)) is much larger than what can be appreciated from Fig 2. Spotting of our evolved lineages on agar plates with increasing levels of ethanol indicates that these adapted strains can grow on concentrations up to 11% (v/v) EtOH, with the haploid control ancestral strain showing almost no growth under these conditions (i.e. an infinite increase in fitness compared to the ancestral haploid strain) (S3 Fig). Ancestral diploid and tetraploid strains show growth on 11% EtOH, although they are still outperformed by their evolved clones under these conditions.

Diploidization of haploid cells is a fast way of increasing ethanol tolerance
Propidium iodide staining and flow cytometry analysis of evolved populations showed that diploid cells appeared relatively quickly in the originally haploid populations (Fig 3; PI staining profiles of ancestral strains can be found in S4 Fig; ploidy of sequenced clones can be found in S1 File). This was observed independently in both reactors started from the same isogenic haploid population. It should be noted that around generation 40 in reactor 1, some diploids are already present in the originally haploid population. By generation 60, haploid cells took over the population again, possibly by acquiring a mutation that made them more fit than these new diploids. In both reactors, diploid cells have taken over the entire population by 130 generations. These diploid variants are still of mating type α, the same as the initial haploid ancestral strain, indicating that they did not arise through mating type switching and subsequent mating.
Our results indicate that the diploids in our turbidostat are likely the result of failed cytokinesis. The relatively rapid evolutionary sweep of the new diploid variants points to a significant , diploid (VK145) and tetraploid (VK202)) for increased ethanol tolerance was performed in a turbidostat. Every 25 generations, the ethanol concentration in the media was increased in a stepwise manner (starting at 6% (v/v) and reaching 12% at 200 generations). Increasing the ethanol concentration from 10% to 11% dramatically reduced growth rate of evolving cells. Therefore, instead of increasing ethanol levels, we first reduced the ethanol level to 10.7% after 100 generations. (B) Red circles represent sampling points (indicated as number of generations) for which whole-genome sequencing was performed. For each circle, heterogeneous populations as well as three evolved, ethanol tolerant clones were sequenced. Sequencing of the population sample of reactor 4 at 200 generations failed, so this data is omitted from the manuscript. For generation 80 of reactor 1, only population data is available. Experimental Evolution for Ethanol Tolerance selective advantage. Competition experiments confirmed that the fitness of a diploid cell is indeed significantly higher than that of an isogenic haploid strain in 9% EtOH (p = 0.0063; unpaired t-test), whereas there is no significant fitness difference in the absence of ethanol (p = 0.469; unpaired t-test) (S5 Fig). Interestingly, our tetraploid starting populations also converged to a diploid state relatively fast during the evolution experiment. The parallel diploidization observed in two reactors with haploid ancestral cells and two reactors with tetraploid ancestral cells might suggest that becoming diploid is one of the main and/or one of the more easily accessible routes leading to increased ethanol tolerance.

Whole-genome sequencing of evolved populations and clones reveals a complex pattern of adaptation
To identify the mutational pathways during adaptation to ethanol, we performed wholegenome sequencing of populations of evolving cells throughout the two-year experiment. Each population sample was sequenced to 500 fold coverage on average, and this for multiple time points (34 samples for all reactors combined, see Fig 1). In addition to this whole-population sequencing strategy, we also sequenced for each time point the genomes of three adapted clones to 80 fold coverage on average, resulting in 102 clonal samples sequenced in total. Details on the sequencing and analysis pipelines can be found in Materials and Methods.
Whole genome sequencing of adapted clones reveals mutators and extensive aneuploidies After 2 years, yielding around 200 generations, evolved clones (excluding clones from reactor 2 and 6, which acquired a mutator phenotype, see below) contained on average 23 SNPs compared to the ancestral strain (S1 File). This number is higher than what would be expected . Fitness is expressed relative to the ancestral strain of each reactor (haploid for reactor 1-2; diploid for reactor 3-4 and tetraploid for reactor [5][6]. Data represent the average of three independent measurements, error bars represent standard deviation. Populations with mutator phenotypes showed higher standard deviations. This might be due to the large supply of mutations generated by these mutators, which could affect mutation-selection balance.  Experimental Evolution for Ethanol Tolerance based on measured rates of spontaneous mutations [50], and could reflect an increased mutation rate under the stressful conditions imposed by the high ethanol levels in our set-up. Across all reactors and clones sequenced, we identified a total of 8932 different sites mutated. The largest fraction are SNPs (6424 out of 8932; 72%); Indels are found mostly in non-coding regions (1830 out of 2508; 73%), whereas SNPs are mostly found inside genes (4971 out of 6424; 77%). Most of these coding SNPs are non-synonymous (3672 out of 4971; 74%). A full list of all mutations (SNPs and Indels) in all sequenced clones for the different reactors and time points can be found in the S1 File.
In two reactors (reactor 2 and reactor 6), we noticed a marked increase in the number of mutations found in individual clones (see Fig 4, S1 and S2 Files). Specifically, clones from reactor 2 contain a significantly higher number of indels (about 929 on average per genome for clones isolated after 200 generations, compared to 21 on average per genome for clones isolated after 200 generations in other reactors), whereas clones from reactor 6 display an increase in the number of SNPs (about 1765 per genome for clones isolated after 200 generations, compared to an average of 31 per clone for other reactors).This high number of mutations points to a so-called "mutator" phenotype typically present in cells with lower DNA replication fidelity and/or DNA repair. Interestingly, such mutators are frequently observed during evolution experiments, probably because mutators are more likely to acquire (combinations of) adaptive mutations faster [23,[51][52][53][54].
Interestingly, the increased mutation rate in reactor 2 is first observed around generation 70-80, coinciding with the appearance of a 21 bp insertion in the MSH2 gene-a key player in DNA mismatch repair. Mutations in MSH2 have been previously reported to confer a mutator phenotype [55,56], and mutations in MutS, the ortholog of MSH2 in E. coli, were identified in a long term evolution experiment and also result in a mutator phenotype [23]. Indeed, deletion of this gene, as well as re-creating this insertion in an otherwise wild type background drastically increases mutation rate (S6 Fig). The MSH2 mutation eventually reaches a frequency of 100% at the end of the evolution experiment (200 generations). We currently do not know the exact cause of the elevated mutation rate observed in reactor 6.
Apart from the convergence towards diploidy (see above), we also detected extensive copy number variation (CNVs) in our evolved clones, comprising both duplicated and deleted chromosomal regions (see Fig 5 and S3 File). Other studies have observed similar copy number variation during adaptation to stress, including heat stress and specific nutrient limitations [24,28,33]. Interestingly, acquisition of an extra copy of chromosome III appeared to be a common feature for most of our evolved clones (S3 File). Some (parts of) other chromosomes, including chromosome XII and chromosome IV, are also duplicated in several of our evolved clones. These frequent occurrences indicate that these specific aneuploidies could be adaptive under ethanol conditions, although the exact mechanistic basis remains to be elucidated and further experiments are needed to validate the (potential) adaptive role of these observed aneuploidies.

Whole-genome sequencing of evolved populations reveals extensive clonal interference
While the sequencing of clones yielded valuable information on individual lineages within the evolving populations; sequencing of evolving populations yielded more information on the complex mutational paths and dynamics between sub-populations within each evolving population. A list of SNPs and Indels identified in the evolved populations, together with their allele frequency in the population at each sampling point, can be found in S4 File. We observe distinct pattern of mutations appearing and disappearing over time in each of the six reactors.
Some of these mutations remain in the population, eventually reaching high levels or even complete fixation (i.e. presence in 100% of all cells in the population). Other mutations only persist for a short time, until lineages carrying these mutations are outcompeted by others, socalled clonal interference. In total, we identified 1637 mutations across all populations and time points. 117 of these mutations are no longer present in the final time points sequenced, and 101 mutations drop more than 10% in frequency after reaching their maximum frequency,  indicative of clonal interference. Interestingly, we also identified some overlap between the mutations found in different independently evolving populations (i.e. in different reactors). Specifically, we find 20 genes that are mutated twice in different generations and populations, 3 genes mutated 3 times, 2 genes mutated 4 times, 2 genes mutated 5 times and 1 gene 6 times (see also S1 Table). This significantly differs from what would be expected by chance (see Materials and Methods and S1 Table for p-values, exact binomial test). Repeatedly hit genes are, amongst others, involved in stress response, cell cycle and heme biosynthesis.
The higher number of sequenced samples from reactors 1 and 2 allowed us to further analyze these population sequences and group mutations based on correlations in the changes in their respective frequencies (based on the pipeline developed by [32]; see also Materials and Genome view of yeast chromosomes and CNV patterns from a sliding window analysis. The y-axis represents log2 ratios of the coverage observed across 500bp genomic windows to the coverage expected in a diploid genome without CNVs. Area of the plot located between the red lines (from -0.23 to -1) marks putative CNV loss events, whereas region between the blue lines (from 0.23 to 0.58) marks putative CNV gain events. Data for other reactors is shown in S3 File. It should be noted that the amplified region of chromosome XII observed in some of our clones does not correspond to the ribosomal DNA genes. Methods). This yields a more detailed picture of the different co-evolving sub-populations present in these reactors, which is depicted in the Muller diagrams of Fig 6 and  In both reactors, selective sweeps mostly consists of groups of mutations that move through the population together. While these reactors were inoculated with the same strain, the type and dynamics of mutations observed during adaptation appear very different. However, both evolving populations of reactor 1 and reactor 2 are characterized by strong clonal interference. In reactor 1, 4 different subpopulations are present around generation 90, each carrying different mutations. By generation 130, these lineages have been outcompeted by another lineage that almost completely dominates the population by 200 generations (S7 Fig). In reactor 2, a lineage carrying a mutation in PDE2 (encoding a high-affinity cAMP phosphodiesterase) is driven to extinction by a subpopulation carrying indels in ASG1 and MSH2. ASG1 is a transcriptional regulator involved in the stress response and has been found mutated in evolved populations from different reactors (including reactor 1, see also Another example of clonal interference is observed in the later generations of this reactor: a subpopulation carrying a mutation in CDC27 (encoding a subunit of the anaphase promoting complex/cyclosome) is driven to extinction by a subpopulation carrying mutations in BNI1 (important for nucleation of actin filaments) and PET123 (encoding a mitochondrial ribosomal protein).

Diverse pathways involved in adaptation to ethanol
The results discussed above revealed extensive variability in the type and number of mutations present in each evolving population. While this could suggest the presence of several, different mutational pathways (and thus lack of parallel evolution), mutations in different genes might affect identical or similar pathways, implying that the physiological adaptation to high ethanol might in fact be more similar than what is immediately apparent from the individual mutations. It should also be noted that some of the mutations identified in our evolved lineages might represent adaptations to the device used, or other aspects of the selection regime, rather than ethanol per se.
To gain insight into the affected biological pathways and investigate the possible similarities in adaptation to increased ethanol, different computational approaches were used (see also S1 Text). First, a term-enrichment analysis was performed (for enriched clusters, see Table 1 and S5 File). For this analyses, we excluded mutations obtained from the populations with a mutator phenotype (reactor 2 and 6) because of their low signal to noise ratio. These enrichment methods have been used as one of the standard functional analysis tools and gave us a first insight into potential adaptive pathways present in our evolved lineages.
In a second step, we used a sub-network-based selection method [58,59] (see Materials and Methods) first developed for E. coli expression data. Here, we adapted and extended this method to select the subnetwork from the global yeast interaction network that best connected the mutated genes in the most parsimonious way. This method also identifies the intermediary genes involved in signaling mechanisms, which are not necessarily mutated in our evolved lineages but mediate the cellular response. We excluded mutations obtained from the populations with a mutator phenotype (reactor 2 and 6) because of their low signal to noise ratio. This analysis identifies genes frequently mutated in the different populations (DSK2, ASG1), as well as genes that are closely connected on the interaction graph (HEM3, HEM12,. . .). This latter set reflects parallelism at the pathway level in the different reactors. 25% of genes from inside our identified networks have been previously linked to ethanol tolerance (broadly defined as fermentation/growth capacity under conditions with EtOH, as reported in literature), whereas such as a connection was found for only 9% of genes outside of the enriched networks. From  Fig 7, it is clear that sometimes multiple mutations occurring in the same pathway originated in the same reactor rather than across different reactors. This could indicate not only a level of parallelism between the different reactors but also between the sub-populations in the same reactor. To confirm this, we also applied the sub-network selection method on the mutations obtained for each of the populations separately. This confirms that within single populations the same pathways seem to be affected as those that are consistently affected between the populations (including cell cycle, DNA repair and protoporphyrinogen metabolism, see also below). All resulting sub-networks as well as the enriched genes (including p-values) can be found in S6 File, Fig 7 and S8-S13 Figs.  Table 1. Top functional clusters identified amongst genes hit by mutations in the evolution experiment. An enrichment score cutoff value of 1.3, equivalent to a p-value of 0.05 for term enrichment was used [57].

Cluster ranking
Cluster name Enrichment score To increase mechanistic insight, we performed several additional analyses, such as interactome analyses of genes containing non-synonymous mutations in our different population samples as well as analyses of the effect of these mutations on protein functional domains (see also S1 Text, S3 Table and S7 File for more details). Together, these analyses helped us identify the different pathways that were affected in our evolved lineages. Pathways identified as affected across different reactors by these different analyses are discussed in more detail in the next paragraphs.
Cell cycle and DNA replication. Our analyses suggest that the cell cycle and DNA replication are affected in our evolved lineages (Fig 7). Ethanol has been previously reported to delay cell cycle progression, probably through depolarization of the actin cytoskeleton [60]. This could be relevant for adaptation in the sense that slower growth could protect cells from stress, perhaps by inducing genes involved in the general stress response [61][62][63].
Respiration. PheNetic analysis shows that protoporphyrinogen metabolism is affected in our evolved populations (Fig 7). Since this is important for heme biosynthesis, this could indicate that respiration is affected downstream of mutations. Non-synonymous SNPs in functional domains or modified sites are expected to have more effect on the cell than mutations outside of these regions. We find such mutations in proteins involved in regulation and in enzymes of heme biosynthesis and proteins that need heme groups for proper functioning, that are themselves involved in respiration (S3 Table and S7 File). These analyses point to a role for changes in respiration in adaptation to ethanol. Together, our results show evidence for parallel evolution, with the same processes (cell cycle and DNA replication) affected by different mutations, and also highlight the plethora of processes involved in increased ethanol tolerance in our adapted lineages.

Several high-frequency mutations increase ethanol tolerance of nonevolved strain
The high number of mutations (both SNPs and Indels) precluded an exhaustive analysis of all mutations present in our tolerant clones and populations and their effect on ethanol tolerance. One commonly used approach to investigate putative beneficial mutations is backcrossing of evolved clones to their ancestor, which does not contain any mutation. However, since each of the evolved populations proved unable to form any spores, this strategy was not accessible to us. Hence, we focused on SNPs reaching high frequencies in our 200 generation population samples. In total, nine SNPs were selected for further study ( Table 2). Several of the mutated genes belong to or are linked to the processes affected across and/or within specific reactors (heme metabolism, protein transport and cell cycle, see also Table 2).
These nine SNPs were subsequently introduced into the ancestral haploid strain. The effect of these mutations was assessed by high-throughput competition experiments [64], in 0, 4, 6 and 8% (v/v) EtOH conditions, with glucose as a carbon source (Fig 8, S14 and S15 Figs and S6 Table). Several mutants show a clear increase in fitness, with the fitness effect often depending on the concentration of ethanol in the medium. Most mutants show slightly increased fitness in medium without added ethanol, with further increases in relative fitness with increasing exogeneous ethanol concentrations. Mutants in PRT1 and MEX67 are less fit than the WT in nonethanol conditions, but show increased fitness in higher ethanol levels. MEX67 is a poly (A) RNA binding protein involved in nuclear mRNA export. PRT1 encodes the eIF3b subunit of the eukaryotic translation initiation factor 3 (eIF3). The increased ethanol tolerance associated with these mutations hints at translation processes as targets of ethanol. Interestingly, a recent study in E. coli showed that ethanol negatively impacts transcription as well as translation [6]. However, if and how mutations in MEX67 and PRT1 might mitigate this effect and contribute to ethanol tolerance remains unknown.
Of all mutations introduced into the ancestral strain, vps70 C595A provided the largest fitness increase. VPS70 is putatively involved in sorting of vacuolar carboxypeptidase Y to the vacuole [65]. A mutation in VPS70 has been recently identified by members of our team as a determinant of ethanol tolerance in a Brazilian bioethanol strain [66]. Interestingly, the VPS70 mutation in our evolved ethanol-tolerant lineages alters the same amino acid as the mutation present in the industrial ethanol tolerant strain (Goovaerts A. and Thevelein J.M., personal communication). The fact that one of the mutations identified in our evolved lineages was also found in an industrially used bio-ethanol strain underscores the potential of our approach to find biologically relevant mutations for increased ethanol tolerance. Other members of the VPS family have been previously implicated in ethanol tolerance as well, but also here the exact molecular mechanism through which they could increase ethanol tolerance is still unclear [10,12,67]. To our knowledge, none of the other genes investigated in this study have been previously implicated in ethanol tolerance nor have these mutations been found in natural or industrial yeasts so far. This implies that these mutations could be prime candidates to improve the ethanol tolerance and production of existing industrial yeasts [68].

Discussion
Many fundamental questions on the dynamics and genetics of adaptation to a complex and severe stress such as ethanol are still unanswered. Which processes contribute to ethanol tolerance? Which type and number of mutations are needed and/or sufficient, and how do they penetrate and fix within populations? To what extent are these mutations and the pathways they affect predictable? To address these questions, we performed high-coverage wholegenome sequencing of clones and populations isolated throughout a two-year evolution experiment and characterized their adaptation to increasing ethanol levels. This allowed us to paint a detailed picture of the mutational dynamics in our evolving lineages.
Our study demonstrates how many different evolutionary mechanisms all come together to provide adaptation to a severe and complex stress. Specifically, we find that adaptation to high ethanol levels involves changes in ploidy, copy number variation and the appearance of mutator phenotypes, with evolving populations showing strong clonal interference. Although these mechanisms have been observed in other studies (see, for example, [24,27,32,[69][70][71], different mechanisms are often observed and/or studied separately. More importantly, in traditional evolution studies, populations are exposed to a fixed concentration of antibiotic or limiting nutrient. Under these conditions, selection pressure is reduced or even eliminated when cells become resistant. In our experimental set-up, ethanol levels were stepwise increased over time so that cells were evolving under constant selection. High concentrations of ethanol also actively kill non-tolerant cells, so that cells not adapted to the higher ethanol concentrations die and/or are washed out of the turbidostat. Taken together, our study combines aspects of traditional evolution studies with principles used in morbidostat experiments [27]. This set-up is likely a better representation of what happens when a population gradually penetrates and adapts to a new niche. Under such severe and continuous stress conditions (near-morbidostats), many cells fail to survive. This also implies that the traditional way of measuring evolutionary time by counting the number of generations may become problematic as cell death and mutations contributing to survival instead of replication become increasingly important. This phenomenon can also result in extremely quick sweeps, because a mutation cannot only penetrate a population because it increases the replication rate, but also because it protects against death. Moreover, as cell replication slows down, the number of mutations that is not associated with DNA replication but rather with DNA damage and repair likely increases. It seems plausible that these mechanisms contribute to the very quick genetic sweeps that we observe in some reactors, most notably the quick sweep of diploidized cells in reactors 1 and 2. This could also explain why the number of mutations that we observe in our two-year experiment might seem high for only 200 generations when compared to experiments where cells are growing at higher rates and where 200 generations are often reached within only one or two months of evolution. While it is difficult to think of a better universal measure of evolutionary time, it is important to realize that in some cases, a combination of physical time and generations may be a more appropriate way to assess evolution. We find extensive variation in genome size in our evolved cells; with aneuploidies in a large number of our evolved clones as well as convergence toward a diploid state in our initially  Table. doi:10.1371/journal.pgen.1005635.g008 haploid and tetraploid populations. Becoming diploid thus appears to be a frequently used strategy by cells of different ploidy which effectively increases ethanol tolerance. Convergence to diploidy has been observed in other laboratory evolution experiments [25,36,72]. In contrast to these studies, we could demonstrate a clear fitness advantage of our diploid strains in ethanol. Although further research is needed to fully understand this fitness advantage, changes in gene expression and/or cell size could be important factors contributing to the increased fitness of diploid cells [41,73,74]. Unfortunately, this convergent evolution prevented us from performing a detailed analysis on the difference in adaptive strategies employed by haploids, diploids and tetraploids to increase ethanol tolerance.
Aneuploidy and copy-number variation are increasingly recognized as common themes in rapid adaptation [24,28,69,70]. It is believed that changes in the copy number of chromosomes or chromosomal fragments provide a relatively easily accessible way to change expression levels of specific key genes [75,76], and these CNVs can provide large, usually condition-specific, fitness effects [77]. However, such large-scale changes in the genome likely also have some unwanted, detrimental side effects, such as imbalance between gene products and genome destabilization [78][79][80]. Evolving populations are believed to gradually replace adaptive CNVs with more specific mutations that show fewer pleiotropic effects [28]. Notably, several of our evolved clones, isolated from different reactors, carry an extra copy of chromosome III and/or ChrXII; pointing towards a potential adaptive benefit of this specific aneuploidy. Interestingly, we find that clones isolated at later time points have a specific, smaller region of chromosome XII amplified (see also S3 File); indicating a more refined solution. GO enrichment and network analyses of the repeatedly amplified region of ChrXII (position 657500-818000) hints at cell wall formation as one of the key processes affected by these amplifications. Previous studies have indeed shown that cell wall stability is a key factor involved in ethanol tolerance [18].
Apart from diploidization of our evolving lineages, another example of parallelism at the phenotypic level is the appearance of a mutator phenotype in two of our six evolving populations. The sweep of the MSH2 mutation in reactor 2 is likely caused by a so-called hitchhiking event, with the high mutation rates in the MSH2 mutant leading to the appearance of one or several beneficial mutations that drive the selective sweep.
Parallelism at the genotypic level is less clear: we find few mutations and mutated genes shared between the different evolved lineages. Interestingly, nine (ACE2, POL3, PUF4, GFA1, UTH1, JID1, RIM15, ATG11 and VPS74) out of the 28 genes (32%) that were found hit in multiple reactors have been previously implicated in ethanol tolerance. Applying different types of network and enrichment analyses revealed functional modules affected in several of the adapted populations. These pathways include response to stress, intracellular signal transduction, cell cycle and pathways related to membrane composition and organization (such as isoprenoid metabolism, glycerophospholipid catabolism and fatty-acyl-coA metabolism). For some of these pathways, further work is needed to clarify their exact involvement in ethanol tolerance.
To investigate the phenotypic effect of mutations present in our evolved lineages, we performed high-throughput fitness measurements in different ethanol concentrations. While our evolved lineages contained multiple mutations, single mutations reaching high frequency in the evolved populations could already significantly increase ethanol tolerance when introduced into the ancestral, non-evolved haploid strain. Moreover, several of these single mutations selected for further study also showed a (modest) fitness benefit in conditions with no ethanol, with greater increases as ethanol levels rise. The number of mutations identified in this study was too large to investigate the fitness effect of each individual mutation. However, our strategy to select mutations that reached fixation in the evolved populations clearly proved successful, identifying as many as 4 mutations (out of 9 tested) that confer a fitness advantage in high ethanol environments. Mutations in genes linked to processes identified as affected across our different reactors-such as protein transport (VPS70)-indeed increased fitness in EtOH. This underscores the potential of PheNetic, a sub-network based selection method for identifying adaptive mutations. Five of the mutations tested did not significantly increase fitness, although they reached high frequency in our adapted populations. This is indicative of hitch-hiking with other, beneficial mutations; or possible epistatic interactions with other mutations.
Why then would not all feral yeasts show high ethanol tolerance, if it appears so easy to attain? Firstly, it seems plausible that not all yeasts are confronted with selection for high ethanol tolerance. Furthermore, it is important to note that we have not tested the fitness of the mutants under many different conditions that mimic the natural habitats of yeasts. It seems likely that some of the mutations identified in this study would result in lower fitness in other environments [81,82]. Moreover, we have not investigated the effect of combined mutations. While it is possible that combining different mutations could increase ethanol tolerance even further, it also seems likely that some mutations and/or specific combinations of mutations could lead to reduced fitness in different low and/or high ethanol environments. Indeed, while our clones have increased fitness in EtOH, we observe that fitness of several of our evolved clones (containing multiple other mutations apart from the ones investigated in this study) decreased in medium without exogenous EtOH (see S16 Fig). These results are indicative of antagonistic pleiotropy: the specific mutations present in our evolved clones increase fitness in one condition (high ethanol, which was selected for), whereas they reduce fitness in other environments.
Ethanol resistance is an important trait for the survival of feral yeasts in nature because the ethanol produced inhibits growth of competing microorganisms, while it serves as a carbon source in later stages of growth, when all fermentable sugars are depleted (the so-called makeaccumulate-consume strategy [83][84][85]). Our results suggest that adaptation to high ethanol is complex and can be reached through different mutational pathways. Apart from yielding insight into the evolutionary mechanisms leading to such complex and ecologically important phenotypes, our study is also of considerable industrial importance. Several of the mutations identified in this study may be useful to increase the ethanol tolerance of industrial strains used for the production of alcoholic beverages or biofuels.

Strains used in this study
Starting strains for the evolution experiment are all derived from the haploid prototrophic S288c strain FY5 [86]. To prevent clumping of cells during the evolution experiment, the flocculation genes FLO1, FLO10 and FLO11 were deleted in this strain using deletion cassettes based on pUG6, conferring resistance to G-418 disulfate [87]. Markers were removed through the Cre/LoxP technique using pSH65 [88]. Mating type switching of this strain was then performed, using plasmid pSB283, to create isogenic diploid and tetraploid strains. Fluorescent versions of strains (YECitrine or mCherry tagged) were constructed by integrating fluorescent markers at an intergenic, neutral region of chromosome II. A full list of strains used, with their complete genotype, can be found in S4 Table. Primers used for strain construction and verification can be found in S5 Table. Long-term selection Populations were founded in 400 mL ethanol containing media. Media contained 10 g/L yeast extract, 20g/L bactopeptone, 4% (w/v) glucose, 0.001% (v/v) Rhodorsil, Antifoam Silicone 46R, chloramphenicol (50 μg/mL) and increasing concentrations of ethanol. Populations were maintained at an average population size of 10 10 cells. After 25 generations, the level of EtOH in the media was increased each time (starting at 6% (v/v) and reaching 12% at 200 generations).
Turbidostat cultures were maintained using Sixfors reactors (Infors) at 30°C, pH was kept constant at 5.0 with continuous mixing at 250 rpm in aerobic conditions. At regular times, a population sample was obtained from each of the cultures for further analyses and stored in glycerol at -80°C. For DNA extraction purposes, a population cell pellet was also frozen down at -80°C.

Fitness determination
Fitness for all evolved strains was determined in rich medium (YP, 2% (w/v) glucose) with 9% (v/v) ethanol, by competing strains against a YECitrine labeled ancestral strain. Cultures were pre-grown in YPD 6% ethanol. After 12h, wells of a 96 deep well plate were inoculated with equal numbers of labeled reference and unlabeled strains (~10 6 cells of each) and allowed to grow for around 10 generations. Outer wells only contained medium and acted as a buffer to prevent ethanol evaporation. Additionally, plates were closed with an adhesive seal and plastic lid, and parafilm was used to prevent ethanol evaporation. Cultures were regularly transferred to new medium to prevent nutrient depletion. The ratio of the two competitors was quantified at the initial and final time points by flow cytometry. Data analysis was done in FlowJo version 10. Measurements were corrected for the small percentage of labeled, non-fluorescent cells that occurred even when the reference strain was cultured separately as well as for the cost of YFP expression in the labeled reference strain. For each fitness measurement, three independent replicates were performed. The selective advantage, s, of each strain was calculated as s = (ln (U f /R f )-ln (U i /R i ))/T where U and R are the numbers of unlabeled and reference strain respectively, the subscripts refer to final and initial populations and T is the number of generations that reference cells have proliferated during the competition. The fitness of the unlabeled WT strain was designated 1, fitness of the evolved strains as 1+s.

Determination of cell ploidy
DNA content of evolved populations and evolved clones was determined by staining cells with propidiumiodide (PI) and analyzing 50 000 cells by flow cytometry on a BD Influx. The ancestral haploid and diploid strains used in the evolution experiment were used for calibration.

Whole genome sequencing
For evolved populations, genomic DNA was directly extracted from pellets that were frozen at the time of sample taking.
Evolved clones were selected from the different population samples by streaking glycerol stocks from the corresponding population samples on YPD plates. Swabs from each population were subsequently grown in YPD 6% EtOH and dilutions were plated on YPD plates with different ethanol concentrations (ranging from 8% to 10%; with a 0.5% stepwise increase in ethanol concentrations). From these plates, ethanol tolerant clones were selected and genomic DNA of these clones was extracted.
Genomic DNA was prepared using the Qiagen genomic tip kit. Final DNA concentrations were measured using Qubit. Paired-end sequencing libraries with a mean insert size of 500 bp were prepared and libraries were run on an Illumina HiSeq2000 (EMBL GeneCore facility, Heidelberg). Average sequencing coverage for clone and population sample is 80X and 500X respectively.
Variant calling, filtration and annotation. Adaptors removal and reads quality control were done by Trim Galore! (http://www.bioinformatics.babraham.ac.uk/projects/trim_galore/) with options (-q 30-length 50). The clean reads were then mapped reference S. cerevisiae genome (S288C, version genebank64) using Burrows Wheeler Alignment allowing maximum 50bp gaps [89]. To identify mutations we employed the BROAD Institute Genome Analysis Toolkit (GATK, version 3.1) [90]. We began by performing local realignment of reads around Indels, in order to eliminate false positives due to misalignment of reads, which was followed by a base recalibration step. We then performed SNP and Indel calling according to the GATK best practice recommendations. Population samples and single clone datasets were analyzed independently.
For population data, Indel calling was performed using the UnifiedGenotyper tool, and the identified variants were filtered based on the recommended criteria (QD < 2.0, FS > 200.0, ReadPosRankSum < -20 and InbreedingCoef < -0.8). For the SNPs, we used the UnifiedGenotyper to perform multi-sample SNP calling on all the population samples together. The resulting multi-sample callset was then used as a prior to call SNPs in all individual samples. The called SNPs were then subject to filtering (QD < 2.0, FS > 60, ReadPosRankSum < -8.0, MQ < 40, MQRankSum < -12.5). Furthermore, we filtered out SNPs that were called inside regions containing Indels. We also generated a list of all known repetitive regions and low complexity regions using the program RepeatMasker (http://www.repeatmasker.org/cgi-bin/ WEBRepeatMasker), which were masked out from the assemblies. Finally, variants present in the ancestral strain were filtered out from all samples.
A similar approach was used to analyze the clonal samples. Ploidy level of each isolated clone was determined by PI staining and variant calling was performed with UnifiedGenotyper for haploid clones or HaplotypeCaller for diploid clones. Subsequently, SNPs and Indels were filtered based on the GATK best practice recommendations as mentioned above. The inbreeding coefficient filter was excluded in the clonal sample filtering for Indels, as it is a populationlevel metric. Annotation and effect prediction of all identified variants was performed using snpEff [91].
Final processing of the data was performed using a R script, which involved filtering out variants called within the sub-telomeric regions (15 kbps from the chromosome ends) and-for population samples-variants whose frequency did not change by more than 10% during the experiment. We used the infrastructure of the VSC-Flemish Supercomputer Center for these analyses. For improved performance, CPU multithreading (-nct) capabilities of GATK were utilized, providing up to 20 cores and maximum memory capacity of 64 GB per process.
Identification of Copy Number Variations (CNV). CNVs in the samples were identified using Nexus Copy Number software, version 7.5 (http://www.biodiscovery.com/software/ nexus-copy-number). Reads were accumulated in 500-base bins along all chromosomes, rejecting bins with less than 10 reads. Log2 ratios of copy numbers were then estimated from the read-depth counts in these intervals. We used the following calling parameters: minimum number of probes per segment of 5, significance threshold of p-value = 1 × 10^-9, percentage of removed outliers of 5% and the limits for copy gain and loss of+0.25 and -0.25, respectively.

Statistical analyses of genes hit multiple times
The p-values probabilities of genes being hit by a mutation a specific number of times were calculated using the binomial distribution. The number of draws was set to the total number of mutations found inside coding regions (i.e. 817); the number of successes was set to the number of times a specific gene was hit by a mutation (2 to 6); the probability of success was set to the specific ratio of the length of each gene hit multiple times (in nucleotides) over the entire coding content of the S. cerevisiae genome (9080922 nucleotides).

Haplotype reconstruction
Haplotype reconstruction was based on the approach described by [32]. Prior to the actual reconstruction, variants identified in the sequencing data were subject to further processing by excluding variants that were multi-allelic, variants that exhibited mixed zygosity in the isolated clones (i.e. homozygous in one clone and heterozygous in another clone) and variants that didn't reach the frequency of 0.2 at any point during the experiment. Haplotypes (or 'mutational cohorts') present in the remaining variants were next reconstructed using a Matlab script (kindly provided by the Desai lab) on a local machine running Matlab 2013a. Briefly, variants present in our dataset were clustered into haplotypes based on the Euclidean distance between their frequencies at specific time-points of the experiment. Afterward, frequencies of individual variants assigned to a specific haplotype were averaged to obtain the frequency of the haplotype itself. Frequencies of the identified haplotypes at specific time points (see S2 Table) were then used as source data to draw their approximate Muller diagram representations with Inkscape 0.48.4.

Enrichment and network analysis
Functionally meaningful terms enriched in the list of genes hit in our evolution experiment were identified using DAVID Tools [57,92]. An enrichment score cutoff value of 1.3 was used, equivalent to a p-value of 0.05 for term enrichment, as recommended by the authors [57].

Phenetic analyses
Intergenic mutations were discarded for the analysis. The interaction network used as input for PheNetic was composed of interactomics data obtained from KEGG [93] for metabolic interactions, String for protein-protein interactions [94] and Yeastract for protein-DNA interactions [95]. The total interaction network contains 6592 genes and 135266 interactions. This interaction network was converted to a probabilistic network using the distribution of the out-degrees of the terminal nodes of the network edges. By doing so, edges connecting nodes with a low out-degree will receive a high probability while edges connecting nodes with a high out-degree receive a low probability. Using lists of mutated genes as input, PheNetic will now infer that sub-network of the probabilistic network that best connects the mutated genes in the list over the probabilistic network. As the probabilistic network penalizes hub nodes, the inferred subnetwork will therefore preferentially connect the mutated genes through the least connected parts of the network. This results in selecting the most specific parts of the network that can be associated with the mutated genes.
PheNetic was used to connect the mutated genes over the interaction network [58] with the following parameters: 100-best paths with a maximum path length of 4 were sampled between the different mutations in combination with a search tree cutoff of 0.01. As the size of the selected sub-network by PheNetic is dependent on both a cost parameter and the number of mutated genes in the input, different costs were used for the sub-network inference from different sizes of mutated gene lists. For the sub-network inference between all the mutated genes from the non-mutator reactors a cost of 0.25 was used, for the non-mutator reactors (1,3,4,5) a cost of 0.05 was used as they all have a similar amount of mutated genes, and for the mutator reactors (2 and 6) a cost of 0.5 was used. For more details, see S1 Text.

Network visualization and enrichment
The resulting networks were visualized using Cytoscape and a functional enrichment using the biological process terms of Gene Ontology in combination with the annotation of SGD of the sub-networks was performed using the Bingo plugin, version 2.44 [96]. The resulting enrichment results are listed in S6 File.

Construction of mutant strains
Selected mutations identified from the whole-genome sequencing data were introduced into the ancestral genetic background using the following protocol. First, a selectable marker conferring resistance to hygromycine was introduced near the genomic location of interest through homologous recombination. Part of this locus was then amplified together with the selectable marker using a forward primer containing the desired mutation. The resulting PCR product was then transformed into the ancestral strain; and presence of the mutation was verified by Sanger sequencing. Primers used can be found in S5 Table. High-throughput competitive fitness measurements YECitrine-or mCherry-tagged site-directed mutant strains were competed with the parental mCherry or YECitrine tagged strains, respectively, as described [64]. In brief, saturated cultures of mutant and parental strains were mixed in equivalent volumes and inoculated onto 150 μl of YNB-low fluorescent medium in 96-well microtiter plates (Corning 3585). Micro-cultures grew without shaking and were serial-diluted every 24 hrs for approximately 28 generations (7 days) in a fully automated robotic system (Tecan Freedom EVO200) that integrates a plate carrousel (Liconic STX110), a plate reader (Tecan Infinite M1000), a 96-channel pipetting head, an orbital plate shaker, and a robotic manipulator arm. The equipment was maintained in an environmental room at constant temperature (30°C) and relative humidity (70%). Fluorescence signal (mCherry: Ex 587 nm/5 nm and Em 610 nm/5 nm; YECitrine: Ex 514 nm/5 nm and Em 529 nm/5 nm) and absorbance at 600 nm were monitored every hour during the entire experiment. The YECitrine-or mCherry-tagged parental strains were competed to each other for normalization and monitored individually to determine background fluorescence signal. Fluorescence and absorbance output data was analyzed in Matlab as described [64] to obtain an average selection coefficient, s mut , with its S.E.M. from three experimental replicates.

Accession numbers
All sequencing data are available from the NCBI Sequence Read Archive: Bioproject, accession number PRJNA292495. Mutations (reaching a frequency of at least 20%) and corresponding frequencies were identified from population sequencing data. Muller diagram represents the hierarchical clustering of these mutations, with each color block representing a specific group of linked mutations. Indels are designated with I , whereas heterozygous mutations are in italics. Mutations present as heterozygous mutations in all clones of a specific time point and present at a frequency of 50% in the population, are depicted as a frequency of 100%. Frequencies of haplotypes can be found in S2 Table. (TIF)  Table. Interactome analyses. Specific pathways hit in the different evolved populations, with FDR values. Mutated genes are the number of genes of this specific pathway mutated in the specific reactor, genes present in genome is the total number of genes in the genome for a particular pathway. As reactors 2 and 6 are mutators for Indels and SNPs respectively, we analyzed enrichments separately for SNPs and Indels to be able to compare between populations. (DOCX) S4