Natural Genetic Variation Can Independently Tune the Induced Fraction and Induction Level of a Bimodal Signaling Response

Abstract

Bimodal responses, where cells in an isogenic population adopt one of two distinct states, are particularly important for microbes coping with fluctuating environments [4,5] and cells of multicellular organisms differentiating into discrete types [6,7].The galactose-utilization (GAL) pathway in Saccharomyces cerevisiae (budding yeast) is a well-characterized bimodal response and a classic model of microbial decision-making [8,9].GAL enzymes are tightly repressed in glucose and activated almost 1000-fold in galactose [10].In mixtures of glucose and galactose, GAL genes induce as a function of the galactose-to-glucose ratio [11] and display complex patterns of bimodal expression [12].Bimodality of GAL gene expression is attributed to bistability arising from positive feedback through the Gal1p kinase and Gal3p transducer [13,14].However, perturbing other pathway components such as Gal2p permease, Gal4p activator, and Gal80p repressor also affect quantitative features of the GAL response [14][15][16][17].Additionally, the modality of the GAL response is affected by the metabolic conditions prior to encountering galactose [12].Despite the complex response of GAL expression distributions to genetic and environmental perturbations, most studies of the pathway have focused on one quantitative feature such as the induced fraction [16,18,19], with a few recent exceptions [13,20].How multiple quantitative features of the pathway are controlled and vary across perturbations is poorly understood.
In previous work, we found that natural yeast isolates differed widely in the inducibility of GAL genes in glucose + galactose mixtures [19,21].In particular, some strains displayed bimodal activation of GAL genes while other strains were unimodal in the same conditions.Similar .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint population heterogeneity has been seen in yeast maltose utilization [22] and bacterial utilization of various sugar mixtures [23].This natural variation provides an opportunity to dissect the genetic variants modulating bimodality in nature and expand our knowledge of the repertoire of quantitative behaviors that can be achieved by this model circuit.
In this work, we showed that natural yeast isolates induce the GAL pathway with a diverse array of bimodal and unimodal expression patterns that vary with sugar conditions.We analyzed this variation in terms of two traits representing the induced fraction of cells and the expression level the induced subpopulation, which vary in an uncorrelated way across natural isolates.Using bulk segregant analysis and CRISPR/cas9 allele replacement, we identified genetic variants underlying these two traits and showed that the variants can affect the traits independently of each other.Additionally, we found that the metabolic history of cells before inducing GAL genes also affects the bimodal response in a trait-specific way.The independent tuning of these two quantitative features of the GAL response can account for the diversity of unimodal and bimodal phenotypes observed in our natural isolates.This genetic flexibility may be advantageous for cells adapting to complex natural nutrient environments.

Natural yeast isolates vary in the degree of bimodality of GAL induction
To study natural variation in the population behavior of the GAL pathway response, we measured the expression of a GAL1 promoter driving YFP (GAL1pr-YFP) in 34 geographically and ecologically diverse yeast strains [21,24,25] grown in a titration of glucose plus a constant level of galactose.As expected, we found that all strains are uninduced in high glucose and fully induced in low or no glucose (Figure 1).However, at intermediate glucose concentrations, some  1B) while other strains display a bimodal mixture of uninduced and partially induced cells (Figure 1C).Additionally, strains with the same modality still have quantitatively different GAL induction profiles (Figure S1), raising the question of what mechanisms can give rise to these diverse signaling phenotypes.

Variation in GAL bimodality phenotypes can be parameterized by two uncorrelated metrics
Upon close inspection, the GAL induction phenotypes generally seem to be a mixture of two components: an induced subpopulation that decreases in YFP level as glucose increases, and an uninduced subpopulation that remains at the same YFP level regardless of glucose concentration (Figure S2A).The mixing of these components can be quantified as an induced fraction that decreases as a function of glucose concentration (Figure S2B).Simply by varying the glucosedependence of the induced fraction and of the induced subpopulation expression level, we can simulate many bimodal phenotypes, as well as unimodal phenotypes, reminiscent of the observed data (Figure S1, S4.2B-C).In this framework, a strain which is unimodal in a particular condition has an induced fraction of one (but a sub-maximal induced level), while a strain that is bimodal in this condition has an induced fraction of less than one.
Applying this population decomposition framework to our data, we computationally separated induced and uninduced cells from each GAL reporter distribution (Figure 2E) and calculated two summary metrics for each strain's phenotype: ‫ܧ‬ ଵ , the glucose concentration where the induced subpopulation reaches 10% of its maximal GAL expression level, and ‫ܨ‬ ହ , the glucose concentration where 50% of cells in the population are induced (Figure 2F).For convenience, these metrics are in log2-transformed units, so a strain with ‫ܧ‬ ଵ ൌ െ1 has an induced .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint subpopulation that reaches 10% of maximal induction at 2 ିଵ ൌ 0.5% w/v glucose.We find that the ‫ܧ‬ ଵ and ‫ܨ‬ ହ are uncorrelated across natural isolates, suggesting the possible existence of genetic changes that can decouple them (Figure 2G).

Bulk segregant analysis identifies genetic loci associated with GAL induction variation
To analyze the genetic basis of ‫ܧ‬ ଵ and ‫ܨ‬ ହ , we crossed strains S288C and DBVPG1106 and phenotyped random haploid segregants from their hybrid.These parent strains differ in both traits, and their segregants display semi-continuous, correlated variation in these traits with a small number of outliers.Therefore, ‫ܧ‬ ଵ and ‫ܨ‬ ହ are likely modulated by multiple genes, at least some of which affect both traits.
To identify these genes, we performed bulk-segregant linkage mapping using a pooled sorting strategy (Figure 2B).We chose a glucose+galactose condition where the parental GAL1pr-YFP distributions were maximally different and used it to induce a pooled mixture of haploid (MATa) segregants (Figure 2C).We then used FACS to sort the segregants into pools of uninduced ("OFF"), induced and low-expression ("LOW"), and induced and high-expression ("HI") cells (Figure 2C), and sequenced each pool to 15-33x coverage.We expected that a genomic locus affecting the induced level (and thus ‫ܧ‬ ଵ ) will differ in allele frequency between the LOW and HI pools, while any locus affecting the induced fraction (and thus ‫ܨ‬ ହ ) would differ in parental allele frequency between the OFF pool and a computationally merged LOW+HI pool ("ON") (Materials and Methods).
We found 5 loci with significantly different allele frequencies between OFF/ON pools or between LOW/HI pools, defined as genomic regions with a peak log-odds (LOD) score > 5 calculated by MULTIPOOL [26] (Figure 3D; Materials and Methods).To look for causal .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint variants, we inspected gene annotations in a region of 2-LOD decrease around each LOD peak. The three most significant loci are centered at chrIV:457Kb, chrXIV:457Kb, and chrXVI:81Kb and contain the genes GAL3, MKT1, and GAL4, respectively (Figure 3D).GAL3 and GAL4 are direct regulators of galactose sensing, while MKT1 is known to have pleiotropic effects in crosses between S288C and natural isolates .The GAL3-associated locus was significant only in the OFF/ON comparison, while the GAL4-associated locus was only significant in the LOW/HI comparison, suggesting that the effect of these loci are specific to either ‫ܧ‬ ଵ or ‫ܨ‬ ହ .The MKT1associated locus was significant in both comparisons but had a higher LOD score in the LOW/HI than in the OFF/ON comparison.Unlike the other loci, the GAL4-associated locus was enriched for the S288C allele in the DBVPG1106-like segregant pool, suggesting the possibility of transgressive segregation.Two other loci, at chrXII:1053Kb and chrXIII:105Kb, were also significant and seemed to have phenotype-specific effects, but did not contain any obvious genes for follow-up.Therefore, we focused on the chrIV (GAL3-associated), chrXIV (MKT1associated), and chrXVI (GAL4-associated) loci for further investigation.

GAL3 and GAL4 alleles specifically affect and while MKT1 alleles affect both traits
To test if GAL3, MKT1, and GAL4 alleles are causal variants in the chrIV, chrXIV, and chrXVI loci, we used CRISPR/cas9 to replace the coding region and flanking regions of each gene (Materials and Methods) in both DBVPG1106 and S288C with the allele from the other parent (Figure S4).In DBVPG1106, replacing the endogenous GAL3 allele with GAL3 S288C shifted ‫ܨ‬ ହ in the direction of the S288C parent (Figure S4A).MKT1 replacement also shifted ‫ܨ‬ ହ and had a small effect on ‫ܧ‬ ଵ as well.GAL4 replacement had a small but clear effect on ‫ܧ‬ ଵ and no detectable effect on ‫ܨ‬ ହ .In the S288C background, allele replacements had similar traitspecificity but much smaller effects (Figure S4B-C).In both parental backgrounds, the GAL4 .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint allele replacement resulted in a change in ‫ܧ‬ ଵ away from the value of the other parent.This is consistent with the sign of allele-frequency differences of the GAL4-containing locus between LOW and HI pools (Figure 2D) and confirms that GAL4 is a transgressive allele in this cross.
Overall, these results show that GAL3, MKT1, and GAL4 are causal variants in their respective loci and corroborate the allelic effects inferred from our bulk segregant analysis.
The single allele replacements only modestly altered the phenotype of the parent strains, suggesting that other genes make substantial contributions to the total phenotypic difference.
Alternatively, there may be genetic interactions between our mapped genes such that allele replacement of 2 or 3 of them is sufficient to achieve conversion of one parental phenotype to the other.To assess these possibilities, we constructed all 16 combinations of strain background, GAL3 allele, MKT1 allele, and GAL4 allele from either the DBVPG1106 or S288C parent, and measured ‫ܧ‬ ଵ and ‫ܨ‬ ହ of 2 independent isolates of each of the 16 genotypes.We examined the resulting phenotypic landscape (Figure 3A) in terms of pairs of strains differing in the allelic status of one gene (or strain background) while other genetic factors are held constant.The effect of switching from the the DBVPG1106 genetic variant to the S288C variant can be visualized as a vector in ‫ܧ‬ ଵ versus ‫ܨ‬ ହ space (Figure 3B-E) or as a trait difference (Figure 3F).This analysis reveals that the trait-specificity of single genetic changes are broadly consistent across different genotypic backgrounds (i.e.combinations of strain background and alleles at the other loci).This can be seen in the fact that effect vectors from DBVPG1106 to S288C variants in ‫ܧ‬ ଵ ‫ܨ-‬ ହ space are parallel (Figure 3B-E), or equivalently, that differential effects cluster by angle to the origin (Figure 3F).Across the combinatorial allele replacement strains, it is clear that GAL3 allele predominantly affects ‫ܨ‬ ହ , GAL4 mostly affects ‫ܧ‬ ଵ , and MKT1 affects both traits.For example, a strain with GAL4 S288C has a lower ‫ܧ‬ ଵ than the congenic strain with .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint GAL4 DBVPG1106 for all such strain pairs.These results show that ‫ܧ‬ ଵ and ‫ܨ‬ ହ can be tuned independently in this cross. In contrast to trait-specificity, the effect sizes of single genetic changes across the combinatorial landscape are more complicated.In general, GAL3 allele effects on ‫ܨ‬ ହ are large, spanning up to half the phenotypic distance between the parents.MKT1 allele effects on ‫ܨ‬ ହ are almost as great, and combined with GAL3 allele replacement, can essentially phenoconvert DBVPG1106 to S288C, but only along the ‫ܨ‬ ହ axis (DSSD versus in DDDD in Figure 3A).However, the reciprocal replacement in the S288C background has a more modest effect (SDDS versus SSSS in Figure 3A).Consistent with these findings, the strain background effect on ‫ܨ‬ ହ (which can be interpreted as the residual variation after allele replacements) varies widely, from negligible to almost as large as that of GAL3 or MKT1.For the ‫ܧ‬ ଵ trait, GAL4 allele effects span between a third and half the phenotypic distance between the parents, but in the opposite direction required for phenoconversion.Therefore, triple allele replacement strains DSSS (DBVPG1106 GAL3 S288C MKT1 S288C GAL4 S288C ) and SSSD still differ substantially in ‫ܧ‬ ଵ from their respective wildtype SSSS and DDDD strains.Overall, strain background has effects on both ‫ܧ‬ ଵ and ‫ܨ‬ ହ in most genotype backgrounds, indicating substantial variation in both traits not accessed by our allele swaps.

Genetic and environmental perturbations that affect do not affect
Our results above show that ‫ܨ‬ ହ can be tuned independently of ‫ܧ‬ ଵ by some genetic variants in the S288C x DBVPG1106 cross.To see if this is true over a larger range of ‫ܨ‬ ହ , we analyzed phenotypic data on S288C, BC187, and DBVPG1106 strains whose GAL3 loci have been replaced with a panel of natural GAL3 alleles that we previously showed to underlie a spectrum .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint of GAL inducibility phenotypes [19].As expected, ‫ܨ‬ ହ varied widely as the GAL3 allele is changed (Figure 4A).However, variation in ‫ܧ‬ ଵ with GAL3 allele was minimal and driven almost entirely by the strain background (Figure 4B).
Previously, a laboratory yeast strain was found to induce GAL genes bimodally or unimodally depending on the carbon source prior to galactose induction [12].To see how metabolic memory affects ‫ܧ‬ ଵ and ‫ܨ‬ ହ across our natural isolates, we pre-grew six strains in raffinose, acetate, or glycerol prior to induction in glucose + galactose.These carbon sources are neither inducers of GAL genes nor signals for glucose catabolite repression [27].Nevertheless, they caused the yeast strains to exhibit different ‫ܨ‬ ହ upon subsequent induction in glucose + galactose (Figure 4C).We did not observe a carbon source pre-conditioning effect on ‫ܧ‬ ଵ (Figure 4D).This parallels the effect of GAL3 alleles, and suggests that the independent tuning of ‫ܧ‬ ଵ and ‫ܨ‬ ହ is a consequence of how the GAL circuit is integrated with carbon metabolism more broadly.

Independent tuning of and modulates the modality of the GAL response
The definitions of the ‫ܧ‬ ଵ and ‫ܨ‬ ହ metrics (Figure 1F, S4.2) imply that tuning either parameter independently should alter the apparent number of modes in GAL expression distributions.Since ‫ܨ‬ ହ varies over a wider range of glucose concentrations than ‫ܧ‬ ଵ does under the perturbations we tested, we asked if independently tuning ‫ܨ‬ ହ affects modality.Indeed, plotting GAL reporter distributions shows that a number of allele replacements are able to convert strains from being bimodal to unimodal and vice versa.For example, DBVPG1106 is bimodal, but replacing alleles with GAL3 S288C and MKT1 S288C increases its ‫ܨ‬ ହ and makes it unimodal (Figure 5A-B).
Conversely, BC187 is unimodal, but decreasing its ‫ܨ‬ ହ by introducing GAL3 YJM978 makes it bimodal (Figure 5C-D).Finally, Y12-WashU, one of the most obviously bimodal strains, is .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint rendered unimodal when pre-conditioned in acetate rather than raffinose before galactose induction (Figure 5E-F).

Independent tuning and the molecular mechanism of bimodality
It is known that positive feedback on GAL gene expression through GAL3 tunes the switching rate of cells between uninduced and induced states [14] and is a key contributor to the bistability of the pathway [13].Changes in GAL3 dosage affects the induced fraction of GAL genes [16], and a panel of natural GAL3 variants confers a spectrum of GAL induction phenotypes [19].We put these previous observations in context by showing that natural GAL3 alleles specifically affect the sugar threshold where individual cells to switch to an induced state, while the level of induction in that state is set by GAL4 and other unknown genes.Both these features combine to yield the population level behavior of the circuit, including apparent patterns of bimodality.
Underscoring this point, we found that GAL3 allele replacement is sufficient to convert a unimodal response to bimodal, and vice versa, while the level of the induced subpopulation remains unchanged.This degree of modularity in the quantitative behavior of the GAL circuit was previously unappreciated.
GAL4 is the transcription factor activating all inducible GAL genes [10,28].Previously, changes in dosage of GAL4 was found to have no effect on the GAL induced fraction [16].The S288C variant of GAL4 contains a non-conservative R95G mutation, as well as a conservative R508K mutation, relative to DBVPG1106 and other natural isolates.Residue 95 is on a loop linking the DNA-binding and regulatory domains of GAL4 and directly participates in interactions with Gal11p [29,30], a component of the RNA polymerase II mediator complex that enhances .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint expression of GAL genes [31].These observations suggest that the S288C and DBVPG1106 GAL4 alleles might differ in their ability to activate transcription of GAL genes.This effect could be specific to induced level if differences in GAL4 activity only affect GAL promoters that are in an active state and the latter variable is separately dictated by feedback loops such as GAL3.An important question for future work is whether this scenario is quantitatively plausible in a mathematical model of the GAL circuit, and what general features of this and other circuits allow for independent tuning.

Modularity of the GAL pathway, genetic background, and metabolic state
We also find that MKT1 alleles affect the GAL response and can play almost as large a role as GAL3.MKT1 is involved in maintaining killer toxin [32], regulating translation [33], and affects numerous traits in crosses between S288C and natural isolates [34][35][36][37][38][39].The S288C allele of MKT1 is a loss-of-function variant relative to natural alleles and causes lower expression of mitochondrial genes [40,41].In turn, deletion mutants of mitochondrial genes tend to exhibit aberrant GAL induction; this effect is more pronounced on the induced fraction than on induced level [42], echoing our observations.Therefore, it is likely that the effect of MKT1 allele on GAL induction is due to perturbations to mitochondrial function.
We found that much of the variation in ‫ܧ‬ ଵ , and to a lesser extent ‫ܨ‬ ହ , must be attributed to unknown alleles in the genetic background.This dovetails with other recent reports that many traits in yeast are dominated by large effects from one or a few loci but can be tuned quantitatively by many small-effect loci [19,[42][43][44].Moreover, MKT1 is not a member of the canonical GAL pathway, and nor are any genes in 2 other loci that reached significance in our linkage mapping.Combined with observations that deletion mutants of up to quarter of all yeast .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint genes have quantitatively perturbed GAL signaling [42], our results indicate that decisionmaking circuits are not as modular with respect to genetic variation as is often assumed. In addition to genetic effects on GAL induction, we found that culturing cells in raffinose, glycerol, or acetate prior to induction in a glucose + galactose led to very different GAL phenotypes.Raffinose is commonly used to pre-culture cells for GAL induction studies (including most experiments here) because unlike glucose, it does not visibly repress GAL genes [14,45].We chose glycerol and acetate by the same criterion.Otherwise, however, these carbon sources elicit very different physiological responses.Raffinose is hydrolyzed to release fructose [46], which can then be fermented [47,48].Glycerol and acetate, by contrast, must be utilized via respiration [49], which entails expression changes in many genes [50] as well as differences in ATP/ADP ratio and redox state [51].Therefore, our results suggest that factors other than canonical glucose catabolite repression may be important in setting the inducibility of GAL genes.
Our results indicate that memory of metabolic state is encoded by the GAL circuit and persists even after the cells have reached steady-state in inducing conditions (Figure 5B, Materials and Methods).This appears to be a distinct phenomenon from the "memory" of glucose or galactose pre-induction conditions previously attributed to bistability of the GAL network [14,52].
However, the fact that pre-induction carbon source specifically affects ‫ܨ‬ ହ , just as GAL3 allele does, suggests that this positive feedback loop may be a nexus of regulation of GAL genes by multiple signals in the cell.Indeed, recently it was shown that NAD(P) can directly bind Gal80p and thereby impact downstream GAL pathway expression [53,54].Since these studies relied on bulk measurements, it will be interesting to revisit these investigations using quantitative, singlecell readouts of pathway behavior.
. CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint

Physiological and ecological role of independent tuning
Our results raise the question of why independent tuning of induced fraction and induced level would exist in nature.Previously we showed that natural variation in the timing of GAL induction during diauxic growth leads to a fitness tradeoff-some strains prepare for glucose exhaustion at an upfront cost while others maximize their growth rate on glucose but suffer a diauxic lag [21].Related work showed that both strategies could be implemented by the same strain as part of a bimodal response [20], and that this may be an evolutionarily stable strategy [55].Under this framework, tuning ‫ܧ‬ ଵ and ‫ܨ‬ ହ separately would allow the timing of the inducing population, and its level of induction, to evolve separately.This could provide fitness benefits in certain conditions, although exactly what these conditions are would depend on the quantitative details of the costs and benefits of induction, an interesting issue to be explored in future work.

Strains and media
Strains were obtained as described in [21].An initial set of 42 strains were assayed in glucose gradients + galactose.Strains CLIB324, L-1528, M22, W303, YIIC17-E5, YJM975, YJM981 were excluded from downstream analysis due to poor growth in our media conditions.Strain 378604X was also excluded due to a high basal expression phenotype that was an outlier in our collection.The genetic basis of this behavior is likely an interesting topic for follow-up studies.
All experiments were performed in synthetic minimal medium, which contains 1.7g/L Yeast Nitrogen Base (YNB) (BD Difco) and 5g/L ammonium sulfate (EMD), plus D-glucose (EMD), D-galactose (Sigma), or raffinose (Sigma).Cultures were grown in a humidified incubator .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint (Infors Multitron) at 30°C with rotary shaking at 230rpm (tubes and flasks) or 999rpm (600uL cultures in 1mL 96-well plates).

Flow cytometry assay in glucose gradient
GAL induction experiments were performed in a 2-fold dilution series of glucose concentration, from 2 0 % to 2 -9 % w/v, with constant 0.25% galactose.2% glucose and 2% galactose conditions were also included with each glucose titration experiment.To assess and control for well-to-well variation, experiments were performed as a co-culture of a "query" strain to be phenotyped and a "reference" strain that was always SLYB93 (natural isolate YJM978 with constitutive mCherry segmentation marker).
To start an experiment, cells were struck onto YPD agar from -80C glycerol stocks, grown to colonies, and then inoculated from colony into YPD liquid and cultured for 16-24 hours.Then, query and reference strain cultures were mixed 9:1 by volume and inoculated in a dilution series (1:200 to 1:6400) in S + 2% raffinose medium.The raffinose outgrowths were incubated for 16-20 hours, and then their optical density (OD 600 ) was measured on a plate reader (PerkinElmer Envision).One outgrowth culture with OD 600 closest to 0.1 was selected for each strain, and then washed once in S (with no carbon sources).Washed cells were diluted 1:200 into glucose + galactose gradients in 96-well plates (600uL cultures in each well) and incubated for 8 hours.
Then, cells were harvested and fixed by washing twice in Tris-EDTA pH 8.0 (TE) and resuspended in TE + 0.1% sodium azide before transferring to a shallow microtiter plate (CELLTREAT) for measurement.Flow cytometry was performed using a Stratedigm S1000EX with A700 automated plate handling system.Data analysis was performed using custom MATLAB scripts, including Flow-Cytometry-Toolkit (https://github.com/springerlab/Flow-Cytometry-Toolkit).
. CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint Experiments using glycerol and acetate as pre-induction carbon sources were done as above, except S + 3% glycerol or S + 2% potassium acetate were used instead of raffinose medium for the outgrowth step.

Crossing and generating segregants
To prepare parent strains for crossing and sporulation, we sporulated diploid natural isolates bearing the hoΔ::GAL1pr-YFP-hphNT1 reporter cassette and isolated random spores that displayed MATa or MATα phenotypes in test crosses.We then introduced a constitutive fluorescent marker in tandem with the GAL reporter, to obtain MATa; hoΔ::GAL1pr-YFP-mTagBFP2-kanMX4 or MATα; hoΔ::GAL1pr-YFP-mCherry-natMX4 parent strains.To the MATa parent we also introduced a pRS413-derived plasmid bearing STE2pr-AUR1-C and hphNT1.This plasmid is maintained by hygromycin selection but also allows selection for MATa cells by Aureobasidin A [56].This plasmid design is inspired by a similar mating-type selection plasmid used in a recent study [57].
To isolate segregants for phenotyping, we crossed a parent with BFP-kanMX + MAT-selection plasmid to a parent with mCherry-natMX and isolated a G418 R Nat R Hyg R diploid hybrid with the plasmid.We sporulated the hybrid by culturing it to saturation in YPD, diluting 1:10 in YP+2% potassium acetate and incubating at 30C for 8 hours, and washing and resuspending into 2% potassium acetate and incubating at 30C until >20% of cells were tetrads, or about 3 days.We incubated ~5x10 6 tetrads in 100uL water with 50U of zymolyase 100T (Zymo Research) for 5 hours at 30C, and then resuspended tetrads in 1mL of 1.5% NP-40 and sonicated for 10 seconds at power setting 3 on a probe sonicator (Fisher Scientific).The resulting segregants were plated on YPD + 0.5ug/mL Aureobasidin A ("AbA"; Clontech) and random colonies were picked into .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint YPD liquid and saved as glycerol stocks.Haploidy was confirmed by mating to tester strains with known mating type.90 segregants were phenotyped for GAL induction as described above.

Sorting-based bulk-segregant analysis
To generate segregant pools, we prepared a diploid hybrid and sporulated it as described above.
To reduce the size of recombination blocks and improve the resolution of linkage mapping [58], we then performed the following "intercross" protocol 4 times: from spore suspension, use Sony SH800 Cell Sorter to sort 4x10 6 BFP+ or mCherry+ (but not +/+ or -/-) cells into 100uL YPD + 40ug/mL tetracycline; incubate for 16 hours at 30C without shaking; add 5mL YPD + 200ug/mL G418 + 100ug/mL ClonNat + 200ug/mL Hygromycin B and incubate 48 hours at 30C with shaking; sporulate cultures and prepare sonicated spore suspension.After the 4 th sporulation cycle, the sonicated spores were resuspended in YPD + 0.5ug/mL AbA and incubated at 30C for 16 hours.This culture was frozen as a glycerol stock, as well as used to inoculate the galactoseinduction sorting experiment.
To sort segregant pools for bulk genotyping, we inoculated the intercrossed, MATa-selected segregants from a saturated YPD culture into S + 2% raffinose + AbA at dilutions of 1:200, 1:400, 1:800, and 1:1600, and incubated at 30C for 16-24 hours.We chose the raffinose culture with OD closest to 0.1, washed once in S (0.17% Yeast Nitrogen Base + 0.5% Ammonium Sulfate), and diluted 1:200 into S + 0.25% glucose + 0.25% galactose + AbA.We incubated the glucose-galactose culture at 30C for 8 hours, and then used a Sony SH800 sorter to isolate pools of 30,000 cells with the 5% lowest ("OFF") and highest ("HI") YFP expression, among cells whose Back Scatter (BSC) signal was between 10 5 and 3x10 5 .The "LOW" pool was similarly obtained, but from the 5% of cells with lowest non-basal expression (Figure S3).The sorted cells were resupended in YPD + AbA and incubated at 30C until saturation, about 16-24 hours.An .CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint aliquot of this culture was saved for -80C glycerol stocks, and another was used to prepare sequencing libraries. To sequence the segregant pools, we extracted genomic DNA from 0.5mL of saturated YPD culture of each segregant pool using the PureLink Pro 96 kit (Thermo Fisher K182104A).From these genomic preps, we made sequencing libraries using Nextera reagents (Illumina FC-121-1030) following a low-volume protocol [59].We adjusted the input DNA concentration so that resulting libraries had mean fragment sizes of 200-300bp as measured on a BioAnalyzer.
Libraries were multiplexed and sequenced in an Illumina NextSeq flow cell to a depth of 16-33x.
Reads from the Illumina sequencing were aligned to the sacCer3 reference genome using bwa mem, and SNP counts were generated using samtools mpileup, on the Harvard Medical School Orchestra cluster .These outputs were processed in MATLAB using custom code as follows: SNPs with coverage less than 2 or more than 1000 were removed.The LOW and HI pools were computationally merged into an ON pool.To make sure the two pools contributed equally to the merged pool, at each SNP, allele counts in the pool with higher coverage were randomly subsampled to the coverage of the other pool.The final allele counts in each pool were output to text files by chromosome and given as inputs to the MULTIPOOL algorithm (mp_inference.pyversion 0.10.2) [60] to compute LOD scores.Loci with maximal LOD>5 were considered significant; previous work showed that this corresponded to an FDR of 5% [61,62].This correspondence may differ under our experimental conditions; therefore, the 2 loci that we did not validate experimentally should be interpreted with caution.
. CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint

CRISPR/Cas9 allele replacement
Allele replacement strains were constructed using 3 rounds of gene knockout followed by CRISPR/Cas9-mediated markerless integration of heterologous allele.Initially, strains were prepared by introducing Cas9 on a CEN/ARS plasmid (SLVF11); this plasmid is derived from a previous one [63], but we replaced the auxotrophic URA3 marker with AUR1-C to allow Aureobasidin A selection on prototrophic natural isolates.In each round of allele replacement, a gene plus upstream and downstream flanking sequences (-784bp to +815bp for GAL3, -449bp to +372bp for MKT1, -191bp to +139bp for GAL4) was deleted by integration of a kanMX6 marker with 40bp flanking homology.Then, a donor DNA, a guide RNA insert, and a guide RNA backbone were simultaneously transformed into the strain [64].The donor DNA contains the new allele, its flanking sequences, and an additional 40bp of homology to target it to the correct genomic locus.The guide RNA insert was a linear DNA containing a SNR52 promoter driving a guide RNA gene containing a 20bp CRISPR/Cas recognition sequence linked to a crRNA scaffold sequence, plus 40bp of flanking homology on both ends to a guide RNA backbone.The guide RNA backbone was a 2u plasmid containing natMX4 (pRS420).This was linearized by NotI + XhoI digestion before transformation.Allele re-integration transformations were plated on cloNAT to select for in vivo assembly of the guide RNA into a maintainable plasmid, and Aureobasidin A to select for presence of Cas9.Successful re-integration was verified by colony PCR and Sanger sequencing was performed on a subset of strains and on all donor DNAs to verify the sequence of allelic variants.
. CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint   (A) ‫ܧ‬ ଵ versus ‫ܨ‬ ହ for all 16 combinations of S288C ("S") or DBVPG1106 ("D") strain background (gray letters), GAL3 allele (red), MKT1 allele (green), and GAL4 allele (blue).
Effects of switching from DBVPG1106 to S288C variant while holding other genetic variables

Figure 1 .
Figure 1.Natural variation in GAL induction can be analyzed in terms of two uncorrelated features Each plot is a series of YFP fluorescence (normalized to side scatter "SSC") histograms from 12 sugar conditions for strains (A) I14, (B) S288C, (C) DBVPG1106, and (D) YJM978.Other phenotyped strains are shown in Figure S1.Darker regions represent more frequently observed YFP values.The middle 10 conditions in each plot are 0.25% galactose + the indicated concentrations of glucose.The first and last conditions contain only one sugar: "D", 2% glucose; "G", 2% galactose.(E) Identification of induced cell subpopulation (green shading) using a reference distribution from 2% glucose (black histogram) (Materials and Methods).(F) Induced level (blue line) and induced fraction (orange line), and the corresponding ‫ܧ‬ ଵ and ‫ܨ‬ ହ metrics, for strain DBVPG1106.(G) Scatterplot of ‫ܧ‬ ଵ versus ‫ܨ‬ ହ across 34 S. cerevisiae natural isolates (mean and S.D.; n=3-10).

Figure 2 .
Figure 2. Bulk segregant analysis of and (A) ‫ܧ‬ ଵ versus ‫ܨ‬ ହ across 90 haploid segregants of the DBVPG1106 x S288C cross.Parent phenotypes are shown as filled circles: DBVPG1106 (red), S288C (blue).(B) Schematic of bulk segregant analysis strategy.(C) GAL reporter histograms of parent strains DBVPG1106 (red) and S288C (blue) and a pool of haploid segregants (gray, bottom) in the sorting conditions, 0.25% glucose + 0.25% galactose.Green boxes are a schematic of the gates used to sort segregant cells into 3 phenotyped pools for sequencing (Gates used in actual sorting experiment are shown in Figure S3).ON pool allele counts are a computational sum of the LOW and HI pool allele counts (Materials and methods).(D) Genome-wide plots of differential allele frequency and log-odds-ratio (LOD) as computed by the MULTIPOOL algorithm (see Main Text, Materials and Methods).Top two panels show the OFF/ON comparison; bottom two panels show the LOW/HIGH comparison.

.
CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under The copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10.1101/131938doi: bioRxiv preprint constant are shown as arrows for switching (B) GAL3 allele, (C) MKT1 allele, (D) GAL4 allele, or (E) strain background.(F) The effects shown in (B)-(E) are plotted as differences in ‫ܧ‬ ଵ versus differences in ‫ܨ‬ ହ .

Figure 4 .
Figure 4. Perturbations that affect do not affect (A) ‫ܨ‬ ହ and ‫ܧ‬ ଵ for allele-replacement strains with S288C, BC187, or DBVPG1106 genetic backgrounds but containing alleles of GAL3 from various other natural isolates.(B) ‫ܨ‬ ହ and ‫ܧ‬ ଵ for 8 natural isolate strains induced in glucose+galactose after being cultured in raffinose, glycerol, or acetate.Raffinose pre-culture is the standard condition used for the other experiments in this paper.

Figure 5 .
Figure 5. Changing changes the number of modes of the GAL response Plotted are GAL reporter histogram series on a glucose gradient + galactose, as in Figure 1F, with induced expression level (blue line), ‫ܧ‬ ଵ (dotted vertical blue line), induced fraction (orange line), and ‫ܨ‬ ହ (dotted vertical orange line).These plots show one representative experiment for each strain/condition, out of the 3-12 replicates plotted in Figures 4.3and 4.4.Strains: (A) and 4.4.Strains: (A) DBVPG1106 with all endogenous alleles; (B) DBVPG1106 with replacements by GAL3 and MKT1 alleles from S288C.(C) BC187 with endogenous alleles; (D) BC187 with replacement by GAL3 allele from YJM978; (E) Y12-WashU cultured in raffinose prior to glucose + galactose (standard protocol); (F) Y12-WashU pre-cultured in acetate.

.
CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available underThe copyright holder for this preprint (which was not this version posted April 28, 2017.; https://doi.org/10. . CC-BY-NC-ND 4.0 International license a certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.It is made available under