Different Levels of Catabolite Repression Optimize Growth in Stable and Variable Environments

1A stable environment generally favors organisms that are well-adapted to that specific niche –. However, in many cases, adaptation to one environment comes at costs to fitness in alternative niches,–. Aside from the fitness tradeoffs, adaptation through mutation is relatively slow. Thus to deal with certain recurring environmental changes, many organisms have evolved the capacity to change gene expression in response to the environment, reducing the need for genetic adaptation.

2Microbial nutrient uptake and metabolism is a prime example of how organisms use transcriptional regulation to optimize fitness in variable environments. Because the expression of nonnecessary metabolic routes and genes is costly,,, microbes often use catabolite repression mechanisms to preferentially consume nutrients that afford high growth rates. This way, nonpreferred nutrient genes are only expressed when other, more preferred nutrients have been depleted. The sensing and signaling cascades required for carbon catabolite repression in the yeast Saccharomyces cerevisiae are particularly well-studied and serve as a model for similar systems in higher eukaryotes –. Glucose acts as a primary signal, triggering a regulatory cascade that results in repression of the consumption of alternative carbon sources, such as maltose, galactose, or ethanol. The main mechanism by which glucose regulates transcription is via the Ras/protein kinase A (PKA) signal transduction pathway. Other effectors include Snf1, the yeast homologue of mammalian AMP-activated PK, and Rgt1. Both of these proteins effect catabolite repression by triggering the transcriptional rewiring of a small subset of genes, many of which are directly involved in the uptake and metabolism of alternative carbon sources –,.

3Like other gene regulation programs, catabolite repression reduces the fitness cost associated with unnecessary gene expression while preserving the possibility of growth in environments with different nutrients. However, the ability to adapt to changing environments appears to be intrinsically opposed to obtaining maximal fitness in stable environments,–. This is partly because maintaining a regulatory mechanism requires energy and provides little benefit in stable environments,–. Moreover, in variable environments, transcriptional reprogramming also requires energy and time. This is clearly manifested when microbes switch from growth on one carbon source to another. During such a switch, cells show temporarily reduced growth rates (a so-called “lag phase”) because transcription is not yet adapted to the new environment,,. As a consequence, an organism's fitness in a given environment depends not only on the maximal reproductive rates in that particular niche, but also on the speed with which gene expression can adapt to the new conditions,,–.

4Because maintaining and operating an environmental sensing system is not always beneficial, it has been suggested that some environments might favor simpler strategies. For example, certain variable and unpredictable environments might stimulate strategies where a clonal population uses random (stochastic) switching between phenotypically different states. Such so-called “bet-hedging” strategies can generate phenotypic diversity independently of present environmental conditions. If this diversity correlates with the frequency of environmental uncertainty, this will be an “evolutionarily stable” strategy that ensures that some portion of the population is always adapted to future conditions. This strategy reduces the duration of the lag phase and also avoids fitness costs associated with maintaining an environmental sensor,,–. Other studies have proposed that microbes can evolve mixed gene regulation strategies that combine sensing with stochastic switching. Such “stochastic sensing” strategies use clues about future environmental changes to induce anticipatory transcriptional changes in a portion of individuals within the population,,,–.

5Whereas the molecular cascades underlying gene regulation have been extensively studied, the natural diversity and fitness costs and benefits of different gene regulation strategies have received less attention. This is in part because accurately measuring fitness across different environments is challenging. Because of the exponential nature of population growth, the long-term expected fitness of an organism is determined by its geometric mean growth rate (GMR) across every environment it encounters, weighted by the frequency and duration with which these environments occur,–,,,. Hence, even short periods of low fitness may have a significant effect on the long-term performance of an organism. “Generalist” strategies affording similar but more modest fitness levels across different environments can therefore result in a higher geometric mean fitness, even if maximal fitness in certain preferred environments is reduced,–.

6In this study, we use a combination of population- and single-cell-level measurements of the model eukaryote S. cerevisiae to explore how different environments shape fitness and transcriptional regulatory strategies. More specifically, we measure fitness as cells grow in environments with a stable supply of glucose compared to environments where cells need to transition from one carbon source to another. We find that different natural yeast strains show large differences in the speed with which they are able to adapt gene expression and growth to changes in carbon sources. Using experimental evolution, we demonstrate that growing a strain that shows slow transcriptional reprogramming in a variable environment frequently results in mutations in key regulatory genes such as HXK2. These mutations give rise to phenotypic “generalists” that thrive well in variable environments, with short lag phases, less stringent catabolite repression, and faster transcriptional reprogramming—at the expense of maximal growth rates (MaxRs) in a stable glucose environment. Many of these generalist isolates implement a transcriptional regulatory strategy mediated by “stochastic sensing” of alternative carbon sources, allowing cultures to maintain consistent fitness across different environments. Alternatively, the same selection regime can favor specialist mutants of an opposite character, which display tight catabolite repression and slow adaptation to new environments (long lag phases), but higher growth rates in stable glucose conditions. An experimentally validated mathematical model reveals how alternative regimes of variable carbon environments will favor one carbon catabolite repression strategy over another. Together, our results reveal that the speed with which genes are induced and repressed in response to environmental signals is a highly variable and evolvable trait. Our study moreover illustrates how distinct strategies of transcriptional reprogramming shape fitness in constant or variable environments.

7To investigate how the lag phase can influence fitness, we compared the growth behavior of 18 different S. cerevisiae strains in stable and variable environments. These included two commonly used laboratory strains (S288c and SK1) as well as 16 genetically diverse strains described by Liti et al. (2009). We used bulk population-level growth measurements in an automated plate reader (see Materials and Methods) to measure fitness in four different environments, including one stable condition with abundant glucose and three variable environments where populations gradually run out of glucose and thus need to switch to a different carbon source to continue growth. To obtain a stable “high glucose” (HG) condition, we supplemented the growth medium with 3% glucose, a condition that allowed relatively constant growth rates until cells entered stationary phase. At the other extreme we supplemented growth media with only 0.5% glucose (low glucose, LG), a condition that allows cells to first grow quickly by fermenting glucose, and then reprogram their metabolic genes to switch to respiratory growth on the ethanol accumulated during the fermentation phase (Figure 1 and Figure S1). Two other variable conditions included supplementation of LG with either maltose or galactose, two nonpreferred fermentable carbon sources whose metabolism is repressed in the presence of glucose (Figure 1 and Figure S1).

8As expected, many yeast strains showed a clear lag phase when grown in media that induce an adaptation to new carbon sources (called a “diauxic shift”) (Figure 1 and Figure S1). The shift can be detected by three characteristic changes in growth rate: a deceleration in growth speed as glucose is depleted, a brief phase where the growth rate reaches a (local) minimum, and subsequent re-acceleration to adapted growth on the alternative carbon source (Figure 1A,B and Figure S1A). The lag phase was especially pronounced in media with LG alone, and more subtle or in some cases completely absent in LG + galactose and LG + maltose. However, even when the growth deceleration during the diauxic shift is very pronounced, it is not trivial to accurately quantify the lag phase because populations rarely arrest growth completely, and because the start and end of the phase cannot be clearly defined. Furthermore, the deceleration and local minimal phases appeared to be affected by the presence of an alternative carbon source (Figure S1A), and this effect was highly variable between strain backgrounds. For example, S288c growing in LG + maltose initially decelerates faster than the LG condition, but subsequently shows significantly higher growth rates across the rest of the experiment. By contrast, the LG + galactose sample has a higher rate of growth throughout the deceleration phase but decelerates later on in the curve as the culture adapts fully to galactose consumption (Figure 1C,D and Figure S1A). Moreover, these differences are strain-dependent. Strain Y55, for example, shows a more pronounced decrease in growth speed during the diauxic shift in LG + galactose compared to LG alone (Figure S1A).

9The results above show that it is difficult to quantify the population-level lag phase by simply measuring its duration. In order to quantify the rate with which strains are able to adapt to variable environments, we therefore use a simple metric that summarizes the overall growth speed (or fitness) of a population as it transitions through a diauxic shift. This GMR is the weighted geometric mean of growth rate values across the shift in carbon conditions and represents the average growth rate of the strain across a specific interval. We chose to calculate the GMR for the interval between O.D. 0.15 and O.D. 0.75, which comprise the complete shift from one carbon source to another (see Materials and Methods and Figure 1A,B). The GMR can approach but never exceed the MaxR achieved by the culture while it was growing on the preferred carbon source glucose. Hence, the longer that a population of cells grows in stable glucose conditions, the closer its GMR approaches the culture's MaxR. In mixed media, by contrast, the GMR can be considerably lower than MaxR because of the decline in growth speed as cultures transition to growth in the nonpreferred carbon sources. Therefore, the ratio of the GMR in stable versus variable conditions is a measure for how efficiently the cells can transition between the different conditions, which in turn largely depends on the severity of the lag phase (see also Materials and Methods).

10Examining the GMR across growth in stable (HG) and variable (LG, LG + maltose and LG + galactose) conditions allowed us to quantify how the carbon environment affected the overall fitness of the different strains. As expected, for all strains, the highest GMR was found in stable (HG) conditions, whereas the LG condition resulted in the lowest GMR (Dataset S1). Remarkably, however, the difference between these bounds was highly variable: whereas some strains only showed a 10% reduction in GMR in LG conditions compared to HG conditions, others showed a 70% reduction in GMR. A similar although less pronounced variation between strains was also observed for LG + galactose and LG + maltose media (Figure S1B and Dataset S1).

11Together, these results indicate that the diauxic shift from glucose to a less preferred carbon source leads to a wide array of growth behaviors, ranging from highly variable growth rates in different media to almost constant growth rates throughout the shift in carbon source. The decrease in fitness during diauxic shift depends both on the yeast strain and the carbon source. In general, the transition from glucose to ethanol (LG medium) induced the strongest decrease in growth rate (and thus also GMR), whereas the glucose to maltose transition (LG + maltose medium) caused the smallest reduction compared to stable HG conditions (Figure S1B).

12To obtain a more detailed picture of the decrease in fitness during the lag phase as cells transition from one carbon source to another, we turned to single-cell measurements. The cells were first grown in glucose, harvested, and then immediately transferred to maltose-containing media. The duration of the lag phases of individual cells after this sudden transfer to maltose was measured using automated time-lapse microscopy (Materials and Methods and Movies S1 and S2). We recorded the lag phases of individual cells as the time it took a cell to begin dividing after transfer from glucose to maltose-containing medium (i.e., the first appearance of a cell bud or the resumption of bud growth). The results of these experiments indicate that the 18 yeast strains showed a striking diversity in the average duration of the lag phase (Figure 1G). Interestingly, even within a population of a given strain, we often observed a large diversity in lag duration among different individual cells. Moreover, in some cases, not all cells seemed to survive the shift in carbon source. The degree of intraclonal variability in lag duration (measured by the standard deviation of single-cell lag phases of a population of cells of a given strain) and fraction of cells surviving within 20 h of recording was significantly related to the average lag phase, an observation that held true for all strains and mutants we examined in this study (Figure S1C–F).

13Strikingly, the average duration (and thereby heterogeneity) of the single-cell lag phase of a given yeast strain was also highly anticorrelated with the fitness (GMR) measured in population-level experiments where strains often displayed pronounced lag phases, such as in LG and LG + galactose conditions. More generally, cells showing long lag phases as measured by single-cell microscopy also showed higher variation in fitness (standard deviation and coefficient of variation in GMR) between different growth media (HG, LG, LG + Gal, LG + Mal; see Figure 1H). Further statistical analysis using proportional hazard regression (see Text S1) confirmed this correlation (Dataset S1B).

14Taken together, these analyses indicate that the efficiency with which populations are able to shift between carbon sources varies significantly between different strains, and is correlated with the average lag phase measured during sudden shifts in carbon source. Specifically, strains that show large differences in fitness between stable (HG) and variable (LG, LG + Maltose and LG + Galactose) environments also show long lag phases.

15The results from the previous section support that lag phases measured for single cells in sudden glucose to maltose shifts are correlated with lag phases measured by monitoring population-level growth in conditions where glucose is more gradually depleted (Figure 1H and Dataset S1). Interestingly, although sudden transitions from glucose to maltose media often led to long and heterogeneous lag phases, most cultures growing in LG + maltose mixed media displayed modest lag phases, suggesting that the yeast cells were able to maintain higher overall growth rates if the transition between glucose and maltose was more gradual. Exploring the molecular details of this shift from glucose to maltose proved to be an ideal model because only three genes are required for maltose consumption, making it a simple system to study. To grow on maltose, yeast cells need to express a maltose transporter (MalT), a maltase (MalS), and a regulator (MalR) that induces the genes in the presence of maltose via positive feedback regulation. To characterize this phenomenon in further depth, we chose to work with the laboratory strain S288c because it is genetically tractable and displays a clear lag phase in glucose-to-maltose transitions.

16It seemed likely that carbon catabolite-mediated repression of the MAL genes was a key factor contributing to the long lag phases in sudden glucose to maltose shifts. Using fusions of the MAL proteins with fluorescent reporters, we observed, as expected, that expression of MAL genes is repressed in the presence of glucose, and induced in the presence of maltose after glucose is depleted (Figure S2). Moreover, constitutive overexpression of the MAL genes resulted in a much shorter lag phase (logrank Chi2 = 591, p<1×10−16) that was comparable to some of the natural strains with short lag phases, suggesting that a long lag phase is due to the slow de-repression of MAL genes (Figure S2B). Taken together, these results suggest that the duration of the lag phase is determined at least in part by the gene expression state of cells upon transition to the new carbon source. In sudden shifts from glucose to maltose, lag phases appear to be longer and more heterogeneous due to the time required to activate the MAL genes, whereas in mixed LG + maltose medium, strains can prepare for maltose fermentation before glucose is completely depleted (Figure 1C,D and Figure S2).Therefore, the different lag phase behaviors are likely to be mediated by a variable extent of carbon catabolite repression across strains and conditions.

17At first sight, it may seem suboptimal for strains to have long lag phases instead of rapidly adapting to a new environment. However, long lag phases might be adaptive under certain conditions. For example, a long lag phase caused by carbon catabolite repression could potentially allow cells to more rapidly resume growth should preferred carbon sources like glucose return to the environment. To test this, we transferred a population of glucose-repressed cells to maltose, waited until half of the population had committed to maltose growth, as indicated by expression of a MalS-YeCitrine fluorescent reporter construct. We then transferred these cultures to glucose medium and measured the initial MAL fluorescence state of individual cells and subsequently tracked these cells' growth rates in glucose using time lapse microscopy (Figure 2 and Figure S2D,E). Compared to isogenic sister cells that had not yet escaped the lag phase, cells that already had activated their MAL genes showed lower growth rates in glucose for at least two cell divisions, showing that commitment to maltose growth came at a fitness cost when glucose reappeared. The large distribution in lag duration between isogenic cells in a population may therefore serve as a way to distribute the costs and benefits of commitment to nonpreferred nutrients across individuals within the population. Strains with tight catabolite repression, with longer and more heterogeneous distribution of single-cell lag times, appear to implement this strategy to a greater extent than strains with shorter and more homogeneous lag time distributions.

18The above results indicate that the lag phase (and thus the speed of transcriptional reprogramming) has significant genetic determinants: there is a wide variation in lag duration between genetically distant yeast strains, and further we can engineer shorter lag phases with a reverse-genetics approach. Moreover, because it appeared that long lag phases could themselves be beneficial when glucose returns frequently to the environment (Figure 2), we reasoned that this trait should be subject to natural selection. To test this, we cycled the strain S288c between glucose and maltose to generate conditions of strong selective pressure. In the first leg of the cycle, to maintain selection on MaxR in preferred nutrients, cells were allowed ∼10 generations of exponential growth in stable glucose conditions. Then, in the second half of the cycle we selected for short lag phases by transferring these cultures into maltose-containing medium, allowing for ∼5 more generations of growth. In a first experiment, 12 populations expressed a constitutively transcribed YeCitrine marker to facilitate later analysis in competition experiments. After six cycles or ∼90 generations, 11 out of 12 evolving populations showed shorter lag phases when compared to the ancestral strain (Figure S3A and Movies S3, S4, S5).

19At the end of the experiment we isolated individual cells from each population. We found that clones isolated from different cultures varied widely in their single-cell lag profiles (Figure 3A), but that the profiles of individual isolates within each evolved population were similar (Dataset S2), suggesting that a single evolved phenotype had come to dominate each independently growing culture. The average single-cell lags of the isolated mutants range from as short as 5 h, comparable to the shortest lag phases observed in the collection of different strains reported in Figure 1, to clones with lags of longer duration than the ancestral strain (Dataset S2). The Malthusian fitness under conditions mimicking selection for each of three short-lagged isolates from independent populations was ∼1.35–1.4-fold higher than the ancestral strain (Figure 3B).

20To investigate how reproducible this result was, and to unravel the underlying genetic and molecular mechanisms that allowed these strains to increase fitness, we repeated the evolution experiment, however this time using 12 populations of cells bearing MALT-YeCitrine and MALS-mCherry constructs to allow measurement of MAL gene expression. After eight cycles or ∼120 generations, all 12 populations in this experiment showed shorter lag phases (Figure S3B). We first carried out extensive growth rate analyses on 36 isolates (three clones from each of the 12 populations). Similarly to what we observed in the previous experiment, the majority of the mutants isolated after cycling populations between glucose and maltose medium showed shortened lag phases (Figure S3B,C). Using population-level growth measurements similar to those reported in Figure 1, we found that these evolved strains had smaller differences in fitness (GMR) between different growth environments than the parental strain, leading to overall higher fitness across the shift in carbon sources relative to the ancestral strain (Figure 4).

21Interestingly, despite the fact that selection was only based on glucose-to-maltose cycles, the mutants also showed dramatic improvement in fitness in media containing LG alone or LG + galactose, conditions where the ancestral strain showed a pronounced lag phase (Figures 4 and S4). For example, several mutants no longer have any lag phase in galactose-containing media, maintaining steady rates of growth throughout the curve with no local growth rate minimum (Figures 4B and S4 and Dataset S2). This reduction in lag phase leads to a 1.2–1.4-fold higher fitness (GMR) during the diauxic shift. Likewise, several isolates have a GMR increase in LG conditions of up to 1.3-fold, an increase in fitness mediated by increased rates of growth throughout the deceleration, local minimum, and re-acceleration phases of the lag phase (Figure 4B,C and Figure S4). In LG + maltose conditions, where the ancestral strain showed a relatively limited lag phase (Figure 1C,D), we found that the evolved strains' fitness also showed a modest increase due to a further reduction of the lag phase.

22Taken together, compared to the ancestor, the majority of evolved isolates developed a low degree of fitness variability (i.e., similar fitness levels in stable HG and variable conditions) and short lag durations that are similar to some of the strains measured in Figure 1 (Figure S4D and Datasets S1 and S2). Intriguingly, however, in addition to the isolates with increased fitness during transition between carbon sources, clones from a few populations showed an increase in lag duration compared to the parental strain (Figure 4D and Figure S4C,D). These isolates appear to have evolved higher MaxRs in stable glucose conditions at the cost of even more pronounced lag phases than the ancestral strain. When the average fitness (GMR) is plotted against the average MaxR, it becomes clear that isolates generally evolved following two different paths: the evolution of shorter lag phases and increased fitness during carbon transitions at the expense of MaxRs in stable glucose conditions, or the evolution of faster growth in stable conditions at the expense of longer lag phases in variable environments (Figure 4D and Dataset S2).

23To test which mutations might have given rise to short lag phases and increased fitness during carbon transitions in the evolved populations, we sequenced the genomes of four evolved isolates showing shorter lag phases. Three mutants of the shortest lag phase length carried mutations in the glucose sensor HXK2, a gene that encodes a protein with multiple genetically dissociable roles in glucose sensing. Specifically, in addition to phosphorylating glucose for entry into glycolysis, Hxk2p plays a signaling role in the SNF1 and Ras/PKA pathway, and further can itself translocate into the nucleus to repress certain nonpreferred carbon catabolite genes. High throughput studies have shown that clean deletions of this gene lead to reduced fitness in YPD media (a condition akin to our HG media, YPD contains 2% glucose) relative to the WT, and increased fitness in diauxic shift from 0.1% glucose to ethanol and glycerol. Another study demonstrated that deletion of this gene leads to genome-wide disruption of transcription, with significant gene ontological (GO) enrichment for genes involved with respiration and nonpreferred carbon metabolism.

24Another isolate (Isolate 3) showing longer lag phase duration carried a mutation in the STD1 gene, encoding a protein that interacts with glucose sensors Snf3 and Rgt2 to regulate RGT1-mediated repression of nonpreferred carbon source genes. Like HXK2, STD1 is also involved in glucose-induced repression of alternative metabolic pathways. Interestingly, however, deletion of this gene typically leads to higher rates of growth, and a disruption of gene regulation that is anticorrelated with that of a strain in which HKX2 has been disrupted.

25Sequencing of HXK2 and STD1 in the 36 different evolved isolates revealed that HXK2 was mutated at different positions in all but one of the different populations that showed shorter lag phases, whereas STD1 was mutated only once (Figure 5A and Datasets S3 and S4). Isolates bearing HXK2 mutations had significantly shorter single-cell lag phases in sudden glucose to maltose shifts than other isolates (12.72±2.85 versus 4.60±0.77 h for non-HXK2mut and HXK2mut strains, respectively) and were 26 times more likely to resume growth after a transition from glucose to maltose (Cox hazard coefficient = 3.26, p<10−10). Moreover isolates with HXK2 mutations had significantly reduced population-level MaxR and increased GMR across the four carbon source environments (p<0.001). To confirm that these mutations were sufficient to confer comparable growth strategies to the evolved strains, we introduced the mutated STD1 allele and two of the HXK2 alleles into the ancestral S288c genome (Materials and Methods). The mutations phenocopied the behavior of the evolved isolates (Figure 5B,C). Furthermore, reverting the mutation back to WT in the evolved clones had the opposite effect, restoring wild-type growth patterns and fitness. Control strains bearing the same dominant marker but that did not incorporate the intended allele all behaved as the parental strain (Figure 5B,C and Datasets S3 and S4). Taken together, these findings reveal how simple mutations in carbon sensing pathways can tune the length of the lag phase in both gradually and suddenly changing environments.

26To determine the molecular mechanisms giving rise to the altered growth characteristics in evolved isolates, we examined whether the mutants display altered MAL gene regulation. Flow cytometry measuring the fluorescence of the MALT-YeCitrine and MALS-mCherry reporter constructs revealed that many short-lagged mutants showed reduced catabolite repression of the MAL genes (indicated by “leaky” MAL gene expression in glucose), possibly explaining why they have shorter lag phases in sudden glucose-to-maltose transitions (Figure 6A). Furthermore, the degree of leaky background expression correlated with high fitness (GMR) in variable carbon environments and was inversely correlated with the MaxR and lag-phase length (Figure S5A and Dataset S2). Although the MAL genes are not necessary per se for growth in alternative carbon sources like ethanol or galactose, this correlation indicates that the leaky expression of the MAL genetic reporter relates more generally to a partial loss of glucose catabolite repression.

27Most interestingly, in contrast to the leaky expression we observed in glucose media alone, we found that when maltose was added to glucose media, some HXK2 mutants expressed their MALT and MALS genes to a great extent (Figure 6A and Figure S5). Deleting the MAL activator protein MalR (Figure S5D) did not affect leaky background expression but did relieve high levels of MAL gene expression in maltose and glucose-containing media, indicating that the MAL genes were being induced by the presence of maltose despite the glucose present in the media. There was significant variability in magnitude between isolates, and further between isogenic cells in the same population (Figure 6A and Dataset S2). However, within single cells, the average MalT and MalS signals were expressed to similar extents, implying that the entire MAL system was activated (Dataset S2). The different HXK2 mutants show a wide range of MAL expression levels and expression noise, suggesting that the different mutations have distinct effects on catabolite repression.

28We reasoned that this heterogeneous MAL expression could serve as a model to elucidate how evolved isolates maintain steady fitness levels in more gradual diauxic shifts between glucose and maltose. To address this, we used both population- and single-cell analyses to track MAL gene activity (using fluorescent reporter constructs) and growth. Figure S5B and S5C and Movies S6, S7, S8 show the result of a time-lapse microscopy experiment of one isolate's cells growing in the presence of glucose and maltose. This experiment suggested that cells in the population switch stochastically between MAL-active and MAL-repressed states. Moreover, flow cytometry of a population of cells indicated that in the absence of glucose repression, similar to other positive feedback-driven networks in microbes,,, the MAL genes of these mutants show hysteresis (history dependence). That is, cultures that are inoculated into maltose/glucose media with repressed MAL genes display fewer induced cells than populations inoculated with MAL genes induced, even after ∼10 generations of growth (Figure 6B). We found the extent of this hysteresis was widely variable across the different HXK2 mutants, with wide variability in the rates at which populations switch from induced to uninduced (ON to OFF) and vice versa (Figures 6B and S6 and S7 and Dataset S2). The per-generation switching rate from ON to OFF and OFF to ON for HXK2 isolates depended on the initial carbon source (maltose or glucose) and the strain's genotype (two-way ANOVA F = 27, p<0.001; see Figure S7). The magnitude of expression of MAL protein for these strains exceeds the generation time, indicating that newborn cells inherit and propagate the MAL activity state from their respective mother cell (Figures 6B and S5, S6, S7 and Movies S6, S7, S8). These results imply that the different HXK2 alleles led to varying rates of switching between MAL induced and repressed states.

29The variability in MAL gene regulation has a profound effect on growth rate (Figure S8). For example, the magnitude of population-level MAL gene expression in maltose/glucose medium is anticorrelated with the population-level growth rate (Figure S8A), suggesting that the expression of MAL genes in medium containing glucose comes at a significant fitness cost. This cost is dependent upon an intact MAL activator, which drives MAL gene expression via positive feedback regulation (Figure S8B). Further single-cell analyses of one isolate confirm that genetically identical cells with transcriptionally active MAL genes grow at slower rates compared to cells that keep their MAL genes inactive (Figure S8C). Although induced cells suffer a fitness defect as long as glucose is present, they show a much reduced or even absent lag phase when glucose is no longer available, increasing their fitness during this transition phase (Figure S8D,E). In diauxic shift conditions, this bimodal gene expression state leads to a diversified growth strategy that distributes the costs and benefits of expressing the genes involved in alternative carbon source metabolism across individuals, allowing the population to maintain steady fitness levels in both stable and variable environments.

30The isolates resulting from our evolution experiment and subsequent population- and single-cell-level analyses indicate that cultures can generally evolve in two directions. In one case, a “glucose-specialist” phenotype emerges, with faster adapted growth rates in glucose together with long lag phases upon a switch to a different carbon source. Alternatively, generalist phenotypes evolve with shorter lag phases in maltose, at an apparent cost to MaxRs in glucose. Furthermore, increased leaky expression of the MAL genes in short-lagged mutants is anticorrelated with MaxR in glucose, and positively correlated with reduced lag phases and fitness variability during diauxic shift (Figure S5A). This further suggests that a molecular cost is paid in order for cells to be prepared for sudden environmental changes (Figures 6 and S5, S6, S7, S8). We reasoned that the predominant growth strategy would be shaped by both (1) how often the culture must switch from stable glucose environments to alternative carbon sources and (2) the relative duration of these environments,,.

31To test this, we first developed a stochastic model of the growth characteristics of population-level behavior based upon a culture of single cells escaping from the lag phase. In the model, each cell of a given strain's population is assigned a time tau corresponding to the point at which the cell will begin growth in maltose. The distribution of tau for a given strain in maltose is equivalent to the cumulative lag time distribution measurements (as reported in Figure 3A). After the time exceeds tau, the cell begins growth at the strain's specific growth rate in maltose. In glucose environments, we simulated that cells would grow without a lag phase at the measured MaxR. At any given time, the sum of the growing and nongrowing cells equals the total population growth as it escapes from the lag phase (Materials and Methods).

32The model allows prediction of the clonal interference patterns that would result between competing strains with different lag characteristics across various maltose-to-glucose switching regimes. For example, the heat map in Figure 7A illustrates how Isolate 1, a “carbon source generalist” HXK2 mutant with short lag phases, would compare in direct head-to-head competition with Isolate 6, a “glucose-specialist” strain with ∼28% faster growth rate in glucose compared to Isolate 1 but with considerably longer and more heterogeneous lag phases (as shown in Dataset S2). The model predicts that an environment that consists of a single shift from glucose to maltose will result in Isolate 1 growing rapidly to high relative frequencies in the population due to its very short lag phase. However, during prolonged growth in glucose, Isolate 1 grows more slowly and thus is rapidly outcompeted. Importantly, certain regimes of maltose-to-glucose shifts are predicted to result in a stable abundance of each isolate relative to the other (green region in Figure 7A).

33We tested the model in a set of experiments where two strains with different lag characteristics were competed in different carbon switching regimes. In a first set of experiments (Figure 7B), we measured Isolate 1 and Isolate 6's abundance relative to one another in single shifts from glucose to maltose, with varying duration of growth in glucose and in maltose. The cultures were grown initially in glucose so that they would have a lag phase upon switching to maltose, or alternatively carry on at steady-state growth rates in glucose. The results confirm the trend suggested by the model: under regimes with longer periods of maltose growth, Isolate 1 shows the highest competitive fitness; conversely, with increasing time in glucose, the fast-growing Isolate 6 performs better. Second, to examine if the model could predict strain performance over multiple cycles of glucose-to-maltose shifts, we carried the experiment forward for another two cycles for the cultures growing under the 4 h∶20 h, 8 h∶16 h, and 12 h∶12 h maltose-to-glucose switching regimes (Figure 7C). The results confirm that different regimes of environmental change result in drastic changes in the frequency of the two competing genotypes within the total population. For example, in the top panel of Figure 7C, when the period of time in glucose exceeds the time in maltose, Isolate 6 comes to dominate because the benefits of faster growth rates in glucose outweigh those of shorter lag phases in maltose. By contrast, as the length of time in maltose increases, the short-lagged Isolate 1 outperforms the slower switching strain despite lower rates of growth in glucose (Figure 7C, bottom panel). Note that the competitor populations depicted in the middle panel remain at relatively equal frequencies when the culture undergoes cycles of 8 h of growth in maltose followed by 16 h of growth in glucose.

34Interestingly, in the 12 h∶12 h regime, we observed that Isolate 1 grew ∼10% more slowly over the glucose leg of the cycle compared to our measurements of the same strain in steady-state conditions (Dataset S8)—thus growing ∼40% more slowly than the Isolate 6 competitor during glucose growth. Across the entire 24-h period of the glucose to maltose cycles, we measured that this slower rate of growth in glucose reduces the average growth rate of this strain by about 5% (Dataset S8). The reduced rate of growth in glucose indicates that Isolate 1 paid a considerable cost upon reintroduction to the glucose-containing environment. This result further supports conclusions from the experiment reported in Figures 2 and S8 where commitment to maltose resulted in slower growth rates for individual cells.

35Taken together, these results demonstrate that the duration and frequency of carbon environments shape the fitness of two of the archetypal phenotypes recovered from the evolution experiments. Environmental regimes with long periods of growth in glucose relative to the duration of time in maltose will favor the growth of the specialist phenotype, with stringent catabolite repression and high growth rates in glucose, but long lag phases upon switching to a different carbon source. By contrast, more frequent shifts of carbon source and longer periods of growth in maltose will favor the growth of strains with less stringent catabolite repression, resulting in short, homogeneous lag phases and slow growth in glucose.

36When Monod (1941) first described the diauxic shift and the corresponding lag phase, he was frustrated by why a culture would reach such slow growth rates despite the abundance of alternative carbon sources. Monod realized that the lag phase was due to the time required for transcriptional reprogramming, which in turn propelled research into the molecular mechanisms underlying gene regulation. However, surprisingly little attention was given to how this transcriptional reprogramming affects growth rate. Our experiments show that the speed with which metabolic genes are reprogrammed in the face of changes in carbon availability is a highly variable and evolvable trait. Strains can maintain high fitness in stable glucose conditions by tightly repressing the costly expression of genes needed for growth in alternative carbon sources. However, this results in slow transcriptional reprogramming when glucose is depleted, leading to reduced fitness during the adaptation to a less-preferred carbon source. By contrast, less stringent catabolite repression and “stochastic sensing” strategies come at a fitness cost in stable conditions, but allow quick and more uniform adaptation, which in turn can lead to higher fitness during the transition phase.

37It is tempting to speculate that feral yeasts are faced both with relatively stable as well as variable carbon supplies, and are therefore subjected to pressure that is similar to the selection in our directed evolution experiments. Moreover, the mutants that we isolated after experimental evolution closely resembles the diversity found in natural isolates, both regarding differences between strains as well as differences between single cells within populations (Figures 1H, S3, and S4D). Furthermore, we observe that mutants isolated after selection in variable environments appear to be constrained by opposing selection for maximal reproductive fitness in stable glucose environments versus adaptability in variable environments. These results are consistent with theories of phenotypic tradeoffs associated with optimality at specific tasks versus the capacity to adapt to new environments,,,–. It is notable, however, that we do not observe the same level of direct tradeoffs between MaxR and lag phase across the strains reported in Figure 1 as we do in the evolved isolates. Given the allelic diversity of these strains, such a result is not unexpected. For example, previous evolution experiments have observed that common phenotypic tradeoffs, such as that between growth rate and carrying capacity, become undetectable with increasing genotypic distance. We therefore speculate that single mutations that give rise to tradeoffs in the evolved isolates could potentially be compensated by additional mutations that might increase overall fitness or that other evolutionary paths are followed in more complex natural environments.

38The natural strains and evolved mutants with short lag phases show similar differences in lag duration and fitness variability. More specifically, strains with short lag phases show similar fitness levels across different environments, a phenotype referred to in the ecological literature as a “generalist” strategy to indicate that these organisms thrive equally well in different environments. It is also worth noting that the HXK2 and STD1 mutants showed increased fitness both in gradually changing conditions (LG, LG + maltose, and LG + galactose) where the preferred carbon source is gradually depleted, as well as in conditions where the cells experience sudden changes in carbon availability (Figure 4 and Dataset S2). In ecological theory, both of these “fine-grained” and “temporally fluctuating” environments are predicted to favor generalist growth strategies, with reduced variability in fitness across different environments,–,,,. However, in contrast to our results, most experimental evolution studies tend only to find niche specialization (i.e., higher rates of growth on one carbon source or the other),. It seems likely that these studies uncovered specialist phenotypes because the selective pressure to improve adapted growth speeds on particular carbon sources was greater than the pressure to reduce lag phases or increase switching rates between carbon sources,.

39The mathematical model and experiments reported in Figure 7 provide further insight how these strategies are shaped and favored by the environment. These analyses illustrate how different regimes of environmental change favor the growth of either specialist strains with stringent catabolite repression or alternatively generalist strains with leaky or even stochastic expression of genes involved in metabolism of different carbohydrates. More broadly, these results confirm theoretical analyses and experimental reports that while stable preferred environments will promote slow switching rates, frequently changing environments can favor the growth of phenotypes with high adaptability, even at the cost of slower rates of growth in preferred environments,,.

40Interestingly, apart from revealing differences in growth strategies between genetically different yeast strains, our results also illustrate how genetically identical cells within a population can also show different growth behaviors that can help to optimize a strain's fitness in variable conditions. Specifically, the results show that an organism can maintain competitiveness by allowing only a fraction of the population to adapt to a new environment. This subpopulation of cells contributes some progeny to exponential growth while keeping others dormant (Figures 1G, 3A, and S3). Although they contribute no progeny in the new environment, these uninduced cells appear to have an advantage upon reintroduction of the previous environment (Figure 2), increasing overall mean growth rates (Figure 7). Such growth strategies have commonly been investigated for the model of seed germination in annual plants, a system that in many ways bears semblance to our model system of the lag phase,,. Heterogeneity within isogenic populations bears further similarities to other microbial systems that lead to variable physiological states between isogenic individuals in the same population. For example, heat-shock resistance, the variable timing of meiosis, or the switch-like commitment to mating in yeasts likely underlie a cost-benefit strategy that increases fitness in the long term. Indeed, these systems are costly to implement,, but offer fitness advantages in stressful environments. However, despite receiving much attention in the literature, few studies have systematically quantified how heterogeneous individual-level behavior can affect evolutionary outcomes,,,.

41Our evolution experiments demonstrated the flexibility of growth strategies and also revealed that simple genetic switches in catabolite repression can regulate lag duration (Figures 4–6). Specifically, HXK2 mutations appear to increase fitness in lag phases in glucose-to-maltose shifts in part by allowing leaky expression of MAL genes (Figure S5). The repression of other genes involved in alternative carbohydrate metabolism is also likely relieved because the fitness of the mutants also improved in mixed carbon conditions different from those used in the selection scheme. This conclusion is supported by whole-genome gene expression profiling of a targeted hxk2 deletion, where genes involved with respiration and alternative carbon source utilization are significantly up-regulated. Although null mutations in HXK2 are known to relieve catabolite repression in S288c, the potential for more subtle mutations in this key regulator to allow “tuning” of switching rates between alternative carbon sources has not been explored extensively,,,,. Furthermore, in the case of the STD1 allele, we observe reduced growth rates, which is the opposite effect to that usually observed in high-throughput studies, where null STD1 alleles grow with higher fitness than the WT S288c strain,. Taken together, this aspect of our results suggests that regulators such as HXK2 sit at an apical position in the regulation of cellular physiology, allowing adaptive reprogramming of cellular fitness strategies in times of environmental change.

42Interestingly, some of the short-lagged isolates show a high degree of heterogeneity in MAL expression within a population (Figures 6 and S5). Specifically, in medium containing both glucose and maltose, some HXK2 mutants exhibit a striking multimodal state, where MAL genes in individual cells are expressed to varying extents ranging from repressed to induced. This behavior emerges because the rate of switching between ON and OFF is slower than the generation time, allowing newly budded cells to inherit their MAL expression state from the mother cell (Figures S6 and S7). This epigenetic behavior is due to the structure of the MAL genetic circuit, which induces via positive feedback (Figures S5D and S8B),,,. Furthermore, it appears that distinct HXK2 mutations can set different “energy barriers” for transitions between induced and uninduced states (Figures 6A,B, S6, and S7). Although such “stochastic switching” networks have been reverse engineered (for example, ), and shortened lag phases observed in natural selection,, no studies have found that mutations in global regulators can give rise to such a wide array of diversified gene regulation strategies.

43The simple genetic architecture of the MAL system has allowed us to closely examine the costs and benefits of different levels of catabolite repression and the outcomes of stochastic gene regulation,,,,,. The stochastic nature of the transition between MAL activation and repression results in diversified growth behavior that appears to be a bet-hedging strategy. However, maltose must be present to induce the positive feedback necessary for the high levels of MAL gene expression shown in Figures 6 and S5, S6, S7, and thus this environmental dependence does not satisfy the most stringent criteria for bet-hedging. Even in the case of stable, constant glucose environments, the low leaky expression of costly nutrient assimilation genes could be viewed as a mechanism of “stochastic sensing,” a term first used by Perkins and Swain (2009) to describe predictive microbial networks,. More recently, Arnoldini et al. (2012) demonstrated analytically that combinations of sensing and stochastic switching strategies are evolutionarily stable when environments provide partially reliable signals about future events. Given that such positive feedback-driven circuits are widespread in microbes, it is likely that nutrient assimilation pathways act as basic sensing tools to maintain long-term fitness in changing environments, without the need for complex sensing and signaling systems,,,,,,,,,.

44Taken together, our results show that individual-level heterogeneity in gene regulation and growth has strong genetic determinants. The speed of metabolic reprogramming in the face of environmental change is a highly regulated trait, and populations can implement catabolite regulatory strategies that fall between traditional sensing/signaling cascades and stochastic switching mechanisms. Specifically, stringent catabolite repression seems favorable in relatively stable environments, whereas less stringent regulation, or even stochastic sensing strategies can increase fitness in variable conditions where cells often need to switch their metabolism. We speculate that similar principles and emergent (epi)genetic switches likely also contribute to other gene regulation systems, including in human diseases involving clonal growth, such as microbial pathogenesis and cancer.

45Standard protocols were used for routine S. cerevisiae strain propagation. A specially engineered maltose-prototrophic S288c strain, bearing a functional MAL regulator allele (MAL63) in place of MAL13 on chromosome VII, was engineered to have a low petite frequency by rescuing a frameshift mutation in SAL1 to reduce the high petite frequency that occurs after extended growth on glucose (Text S1 and ). Other feral strains were part of the SGRP collection. Special attention was given to standardization of pregrowth conditions, in particular to avoid cells that would experience carbon depletion prior to transfer to maltose for lag phase measurements. Specifically, this entails keeping cultures at low densities throughout the experiments. Moreover, where appropriate, care was taken to measure steady-state conditions (where the growth speed of the population was stable). Please refer to specific experimental details provided in Text S1 for the precise conditions for each experiment.

46Cells from a turbid culture grown in YPD for 14 h were inoculated to a final density of 1×105 (haploid S288c) cells per ml in 150 microliters of YP media containing the appropriate carbon source and allowed to grow in the Bioscreen C (Growthcurves USA) at 30°C and continuous medium-amplitude shaking until stationary phase. All media for growth rate measurements were prepared starting from the same batch of double concentrated YP medium (20 g/l yeast extract, 40 g/l bacterial peptone), which was supplemented with an equal volume of filter-sterilized sugar solutions to generate 1× YP medium containing the required mixture of carbon sources to obtain HG (30 g/l glucose), LG (5 g/l glucose), LG + Gal (5 g/l glucose and 25 g/l galactose), or LG + Mal (5 g/l glucose and 50 g/l maltose). We found that variation in osmolarity is a significant factor affecting the lag phase, and we therefore supplemented LG medium with 0.14 molar sorbitol to match the osmolarity of the HG, LG + Mal, and LG + Gal media. All media were divided into smaller batches that were kept frozen until the day of use.

47All growth measurements represent the averages of at least three biological replicates. In general, growth measurements were highly reproducible, with standard errors generally below 5% of the measured growth rates. Standard errors are reported in detail in Datasets S1, S2, S3, S4, S5, S6, S7, S8. Growth of populations of S. cerevisiae were measured by OD600 readings every 15 min in the Bioscreen C automated OD meter at 30°C with constant medium amplitude shaking. This plate reader uses 100-well microcultivation of microbial cultures covered by a heated lid to prevent evaporation. R and MS Excel software were used for all data analyses. Briefly, all growth curves were smoothed using R's smooth.spline function, and then the first derivative of the log-transformed smoothed data was plotted as a function of population size between 0.15 and 0.75 OD600 units, corresponding to 1.0×107 to 5×107 haploid S288c cells per ml (see Datasets S6 and S7). These values are linearly correlated with cell density, and further correspond to lowest and highest OD at which we were able to obtain reproducible measurements, and further capture most of the active growth phase of the cultures before other noncarbon growth resources become depleted.

48Reported maximum growth rates were calculated as the slope of a linear regression model to ln-transformed OD600 values between 0.15 and 0.3 A.U. To calculate GMR, we used R's splinefun function to determine the amount of time t that the culture spent between interpolated OD600 values of 0.15 and 0.75; the GMR is thus equal to ln(0.75/0.15)/t. This interpolation approach significantly reduces the coefficient of variation between biological replicates (error) compared to a simpler approach that would only use the raw OD measurements. These statistics were further analyzed to determine relative variability across the different environments (see Datasets S1 and S2).

49Note that the GMR is an average growth rate that could be calculated for many different initial and final population densities. Here, we use the GMR to reflect the average growth rate across the linear range of our spectrophotometer (i.e., between O.D. 0.15 and 0.75). This is a relatively wide interval that comprises the initially high growth rate in glucose, the lag phase, and the resumption of growth on the new carbon source (Figure 1A,B; Materials and Methods). This interval is easy to standardize and yielded a high reproducibility (Datasets S1 and S2). However, to explore the role that the lag phase played in the GMR, we also calculated average growth rates over other more narrow windows centered on the middle of the lag phase (defined as the point where the culture reaches a minimal growth rate). These measures correlate strongly with the GMR. For example, the narrowest possible window—the minimal growth rate (reported in Datasets S1)—explained 55% of the variance in GMR for the SGRP strains that had lag phases (Dataset S1, R2 = 0.55, p<0.001). Computing the average growth rate across a wider custom-drawn window encompassing 0.1 to 1 doublings prior to and 0.1 doublings after the minimal growth rate explained from 66% to 79% of the variance in GMR (p<0.001). In summary, the GMR in between OD 0.15 and 0.75 seems to be largely influenced by the actual lag phase. Moreover, calculating the GMR across more narrow intervals around the center of the lag phase is much more complex, is not always possible for all growth conditions where strains do not always show a local minimal growth rate, and yields similar results. Further details and example raw data and analyses are available in Text S1.

50We devised a system that traps cells between a coverslip and an agar pad containing media necessary for growth. This allowed continuous monitoring of cellular growth for long periods of time using an inverted automated Nikon TiE fluorescence microscope placed in a temperature-controlled incubator. A 60×, 1.40 NA oil immersion lens was used to monitor up to 120 XY positions per experiment, and lag measurements were made using the microscope's automated Z-plane focusing.

51Lag phases for a given strain become longer as the culture grows for longer periods of time in glucose (Figure S2A). Thus, for analysis of lag phases, we varied the length of time that cells were grown in glucose to regulate the severity of single-cell lag phases. Apart from the results reported in Figure S2A, we either grew populations in glucose for 6 h (reported in Figure 1) or 20 h (reported in Figure 3—the length of time cells grew in glucose during the evolution experiment). After growth in this glucose environment, we transferred cells by two brief (2 min at 1,250×g) centrifugations and resuspension in maltose-containing media. Cells were then transferred to the custom growth chamber (see Text S1) and transferred to the microscope for analysis. Under the microscope, brightfield imaging proceeded every 15 min. After the experiment, cell budding events were scored manually as the hour at which the first morphological change leading to a new bud occurred, or when an already existing bud began to grow. Examiners were blind to experimental conditions at the time of scoring, and separate investigators independently replicated results of preliminary analyses.

52For statistical analysis of single-cell lag data, we used survival analysis: log-rank tests for pairwise comparisons between different WT strains or between ancestral strains and mutants, and a Cox proportional hazards test for datasets for which we had covariate information (see Text S1). For these latter analyses, single covariates were used (1 degree of freedom), and the most significant predictors were then used in paired analyses using other covariates (2 degrees of freedom). The majority of single-cell lag variance was explained by single covariates that reflected relative fitness (GMR) or fitness variability (Dataset S1).

53For doubling time measurements presented in Figure 2, an initial fluorescence image was acquired, followed by brightfield imaging every 5 min. Mother cell doubling times were recorded as the time t taken for a cell to complete two cell divisions (equal to t/2). Brightfield images reported in Figures S5 and S8 were acquired every 5 min, with fluorescence images in the mCherry field every hour. Growth rate measurements of microcolonies reported in Figure S8C account for the fold change in area of microcolonies between 3 and 6 h after recording began (equal to Δ(ln(area))/3 h).

54To measure the relative MAL expression reported in Figure 6 and Dataset S2, we pregrew cultures in maltose media (Figure 6A) or either maltose or glucose media (Figure 6B) and diluted them as exponentially growing cultures into a mixture of 2% glucose + 5% maltose in YP media to a final population density of 1–2,000 cells/ml and allowed them to grow for 20 h for final cell densities between 1–5×106 cells per ml. Cultures were then centrifuged to concentrate cells, and these were frozen at −80°C in 1× phosphate buffered saline (Sigma-Aldrich no P5493) in 25% glycerol until flow cytometric analysis. The low densities at which these cultures grew did not measurably affect glucose concentrations (unpublished data). For gene expression measurements, cell samples were thawed on ice until analysis on a BD Bioscience Influx flow cytometer. mCherry signal detection used a 561 nm laser coupled to a 610/20 nm detector and YeCitrine signal detection used a 488 nm laser coupled to a 580/30 nm detector. R's flowCore package was used to first filter out ∼70% of events using a filter (curv2Filt) that selected the highest-density regions in side- and forward-scatter dimensions. mCherry and YeCitrine intensities for each filtered sample were stored as binned fluorescent measurements and summary statistics.

55For each evolution experiment, the methods established by Lenski were largely followed. All growth for the experimental evolution experiment was in 5 ml of YP media containing either 10% glucose or 20% maltose at 30°C on a rotating wheel. The protocol was followed for two founding S288c strains derived from the modified S288c strain (see above): AN296 constitutively expressed a YeCitrine marker, and AN148 contained fusion constructs of MAL11-YeCitrine (encoding a MalT) and MAL12-mCherry (encoding a MalS). Populations of ∼30,000 initial cells were grown exponentially for 20 h in 5 ml glucose YP to reach population sizes of ∼1–5×107, and then cultures were centrifuged and resuspended in 20% maltose YP and then put back on the wheel for another 3 d until the populations reached high densities (∼5×108 cells/ml, with a final population size of ∼2.5×109). After each round of selection in maltose, we froze an aliquot at −80 for future analysis and resurrection. At the end of the selection experiments, we resurrected individual clones for analysis by diluting the populations to single colonies, and then restreaking random single colonies again to single colonies before phenotypic characterization and long-term storage at −80°C in glycerol.

56For the experiment reported in Figure 3B, we pregrew constitutively YeCitrine-labeled query strains (isolates 1, 3, and 4) in sextuplicate and mixed these 1∶1 with an mCherry fluorescently labeled S288c reference strain (AN74; see strain list in Dataset S4) exactly as in the selection protocol in 20% maltose YP media for 24 h. OD600s were determined and query cultures were mixed at a 1∶1 ratio between reference and query strains. Samples were frozen for initial ratio measurements, and initial population densities determined by diluting cells down such that 100–200 single colonies would grow in 2 d time on solid YPD plates. Thereafter, the exact protocol from the selection procedure was then followed: a 20-h growth in glucose, followed by 2 d of growth to high turbidity in maltose media. At the end, samples were frozen in 1× PBS + 25% glycerol for later analysis, and final population densities determined by plating as before. For flow cytometry analysis, see below.

57We calculated the Malthusian growth rate w of the query as:where N is the total population size determined by plating. Likewise w(reference) was calculated for the fluorescently labeled reference. The w(query)/w(reference) was taken as the fitness of the query strain. This value divided by the ancestral strain's fitness (calculated identically against the same reference) gives the fitness of the query strain. All fitness or relative growth rate measurements are the ratio of the query strain's fitness relative to the reference strain divided by the control query strain's fitness relative to the reference strain.

58For a given culture of cells, the population total Ntotal is equal to the sum of each ith growing and nongrowing cell. Within the population, there are J phenotypes (or strains) and each cell belongs to the jth strain or phenotype such that Ntotalj is the number of cells of the given phenotype or strain. At any given time, the proportion of cells in the jth phenotype is:

59Each strain j in environment k has a specific growth rate equal to μj,k.

60The model assumes specifically that cells entering into glucose from maltose do not have lag phases. Cells that have no lag phase in environment k grow at steady-state growth rates; thus, at time t the total population size of strain j is determined by: with total population size.

61The model assumes specifically that cells going from glucose to maltose do have a lag phase. In a lag phase within environment k, cells of strain j have lag phase durations tau equal to a vector of lag phase durations drawn from experimentally determined distributions dist(τj,k).

62In the lag phase, cells of strain j only begin growth at steady-state growth rate μ(j,k) when t = τj,k(cellij).

63Each celli,j has an initial population size equal to 1. Thus, in the lag phase the population size N for the j th strain at time t is the sum of growth of all cells i.

64And the total populations size N at time t is the sum of all j competitors.

65The model was implemented as a stochastic simulation in R (functions and scripts available upon request). Time was iterated in glucose and maltose environments in 0.1-h increments. A total of 50,000 cells were chosen as initial population sizes for maltose lag phase simulations, and each cell was given a lag time tau and a status (1 or 0) that indicated whether this cell would begin growth within 24 h (the period of time for which we had data about lag time distributions for these strains). A normal distribution of tau was chosen for Isolate 1, and a uniform distribution for Isolate 6. After computing lag phase growth as a function of time in maltose, growth in the glucose environment was computed for each time point in maltose. The resulting matrices of population sizes for the two competitors were used to compute the proportion of each strain in the total population.

66In the text and in Figure 7, p(i) is the only parameter reported. In Figure 7A, the plot reflects the initial conditions of the experiment pictured in Figure 7B and 7C, where the average initial proportion of Isolate 1 relative to Isolate 6 for the six independently competing populations was 0.53317 and the cells had been in maltose for 1 h prior to time point 0. See Dataset S8 for the exact parameters used in the model.

67To begin the experiment where strains were competed in various carbon switching regimes, six independent replicate populations of Isolate 1, a strain constitutively expressing YeCitrine, and Isolate 6, a MALT-YeCitrine and MALS-mCherry strain, were inoculated from turbid YPD overnight cultures into glucose media and allowed to grow for 20 h until they had reached population densities of <5×106 cells per ml, such that both cultures would have lag phases upon transfer to maltose. The cultures were washed into maltose media, OD600 values measured, and then mixed 1∶1 and either kept in maltose or were transferred to glucose to generate the results shown in Figure 7B. This initial time point also served as the beginning of the glucose-to-maltose cycling populations shown in Figure 7C. Growth was at maximum 5×106 cells per ml, and minimal population sizes for long periods of glucose growth were ∼20,000 cells. At each time point, cells were centrifuged and frozen at −80°C in 1× PBS + 25% glycerol for later analysis at the flow cytometer. We used the same glucose and maltose YP media and volumes in this experiment as in the selection procedure.

68For analysis of competitions, 50,000 single-cell events were acquired by a BD Biosciences Influx flow cytometer. We used a 561 nm laser coupled to a 610/20 nm detector for mCherry, and for YeCitrine signal detection, we used a 488 nm laser coupled to a 580/30 nm detector. We used R's Flowcore package to identify subpopulations of unlabeled, YeCitrine-labeled, or mCherry-labeled cells using polygonal gates. We used control cultures of each competitor growing on its own in identical conditions to those of the competition to determine the fraction of events that were incorrectly determined to be one competitor when in fact they were from the other. In general these error rates were below 1/5,000 events.

69Selected samples were whole-genome sequenced using Illumina HiSeq 2000 with 500 bp inserted library. Quality assessment of resulted short reads was performed using FASTX-Toolkit (http://hannonlab.cshl.edu/fastx_toolkit/index.html). After removing the low-quality reads (below Q30) and adaptors, pair-end reads were then mapped onto the reference S. cerevisiae genome (S288C, version genebank64) using Burrows–Wheeler Alignment. Default settings were used except the maximum edit distance was set to 0.01 (−n 0.01). The MarkDuplicates command in Picard (http://picard.sourceforge.net/) was used to remove the reads that mapped to the same positions in the reference genome (PCR duplications). Consensus single-nucleotide variations (SNPs) and small insertions and deletions (Indels) were called for each chromosome using SAMtools and GATK,. Default settings were used, except the maximum read depth in SAMtools was set to 150× (−D 150). The generated SNPs and Indels were then filtered to minimize the false positive mutation calls. First, SNPs and Indels lying in low complexity sequences (such as telomeric, subtelomeric, transposon, repeat regions, etc.) were filtered out. Second, mutations with a total read depth below 20× were discarded. Third, SNPs and Indels with a quality score below 30 were removed. Fourth, mutation calls were only kept when at least 80% of the reads were positive for the SNP sites. Only the SNPs/Indels that were verified by both GATK and SAMtools were kept as confident sites. The lists of SNPs/Indels were then annotated by in-house Perl scripts with the yeastgenome database. CNV-seq was used to identify consecutive regions along the chromosome that show abnormal log2-ratios, which indicated the potential copy number variation (CNV). Only regions larger that 1 Kb were considered as CNV regions.

70Mutations identified by whole genome sequencing that lead to nonsynonymous or frame-shifted protein products in HXK2 and STD1 were first confirmed with dye-terminator Sanger sequencing. To test whether the mutations caused the observed phenotypes, we first integrated a dominant (KANMX) marker downstream of the ancestral and evolved alleles. This marker, including the upstream coding region containing the mutated or WT allele, was used as a template for PCR, which was then transformed via homologous recombination into the corresponding loci in the evolved or ancestral clones. The mutations were subsequently confirmed using Sanger sequencing, and the phenotypes of the genetically transformed strains was compared to that of strains bearing the same KANMX marker at the same locus, but lacking the mutation.