Stochastic tuning of gene expression enables cellular adaptation in the absence of pre-existing regulatory circuitry

Cells adapt to familiar changes in their environment by activating predefined regulatory programs that establish adaptive gene expression states. These hard-wired pathways, however, may be inadequate for adaptation to environments never encountered before. Here, we reveal evidence for an alternative mode of gene regulation that enables adaptation to adverse conditions without relying on external sensory information or genetically predetermined cis-regulation. Instead, individual genes achieve optimal expression levels through a stochastic search for improved fitness. By focusing on improving the overall health of the cell, the proposed stochastic tuning mechanism discovers global gene expression states that are fundamentally new and yet optimized for novel environments. We provide experimental evidence for stochastic tuning in the adaptation of Saccharomyces cerevisiae to laboratory-engineered environments that are foreign to its native gene-regulatory network. Stochastic tuning operates locally at individual gene promoters, and its efficacy is modulated by perturbations to chromatin modification machinery.

The capacity to adapt to changes in the external environment is a defining feature of living systems. Cells can rapidly adapt to familiar changes that are commonly encountered in their native habitat by sensing the parameters of the environment and engaging dedicated regulatory networks that have evolved to establish adaptive gene expression states (Jacob and Monod, 1961; Thieffry et al., 1998). However, dedicated sensory, signaling, and regulatory networks become inadequate, or even detrimental, when cells are exposed to unfamiliar environments that are foreign to their evolutionary history (Tagkopoulos et al., 2008). In principle, at least one gene expression state that maximizes the health/fitness of the cell always exists, despite the inability of the native regulatory network to establish such a state. This is true because under any conceivable environment, the activities of some genes are beneficial, whereas those of others are futile or even actively detrimental (Jacob and Monod, 1961; Tagkopoulos et al., 2008; Hottes et al., 2013). In fact, if the initial fitness defect is not lethal, a population of cells may slowly adapt to an unfamiliar environment through the accumulation of genetic mutations that rewire regulatory networks, thereby achieving more optimal gene expression states (Tagkopoulos et al., 2008; Applebee et al., 2008; Philippe et al., 2007; Goodarzi et al., 2010; Tenaillon et al., 2012; Rodríguez-Verdugo et al., 2016; Blount et al., 2012; Van Hofwegen et al., 2016; Damkiær et al., 2013).

In this work we speculate whether cells have evolved alternative strategies for finding adaptive gene expression states, on more physiological timescales, without relying on their hard-coded sensory and regulatory systems. Since the perception of the external world may be of limited value under unfamiliar conditions, perhaps a more effective strategy would be to focus on maximizing the internal health of the cell—without regard to the specific parameters of the outside world. This would be a challenging strategy, as every gene in the genome would need to independently reach the expression level that maximizes the overall health of the cell, and these expression levels could vary significantly from condition to condition. In particular, we asked whether individual genes could, in principle, carry out a search process equivalent to gradient descent (Cauchy, 1847), where the health consequence of stochastic alterations in gene expression could gradually tune the expression of individual genes towards a level that is optimal for internal health. We reasoned that such an optimization process would require the existence of: (1) a source of stochastic transitions in gene expression; (2) the ability of local chromatin to maintain a record of recent changes in transcription; and (3) a central metabolic hub that integrates diverse parameters of intracellular health and continuously broadcasts whether the overall health of the cell is improving or deteriorating.

In fact, we find that the foundations for meeting these requirements are already present in eukaryotic cells: (1) The expression of many genes is dominated by noisy bursts of transcription—a widespread phenomenon of largely unknown functional significance (Sanchez and Golding, 2013; Raj and van Oudenaarden, 2008; Blake et al., 2006; Raser and O'Shea, 2005; Elowitz et al., 2002); (2) Co-transcriptional histone modification can modify eukaryotic chromatin in promoters and gene bodies, establishing a short-term memory of recent transcriptional events (Li et al., 2007; Rando and Winston, 2012); and (3) Global integrators of cell health have evolved in eukaryotes. A classic example is the mTOR pathway, which integrates a vast array of intracellular parameters reflecting nutrient availability, energy, and the presence of diverse stresses (Conrad et al., 2014; González and Hall, 2017; Albert and Hall, 2015; Saxton and Sabatini, 2017).

With the necessary components for gradient-based optimization of gene expression in place (Figure 1A), the promoter of each gene would be able to conduct a simple search process that culminates in finding the expression level that maximizes the overall health of the cell: if global fitness/health is increasing and there was a previous increase in transcriptional output (representing larger or more frequent transcriptional bursts), the promoter further increases its transcriptional activity (Figure 1B). If fitness is decreasing and there was a previous increase in transcriptional activity, the promoter decreases its transcriptional output. Transcriptional output is altered in the opposite direction in the event that there was a previous decrease in transcriptional output. For each gene, this tuning process can be expressed as: ΔEt = k∙sgn(ΔFt∙ΔEt-1)+η (see Figures 1 and 2A); here, E denotes the vector of gene-level transcription rates, F the current fitness/health of the cell, k is a proportionality constant, η a noise term, and sgn is a function yielding −1 if its argument is negative, 0 if its argument is zero, and +1 if its argument is positive. One can easily see how the process described here can tune the optimal expression of a single gene. What is remarkable, however, is the ability of this hypothetical stochastic tuning process to find near-optimal gene-expression states for a system with thousands of genes. As can be seen in the simulations presented in Figure 2, this is achieved through a fitness-directed stochastic search culminating in individual genes reaching specific gene expression levels that maximize the health/fitness of the cell.

Such a stochastic tuning mechanism would be highly valuable to free-living organisms, enabling them to optimize their global gene expression patterns to match the specific requirements of any environment in which their dedicated sensory and regulatory networks are inadequate or sub-optimal.

Informed by the simulations above, we sought to test for evidence of stochastic tuning in the eukaryotic model organism Saccharomyces cerevisiae. We engineered conditions in which the expression of a single gene was required for growth, but for which no regulatory input existed to drive appropriate expression levels. This was achieved by using a yeast strain (BY4743) that lacks the URA3 gene, which is essential when cells are grown in the absence of uracil. We placed a chromosomally integrated copy of URA3 at a different locus under the control of a weak synthetic promoter, consisting primarily of a pseudorandom sequence. All recognizable binding sites for native transcription factors were removed from the generated promoter sequence (see Materials and Methods and Supplementary file 1 for details), in an attempt to decouple it from any existing sensory and regulatory input. We henceforth refer to this synthetic promoter sequence as synprom (see Supplementary file 1 for sequence). In the experiments described below, URA3 is typically tagged with a fluorescent fusion, either mRuby (Kredel et al., 2009) or a superfolder GFP (Pédelacq et al., 2006), and a copy of a mouse DHFR gene coupled to a different fluorescent protein is inserted at the same location on the sister chromosome to act as an internal control. A schematic of the insertion constructs is shown in Figure 3A. We also added the URA3 competitive antagonist 6-azauracil (6AU) to the media to control the threshold level of URA3 production required for growth. The growth condition, SC+glu-ura media, containing x µg/ml of 6AU, will henceforth be referred to as ura-/6AUx.

Even with the challenging and specific experimental layout described here, with growth highly dependent on URA3 expression, we expect that stochastic tuning might contribute to fitness through mechanisms acting in cis at the promoter driving URA3, those acting in trans through modulation of factors that (despite our best efforts) weakly affect the promoter driving URA3, and through tuning of unrelated pathways that benefit survival and growth in the –URA condition. Nevertheless, URA3 expression itself will clearly be the key driver of growth since it is the critical bottleneck for nucleotide biosynthesis in the absence of uracil supplementation.

To look for evidence of fitness-directed stochastic tuning, we tracked the colony formation of cells containing synprom-driven URA3 after plating on ura-/6AU15 plates. Lacking sufficient URA3 expression to overcome high 6AU levels, these non-growing cells would be expected to succumb to starvation and die. Remarkably, however, after prolonged incubation we observed apparently stochastic transitions to rapid growth, leading to the formation of macroscopic colonies over time (Figure 3B). We eventually observed colony formation by roughly one cell in 103, a rate too high to be driven by mutation-driven adaptation in the absence of growth.

The synthetic promoter referred to as ‘synprom’ throughout the text is the combination of a pseudorandom sequence with a small natural promoter-proximal region taken from the SAM3 gene, with both stripped of all recognizable matches to known transcription factor binding sites (see Materials and Methods for details). We also tested all combinations of five other synthetic promoter sequences and one other promoter proximal region, enumerated in Supplementary file 2. As shown in Figure 3—figure supplement 1, four of the six synthetic promoters support stochastic tuning, and the ability of synprom5 (the purely artificial component of the synprom referred to in the remainder of the text; see Supplementary file 2 for all synthetic promoter sequences) to undergo tuning remains even with a different promoter proximal region. These findings highlight the universality of the observed tuning phenomenon and minimize the possibility that our observations actually arise due to the presence of some residual sequence-specific transcription factor binding site present in synprom.

As shown in Figure 3C, we also observed similar tuning behavior for two high-noise natural promoters, PHSP12 and PRGI1 (Tirosh et al., 2009; Tirosh et al., 2006), indicating that stochastic tuning can function even when superimposed on naturally evolved regulatory sites. Across all promoters (natural and synthetic) tested here, the observed tuning rates, relative to the number of viable plated cells, varied from 1 in 101 (PHSP12) to 1 in 105 (synprom5-arf1).

The apparently stochastic nature of colony formation in our experiments is reflected both in the steady emergence of colonies over the course of days or weeks (Figure 3B–C and Figure 3—figure supplement 1), and in the wide variance of colony sizes observed on ura-/6AU15 plates (Figure 3D). Microscopy revealed that cells remain quiescent for days before transitioning to URA3 expression and rapid growth, with a transition rate dependent on the choice of promoter (Figure 3E). Furthermore, the change that enables growth under the ura-/6AU15 condition must be passed from mother to daughter cells, as colonies expand from a few points of initiation instead of showing random division of cells throughout the microscopic field over time. While the presence of some deterministic process, yielding colony formation over the observed timescales (dependent on the initial state of each cell), cannot be ruled out, a far simpler explanation for the observed phenomenon of a long lag followed by appearance of colonies over a wide range of times is that each cell independently undergoes a random process that can eventually lead to growth. We confirmed that the appearance of colonies is not simply due to aging of the plates; 6AU-containing plates which were pre-incubated for a week or longer prior to plating of cells showed no change in colony formation rates (data not shown).

To provide further insights into the regulatory changes occurring during the onset of cell growth, we performed flow cytometry time courses on cells challenged by, and subsequently growing in, liquid ura-/6AU5 media, using cells with synprom-driven URA3-mRuby, and with a DHFR-GFP fusion driven by either the constitutive ADH1 promoter (Figure 4A–B) or synprom (Figure 4C–D) itself. The use of PADH1 to drive the second reporter allows us to control for extrinsic noise and global changes in gene expression, whereas coupling synprom to the non-beneficial DHFR-GFP fusion allows us to test whether the observed stochastic tuning is driven by any trans-acting input from some existing regulatory network or whether it is truly specific to the allele needed for growth, as required by our proposed tuning model.

Several patterns in the growth curves and flow cytometry data are immediately apparent. First, as with the agar-based growth discussed above, cells show a lag of at least 72 hr with undetectable growth, followed by the onset of steady growth until saturation. In the case of URA3-mRuby driven by synprom and DHFR-GFP by the constitutive promoter PADH1, URA3-mRuby fluorescence increases substantially in tandem with the onset of cell growth, and expression subsequently remains high until saturation; in contrast, DHFR-GFP signals do not even recover to their initial levels (Figure 4B; compare dashed and solid line distributions). This demonstrates that the URA3 induction resulting in growth is promoter-specific and does not simply reflect a general increase in protein expression. We observed qualitatively equivalent behavior when URA3 was driven by PRGI1 or PHSP12 (Figure 4—figure supplement 1). Even more strikingly, for cells with synprom driving both fluorescent fusions, we observed a specific enhancement of URA3-mRuby expression over that of DHFR-GFP (Figure 4D), showing that the transition to high URA3 expression is not only promoter-specific but allele-specific, and thus must be driven at least partly by changes occurring in cis at the specific locus whose expression is required for growth. As an additional test, we performed quantitative RT-PCR experiments to measure the ratio of URA3 and DHFR mRNA expression in tuned cells either in liquid ura-/6AU5 media or on ura-/6AU15 plates (see Figure 4—figure supplement 2).

In both cases, we observed a substantial increase in the URA3:DHFR ratio in the tuned cells, indicating that the observed tuning occurs at least partly through a local cis-acting process at the locus required for growth (although we cannot rule out additional changes in other promoters that also contribute to survival and growth, which may account for the observed heterogeneity in expression levels between replicates). Consistent with our proposed tuning model, the allele-specific nature of the transcriptional induction supports a key role for a local tuning process that is independent of dedicated sensory and regulatory input.

The presence of the competitive URA3 inhibitor 6-azauracil allows us to vary the threshold level of URA3 required for growth. Thus, it is instructive to consider how the concentration of 6AU may alter stochastic tuning behavior, both in the context of the computational model described above and in the actual behavior of the system. We made two crucial modifications to the numerical model employed in Figure 2 to mimic our experimental setup. First, rather than having the entire gene expression profile begin far from the optimal point, we begin with all genes but one (representing URA3) at their optimal values, reflecting the fact that aside from the artificial stress of lacking appropriate URA3 regulation, the cells’ native regulatory network can provide an appropriate response to ura-/6AU media. Second, we note that due to the presence of the competitive inhibitor 6AU, the URA3 in the cell will not even be able to contribute meaningfully to nucleotide biosynthesis (and thus impact the cell’s health/fitness) until it passes a threshold level. Thus, the tuning term (Figure 2A) is not applied to the gene representing URA3 until after the concentration of URA3 passes a threshold. Aside from the modifications noted above, we model tuning in the ura-/6AU environment as we did for the general case in Figure 2, and in particular, the fitness effects of changing URA3 expression must compete with noisy gene expression from the other 999 genes in the model gene expression profile to impact the direction of tuning.

The resulting URA3 expression profiles during simulated tuning in the presence of low or high concentrations of 6AU are shown in Figure 5A. In the low 6AU case, the tuning mechanism pushes URA3 expression almost deterministically to its optimal (high) value, whereas in the presence of high 6AU, the URA3 expression level undergoes a random walk until expression becomes high enough to allow the tuning mechanism to ‘sense’ the gradient and drive the cells into a URA3+ state. The effects on tuning rates of varying the 6AU concentration are plotted in Figure 5B, where we observe that increasing 6AU concentrations both slow tuning and dramatically increase the variance in the amount of time required for each individual cell to reach a URA+ state. This is precisely the behavior observed experimentally with high 6AU concentrations (Figure 3). On the other hand, tuning in our experimental system switched from slow and stochastic to rapid and deterministic in the presence of low 6AU concentrations, with observable tuning occurring over the course of a few hours (Figure 5C). Importantly, the tuning process is confined to the URA3-mRuby allele, despite the fact that DHFR-GFP is also being driven by the same synthetic promoter. This again demonstrates that the tuning process occurs independently of conventional gene regulation by dedicated sensory and regulatory input.

We utilized time-lapse fluorescence microscopy to monitor the correspondence between expression of URA3-mRuby and cell division in PHSP12-URA3-mRuby/PADH1-DHFR-GFP cells that initiated the tuning process. Consistent with our proposed tuning model, gene expression fluctuations that surpassed a threshold for alleviating the URA3 deficit were reinforced over long timescales and were sustained (inherited) across multiple generations as the tuned colony expanded (Figure 6). As expected, there is no accompanying increase in DHFR-GFP. Similar trajectories were observed for other tuning micro-colonies (Figure 7A). The apparently long autocorrelation time of URA3-mRuby fluctuations through the duration of a tuning trajectory is consistent with our proposed fitness feedback reinforcement mechanism. In order to quantitatively determine the timescale of gene expression fluctuations, also known as mixing times (Sigal et al., 2006), we utilized fluorescence-activated cell sorting (FACS) to sort a population of cells for the bottom 20%, the top 20%, and complete (mock-sorted) distribution of URA3-mRuby expression and measured the timescales over which the sorted fluorescence distributions converged to each other (Figure 7—figure supplement 1). For cells growing under uracil-replete conditions (SC+glu), we observed a relatively fast mixing time on the order of ~100 min (Figure 7—figure supplement 1; Supplementary file 3). On the other hand, cells starving in ura-/6AU10 media had mixing times that ranged from 400 to 1200 min (Figure 7—figure supplement 1; Supplementary file 3).

To determine the association of URA3-mRuby levels across generations with growth, we primed cells with 12 hr of exposure to ura-/6AU5 media and then tracked the division of tuned vs. untuned microcolonies of PHSP12-URA3-mRuby/PADH1-DHFR-GFP cells over 24 hr time courses in ura-/6AU5 media. By comparing the fluorescence of cells that are about to divide with those that are not, we found that dividing cells have significantly higher levels of mRuby than non-dividing cells, whereas the separation was much smaller for GFP (Figure 7C). Furthermore, the URA3-mRuby levels within the tuning colony were highly heritable; as seen in Figure 7D, as the indicated colony tunes and grows, cells within that colony maintain a high-mRuby state through subsequent divisions, and even their internal rankings are mostly preserved. mRuby levels in other, non-tuned microcolonies are almost uniformly lower than cells in the tuned colony. The fitness-driven optimization component of our model (Figure 1) further predicts that fluorescence levels should not only be heritable, but also that cells will continue to increase URA3 expression (possibly noisily) until they reach either a local optimum fitness or some biological constraint on maximum gene expression. Consistent with our expectation, we observed that the ratio of mRuby to GFP levels (the latter of which is fused to a gene whose product is not needed for growth) became steadily higher in cell lineages that had been dividing for longer (Figure 7E). These observations demonstrate that the level of URA3 expression is correlated with fitness, is transmitted across several generations, and shows an ongoing upward trend in tuned cells over the course of time.

That last finding is particularly important because a directed increase in URA3 once a lineage begins growing is predicted by our model for fitness-directed tuning, but cannot be explained by other competing hypotheses. The images and data shown in Figure 7 were taken for colonies within a single field of view of a 40x objective to ensure internal consistency in illumination and normalization, but their behavior is representative of our observations across multiple such windows. (e.g., Figure 7—figure supplement 2, panel A). Similar quantitative analysis from another experiment beginning directly from growth in SC+glu (instead of short-term pregrowth in ura-6AU media) is shown in Figure 7—figure supplement 2, panels B-D.

It is crucial to exclude the possibility that genetic mutations underlie the observed tuning transition on –ura/6AU plates. The ongoing emergence of the tuned state in non-growing cells, over the course of many days, makes mutational mechanisms unlikely. In addition, as seen by microscopy (Figure 3E), no more than 1–3 cell divisions occur prior to the onset of sustained growth in a small fraction of cells.

Nevertheless, given the phenomenon of stress-induced mutagenesis in non-growing bacterial cells (Al Mamun et al., 2012), we wished to conclusively exclude any possibility of mutational mechanisms. To this end, we note that changes in URA3 expression occurring due to mutations should be stably heritable in the progeny of the tuned cells, which would not be expected to revert to a URA3 low state even after restoration of uracil in the media. To test the reversibility of the URA3 high state, we designed an experimental setup in which tuned colonies isolated from ura-/6AU plates were grown for varying numbers of passages in uracil-replete media (SC+glu including uracil) and then re-exposed to uracil starvation (Figure 8—figure supplement 1). If any genetic mutation were responsible for increasing URA3 expression in the tuned cells, the phenotype should be stable for many generations. On the other hand, stochastic tuning would predict that cells revert to a naïve state following sufficient growth in uracil-containing conditions, as they no longer benefit from URA3 expression. As seen in Figure 8A, cells with synprom-driven URA3 show reversion toward the naïve colony formation rates upon growth in (uracil containing) SC+glu media, with recovery apparent even after a single round of growth on an SC+glu plate, and subsequently becoming stronger with additional SC+glu passages.

To conclusively exclude mutational mechanisms, we performed untargeted whole-genome re-sequencing of a total of eight isolates with synprom-driven URA3-mRuby (four colonies from 6AU15 plates and four separate biological replicates taken after the onset of growth in 6AU5 liquid media; see Materials and Methods for details). For each case, we scanned the region within 25 kb of the LEU2 locus (where the URA3 cassettes were integrated) for mutations, since control of URA3 expression was shown in these cells to operate locally in cis (Figure 4 and Figure 4—figure supplement 2). The results are summarized in Supplementary file 4: Of the eight isolates, five show no mutations within 25 kb of the URA3-mRuby insertion, two show SNPs of unknown fitness contribution in a minority of the population, and one shows a duplication of the URA3-mRuby cassette (based on the presence of a read density that is twice the level observed elsewhere for the same chromosome). These data clearly indicate that the origin of growth-supporting URA3 expression levels in these cells cannot be reliant on a mutational mechanism, as only one of the eight cases – that with the URA3 duplication -- shows a mutation at high enough levels in the population to explain the onset of growth (mutations present in less than half of the population must have arisen after one or more cells in the population had already tuned and began growing, and thus by definition could not be responsible for the initial onset of the growing state). The phenotypes caused by the sequence variants observed in populations C2 and L4 are not immediately obvious, but even if they are beneficial, their presence in a minority of cells excludes the possibility that they were responsible for the onset of tuning.

Note that it should not be surprising (and, indeed, would be expected) that beneficial mutations might arise in a population once it had begun expanding in a new environment due to stochastic tuning. Our findings are consistent with a non-genetically heritable basis for the observed tuning in seven out of eight of the cases examined, as in all other growing lines, mutations near the URA3 gene were either non-existent or present only in a minority of the population.

A formal possibility for colony formation in a subset of the population is that growth occurs solely on the basis of pre-existing URA3 levels in cells prior to being exposed to uracil deprivation. Microscopic observations of starving cells (Figure 3E) argue against this possibility, as a substantial lag passes before any cells begin sustained growth. Also, colony formation continues over the course of many days (Figure 3B–D), demonstrating that even cells that were non-growing for a substantial time period after exposure to URA- stress can eventually grow under this condition. Nevertheless, to conclusively discount the possibility of pre-existing URA3 levels determining tuning, we sorted populations of cells on the basis of initial URA3 expression, isolated those with the highest mRuby levels (the top 0.5–1%, well outside of the main distribution of the population) and plated them. These experiments clearly showed that the ability to form colonies on ura-/6AU plates is not restricted to cells with initially high URA3-mRuby expression (Supplementary file 5), as the highly fluorescent cells do not form colonies on ura-/6AU plates at rates substantially higher than unsorted cells, and certainly not at a sufficiently higher rate to fully explain the observed colony formation rates. These data argue against the possibility that growth occurs only in cells that, by chance, already have high levels of URA3 expression at the time of plating (although such cells may have some slight advantage, given the nature of their initial state).

The proposed fitness-directed tuning mechanism relies on the capacity of local chromatin to maintain a memory of recent changes in transcription, and to modulate the transcription rate based on the fitness consequences of those changes, as conveyed by the proposed central metabolic integrator of health/fitness. We hypothesized that chromatin modification machinery may be intimately involved in these processes.

To probe the mechanistic basis of stochastic tuning, we focused on perturbations to histone acetylation/deacetylation (deletions of GCN5, SIN3, HST3, HST4), and chromatin remodeling (deletions of ASF1, ISW2, SWR1, UBP8), all of which provide potential pathways for coupling feedback from the cell’s physiological state to allele-specific modulation of chromatin and transcription (See Table 1 for details). We selected these targets because of their association with genes showing particularly high levels of noise (and thus, more likely to be driven by tuning) in single-cell proteomic analysis (Newman et al., 2006). In our screening, homozygous replacements of HST3, HST4, SWR1, ISW2, and UBP8 with a kanMX cassette showed little effect on colony formation rates on ura-/6AU plates, and SIN3::kanMX/SIN3::kanMX strains showed severely compromised cell survival under growth-arrested conditions; all were excluded from further analysis. On the other hand, we found that genetic perturbations to the histone acetylation machinery through deletion of the key histone acetyltransferase GCN5 essentially abolished tuning, whereas deletion of the histone chaperone ASF1, in contrast, increased tuning rate by more than an order of magnitude (Figure 8B). At the same time, we show that the observed tuning process does not rely on transcriptional memory mechanisms grounded in chromatin localization, given the lack of effect of a NUP42 deletion (Figure 8B; cf. (Guan et al., 2012)).

In interpreting our data on the effects of genetic perturbations on tuning (Figure 8B), it was crucial to consider the possibility that cells may lose viability at variable rates under different conditions, which could contribute to the observed differences in colony formation rates. We thus performed experiments to measure the rate of cell death in the presence of uracil starvation and compared the results with the different colony formation rates observed. As shown in Figure 8—figure supplement 2, the effects of a mutation on survival and tuning rates are not significantly correlated. For example, deletion of GCN5 resulted in the nearly complete loss of stochastic tuning, deletion of NUP42 had no effect, and deletion of ASF1 substantially enhanced tuning, yet none of these mutations shows a change in survival rates during incubation in uracil-free media compared with wild type cells sufficient to explain the observed change in colony formation rate (Figure 8—figure supplement 2). Even for the poorest surviving strain, GCN5::kanMX/GCN5::kanMX, colony formation rates after ten days are 100−1000 times lower than wild type cells even though survival rates are lower only by a factor of ten.

Given the apparent importance of chromatin modifications in fitness-directed tuning, we also tested the effects of nicotinamide treatment (which inhibits the sirtuin class of histone deacetylases, or HDACs (Bitterman et al., 2002)) on reversion of the tuned cells back to a naïve state. As shown in Figure 8A, we found that chemical inhibition of sirtuin HDACs by nicotinamide treatment substantially accelerated the decay of a tuned population to the naïve state, further highlighting the importance of histone modification in stochastic tuning. Combined with the data on knockout strains described above, our results suggest a central role for chromatin modifications in the establishment and maintenance of the tuning process, although the molecular details cannot be discerned from these data alone.

The abstract model introduced in Figures 1–2 demonstrates the potential utility of fitness-directed stochastic tuning to establish adaptive gene expression states without directly sensing the external environment. In order to substantiate the biological feasibility of stochastic tuning, we implemented its critical components in a plausible simulation incorporating generic features of chromatin modification and the information flow of the Central Dogma of Molecular Biology. We therefore designed and simulated a dynamical model tracking transcription rates, transcript levels, protein levels, and histone modifications in a single cell, with parameter distributions sampled from experimental data (Figure 9A; see Methods for details). We incorporated the possibility of adding or removing chromatin marks that can alter the transcription rates of the associated genes. Our model incorporates two classes of marks: tuning marks (T), which link cellular fitness to transcriptional output by having mark addition rates that are a function of the recent direction of change in global fitness and current number of such marks at each promoter; and stabilizing marks (S), which are added at a rate dependent on the number of tuning marks at each promoter (Figure 9B). At any time, the transcriptional output of the promoter is a function of the density of both tuning marks and stabilizing marks. As such, the tuning marks provide a critical connection between changes in global fitness and transcription rates, whereas the more slowly changing stabilizing marks capture the average transcriptional output over longer timescales, enabling a more stable optimization trajectory. Both T and S chromatin marks come in two varieties: positive (activating) and negative (repressive).