Genomics of cellular proliferation under periodic stress

Living systems control cell growth dynamically by processing information from their environment. Although responses to one environmental change have been intensively studied, little is known about how cells react to fluctuating conditions. Here we address this question at the genomic scale by measuring the relative proliferation rate (fitness) of 3,568 yeast gene deletion mutants in out-of-equilibrium conditions: periodic oscillations between two salinity conditions. Fitness and its genetic variance largely depended on the stress period. Surprisingly, dozens of mutants displayed pronounced hyperproliferation at short periods, identifying unexpected controllers of growth under fast dynamics. We validated the implication of the high-affinity cAMP phosphodiesterase and of a regulator of protein translocation to mitochondria in this control. The results illustrate how natural selection acts on mutations in a fluctuating environment, highlighting unsuspected genetic vulnerabilities to periodic stress in molecular processes that are conserved across all eukaryotes.

little is known about how cells react to fluctuating conditions. Here we address this question 23 at the genomic scale by measuring the relative proliferation rate (fitness) of 3,568 yeast gene 24 deletion mutants in out-of-equilibrium conditions: periodic oscillations between two salinity 25 conditions. Fitness and its genetic variance largely depended on the stress period. The control of cellular proliferation is essential to life and is therefore the focus of 47 intense research, but its coupling to environmental dynamics remains poorly characterized. In 48 addition, proliferation drives evolutionary selection, and the properties of natural selection in 49 fluctuating environments are largely unknown. Although experimental data exist 1,2 , they are 50 scarce and how mutations are selected in fluctuating conditions have mostly been studied 51 under theoretical frameworks 3-6 . Repeated stimulations of a cellular response may have 52 consequences on growth that largely differ from the consequences of a single stimulus. First, 53 a small growth delay after the stimulus may be undetectable when applied only once, but can 54 be highly significant when cumulated over multiple stimuli. Second, growth rate at a given 55 time may depend on past environmental conditions that cells 'remember', and this memory 56 can sometimes be transmitted to daughter cells 7 . These two features are well illustrated by the 57 study of Razinkov et al., who reported that protecting yeast GAL1 mRNA transcripts from 58 their glucose-mediated degradation resulted in a growth delay that was negligible after one 59 galactose-to-glucose change but significant over multiple changes 8 . This effect is due to short-60 term 'memory' of galactose exposures, which is mediated by GAL1 transcripts that are 61 produced during the galactose condition and later compete for translation with transcripts of 62 the CLN3 cyclin during the glucose condition. Other memorization effects were observed on 63 bacteria during repeated lactose to glucose transitions, this time due to both short-term 64 memory conferred by persistent gene expression and long-term memory conferred by protein 65 stability 9 . 66 67 The yeast response to high concentrations of salt is one of the best studied mechanism 68 of cellular adaptation. When extracellular salinity increases abruptly, cell-size immediately 69 reduces and yeast triggers a large process of adaptation. The translation program 10,11 and 70 turnover of mRNAs 12 are re-defined, calcium accumulates in the cytosol and activates the 71 calcineurin pathway 13 , osmolarity sensors activate the High Osmolarity Glycerol MAPK 72 pathway 13,14 , glycerol accumulates intracellularly as a harmless compensatory solute 14 , and 73 membrane transporters extrude excessive ions 13 . Via this widespread adaptation, hundreds of 74 genes are known to participate to growth control after a transition to high salt. What happens 75 in the case of multiple osmolarity changes is less clear, but can be investigated by periodic 76 stimulations of the adaptive response. For example, periodic transitions between 0 and 0.4M 77 NaCl showed that MAPK activation was efficient and transient after each stress except in the 78 range of ~8 min periods, where sustained activation of the response severely hampered cell 79 growth 15 . How genes involved in salt tolerance contribute to cell growth in specific dynamic 80 regimes is unknown. If a protein participates to the late phase of adaptation its mutation may 81 have a strong impact at large periods and no impact at short ones. It is also possible that 82 mutations affecting growth in dynamic conditions have been missed by long-term adaptation 83 screens. As mentioned above, a slight delay of the lag phase of adaptation may remain 84 unnoticed after a single exposure, but its effect would likely cumulate over multiple 85 exposures and be under strong selection in a periodic regime. Thus, even for a well-studied 86 system such as yeast osmoadaptation, our molecular knowledge of cellular responses may be 87 modest when dynamics are to be understood. 88

89
Although microfluidics enables powerful gene-centered investigations, its limited 90 experimental throughput is not adapted to systematically search for genes involved in the 91 dynamics of a cellular response. Identifying such genes can be done by applying stimulations 92 to mutant cells periodically and testing if the effect of the mutation on proliferation is 93 averaged over time. In other word, does fitness (proliferation rate relative to wild-type) of a 94 mutant under periodic stress match the time-average of its fitness in each of the alternating 95 condition? This problem of temporal heterogeneity is equivalent to the homogenization 96 problem commonly encountered in physics for spatial heterogeneity, where microscopic 97 heterogeneities in materials modify macroscopic properties such as their stiffness or 98 conductivity 16 . If fitness is homogeneous (averaged over time), it implies that the effect of the 99 mutation on the response occurs rapidly as compared to the frequency of environmental 100 changes, that is does not affect the response lag phase and that the mutated gene is not 101 involved in specific memory mechanisms. In contrast, fitness inhomogeneity (deviation from 102 time-average expectation) is indicative of a role of the gene in the response dynamics. 103

104
In this study, we present a genomic screen that addresses this homogenization problem 105 for thousands of gene deletion mutations in the context of the yeast salt response. The results 106 reveal how selection of mutations can depend on environmental oscillations and identify 107 molecular processes that unsuspectedly become major controllers of proliferation at short 108 periods of repeated stress. We measured experimentally the contribution of thousands of yeast genes on 115 proliferation in two steady conditions of different salinity, and in an environment that 116 periodically oscillated between the two conditions. We used a collection of yeast mutants 117 where ~5,000 non-essential genes have been individually deleted 17 . Since every mutant is 118 barcoded by a synthetic DNA tag inserted in the genome, the relative abundance of each 119 mutant in pooled cultures can be estimated by parallel sequencing of the barcodes (BAR-120 Seq) 18,19 . We set up an automated robotic platform to culture the pooled library by serial 121 dilutions. Every 3 hours (average cell division time), populations of cells were transferred to a 122 standard synthetic medium containing (S) or not (N) 0.2M NaCl. The culturing program was 123 such that populations were either maintained in N, maintained in S, or exposed to alternating 124 N and S conditions at periods of 6, 12, 18, 24 or 42 hours (Fig. 1A). Every regime was run in 125 quadruplicates to account for biological and technical variability. Duration of the experiment 126 was 3 days and populations were sampled every day. After data normalization and filtering 127 we examined how relative proliferation rates compared between the periodic and the two 128 steady environments. We observed that genes involved in salt tolerance during steady conditions differed in 133 the way they controlled growth under the periodic regime. As shown in Fig. 1B, differences 134 were visible both among genes inhibiting growth and among genes promoting growth in high 135 salt. For example, NBP2 is a negative regulator of the HOG pathway 20 and MOT3 is a 136 transcriptional regulator having diverse functions during osmotic stress 21,22 . Deletion of either 137 of these genes improved tolerance to steady 0.2M NaCl (condition S). In the periodic regime, 138 the relative growth of mot3Δ cells was similar to the steady condition N, as if transient 139 exposures to the beneficial S condition had no positive effect. In contrast, the benefit of 140 transient exposures was clearly visible for nbp2Δ cells. Differences were also apparent among 141 protective genes. The Rim101 pathway has mostly been studied for its role during alkaline 142 stress 13 , but it is also required for proper accumulation of the Ena1p transporter and efficient 143 Na + extrusion upon salt stress 23 . Eight genes of the pathway were covered by our experiment. 144 Not surprisingly, gene deletion decreased (resp. increased) proliferation in S (resp. N) for all 145 positive regulators of the pathway ( Fig. 1B and Fig1-supplement-1). This is consistent with 146 the need of a functional pathway in S and the cost of maintaining it in N where it is not 147 required. The response to periodic stimulation was, however, different between mutants 148 (Fig1-supplement-1). Although RIM21, DFG16 and RIM9 all code for units of the 149 transmembrane sensing complex 24 , proliferation was high for rim21Δ and dfg16Δ cells but not 150 for rim9Δ cells. Similarly, Rim8 and Rim20 both mediate the activation of the Rim101p 151 transcriptional repressor 25,26 ; but rim8Δ and rim101Δ deletions increased proliferation under 152 periodic stress whereas rim20Δ did not. This pathway was not the only example displaying 153 such differences. Cells lacking either the HST1 or the HST3 NAD(+)-dependent histone 154 deacetylase 27 grew poorly in S, but hst1Δ cells tolerated periodic stress better than hst3Δ cells 155 We then systematically asked, for each of the 3,568 gene deletion mutants, whether its 162 fitness in the periodic regime matched the time-average of its fitness in conditions N and S. 163 We both tested the statistical significance and quantified the deviation from the time-average 164 expectation. For statistical inference, we exploited the full BAR-seq count data, including all 165 replicated populations, by fitting to the data a generalized linear model that included a non-166 additive term associated to the fluctuations (see methods). The models obtained for the six 167 genes discussed above are shown in Fig. 1C. Overall, we estimated that deviation from time-168 average fitness was significant for as many as ~2,000 genes, because it was significant for 169 2,497 genes at a False-Discovery Rate (FDR) of 0.2 (Supplementary Table 1 Fitness is inhomogeneous when this ratio deviates from 1. Plotting the distribution of this 201 ratio at each period of fluctuation showed that, as expected, inhomogeneity was less and less 202 pronounced as the period increased ( Fig. 2A). We examined more closely three mutants 203 displaying the highest inhomogeneity at the 6h period. Plotting their relative abundance in the 204 different populations over the time of the experiment clearly showed that fitness of these 205 mutants was unexpectedly extreme at short periods but less so at larger periods (Fig. 2B). 206

207
The fact that some mutants but not all were extremely fit to short-period fluctuations 208 raised the possibility that the extent of differences in fitness between mutants may change 209 with the period of environmental fluctuation. To see if this was the case, we computed the 210 genetic variance in fitness of each pooled population of mutants (see methods). Fitness 211 variation between strains was more pronounced when populations were grown in S than in N, 212 which agrees with the known effect of stress on fitness differences 31 . Remarkably, differences 213 were even larger in fast-fluctuating periodic regimes, but not slow-fluctuating ones ( Fig. 2A). other. This phenomenon is a special case of gene x environment interaction and is called 221 antagonistic pleiotropy (AP) 28 . It is difficult to anticipate whether such mutations have a 222 positive or negative impact on long-term growth in a periodic regime that alternates between 223 favorable and unafavorable conditions, especially since fitness is not necessarily 224 homogenized over time. We therefore studied these cases in more detail. 225 226 First, we examined if fitness inhomogeneity was related to the difference in fitness 227 between the steady conditions ( Fig. 3A). Interestingly, gene deletions conferring higher 228 fitness in N than in S tended to be over-selected in the 6h-periodic regime, revealing a set of 229 yeast genes that are costly in standard laboratory conditions as well as in the fast-fluctuating 230 regime. We then searched for gene deletions that were advantageous in one steady condition 231 and deleterious in the other (AP deletions). We found 48 gene deletions with statistically-232 significant AP between the N and S conditions (FDR = 0.01, Supplementary Table 3, see  233 methods and Fig3-supplement-1). Interestingly, three of these genes coded for subunits of the 234 chromatin-modifying Set1/COMPASS complex (Supplementary Table 2  This may have important implications on the spectrum of mutations found in 259 hyperproliferative clones that experienced repetitive stress (see discussion). It is also 260 remarkable that the gene deletions displaying this effect were associated to various cellular 261 and molecular processes: cAMP/PKA (pde2Δ), protein import into mitochondria (tom7Δ), 262 autophagy (atg15Δ), tRNA modification (trm1Δ), phosphatidylcholine hydrolysis (srf1Δ) and 263 MAPK signalling (ssk1Δ, ssk2Δ); and some of these molecular functions were not previously 264 associated to salt stress. 265

266
The high-affinity cAMP phosphodiesterase and Tom7p are necessary to limit 267 hyperproliferation during periodic salt stress 268 269 As mentioned above, several gene deletions impairing the cAMP/PKA pathway 270 displayed inhomogeneous fitness (Supplementary Table 2). One of them, pde2Δ, had a 271 particularly marked fitness transgressivity (Fig. 4B). To determine if this effect truly resulted 272 from the loss of PDE2 activity, and not from secondary mutations or perturbed regulations of 273 neighboring genes at the locus, we performed a complementation assay. Re-inserting a wild-274 type copy of the gene at another genomic locus reduced hyperproliferation and fully abolished 275 fitness transgressivity (Fig. 4F). Thus, the observed effect of pde2Δ directly results from the 276 loss of Pde2p, the high-affinity phosphodiesterase that converts cAMP to AMP 32 , showing 277 that proper cAMP levels are needed to limit proliferation during repeated salinity changes. 278 279 Unexpectedly, we found that deletion of TOM7, which has so far not been associated 280 to saline stress, also caused fitness transgressivity in the 6h-periodic environment (Fig. 4C). We quantified the contribution of 3,568 yeast genes to cell growth during periodic salt 299 stress. This survey showed that for about 2,000 genes, fitness was not homogenized over 300 time. In other words, the observed fitness of these genes in periodic stress did not match the 301 time-average of the fitness in the two alternating conditions. This widespread and sometimes 302 extreme time-inhomogeneity of the genetic control of cell proliferation has several important 303 implications. 304 305 Novel information is obtained when studying adaptation out of equilibrium. 306 307 A large part of information about the properties of a responsive system is hidden at 308 steady state. For example, a high protein level does not distinguish between fast production 309 and slow degradation. For this reason, engineers working on control theory commonly study 310 complex systems by applying periodic stimulations, a way to explore the system's behaviour 311 out of equilibrium. Determining the frequencies at which a response is filtered or amplified is 312 invaluable to predict the response to various types of stimulations. Such spectral analysis can 313 sometimes reveal vulnerabilities, and it has also been applied to biological systems 36  below which prediction is challenging. We showed that for a given system (yeast tolerance to 337 salt) this limit differed between mutations. Future experiments that track the growth of 338 specific mutants in microfluidic chambers may reveal the bandwith of frequencies at which 339 GxE interactions take place. 340 341 It is important to distinguish a periodic stress that is natural to an organism from a 342 periodic stress that has never been experienced by the population (as considered here). In the 343 first case, populations can evolve molecular clocks adapted to the stress period. This capacity 344 is well known: nature is full of examples, and artificial clocks can be obtained by 345 experimental evolution of micro-organisms 37 . In the case of periodic stress, an impressive 346 result was obtained on nematodes evolving under anoxia/normoxia transitions at each 347 generation time. An adaptive mechanism emerged whereby hermaphrodites produced more 348 glycogen during normoxia, at the expense of glycerol that they themselves needed, and 349 transmitted this costly glycogen to their eggs in anticipation to their need of it in the 350 upcoming anoxia condition 38 . In contrast, when a periodic stress is encountered for the first 351 time, cells face a novel challenge. The dynamic properties of their stress response can then 352 generate extreme phenotypes, such as hyperproliferation, as described here (Fig. 2B, Fig. 4A suspension were transferred to each of the 4 source plates which were then incubated at 30°C 532 with 270 rpm for another 3 hours. Every 6 hours, cell density was monitored for one of the 533 four replicate plates by OD 600 absorbance. Every 24 hours, 120 µl of cultures from each 534 replicate plate were sampled, pooled in a single microplate, centrifuged 10 minutes at 5000g 535 and cell pellets were frozen at -80°C. Dilution rates of the populations were: 85% when the 536 action was only to replace the media, 55% when it was to replace the media and to measure 537 OD, and 32% when it was to replace the media, to measure OD and to store samples. 538 The experiment lasted 78 hours in total and generated samples from 28 independent 539 populations at time points 6h (end of initialization), 30h, 54h and 78h.  Table 4) were 583 considered to be pseudogenes or genes with no effect on growth, and the data from the 584 corresponding deletion mutants were combined and used as an artificial "wild-type" 585 reference. For each mutant strain M, w was calculated as : 586 with M b , M e , WT b , and WT e being the frequencies of strain M and artificial wild type strain 587 (WT) at the beginning (b) or end (e) of the experiment, and g the number of generations in 588 between. g was estimated from optical densities at 600nm of the entire population. It poorly 589 differed between conditions and we fixed g = 24 (8 generations per day, doubling time of 3h). 590 591 Deviation from time-average fitness. We analyzed fitness inhomogeneity by both 592 quantifying it and testing against the null hypothesis of additivity. The quantification was 593 done by computing = , where w observed was the fitness of the mutant strain 594 experimentally measured in the periodic environment and w expected was the fitness expected 595 given the fitness of the mutant strain in the two steady environments (N and S), calculated as 596 was the total variance, and 614 was an estimate of the non-genetic variance in fitness, with N being the number of gene 616 deletions, w i,j the fitness of gene deletion i in replicate j, ! the mean fitness of gene deletion i 617 and the global mean fitness. The 95% confidence intervals of V G were computed from 618 1,000 bootstrap samples (randomly picking mutant strains, with replacement). 619 620 Antagonistic Pleiotropy. We used the observed w N and w S values (fitness in the N and S 621 steady conditions, respectively) of the deletion mutants to determine if a mutation was 622 antagonistically pleiotropic (AP). Our experiment provided, for each mutant, 3 independent 623 estimates of w N and 4 independent estimates of w S (replicate populations). For each mutant, 624 we combined these estimates in 3 pairs of (w N , w S ) values by randomly discarding one of the 625 4 available w S values, and these pairs were considered as 3 independent observations. We 626 considered that an observation supported AP if the fitness values (w N , w S ) showed (1) an 627 advantage in one of the conditions and a disadvantage in the other, and (2) deviation from the 628 distribution of observed values in all mutants, since most deletions are not supposed to be AP. 629 Condition (1) corresponded to: ( w N > 1 AND w S < 1) OR (w N < 1 AND w S > 1). Condition 630 (2) was tested by fitting a bivariate Gaussian to all observed (w N , w S ) pairs and labelling those 631 falling 2 standard deviations away from the model (Fig3-supplement-1). A deletion was 632 considered AP if all 3 observations supported AP, which was the case for 48 deletions. A 633 permutation test (re-assigning observations to different deletions replicates) determined that 634 less than one deletion (0.54 on average) was expected to have three observations supporting 635 AP by chance only (Supplementary Table 3). For the selected 48 deletions, the magnitude of 636 AP was computed as w N / w S . For each deletion, the direction of selection (Fig. 3C)  (resp. ! ) was the mean fitness value in steady condition N and S, respectively, and ! (resp. 648 ! ) the corresponding standard deviation. A permutation test (re-assigning observations to 649 random mutants) determined that less than three mutants (2.24 on average) were expected to 650 display three replicates supporting transgressivity by chance only (Supplementary Table 5