DFEs across All Environments Show Many Wild-Type Like Mutations

The shape of the DFE indicates the relative importance of purifying or directional selection as compared with neutral evolution. In apparent contrast to the strong conservation of the Hsp90 middle domain in natural populations, we observed DFEs with mostly wild-type like mutations across all environments (supplementary figs. 2 and 3, Supplementary Material online). Throughout the manuscript, we use the term “wild-type like” to denote mutations that are indistinguishable from the wild-type reference in the limit of experimental accuracy, see Materials and Methods section. We categorized mutants as wild-type like if the 95% confidence interval (CI) of the estimated selection coefficient overlapped with 0 (see also supplementary table 1, Materials and Methods). According to this criterion, between 50% (in the H₂O₂ environment) and 65% (in the standard environment) of mutations showed a fitness effect that is indistinguishable from the reference type. Large numbers of wild-type like mutations have been observed previously in deep mutational scanning studies of DFEs (Soskine and Tawfik 2010; Hietpas et al. 2013; Melamed et al. 2013; Bank et al. 2014; Hom et al. 2019). Both biological and technical factors can be invoked as an explanation for the large number of wild-type like mutations. First, the resolution of the experiment is likely much lower than the resolution at which natural selection may act in large yeast populations (Ohta 1992; Boucher et al. 2016). Second, selection pressures in the laboratory might differ greatly from those in nature (Reznick and Ghalambor 2005; Kvitek and Sherlock 2013). Finally and relatedly, natural environments might be fundamentally different and rapidly fluctuating (Mustonen and Lässig 2009). For example, a recent study of the DFE of the full Hsp90 sequence found that mutations which were tolerated across a set of diverse environments were those most likely observed in natural sequences (Flynn et al. 2020). We repeated the analysis of Flynn et al. (2020) and found the same pattern: our estimated DFEs across all environments of the subset of variants observed in nature (amino acid mutations observed across 267 eukaryotic sequences, see Materials and Methods section) showed an elevated peak around 0. This indicates a further enrichment of wild-type like mutations in the subset of naturally observed variants as compared with the full data set (supplementary fig. 5, Supplementary Material online).

Our approach provides a dense scan of the local fitness landscape by measuring the selection coefficient of all amino-acid changing mutations that are available in a single mutational step from the ancestral state. The presence of a large number of wild-type like mutations in various different environments suggests that the local fitness landscape is rather flat, which is at odds with the strong conservation of the protein in yeast. However, further away from the wild-type, epistatic interactions may condition the following mutational steps and thereby change the configuration of the fitness landscape, creating fewer attainable mutational paths, and constraining evolution on longer time scales (Weinreich et al. 2006; Kryazhimskiy et al. 2014).

Correlations across Conditions Reveal That Diamide Stands Out with Respect to Mutational Effects

We next quantified the correlations of fitness effects across replicates and environments (fig. 2 and supplementary table 2, Supplementary Material online). Consistent with the high accuracy of the experiment, we observed strong correlations of estimated mutational fitness effects between replicates (supplementary fig. 6, Supplementary Material online, mean Pearson correlation r = 0.91). Across pairs of environments, we observed a large variation of correlations ranging from r = 0.48 between diamide and salt to r = 0.98 between H₂O₂ and the standard environment. In this analysis, diamide clearly stood out by showing consistently lower correlations of fitness effects with other environments than all others (mean correlation including diamide r = 0.65, mean correlation of all others r = 0.86). This could be because diamide exerts multiple negative effects on the yeast cells. For example, among the environments we investigated, diamide was the only condition that induces the expression of cell wall biosynthesis genes and genes involved in protein secretion and processing in the Endoplasmic Reticulum, which indicates its role in cell wall damage (Gasch et al. 2000). Diamide was also shown to affect individual transcription factors differently from H₂O₂, the other oxidative environment in our selection. One example that has been described in the literature is the Yap1 transcription factor, which is nuclear localized and active in diamide, but cytoplasmic and inactive in H₂O₂ (Gulshan et al. 2011). Interactions of Hsp90 with other genes may also contribute to fitness differences of the mutations between the tested environmental conditions and mutants that we tested. Further global experimental analyses beyond the scope of this work will be required to determine the molecular features of diamide stress that elicit distinct selection on Hsp90 sequence.

Fig. 2.

Histogram of correlations of fitness effects among replicates and environments reveals that diamide stands out as different. The histograms correspond to comparisons of replicates (A), comparisons that include diamide (B), and all other comparisons (C). Correlations between replicates and between pairs of environments are in general high, whereas correlations between diamide and the other environments tend to be lower. Altogether, this suggests that cells react in unique ways to the oxidative stress enacted by exposure to diamide, and that Hsp90 plays a specific role in the response to this stress.

Previous studies have reported that mutations had similar fitness effects across environments that shared metabolic features (Ostrowski et al. 2005; Dillon et al. 2016; Sane et al. 2018). Thus, we expected to see stronger correlations of fitness effects between environments that induce the same type of stress (salt vs. sorbitol and H₂O₂ vs. diamide) than between different types of stress. Indeed, we observed a strong correlation of the fitness effects of mutations between salt and sorbitol (r = 0.79) and between H₂O₂ and diamide (r = 0.74). However, these correlations are not much different from comparing any other pairs of environments (supplementary table 1). Again, diamide is the one environment that stands out with respect to this pattern. Here, the average correlation between diamide and other non-H₂O₂ environments is 0.62 (vs. r = 0.74 between H₂O₂ and diamide), suggesting that, in this case, H₂O₂ and diamide may share some metabolic features where Hsp90 is involved. One possible explanation for this is that the oxidative stresses of diamide and H₂O₂ cause an increase in the expression of chaperones, including Hsp90 (Gasch et al. 2000), inducing similar gene expression responses to those observed during thermal stress in yeast (Gasch et al. 2000).

We also observed a tendency for selection coefficients from the same batch of experiments to be more strongly correlated than selection coefficients inferred from different batches of experiments (supplementary table 2). This is not unusual (Venkataram et al. 2016) and may be due to small differences in the growth conditions between batches or during the sequencing steps. To mitigate this effect we performed the analyses to detect beneficial and deleterious hotspots either based on at least two replicates, or using the average of the estimated fitness effects between replicates (as indicated below) and recategorizing mutations based on the mutant category of both replicates (see also Materials and Methods section).

The large correlations observed between most environments suggest that potential costs of adaptation (discussed in detail further below) should be the exception rather than the rule. This is at odds with previous results from a smaller region of Hsp90 (Hietpas et al. 2013), where most beneficial mutations detected at high salinity were deleterious in other environments. A potential explanation for this discrepancy lies in the choice of the protein region. Whereas the previous study focused on only nine amino acids and thus chose a set of positions that are likely to show functionally specific patterns, we here had the experimental power to scan a larger region of the protein. Along a larger, less specifically chosen, stretch of the protein, it is likely that many positions and mutations play the same role across many environments. In this case, costs of adaptation of beneficial mutations could still be large, but due to their low proportion in the overall pool of observed mutations they barely affect the correlation of selection coefficients across environments. Indeed, when computing the correlation in a ten amino acid sliding window, we see high variability in the correlations (supplementary fig. 7, Supplementary Material online), for example with a range between 0.36 and 0.84 for the correlation between Diamide and Standard environments. This suggests that specific regions may represent functionally important positions (see below).

Beneficial Mutations Are Present across All Environments and Are Enriched in a Region of Hsp90 That Is Implicated in Stabilization of Interdomains and Client-Binding Interfaces

Next, we identified the proportion and identity of putatively beneficial mutations across all treatments (supplementary table 1 and figs. 8 and 9, Supplementary Material online). We categorized mutants as putatively beneficial if the lower limit of the 95% CI of the selection coefficient was larger than 0, and we considered as strongest candidates those that overlapped between replicates from the same environment. As reported above, wild-type like mutations are the most abundant category in all environments, with a low proportion but considerable number of beneficial mutations in almost all environments (supplementary figs. 8 and 9, Supplementary Material online). We found the lowest number of beneficial mutations in H₂O₂ and salt environments (n_H2O2=65 and n_Salt=64, around 2.5% of the mutations) and the highest in diamide (n_Diamide=307, 12.3%).

The proportion of beneficial mutations, interpreted in the light of Fisher's Geometric Model (Fisher 1930; Tenaillon 2014), should be informative about the “harshness” of the experimental environments and the resulting potential for adaptation. Specifically, Fisher imagined that populations evolve in a multidimensional geometric phenotype space, where the same mutation is more likely to be beneficial if it happens in an individual that is far from the phenotypic optimum (Fisher 1930; Tenaillon 2014). Thus, we expected that in environments with large doubling times, corresponding to lower absolute fitness of the reference type (supplementary fig. 1, Supplementary Material online), we should observe a larger number of beneficial mutations. Moreover, we expected the lowest number of beneficials in the standard environment, to which the wild-type is well adapted and in which it has the lowest doubling time. However, this expectation was met only in the diamide environment. At high salinity and in sorbitol, the observed proportion of beneficial mutations was lower than in the standard environment. This discordance with expectations from FGM could be due to the model's very general assumptions which, for example, include that mutations affect all phenotypic dimensions with equal weight (see Harmand et al. 2017). Many of these assumptions are likely violated here, because, for example, Hsp90 plays a very different functional role in the osmotic stress response in comparison with diamide-induced oxidative stress—probably resulting in different phenotypic distributions of the same mutations in the (anyways rather abstract) phenotype space.

Specifically, it is known that Hsp90 basal function together with its co-chaperone Cdc37 is required for the induction of the osmotic stress response in yeast via activation of the Hog1 kinase in the HOG pathways (Hawle et al. 2006; Yang et al. 2007). However, increased salinity or sorbitol do not cause an increase in Hsp90 mRNA expression in comparison with diamide exposure, which causes an increase in mRNA expression similar to what is observed during heat shock (Gasch et al. 2000). The decoupling of function and expression, and the low number of beneficial mutations suggest that Hsp90's role during osmotic stress may be related to the activation of the general stress response mechanism of the cell (Mager and Siderius 2002). Another possibility is that osmotic stress does not cause heat shock-induced protein unfolding (Richter et al. 2010) or diamide-induced protein modifications that may compromise protein and cellular function (Gasch et al. 2000). Thus, under osmotic stress conditions the cell would not require increased expression of Hsp90. This suggests that mutations may play distinct roles for different conditions, perhaps under osmotic stress activation of clients and under diamide could be activation of clients and/or increase in Hsp90 expression and function.

Interestingly, beneficial mutations in diamide were dispersed across the whole middle domain (supplementary fig. 8, Supplementary Material online), whereas most of the other environments showed a clear enrichment of beneficial mutations in the region of positions 356–366 (fig. 3 and supplementary fig. 8, Supplementary Material online). This hotspot of beneficial mutations was particularly strong in sorbitol (supplementary fig. 8, Supplementary Material online). Out of the 31 beneficial mutations that showed a beneficial effect in at least 2 environments, 18 were located in this hotspot.

Fig. 3.

Hotspots of beneficial and deleterious mutations in the middle domain do not coincide with known client-binding loop. (A) Average proportion of beneficial and deleterious mutations along the genome across all environments, illustrated using a ten-amino-acid sliding window. The region with the 10% largest proportions of beneficial mutations is highlighted in light orange (positions 356–366), the region with the 10% largest proportions of deleterious mutations is highlighted in light blue (positions 347 and 376–385). A known client-binding loop (positions 327–341) is highlighted in light gray. To avoid biases, the analysis was restricted to subsets of two replicates of all environments. (B) Structural representation of the Hsp90 middle domain (1HK7.PDB) illustrating the beneficial hotspot in magenta with residue 364 highlighted in cyan. Biochemical, structural, and mutational studies identified positions 327–341 as a client-binding loop, which is shown in dark blue. The beneficial hotspot that was identified in our analyses is adjacent to the above mentioned client-binding loop. (C) Structural representation of the ATPase and middle domains of Hsp90 from 2CG9.PDB. The deleterious hotspot (in cyan) contains a catalytic amino acid and is located at a dimerization interface. It includes residue R380 that stabilizes the leaving phosphate during ATP hydrolysis. Subunits are distinguished with different shading. (D) The deleterious hotspot at position 347 is also located at the interface of the ATPase domain and the middle domain and is highlighted in pink and blue.

The beneficial hotspot contains residues that are part of an allosteric center which mediates distinct conformations in structures with and without client (Blacklock and Verkhivker 2013, 2014; Czemeres et al. 2017). Despite its different structural arrangements, the hydrophobic amino acids in this region are mostly buried from solvent (fig. 3B), indicating that the primary role of this region is to mediate the stability and rearrangement of different conformations. One position that particularly stands out with respect to the number of beneficial mutations is 364. Position 364 lies at the center of the cluster of the 10% largest average proportions of beneficial mutations (fig. 3A) and shows a clear deviation from the usually observed shape of the DFE (supplementary fig. 10, Supplementary Material online). Specifically, we identified six mutations at this position that are beneficial in at least two environments (supplementary fig. 10, Supplementary Material online). The burial of the large hydrophobic side chain of F364 should provide local stability. The beneficial mutations at position 364 and in this region in general suggest that disruption of local stability and conformational dynamics may alter Hsp90 function. Because this beneficial region partially overlaps with an allosteric center involved in Hsp90 conformational stability and dynamics (Blacklock and Verkhivker 2013, 2014) we hypothesize that Hsp90 function may be altered by the identified candidate beneficial mutations in a manner that disrupts local stability and dynamics of this region. Furthermore, these disruptions may promote conformational changes in the neighboring client-binding loop that allow this region to sample a larger and/or different conformational space in a manner that changes the relative affinity and thereby priority of different clients—a property that we speculate could provide benefits in specific conditions. Consistent with this hypothesis, the beneficial hotspot is adjacent to a known client-binding loop (fig. 3) such that alterations in the structure or dynamics of the hotspot are likely to influence client-binding nearby. It is possible that natural selection balances client priorities integrated over multiple conditions, which provides opportunities for Hsp90 mutations to improve priorities for individual conditions (Flynn et al. 2020).

The Proportion of Deleterious, Wild-Type Like, and Beneficial Mutations Varies Greatly along the Protein Sequence

We showed above that the overall correlations of fitness effects across environments were generally large, which indicates similar effects of the same mutations across environments. In contrast and similar to the local correlations of fitness effects (supplementary fig. 7, Supplementary Material online), the proportions of beneficial, wild-type like, and deleterious mutations vary greatly along the protein sequence. Whereas the overall pattern is similar between environments, the relative proportions differ between environments (supplementary figs. 8 and 9, Supplementary Material online). Our finding that the same positions are enriched for deleterious and beneficial mutations across environments suggests that the structural properties of the middle domain in these regions, rather than its binding partners, might be the most important factor for predicting the fitness effects of mutations.

Interestingly, the regions with the largest proportions of deleterious mutations are located near the ends of the beneficial hotspot region (positions 347 and 376–385, respectively) (supplementary fig. 8, Supplementary Material online). At the latter deleterious hotspot around 70% of all mutations are deleterious in a sliding window of ten amino-acid positions. Such a shift of the DFE towards deleterious mutations suggests that this protein region is under strong purifying selection. Structurally, the region 376–385 is part of a catalytic loop required for ATP hydrolysis which is necessary for the activation of all known clients (Wolmarans et al. 2016). Residue R380 in the catalytic loop binds to and stabilizes the leaving phosphate of ATP (fig. 3C) and mutations at this position compromise Hsp90 function and cell viability (Meyer et al. 2003). The efficiency of catalysts depends strongly on geometry, and the precise location of R380 relative to ATP is likely tightly linked to Hsp90 function. The regions adjacent to R380 appear to be important for positioning the catalytic arginine, providing a rationale for the strong purifying selection that we infer. The second deleterious hotspot at position 347 is also located at the interface of the ATPase domain and the middle domain (fig. 3D), consistent with current understanding that Hsp90 mechanism requires precise ATP-dependent interactions between these domains (Schopf et al. 2017).

Few Mutations Show Costs of Adaptation

Among the identified beneficial mutations, we were specifically interested in those that are deleterious in other environments, which results in a so-called “cost of adaptation.” The same phenomenon has also been termed antagonistic pleiotropy. In previous work, we reported a large prevalence of costs of adaptation in a nine-amino-acid region of Hsp90 (Hietpas et al. 2013). In this study, we found that costs of adaptation were not pervasive (fig. 4 and supplementary table 3, Supplementary Material online). Consistent with the comparatively low correlations of fitness effects and the special role of the diamide stress discussed above, we observed the largest proportion of mutations that show costs of adaptation between diamide and other environments. Whereas the mean proportion of mutations displaying costs of adaptation across all comparisons of environments was 14.3%, comparisons that included diamide showed on average 22.7% mutations with costs of adaptation.

Fig. 4.

Histogram of proportions of mutations that display costs of adaptation illustrates that costs of adaptation are generally rare, and most likely if comparisons involve diamide environments. The histograms correspond to comparisons of replicates (A), the subset of other comparisons that include diamide (B), and all other comparisons (C).

Mapping the proportion of beneficial mutations that are deleterious in other environments along the protein sequence, we observed a hotspot for costs of adaptation at amino-acid positions 381–391 (fig. 5). Interestingly, this region showed costs of adaptation across various environments, which indicates that each environment has specific beneficial mutations which are deleterious in other environments. Structurally, this region also belongs to the catalytic loop involved in ATP hydrolysis discussed above. Indeed, the identified region partly overlaps with the above-discussed hotspot of deleterious mutations (positions 376–385), which is unsurprising since a larger proportion of deleterious mutations also increases the (statistical) proportion of mutations that can be classified as costly.

Fig. 5.

A hotspot of costs of adaptation is located at positions 381–391. The 10% positions with the largest mean proportion of mutations that display costs of adaptation (dark purple line) are highlighted in light purple. This region greatly overlaps with the hotspots of deleterious mutations from figure 3 (highlighted in blue). Thin curves indicate the mean proportion of beneficial mutations in a focal environment that is deleterious in another environment. A ten-amino-acid sliding window was used to locate and display region-specific effects. To avoid biases, the analysis was restricted to subsets of two replicates of all environments.

We next computed whether there was a correlation between the effects of the subset of beneficial mutations in one environment with their effect in the other environments. Indeed, the effects of beneficial mutations in diamide were negatively correlated with the same mutation's effect in all other environments (fig. 5 and supplementary fig. 11, Supplementary Material online) except H₂O₂. In other words, the more beneficial a mutation was in diamide, the more deleterious it tended to be in salt, standard, and sorbitol. Again, this suggests that the beneficial mutations we identified in diamide may be involved in the response to a very specific type of stress. Across all other environments, we did not see any evidence that stronger beneficial mutations tended to have a more deleterious effect in other environments. In fact, there is a suggestive positive correlation between fitness effects of beneficial mutations with their respective effect in all other nondiamide environments, which points to the presence of synergistic pleiotropy. Although generally defined as mutations having the same effect on more than one trait, in the context of this study synergistic pleiotropy would mean that beneficial mutations tended to have a similar ranking across environments. This is in line with results from studies of Escherichia coli using experimental evolution (Ostrowski et al. 2005; Dillon et al. 2016) or mutation accumulation (Sane et al. 2018). These studies reported synergistic pleiotropy between different carbon source environments, and rare presence of antagonistic pleiotropy. Specifically, Sane et al. (2018) found that this pattern was maintained also for the categories of neutral and deleterious mutations, and that mutations with larger effect were more likely to show antagonistic pleiotropy. The reported prevalence of synergistic pleiotropy was associated with a sharing of the metabolism and transport of resources (Ostrowski et al. 2005; Dillon et al. 2016; Sane et al. 2018).

Generating Point Mutations

To accurately measure the fitness effects of all possible point mutations in a large portion of the middle (client-binding) domain of yeast Hsp90, we used saturation mutagenesis at positions 291–409. We used a cassette ligation strategy to introduce mutations as previously described (Hietpas et al. 2012). This strategy reduces the likelihood of secondary mutations because it avoids the potential for errors during PCR steps. As a control for the mutational procedure, 12 positions were randomized in isolation and Sanger sequenced to assess the level of incorporation of all four nucleotides at each position in the target codon. At all randomized nucleotide positions within these 12 samples, we observed a similar magnitude of signal for each of the four nucleotides. For larger scale production, libraries were generated at ten positions at a time as previously described (Hietpas et al. 2012). As additional controls, we generated a sample containing individual stop codons as well as the parental wild-type Hsp90 sequence. All variants of Hsp90 were generated in a plasmid (pRS414GPD) previously shown to produce endogenous levels of Hsp90 protein in yeast (Chang and Lindquist 1994).

Constructing Barcoded Libraries

To improve the efficiency and accuracy of fitness estimates, we added barcodes to a nonfunctional region of the plasmid, ∼200 NTs downstream from the 3′ untranslated region of Hsp90. Barcodes were introduced using a cassette ligation strategy (supplementary fig. 12, Supplementary Material online). Plasmid libraries were treated with NotI and AscI to generate directional sticky ends. NotI and AscI recognize 8-base cut sites that are unlikely to cut any of the Hsp90 variants in the library. We designed and annealed barcode forward and barcode reverse oligos together such that the resulting duplex product included a central N18 region bracketed by constant regions that facilitate annealing and overhangs that direct directional ligation into NotI and AscI overhangs. One of the constant ends in the designed oligo cassette contained an annealing region for an Illumina sequencing primer. Barcoded libraries were transformed into E. coli and pooled into a bulk culture that contained about 10-fold more transformants than Hsp90 variants in the library. We purified barcoded plasmids from this bulk culture. This procedure resulted in approximately 20 barcodes for each Hsp90 codon variant in the library (supplementary fig. 13, Supplementary Material online). The potential diversity in the N18 barcode that we used (4¹⁸∼10¹¹) far exceeds the number of barcodes that we utilized (∼64*119*10∼10⁵), which makes it likely that each Hsp90 variant receives a barcode that differs from all other barcodes at multiple nucleotides. With this setup, errors in sequencing of barcodes can be detected and eliminated from further analysis, which reduces the impact of sequencing misreads on estimates of variant frequency and fitness. Additional controls consisting of individual stop codons and wild-type Hsp90 were barcoded separately. For these controls, we isolated barcoded plasmid DNA from individual bacterial colonies and determined the barcodes by Sanger sequencing (supplementary fig. 12, Supplementary Material online).

Associating Barcodes to Mutants

To identify the barcodes associated with each Hsp90 variant in our libraries, we used a paired-end sequencing approach essentially as previously described (Hiatt et al. 2010, see also supplementary fig. 14, Supplementary Material online). Using paired-end sequencing on an Illumina MiSeq Instrument barcodes were associated with variant genotypes via reading from the Hsp90 gene with a 250 bp read and the associated N18 barcode with a 50 bp read. To reduce the size of the DNA fragments for efficient Illumina sequencing, we removed a portion of the  Hsp90 gene such that the randomized regions were closer to the N18 barcode. This was done to increase the density of DNA on the sequencer by reducing the radius of clonal clusters during sequencing. Plasmid DNA was linearized with StuI and NotI endonucleases that removed ∼400 bp’s of DNA. The linearized products were circularized by blunt ending followed by ligation at DNA concentrations that predominantly lead to unimolecular ligations (Revie et al. 1988). The resulting DNA products were amplified using a PE2_F primer and the standard Illumina PE1 primer that anneals next to the N18 barcode. Two PE2_F primers were designed to read across the region of Hsp90 that we randomized. PCR products were gel purified and submitted for paired end sequencing. We obtained sufficient paired-end reads such that the average barcode was read more than ten times (supplementary fig. 13, Supplementary Material online). The paired-end sequencing data was subjected to a custom data analysis pipeline to associate Hsp90 variants with barcodes. First, very low-quality reads with any Phred score less than 10 in reads 1 or 2 were discarded. Next, the data were organized by the barcode sequence. For barcodes with three or more reads, we constructed a consensus of the Hsp90 sequence read. We compared the consensus sequence to the parental Hsp90 sequence to determine mutations. Consensus sequences containing more than one protein mutation were discarded. The pipeline output generated a file organized by point mutation that lists all barcodes associated with that mutation (supplementary data). This file was used as the basis for calling a variant based on barcode reads.

Yeast Competitions

As in previous work (Hietpas et al. 2012, 2013), we performed Hsp90 competitions using a shutoff strain of Saccharomyces cerevisiae (DBY288) (supplementary fig. 14, Supplementary Material online). The sole copy of Hsp90 in DBY288 is driven by a galactose-inducible promoter, such that the strain requires galactose for viability and cannot grow on glucose. The introduction of a functional Hsp90 variant driven by a constitutive promoter rescues the growth of DBY288 in glucose media. We introduced the library of middle domain variants of Hsp90 driven by a constitutive promoter into DBY288 cells using the lithium acetate method (Gietz and Schiestl 2007). The efficiency of transformation was more than 10-fold higher than the number of barcodes, such that the average barcode was transformed into more than ten individual cells. To enable the analyses of full biological replicates, transformations were performed in replicate such that a separate population of yeast was transformed for each biological replicate. Transformed yeast were initially selected in RGal-W (synthetic media that lacked tryptophan and contained 1% raffinose and 1% galactose supplemented with 100 μg/ml ampicillin to hinder bacterial contamination) to select for the plasmid, but not function of the plasmid encoded Hsp90 variant. This enabled us to generate a yeast library containing Hsp90 variants that could support a full range of fitness from null to adaptive. Cells were grown for 48 h at 30 °C in liquid SRGal-W until the culture was visibly opaque compared with a control transformation sample that lacked plasmid DNA but was otherwise identical to the library sample.

To initiate the competition, cells were transferred to shut-off conditions and different environmental stresses. To deplete the pool of wild-type Hsp90 protein, the library sample was diluted into SD-W (synthetic media lacking tryptophan with 2% glucose and 100 μg/ml of ampicillin) and grown at 30 °C for 10 h. After depletion of wild-type Hsp90 protein, the cells were split into different stress conditions in SD-W media at 30 °C: salt stress (0.5 M NaCl), osmotic stress (0.6 M sorbitol), oxidative stress (0.6 mM hydrogen peroxide or 0.85 mM diamide), as well as a control nonstress condition. We used a custom built turbidostat (supplementary fig. 14, Supplementary Material online) to provide constant growth conditions during this competition phase of the experiment. Cells were grown under consistent density and population size. 10⁹cells were maintained in a 50 ml volume over the course of 40 h for each condition. Rapid magnetic stirring was used to provide aeration to the media and cells. Samples of 10⁸cells were collected after 0, 4, 8, 12, 24, 32, 40, and 48 h of competition in the different conditions. At the time of collection, samples were centrifuged and the pellets immediately frozen at −80 °C.

Sequencing of Competition Samples

The fitness effects of Hsp90 variants were estimated based on frequencies observed in next-generation sequencing analyses essentially as previously described (Hietpas et al. 2012). DNA was isolated from each timepoint sample as described (Hietpas et al. 2013) and the barcodes were amplified and sequenced. Barcodes were amplified using the standard Illumina PE1 primer that anneals next to the N18 barcode and custom designed barcode_forward primers. A set of barcode_forward primers were designed with identifier sequences that could be read during sequencing and used to distinguish each timepoint sample (supplementary fig. 14C, Supplementary Material online). Each identifier sequence was eight nucleotides in length and differed by at least two bases from all other identifier sequences. Twenty cycles of PCR were sufficient to generate clear products from all samples. These PCR products were purified on silica columns (Zymo Research) and sequenced for 100 base single reads on an Illumina NextSeq instrument. The barcode corresponded to the first 18 bases and the identifier at positions 91–98. The resulting fastq files were processed and analyzed using customized software tools. First, poor quality reads containing any positions with a Phred score less than 20 were discarded. Reads were tabulated if the barcode matched a barcode associated with a point mutation and if the identifier matched with a timepoint. The analyses scripts output a file with the number of reads for each amino acid point mutant at each timepoint in each condition (supplementary data). Mutants with 50 or less total reads at the first time point were removed from the data set.

Estimation of Selection Coefficients

Inference of selection coefficients was performed via log-linear regression as described in Matuszewski et al. (2016) (see also supplementary material). To improve estimation accuracy by incorporating information from each individual barcode, the linear model for each amino-acid changing mutation included barcode identities as nominal variables. For each amino-acid changing mutation we obtained a selection coefficient and 95% CI of the estimate (supplementary data 2), representing the variation within amino acid due to differences in codon*Barcode tag throughout time. This reduces the impact of potential outliers and averages over synonymous codons within amino acid. Finally, we normalized all selection coefficients by subtracting the median of all mutations that were synonymous to the wild-type. This ensures that the average of the selection coefficients of wild-type synonyms represents a selection coefficient of 0. To categorize mutations as beneficial, deleterious, or wild-type we tested the overlap of the CI with 0. Namely, if the lower CI limit was larger than 0, a mutation was considered beneficial, and if the upper CI limit was smaller than 0, a mutation was considered deleterious. All other mutations were considered wild-type like. We use this terminology to distinguish between neutral and wild-type like mutations. Specifically, we expect that at large population sizes the number of wild-type like mutations is larger than the number of neutral mutations. This is because neutral mutations should behave neutrally with respect to the neutral or nearly neutral theories, that is, when 2Ns < 1, where s denotes the selection coefficient and N the population size (Ohta 1992). Considering the large experimental population sizes of yeast, this threshold for neutrality is far below the measurement accuracy of our experimental setup. Therefore, we expect that neutral mutations are a subset of the wild-type like mutations. Thus, we instead consider as “wild-type like” those mutations that are indistinguishable from the wild-type according to our inference accuracy.

Natural Variants

We extracted observed amino acid mutations from Hsp90 sequences of 261 eukaryotic organisms (Starr et al. 2018). Then, for each environment, we compared the DFEs of the identified natural variants with the DFE from all mutations. Finally, we also computed the overall proportion of beneficial, deleterious, and wild-type like mutations across environments and for the natural variants.

Costs of Adaptation

To compute the costs of adaptation for figures 4 and 5, we first selected all mutations in a focal environment that were categorized as beneficial. We then computed the proportion of this subset of mutations that was deleterious either in one other environment (for fig. 4) or across all other environments (excluding replicates of the focal environment for fig. 5). Calculating the proportion of mutations with costs of adaptation compensates for variable numbers of beneficial mutations across environments, which may occur due to both biological and experimental reasons (e.g. different error margins). For figures 2 and 4, all available data sets were considered for the analysis. For figures 3 and 5, only two replicates of each environment were considered. That was done to avoid biases due to the under- or overrepresentation of environments in the analysis.

Analyses were performed using R version 3.5.1 (R Core Team 2018), Mathematica 12.0.0.0, Python 2.7.12 and Pymol 0.99. Code is available as supplementary material.