A Bayesian MCMC Approach to Assess the Complete Distribution of Fitness Effects of New Mutations: Uncovering the Potential for Adaptive Walks in Challenging Environments

EMPIRIC experimental procedure and environmental challenges

Library construction, bulk competitions, and sequencing were performed as previously described (Hietpas et al. 2013). Briefly, codon substitution libraries for the amino acid 582–590 region of yeast Hsp90 (Hsp82) were transformed into the DBY288 shutoff strain (can1-100 ade2-1 his3-11,15 leu2-3,12 trp1-1 ura3-1 hsp82::leu2 hsc82::leu2 ho::pgals-hsp82-his3) of S. cerevisiae by the lithium acetate method (Gietz and Woods 2002). The yeast library was amplified in nonselective galactose (Gal) medium with 100 μg/ml ampicillin (1.7 g yeast nitrogen base without amino acids, 5 g ammonium sulfate, 0.1 g aspartic acid, 0.02 g arginine, 0.03 g valine, 0.1 g glutamic acid, 0.4 g serine, 0.2 g threonine, 0.03 g isoleucine, 0.05 g phenylalanine, 0.03 g tyrosine, 0.04 g adenine hemisulfate, 0.02 g methionine, 0.1 g leucine, 0.03 g lysine, 0.01 g uracil per liter with 1% raffinose and 1% galactose) for 12 hr at 30° and transferred to selective dextrose medium (identical to Gal medium but raffinose and galactose replaced with 2% dextrose) at 30° for 8 hr to begin shutoff of the wild-type copy of Hsp90 . After 8 hr, the library culture was split into six distinct environmental conditions (25°, 30°, 36°, 25° + S, 30° + S, and 36° + S, where S represents the addition of 0.5 M sodium chloride), and the cultures were grown for 12–20 generations with samples taken at distinct time points and stored at −80°. Wild-type generation time was calculated by monitoring the increase in OD₆₀₀ of a monoculture of DBY288 cells harboring the wild-type Hsp90 construct under each environmental condition. Yeast lysis, DNA isolation, and preparation for Illumina sequencing were performed as previously described (Hietpas et al. 2012). Sequencing was performed by the University of Massachusetts deep sequencing core facility and generated ≈30 million reads of 99% confidence at each read position as judged by PHRED scoring (Ewing and Green 1998; Ewing et al. 1998).

Timescaling and outlier detection

To account for the slower growth of S. cerevisiae in stressful environments, we scaled the data such that sampling times are specified in hours. In other words, we measure the growth rate per hour instead of per wild-type generation (as commonly done), which results in a less intuitive interpretation of the growth rates but at the same time allows for a more consistent comparison of growth rates across environments. Missing data points are replaced with zeros.

Similar to previous studies, 10 of the randomized codons that result in internal HpaII sites (endonuclease was used in processing samples) were removed from further analysis.

To correct for sequencing errors at single sampling points that can, for example, arise from amplification errors during the sequencing procedure (for some example trajectories, see Supporting Information, Figure S1), we performed an outlier-detection step on the basis of an individual mutant’s trajectory. We acknowledge that some fraction of these removed data points could reflect biologically interesting non-log-linear effects; however, it is likely that the majority are errors and so their removal is necessary. In Results and Discussion, we show that performance is optimized when using the log ratio of the mutant’s read number to the total number of reads at each individual time point. These ratios serve as the basis for an outlier detection using the DFBETA statistic, using a cutoff of 2 (Belsley et al. 1980).

Model underlying the MCMC and estimation procedure

Each mutant i ∈ {1, … , K} in the culture has an initial population size c_i at the beginning of the experiment and is assumed to grow exponentially at a constant rate r_i (cf. Figure S2). The true abundance of an exponentially growing mutant in the population, N_i(t), follows the equation $N_i\left(t\right)=c_i{e}^{r_{i}t},$ where c_i is the (true) population size of that mutant at time 0, r_i is the growth rate of a particular mutant i, t ∈ {1, … , T} is the sampling time point in hours, and T is the number of sampling time points.

At each sampling point and for each mutant i, sequencing reads are drawn from a multinomial distribution with parameters n_t (total number of reads at time t) and {p₁, p₂, p_k}, where $p_i\left(t\right)=c_i{e}^{r_{i}t}/{\displaystyle {\sum }_{k=1}^K{c}_k{e}^{r_{k}t}}$ is the relative amount of mutant i in the population after t generations of exponential growth at rate r_i. We assume that each sampling point t represents an independent drawing experiment of n_i,t reads for each mutant i from the total amount of reads ${\displaystyle {\sum }_{i=1}^K{n}_{i,t}}=n_t.$ Hence, the overall likelihood is the product of the individual likelihoods, $L\left(c,r|\text{Data}\right)={\displaystyle {\prod }_{t\in T}L\left(c,r,t|\left\{n_{1,t},n_{2,t},\dots ,n_{K,t}\right\}\right)}.$ For each mutant, we estimate its initial population size c_i and the corresponding growth rate r_i. However, since we measure relative abundance of each mutation, only relative sizes and growth rates can be estimated, which reduces the effective number of estimated parameters by 2. Sampling points that were detected as outliers were excluded from the analysis for the respective mutant and time point. To preserve the overall read number for the multinomial sampling (and hence the multinomial probabilities), the sum of these dropped counts was added in an additional row for each time point at the end of the data set. This creates a set of at most T additional parameters that account for all outliers of the respective time point.

To allow for arbitrary distributions of growth rates and initial population sizes, we use improper priors for all estimated parameters that give equal prior probabilities to all attainable values of r_i and c_i (i.e., r_i ∈ ℝ⁺ and c_i ∈ ℕ), but the method allows for the implementation of any kind of prior assumptions. We implemented a Metropolis–Hastings algorithm in R to compute the stationary distribution of the MCMC. The variances of the proposal distributions were set such that the acceptance ratio was kept between 20 and 30%. A burn-in phase of 5,000,000 accepted parameter combinations and a subsequent estimation phase of 10,000,000 accepted values ensured convergence of the MCMC for the high number of parameters that are estimated simultaneously. Subsampling of every 1000th parameter combination resulted in the data sets used for further analysis. The R package “coda” was used to confirm convergence and mixing of the MCMC.

During the estimation phase, the wild-type sequence was normalized to growth rate r = 1 in every generation and its population size to 10,000—both values representing an arbitrary choice that does not affect the outcome of the estimation. A corrected mean of the growth rates of all mutations synonymous to the wild type is later used as a normalization constant, and all growth rates are rescaled accordingly. This approach improves the reproducibility of fitness estimates from full experimental replicates (cf. Results and Discussion). The original wild-type sequence is excluded from further analysis. The corrected mean is determined as the mean growth rate of those mutants that are synonymous to the wild type (excluding the original wild type itself) and that are within the mean plus or minus two standard deviations from the mean of the set of all mutants synonymous to the wild-type sequence. This accounts for potential outliers of the distribution of synonyms (cf. Results and Discussion). After this procedure, 560 single-nucleotide mutations (in all environments) remain for further analysis.

Test data sets

To evaluate experimental noise, the power of the MCMC, and the reliability of the outlier detection, we created 800 simulated data sets according to the above-described model with varying amounts of artificial noise. Distributions of fitness effects and initial population sizes were drawn from the empirical distributions of the respective parameters obtained from the 30° and 30° + S data sets to represent a fast- and a slow-growing sample data set. A distribution of artificial wild-type-synonym growth rates was drawn from a normal distribution with mean 1 and variance obtained from the empirical distribution of wild-type synonyms. Of note, we are not able to detect any sequencing errors that lie within the range of the distribution that arises from multinomial sampling. Therefore, we obtained the expected mean sampling number at a time point, m_t, and its standard deviation σ. We then introduced outliers drawn as random values from a log-uniform distribution in the interval [1/10m_t, 10m_t], excluding the interval of the expected mean sampling number at a data point plus or minus four times its standard deviation (i.e., [m_t − 4σ, m_t + 4σ]). We used four different probabilities of outliers, f = 0, 0.01, 0.05, 0.1. Outliers were assigned independently for each data point. We then performed the outlier detection on all data sets, using the DFBETA statistic, and ran the MCMC for 50 simulated data sets for 30° + S both with and without outlier detection for f = 0, 0.01, 0.05.

Evaluation of the MCMC approach

We evaluated two different methods to perform a linear regression as a basis for outlier detection, utilizing the simulated data sets detailed in Materials and Methods. First, we based the regression on log ratios of mutant to wild-type reads (in the following, we abbreviate this approach to WT), which is the commonly used method (e.g., Hietpas et al. 2011)—however, we hypothesized that potential noise in the wild-type measurement that transfers to all measurements could cause deviations in the estimates. For comparison, we based the regression on log ratios of mutant reads to the sum of all reads for that time point (abbreviated by TOT). Compared with the former procedure, this introduces an error that is due to the nonexponentiality of a sum of exponentially growing individuals with different growth rates. Figure 1 and Figure 2 (as well as Figure S3 and Figure S4) show that this error is small compared with the potential noise that sampling variance and mismeasurement of the wild-type sequence introduce into each measurement—which suggests that for similar next-generation sequencing data sets containing many mutants with individually low read numbers, the log ratio of mutant-to-total read number (or to another, ideally large subset of read numbers) provides much better estimates than the commonly used mutant-to-wild-type ratio.

Furthermore, we tested three commonly used outlier detection methods, based on the residuals of a linear regression of the above-described log ratios: a modified z-score, the DFBETA statistic, and Cook’s distances. These were chosen to represent different levels of complexity and differently conservative statistics. Whereas the modified z-score (MODZ) is solely based on the residuals of a simple linear regression, both DFBETA and Cook’s distances (COOK) are based on the change in parameters upon rejecting individual data points. Hereby, DFBETA calculates the effect on the estimated parameters (i.e., slope and intercept of the linear regression) separately, whereas COOK computes an aggregate measure of influence of a data point. As suggested in the literature (Belsley et al. 1980), we chose cutoffs of 3.5 for MODZ, 2 (as a conservative cutoff for small data sets) for DFBETA, and 4/n (where n is the number of sampling points) for COOK.

We classified the fraction of true positives (TP) (i.e., sampling points correctly classified as outliers), false positives (FP) (i.e., sampling points wrongly classified as outliers), true negatives (TN) (i.e., sampling points correctly classified as not being an outlier), and false negatives (FN) (i.e., true outliers that were not identified as such), as well as the sensitivity and the specificity for each of the regression methods combined with each of the detection methods. The results are shown in Figure 1 and Figure S3.

Next, we evaluated the root mean square error (RMSE) of growth rate estimates based on a linear regression with different outlier detection methods and under different normalization (cf. Figure 2 and Figure S4). Figure 2A shows that even if no outliers are present in the data, the outlier detection methods DFBETA and COOK improve the estimates. This owes to the fact that they filter a part of the sampling noise from the multinomial sampling. However, because not many real data points are discarded using DFBETA (cf. fraction of false positives in Figure S3), we consider this effect unproblematic. Regardless, the regression based on TOT demonstrates much better results than WT, likely owing to the inability to detect outliers in the wild-type sequence reliably. For the same reason, it is favorable to use a normalization of growth rates to the distribution of synonyms instead of a normalization to the wild-type sequence, as is shown in Figure 2B. Although the median improvement in the RMSE stays close to 0, the quality of estimates can, in some cases, increase by up to 80% if the normalization is based on the distribution of synonyms. Naturally, and apart from the above-described filtering of sampling noise, the knowledge of the true outliers is favorable in the case of many outliers (cf. Figure S4B).

We also evaluated the fraction of mutations for which the 95% credibility interval of the MCMC estimate includes the true growth rate (Figure 3), comparing no outlier detection with the DFBETA detection method and the knowledge of the true outliers (notably a case that is unachievable for real data). Whereas the performance is similar if no outliers are present in the data set, there is a steep decline in the proportion of well-estimated growth rates if 1 or 5% of outliers are present and no outlier detection is conducted. At the same time, the performance with outlier detection is almost as good as if true outliers are known. However, we note that, even if the estimation of growth rates from EMPIRIC bulk competition experiments is on average very accurate, statements about single mutations should ideally be reinforced by experimental replicates or separate competition assays.

In conclusion, all obtained results point toward DFBETA in combination with the regression approach TOT for outlier detection and a normalization based on the distribution of synonyms instead of the wild-type sequence only being the most reliable data processing steps that should be undertaken to filter sequencing errors.

Evaluation of the DFE in six environments

We next applied our novel MCMC approach to characterize changes in the DFE across six environmental conditions. Following Hietpas et al. (2011), deep sequencing of systematic mutant libraries was used to evaluate the fitness of 560 point mutants in a nine-amino-acid window of Hsp90. Hietpas et al. (2013) evaluated four environmental conditions and performed analyses only at the amino acid level based on log-linear regressions. Thus, we here add two additional environmental conditions, perform analyses based on the nucleotide level, and implement a Bayesian MCMC approach to incorporate uncertainty into our estimates based on an exponential growth model.

Hsp90 is a well-studied chaperone that aids in protein folding, especially at high temperatures (Borkovich et al. 1989). As such, the experimental treatments included three different temperatures (25°, 30°, and 36°), with the expected importance of Hsp90 increasing with temperature. Three additional treatments consisted of the addition of 0.5 M sodium chloride (NaCl) at each temperature (hereafter referred to as the 25S, 30S, and 36S treatments). Elevated NaCl does not induce increased Hsp90 levels (Gasch et al. 2000; Causton et al. 2001; Berry and Gasch 2008), although its basal activity is required for activation of the high-osmolarity glycerol pathway that provides for efficient growth in elevated salinity environments (Hohmann 2002; Yang et al. 2006, 2007; Hawle et al. 2007). Based on these observations and the absolute growth rates of wild-type yeast in the six environments (Table 1), we state the following (cf. also Hietpas et al. 2013):

1. Thirty degrees corresponds to the standard condition. Therefore, this environment represents the one that yeast is best adapted to. Every other environment represents some kind of “adaptive challenge”, which is reflected in lower absolute growth rates.

2. The constraint on the protein function of Hsp90 grows with increasing temperature. However, absolute growth rates imply that wild-type yeast is better able to cope with increased than with decreased temperature.

3. Absolute growth rates indicate that elevated salinity represents a greater environmental challenge than a change in temperature. However, elevated salinity does not result in higher constraint on Hsp90 protein function.

A description of our findings concerning the shape of the DFE in these six environments and the implications for the potential of adaptation to challenging environments follow.

The shape of the full DFE

Figure 4 shows the histograms of the full DFEs obtained for all six environments. All DFEs but that for the 25S environment are bimodal, with one mode centered at wild-type-like fitness and the other mode centered at stop-codon-like fitness. It is likely that the bimodal shape is not seen for 25S because the absolute growth rate of the wild type is the lowest (cf. Table 1). Therefore, the difference between wild-type-like and stop-codon-like fitnesses is smaller. Second, we find the largest beneficials in the 25S environment, resulting in many low read numbers at later time points of the experiment that frequently induce overestimation of deleterious selection coefficients (cf. Figure S2). However, Figure 4 shows that we can clearly distinguish wild-type-like and strongly deleterious mutants although the relative location of the deleterious mode differs between environments dependent on the relative growth of the wild type between environments (generally slower growth at high salinity) and on the duration of the experiment.

The strong mutational constraint on the protein in the high-temperature environment (36°) in conjunction with a relatively high relative growth rate of the wild type results in the most clearly different DFE among the six environments, with a huge number of deleterious mutants, and almost no mutants better than wild-type average. In contrast, the 25°, 25S, and 30S environments, which present challenging environments, but wherein Hsp90 is potentially under less constraint, harbor adaptive potential reflected in many mutations that have higher growth rates than the wild-type average.

To compare the absolute numbers of beneficial and deleterious mutations, we chose a global cutoff of |r| < 0.004 that we define as the wild-type-like (“wt-like”) growth rate. This is a conservative cutoff based on the largest mean 95% credibility interval of wild-type-synonymous mutations across all six environments. Of note, the wt-like category is not necessarily the same as a neutral class of mutations. How many of our observed mutations are effectively neutral is highly dependent on the effective population size, which is probably very high in the bulk competition experiment (cf., e.g., Skelly et al. 2009, who estimate N_e ≈ 2.6 × 10⁷). Since neutral mutations are expected to be those with s ≤ 1/N_e, the threshold for effective neutrality is probably very low and unlikely to be captured by any existing experimental method. In the following, we characterize and compare the DFEs in more detail, based on the above-described categorization.

Evaluating changes in selective pressure on the DFE—one-, two-, and three-step mutations

Because every possible codon was tested at every amino acid position, we can split the full data set into subsets containing mutations that are one-, two-, and three-nucleotide (nt) changes away from the wild-type sequence. These can be interpreted as representing different step numbers in a potential adaptive walk. Of note, only the one-step subset comprises all 81 possible single-nt mutations, whereas the other categories contain subsets of all possible two- and three-step walks in sequence space.

Figure 5 shows bar charts of the proportions of beneficial, wt-like, deleterious, and strongly deleterious mutants for each of the mutational step classes. We observe the following:

1. The proportion of wt-like mutations decreases with the number of nt changes, whereas the number of deleterious and strongly deleterious mutations increases. This is always true for the difference between one- and two-step mutations. This is consistent with a generally lower effect size of single-nt mutations compared with multiple-nt steps—but also with the population residing in a position in the sequence space that confers high mutational robustness (see also Figure 3D in Hietpas et al. 2011). This observation is consistent between all environments.

2. The proportion of beneficial mutations consistently increases with the number of nt steps in the 25°, 25S, 30S, and 36S environments (i.e., environments imposing a selective challenge that does not require higher Hsp90 expression). Conversely, the number of beneficials tends to decrease in the standard environment and the 36C environment. These observations show a stunning correspondence with predictions of the FGM, if one takes into account that we cannot categorize mutations of small fitness effect. In the standard environment, the optimum is close to the current phenotype, and each successive mutational step is likely to overshoot or move away from the optimum, resulting in fewer beneficial and wt-like mutations. On the other hand, if the optimum is shifted away from the current phenotype by presenting a new environmental challenge, there is more potential for mutations to have a beneficial effect, because they are more likely to leave the wt-like area. Hence more steps (or, alternatively, larger arrows) may more likely result in advantageous mutants.

3. The variance of the DFE is higher at 30° and 36° than it is at 25°, as might be expected if Hsp90 is less important for fitness in the 25° environment. Variance is substantially lower for the one-step mutants in comparison to all mutants, again suggesting that the wild-type sequence is fairly robust to nearby changes in sequence space (Figure 6).

We used a Kolmogorov–Smirnov (KS) test to evaluate whether the distributions of one-, two-, and three-step mutations are significantly different from each other and from the complete DFE. The resulting P-values are summarized in Table 2. We observe that the single-step DFE is significantly different from the other distributions, in particular for the 30° and 36° data sets, which are those that are suggestive of being fairly well adapted and close to the phenotypic optimum (cf. the following subsection). Hence, the difference in the distributions likely results not from differences in the beneficial tails of the distributions, but rather from a difference in the proportions of deleterious vs. wt-like mutations (cf. Figure 5). This implies that during adaptation, populations move not only to a local fitness peak but also to a location in the sequence space that confers mutational robustness, such that single-nt mutations are less likely to result in unfit progeny. This is in accordance with a recent study of an RNA virus, showing that mutational robustness evolves even on the level of codon usage (Lauring et al. 2012).

The distance to the optimum according to the FGM

Martin and Lenormand (2006) proposed that the distance to the phenotypic optimum according to the FGM can be estimated by fitting a so-called displaced-gamma distribution to the distribution of fitness effects. A displaced-gamma distribution has the shape of a negative gamma distribution, shifted by a parameter μ to accommodate beneficial mutations. To determine the respective distances to the phenotypic optimum in each environment, which is given by μ, we used a maximum-likelihood approach to fit displaced-gamma distributions to the wild-type-like mode of our DFEs for all six environments [the wild-type-like mode of the DFE for each environment was defined based on the cutoff determined during the initial data preparation (cf. Materials and Methods)]. The resulting distances to the optimum are summarized in Table 3 and a graphical representation of the fitted curves can be found in Figure S6. Largely consistent with what we have reported on the basis of amino acids (Hietpas et al. 2013), the distance to the optimum is lowest for the 36S and 36° environments that are under the highest Hsp constraint. The largest distances to the optimum are estimated for the 25S and 25° environments.

To test whether the displaced-gamma estimates from the single-nt subset differ significantly from those obtained from the full data set, we compared the posterior estimates of the distance to the optimum between data sets. To take into account the effect of the smaller single-step data sets, we repeated the fitting exercise with random subsets of the full data set of the same length as the single-step data. The results are summarized in Figure 7.

Clearly, the estimated distances to the optimum between the full and the single-step data set differ dramatically when the distance to the optimum is large (i.e., particularly in the 25° and 25S environments). Even though the estimation from subsets of the full data naturally results in a higher variance of estimates, this cannot account for the observed differences in the estimates. This has a number of implications:

1. The estimate of the distance to the optimum seems to be primarily determined by the maximum growth rate in the data set. Therefore, the estimates from the subsets of the full data are on average lower than those from the complete data. This implies that the estimate of μ is critically dependent on the successful sampling of beneficial mutants. This can be ensured only in large data sets or lengthy laboratory selection experiments.

2. In those environments that provide potential for adaptation (as suggested by a high number of beneficials), μ is generally underestimated when only the single-nt subset of the data is used—in fact, the maximum observed effect lies frequently above the proposed distance to the optimum estimated from the single-step data set. Hence, single-nt mutations are not sufficient to explore the adaptive landscape well enough to make a statement about the distance to the optimum. This corresponds well to the above observations of larger potential for adaptation to challenging environments in multiple-step mutations (cf. Figure 5 and discussion thereof).

The shape of the beneficial tail of the DFE

The shape of the beneficial tail of the DFE harbors important information about the course and the predictability of evolution. The FGM predicts bounded fitness tails if the optimum is close and an exponential distribution of beneficial fitness effects when the optimum is far away (Martin and Lenormand 2008). However, given that the distribution of beneficial effects indeed represents a tail distribution, all possible tail shapes can be summarized in the generalized Pareto distribution (GPD), and its shape parameter κ yields information about the domain of attraction of the underlying DFE (Beisel et al. 2007; Joyce et al. 2008). So far, most experimental studies have reported exponential tail distributions (e.g., Kassen and Bataillon 2006; MacLean and Buckling 2009), but there exist also reports of bounded beneficial tails (Rokyta et al. 2008; Bataillon et al. 2011), and a recent example of a potentially heavy-tailed distribution of beneficial effects that would not be accounted for by the FGM (Foll et al. 2014).

We fitted the GPD to the beneficial part of our data to compare the tail shapes across environments. The resulting shape parameter estimates are shown in Figure 8. For all but the 25S environment, the negative shape parameter estimates indicate that the DFE most likely belongs to the Weibull domain, i.e., that the DFE is bounded to the right. This would indicate that there is indeed a close optimum and that mutational effects cannot exceed a certain value. However, for the 25S environment, the shape estimates indicate that the DFE might belong to the Frechet domain, i.e., that it has a heavier-than-exponential tail. If the beneficial DFE is heavy-tailed, the size of beneficial fitness effects can be very large and the step size of adaptive evolution will be highly unpredictable.

Effects of synonymous mutations

Many population-genetic statistics rely on the assumption that synonymous mutations in coding regions (i.e., mutations that do not change the amino acid and hence the protein sequence) have equal fitness [e.g., the McDonald–Kreitman test (McDonald and Kreitman 1991) and the HKA test (Hudson et al. 1987)]. However, an increasing number of studies have reported fitness differences between synonymous mutations (e.g., Singh et al. 2007). It is still unclear, however, how these fitness differences arise, although one common hypothesis suggests that, e.g., codon composition can affect RNA packaging (Sauna and Kimchi-Sarfaty 2011). The high precision of the EMPIRIC estimates enables us to explicitly study the distribution of synonymous fitness effects. Each of our data sets comprises 15 mutations that are synonymous to the wild-type sequence, and even within this small selection, we identify a deleterious outlier that persists across replicates and in three of the six different environments (cf. Figure 9 and Figure S7). The identified mutation 588aac represents an unpreferred change coding for asparagine.

Genotype × environment interactions

While the overall shape of the DFE is similar between environments (Figure 4), different individual mutations may nonetheless have very different effects, depending on the environment. Such a scenario is referred to as a genotype × environment (G × E) interaction, wherein the fitness effect of a mutation depends on its environment. The consistency of evolutionary responses of populations to different environments will depend in large part on the extent of G × E, which can be captured as the covariance of fitness effects between environments. If covariance is high, then we expect populations to follow very similar trajectories, since the same mutations are beneficial or deleterious in each environment. By contrast, if covariance is low or negative, then different mutations would be expected to accumulate between populations. The extent and magnitude of G × E is of substantial current interest (e.g., Lazzaro et al. 2008; Gerke et al. 2010) and is typically inferred from studies of standing variation in natural populations. Thus, the current data set represents a unique opportunity to examine the extent to which fitness varies with environment for new mutations.

We find strong between-environment correlations for all mutants (cf. Table 4) and for the 81 single-step mutants (data not shown), with correlation coefficients ranging from 0.76 to 0.97 (all mutants) or from 0.80 to 0.97 (single-step mutants). This result suggests that, on the whole, the sign and magnitude of selection coefficients are consistent between environments. Further consideration, however, suggests that there is substantial variation in fitness effects between environments for particular classes of mutation. For example, among the 76 mutants that are beneficial in at least one environment, correlation coefficients range from −078 to 0.87 (cf. Table S1). Correlations between high salt and normal salt environments tend to be negative, indicating that mutations that are neutral or beneficial in salt tend to be deleterious in normal salinity and vice versa. Mutations that are deleterious in at least one environment (cf. Table S2) also have heterogeneous effects, albeit to a lesser extent than beneficial mutations, with correlation coefficients ranging from −0.12 to 0.86.