Repeatability of gene expression evolution in experimental environmental adaptation

Abstract

Assessing the relative roles of chance and necessity in evolution is of wide interest, but it requires evolving the same population under the same environment multiple times—a virtually impossible task in nature that has been repeatedly accomplished in the laboratory. Capitalizing on the transcriptome data collected in 10 laboratory evolution studies conducted in 22 distinct environments, we investigate the evolutionary repeatability of a total of 182,103 gene expression traits in a prokaryotic and five eukaryotic species. The number of gene expression traits exhibiting concordant changes in direction and magnitude between two replicate populations typically exceeds the chance expectation by 10 to 100 standard deviations. Contrasting replicate evolution in the same environment with that in different environments suggests that the concordance in expression evolution is largely attributable to environment-specific selections, a finding that is further supported by comparing with the outcome of mutation accumulation experiments where the efficacy of selection is minimized. Additionally, genes controlled by more transcription factors tend to show more repeatable expression evolution, likely due to a higher certainty in the occurrence of mutations altering their expressions. In conclusion, contrary to almost unrepeatable genotypic evolution, phenotypic evolution during environmental adaptation is quite repeatable and deterministic, at least for gene expressions.

Experimental evolution enables evaluation of the relative roles of chance and necessity in evolution. This study compiles transcriptomic data from experimental evolution of a prokaryotic and five eukaryotic species in 22 environments to reveal that gene expression evolution is often repeatable and deterministic.

Introduction

The relative importance of chance and necessity in evolution has long been a favorite topic of biologists^1–11. On the one hand, even for the same population evolving under the same environment, the outcome can vary due to several stochastic processes. First, mutations occur randomly. Second, the fixation or loss of a mutant allele is influenced by genetic drift, another stochastic process. Third, the appearance and fixation of a mutation could influence the selection coefficients of future mutations and future evolution, a phenomenon known as contingency^9,12. While contingency itself is not stochastic, the outcome is, because of the randomness of the initial mutation and possibly its fixation. Hence, in a fitness landscape¹³, which maps genotypes to fitness, two identical populations starting from the same location in the landscape could evolve to different locations due to the stochastic processes mentioned, especially when the landscape is rugged¹⁴. On the other hand, there are also deterministic factors in evolution, such as the selection imposed by the environment and the prediction that the population fitness generally increases through evolution¹⁵. The relative importance of chance and necessity in evolution is an empirical question. Furthermore, because fitness is directly determined by the phenotype and because one phenotype can be realized by many genotypes, the relative importance of chance and necessity in evolution likely depends on whether genotypic or phenotypic evolution is concerned and what phenotypic traits are concerned^5,16.

Because repeatability implies necessity, comparative biology has addressed the empirical question of the relative roles of chance and necessity in genotypic and phenotypic evolution by looking for convergence, which refers to evolution to the same state in different lineages^17,18. While cases of convergence are known at genotypic and phenotypic levels, divergent evolution is far more common than convergent evolution^9,19–21, which might lead one to conclude that evolutionary repeatability is generally low and thereby chance plays a bigger role than necessity in evolution. However, comparative biology is perhaps not an ideal approach to this question, because there may simply be too few opportunities for evolution to repeat itself due to the variation in genotype and/or environment among different evolutionary lineages.

To better evaluate the relative importance of chance and necessity in evolution, Gould famously introduced the metaphor of “replaying life’s tape” in his book Wonderful Life¹². Gould described his gedankenexperiment as replays from an identical starting condition, but he lamented that such experiments can never be performed¹². (Note that Gould also asked how differences in starting conditions would influence evolutionary outcomes¹², but we will not address this question here.) While it is indeed extremely difficult to evolve the same population under the same environment multiple times in nature, such experiments have been conducted many times in the laboratory by the experimental evolution approach^22–24. In experimental evolution, one can evolve the same genotype or population under the same laboratory conditions multiple times and record the genotypic and phenotypic evolution that ensues. It is therefore an ideal approach to gauging evolutionary repeatability and understanding the relative roles of chance and necessity in evolution⁹.

In fact, experimental evolution has convincingly demonstrated that genotypic evolution is generally unrepeatable—the same mutation is very rarely fixed in replicate populations evolving under the same environment, even though different mutations in the same gene or in different genes of the same biological process/pathway are often fixed in replicate populations^25,26. The low repeatability of genotypic evolution supports the fitness landscape-based theoretical prediction.

At the phenotypic level, one trait—fitness (or its proxy)—almost always increases in experimental evolution^25,27, but because this trend is virtually independent of the genotype and environment, it is irrelevant to the type of evolutionary repeatability concerned here. For other phenotypic traits, results from experimental evolution are mixed^9,28–31. For example, the evolution of antibiotic resistance appears to be highly repeatable, yet the evolution of the capacity to grow aerobically on citrate occurred in only one of 12 replicate populations of Lenski’s long-term evolution experiment (LTEE) of Escherichia coli in a low glucose medium⁹. Notwithstanding, because the number of phenotypic traits analyzed per study has been small, we lack a general understanding of the repeatability of phenotypic evolution.

An opportunity has now arrived to fill the above gap, thanks to a number of experimental evolution studies that collected transcriptome data. Because the expression level of every gene is a phenotypic trait, these transcriptome data allow probing thousands to tens of thousands of traits per study³². Here we investigate the evolutionary repeatability of a total of 182,103 gene expression traits from 10 experimental studies of six species in 22 environments. Our results paint a general picture that, during environmental adaptation, evolution of gene expression is quite repeatable and deterministic and that the concordant expression changes across replicate populations are driven largely by environment-specific selections with additional contributions from mutation bias (i.e., the rate of mutation affecting the expression of a gene varies across genes).

Results

Experimental evolution and associated transcriptome data

To study the evolutionary repeatability of gene expression traits, we identified from the literature 10 experimental evolution studies that included transcriptome profiling (Table 1). These studies involved the adaptation of a bacterium (E. coli), a fungus (yeast S. cerevisiae), two invertebrates (fly Drosophila simulans and copepod Acartia tonsa), a vertebrate (guppy Poecilia reticulata), and a plant (morning glory Ipomoea purpurea) in a total of 22 environments. In each environment, two or more replicate strains/populations evolved for an equal number of generations. Gene expressions were measured for each replicate strain/population when it first entered the environment concerned (i.e., ancestral expressions) and at the end of the experimental evolution in that same environment (i.e., evolved expressions). Hence, we assume that any gene expression changes detected were due to genetic changes (i.e., evolution) rather than plastic changes³³, although we cannot definitively exclude the involvement of phenotypic plasticity (see Discussion).

Table 1.

Gene expression datasets analyzed

Organism	No. of environments	No. of replicate populations per environment	Biological replicates in transcriptome profiling?	Transcriptome profiling method	Reference
Experimental evolution
E. coli K-12 MG1655	2	7	Yes	Microarray	Fong et al.28
E. coli K-12 MG1655	1	10	No	RNA-seq	Sandberg et al.34
E. coli B REL1206	1	2	No	RNA-seq	Rodríguez-Verdugo et al.35
E. coli MDS42	11	5	No	Microarray	Horinouchi et al.36
E. coli B REL606/607	1	11	Yes	RNA-seq	Favate et al.38
Yeast	2	3	No	Microarray	Dhar et al.39
Fly	1	5	No	RNA-seq	Mallard et al.40,41
Copepod	1	4	No	RNA-seq	Brennan et al.43
Guppy	1	2	No	RNA-seq	Ghalambor et al.44
Morning glory	1	8	No	RNA-seq	Josephs et al.45
Mutation accumulation
E. coli K-12 MG1655 lacking dnaQ	1	3 per founder	No	Microarray	Tsuru and Furusawa50

We addressed the following questions when analyzing the transcriptome data. First, did the number of genes that showed an expression change in two or more replicate populations during experimental evolution exceed the chance expectation, which is the number under the null hypothesis of independent phenotypic evolution among replicates? Second, did the number of genes that showed the same direction of expression change in two or more replicates exceed the chance expectation? Third, did the number of genes that showed the same direction and magnitude of expression change in two or more replicates exceed the chance expectation? Fourth, if an experimental evolution study involved two or more different environments, we further asked whether the evolutionary repeatability of gene expression was higher in the same environment than in different environments. A positive answer would suggest that the observed evolutionary repeatability among replicates in the same environment was at least in part attributable to the selection imposed by the common environment. Otherwise, the observed evolutionary repeatability might have other causes, such as mutation bias. For one species (E. coli), we additionally had access to transcriptome data from a mutation accumulation (MA) experiment where the populations went through repeated severe bottlenecks such that the efficacy of selection was minimized. We thus addressed the fifth question: how repeatable was gene expression evolution in the virtual absence of selection? Sixth, the inclusion of many more than two replicate populations in some of the experimental evolution studies allowed quantifying the level of expression evolution repeatability for each gene, enabling us to address whether different gene expression traits have different levels of evolutionary repeatability and why. We start by describing our analysis of the first experimental evolution study in detail, which is followed by a summary of the results from all 10 studies.

E. coli adaptation to a glycerol medium and a lactate medium

Fong et al. evolved an E. coli K-12 strain that had been adapted to a glucose medium in a glycerol medium for 44 days (~600 generations), with seven replicates²⁸. They similarly evolved the same ancestral strain in a lactate medium for 60 days (~1000 generations), with seven replicates²⁸. Adaptation was evident in the two evolution experiments, because for each of the 14 E. coli populations, the growth rate more than doubled at the end of the evolution compared with that at the beginning. The authors used microarrays to quantify gene expressions of the ancestral line and the seven glycerol-evolved lines in the glycerol medium, and the ancestral line and the seven lactate-evolved lines in the lactate medium, respectively. Biological replicates were included in the microarray experiment, permitting identifying from each evolved line genes that showed a significant expression change from the ancestral level at a false discovery rate (FDR) of 0.05. These genes are referred to as differentially expressed genes (DEGs).

The number of DEGs identified from the seven replicate populations in glycerol varied between 1182 and 2778. For each pair of replicate populations, we computed Dice’s coefficient of similarity between their DEGs, which is the number of DEGs shared by the two replicates divided by the mean number of DEGs in the two replicates (see Methods). Dice’s coefficient is quite high, with an average of 0.781 for the 21 replicate pairs in glycerol (left-most blue bar in Fig. 1a). For comparison, for each replicate pair, we estimated the chance expectation of Dice’s coefficient of similarity in DEGs, which is the coefficient when the DEGs of one replicate are independent from those of the other replicate under the assumption that all genes are equally likely to be DEGs (see Methods). We found that this expectation is on average 0.446 for the 21 replicate pairs (left-most gray bar in Fig. 1a). For each replicate pair, we additionally estimated the mean and standard deviation (SD) of the number of shared DEGs by chance (see Methods). These values allowed calculating a Z-score, which is the number of SDs by which the observed number of shared DEGs exceeds the mean chance expected number of shared DEGs; a Z-score of 1.96 corresponds to a P-value of 0.05 in a two-tailed Z-test. The Z-score ranged from 29.1 to 63.7 for the 21 pairs of replicates (left-most column of blue dots in Fig. 1a), indicating that the number of shared DEGs significantly exceeds the chance expectation for every replicate pair.

Fig. 1. Repeatability of gene expression evolution in E. coli populations adapting to a glycerol medium and those adapting to a lactate medium under three different definitions of repeatability.

Evolutionary repeatability assessed by sharing of differentially expressed genes (DEGs), which show significant expression changes in experimental evolution, among replicate populations evolving in glycerol (a), among replicate populations evolving in lactate (b), and between a population evolving in glycerol and another evolving in lactate (c). X-axis shows the number of populations per comparison. The left Y-axis (bars) indicates the mean Dice’s coefficient of similarity in DEGs, while the right Y-axis (dots) indicates the Z-score, which is the number of SDs by which the observed number of shared DEGs exceeds the chance expectation. Blue and gray bars indicate the mean observed and randomly expected values, respectively, with the error bar showing one standard error. Each dot shows the Z-score from a comparison, with solid blue, solid gray, and open dots indicating Z-scores >1.96, between −1.96 and 1.96, and <−1.96, respectively. There are no open dots in any panel. In (c), the two horizontal blue lines indicate the levels of the blue bar for the 2-line comparison in (a) and (b), respectively. Evolutionary repeatability assessed by sharing of DEGs with the same direction of expression evolution among replicate populations evolving in glycerol (d), among replicate populations evolving in lactate (e), and between a population evolving in glycerol and another evolving in lactate (f). All symbols follow (a–c) except that blue is replaced with orange. Evolutionary repeatability assessed by sharing of DEGs with the same direction and magnitude of expression evolution among replicate populations evolving in glycerol (g), among replicate populations evolving in lactate (h), and between a population evolving in glycerol and another evolving in lactate (i). All symbols follow (a–c) except that blue is replaced with red. Dots are not presented when Z-scores are infinite. In all panels, the bin of “2 lines” contains 21 pairs of replicate populations; that of “3 lines” contains 35 groups of three populations; that of “4 lines” contains 35 groups; that of “5 lines” contains 21 groups; that of “6 lines” contains 7 groups; and that of “7 lines” contains 1 group. In (c, f, i), the colored bar is significantly lower than each colored line. The blue bar in (c) is significantly lower than the top blue line (P = 3.85 × 10⁻³) and the bottom blue line (P = 5.09 × 10⁻¹¹). The orange bar in (f) is significantly lower than the top orange line (P = 1.19 × 10⁻³) and the bottom orange line (P = 4.90 × 10⁻¹⁰). The red bar in (i) is significantly lower than the top red line (P = 8.76 × 10⁻⁸) and the bottom red line (P = 1.40 × 10⁻⁹). P values are from a two-tailed Wilcoxon rank-sum test.

Similar analyses of Dice’s coefficient of similarity in DEGs were performed for all groups of three replicate populations in glycerol, all groups of four replicate populations, and so on, and the results obtained match those from replicate pairs qualitatively (Fig. 1a). When all seven replicates are simultaneously considered, the randomly expected Dice’s coefficient is only 0.0076, yet the observed value is 0.562 (i.e., on average 56.2% of DEGs in a replicate are also DEGs in all other replicates), demonstrating a substantial level of evolutionary repeatability of gene expression in E. coli’s adaptation to the glycerol medium. Similarly, for each group of three or more replicates, we measured the number of shared DEGs by a Z-score. These Z-scores ranged from 46.8 to 297 (Fig. 1a).

We similarly analyzed the seven replicate populations adapted to the lactate medium and obtained qualitatively comparable results (Fig. 1b). Z-scores are generally even greater in lactate (Fig. 1b) than in glycerol (Fig. 1a).

The relatively high evolutionary repeatability of gene expression in each of the two environments raises the question of whether DEGs in glycerol overlap with those in lactate. There are a few reasons why this could be true. First, there may be a mutation bias that causes unequal expression evolution among genes regardless of the specific selection imposed by the environment. For example, expression variation created by mutation may be particularly large for certain genes, which could create shared DEGs between populations evolving in different environments. Second, some shared DEGs may reflect adaptation to the shared components of the glycerol and lactate media or, to a lesser degree, reflect similar intensities of purifying selection on a gene expression trait in the two media for most expression traits. It may also reflect other common aspects of the evolution experiments, such as temperature and serial batch culture conditions. Third, E. coli cells may employ a general response involving the same set of genes to a broad spectrum of stress, such that even though glycerol and lactate impose different stresses, the gene expression responses could be similar. Finally, in both media, E. coli fitness rose by adaptation; some gene expression changes may be consequences of a fitness increase, so could be shared between populations adapting to different environments. Indeed, Dice’s coefficients and Z-scores both reveal that the sharing of DEGs between two populations respectively adapting to the glycerol and lactate media is greater than the chance expectation (Fig. 1c). Nonetheless, Dice’s coefficients are significantly lower between the two environments (blue bar in Fig. 1c) than within each environment (two horizontal lines in Fig. 1c). Further, Z-scores for the number of shared DEGs between environments (Fig. 1c) are smaller than those within each environment (Fig. 1a, b). These observations suggest that the similarity in DEGs between replicate populations evolving in the same environment is, in a large part, attributable to the environment-specific selection imposed on these populations.

In the above, we considered gene expression evolution without specifying whether the expression level has increased or decreased. Next, we assigned a DEG to one of two categories that respectively included expression increases and decreases. We considered the expression change of a gene repeated if it is a DEG and has the same direction of expression change in all populations compared. Under the new definition, we re-estimated Dice’s coefficients and Z-scores (Fig. 1d–f). As expected, Dice’s coefficients are lower now than in Fig. 1a–c, but they remain higher than the chance expectations. Similarly, Z-scores tend to be lower than those in Fig. 1a–c but remain greater than 1.96 in all cases. These results reveal evolutionary repeatability of gene expression even under a more stringent definition of repeatability.

We further increased the stringency in the definition of expression evolution repeatability by considering not only the direction but also the magnitude of expression evolution. If a DEG shows the same direction of expression change in each population concerned and if the difference in log₂(expression ratio) between any two populations concerned is no greater than 0.5, where expression ratio is the evolved expression level divided by the ancestral level, we consider the gene to have shown repeated expression changes. We then reperformed all earlier analyses. Again, we observed qualitatively similar results (Fig. 1g–i).

Hence, by all three definitions of repeatability with different levels of stringency, we found gene expression trait evolution to be quite repeatable in E. coli’s adaptation in glycerol and lactate, respectively. Furthermore, this evolutionary repeatability is significantly greater between populations evolving in the same environment than those evolving in different environments, confirming a major role of the selection imposed by a common environment in the observed evolutionary repeatability.

Analyses of nine additional experimental evolution studies

Below, we briefly describe the nine additional experimental evolution studies considered (Table 1), followed by a summary of the results from the analyses of all 10 studies. In the second experimental evolution study, Sandberg et al. evolved an E. coli K-12 strain that had been adapted to 37 °C in 42 °C for ~1500 generations, with 10 replicates³⁴. Adaptation to the high temperature was evident from increased population growth rates. RNA-seq transcriptomic data were collected from the ancestral line and 10 evolved lines at 42 °C.

In the third study, Rodríguez-Verdugo et al. evolved a strain of E. coli B (REL1206) that had been adapted to 37 °C in 42 °C for ~2000 generations, with two replicates³⁵. They reported increased cell density for the evolved strains relative to the ancestor at high temperature. RNA-seq was used to profile the transcriptome of the ancestral strain and the two evolved lines at 42 °C. Although both the second and third studies investigated E. coli adaptation to a high temperature, they used different growth media (in addition to different strains and different numbers of generations of evolution). Hence, the two evolution experiments took place in different environments.

In the fourth study, Horinouchi et al. evolved an E. coli strain that had been adapted to the M9 minimal medium in each of 11 different harsh environments for 350–800 generations, with five replicates per harsh environment³⁶. The authors observed drastically increased growth rates of these lines in the harsh environments after the experimental evolution. Transcriptomes were microarray-profiled for the ancestral line in each harsh environment and for all evolved lines in the respective environments where they evolved. Note that, although one of the harsh environments contained lactate, as in the first study²⁸, the lactate concentration differed in the two studies; hence we treated the two as different environments.

In the E. coli LTEE, Lenski established 12 parallel E. coli lines from a single clone and allowed them to continuously evolve in a low glucose medium³⁷. These lines have been evolving for over 80,000 generations, and their fitness has been continuously rising²⁷. In the fifth study, transcriptomes from the ancestor and the 12 parallel lines at ~50,000 generations were collected in the low glucose medium by RNA-seq³⁸. Because one transcriptome (Ara+6) was contaminated³⁸, we analyzed the remaining 11 replicate lines.

In the sixth study, Dhar et al. respectively evolved S. cerevisiae that had been adapted to a benign environment in salt stress (NaCl) and in oxidative stress (H₂O₂) for 300 generations, each with three replicates³⁹. They observed a significant increase in yeast fitness in each stressful environment after the experimental evolution. Microarray-based transcriptome profiling was performed for the ancestral strain under the two stresses and each evolved line under the respective stress.

In the seventh study, Mallard et al. exposed five replicate populations of D. simulans to a hot environment with fluctuating temperature between 18 and 28 °C for 64 generations and another set of five replicate populations to a cold environment with fluctuating temperature between 10 and 20 °C for 39 generations, respectively^40,41. They collected eggs from each population and reared them at 23 °C in a common garden experiment. Transcriptomes of five cold-evolved populations in 23 °C and five hot-evolved populations in 23 °C were profiled by RNA-seq. As previously demonstrated, the cold-evolved populations can be considered ancestral in their gene expression levels⁴². Hence, we compared gene expressions at 23 °C between the cold-evolved and hot-evolved populations.

To investigate how climate change might affect the marine copepod A. tonsa, in the eighth study, Brennan et al. experimentally evolved copepods in an environment reflecting today’s oceanic pH and temperature and in a potential future ocean environment with a lower pH and a higher temperature for 20 generations, each with four replicates⁴³. Copepods showed rapid adaptation to the stressful future environment by recovering egg production and hatching success in 20 generations. Each population was then separated into two, one of which was maintained in the same environment as the previous 20 generations for three more generations, while the other was transplanted to the alternative environment for three generations. We compared RNA-seq transcriptomic data from two groups of populations collected after 23 generations: group 1 populations were continuously maintained in the future environment for 23 generations (reflecting evolved expressions), while group 2 populations were transplanted to the future environment after being maintained in today’s environment for 20 generations (reflecting ancestral expressions).

In the ninth study, Ghalambor et al. investigated the adaptation of the Trinidadian guppy P. reticulata by introducing individuals sampled from a high-predation ancestral environment to two low-predation streams⁴⁴. Evidence for adaptation to the low-predation environment was observed after a year. They reared guppies caught from the ancestral environment and those caught from the low-predation streams in an artificial low-predation environment. RNA-seq was performed using the brain tissue from these guppies.

In the tenth study, Van Etten and colleagues selected the common morning glory I. purpurea for herbicide resistance over three generations, with eight replicates^45,46. They also maintained a control line without herbicide for the same number of generations. Then, they applied herbicide on both the resistant lines and the control line. RNA-seq transcriptomes were collected from leaf tissues at 8 and 32 h after herbicide treatment. We analyzed the transcriptomes collected at 32 h post-treatment, by which point the gene expression response was expected to be fully induced.

In total, the 10 studies included the evolution of 117 populations of one prokaryote and five eukaryotes in 22 distinct environments. Following the analysis in Fig. 1, we examined the evolutionary repeatability of gene expression in an environment using Dice’s coefficients of similarity and Z-scores, by considering the three definitions of repeatability introduced earlier. Below, we describe the findings from the analyses of replicate pairs (Fig. 2), because results from the analyses of three or more replicates are qualitatively similar (Figs. S1–S7). First, when sharing a DEG between replicates is considered repeatable expression evolution, the average Dice’s coefficient is greater than the chance expectation in each experiment, although a substantial variation in Dice’s coefficient exists among the 22 experiments (Fig. 2a). Almost all Z-scores are greater than 1.96, with the median being 26.1 (Fig. 2b). Second, when sharing a DEG with the same direction of expression change between replicates is considered repeated evolution, the average Dice’s coefficient is again greater than the chance expectation in each of the 22 experiments (Fig. 2c) and almost all Z-scores are greater than 1.96, with the median being 30.5 (Fig. 2d). Third, when sharing a DEG with the same direction and magnitude of expression change between replicates is considered repeated evolution, the mean Dice’s coefficient is again greater than the chance expectation in each of the 22 experiments (Fig. 2e) and almost all Z-scores exceed 1.96 (median = 23.6) (Fig. 2f). These results demonstrate that gene expression evolution is generally quite repeatable.

Fig. 2. Repeatability of gene expression evolution based on 10 studies of experimental evolution of six species in 22 environments.

a Mean Dice’s coefficient of similarity of DEGs between replicate populations. Yeast in H₂O₂ and in NaCl each contains 3 pairs of replicate populations. Copepod in the future environment contains 6 pairs of populations. Fly in hot temperature contains 10 pairs of populations. Morning glory with herbicide contains 28 pairs of populations. E. coli in NaCl, KCl, CoCl₂, Na₂CO₃, Lac (Horinouchi et al.), Mal, MCL, MG, BuOH, CPC, and Cro each contain 10 pairs of populations. E. coli in glycerol and in lactate (Fong et al.) each contains 21 pairs of populations. E. coli K-12 in 42 °C contains 45 pairs of populations. E. coli LTEE contains 55 pairs of populations. Guppy in low-predation streams and E. coli B REL1206 in 42 °C each contain 1 pair of populations. Blue and gray dots, respectively, show the mean observed and randomly expected Dice’s coefficients, with error bars indicating the standard error. A thin horizontal line connects the blue and gray dots for the same experiment. b Z-scores measuring the extent of sharing of DEGs between replicate populations that goes beyond the chance expectation. Each dot represents a comparison between two replicate populations evolving in the same environment. Blue, gray, and open dots respectively show Z-scores >1.96, between −1.96 and 1.96, and <−1.96. No open dots are present in any panel. The box plot summarizes the distribution of the dots: the left and right edges of the box represent the first and third quartiles, respectively, the vertical line inside the box indicates the median, and the whiskers extend to the most extreme values inside inner fences (median ± 1.5× interquartile range). c Mean Dice’s coefficient of similarity of DEGs with the same direction of expression evolution between replicate populations. Symbols follow (a) except that blue is replaced with orange. d Z-scores measuring the extent of sharing of DEGs with the same direction of expression evolution between replicate populations. Symbols follow (b) except that blue is replaced with orange. e Mean Dice’s coefficient of similarity of DEGs with the same direction and magnitude of expression evolution between replicate populations. Symbols follow (a) except that blue is replaced with red. f Z-scores measuring the extent of sharing of DEGs with the same direction and magnitude of expression evolution between replicate populations. Symbols follow (b) except that blue is replaced with red. In (a, c, e), the 22 experiments are ordered on the Y-axis based on their mean Dice’s coefficients. In (b, d, f), the 22 experiments are ordered on the Y-axis following the order in (a, c, e), respectively.

In addition to the first experimental evolution study analyzed (Fig. 1), two other studies respectively performed experimental evolution in more than one environment (Table 1). We thus similarly examined evolutionary repeatability between populations evolving in different environments (Figs. S8 and S9). The general trend observed is similar to that found from the first study (Fig. 1c, f, i). That is, the similarity in gene expression evolution between populations evolving in different environments is greater than the chance expectation, yet is lower than that between populations evolving in the same environment.

Of the 10 studies, only two^28,38 contained biological replicates when profiling the transcriptome of each strain. For these two studies, we identified DEGs using an adjusted P-value of 0.05 after FDR corrections as the cutoff. For the remaining studies, we computed the expression ratio by dividing the evolved expression level of a gene by its ancestral expression level and used a cutoff of |log₂(expression ratio)| >0.5 for microarray data and >2 for RNA-seq data to identify DEGs (see Methods). We used different cutoffs for the two types of data because, compared with RNA-seq, microarray has a smaller dynamic range and is less powerful in detecting expression differences^47,48. Regardless, because we always contrast observed evolutionary repeatability with its chance expectation under the same cutoff, the specific cutoff used is not expected to affect our results qualitatively. To confirm this prediction, we employed the |log₂(expression ratio)| cutoff appropriate for the respective transcriptomic data type on the two datasets with biological replicates in transcriptome profiling^28,38. Overall, we found that using |log₂(expression ratio)| and adjusted P-value cutoffs yielded qualitatively similar results (Figs. S10 and S11), suggesting that our results from the eight datasets without biological replicates in transcriptome profiling are likely reliable.

Mutation bias contributes to the evolutionary repeatability of gene expression

Our analyses not only revealed a similarity in gene expression evolution among replicate populations adapting to the same environment but also discovered a significant albeit lower level of similarity between populations adapting to different environments (Fig. 1c, f, i; Figs. S8 and S9). These observations suggest the possibility that evolutionary repeatability is in part due to mutation bias⁴⁹, which makes the rate of mutation influencing the expression of a gene different for different genes. To test this hypothesis, we analyzed transcriptomic data collected from an E. coli MA experiment (Table 1). Specifically, Tsuru and Furusawa established two founder strains (A and B) from an E. coli mutator strain (K-12 MG1655) lacking the dnaQ gene⁵⁰. They then propagated three parallel MA lines per founder strain for 30 rounds of single-cell bottlenecks on LB plates and collected the transcriptomes of the two founder strains and six MA lines using microarrays. We analyzed these data using the three different definitions of evolutionary repeatability.

For the three MA lines derived from Founder A, under each definition of evolutionary repeatability, the mean Dice’s coefficient is greater than the corresponding chance expectation, and every Z-score is greater than 1.96 for replicate pairs and replicate triplets (Fig. 3a–c). Similar results were obtained for the three MA lines derived from Founder B, although some Z-scores are below 1.96 (Fig. 3d–f). These results demonstrate that expression evolution is somewhat repeatable even in the virtual absence of selection. This said, we note that the median Z-scores observed here for replicate pairs are much lower than those in Fig. 2. Similarly, the mean Dice’s coefficients observed here are smaller than those of most datasets in Fig. 2. Furthermore, the evolutionary repeatability of gene expression due to mutation bias is probably lower in experimental evolution than observed here because of the use of a mutator in the MA experiment. Hence, our results suggest that mutation bias is a contributor to the repeatability of gene expression evolution observed in experimental evolution, yet its contribution is minor. Furthermore, because the mean Dice’s coefficients observed here (Fig. 3) are similar to those observed between populations adapting to different environments (Fig. 1c, f, i; Figs. S8 and S9), the similarity in expression evolution in different environments may be largely due to mutation bias.

Fig. 3. Repeatability of E. coli gene expression evolution during mutation accumulation.

Evolutionary repeatability of Founder A lines assessed by sharing of DEGs (a), sharing of DEGs with the same direction of expression evolution (b), and sharing of DEGs with the same direction and magnitude of expression evolution (c) among replicate lines. X-axis shows the number of replicate lines per comparison. The left Y-axis (bars) indicates the mean Dice’s coefficient of similarity, while the right Y-axis (dots) indicates the Z-score. Colored and gray bars indicate the mean observed and randomly expected values, respectively, with the error bar showing one standard error. Each dot shows the Z-score from a comparison, with colored, gray, and open dots indicating Z-scores >1.96, between −1.96 and 1.96, and <−1.96, respectively. There are no open dots in any panel. Evolutionary repeatability of Founder B lines assessed by sharing of DEGs (d), sharing of DEGs with the same direction of expression evolution (e), and sharing of DEGs with the same direction and magnitude of expression evolution (f) among replicate lines. All symbols follow (a–c). In all panels, the bin of “2 lines” contains 3 pairs of replicate lines, whereas the bin of “3 lines” contains 1 group of three populations.

Expression evolution is more repeatable for genes controlled by more transcription factors

Because gene expression evolution is not 100% repeatable (Fig. 2), one wonders whether different genes show different levels of repeatability. To this end, we focused on the E. coli LTEE dataset³⁸ because of its large number of replicate populations. The relative level of expression repeatability of a gene is measured by the number of replicates in which a gene is found to be a DEG. If all genes have equal repeatability, the number of replicates in which a gene is a DEG should follow a Poisson distribution, whose variance equals the mean. The variance of the observed distribution of the number of replicates in which a gene is a DEG is much greater than that of the Poisson expectation (P < 0.001, one-tailed chi-squared test comparing the observed variance with the Poisson variance), with overrepresentations of genes that show repeatability in 0 and in at least 4 replicates (Fig. 4a). This result indicates that different genes vary significantly in their expression evolution repeatability.

Fig. 4. Expression evolution of E. coli genes controlled by more transcription factors (TFs) is more repeatable.

a Frequency distribution of the number of replicate populations (out of 11) where a gene is a DEG based on E. coli LTEE³⁸ (blue bars), compared with the Poisson distribution with the same mean (gray bars). The P-value is from a one-tailed chi-squared test comparing the observed variance with the corresponding Poisson variance (P = 0.00). Repeatability of the expression evolution of a gene is positively correlated with the number of TFs controlling the gene, based on the experimental evolution data of E. coli LTEE³⁸ (11 replicates) (b), K-12 strain evolving in 42 °C³⁴ (10 replicates) (c), MDS42 strain evolving in one of the 11 harsh environments (MG)³⁶ (5 replicates) (d), and MDS42 strain evolving in all 11 harsh environments³⁶ (e). The dotted line shows the linear regression based on unbinned data, with the shaded area showing the 95% confidence intervals of the mean number of TFs controlling a gene. Spearman’s rank correlation (ρ) and associated P-value are based on unbinned data. The width of each violin is proportional to the number of genes. The bolded line inside each violin represents the mean. Blue and gray violins indicate repeatable genes (being a DEG in ≥ 2 replicates) and unrepeatable genes (being a DEG in only 1 replicate), respectively. In (e), the 55 populations across the 11 environments are considered 55 replicates.

What types of genes tend to have higher repeatability? Genes controlled by more TFs are expected to have larger targets for expression-altering mutations⁵¹. Hence, there is a higher certainty for genes controlled by more TFs to have genetic expression variants in replicate populations, which will translate to a higher evolutionary repeatability. To confirm this prediction, we focused on three large datasets from E. coli experimental evolution^34,36,38 because (i) the information about the number of TFs controlling each gene is readily available in E. coli⁵², and (ii) two of the three large datasets each have ≥10 replicates per environment^34,38 while the remaining one has 11 environments each with five replicates, which may be considered to have 55 approximate replicates³⁶. Indeed, in each of these datasets, we observed a significant positive correlation between the number of TFs controlling a gene and its repeatability in expression evolution (i.e., the number of replicates in which the gene is a DEG) (Fig. 4b–e). Notwithstanding, the correlations were relatively small, suggesting that the number of TFs controlling a gene is but one of potentially many factors influencing the expression evolution repeatability of the gene. It would be interesting to tease apart the contributions of these factors when more data become available in the future.

Discussion

By analyzing the evolution of a total of 182,103 gene expression traits from 10 experimental evolution studies performed in 22 environments, we found that gene expression evolution is quite repeatable for the same population or genotype in the same environment. This repeatability is reflected in the between-replicate sharing of (i) genes showing expression evolution, (ii) genes with a particular direction of expression evolution, and (iii) genes with a particular direction and magnitude of expression evolution. While the outcomes of gene expression evolution of replicates in the same environment are almost always significantly more similar than the chance expectation (Fig. 2b, d, f), the level of similarity measured by Dice’s coefficient of similarity varies greatly among environments and organisms (Fig. 2a, c, e). Interestingly, the mean Dice’s coefficient per environment appears lower for the four multicellular species than for the two unicellular species under each definition of evolutionary repeatability (two-tailed P = 0.061, 0.074, and 0.011, respectively, Wilcoxon rank-sum test; Fig. 2a, c, e). However, because the 22 evolution experiments from the 10 studies varied in so many parameters (species, environment, duration of the evolution, and transcriptome profiling method), much more data are required to evaluate the influences of individual factors.

Our results suggest that, at least for gene expression traits, phenotypic evolution is quite repeatable and deterministic, contrasting genotypic evolution, which has been found by both comparative biology^9,19–21 and experimental evolution^25,26 to be minimally repeatable and highly stochastic. This disparity may be because genotype-to-phenotype mapping is many-to-one, meaning that a phenotype can be achieved by multiple genotypes. Hence, different genetic changes may yield the same or similar phenotypic changes. Of course, ultimately, what matters in evolution is fitness, and phenotype-to-fitness mapping is also many-to-one, which is at least part of the reason why phenotypic evolution is not 100% repeatable. Therefore, it is possible that the higher (yet not 100%) repeatability of phenotypic evolution relative to that of genotypic evolution is a natural consequence of the genotype-to-phenotype-to-fitness mapping that is many-to-one at each step of the way.

We observed a level of similarity in gene expression evolution that exceeds the chance expectation, even between populations evolving in different environments, and provided evidence for one potential cause. Specifically, we found that, even in the absence of selective effects, gene expression evolution showed higher-than-expected repeatability, suggesting mutation bias as a contributor to the repeatability of expression evolution in the same or different environments. However, it is not possible to estimate precisely the contribution of mutation bias to the level of evolutionary repeatability observed in experimental evolution, because the strains and the experimental conditions were not the same in MA and experimental evolution. Other potential causes of the similarity in gene expression evolution in different environments include the possibility that the different environments concerned share some elements of selection because of the similarity of the environments, the possibility that organisms use a common gene expression program to respond to a broad spectrum of stress, and the possibility that some gene expression changes are consequences of fitness improvements. While we do not have direct evidence for these other causes, they seem plausible.

Our study has several caveats that are worth discussing. First, in our analysis, the ancestral transcriptome was profiled some time after the strain/population was placed in the new environment. If plastic gene expression changes had not been fully induced at the time of transcriptome profiling, some plastic changes could be included in the measured evolutionary changes, which could affect the estimated repeatability of gene expression evolution among replicates. While this possibility exists, for the following reasons, we believe that our conclusion is unlikely to have been qualitatively affected. First, the authors of the original studies profiled the transcriptomes with the explicit purpose of measuring gene expression evolution. Hence, they presumably collected the transcriptomic data after the plastic expression changes had been fully induced. For example, in the study of E. coli adaptations to 11 different harsh environments, the authors allowed E. coli populations to grow for 60 h in the new environments before transcriptome profiling³⁶. Similarly, two generations of acclimation to the new environment took place before the transcriptome was profiled in the fly^40,41 and guppy⁴⁴ evolution experiments, respectively. Second, some studies performed transcriptome profiling at several timepoints after placing the organisms in the new environment for acclimation^43,45,46, and we always used the data collected after the longest acclimation to maximize the induction of plastic gene expression changes. This practice presumably minimized the influences of plastic changes on our analysis. Third, while the acclimation time measured in hours or number of generations varied across the 10 studies, all studies—including those with sufficient acclimations—exhibited qualitatively similar results of high repeatability of gene expression evolution, rendering it unlikely that some of the studies were qualitatively affected by plastic gene expression changes. Finally, because genetic expression changes tend to reverse plastic expression changes³³, our measured expression changes are unlikely to be greater in magnitude than genetic expression changes; consequently, evolutionary repeatability is unlikely to have been overestimated. We note that the finding of a large number of DEGs in the adaptation to an environment (e.g., >1000 in the first dataset analyzed) does not imply an influence of plastic gene expression changes, because a mutation observed in experimental evolution was shown to be able to alter the expressions of over a thousand genes in E. coli³⁵.

Second, organisms at the end of experimental evolution are only partially adapted to that environment. This is evident from the E. coli LTEE, in which the fitness of the 12 E. coli populations continued to rise in evolution⁵³ after the timepoint of transcriptome profiling³⁸. That is, even after 50,000 generations of evolution in a constant environment, the E. coli populations were still adapting. Thus, it is highly likely that the gene expression levels at the end of the experimental evolution in each study do not represent those when the populations become fully adapted to the environment. However, because the replicate populations compared evolved for the same number of generations, comparing them is appropriate. Furthermore, the concept of evolutionary repeatability need not be restricted to fully adapted populations. Hence, our comparison among replicate populations in experimental evolution is meaningful in terms of informing about evolutionary repeatability in nature, especially because natural populations are most likely in the process of adaptation rather than at the end of adaptation⁵⁴.

Third, we treated the expression level of each gene as a phenotypic trait. While the measurement of each trait is independent, the traits are correlated because the expressions of all genes are controlled by the same gene regulatory network. In other words, changes of traits are not independent from one another. However, this situation is no different from the analysis of multiple morphological or physiological traits, which also tend to be correlated⁵⁵. The advantage of analyzing gene expression traits is the availability of a huge number of traits that have been measured in a uniform and accurate fashion in multiple studies. No other group of traits has such large datasets from experimental evolutionary studies.

Fourth, although we detected repeatability of gene expression evolution, it is unclear whether the observed gene expression changes contribute to organismal adaptation or are consequences of organismal adaptation. It is likely that both situations exist, but the relative proportions are unknown. While answering this question would not influence the assessment of evolutionary repeatability, it would help deepen our understanding of the cause of evolutionary repeatability.

The relatively high evolutionary repeatability of gene expression discovered in the present work contrasts the prior observations from comparative studies^56,57, suggesting that the seemingly low evolutionary repeatability of gene expression in natural evolution is likely due to the variation in the starting genotype and/or environment across evolutionary lineages (i.e., a lack of opportunity for repeated evolution) rather than a genuinely low repeatability. In other words, the relative importance of chance and contingency in evolution may have been overestimated. It is worth mentioning that while we primarily considered intrinsic stochastic evolutionary processes such as mutation and drift as chance in the present study, chance could also arise from random environmental fluctuations such as natural disasters and weather events. Although most of the experimental evolution studies considered here were performed in laboratories and hence were deprived of such environmental fluctuations, at least one of the studies was conducted in the wild⁴⁴ so was subjected to environmental fluctuations.

To what extent our observation of a relatively high degree of evolutionary repeatability of gene expression traits holds true for other phenotypic traits is unclear (except for fitness), because no other traits have been systematically investigated by experimental evolution at such a large scale. Therefore, it is unknown whether the relative importance of chance in the evolution of other phenotypic traits has also been overestimated. It has been proposed that, when phenotypic traits are stratified according to a hierarchy of biological organization, the fraction of evolutionary changes in phenotype that are adaptive increases with the phenotypic level considered⁵, which would mean that the repeatability of phenotypic evolution also rises with the phenotypic level considered⁵⁸. With the advancement of phenotyping methods (e.g., for measuring cell morphologies)^59,60, we may soon be able to quantify many other traits in a high-throughput fashion in experimental evolution studies, which will allow comparing the repeatability of phenotypic evolution across different types of traits and provide a broader picture of the relative importance of chance and necessity in phenotypic evolution.

Methods

E. coli transcriptomic data

Six different E. coli transcriptomic datasets were used in this study. The first dataset is from the experimental evolution performed by Fong et al.²⁸. Specifically, the authors evolved an E. coli K-12 MG1655 strain that had been adapted to a glucose medium in a glycerol medium and in a lactate medium, respectively, each with seven replicates. They used microarray to quantify gene expressions of (i) the ancestral line evolving in the glycerol medium on day 1 (five biological replicates) and the same ancestral line evolving in the lactate medium on day 1 (six biological replicates), and (ii) seven parallel lines that had evolved in the glycerol medium for 44 days (three biological replicates per line) and seven parallel lines that had evolved in the lactate medium for 60 days (three biological replicates per line), respectively. In total, expressions of 4392 genes were analyzed in each environment. This study investigated the repeatability of gene expression evolution but concluded that the repeatability was low, apparently because the authors required a gene to be a DEG in the vast majority of replicate populations for it to be considered a case of repeated evolution and because they did not compare the observed number of cases of repeated evolution with that expected by chance²⁸.

The second dataset is from the experimental evolution of E. coli strain K-12 MG1655 that had been adapted to 37 °C in glucose M9 minimal medium at 42 °C, with 10 replicates³⁴. After approximately 1500 generations of growth, the authors performed RNA-seq on (i) the ancestral line of E. coli at 42 °C with 10 biological replicates (the mean expression level from the 10 replicates was considered for each gene), and (ii) 10 evolved lines (with one replicate per line) at 42 °C. In total, the expressions of 3815 genes were considered.

The third dataset is from the experimental evolution of E. coli strain B REL1206 that had been adapted to 37 °C in 42 °C, with two replicates³⁵. After 2000 generations, the authors performed RNA-seq on (i) the ancestral strain at 42 °C and (ii) two evolved lines at 42 °C. In total, the expressions of 4180 genes were analyzed.

The fourth dataset is from the experimental evolution of E. coli strain MDS42 preadapted to the M9 minimal medium with glucose in 11 different harsh environments for 350–800 generations³⁶. The 11 harsh environments are M9 medium with 400 mM NaCl (NaCl), 210 mM potassium chloride (KCl), 16 μM cobalt chloride (CoCl₂), 32.5 mM sodium carbonate (Na₂CO₃), 40 mM L-lactate (Lac), 30 mM L-malate (Mal), 8.75 mM methacrylate (MCL), 50 mM crotonate (Cro), 350 μM methylglyoxal (MG), 1.25% n-butanol (BuOH), and 4.8 μM cetylpyridinium chloride (CPC), respectively. The authors used microarray to measure the gene expressions of (i) the ancestral strain in the 11 harsh environments, and (ii) five parallel lines that evolved in each harsh environment in the respective harsh environment. In total, the expressions of 4492 genes were considered.

The fifth dataset is from the E coli LTEE³⁷. Briefly, 12 populations were founded from a single clone of E coli strain B in 1988 and allowed to continuously evolve in a low-glucose medium. Favate et al. performed RNA-seq on a single clone (with two biological replicates) from each of the 12 parallelly evolving lines⁶¹. Because one of the samples (Ara+6) was contaminated, we excluded it from our analysis. In total, the expressions of 4133 genes were considered. Gene expression data are available as an output file from DESeq2⁶¹. This study also investigated the repeatability of gene expression evolution and concluded that the repeatability was high⁶¹.

The sixth dataset is from an MA experiment of E. coli K-12 MG1655⁵⁰. The authors established two founder strains (A and B) from an E. coli mutator strain lacking dnaQ. Single colonies were randomly picked and transferred daily to fresh LB agar plates for 30 rounds. Transcriptomes were collected using microarray for (i) two biological replicates of each founder strain, and (ii) three parallel MA lines diverged from each founder strain after 10 rounds and 30 rounds (one biological replicate per line). We analyzed the transcriptomes at the later timepoint and used the mean expression level of the ancestral line from the two biological replicates.

Yeast transcriptomic data

Dhar et al. respectively evolved the haploid yeast strain BY4741 in salt stress (NaCl) and oxidative stress (H₂O₂) for 300 generations³⁹. After the experimental evolution, they profiled the transcriptomes using a microarray on (i) the ancestral strain in NaCl and H₂O₂, respectively, and (ii) three parallel lines evolved in NaCl (in NaCl) and three parallel lines evolved in H₂O₂ (in H₂O₂), respectively. The microarray experiment had one biological replicate per evolved line and two biological replicates for the ancestral line (the mean from the two biological replicates was used in the analysis). In total, the expressions of 5814 genes were considered.

Fly transcriptomic data

Mallard et al. collected Drosophila simulans in Northern Portugal in 2008^40,41. They evolved five replicate populations in a hot environment with 12 h at 18 °C (dark) and 12 h at 28 °C (light) for 64 generations. They evolved another five replicate populations in a cold environment with 12 h at 10 °C (dark) and 12 h at 20 °C (light) for 39 generations (cold). They performed RNA-seq on (i) the five cold-evolved populations after rearing them at 23 °C and (ii) five hot-evolved populations after rearing them at 23 °C. In total, expressions of 11,870 genes were considered.

Copepod transcriptomic data

Brennan et al. collected marine copepods Acartia tonsa from Esker Point Beach in June 2016 and evolved them for 20 generations in an ambient environment (AM) (18 °C, pH ~8.2, and pCO2 ~ 400 μatm) and a potential future environment with ocean warming and acidification (OWA) (22 °C, pH ~7.5, and pCO2 ~ 2000 μatm)⁴³. In the 21st generation, they started to reciprocally transplant some copepods between AM and OWA, while others remained in their respective original environments as control populations. The reciprocal transplant and the controls continued for three more generations (22nd, 23rd, and 24th). RNA-seq libraries were prepared with a pool of copepods after each reciprocal transplant for each condition. We analyzed the transcriptomes from the 24th generation of (i) the four parallel populations of copepods in AM that were transplanted to OWA at the 21st generation, and (ii) the four parallel populations of copepods in OWA since the beginning of the experiment. In total, expressions of 25,218 genes were considered.

Guppy transcriptomic data

Ghalambor et al.⁴⁴, introduced Trinidadian guppies (P. reticulata) sampled from a high-predation natural environment to two low-predation environments in March 2008. After 1 year of experimental evolution (about three to four generations), Ghalambor et al. performed RNA-seq on the brain tissues of guppies from the two introduced populations after acclimation in an artificial low-predation environment. They also performed RNA-seq on the brain tissues of guppies that were sampled from the high-predation environment and acclimated in the same artificial low-predation environment. In total, expressions of 37,444 genes were considered.

Morning glory transcriptomic data

Van Etten and colleagues selected common morning glories (I. purpurea) for herbicide resistance and performed RNA-seq using leaf tissues after the experimental evolution^45,46. Seeds from an ancestral population were sampled from the University of Georgia Plant Sciences Farm in Oconee, GA in 2000. The offspring of this population were screened in a greenhouse for high resistance to the herbicide glyphosate, and the top 20% resistant lines were selected for herbicide resistance for three generations. Control lines were also developed by randomly choosing 20% of the ancestral population and growing them for three generations in the same greenhouse with no herbicide. Seeds from both the resistant lines and the control lines were planted into two blocks with a randomized treatment design (herbicide or no herbicide). After 6 weeks of growth, the authors applied herbicide and collected leaf tissues 8 and 32 h after the treatment, respectively. In this study, we used leaf transcriptomic data from (i) 16 inbred individuals from the control lines 32 h after herbicide treatment, and (ii) 8 inbred individuals from the resistant lines 32 h after herbicide treatment. In total, expressions of 25,619 genes were considered.

Identification of DEGs

Biological replicates were included in transcriptome profiling of both ancestral and evolved lines in two datasets^28,38. We identified DEGs using the R package limma⁶² at an FDR of 0.05 in each test for microarray data provided by Fong et al. For RNA-seq data from Favate et al. (available as DESeq2 output), we similarly used an FDR cutoff of 0.05 to identify DEGs. For the rest of the datasets, where biological replicates were not included in transcriptome profiling of both ancestral and evolved lines, DEGs were identified using |log₂(evolved expression/ancestral expression)| cutoffs. For microarray data, the cutoff of 0.5 was applied. For RNA-seq data, the cutoff of 2 was applied.

Dice’s coefficient of similarity

We followed a previous study²⁵ to employ Dice’s coefficient of similarity to measure the similarity between two gene sets. Dice’s coefficient equals $\frac{2∣{\text{X}}_1\cap {\text{X}}_2∣}{∣{\text{X}}_1∣+∣{\text{X}}_2∣}$, where X₁ and X₂ refer to two gene sets, $∣{\text{X}}_1∣\text{and}∣{\text{X}}_2∣$ represent the numbers of genes in X₁ and X₂, respectively, and $∣{\text{X}}_1\cap {\text{X}}_2∣$ represents the number of genes in the intersection of X₁ and X₂. Similarly, Dice’s coefficient of similarity among n gene sets equals $\frac{n∣{\text{X}}_1\cap {\text{X}}_2\cap \dots \cap {\text{X}}_n∣}{{\sum }_{i=1}^n∣{\text{X}}_i∣}$. Dice’s coefficient ranges from 0 (no gene appears in all sets) to 1 (all sets are the same).

Similarity in gene expression evolution between replicate populations by chance

Let us assume that n evolved populations each have the same m genes. In each population, we marked a subset of genes that showed significant expression changes in evolution, which allowed computing the number (N) of genes marked in all n populations as well as Dice’s coefficient of similarity among the n sets of marked genes. To assess the number of genes marked in all n populations simply by chance, we performed the following analysis. First, for each population, we randomly marked the same number of genes as originally marked. Second, we counted the number of genes now marked in all n populations. We repeated this process 1000 times to obtain the mean (M) and SD of the number of genes marked in all populations by chance. We then computed Z-score = (N-M)/SD. Based on the same 1000 sets of randomly marked genes of the n populations, we computed 1000 Dice’s coefficients of similarity among the n populations by chance and their mean. Similar analyses were performed when we also considered the direction of a gene expression change or considered both the direction and magnitude of a gene expression change in defining evolutionary repeatability.

Transcription factors

The number of transcription factors controlling each E. coli gene was downloaded from RegulonDB version 13.6.0 (released 03/19/2025)⁵². In total, there are 241 TFs that control 2432 genes via 5732 interactions.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

J.Z. conceived the project and secured funding; J.L. and J.Z. designed the study; J.L. analyzed the data; J.L. and J.Z. wrote the paper.

Peer review

Peer review information

Nature Communications thanks Zachary Blount and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

Gene expression data of E. coli evolving in glycerol and lactate can be found at NCBI GEO under the accession code GSE33147. Gene expression data of E. coli strain K-12 MG1655 adapting to 42 °C were from Data S3 in Sandberg et al.³⁴. Gene expression data of E. coli strain B REL1206 adapting to 42 °C were from Rodríguez-Verdugo et al.³⁵. Gene expression data of E. coli LTEE can be found at GSE164308. Gene expression data of E. coli evolving in 11 harsh environments were from Table S2 in Horinouchi et al.³⁶ and at GSE89746. Gene expression data of E. coli mutation accumulation can be found at GSE260863. Yeast gene expression data were from Dhar et al.³⁹. Fly gene expression data were from Mallard et al.⁴⁰. Copepod gene expression data were from Brennan et al.⁴³. Guppy gene expression data were from Ghalambor et al.⁴⁴. Morning glory gene expression data were from Josephs et al.⁴⁵. Computer code and relevant data for producing all figures presented in the main text and supplementary materials are available via Zenodo (10.5281/zenodo.17965753)⁶³.

Code availability

Custom code can be found via GitHub (https://github.com/jiachenli12/evo_repeatability/tree/v1.0).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-68838-x.