Introduction

It is often convenient to describe a mutation as deleterious, neutral, or beneficial. In fact, there exists a continuous distribution of fitness effects caused by mutations, from lethal to highly beneficial, and every value in between. This distribution of fitness effects is important for many theories in the fields of genetics and evolution (Whitlock et al. 1995; Fenster et al. 1997; Ostman et al. 2012). For example, the rate at which a population adapts to an environment depends on the frequency of beneficial mutations relative to the more common deleterious and neutral mutations (Desai et al. 2007). In addition, whether a mutation is beneficial or deleterious depends on the presence of other mutations (genetic background). Epistasis is a term used broadly to describe instances when the effects of combinations of mutations are not easily predicted by the effect of each individual mutation. Positive epistasis is used to describe situations when combinations of mutations produce higher fitness than expected from their individual effects. Negative epistasis occurs when multiple mutations produce lower fitness together than expected from their individual effects (Fig. 1). In some situations, a mutation can even be beneficial in some genetic backgrounds and deleterious in others. This is referred to as sign epistasis, and highlights the difficulty in classifying mutations as beneficial, deleterious, or neutral. Sign epistasis has received considerable attention because it can make certain sequential combinations of mutations unlikely to persist in a population, even though they may be subsets of extremely beneficial combinations of mutations. The presence of sign epistasis means that some pathways to higher fitness are highly improbable (Weinreich et al. 2005; Weissman et al. 2009). Populations evolving in the laboratory, as well as in the wild, must navigate the peaks and valleys of fitness landscapes caused by selection pressure and epistatic interactions. Our ability to forecast evolutionary outcomes will require advancements in our understanding of epistasis within and between genes.

Fig. 1
figure 1

Experimental approaches to uncovering epistatic mutational interactions. Mutation accumulation and curve fitting. A Predominantly negative (red) or positive (green) epistasis can be detected by mutation accumulation followed by non-linear curve fitting. The average fitness (w) is determined for populations of genotypes with a given number of mutations (n). No epistasis is inferred by β = 1 (black curve). Decreasing β results in positive epistasis, while increasing β results in negative epistasis. All curves have the same average mutation effect (α = 0.2). B Examples of curves with no epistasis (β = 1) and different values for α. Comparing effects of individual versus pairs of mutations. Mutation effects are compared to a wild-type reference (ab). Two mutations are indicated as a to A and b to B. Assuming that the effects of each mutation is multiplicative, epistasis (gray box) is identified as a deviation from this prediction. Positive epistasis is observed as higher than expected fitness. Negative epistasis is observed as lower than expected fitness. C Epistasis between deleterious mutations (red). D Epistasis between beneficial mutations (blue)

Full size image

Two experimental approaches for detecting epistasis are commonly used (Fig. 1). One way is by determining the average fitness effects of increasing numbers of mutations relative to a reference. The genetic variants are often generated in the laboratory by mutation accumulation, utilizing extreme population bottlenecks to enhance genetic drift and force random mutation fixation (Halligan and Keightley 2009). Mutations are more frequently deleterious than beneficial, and the random accumulation thus causes a decline in the average fitness of individuals with increasing numbers of mutations. The decline can be expected to follow a simple exponential curve if epistasis is absent or if there exists a balance of positive and negative interactions (note that fitness decline will follow a linear curve for logarithmically transformed fitness values). Deviations from a simple exponential can identify epistasis when it is predominantly negative or positive (Fig. 1A). A second experimental approach to revealing epistasis is a pairwise comparison of mutational effects. For this, the effect of several individual mutations is determined and compared to pairs of these mutations. The epistasis between each pair of mutations can be determined by comparing their actual fitness effect relative to what is expected from each individual effect (Fig. 1). Both of these types of experiments have been conducted on the scale of individual genes (Bershtein et al. 2006; Hayden et al. 2015) and whole genomes (Kouyos et al. 2007; Halligan and Keightley 2009).

These experimental approaches can be thought of as ways to understand the complex mapping of genotypes to phenotypes. Ideally, a fitness value could be assigned to every possible genotype, generating a comprehensive mapping of genotype to phenotype. However, this is not possible because of the vast numbers of potential genotypes. For example, a genome of length 100 bp has 4100 > 1060 possible sequence variants. As the length of the genome increases, or the number of mutated sites increases, the effort required to identify, isolate, or synthesize, and assay each variant increases exponentially. For this reason, individual gene products, including proteins and non-coding RNA (ncRNA) molecules, are an attractive model system (Jiménez et al. 2013). For the smaller genotype space of genes, a large subset of interesting mutations can be identified and many possible combinations can be produced and assayed. Understanding epistasis within genes has relevance to directed evolution approaches aimed at optimizing gene functions (Bloom and Arnold 2009). In addition, as the expressed components of genomes, the genotype to phenotype maps of individual genes may inform our understanding of epistasis at the level of whole genomes (Soskine and Tawfik 2010). A better understanding of epistasis may improve evolutionary theories required to assess the vast majority of genotype space that will not be studied with these experimental approaches. In the study of epistasis within a gene product, protein molecules have been the subject of the majority of empirical investigations. However, our growing understanding of the numerous and critical roles carried out by ncRNA molecules warrants investigation of how mutational effects contribute to the evolution of new RNA functions.

Several recent experiments have produced extensive data on epistatic interactions in ncRNA molecules (Fig. 2). Each study produced large numbers of mutational combinations and utilized a high-throughput assay for the effects of individual and combinations of mutations. Two studies used ribozymes (RNA enzymes), where fitness is defined as the ribozyme activity of a variant relative to that of a wild-type reference. In these experiments ribozyme fitness was determined in vitro, outside of a cellular context. Two other studies transformed yeast with libraries of specific ncRNA molecules, and used the RNA-dependent growth rate as the in vivo fitness metric. Here we characterize the epistatic mutational effects observed in each of these experimental systems. On average, all of the RNA molecules show a predominance of negative epistatic interactions between random mutations, whether the assay was carried out in vivo or in vitro. Underlying this average effect are similarities in the distribution of individual mutational effects and the distribution of pairwise mutational interactions. The predominance of negative epistasis is observed despite the fact that the RNA molecules reviewed have very different structures, and were studied with different experimental approaches. Similar epistasis has been seen in protein enzymes, suggesting that negative epistasis between mutations within genes may be a common property of biological parts (Bloom et al. 2004; Bershtein et al. 2006; Wylie and Shakhnovich 2011).

Fig. 2
figure 2

Structures of the RNA molecules. The size of the RNA, name, and genomic source is given below each structure. The assay conditions used to measure mutational effect (in vitro vs. in vivo) are also indicated. Structures were rendered in Pymol. Crystal structure coordinates are from the Oryza sativa Twister ribozyme (4OJI), the Azoarcus group I intron (1ZZN), an Asparagine tRNA from yeast (1VTQ). The U3 snoRNA structure is taken from the context of a cryo-EM structure of the 90S pre-ribosome (5JPQ). Structures are not to scale

Full size image

The RNA Molecules and Experimental Approaches

We have collected data from several recent experiments in order to facilitate direct comparison between different RNA molecules with different functions and between in vitro and in vivo experiments. We will briefly describe some of the pertinent details of each experimental system. The in vitro ribozyme data are from a 54-nt self-cleaving ribozyme that belongs to a structural family of ribozymes called Twister (Roth et al. 2014) and a 197-nt Azoarcus self-splicing group I intron ribozyme (Reinhold-Hurek and Shub 1992). For the Twister ribozyme, the mutants were generated by gene randomization during DNA synthesis, and therefore the analysis included all single and double mutants of the wild-type ribozyme, as well as a random sampling of sequences with three or more mutations (Kobori and Yokobayashi 2016). The fitness of each ribozyme was determined by the amount that each variant was able to self-cleave during in vitro transcription. All variants were transcribed simultaneously, and high-throughput sequencing was used to determine the fraction of each variant found in the cleaved form. Data were openly shared for the single and double mutants, enabling nearly exhaustive analysis of interactions between the effects of pairs of mutations. However, the average fitness of combinations of three or more mutations was not available. In contrast, the data for the Azoarcus ribozyme included average mutational effects even for high numbers of mutations, but the individual and pairwise effects were not exhaustively determined. For the Azoarcus ribozyme, average fitness effects were determined by producing separate populations of ribozymes with incrementally increasing numbers of mutations by consecutive rounds of error-prone PCR, and then determining the activity of each population relative to the wild-type ribozyme (Hayden et al. 2015). The number of mutations in each population was determined by high-throughput sequencing.

The ncRNA molecules studied in vivo include the Arginine tRNACCU, and the U3 small nucleolar RNA (snoRNA), both from Saccharomyces cerevisiae. The sequence variants, in both of these datasets were produced by randomization during DNA synthesis and contained most of the possible single and double mutants, and a random sampling of three or more mutations. The tRNA experiments used a yeast strain with this non-essential tRNA replaced by a tRNA sequence variant at its native genomic location (Li et al. 2016). The growth rate was determined for more than 105 yeast strains, each with a different tRNA variant. This was accomplished by counting the population frequency of each tRNA before and after growth competition using high-throughput sequencing. The enrichment rate of each tRNA variant is assumed to arise from the growth advantage, or disadvantage, provided by the specific tRNA. The experiments were carried out at 37 °C in YPD media, conditions that inhibited the growth rate of a tRNACCU knockout strain to ~ 20% of the wild type. The U3 snoRNA experiments used a yeast strain with the single copy of the U3 gene under a galactose-inducible promoter (Puchta et al. 2016). This strain could grow in galactose media, but showed growth arrest when transferred to glucose. Growth in glucose was then recovered by transformation with a plasmid constitutively expressing the wild-type U3 sequence. Mutational variants were assayed by transforming this yeast strain with a library of U3 variants contained on a plasmid. Each plasmid also contained a unique 20nt DNA sequence or “barcode.” The barcodes linked to each U3 variant were verified in a separate sequencing reaction. This enabled only the barcodes to be sequenced during growth competitions. The population frequency of each barcode before and after growth competition was used to determine the fitness effect of each U3 variant associated with this barcode. For these in vivo experiments, the fitness distributions of mutations and epistasis were previously reported, but in different formats. The authors kindly shared their fitness data for this review, so that the data could be presented in a common format for comparison.

Fitness Declines to Accumulated Mutations Reveals Negative Epistasis and Robustness–Epistasis Link

Epistasis can be detected from the average effects of increasing numbers of mutations. Typically, this is done by categorizing many sequence variants based on the number of mutations per molecule n and then determining the average fitness at each value of n. The data are fit to the equation w(n) = exp(− αn β) (Wilke and Adami 2001). Upon fitting this equation to the data, the parameter α indicates the average deleterious effects of mutations, and determines how rapidly fitness declines as more mutations are introduced. Lower values of α require that higher values of fitness are preserved upon mutation, a property referred to as mutational robustness (Wagner 2005). The parameter β indicates epistatic interactions in the following way. If there are no epistatic interactions, or a perfect balance of positive and negative interactions, then β = 1 and the average fitness declines exponentially with increasing numbers of mutations. Values of β > 1 indicates a predominance of negative epistasis (Wilke et al. 2003), and the deviation from a pure exponential is such that the fitness of genotypes with multiple mutations is lower than expected from the average fitness at n = 1 (Fig. 1). Values of β < 1 indicates predominantly positive epistasis. In this case the decline in fitness is less rapid and multiple mutations are less deleterious than expected from the average fitness at n = 1.

In order to facilitate comparison, we have collected the data for each of the four RNA molecules, and plotted fitness as a dependent variable and number of mutations as the independent variable. We used non-linear least squares to fit the data to the above equation that models fitness decline as a function of the number of mutations (Fig. 3). All four of the RNA molecules analyzed in this way show a predominance of negative epistasis with β > 1. For comparison, we have also plotted fitness decline curves without epistasis (β = 1), but with similar deleterious effects of individual mutations (Fig. 3 dashed lines). Similar epistasis is observed for the Azoarcus ribozyme (β = 1.3 ± 0.20) and the U3 snoRNA (β = 1.2 ± 0.02), and the curves are nearly overlapping. More extreme average epistasis is seen in the Twister ribozyme (β = 1.4 ± 0.11) and the tRNACCU (β = 2.7 ± 0.02). Taken together, all molecules fit into the previously observed negative correlation between α and β (van Nimwegen et al. 1999; Wilke and Adami 2001; Bershtein et al. 2006), in that all molecules show relatively high robustness (α < 0.6) and negative epistasis (β > 1). It is important to point out that we are only noting general trends in the data, and have not evaluated the significance of the differences in α and β between the datasets. This is especially true for the Twister data, which come from curve fitting to only three available data points of average fitness at n = 0, n = 1, and n = 2.

Fig. 3
figure 3

Decline in the average fitness of ncRNA variants caused by increasing numbers of mutations. The average fitness w is plotted as a function of the number of mutations per molecule n determined by the number of nucleotide differences relative to a wild-type reference sequence. Solid lines represent epistatic equations of the form w(n) = exp(− αn β), with parameters that produce the best fit to the experimental data by non-linear least squares curve fitting (Python). Individual data points are excluded for visual clarity; however, the full data sets are included in the Supplemental Material (Fig. S2). For comparison, dashed lines show curves with no epistasis β = 1, and an activity at n = 1 similar to the twister ribozyme (gray-dashed line) or the other three RNA molecules (blue-dashed line). Data from the computational folding of RNA sequences (RNA comp) with positive epistasis are also shown for comparison (green)

Full size image

Distributions of Individual Mutational Effects and Pairwise Interactions

Underlying the average epistatic effects described above is a range of fitness effects from individual mutations, and combinations of mutations. For example, mutational robustness (small α) could come from either many mutations with neutral effects, or a balance of beneficial and deleterious effects. Similarly, average negative epistasis could arise from a combination of positive and negative epistatic effects, if the negative epistatic effects are either more frequent or more extreme than the positive epistatic effects. Fortunately, three of the datasets (snoRNA, tRNA, and Twister) reported the fitness consequence of nearly every possible single and double mutation. This allows us to compare the distributions of fitness effects and pairwise epistatic interactions in each of these RNA molecules. The snoRNA data were reported as log(w), and we transformed this to w to facilitate comparison. In addition, we normalized all the data so that the variants with the lowest detected fitness had fitness = 0 and the wild-type variants all had fitness = 1.

We show the distribution of fitness effects caused by mutations as histograms (Fig. 4). The distributions of individual mutational effects are quite similar, despite the differences between the size and structures of these ncRNA molecules (Fig. 2), as well as the differences in experimental approaches. All three distributions are characterized by a large number of neutral mutations, indicated by the modal peak at fitness = 1. There is a long tail of deleterious mutational effects (0 < fitness < 1). There are very few beneficial mutations (fitness > 1), although more were detected for the U3 snoRNA. The U3 snoRNA shows the highest fraction of neutral mutations, which is consistent with the lowest α values obtained from curve fitting. In addition, the Twister ribozyme has a higher fraction of extremely deleterious effects (fitness < 0.2), which is consistent with the larger values of α obtained by curve fitting. Taken together, the differences in the robustness of these ncRNA molecules to the effects of individual mutations is observable in the differences in the distributions of fitness effects.

Fig. 4
figure 4

Distribution of fitness effects and pairwise epistasis. A Distributions of fitness effects for individual mutations (dark blue) and pairs of mutations (light blue) for tRNACCU (Li et al. 2016). B Distributions of individual (dark red) and pairwise (light red) mutational effects in the U3 snoRNA (Puchta et al. 2016). C Distributions of individual (dark gray) and pairwise (light gray) mutational effects in the Twister ribozyme (Kobori and Yokobayashi 2016). D Distribution of epistatic values for pairs of mutations for the tRNACCU. E Distribution of epistatic values for pairs of mutations for the U3 snoRNA. F Distribution of epistatic values for pairs of mutations for the Twister ribozyme. Epistatic values were calculated as ε = log10 (W AB × W wt/W A × W B), where W A and W B are the fitness of RNA variants with a single mutation, W AB is the fitness of the variant with both mutations, and W wt is the fitness of the wild type. All distributions are set to the same scale on the x-axis. Inset in F shows the full distribution of epistatic effects in the Twister ribozyme

Full size image

We also plotted the distribution of effects caused by two mutations as lighter colored bar graphs in Fig. 4. The distributions are clearly shifted to the left relative to the single mutation distributions, which indicates that two mutations are typically more deleterious than one mutation. This is expected, and in itself does not indicate epistasis. However, the differences between the distributions of each RNA molecule are informative. We will describe the change that is observed for each ncRNA by comparing the distribution of single mutational effects to double mutational effects, i.e., from the dark to light histograms in Fig. 4. The Twister data (Fig. 4E gray) show a dramatic increase in the number of non-functional variants, such that the modal fitness actually changes from fitness = 1 to fitness = 0 when comparing single to double mutations. In contrast, the most apparent change in the tRNACCU data (Fig. 4A blue) is a dramatic decrease in the proportion of neutral variants (fitness = 1). The distribution of double-mutant fitness values for the tRNACCU data is more broadly distributed over intermediate non-zero values, as compared to the Twister data. Finally, the U3 snoRNA data have the least pronounced change, with a slight increase in non-functional variants as well as intermediate low fitness variants, but the model peak near fitness = 1 remains. The cause of differences in the fragility of the different RNA molecules to the effects of two mutations remains unknown, but may involve the thermodynamics of the specific structure, or differences in the environment, such as the presence of chaperone proteins in vivo, which will be discussed in more detail below.

To detect epistasis in this data requires a comparison between the measured effects of pairs of mutations and what would be predicted from the effects of each individual mutation (see Fig. 1 for explanation). The calculated pairwise epistatic values are plotted as histograms in Fig. 4. All distributions indicate that there exists both positive and negative epistasis. However, the mean and skew of the data support a predominance of negative epistasis. The distribution of effects in the Twister ribozyme is much broader than the other two distributions, with a heavy tail in the negative direction. The nearly balanced distribution of positive and negative effects in the pairwise interactions of U3 snoRNA emphasizes the importance of higher-order epistatic interactions in this ncRNA (Weinreich et al. 2013; Sailer and Harms 2017), and is consistent with the curve fitting analysis to the average effects of higher numbers of mutations (Fig. 3).

Discussion

Laboratory and Computational Studies Find Predominantly Different Epistatic Interactions

The predominance of negative epistasis is notable because it differs from previous results from computational folding of RNA that uncovered a predominance of positive epistasis (Fig. 3 green curve). Specifically, Wilke et al. studied the computationally predicted folding of 76nt long RNA molecules (Wilke et al. 2003). They randomly generated 100 reference sequences of this length, and determined their computationally predicted secondary structure. Then they produced sequences at incrementally increasing numbers of mutations relative to each reference. They defined fitness as the fraction of sequences that folded into the same structure as the reference, and determined fitness for up to 106 sequences at each number of mutations. They fit this data to the epistasis equation used above, and extracted α and β parameters. In contrast to the laboratory studies reviewed here, all values of β fell into the positive epistasis range (β < 1), and all molecules also showed a lower tolerance to mutations (α > 0.6). One intriguing possibility for this difference is that natural selection has favored molecules with many mutational neighbors that maintain function, leading to the evolution of mutational robustness in naturally occurring RNA (Meyers et al. 2004; Kun et al. 2005). The random sequences used in the computational studies would not be expected to have mutational robustness if it is the product of evolution. Future high-throughput experiments with natural and artificially selected ribozymes could be designed to directly test this hypothesis. Another possible explanation for this discrepancy is that computationally folded RNA secondary structures are bimodal and predict either a fitness of 1 or 0 (Wilke et al. 2003). The empirical fitness measurements, on the other hand are continuous, and a large fraction of the fitness effects fall into intermediate values of fitness effects. This lack of intermediate values in computational structure prediction could lead to an overestimation of deleterious mutational effects, and an underestimation of negative epistasis.

As noted, several previous studies involving both theoretical prediction and experimental data found a negative correlation between the parameters α and β (Wilke and Adami 2001; Bershtein et al. 2006; Hayden et al. 2015). This suggests that epistasis becomes more predominantly negative as the average mutational effects are decreased. The data previously reported for the Azoarcus ribozyme involved three different conditions where selection pressure was intentionally decreased. As expected, decreased strength of selection resulted in increased β, and decreased α (Hayden et al. 2015). Therefore, it appears that the negative correlation between α and β holds whether the intensity of mutational effects is altered by changes in the specific molecule, such as the different RNA molecules reviewed here, or changes in the environment.

A direct comparison of in vivo fitness and in silico prediction was also previously reported for the specific Arginine tRNACCU (Li et al. 2016). The authors found that the propensity to fold properly is a poor indicator of fitness in their experiments. They found that their strain of S. cerevisiae is more robust to mutations in this particular tRNA than predicted from computational folding of these sequences. This could indicate that this tRNA can still function properly despite small deviations from a canonical structure, such as a single missing base pair. This would be somewhat surprising given the numerous interactions involving tRNA molecules during its lifetime in the cell (Maraia and Arimbasseri 2017). In addition, other tRNAs can also decode this codon, and it is possible that only a small amount of properly folded tRNACCU is required to recover normal growth. The AGG codon that is decoded by tRNACCU is the second most common Arginine codon in S. cerevisiae, representing ~ 20% of the codons for this amino acid (Cherry et al. 2012).

Another possibility is that RNA chaperones alter the folding such that some of sequence variants fold into the native structure even when it is not the most stable (MFE) structure. It is well established that an RNA chaperone called the La protein can assist the proper folding of several RNA molecules, including tRNAs, and has been shown to hide the deleterious effects of point mutations (Chakshusmathi et al. 2003). In addition, several RNA chaperones have been shown to facilitate the native folding of numerous RNA molecules (Herschlag 1995; Russell 2008), including several ribozymes (Herschlag et al. 1994; Halls et al. 2007; Sinan et al. 2011). The U3 snoRNA is a part of an RNA–protein complex (SSU processome) that processes ribosomal RNA from primary transcripts. Assembly of this complex has been shown to involve several RNA chaperones and helicases (Soltanieh et al. 2015; Hunziker et al. 2016). In addition to normal folding pathways, several chaperones have recently been shown to buffer the deleterious effects of many different mutations (Rudan et al. 2015). Similar to chaperones for protein folding, such as HSP90, RNA chaperones could enable a property referred to as phenotypic capacitance, where otherwise deleterious mutations are maintained in the population, providing the potential to produce novel adaptations upon environmental change (Rutherford and Lindquist 1998; Queitsch et al. 2002; Jarosz and Lindquist 2010). In fact, this has recently been proposed as a factor in tRNA diversification in Eukaryotes (Maraia and Arimbasseri 2017). Despite this critical role of RNA chaperones in the mapping of RNA genotypes to phenotypes, the role of RNA chaperones in promoting the evolution of novel ncRNA structures and functions remains poorly understood.

However, the prevalence in negative epistatic interactions for RNA molecules studied in vitro and in vivo suggests a common mechanism that cannot be explained by the different experimental environments. For example, besides the presence of chaperone proteins, negative epistasis in vivo could result from a cooperative destabilization of the multi-component complexes involving the studied RNA molecules. While this contribution is not ruled out, it cannot be the cause of negative epistasis for the ribozymes studied in vitro where only the RNA is present. This suggests that a property of the RNA structures themselves can account for negative epistasis. RNA structures may provide buffering against individual mutations, yet be sensitive to many mutations. We note that protein enzymes have also shown a predominance of negative epistasis that has been attributed to the crossing of a thermodynamic threshold (Bershtein et al. 2006). The combined results suggest that a similar phenomenon is occurring in these RNA structures. A thermodynamic framework may provide a prediction of epistasis within gene products and multi-component RNA–protein complexes.

It is interesting to note that several previous studies have found a predominance of positive epistasis in the genomes of RNA viruses when they are randomly mutated away from the wild type (Bonhoeffer et al. 2004; Sanjuán 2010). The distribution of individual mutational effects in RNA viruses has also been studied, and uncovered a very high fraction of mutations with a lethal effect (fitness = 0). For example, about 40% of mutations were found to be lethal in both the tobacco etch virus and vesicular stomatitis virus (Sanjuán et al. 2004; Iglesia and Elena 2007). These findings are consistent with the negative correlation between epistasis and robustness (larger α and β < 1). More generally, the distribution of fitness effects in viral genomes (Sanjuán 2010) appears to be quite similar to findings in some cellular genomes (Eyre-Walker and Keightley 2007), in that both show large fractions of lethal mutations, and predominately positive epistatic interactions or no predominant direction of epistasis (Elena and Lenski 1997; He et al. 2010). Individual proteins, on the other hand, have fitness distributions very similar to those seen here in RNA molecules, with very few lethal mutations and negative epistasis (Soskine and Tawfik 2010). Given the apparent robustness of individual macromolecules, and the multiple forms of robustness in living organisms (Wagner 2011), the mutational targets that are the source of lethal mutations in the genomes of viruses and cellular organisms is not completely understood. It is possible that these few well-studied examples of proteins and RNA are not typical in their distribution of fitness effects, but this seems unlikely given the variety of forms and functions as well as the multiple ways they were assayed. Another possibility is that pleiotropic mutations that affect multiple traits or processes are more likely to be lethal. More research is needed to understand the mechanisms and implications of differences in the mutational effects in organisms and their genetically encoded biological parts.

The presence of negative epistasis is important for models that attempt to explain the evolution and maintenance of recombination and sexual reproduction (Kouyos et al. 2007). The RNA molecules reviewed here meet several of the requirements for this theory, namely a predominance of negative epistasis, correlation between epistasis and strength of selection, and probably tight linkage between mutations within the small RNA molecules. At first these small RNA genes seem unlikely to be a large enough target for recombination to explain the evolution of recombination at the organismal level. However, when one considers the pervasiveness of transcription, and the discovery of long non-coding RNA, the breaking up of linked mutations with negative epistasis within RNA molecules cannot be ruled out as a driving force for the evolution of recombination. Interestingly, it has been proposed that genetic recombination may have contributed to the origin of life (Lehman 2003). In fact, RNA recombinase ribozymes have been demonstrated in the laboratory as models of the RNA World versions of modern protein recombinase enzymes (Hayden et al. 2005; Vaidya et al. 2012; Pesce et al. 2016). The predominance of negative epistasis in the RNA molecules reviewed here suggest that the earliest RNA genomes could have benefited from recombination, and supports the theory that genetic recombination may be as old as life itself.