Experimental estimates of germline mutation rate in eukaryotes: a phylogenetic meta-analysis

Abstract

Mutation is the ultimate source of all genetic variation, and over the last 10 years the ready availability of whole-genome sequencing has permitted direct estimation of mutation rate for many non-model species across the tree of life. In this meta-analysis, we make a comprehensive search of the literature for mutation rate estimates in eukaryotes, identifying 140 mutation accumulation (MA) and parent–offspring (PO) sequencing studies covering 134 species. Based on these data, we revisit differences in the single-nucleotide mutation (SNM) rate between different phylogenetic lineages and update the known relationships between mutation rate and generation time, genome size, and nucleotide diversity—while accounting for phylogenetic nonindependence. We do not find a significant difference between MA and PO in estimated mutation rates, but we confirm that mammal and plant lineages have higher mutation rates than arthropods and that unicellular eukaryotes have the lowest mutation rates. We find that mutation rates are higher in species with longer generation times and larger genome sizes, even when accounting for phylogenetic relationships. Moreover, although nucleotide diversity is positively correlated with mutation rate, the gradient of the relationship is significantly less than one (on a logarithmic scale), consistent with higher mutation rates in populations with smaller effective size. For the 29 species for which data are available, we find that indel mutation rates are positively correlated with nucleotide mutation rates and that short deletions are generally more common than short insertions. Nevertheless, despite recent progress, no estimates of either SNM or indel mutation rates are available for the majority of deeply branching eukaryotic lineages—or even for most animal phyla. Even among charismatic megafauna, experimental mutation rate estimates remain unknown for amphibia and scarce for reptiles and fish.

Introduction

The per-nucleotide per-generation de novo mutation rate, µ, is a key parameter in population and evolutionary genetics, appearing either alone or combined with effective population size (N_e) in the compound parameter θ. However, it seems implicit that µ is of no interest to many evolutionary studies; researchers often use µ from one species when analyzing another—as if µ were exchangeable in a way that N_e is not (e.g., Wilding, 2017). This is partly because the historically high cost of sequencing made estimates of µ prohibitively expensive, but may also be because variation in µ among species is widely thought to be negligible compared with variation in N_e. But, if the mutation rate varied as widely among multicellular eukaryotes as many other traits, there would be much less need to invoke variation in N_e when explaining variation in (e.g.) genetic diversity. One reason why µ may vary relatively little is if it is minimized to a lower limit imposed by drift: the “drift barrier” hypothesis (Lynch, 2008, 2010). However, it might be that in reality mutation rate is optimized to well above the drift-barrier limit by selection in favor of adaptability or the cost of replication fidelity (Kimura, 1967; Liu & Zhang, 2021; Peck & Lauring, 2018). If this were the case, then we might find that mutation rates varied dramatically, even among closely related taxa.

Early estimates of the mutation rate leveraged visible phenotypes, such as the occurrence of genetic diseases in humans (Nute & Stamatoyannopoulos, 1984; Trimble & Doughty, 1974). For example, using hemoglobin M disease, Stamatoyannopoulos and Nute (1982) estimated the mutation rate in the human hemoglobin beta gene at 7.4 × 10⁻⁹ per bp per generation. However, such estimates are prone to bias, for example, if the mutation is required to be autosomal dominant and nonlethal (Nute & Stamatoyannopoulos, 1984), and such approaches are not easily applicable in organisms for which disease is hard to observe. More commonly, using the equivalence between mutation and neutral substitution rates (Kimura, 1968), many studies have used time-calibrated sequence divergence to estimate µ. For example, comparison between humans and chimpanzees gives rise to estimates on the order of 10⁻⁸ per bp per generation (Drake et al., 1998; Kondrashov & Crow, 1993; Nachman & Crowell, 2000). However, such indirect estimates are also limited (Kondrashov, 2003). First, because generation time and calibration dates are prone to substantial uncertainty (e.g., Obbard et al., 2012). Second, because it is hard to distinguish between unconstrained and weakly constrained sites (Harmon et al., 2021). Third, because mutation saturation in hotspots may lead to an underestimation, especially for sequences with a long divergence time (Sigurðardóttir et al., 2000). In addition, phylogenetic-calibration approaches reflect the long-term mutation rates, averaging across biologically important factors such as generation time, sex differences, life-stage differences, and interindividual variation.

In contrast to phenotypic and phylogenetic approaches, mutation–accumulation (MA) and parent–offspring (PO) sequencing can provide direct estimates of the per-generation mutation rate. MA utilizes the accumulation of spontaneous mutations in a single inbred or asexual genome over multiple generations, by minimizing the effectiveness of natural selection (Halligan & Keightley, 2009). Given the low mutation rates in eukaryotes, the multiple generations in an MA experiment allow sufficient mutations for them to be detected at the end. Although the first use of MA to estimate µ phenotypically dates back to a study by Mukai and Cockerham (1977), the availability of whole-genome sequencing and its decreasing price have made sequencing approaches more widely applicable. MA best suits easily maintained species with short generation times, for example, nematodes (Denver et al., 2004; Konrad et al., 2019; Meier et al., 2014; Saxena et al., 2019), fruit flies (Haag-Liautard et al., 2007; Huang et al., 2016; Keightley et al., 2009; Schrider et al., 2013), green algae (Lopez-Cortegano et al., 2021; Ness et al., 2012), and water fleas (Ho et al., 2020; Keith et al., 2021). However, MA may under- or overestimate mutation rates (Lynch & Walsh, 1998) because natural selection will still come into play for mutations that are very highly deleterious, such as those that cause complete sterility or are lethal (Baer et al., 2007).

By sequencing parents and offspring, and counting the mutations arising over a single generation, the PO approach avoids purging deleterious mutations, except those that are dominant lethal, and so provides the most direct estimate of mutation rate. It is not limited to model species (Keightley et al., 2014; Yang et al., 2015) and is suitable for large animals outside of the laboratory (Harland et al., 2017; Wang et al., 2022). PO sequencing has been particularly widely used in humans, where MA is not applicable. However, the challenge of robustly identifying new mutations in natural heterozygous genomes is formidable. First, because the number of mutations arising within one generation is small when compared with the impact of sequencing/mapping errors and the genetic variation of heterozygotes. Second, because unless sequencing coverage is high or multiple offspring are sequenced, pre-existing heterozygous sites in parental genotypes can be missed. For example, Bergeron et al. (2022) compared the estimates obtained by five different research groups for the same family of rhesus macaques and found that the highest estimates could be twice the lowest ones. Minimizing the rates of false positive and false negative mutation calls is therefore critical for the PO approach (Yoder & Tiley, 2021).

Relaxed-clock phylogenetic studies have generally indicated there is substantial among-lineage variation in the evolutionary rate, suggestive of variation in mutation rates among taxa (Ho et al., 2015). But such studies cannot separate the impact of generation time and mutation rate and may be biased by variation in the action of (weak) selection (Harmon et al., 2021). As more direct experimental estimates have become available, it has become possible to draw comparisons within particular clades such as primates (Chintalapati & Moorjani, 2020) or vertebrates (Bergeron et al., 2022, 2023; Yoder & Tiley, 2021). However, there has not been a wider analysis, nor one that includes both PO and MA studies.

In addition, many previous studies have not taken an explicitly phylogenetic analytical approach (e.g., Katju & Bergthorsson, 2019). This could in principle lead to pseudoreplication in the analyses that is caused by the nonindependence among taxa that arises from their shared ancestry (reviewed in Freckleton, 2009). However, this problem can be mitigated by explicitly modeling covariance among species that is contributed by their patterns of relationship, and generalized linear mixed modeling approaches are available to do this (Hadfield & Nakagawa, 2010; Halliwell et al., 2022; Lynch, 1991).

Here we perform a comprehensive phylogenetic meta-analysis of published eukaryotic single-nucleotide mutation (SNM) and insertion–deletion (indel) mutation rate estimates that use direct sequencing of multiple generations, through either MA or PO comparisons. We focus on the variation in mutation rate among species, and the relationship between mutation rate estimates and the experimental method, generation time, genome size, and genetic diversity.

Methods

We made an exhaustive literature search for publications on the experimental estimation of mutation rate. We first searched Clarivate Web of Science up to 21 September 2022 for “Title/Keywords/Abstract” fields containing (“mutation rate” | “mutation rates” | “mutational rate” | “mutational rates”). As much of the resulting literature was related to somatic mutations (e.g., cancers) or virus mutations, we excluded “(‘tumor’ | ‘cancer’ | ‘clinical’ | ‘virus’)” in the search field of “Title/Keywords/Abstract.” This search strategy resulted in 9,462 studies with the earliest one dated 1928. We also searched for references to mutation rate estimates in other papers not captured by our search. As bacteria are outside of our study, we manually removed studies related to bacterial mutations. These filters resulted in 174 studies. Of these studies, we further excluded 20 studies that used phylogenetic approaches to infer mutation rates, 14 studies that used phenotypic markers, and 2 other studies that used recombination landscape or the site frequency spectrum to infer mutation rates statistically. In addition, we included one more recent study published after our initial submission (Bergeron et al., 2023), which estimated the mutation rates for 68 vertebrates using PO sequencing. This resulted in a total of 140 studies in our final data set.

From each of the potential studies, we extracted the year of the publication, the species, the population, reproductive mechanism (sexual, asexual), mutation type (SNM, short insertion, short deletion), the units of mutation rate (per bp per year, per bp per generation), the mean and 95% CI of mutation rate, the number of identified mutation events, the number of callable sites, the methods used to infer mutation rate (PO, MA, or others), the sample size, the number of generations (if MA), the sequencing technology, the sequencing depth, and other pertinent observations. Note that for multigeneration studies, such as MA, the number of “callable sites” incorporates the possibility that a mutation may have arisen in any of several generations (i.e., the product of genomic sites and the number of generations). If a 95% confidence or credible interval was not provided in the original study, we calculated a 95% confidence interval using “binconf()” from the Hmisc R package (version 4.7-0), based on the number of mutation events and callable sites and assuming a binomial distribution (Girard et al., 2011). If a study did not provide the mutation count, we inferred an effective number of mutation events based on the reported mutation rate and its confidence interval. For studies in which the number of callable sites was not reported, an estimate was made from the number of mutation events and the reported mutation rate. If a study only reported the mutation rate, we assumed one mutation event and calculated the corresponding callable sites based on mutation rates. This latter approach will necessarily underestimate the precision with which rates are estimated (minimizing their weight in the analysis), while keeping the estimate in the meta-analysis with unchanged mean values.

To compare the SNM rates across different species, we conducted a phylogenetic meta-analysis (i.e., a combined analysis of published data) using the mutation data from MA or PO studies. We obtained the phylogenetic relationships and relative divergence dates from http://www.timetree.org/ (Kumar et al., 2022) and visualized the tree using ggtree (Yu, 2020). We performed the analyses using a phylogenetic generalized linear mixed model (PGLMM), with a unit-scaled phylogeny as a random effect, assuming a Poisson distribution for the occurrence of de novo mutations offset by the number of callable sites (i.e., enforcing a directly proportional relationship between the expected number of mutations and the number of sites observed). The analysis was conducted using the Bayesian mixed-model R-package MCMCglmm (version 2.32) (Hadfield, 2010) (see Supplementary Materials). To investigate the effects of methods (MA or PO) used to estimate mutation rate, we additionally included “Method” as an additional fixed effect in the above model.

We also investigated the relationships between mutation rate and generation time and genome size using univariate phylogenetic mixed models that included these factors separately as fixed effects and a phylogenetic species term as a random effect. A log-transformation was made on these predictors before fitting the model (see Supplementary Materials), and to illustrate the results, model predictions were made conditioning on each of the different species as random effect levels. The estimated generation time for each species was gathered from the literature (see Data availability), and we obtained estimates for 116 species. Genome sizes were taken to be the genome assembly size and gathered from the NCBI database, which resulted in 108 species with genome size available.

To test for a relationship between nucleotide diversity (π) and the per generation per-site SNM rate (µ) while accounting for uncertainty in both, we used MCMCglmm to fit a bivariate PGLMM, with mutation number and π as responses, again including the number of callable sites as an offset on the number of mutations and generation time as a fixed effect (see Supplementary Materials). We collected nucleotide diversity data for 102 species from Buffalo (2021) and Bergeron et al. (2023) or the wider literature (see Data availability).

Results and discussion

Studies included

We searched for eukaryotic mutation rate studies in Clarivate Web of Science published prior to 21 September 2022, and then we manually examined these studies and identified those containing mutation rate estimates (see Methods). For the majority of our analyses, we also included data from one further recent study of 68 vertebrate species by Bergeron et al. (2023), published after our initial submission. In total, we included 140 studies using either MA or PO approaches, covering 134 species (see Data availability). Studies on primate mutation rate accounted for over a third of this literature, and 82% (45/55) of the primate studies were on humans. Arthropods accounted for 17% of the identified literature, followed by fungi (11%), plants (8%), nonprimate mammals (7%), and nematodes (5%) (Figure 1A). Excluding the recent study of Bergeron et al. (2023), we only identified four studies on fish, two studies on birds, and one study on reptiles. Across all groups, most studies focused on model species: yeast in fungi, Drosophila in arthropods, mice in mammals, and Arabidopsis in plants. These estimates tended to derive from MA analyses rather than PO studies (61 vs. 43 among nonhuman studies). However, as sequencing has become cheaper, there has been a rapid increase in the number of published experimental estimates of the mutation rate, with an average of 11 studies each year in eukaryotic species between 2010 and 2022 (Figure 1B).

Figure 1.

A summary of de novo mutation rate studies published up to 21 September 2022. (A) The proportion of studies in different taxonomic groups. (B) The number of studies by year. MA = mutation accumulation studies, PO = parent–offspring studies.

SNM rates across eukaryotes

To summarize the SNM rates in different lineages, we obtained a time-scaled phylogenetic tree of all represented species from http://www.timetree.org/ (Kumar et al., 2022) and used this to fit the covariance among species in a PGLMM using the Bayesian mixed model package MCMCglmm (Hadfield & Nakagawa, 2010). Treating species as a random effect, this provides a posterior estimate of mutation rate for each species, given the observations and relationships to other species (see Methods). In these data, the phylogenetic effect accounts for 96.4% of the random-effects variance, with a 95% highest posterior density (HPD) interval of 94.9%–97.9%.

Among all the studied species, a ciliate (Tetrahymena thermophila) had the lowest mutation rate, estimated at 0.01 × 10⁻⁹ per generation per base pair, while a fungus (Marasmius oreades) had the highest rate, estimated at 55.58 × 10⁻⁹—a difference of over 5,000-fold. The estimates for unicellular species were generally more than an order of magnitude lower than rates in multicellular species (below; Figure 2), even compared with species in the same clade, for example, fungi. This may either be attributable to there being a single cell division per generation, or possibly to a limit set by the drift barrier in species with larger N_e (Sung et al., 2012). We also estimated relatively low SNM rates for nematodes at 2.44 × 10⁻⁹, arthropods at 2.82 × 10⁻⁹, and fish at 5.55 × 10⁻⁹. For plants, the clade root was predicted as 3.66 × 10⁻⁹, but a large difference was observed between duckweed and other plant species (0.22 × 10⁻⁹ vs. 19.8 × 10⁻⁹). The high mutation rates in most plant species might be attributable to the lack of a segregated germline (Burian, 2021; Cruzan et al., 2022; Lanfear, 2018; Wang et al., 2019). Mammals had moderate mutation rates with a root estimate at 8.36 × 10⁻⁹ and species (excluding primates) ranging from 5.64 × 10⁻⁹ in pigs to 11.68 × 10⁻⁹ in killer whales. Primates had higher mutation rates, with a root estimate at 9.37 × 10⁻⁹. Of note, as the most studied species, the human mutation rate was estimated as 13.41 × 10⁻⁹, with 95% CI 11.61–15.28 × 10⁻⁹, which is also the highest among mammals. Similar to mammals, birds and reptiles had a high mutation rate, with root estimates at 10.43 × 10⁻⁹ and 12.41 × 10⁻⁹, respectively.

Figure 2.

A phylogenetic meta-analysis of SNM rates for different species identified from the literature. Mutation rates and 95% CIs were estimated using a PGLMM. For each annotated clade, we estimated the rate for their most recent common ancestor (clades marked in color; estimates in dashed lines and below taxon icons to the right). The vertical black dashed line indicates the mutation rate of common ancestor of the tree. PGLMM = phylogenetic generalized linear mixed model; SNM = single-nucleotide mutation.

To assess a possible impact of the experimental method on rate estimates, we included the experimental approach (MA or PO) as a fixed effect in the PGLMM. Overall, we did not find a significant difference between the two methods (MCMC p = .155). We then examined comparisons between MA and PO in fruit flies and mice more closely, as these two species have been investigated using both methods. The rate estimated using MA in fruit flies was 5.28 × 10⁻⁹, but 3.20 × 10⁻⁹ using PO (Supplementary Figure S1), and in mice, the MA estimate was 5.40 × 10⁻⁹, while the PO estimate was 4.05 × 10⁻⁹ (Supplementary Figure S2), and these differences were similarly not significant. Note, however, that the mutation rate may be heterogeneous among different populations: for Drosophila melanogaster, the estimated mutation rate from African populations was lower than that from European and North American populations (Chan et al., 2012; Wang et al., 2023). Making a more robust analysis would require samples of the same population but using different methods (MA or PO), which are not yet available.

Biological predictors of the SNM rate

Many biological factors have been proposed to be associated with mutation rate, including generation time (Bailey et al., 1991; Keightley & Eyre-Walker, 2000; Li et al., 1987; Wu & Li, 1985) and genome size (Drake, 1991; Lynch, 2010). To test for a dependence of mutation rate on these factors, we ran three further models: the first two including these factors each separately as an additional fixed effect in a PGLMM and the last one including both of them simultaneously.

We found a significant positive relationship with generation time (Bayesian PGLMM MCMC p < .0001; Figure 3A), such that a longer generation time predicted a higher per-site mutation rate per generation. If mutations resulted only from replication errors, and most species had a similar number of cell divisions per generation, then the per-generation mutation rate would not scale with generation time (Thomas & Hahn, 2014; Wu & Li, 1985). However, although few estimates are available, in reality the number of germline cell divisions per generation varies within and between species. In humans, there are around 401 cell divisions per 30-year generation in males and 31 in females (Drost & Lee, 1995; Ohno, 2019), but in mice, it is 62 per 9-month generation in males and 25 in females. In Drosophila, the sexes are much more similar, with an estimated 35.5 cell divisions in an 18-day generation for males but 36.5 in a 25-day generation in females in D. melanogaster (Drost & Lee, 1995). Within a species, a long generation time is likely to permit more cell divisions and thus a higher mutation rate per generation, and many studies in humans and primates have shown that mutations are heavily male biased and correlated with paternal ages (Kaplanis et al., 2022; Kong et al., 2012; Thomas et al., 2018; Wang et al., 2020).

Figure 3.

Phylogenetic linear regressions of SNM rate on (A) generation time for 116 species and (B) genome size for 108 species. The solid black lines represent simple linear regressions; the colored lines represent different PGLMM regressions conditional on each of the species random effects. (C) A regression of pairwise nucleotide diversity (π) on SNM rate (µ), fitted as a bivariate model in MCMCglmm for 102 species. The solid gray line indicates the model regression line; the dashed gray line indicates the expected slope for θ_π = 4N_eμ on the log scale. All axes are plotted on a log scale. PGLMM = phylogenetic generalized linear mixed model; SNM = single-nucleotide mutation.

Alternatively, mutations may not result predominantly from replication errors. Recent studies suggest that not all germline mutations track cell divisions, and DNA damage may contribute to the mutation rate (Wu et al., 2020; Wu & Przeworski, 2022). Independent of cell divisions, this mutation rate would be correlated with calendar time, and thus species with a longer generation time would have a higher rate. The relationship we observe between generation time and per-generation mutation rate could therefore be a consequence either of a greater number of cell divisions or of accumulating damage over time, and it will be hard to separate these potential causes given the limited data available on the number of germline cell divisions for different species (Gao et al., 2016; Wu & Przeworski, 2022).

We also found evidence that larger genome size predicts a higher per-generation mutation rate (MCMC p = .0020; Figure 3B). Mutation rate has previously been found to correlate negatively with genome size in prokaryotes, a phenomenon termed “Drake’s rule” (Bourguignon et al., 2020; Drake, 1991; Marais et al., 2020), but positively in eukaryotes (Lynch, 2010).

In each of the two models above, we fitted generation time and genome size individually as predictors of mutation rate in a regression model. However, these factors might be highly correlated: large organisms tend to have longer generation times and often larger genomes (Buffalo, 2021; Martin & Palumbi, 1993), and it is therefore hard to distinguish between casual relationships and confounding correlations. Including generation time and genome size together as fixed terms in a single model resulted in only generation time being significant with MCMC p < .001 (vs. MCMC p = .301 for genome size).

The drift-barrier hypothesis

The most direct test of the drift-barrier hypothesis would be a regression of SNM rate on N_e. However, long-term N_e can only be estimated from neutral genetic diversity (e.g., π_s) by making use of the relationship θ_π = 4N_eμ, and any uncertainty in μ could induce a spurious correlation in a regression of μ on π/4μ. Instead, error in both π and μ can be accounted for by fitting a bivariate linear mixed model, parameterized in terms of a regression of π on μ (see Methods). Then, on a logarithmic scale, the gradient of this relationship would be expected to be 1 if μ and N_e are independent of each other, and lower than 1 if species with large N_e tended to have lower μ. Nevertheless, such a relationship could be driven by other factors. For example, if species with small N_e tend to have a longer generation time, and a longer generation time causes higher mutation rates (i.e., Figure 3A), then a higher μ in species with low N_e could be driven by a mechanistic generation-time effect, rather than the drift barrier. We attempted to account for this by including generation time in our analysis as a predictor of both μ and π, thereby accounting for the direct impact of generation time on the relationship between them. In this bivariate phylogenetic model, μ increased with generation time (MCMC p < .001), whereas π did not (MCMC p = .501), and the gradient of the regression of π on μ was significantly less than 1 (gradient = 0.535, 95% CI [0.259, 0.832]; MCMC p = .004). This suggests that populations with larger N_e tend to have a lower mutation rate even after accounting for their shorter generation times—as predicted by the drift-barrier hypothesis (Lynch, 2010). Note that excluding generation time as a predictor resulted in an even stronger effect (gradient = 0.300 [0.098, 0.486]; MCMC p < .001; Figure 3C), whereas removing unicellular species from the analysis (because of their low mutation rates) resulted in a gradient that was still much less than 1 (gradient = 0.439 [0.155, 0.746]; MCMC p < .001). When we excluded the vertebrate data from the study of Bergeron et al. (2023), for which π_all but not π_s was available, the gradient did not increase (gradient = 0.494 [0.185, 0.829]; MCMC p = .002), demonstrating that this is not driven by constraint acting on coding sites in those data (perhaps expected, given the relatively low coding density in vertebrates).

The mutation rates of short indels across eukaryotes

Although the rate of indel mutation is much less commonly estimated than the SNM rate, our literature search identified estimates for 29 species. We found that the indel mutation rate across eukaryotes displayed a similar pattern to that of the SNM rate (Figure 4), with a correlation coefficient between the two of 0.75 (95% HPD CI: 0.44, 0.98; MCMC p < .001) using a PGLMM, with the SNM rate being generally higher than the indel mutation rate, except in two lineages (nematodes and amebae) for which the indel mutation rate was apparently higher (Figure 5A). However, the mutations in the nematode study (Denver et al., 2004) were specifically collected from mononucleotide repeats, which may artificially inflate the indel mutation rate estimate in that study (Konrad et al., 2019; Rajaei et al., 2021). The unicellular species had the lowest indel mutation rate, from 0.02 × 10⁻⁹ in yeast to 0.26 × 10⁻⁹ in algae, followed by arthropods with 0.55 [0.08, 1.24] × 10⁻⁹. Mammals had a relatively higher indel mutation rate with 1.13 [0.43, 2.05] × 10⁻⁹, with humans at the higher end of this range at 1.74 [1.16, 2.50] × 10⁻⁹. Of all the studied species, plants had the highest indel mutation rate, which was as high as 2.38 [0.51, 4.96] × 10⁻⁹.

Figure 4.

An analysis of short indel mutation rates for different species. The estimated mutation rates are presented on a log scale for clarity.

Figure 5.

The correlation between indel mutation rates and SNM rates plotted on a log scale with a 1:1 ratio shown as a dashed line (A). The aggregated log ratios of deletion/insertion in eukaryotes (B). SNM = single-nucleotide mutation.

Short deletions generally outnumbered short insertions across eukaryotes (Figure 5B, Supplementary Figure S3). The ratio of deletions over insertions was as high as 15.0 in the water flea, 6.0 in mouse, 3.4 in D. melanogaster, 2.6 in humans, 1.8 in Caenorhabditis elegans, and 1.6 in Arabidopsis thaliana. Honeybees, diatoms, and two plants (rice and peach) each exhibited more insertions than deletions, but the difference was not statistically significant (p > .05, binomial test). Fission yeast showed significantly more insertions overall.

A bias toward deletion events has been known for at least 40 years, since de Jong and Rydén (1981) found a fourfold excess of deletions over insertions in protein sequences. Petrov (2002) argued that the deletion bias was likely due to a thermodynamic disparity between short deletions and insertions that makes deletion events easier. However, studies in plants have suggested that unequal homologous recombination and illegitimate recombination may be another important cause of small deletions (Bennetzen et al., 2005). Whatever the cause, the dominance of deletions over insertions seems convincing, and many early studies argued that these short deletions resulted in genome contraction (Gregory, 2003; Petrov, 2002; Thomas et al., 2003). However, the role of deletions in shaping genome size remains controversial (Gregory, 2003, 2004). A recent study in birds and mammals indicated that short deletions alone cannot explain the DNA content loss (Kapusta et al., 2017), and we fail to find any evidence of a relationship between deletion rate and genome size here (Supplementary Figure S4; but note that power is likely to be low). It seems probable that other structural mutations, such as large deletions, gene duplication, and transposable elements, may have greater impacts on genome size. For example, in Drosophila, Wang et al. (2023) found transposable element insertions to be more frequent than SNMs, and an order of magnitude more frequent than short indels. Together, these studies suggest that the impact of short deletions on genome size is likely to be small, despite their higher frequency relative to short insertions.

Concluding remarks

All genetic variation arises from de novo mutations. To estimate the rate at which these arise, studies have moved from investigating a few genes that result in phenotypic change, to whole-genome sequencing of MA lines or families. Ten years ago, Campbell and Eichler (2013) called for more efforts to sequence genomes from nonhuman primate families “…to understand how the rate has changed in different lineages.” And, in the present data-rich and tool-rich era, we have now seen the sequencing of not just multiple non-human primates, but many taxa representing major clades across the tree of life. Nevertheless, the “tree of mutations” is far from complete, at either the large or small scale. Even within animals, the vast majority of phyla have never been examined, and with the exception of primates, no large groups of very close relatives have been examined. Even for vertebrates—and despite the recent large study by Bergeron et al. (2023) who examined 68 species—data are still scarce for reptiles and entirely lacking for amphibians.

However, as progress is being made toward a tree of mutation rates, attention still needs to focus on methodological heterogeneity. Different data filtering strategies may result in non-negligible differences in mutation rate estimates (Bergeron et al., 2022). Filters that are too conservative may discard some true mutations, while relaxed filters may inflate the false positive rate. Although many strategies have been proposed (Bergeron et al., 2022), no standardized pipeline has yet been agreed on or widely adopted. All studies must therefore carefully address the issues of false positives by manual curation and/or PCR follow-up, and false negatives by simulation. Here, we did not examine the impact of the false positive rate or false negative rate, but we remind the readers of the potential impact of methodological difference and caution that attention should be paid to the methods or parameters used in each study when comparisons are made.

Another potential source of bias may also come from the specific choice of samples used in each study. Besides the biological factors that affect mutation rates (e.g., age, sex, genetic background), environmental factors can also influence mutation rates, for example, temperature (Belfield et al., 2021; Waldvogel & Pfenninger, 2021), osmotic stress (Hasan et al., 2022), or ultraviolet light (Xu et al., 2019). Samples collected from the wild rather than reared in laboratories may better reflect the “real” mutation rate, but the variation that this introduces may also make it harder to draw broad conclusions. Despite these challenges, the data to come are exciting. In particular, as sequencing technology advances, we expect improved accuracy in long-read sequencing facilitating the phasing of mutations and permitting the easy detection of complex mutations and larger-scale structural changes.

Data availability

The data sets and the code to analyze the data are available on GitHub: https://github.com/Yiguan/mutation_literature.

Author contributions

Y.W. participated in data collection, performed analyses, and wrote the first draft. D.J.O. wrote a part of the first draft and made extensive contributions to the development, drafting, and editing of the manuscript and supporting material, as well as model explanation. All authors gave final approval for publication.

Conflict of interest: The authors declare no conflict of interest.