Did Genetic Drift Drive Increases in Genome Complexity?

Summary

Mechanisms underlying the dramatic patterns of genome size variation across the tree of life remain mysterious. Effective population size (N_e) has been proposed as a major driver of genome size: selection is expected to efficiently weed out deleterious mutations increasing genome size in lineages with large (but not small) N_e. Strong support for this model was claimed from a comparative analysis of N_eu and genome size for ≈30 phylogenetically diverse species ranging from bacteria to vertebrates, but analyses at that scale have so far failed to account for phylogenetic nonindependence of species. In our reanalysis, accounting for phylogenetic history substantially altered the perceived strength of the relationship between N_eu and genomic attributes: there were no statistically significant associations between N_eu and gene number, intron size, intron number, the half-life of gene duplicates, transposon number, transposons as a fraction of the genome, or overall genome size. We conclude that current datasets do not support the hypothesis of a mechanistic connection between N_e and these genomic attributes, and we suggest that further progress requires larger datasets, phylogenetic comparative methods, more robust estimators of genetic drift, and a multivariate approach that accounts for correlations between putative explanatory variables.

Introduction

The vast array of genome sizes is a pattern that begs for explanation [1], [2]. Haploid (1C) genome size (measured either in base pairs or mass, where 10⁶ Kb ≈1 picogram) spans eight orders of magnitude: the known eukaryotic range is ≈2,249–978,000,000 Kb [3], while Archaea and Bacteria range from 491–5,751 Kb and 76–13,034 Kb, respectively [4].

Lynch and colleagues [5]–[7] have argued strongly for a central role for nonadaptive processes such as mutation and drift in the evolution of genome size and complexity. In contrast to proposed neutral and adaptive models of genome size evolution (see, e.g. [8], [9]), they outline a model positing that mutations increasing genome size are slightly deleterious. Under this model, lineages differ in effective population size (N_e) and, as a result, differ in the efficacy with which natural selection will counteract genome expansion. Thus, lineages with small N_e will experience drift towards larger genomes [7]. As support for their argument, they presented a comparative analysis of roughly 30 taxa, ranging from bacteria to angiosperms, fungi, and mammals. Among these taxa, they reported a statistically significant negative relationship between N_eu (a composite parameter including effective population size and nucleotide mutation rate) and genome size. Strikingly, the relationship was quite strong: 66% of the variation in genome size was explained by N_eu [7]. This is truly an astounding result, considering the widely divergent selective regimes, life histories, and modes of reproduction found across these diverse organisms.

The Lynch & Conery model has sparked intense interest and >330 citations. Some objections on theoretical and methodological grounds have been voiced. Charlesworth and Barton [10] point out that N_e is confounded with many different aspects of organismal biology (e.g., developmental rate, body size), and thus that both N_e and genome size may be correlated effects of one or more other causal factors. Daubin and Moran [11] outline several objections, including that taxon differences in mutation rates make N_eu a poor proxy for N_e that estimates of N_e from silent-site nucleotide diversity in bacteria (as in [7]) are skewed by population subdivision and cryptic species, and further that such N_e estimates are overly sensitive to recent evolutionary history. Nevertheless, the idea that N_e drives genome size and complexity seems to have gained acceptance [12]–[14], with some going so far as to characterize it as “the principal explanatory framework for understanding the evolution of genome organization” ([12], p. 303).

Here, we argue that such conclusions are premature without phylogenetic comparative analyses of genome size evolution. When species are used as data points, relationships between raw values of any two traits (e.g., N_e and genome size) are difficult to interpret, as shared phylogenetic history means that assumptions of statistical independence are likely to be violated [15]–[17]. Special methods are required to recover independence of observations and to test for evolutionary associations between traits. Frequently, conventional (nonphylogenetic) analyses overestimate the strength of the association between traits relative to phylogenetic methods [18]. In an extreme case, a strong correlation in the raw data can be driven by a single association at the base of the phylogenetic tree, e.g., it can reflect a single instance of correlated change in the traits, followed by uncorrelated changes and/or stasis in trait values during subsequent evolutionary history (Figure 1). In this study, we revisit the Lynch & Conery dataset with a phylogenetic perspective, taking advantage of new phylogenetic data and analysis tools.

Results

Model fitting

A phylogenetic topology and reconstruction of genome sizes is presented in Figure 2, illustrating that close relatives have similar genome sizes. Initial simple linear regressions of genome size on N_eu explored four branch length models and found that the phylogenetic generalized least squares (PGLS) model with all branches  = 1.0 provided a better fit than the nonphylogenetic ordinary least squares (OLS) model (Table 1). Subsequent analyses therefore used branch lengths of 1.0. For all variables except intron number, phylogenetic models (PGLS) exhibited better fit than nonphylogenetic (OLS) models (Table 1). For genome size and gene number, estimation of the Ornstein-Uhlenbeck transformation parameter d indicated substantial phylogenetic signal (d = 1.31 and 1.16, respectively), and the resulting RegOU models fit significantly better than the OLS models (ln likelihood ratio tests (LRTs), χ² = 5.88, P = 0.015 and χ² = 7.90, P = 0.005, respectively). In comparing the two phylogenetic models, the RegOU model did not produce significantly better fit vs. PGLS (LRTs, χ² = 1.84, P = 0.175 and χ² = 0.46, P = 0.498 for genome size and gene number, respectively).

Phylogenetic regressions do not detect relationships between N_eu and genomic attributes

Although there were strong negative relationships between N_eu and six of the seven genomic attributes in nonphylogenetic regressions, the patterns disappeared when phylogenetic models were applied (Table 1). For example, the strong negative relationship between N_eu and genome size (OLS, P<0.001, Figure 3A) was replaced with a nonsignificant relationship under better-fitting phylogenetic models (PGLS, P = 0.137, Figure 3B; RegOU, P = 0.328). Similar patterns were evident for gene number, the half-life of gene duplicates, intron size, intron number, transposon number, and transposon fraction (Table 1).

Discussion

Accounting for phylogenetic history substantially altered the perceived strength of the relationship between N_eu and genomic attributes. In phylogenetic analyses, there were no consistent evolutionary associations between N_eu and gene number, intron size, intron number, the half-life of gene duplicates, transposon number, transposons as a fraction of the genome, or overall genome size. Thus, a phylogenetically controlled reanalysis of the Lynch & Conery dataset [7] does not support the conclusion that N_e drives genome size patterns across the tree of life.

The few existing comparative analyses of more phylogenetically restricted datasets either do not support or provide only equivocal support for the Lynch & Conery model. Whitney et al. [19] conducted a phylogenetically controlled analysis of 205 species of seed plants and found no association between N_e and genome size. Kuo et al. [20] analyzed 42 paired bacterial genomes, using the efficacy of purifying selection in coding regions to quantify genetic drift. Bacterial taxa experiencing greater levels of genetic drift – implying a smaller evolutionary N_e – had smaller genomes, a pattern opposite that predicted by the Lynch & Conery model as articulated in [7]. Finally, in putative support of the model, Yi & Streelman [21] reported a significant negative relationship between N_e and genome size in a phylogenetically corrected analysis of 33 species of ray-finned fish. However, this analysis has been challenged as artifactual. Gregory & Witt [22] argue that Pleistocene population bottlenecks and polyploidy shaped both N_e and genome size of fishes in such a way as to generate a non-causal correlation between N_e and genome size in this particular dataset.

Future investigations of the role of genetic drift in determining genome size across the tree of life would benefit from several approaches. First, utilizing phylogenetic comparative methods, for which we advocate here, is an important step towards drawing robust inferences from species-level comparative analyses. Second, larger datasets would certainly increase confidence in our interpretations. While statistically nonsignificant, we note the relationships between N_eu and genomic attributes (Table 1) are negative and thus are at least qualitatively consistent with the Lynch & Conery model, suggesting that power may be an issue. Furthermore, given that the N_eu estimates in the current analysis required sequence data, species with small genomes relative to averages within clades are likely overrepresented; thus it would be important to ensure that species with large genomes are included in future analyses. Third, future studies would benefit from more robust estimates of genetic drift, as N_eu estimated from silent-site diversity (as in [7] and the present reanalysis) has several undesirable properties. Because the mutation rate u differs among lineages [11], [23], [24], using N_eu as a proxy for N_e could obscure any relationship between N_e and genome size. Further, N_e estimated from silent-site diversity may signal the effects of recent evolutionary events more than the long-term history under which genome size evolved [11]. K_a/K_s ratios (ratios of nonsynonymous to synonymous substitutions per site) are a promising alternative to N_eu for estimating genetic drift [11], [20]. Finally, genome size is a complex trait that is unlikely to be explained by univariate analyses [10]. Phylogenetic comparative methods should be combined with multivariate models that are capable of distinguishing the contributions of highly correlated predictor variables. A recent analysis [19] is a step in the right direction: plant outcrossing rate and N_e were simultaneously examined in a multiple regression analysis of phylogenetically independent contrasts, allowing the partial contribution of each variable to be characterized. To make further progress on the population genetics of genome size and complexity, we clearly need phylogenetic comparative analyses of large datasets capable of distinguishing the contributions of N_e and its multiple correlates, including body size, developmental rate, and metabolic rate.

Materials and Methods

Data sources

Data on N_eu and genome sizes for 22 eukaryotic and 7 prokaryotic species were obtained from the Supporting Online Material of [7]. For a subset of these species, data on gene number, intron size, intron number, and the half-life of gene duplicates were also obtained from the same source. Data on total transposon number and fraction of the genome occupied by transposons were obtained directly from M. Lynch; these data combine counts of LTR, non-LTR, and DNA transposons and correspond to the fourth panel of Fig. 4 of [7]. All traits were log₁₀ transformed prior to analysis; for total transposon number and transposon fraction, constants of 1.0 and 0.01, respectively, were added prior to log-transformation.

Phylogeny construction

A composite tree for the species was constructed in Mesquite v. 2.71 [25] based on phylogenetic trees reported in [26]–[28]. As a visual heuristic, genome sizes were traced onto the phylogeny using the Parsimony Ancestral States method [29] with an assumption that all branch lengths equal 1.0.

Phylogenetic comparative analyses

All dependent variables were regressed on N_eu using REGRESSIONv2.m [30] running in MATLAB v. 7.9.0. Three types of models were examined: ordinary least squares (OLS), phylogenetic generalized least squares (PGLS), and phylogenetic regression under an Ornstein-Uhlenbeck process (RegOU) [30], [31]. OLS is traditional ‘nonphylogenetic’ regression, which in effect assumes a star phylogeny in which all species are equally unrelated, and corresponds to the N_eu vs. genome size analysis reported in [7]. PGLS assumes that residual variation among species is correlated, with the correlation given by a Brownian-motion like process along the specified phylogenetic tree (topology and branch lengths). PGLS is functionally equivalent to Felsenstein's [15] phylogenetically independent contrast method [31]. Finally, the RegOU model estimates (via restricted maximum likelihood) the strength of phylogenetic signal in the residual variation simultaneously with the regression coefficients; the former is given by d, the Ornstein-Uhlenbeck transformation parameter. An OU evolutionary model is typically used to model the effects of stabilizing selection around an optimum [30]. When d = 0, there is no phylogenetic signal in the residuals from the regression model; when d is significantly greater than 0, significant phylogenetic signal exists [30], [32].

Following [33], starter branch lengths corresponding to all branches  = 1.0, Grafen's arbitrary lengths, Pagel's arbitrary lengths, and Nee's arbitrary lengths were compared in PGLS and RegOU regressions of genome size on N_eu. Based on their likelihoods, the models with all branches  = 1.0 achieved the best fit, and thus these branch lengths were used in all subsequent phylogenetic analyses. Model selection for each variable then proceeded in two steps. First, we compared the likelihoods of the PGLS model and the OLS model, with a higher likelihood taken as evidence of a better-fitting model. Second, we used ln likelihood ratio tests (LRTs) to compare the RegOU model with the PGLS and OLS models with 1 d.f. [30]. Given the issue of small sample sizes (see [32]) for most dependent variables and the fact that RegOU models require estimation of an extra parameter, RegOU models were examined only for genome size and gene number.