Experimental Evolution with Caenorhabditis Nematodes

The hermaphroditic nematode Caenorhabditis elegans has been one of the primary model systems in biology since the 1970s, but only within the last two decades has this nematode also become a useful model for experimental evolution. Here, we outline the goals and major foci of experimental evolution with C. elegans and related species, such as C. briggsae and C. remanei, by discussing the principles of experimental design, and highlighting the strengths and limitations of Caenorhabditis as model systems. We then review three exemplars of Caenorhabditis experimental evolution studies, underlining representative evolution experiments that have addressed the: (1) maintenance of genetic variation; (2) role of natural selection during transitions from outcrossing to selfing, as well as the maintenance of mixed breeding modes during evolution; and (3) evolution of phenotypic plasticity and its role in adaptation to variable environments, including host–pathogen coevolution. We conclude by suggesting some future directions for which experimental evolution with Caenorhabditis would be particularly informative.

achieve this is not a trivial problem (LYNCH and WALSH 1998). Mutational variances (VM), and genetic variances in general (VG), can be scaled in two ways, each of which having advantages and limitations. Traditionally, VM or VG is scaled relative to the within-line variance of the trait that has a stochastic environmental origin (VE), due to measurement error and developmental instability. The ratio VM/VE is called the "mutational heritability", and is analogous to the "narrow-sense heritability" of a trait, VA/VE (where VA is the additive genetic variance, i.e. variation among individuals in their breeding value). The narrow-sense heritability predicts the extent to which the trait will respond to selection in the short-term from standing genetic variation, and similarly the mutational heritability predicts selection response from a situation without standing genetic variation. Alternatively, VM and VA can be scaled relative to the square of the trait mean, VM/ ̅ 2 , which provides a sense for the long-term evolvability of the trait (HOULE 1992). In most circumstances scaling by the mean rather than by the variance is preferable , but in some circumstances scaling by the mean is not meaningful, as when the trait can take on values that are either positive or negative.
In MA experiments, it is customary to divide the trait variance of derived lines by the trait mean of the ancestral because this ancestral is fixed. Likewise, in EE from standing genetic variation, the trait mean of derived populations can be divided by the ancestral mean. But in cases where the trait mean evolves substantially, dividing by the ancestral mean can give an unrealistically small estimate of the within population variance (VE). In addition, since the ancestor mean trait value is taken as fixed, initial sampling of individuals to begin EE, particularly from standing genetic variation, must be large in order to avoid idiosyncratic responses. One solution is to divide each group (MA lines or derived populations and ancestor) by its own mean, which is equivalent to assuming that mutational or allelic trait effects are multiplicative. An alternative strategy is to log-transform the data before calculating the means and variances, which is almost equivalent to scaling by the treatment mean (FRY and HEINSOHN 2002). Yet an alternative is to repeat EE from several ancestors with different past histories, though how many are sufficient is an open question (WHITLOCK et al. 2002). Further note that having one or few ancestors necessarily makes experimental design unbalanced. Houle, Hansen, and their colleagues have written thoughtfully and in considerable depth about scaling in the context of evolutionary biology, and we strongly recommend that students of EE read their work on measurement theory ).
Yet another issue to consider in the design and interpretation of evolution experiments is the ability to detect low frequency genetic variants. The most straightforward and unbiased way to detect low frequency variants in multicellular organisms is to sequence the genomic DNA of multiple gametes taken from single individuals. Since the transcripts expressed in gametes are provided by the parent, the gamete's own genome is typically invisible to selection and thus any mutations present in the gamete genome are neutral to a first approximation and the mutational spectrum should be unbiased. However, there are practical difficulties involved in sequencing whole genomes of individual gametes. Isolating individual gametes is technically challenging and it is not yet possible to obtain whole-genome sequence from a single cell without an initial amplification step. The most accurate Taq polymerase has a per-nucleotide error rate on the order of 10 -6 , whereas the per-nucleotide and per-generation mutation rate is on the order of 10 -8 or less. PCR errors will thus outnumber real mutations by at least hundred-fold, but probably by much more. This problem can be avoided in principle by splitting the sample prior to the amplification step (if the amplification step does not itself introduce biases), but doing so obviously requires that each genome be sequenced twice. This can actually be accomplished fairly easily using an overlapping pair-end approach, which has been recently shown to be sensitive to very low frequency alleles in a C. remanei EE population (PRESTON et al. 2016).
Another method to detect low frequency variants is to sequence parents and offspring, or more generally, relatives within a pedigree that can be inferred. This is the technique by which human mutational properties have been characterized (CONRAD et al. 2011), and it can be done with individual Caenorhabditis, but an initial amplification step is still necessary to obtain enough material for whole-genome sequencing. A complicating factor is that, depending on when during the ontogeny of the parental germline a particular mutation occurred, the offspring may be either homozygous or heterozygous for a new mutation. The possibility that a new mutation segregates as a heterozygote adds the additional complication of binomial sampling. At any individual heterozygous nucleotide position, the probability of not sampling one of the two alleles is the binomial probability, = ( 0 ) (1 − ) , where N is the number of times the nucleotide position is represented in the data (the coverage) and p is 0.5 (the frequency of the allele). To achieve a 95% probability that no heterozygous site is misidentified as homozygousor in other words, (1-u) X > 0.95, where X is 10 8 -a genome of 10 8 bases needs to be sequenced to at least 31X minimum coverage, or at least 36X to identify each allele at least twice, and thereby be able to distinguish a true mutation from a sequencing error.
One last consideration that is relevant in any kind of evolution experiment is how many genomes are actually being sequenced? As noted, sequencing single individuals, and thus single genomes, requires an amplification step. To date, sequencing     (LYNCH 1993;LYNCH et al. 1995).
In the era of whole-genome sequencing, MA experiments provide a costeffective way to infer mutational properties at the level of the genome itself. (DENVER et al. 2000) reported the first direct estimate of the mutational properties of the mitochondrial genome in any organism, based on sequencing the mtDNA of a set of N2 strain MA lines. In 2004, the same group reported the first direct estimate of the nuclear genome-wide mutation rate in any multicellular organism (DENVER et al. 2004b).
Additional studies have built on that work, including characterizations of the mtDNA and nuclear base-substitution spectrum in other strains and species DENVER et al. 2012) and features of the genome beyond base-substitutions, including short tandem repeats (DENVER et al. 2004a;SEYFERT et al. 2008; and with hermaphrodites or females XX and males X0. Males can be produced by mutation through the non-disjunction of the X-chromosome during hermaphrodite gametogenesis (NIGON 1949;HODGKIN et al. 1979), but their appearance in natural isolates is relatively rare, on the order of 10 -3 -10 -4 (TEOTÓNIO et al. 2006;TEOTÓNIO et al. 2012).
Population genetic dynamics of multiple loci under (partial) selfing are highly complex and still poorly understood when they depend on the degree of dominance, epistasis, and linked selection (e.g., (WEIR et al. 1980;ZIEHE and ROBERDS 1989;ROZE 2015;ROZE 2016). For neutral loci unlinked to any selected alleles, the effective population size under partial selfing and at genetic drift-mutation equilibrium is expected to be Ne = N/(1+F)=N(2-S)/2 (POLLAK 1987;NORDBORG 2000), where N is population size. Selfing is expected to halve heterozygosity according to Ht = Ht0 (1-S) + S/2 (Ht-1), with Ht0 being the observed heterozygosity of the population at time t=0 before selfing started. Heterozygosity will only be a fraction 1-F = 1-1/2Ne =1-1/N(2-S) of the ancestral levels, with great loss of heterozygosity and concomitant increase in homozygosity occurring only if relatively high selfing rates are maintained for long periods. At the genome-wide level, selfing increases "identity disequilibrium" between loci, by increasing homozygosity correlations among any two neutral loci as: where =1-2r and r recombination rate (WEIR and COCKERHAM 1973;CHRISTIANSEN 1989).
As a consequence of increased homozygosity at multiple loci, selfing is also expected to decrease the "effective" recombination rates (rs) by the fraction 1-1/N(2-S) (NORDBORG 1997;NORDBORG 2000). An important application of this model is to measure historical selfing rates (NORDBORG and DONNELLY 1997), which has indicated that natural populations of C. elegans have predominantly inbred by selfing for thousands of generations (BARRIERE and FELIX 2005). In particular, assuming an equilibrium between recombination and genetic drift, the expected average correlation of allele diversity between two neutral loci-r 2 , a linkage disequilibrium metric that can be estimated by sampling genotypes-is E(r 2 )=1/1+4Ners (HUDSON  Similar to its effect on segregation and recombination, selfing is expected to reduce the overall size of the G-matrix --the matrix whose entries are the additive genetic variances and covariances between traits (LANDE 1980;PHILLIPS and MCGUIGAN 2006) --by a fraction 1-F=1-1/N(2-S) of the ancestral size while causing no change in G-matrix orientation on average. After a period of selfing sufficient to achieve complete homozygosity, additive genetic variances will be composed of all but strictly genetic variance that can be estimated from trait differentiation among selfing lineages (LYNCH and WALSH 1998). However, the expected genetic variance between selfing populations differentiating only by genetic drift will be 2F=2/N(2-S) times the genetic variation of the ancestral population (assumed to be panmictic and outcrossing); again reflecting the higher segregation of homozygotes when compared to randomly mating populations. This has the important consequence for EE design since more replication is necessary under selfing than outcrossing to detect an evolutionary trait response of similar magnitude.